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 2011-2012 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
60 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
61 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
67 #include <semaphore.h>
72 #include <valgrind/memcheck.h>
73 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
74 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
75 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
76 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
77 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
79 #define VGMEMP_CREATE(h,r,z)
80 #define VGMEMP_ALLOC(h,a,s)
81 #define VGMEMP_FREE(h,a)
82 #define VGMEMP_DESTROY(h)
83 #define VGMEMP_DEFINED(a,s)
87 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
88 /* Solaris just defines one or the other */
89 # define LITTLE_ENDIAN 1234
90 # define BIG_ENDIAN 4321
91 # ifdef _LITTLE_ENDIAN
92 # define BYTE_ORDER LITTLE_ENDIAN
94 # define BYTE_ORDER BIG_ENDIAN
97 # define BYTE_ORDER __BYTE_ORDER
101 #ifndef LITTLE_ENDIAN
102 #define LITTLE_ENDIAN __LITTLE_ENDIAN
105 #define BIG_ENDIAN __BIG_ENDIAN
108 #if defined(__i386) || defined(__x86_64)
109 #define MISALIGNED_OK 1
115 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
116 # error "Unknown or unsupported endianness (BYTE_ORDER)"
117 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
118 # error "Two's complement, reasonably sized integer types, please"
121 /** @defgroup internal MDB Internals
124 /** @defgroup compat Windows Compatibility Macros
125 * A bunch of macros to minimize the amount of platform-specific ifdefs
126 * needed throughout the rest of the code. When the features this library
127 * needs are similar enough to POSIX to be hidden in a one-or-two line
128 * replacement, this macro approach is used.
132 #define pthread_t DWORD
133 #define pthread_mutex_t HANDLE
134 #define pthread_key_t DWORD
135 #define pthread_self() GetCurrentThreadId()
136 #define pthread_key_create(x,y) (*(x) = TlsAlloc())
137 #define pthread_key_delete(x) TlsFree(x)
138 #define pthread_getspecific(x) TlsGetValue(x)
139 #define pthread_setspecific(x,y) TlsSetValue(x,y)
140 #define pthread_mutex_unlock(x) ReleaseMutex(x)
141 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
142 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
143 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
144 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
145 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
146 #define getpid() GetCurrentProcessId()
147 #define fdatasync(fd) (!FlushFileBuffers(fd))
148 #define ErrCode() GetLastError()
149 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
150 #define close(fd) CloseHandle(fd)
151 #define munmap(ptr,len) UnmapViewOfFile(ptr)
154 #define LOCK_MUTEX_R(env) sem_wait((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) sem_wait((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
158 #define fdatasync(fd) fsync(fd)
161 #define fdatasync(fd) fsync(fd)
163 /** Lock the reader mutex.
165 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
166 /** Unlock the reader mutex.
168 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
170 /** Lock the writer mutex.
171 * Only a single write transaction is allowed at a time. Other writers
172 * will block waiting for this mutex.
174 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
175 /** Unlock the writer mutex.
177 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
178 #endif /* __APPLE__ */
180 /** Get the error code for the last failed system function.
182 #define ErrCode() errno
184 /** An abstraction for a file handle.
185 * On POSIX systems file handles are small integers. On Windows
186 * they're opaque pointers.
190 /** A value for an invalid file handle.
191 * Mainly used to initialize file variables and signify that they are
194 #define INVALID_HANDLE_VALUE (-1)
196 /** Get the size of a memory page for the system.
197 * This is the basic size that the platform's memory manager uses, and is
198 * fundamental to the use of memory-mapped files.
200 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
203 #if defined(_WIN32) || defined(__APPLE__)
210 /** A flag for opening a file and requesting synchronous data writes.
211 * This is only used when writing a meta page. It's not strictly needed;
212 * we could just do a normal write and then immediately perform a flush.
213 * But if this flag is available it saves us an extra system call.
215 * @note If O_DSYNC is undefined but exists in /usr/include,
216 * preferably set some compiler flag to get the definition.
217 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
220 # define MDB_DSYNC O_DSYNC
224 /** Function for flushing the data of a file. Define this to fsync
225 * if fdatasync() is not supported.
227 #ifndef MDB_FDATASYNC
228 # define MDB_FDATASYNC fdatasync
231 /** A page number in the database.
232 * Note that 64 bit page numbers are overkill, since pages themselves
233 * already represent 12-13 bits of addressable memory, and the OS will
234 * always limit applications to a maximum of 63 bits of address space.
236 * @note In the #MDB_node structure, we only store 48 bits of this value,
237 * which thus limits us to only 60 bits of addressable data.
241 /** A transaction ID.
242 * See struct MDB_txn.mt_txnid for details.
246 /** @defgroup debug Debug Macros
250 /** Enable debug output.
251 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
252 * read from and written to the database (used for free space management).
257 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
258 # define DPRINTF (void) /* Vararg macros may be unsupported */
260 /** Print a debug message with printf formatting. */
261 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
262 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
264 # define DPRINTF(fmt, ...) ((void) 0)
266 /** Print a debug string.
267 * The string is printed literally, with no format processing.
269 #define DPUTS(arg) DPRINTF("%s", arg)
272 /** A default memory page size.
273 * The actual size is platform-dependent, but we use this for
274 * boot-strapping. We probably should not be using this any more.
275 * The #GET_PAGESIZE() macro is used to get the actual size.
277 * Note that we don't currently support Huge pages. On Linux,
278 * regular data files cannot use Huge pages, and in general
279 * Huge pages aren't actually pageable. We rely on the OS
280 * demand-pager to read our data and page it out when memory
281 * pressure from other processes is high. So until OSs have
282 * actual paging support for Huge pages, they're not viable.
284 #define MDB_PAGESIZE 4096
286 /** The minimum number of keys required in a database page.
287 * Setting this to a larger value will place a smaller bound on the
288 * maximum size of a data item. Data items larger than this size will
289 * be pushed into overflow pages instead of being stored directly in
290 * the B-tree node. This value used to default to 4. With a page size
291 * of 4096 bytes that meant that any item larger than 1024 bytes would
292 * go into an overflow page. That also meant that on average 2-3KB of
293 * each overflow page was wasted space. The value cannot be lower than
294 * 2 because then there would no longer be a tree structure. With this
295 * value, items larger than 2KB will go into overflow pages, and on
296 * average only 1KB will be wasted.
298 #define MDB_MINKEYS 2
300 /** A stamp that identifies a file as an MDB file.
301 * There's nothing special about this value other than that it is easily
302 * recognizable, and it will reflect any byte order mismatches.
304 #define MDB_MAGIC 0xBEEFC0DE
306 /** The version number for a database's file format. */
307 #define MDB_VERSION 1
309 /** The maximum size of a key in the database.
310 * While data items have essentially unbounded size, we require that
311 * keys all fit onto a regular page. This limit could be raised a bit
312 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
314 #define MAXKEYSIZE 511
319 * This is used for printing a hex dump of a key's contents.
321 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
322 /** Display a key in hex.
324 * Invoke a function to display a key in hex.
326 #define DKEY(x) mdb_dkey(x, kbuf)
328 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
332 /** @defgroup lazylock Lazy Locking
333 * Macros for locks that are't actually needed.
334 * The DB view is always consistent because all writes are wrapped in
335 * the wmutex. Finer-grained locks aren't necessary.
339 /** Use lazy locking. I.e., don't lock these accesses at all. */
343 /** Grab the reader lock */
344 #define LAZY_MUTEX_LOCK(x)
345 /** Release the reader lock */
346 #define LAZY_MUTEX_UNLOCK(x)
347 /** Release the DB table reader/writer lock */
348 #define LAZY_RWLOCK_UNLOCK(x)
349 /** Grab the DB table write lock */
350 #define LAZY_RWLOCK_WRLOCK(x)
351 /** Grab the DB table read lock */
352 #define LAZY_RWLOCK_RDLOCK(x)
353 /** Declare the DB table rwlock. Should not be followed by ';'. */
354 #define LAZY_RWLOCK_DEF(x)
355 /** Initialize the DB table rwlock */
356 #define LAZY_RWLOCK_INIT(x,y)
357 /** Destroy the DB table rwlock */
358 #define LAZY_RWLOCK_DESTROY(x)
360 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
361 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
362 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
363 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
364 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
365 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x;
366 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
367 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
371 /** An invalid page number.
372 * Mainly used to denote an empty tree.
374 #define P_INVALID (~0UL)
376 /** Test if a flag \b f is set in a flag word \b w. */
377 #define F_ISSET(w, f) (((w) & (f)) == (f))
379 /** Used for offsets within a single page.
380 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
383 typedef uint16_t indx_t;
385 /** Default size of memory map.
386 * This is certainly too small for any actual applications. Apps should always set
387 * the size explicitly using #mdb_env_set_mapsize().
389 #define DEFAULT_MAPSIZE 1048576
391 /** @defgroup readers Reader Lock Table
392 * Readers don't acquire any locks for their data access. Instead, they
393 * simply record their transaction ID in the reader table. The reader
394 * mutex is needed just to find an empty slot in the reader table. The
395 * slot's address is saved in thread-specific data so that subsequent read
396 * transactions started by the same thread need no further locking to proceed.
398 * Since the database uses multi-version concurrency control, readers don't
399 * actually need any locking. This table is used to keep track of which
400 * readers are using data from which old transactions, so that we'll know
401 * when a particular old transaction is no longer in use. Old transactions
402 * that have discarded any data pages can then have those pages reclaimed
403 * for use by a later write transaction.
405 * The lock table is constructed such that reader slots are aligned with the
406 * processor's cache line size. Any slot is only ever used by one thread.
407 * This alignment guarantees that there will be no contention or cache
408 * thrashing as threads update their own slot info, and also eliminates
409 * any need for locking when accessing a slot.
411 * A writer thread will scan every slot in the table to determine the oldest
412 * outstanding reader transaction. Any freed pages older than this will be
413 * reclaimed by the writer. The writer doesn't use any locks when scanning
414 * this table. This means that there's no guarantee that the writer will
415 * see the most up-to-date reader info, but that's not required for correct
416 * operation - all we need is to know the upper bound on the oldest reader,
417 * we don't care at all about the newest reader. So the only consequence of
418 * reading stale information here is that old pages might hang around a
419 * while longer before being reclaimed. That's actually good anyway, because
420 * the longer we delay reclaiming old pages, the more likely it is that a
421 * string of contiguous pages can be found after coalescing old pages from
422 * many old transactions together.
424 * @todo We don't actually do such coalescing yet, we grab pages from one
425 * old transaction at a time.
428 /** Number of slots in the reader table.
429 * This value was chosen somewhat arbitrarily. 126 readers plus a
430 * couple mutexes fit exactly into 8KB on my development machine.
431 * Applications should set the table size using #mdb_env_set_maxreaders().
433 #define DEFAULT_READERS 126
435 /** The size of a CPU cache line in bytes. We want our lock structures
436 * aligned to this size to avoid false cache line sharing in the
438 * This value works for most CPUs. For Itanium this should be 128.
444 /** The information we store in a single slot of the reader table.
445 * In addition to a transaction ID, we also record the process and
446 * thread ID that owns a slot, so that we can detect stale information,
447 * e.g. threads or processes that went away without cleaning up.
448 * @note We currently don't check for stale records. We simply re-init
449 * the table when we know that we're the only process opening the
452 typedef struct MDB_rxbody {
453 /** The current Transaction ID when this transaction began.
454 * Multiple readers that start at the same time will probably have the
455 * same ID here. Again, it's not important to exclude them from
456 * anything; all we need to know is which version of the DB they
457 * started from so we can avoid overwriting any data used in that
458 * particular version.
461 /** The process ID of the process owning this reader txn. */
463 /** The thread ID of the thread owning this txn. */
467 /** The actual reader record, with cacheline padding. */
468 typedef struct MDB_reader {
471 /** shorthand for mrb_txnid */
472 #define mr_txnid mru.mrx.mrb_txnid
473 #define mr_pid mru.mrx.mrb_pid
474 #define mr_tid mru.mrx.mrb_tid
475 /** cache line alignment */
476 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
480 /** The header for the reader table.
481 * The table resides in a memory-mapped file. (This is a different file
482 * than is used for the main database.)
484 * For POSIX the actual mutexes reside in the shared memory of this
485 * mapped file. On Windows, mutexes are named objects allocated by the
486 * kernel; we store the mutex names in this mapped file so that other
487 * processes can grab them. This same approach is also used on
488 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
489 * process-shared POSIX mutexes. For these cases where a named object
490 * is used, the object name is derived from a 64 bit FNV hash of the
491 * environment pathname. As such, naming collisions are extremely
492 * unlikely. If a collision occurs, the results are unpredictable.
494 typedef struct MDB_txbody {
495 /** Stamp identifying this as an MDB lock file. It must be set
498 /** Version number of this lock file. Must be set to #MDB_VERSION. */
499 uint32_t mtb_version;
500 #if defined(_WIN32) || defined(__APPLE__)
501 char mtb_rmname[MNAME_LEN];
503 /** Mutex protecting access to this table.
504 * This is the reader lock that #LOCK_MUTEX_R acquires.
506 pthread_mutex_t mtb_mutex;
508 /** The ID of the last transaction committed to the database.
509 * This is recorded here only for convenience; the value can always
510 * be determined by reading the main database meta pages.
513 /** The number of slots that have been used in the reader table.
514 * This always records the maximum count, it is not decremented
515 * when readers release their slots.
517 unsigned mtb_numreaders;
518 /** The ID of the most recent meta page in the database.
519 * This is recorded here only for convenience; the value can always
520 * be determined by reading the main database meta pages.
522 uint32_t mtb_me_toggle;
525 /** The actual reader table definition. */
526 typedef struct MDB_txninfo {
529 #define mti_magic mt1.mtb.mtb_magic
530 #define mti_version mt1.mtb.mtb_version
531 #define mti_mutex mt1.mtb.mtb_mutex
532 #define mti_rmname mt1.mtb.mtb_rmname
533 #define mti_txnid mt1.mtb.mtb_txnid
534 #define mti_numreaders mt1.mtb.mtb_numreaders
535 #define mti_me_toggle mt1.mtb.mtb_me_toggle
536 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
539 #if defined(_WIN32) || defined(__APPLE__)
540 char mt2_wmname[MNAME_LEN];
541 #define mti_wmname mt2.mt2_wmname
543 pthread_mutex_t mt2_wmutex;
544 #define mti_wmutex mt2.mt2_wmutex
546 char pad[(sizeof(pthread_mutex_t)+CACHELINE-1) & ~(CACHELINE-1)];
548 MDB_reader mti_readers[1];
552 /** Common header for all page types.
553 * Overflow records occupy a number of contiguous pages with no
554 * headers on any page after the first.
556 typedef struct MDB_page {
557 #define mp_pgno mp_p.p_pgno
558 #define mp_next mp_p.p_next
560 pgno_t p_pgno; /**< page number */
561 void * p_next; /**< for in-memory list of freed structs */
564 /** @defgroup mdb_page Page Flags
566 * Flags for the page headers.
569 #define P_BRANCH 0x01 /**< branch page */
570 #define P_LEAF 0x02 /**< leaf page */
571 #define P_OVERFLOW 0x04 /**< overflow page */
572 #define P_META 0x08 /**< meta page */
573 #define P_DIRTY 0x10 /**< dirty page */
574 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
575 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
577 uint16_t mp_flags; /**< @ref mdb_page */
578 #define mp_lower mp_pb.pb.pb_lower
579 #define mp_upper mp_pb.pb.pb_upper
580 #define mp_pages mp_pb.pb_pages
583 indx_t pb_lower; /**< lower bound of free space */
584 indx_t pb_upper; /**< upper bound of free space */
586 uint32_t pb_pages; /**< number of overflow pages */
588 indx_t mp_ptrs[1]; /**< dynamic size */
591 /** Size of the page header, excluding dynamic data at the end */
592 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
594 /** Address of first usable data byte in a page, after the header */
595 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
597 /** Number of nodes on a page */
598 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
600 /** The amount of space remaining in the page */
601 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
603 /** The percentage of space used in the page, in tenths of a percent. */
604 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
605 ((env)->me_psize - PAGEHDRSZ))
606 /** The minimum page fill factor, in tenths of a percent.
607 * Pages emptier than this are candidates for merging.
609 #define FILL_THRESHOLD 250
611 /** Test if a page is a leaf page */
612 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
613 /** Test if a page is a LEAF2 page */
614 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
615 /** Test if a page is a branch page */
616 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
617 /** Test if a page is an overflow page */
618 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
619 /** Test if a page is a sub page */
620 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
622 /** The number of overflow pages needed to store the given size. */
623 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
625 /** Header for a single key/data pair within a page.
626 * We guarantee 2-byte alignment for nodes.
628 typedef struct MDB_node {
629 /** lo and hi are used for data size on leaf nodes and for
630 * child pgno on branch nodes. On 64 bit platforms, flags
631 * is also used for pgno. (Branch nodes have no flags).
632 * They are in host byte order in case that lets some
633 * accesses be optimized into a 32-bit word access.
635 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
636 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
637 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
638 /** @defgroup mdb_node Node Flags
640 * Flags for node headers.
643 #define F_BIGDATA 0x01 /**< data put on overflow page */
644 #define F_SUBDATA 0x02 /**< data is a sub-database */
645 #define F_DUPDATA 0x04 /**< data has duplicates */
647 /** valid flags for #mdb_node_add() */
648 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
651 unsigned short mn_flags; /**< @ref mdb_node */
652 unsigned short mn_ksize; /**< key size */
653 char mn_data[1]; /**< key and data are appended here */
656 /** Size of the node header, excluding dynamic data at the end */
657 #define NODESIZE offsetof(MDB_node, mn_data)
659 /** Bit position of top word in page number, for shifting mn_flags */
660 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
662 /** Size of a node in a branch page with a given key.
663 * This is just the node header plus the key, there is no data.
665 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
667 /** Size of a node in a leaf page with a given key and data.
668 * This is node header plus key plus data size.
670 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
672 /** Address of node \b i in page \b p */
673 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
675 /** Address of the key for the node */
676 #define NODEKEY(node) (void *)((node)->mn_data)
678 /** Address of the data for a node */
679 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
681 /** Get the page number pointed to by a branch node */
682 #define NODEPGNO(node) \
683 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
684 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
685 /** Set the page number in a branch node */
686 #define SETPGNO(node,pgno) do { \
687 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
688 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
690 /** Get the size of the data in a leaf node */
691 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
692 /** Set the size of the data for a leaf node */
693 #define SETDSZ(node,size) do { \
694 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
695 /** The size of a key in a node */
696 #define NODEKSZ(node) ((node)->mn_ksize)
698 /** Copy a page number from src to dst */
700 #define COPY_PGNO(dst,src) dst = src
702 #if SIZE_MAX > 4294967295UL
703 #define COPY_PGNO(dst,src) do { \
704 unsigned short *s, *d; \
705 s = (unsigned short *)&(src); \
706 d = (unsigned short *)&(dst); \
713 #define COPY_PGNO(dst,src) do { \
714 unsigned short *s, *d; \
715 s = (unsigned short *)&(src); \
716 d = (unsigned short *)&(dst); \
722 /** The address of a key in a LEAF2 page.
723 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
724 * There are no node headers, keys are stored contiguously.
726 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
728 /** Set the \b node's key into \b key, if requested. */
729 #define MDB_SET_KEY(node, key) { if ((key) != NULL) { \
730 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
732 /** Information about a single database in the environment. */
733 typedef struct MDB_db {
734 uint32_t md_pad; /**< also ksize for LEAF2 pages */
735 uint16_t md_flags; /**< @ref mdb_open */
736 uint16_t md_depth; /**< depth of this tree */
737 pgno_t md_branch_pages; /**< number of internal pages */
738 pgno_t md_leaf_pages; /**< number of leaf pages */
739 pgno_t md_overflow_pages; /**< number of overflow pages */
740 size_t md_entries; /**< number of data items */
741 pgno_t md_root; /**< the root page of this tree */
744 /** Handle for the DB used to track free pages. */
746 /** Handle for the default DB. */
749 /** Meta page content. */
750 typedef struct MDB_meta {
751 /** Stamp identifying this as an MDB data file. It must be set
754 /** Version number of this lock file. Must be set to #MDB_VERSION. */
756 void *mm_address; /**< address for fixed mapping */
757 size_t mm_mapsize; /**< size of mmap region */
758 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
759 /** The size of pages used in this DB */
760 #define mm_psize mm_dbs[0].md_pad
761 /** Any persistent environment flags. @ref mdb_env */
762 #define mm_flags mm_dbs[0].md_flags
763 pgno_t mm_last_pg; /**< last used page in file */
764 txnid_t mm_txnid; /**< txnid that committed this page */
767 /** Auxiliary DB info.
768 * The information here is mostly static/read-only. There is
769 * only a single copy of this record in the environment.
771 typedef struct MDB_dbx {
772 MDB_val md_name; /**< name of the database */
773 MDB_cmp_func *md_cmp; /**< function for comparing keys */
774 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
775 MDB_rel_func *md_rel; /**< user relocate function */
776 void *md_relctx; /**< user-provided context for md_rel */
779 /** A database transaction.
780 * Every operation requires a transaction handle.
783 MDB_txn *mt_parent; /**< parent of a nested txn */
784 MDB_txn *mt_child; /**< nested txn under this txn */
785 pgno_t mt_next_pgno; /**< next unallocated page */
786 /** The ID of this transaction. IDs are integers incrementing from 1.
787 * Only committed write transactions increment the ID. If a transaction
788 * aborts, the ID may be re-used by the next writer.
791 MDB_env *mt_env; /**< the DB environment */
792 /** The list of pages that became unused during this transaction.
796 ID2L dirty_list; /**< modified pages */
797 MDB_reader *reader; /**< this thread's slot in the reader table */
799 /** Array of records for each DB known in the environment. */
801 /** Array of MDB_db records for each known DB */
803 /** @defgroup mt_dbflag Transaction DB Flags
807 #define DB_DIRTY 0x01 /**< DB was written in this txn */
808 #define DB_STALE 0x02 /**< DB record is older than txnID */
810 /** Array of cursors for each DB */
811 MDB_cursor **mt_cursors;
812 /** Array of flags for each DB */
813 unsigned char *mt_dbflags;
814 /** Number of DB records in use. This number only ever increments;
815 * we don't decrement it when individual DB handles are closed.
819 /** @defgroup mdb_txn Transaction Flags
823 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
824 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
826 unsigned int mt_flags; /**< @ref mdb_txn */
827 /** Tracks which of the two meta pages was used at the start
828 * of this transaction.
830 unsigned int mt_toggle;
833 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
834 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
835 * raise this on a 64 bit machine.
837 #define CURSOR_STACK 32
841 /** Cursors are used for all DB operations */
843 /** Next cursor on this DB in this txn */
845 /** Original cursor if this is a shadow */
847 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
848 struct MDB_xcursor *mc_xcursor;
849 /** The transaction that owns this cursor */
851 /** The database handle this cursor operates on */
853 /** The database record for this cursor */
855 /** The database auxiliary record for this cursor */
857 /** The @ref mt_dbflag for this database */
858 unsigned char *mc_dbflag;
859 unsigned short mc_snum; /**< number of pushed pages */
860 unsigned short mc_top; /**< index of top page, mc_snum-1 */
861 /** @defgroup mdb_cursor Cursor Flags
863 * Cursor state flags.
866 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
867 #define C_EOF 0x02 /**< No more data */
868 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
869 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
870 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
872 unsigned int mc_flags; /**< @ref mdb_cursor */
873 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
874 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
877 /** Context for sorted-dup records.
878 * We could have gone to a fully recursive design, with arbitrarily
879 * deep nesting of sub-databases. But for now we only handle these
880 * levels - main DB, optional sub-DB, sorted-duplicate DB.
882 typedef struct MDB_xcursor {
883 /** A sub-cursor for traversing the Dup DB */
884 MDB_cursor mx_cursor;
885 /** The database record for this Dup DB */
887 /** The auxiliary DB record for this Dup DB */
889 /** The @ref mt_dbflag for this Dup DB */
890 unsigned char mx_dbflag;
893 /** A set of pages freed by an earlier transaction. */
894 typedef struct MDB_oldpages {
895 /** Usually we only read one record from the FREEDB at a time, but
896 * in case we read more, this will chain them together.
898 struct MDB_oldpages *mo_next;
899 /** The ID of the transaction in which these pages were freed. */
901 /** An #IDL of the pages */
902 pgno_t mo_pages[1]; /* dynamic */
905 /** The database environment. */
907 HANDLE me_fd; /**< The main data file */
908 HANDLE me_lfd; /**< The lock file */
909 HANDLE me_mfd; /**< just for writing the meta pages */
910 /** Failed to update the meta page. Probably an I/O error. */
911 #define MDB_FATAL_ERROR 0x80000000U
912 uint32_t me_flags; /**< @ref mdb_env */
913 uint32_t me_extrapad; /**< unused for now */
914 unsigned int me_maxreaders; /**< size of the reader table */
915 MDB_dbi me_numdbs; /**< number of DBs opened */
916 MDB_dbi me_maxdbs; /**< size of the DB table */
917 char *me_path; /**< path to the DB files */
918 char *me_map; /**< the memory map of the data file */
919 MDB_txninfo *me_txns; /**< the memory map of the lock file */
920 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
921 MDB_txn *me_txn; /**< current write transaction */
922 size_t me_mapsize; /**< size of the data memory map */
923 off_t me_size; /**< current file size */
924 pgno_t me_maxpg; /**< me_mapsize / me_psize */
925 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
926 unsigned int me_db_toggle; /**< which DB table is current */
927 txnid_t me_wtxnid; /**< ID of last txn we committed */
928 MDB_dbx *me_dbxs; /**< array of static DB info */
929 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
930 MDB_oldpages *me_pghead; /**< list of old page records */
931 pthread_key_t me_txkey; /**< thread-key for readers */
932 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
933 /** IDL of pages that became unused in a write txn */
935 /** ID2L of pages that were written during a write txn */
936 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
937 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
938 LAZY_RWLOCK_DEF(me_dblock)
940 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
944 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
948 /** max number of pages to commit in one writev() call */
949 #define MDB_COMMIT_PAGES 64
950 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
951 #undef MDB_COMMIT_PAGES
952 #define MDB_COMMIT_PAGES IOV_MAX
955 static MDB_page *mdb_page_alloc(MDB_cursor *mc, int num);
956 static MDB_page *mdb_page_new(MDB_cursor *mc, uint32_t flags, int num);
957 static int mdb_page_touch(MDB_cursor *mc);
959 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
960 static int mdb_page_search_root(MDB_cursor *mc,
961 MDB_val *key, int modify);
962 static int mdb_page_search(MDB_cursor *mc,
963 MDB_val *key, int modify);
964 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
965 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
966 pgno_t newpgno, unsigned int nflags);
968 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
969 static int mdb_env_read_meta(MDB_env *env, int *which);
970 static int mdb_env_write_meta(MDB_txn *txn);
972 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
973 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
974 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
975 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
976 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
977 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
978 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
979 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
980 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
982 static int mdb_rebalance(MDB_cursor *mc);
983 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
985 static void mdb_cursor_pop(MDB_cursor *mc);
986 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
988 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
989 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
990 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
991 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
992 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
994 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
995 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
997 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
998 static void mdb_xcursor_init0(MDB_cursor *mc);
999 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1001 static int mdb_drop0(MDB_cursor *mc, int subs);
1002 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1005 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1009 static SECURITY_DESCRIPTOR mdb_null_sd;
1010 static SECURITY_ATTRIBUTES mdb_all_sa;
1011 static int mdb_sec_inited;
1014 /** Return the library version info. */
1016 mdb_version(int *major, int *minor, int *patch)
1018 if (major) *major = MDB_VERSION_MAJOR;
1019 if (minor) *minor = MDB_VERSION_MINOR;
1020 if (patch) *patch = MDB_VERSION_PATCH;
1021 return MDB_VERSION_STRING;
1024 /** Table of descriptions for MDB @ref errors */
1025 static char *const mdb_errstr[] = {
1026 "MDB_KEYEXIST: Key/data pair already exists",
1027 "MDB_NOTFOUND: No matching key/data pair found",
1028 "MDB_PAGE_NOTFOUND: Requested page not found",
1029 "MDB_CORRUPTED: Located page was wrong type",
1030 "MDB_PANIC: Update of meta page failed",
1031 "MDB_VERSION_MISMATCH: Database environment version mismatch"
1035 mdb_strerror(int err)
1038 return ("Successful return: 0");
1040 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
1041 return mdb_errstr[err - MDB_KEYEXIST];
1043 return strerror(err);
1047 /** Display a key in hexadecimal and return the address of the result.
1048 * @param[in] key the key to display
1049 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1050 * @return The key in hexadecimal form.
1053 mdb_dkey(MDB_val *key, char *buf)
1056 unsigned char *c = key->mv_data;
1058 if (key->mv_size > MAXKEYSIZE)
1059 return "MAXKEYSIZE";
1060 /* may want to make this a dynamic check: if the key is mostly
1061 * printable characters, print it as-is instead of converting to hex.
1064 for (i=0; i<key->mv_size; i++)
1065 ptr += sprintf(ptr, "%02x", *c++);
1067 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1074 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1076 return txn->mt_dbxs[dbi].md_cmp(a, b);
1080 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1082 if (txn->mt_dbxs[dbi].md_dcmp)
1083 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1085 return EINVAL; /* too bad you can't distinguish this from a valid result */
1088 /** Allocate a single page.
1089 * Re-use old malloc'd pages first, otherwise just malloc.
1092 mdb_page_malloc(MDB_cursor *mc) {
1094 size_t sz = mc->mc_txn->mt_env->me_psize;
1095 if (mc->mc_txn->mt_env->me_dpages) {
1096 ret = mc->mc_txn->mt_env->me_dpages;
1097 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1098 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1099 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1102 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1107 /** Allocate pages for writing.
1108 * If there are free pages available from older transactions, they
1109 * will be re-used first. Otherwise a new page will be allocated.
1110 * @param[in] mc cursor A cursor handle identifying the transaction and
1111 * database for which we are allocating.
1112 * @param[in] num the number of pages to allocate.
1113 * @return Address of the allocated page(s). Requests for multiple pages
1114 * will always be satisfied by a single contiguous chunk of memory.
1117 mdb_page_alloc(MDB_cursor *mc, int num)
1119 MDB_txn *txn = mc->mc_txn;
1121 pgno_t pgno = P_INVALID;
1124 if (txn->mt_txnid > 2) {
1126 if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
1127 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1128 /* See if there's anything in the free DB */
1131 txnid_t *kptr, oldest;
1133 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1134 mdb_page_search(&m2, NULL, 0);
1135 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1136 kptr = (txnid_t *)NODEKEY(leaf);
1140 oldest = txn->mt_txnid - 1;
1141 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1142 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1143 if (mr && mr < oldest)
1148 if (oldest > *kptr) {
1149 /* It's usable, grab it.
1155 mdb_node_read(txn, leaf, &data);
1156 idl = (ID *) data.mv_data;
1157 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1158 mop->mo_next = txn->mt_env->me_pghead;
1159 mop->mo_txnid = *kptr;
1160 txn->mt_env->me_pghead = mop;
1161 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1166 DPRINTF("IDL read txn %zu root %zu num %zu",
1167 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1168 for (i=0; i<idl[0]; i++) {
1169 DPRINTF("IDL %zu", idl[i+1]);
1173 /* drop this IDL from the DB */
1174 m2.mc_ki[m2.mc_top] = 0;
1175 m2.mc_flags = C_INITIALIZED;
1176 mdb_cursor_del(&m2, 0);
1179 if (txn->mt_env->me_pghead) {
1180 MDB_oldpages *mop = txn->mt_env->me_pghead;
1182 /* FIXME: For now, always use fresh pages. We
1183 * really ought to search the free list for a
1188 /* peel pages off tail, so we only have to truncate the list */
1189 pgno = MDB_IDL_LAST(mop->mo_pages);
1190 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1192 if (mop->mo_pages[2] > mop->mo_pages[1])
1193 mop->mo_pages[0] = 0;
1197 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1198 txn->mt_env->me_pghead = mop->mo_next;
1205 if (pgno == P_INVALID) {
1206 /* DB size is maxed out */
1207 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1208 DPUTS("DB size maxed out");
1212 if (txn->mt_env->me_dpages && num == 1) {
1213 np = txn->mt_env->me_dpages;
1214 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1215 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1216 txn->mt_env->me_dpages = np->mp_next;
1218 size_t sz = txn->mt_env->me_psize * num;
1219 if ((np = malloc(sz)) == NULL)
1221 VGMEMP_ALLOC(txn->mt_env, np, sz);
1223 if (pgno == P_INVALID) {
1224 np->mp_pgno = txn->mt_next_pgno;
1225 txn->mt_next_pgno += num;
1229 mid.mid = np->mp_pgno;
1231 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1236 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1237 * @param[in] mc cursor pointing to the page to be touched
1238 * @return 0 on success, non-zero on failure.
1241 mdb_page_touch(MDB_cursor *mc)
1243 MDB_page *mp = mc->mc_pg[mc->mc_top];
1246 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1248 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1250 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1251 assert(mp->mp_pgno != np->mp_pgno);
1252 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1254 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1257 mp->mp_flags |= P_DIRTY;
1260 /* Adjust other cursors pointing to mp */
1261 if (mc->mc_flags & C_SUB) {
1262 MDB_cursor *m2, *m3;
1263 MDB_dbi dbi = mc->mc_dbi-1;
1265 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1266 if (m2 == mc) continue;
1267 m3 = &m2->mc_xcursor->mx_cursor;
1268 if (m3->mc_snum < mc->mc_snum) continue;
1269 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1270 m3->mc_pg[mc->mc_top] = mp;
1276 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1277 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1278 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1279 m2->mc_pg[mc->mc_top] = mp;
1283 mc->mc_pg[mc->mc_top] = mp;
1284 /** If this page has a parent, update the parent to point to
1288 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1290 mc->mc_db->md_root = mp->mp_pgno;
1291 } else if (mc->mc_txn->mt_parent) {
1294 /* If txn has a parent, make sure the page is in our
1297 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1298 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1299 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1300 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1301 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1302 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1303 mc->mc_pg[mc->mc_top] = mp;
1309 np = mdb_page_malloc(mc);
1310 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1311 mid.mid = np->mp_pgno;
1313 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1321 mdb_env_sync(MDB_env *env, int force)
1324 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1325 if (MDB_FDATASYNC(env->me_fd))
1331 /** Make shadow copies of all of parent txn's cursors */
1333 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1335 MDB_cursor *mc, *m2;
1336 unsigned int i, j, size;
1338 for (i=0;i<src->mt_numdbs; i++) {
1339 if (src->mt_cursors[i]) {
1340 size = sizeof(MDB_cursor);
1341 if (src->mt_cursors[i]->mc_xcursor)
1342 size += sizeof(MDB_xcursor);
1343 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1350 mc->mc_db = &dst->mt_dbs[i];
1351 mc->mc_dbx = m2->mc_dbx;
1352 mc->mc_dbflag = &dst->mt_dbflags[i];
1353 mc->mc_snum = m2->mc_snum;
1354 mc->mc_top = m2->mc_top;
1355 mc->mc_flags = m2->mc_flags | C_SHADOW;
1356 for (j=0; j<mc->mc_snum; j++) {
1357 mc->mc_pg[j] = m2->mc_pg[j];
1358 mc->mc_ki[j] = m2->mc_ki[j];
1360 if (m2->mc_xcursor) {
1361 MDB_xcursor *mx, *mx2;
1362 mx = (MDB_xcursor *)(mc+1);
1363 mc->mc_xcursor = mx;
1364 mx2 = m2->mc_xcursor;
1365 mx->mx_db = mx2->mx_db;
1366 mx->mx_dbx = mx2->mx_dbx;
1367 mx->mx_dbflag = mx2->mx_dbflag;
1368 mx->mx_cursor.mc_txn = dst;
1369 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1370 mx->mx_cursor.mc_db = &mx->mx_db;
1371 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1372 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1373 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1374 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1375 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1376 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1377 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1378 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1381 mc->mc_xcursor = NULL;
1383 mc->mc_next = dst->mt_cursors[i];
1384 dst->mt_cursors[i] = mc;
1391 /** Merge shadow cursors back into parent's */
1393 mdb_cursor_merge(MDB_txn *txn)
1396 for (i=0; i<txn->mt_numdbs; i++) {
1397 if (txn->mt_cursors[i]) {
1399 while ((mc = txn->mt_cursors[i])) {
1400 txn->mt_cursors[i] = mc->mc_next;
1401 if (mc->mc_flags & C_SHADOW) {
1402 MDB_cursor *m2 = mc->mc_orig;
1404 m2->mc_snum = mc->mc_snum;
1405 m2->mc_top = mc->mc_top;
1406 for (j=0; j<mc->mc_snum; j++) {
1407 m2->mc_pg[j] = mc->mc_pg[j];
1408 m2->mc_ki[j] = mc->mc_ki[j];
1411 if (mc->mc_flags & C_ALLOCD)
1419 mdb_txn_reset0(MDB_txn *txn);
1421 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1422 * @param[in] txn the transaction handle to initialize
1423 * @return 0 on success, non-zero on failure. This can only
1424 * fail for read-only transactions, and then only if the
1425 * reader table is full.
1428 mdb_txn_renew0(MDB_txn *txn)
1430 MDB_env *env = txn->mt_env;
1433 if (txn->mt_flags & MDB_TXN_RDONLY) {
1434 MDB_reader *r = pthread_getspecific(env->me_txkey);
1437 pid_t pid = getpid();
1438 pthread_t tid = pthread_self();
1441 for (i=0; i<env->me_txns->mti_numreaders; i++)
1442 if (env->me_txns->mti_readers[i].mr_pid == 0)
1444 if (i == env->me_maxreaders) {
1445 UNLOCK_MUTEX_R(env);
1448 env->me_txns->mti_readers[i].mr_pid = pid;
1449 env->me_txns->mti_readers[i].mr_tid = tid;
1450 if (i >= env->me_txns->mti_numreaders)
1451 env->me_txns->mti_numreaders = i+1;
1452 UNLOCK_MUTEX_R(env);
1453 r = &env->me_txns->mti_readers[i];
1454 pthread_setspecific(env->me_txkey, r);
1456 txn->mt_toggle = env->me_txns->mti_me_toggle;
1457 txn->mt_txnid = env->me_txns->mti_txnid;
1458 /* This happens if a different process was the
1459 * last writer to the DB.
1461 if (env->me_wtxnid < txn->mt_txnid)
1462 mt_dbflag = DB_STALE;
1463 r->mr_txnid = txn->mt_txnid;
1464 txn->mt_u.reader = r;
1468 txn->mt_txnid = env->me_txns->mti_txnid;
1469 if (env->me_wtxnid < txn->mt_txnid)
1470 mt_dbflag = DB_STALE;
1472 txn->mt_toggle = env->me_txns->mti_me_toggle;
1473 txn->mt_u.dirty_list = env->me_dirty_list;
1474 txn->mt_u.dirty_list[0].mid = 0;
1475 txn->mt_free_pgs = env->me_free_pgs;
1476 txn->mt_free_pgs[0] = 0;
1477 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1481 /* Copy the DB arrays */
1482 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1483 txn->mt_numdbs = env->me_numdbs;
1484 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1485 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1486 if (txn->mt_numdbs > 2)
1487 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1488 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1489 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1491 memset(txn->mt_dbflags, mt_dbflag, env->me_numdbs);
1497 mdb_txn_renew(MDB_txn *txn)
1504 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1505 DPUTS("environment had fatal error, must shutdown!");
1509 rc = mdb_txn_renew0(txn);
1510 if (rc == MDB_SUCCESS) {
1511 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1512 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1513 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1519 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1524 if (env->me_flags & MDB_FATAL_ERROR) {
1525 DPUTS("environment had fatal error, must shutdown!");
1529 /* parent already has an active child txn */
1530 if (parent->mt_child) {
1534 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1535 if (!(flags & MDB_RDONLY))
1536 size += env->me_maxdbs * sizeof(MDB_cursor *);
1538 if ((txn = calloc(1, size)) == NULL) {
1539 DPRINTF("calloc: %s", strerror(ErrCode()));
1542 txn->mt_dbs = (MDB_db *)(txn+1);
1543 if (flags & MDB_RDONLY) {
1544 txn->mt_flags |= MDB_TXN_RDONLY;
1545 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1547 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1548 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1553 txn->mt_free_pgs = mdb_midl_alloc();
1554 if (!txn->mt_free_pgs) {
1558 txn->mt_u.dirty_list = malloc(sizeof(ID2)*MDB_IDL_UM_SIZE);
1559 if (!txn->mt_u.dirty_list) {
1560 free(txn->mt_free_pgs);
1564 txn->mt_txnid = parent->mt_txnid;
1565 txn->mt_toggle = parent->mt_toggle;
1566 txn->mt_u.dirty_list[0].mid = 0;
1567 txn->mt_free_pgs[0] = 0;
1568 txn->mt_next_pgno = parent->mt_next_pgno;
1569 parent->mt_child = txn;
1570 txn->mt_parent = parent;
1571 txn->mt_numdbs = parent->mt_numdbs;
1572 txn->mt_dbxs = parent->mt_dbxs;
1573 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1574 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1575 mdb_cursor_shadow(parent, txn);
1578 rc = mdb_txn_renew0(txn);
1584 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1585 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1586 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1592 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1593 * @param[in] txn the transaction handle to reset
1596 mdb_txn_reset0(MDB_txn *txn)
1598 MDB_env *env = txn->mt_env;
1600 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1601 txn->mt_u.reader->mr_txnid = 0;
1607 /* close(free) all cursors */
1608 for (i=0; i<txn->mt_numdbs; i++) {
1609 if (txn->mt_cursors[i]) {
1611 while ((mc = txn->mt_cursors[i])) {
1612 txn->mt_cursors[i] = mc->mc_next;
1613 if (mc->mc_flags & C_ALLOCD)
1619 /* return all dirty pages to dpage list */
1620 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1621 dp = txn->mt_u.dirty_list[i].mptr;
1622 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1623 dp->mp_next = txn->mt_env->me_dpages;
1624 VGMEMP_FREE(txn->mt_env, dp);
1625 txn->mt_env->me_dpages = dp;
1627 /* large pages just get freed directly */
1628 VGMEMP_FREE(txn->mt_env, dp);
1633 if (txn->mt_parent) {
1634 txn->mt_parent->mt_child = NULL;
1635 free(txn->mt_free_pgs);
1636 free(txn->mt_u.dirty_list);
1639 if (mdb_midl_shrink(&txn->mt_free_pgs))
1640 env->me_free_pgs = txn->mt_free_pgs;
1643 while ((mop = txn->mt_env->me_pghead)) {
1644 txn->mt_env->me_pghead = mop->mo_next;
1649 /* The writer mutex was locked in mdb_txn_begin. */
1650 UNLOCK_MUTEX_W(env);
1655 mdb_txn_reset(MDB_txn *txn)
1660 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1661 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1662 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1664 mdb_txn_reset0(txn);
1668 mdb_txn_abort(MDB_txn *txn)
1673 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1674 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1675 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1678 mdb_txn_abort(txn->mt_child);
1680 mdb_txn_reset0(txn);
1685 mdb_txn_commit(MDB_txn *txn)
1696 assert(txn != NULL);
1697 assert(txn->mt_env != NULL);
1699 if (txn->mt_child) {
1700 mdb_txn_commit(txn->mt_child);
1701 txn->mt_child = NULL;
1706 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1707 if (txn->mt_numdbs > env->me_numdbs) {
1708 /* update the DB tables */
1709 int toggle = !env->me_db_toggle;
1713 ip = &env->me_dbs[toggle][env->me_numdbs];
1714 jp = &txn->mt_dbs[env->me_numdbs];
1715 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1716 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
1720 env->me_db_toggle = toggle;
1721 env->me_numdbs = txn->mt_numdbs;
1722 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1728 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1729 DPUTS("error flag is set, can't commit");
1731 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1736 /* Merge (and close) our cursors with parent's */
1737 mdb_cursor_merge(txn);
1739 if (txn->mt_parent) {
1745 /* Update parent's DB table */
1746 ip = &txn->mt_parent->mt_dbs[2];
1747 jp = &txn->mt_dbs[2];
1748 for (i = 2; i < txn->mt_numdbs; i++) {
1749 if (ip->md_root != jp->md_root)
1753 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1755 /* Append our free list to parent's */
1756 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1758 mdb_midl_free(txn->mt_free_pgs);
1760 /* Merge our dirty list with parent's */
1761 dst = txn->mt_parent->mt_u.dirty_list;
1762 src = txn->mt_u.dirty_list;
1763 x = mdb_mid2l_search(dst, src[1].mid);
1764 for (y=1; y<=src[0].mid; y++) {
1765 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1769 dst[x].mptr = src[y].mptr;
1772 for (; y<=src[0].mid; y++) {
1773 if (++x >= MDB_IDL_UM_MAX) {
1780 free(txn->mt_u.dirty_list);
1781 txn->mt_parent->mt_child = NULL;
1786 if (txn != env->me_txn) {
1787 DPUTS("attempt to commit unknown transaction");
1792 if (!txn->mt_u.dirty_list[0].mid)
1795 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1796 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1798 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1800 /* should only be one record now */
1801 if (env->me_pghead) {
1802 /* make sure first page of freeDB is touched and on freelist */
1803 mdb_page_search(&mc, NULL, 1);
1805 /* save to free list */
1806 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1810 /* make sure last page of freeDB is touched and on freelist */
1811 key.mv_size = MAXKEYSIZE+1;
1813 mdb_page_search(&mc, &key, 1);
1815 mdb_midl_sort(txn->mt_free_pgs);
1819 ID *idl = txn->mt_free_pgs;
1820 DPRINTF("IDL write txn %zu root %zu num %zu",
1821 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1822 for (i=0; i<idl[0]; i++) {
1823 DPRINTF("IDL %zu", idl[i+1]);
1827 /* write to last page of freeDB */
1828 key.mv_size = sizeof(pgno_t);
1829 key.mv_data = &txn->mt_txnid;
1830 data.mv_data = txn->mt_free_pgs;
1831 /* The free list can still grow during this call,
1832 * despite the pre-emptive touches above. So check
1833 * and make sure the entire thing got written.
1836 i = txn->mt_free_pgs[0];
1837 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1838 rc = mdb_cursor_put(&mc, &key, &data, 0);
1843 } while (i != txn->mt_free_pgs[0]);
1844 if (mdb_midl_shrink(&txn->mt_free_pgs))
1845 env->me_free_pgs = txn->mt_free_pgs;
1847 /* should only be one record now */
1848 if (env->me_pghead) {
1852 mop = env->me_pghead;
1853 env->me_pghead = NULL;
1854 key.mv_size = sizeof(pgno_t);
1855 key.mv_data = &mop->mo_txnid;
1856 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1857 data.mv_data = mop->mo_pages;
1858 mdb_cursor_put(&mc, &key, &data, 0);
1862 /* Update DB root pointers. Their pages have already been
1863 * touched so this is all in-place and cannot fail.
1868 data.mv_size = sizeof(MDB_db);
1870 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
1871 for (i = 2; i < txn->mt_numdbs; i++) {
1872 if (txn->mt_dbflags[i] & DB_DIRTY) {
1873 data.mv_data = &txn->mt_dbs[i];
1874 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1879 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1885 /* Windows actually supports scatter/gather I/O, but only on
1886 * unbuffered file handles. Since we're relying on the OS page
1887 * cache for all our data, that's self-defeating. So we just
1888 * write pages one at a time. We use the ov structure to set
1889 * the write offset, to at least save the overhead of a Seek
1893 memset(&ov, 0, sizeof(ov));
1894 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1896 dp = txn->mt_u.dirty_list[i].mptr;
1897 DPRINTF("committing page %zu", dp->mp_pgno);
1898 size = dp->mp_pgno * env->me_psize;
1899 ov.Offset = size & 0xffffffff;
1900 ov.OffsetHigh = size >> 16;
1901 ov.OffsetHigh >>= 16;
1902 /* clear dirty flag */
1903 dp->mp_flags &= ~P_DIRTY;
1904 wsize = env->me_psize;
1905 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
1906 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
1909 DPRINTF("WriteFile: %d", n);
1916 struct iovec iov[MDB_COMMIT_PAGES];
1920 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1921 dp = txn->mt_u.dirty_list[i].mptr;
1922 if (dp->mp_pgno != next) {
1924 DPRINTF("committing %u dirty pages", n);
1925 rc = writev(env->me_fd, iov, n);
1929 DPUTS("short write, filesystem full?");
1931 DPRINTF("writev: %s", strerror(n));
1938 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
1941 DPRINTF("committing page %zu", dp->mp_pgno);
1942 iov[n].iov_len = env->me_psize;
1943 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
1944 iov[n].iov_base = (char *)dp;
1945 size += iov[n].iov_len;
1946 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
1947 /* clear dirty flag */
1948 dp->mp_flags &= ~P_DIRTY;
1949 if (++n >= MDB_COMMIT_PAGES) {
1959 DPRINTF("committing %u dirty pages", n);
1960 rc = writev(env->me_fd, iov, n);
1964 DPUTS("short write, filesystem full?");
1966 DPRINTF("writev: %s", strerror(n));
1973 /* Drop the dirty pages.
1975 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1976 dp = txn->mt_u.dirty_list[i].mptr;
1977 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1978 dp->mp_next = txn->mt_env->me_dpages;
1979 VGMEMP_FREE(txn->mt_env, dp);
1980 txn->mt_env->me_dpages = dp;
1982 VGMEMP_FREE(txn->mt_env, dp);
1985 txn->mt_u.dirty_list[i].mid = 0;
1987 txn->mt_u.dirty_list[0].mid = 0;
1989 if ((n = mdb_env_sync(env, 0)) != 0 ||
1990 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
1994 env->me_wtxnid = txn->mt_txnid;
1998 /* update the DB tables */
2000 int toggle = !env->me_db_toggle;
2004 ip = &env->me_dbs[toggle][2];
2005 jp = &txn->mt_dbs[2];
2006 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
2007 for (i = 2; i < txn->mt_numdbs; i++) {
2008 if (ip->md_root != jp->md_root)
2013 env->me_db_toggle = toggle;
2014 env->me_numdbs = txn->mt_numdbs;
2015 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
2018 UNLOCK_MUTEX_W(env);
2024 /** Read the environment parameters of a DB environment before
2025 * mapping it into memory.
2026 * @param[in] env the environment handle
2027 * @param[out] meta address of where to store the meta information
2028 * @return 0 on success, non-zero on failure.
2031 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2033 char page[MDB_PAGESIZE];
2038 /* We don't know the page size yet, so use a minimum value.
2042 if (!ReadFile(env->me_fd, page, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2044 if ((rc = read(env->me_fd, page, MDB_PAGESIZE)) == 0)
2049 else if (rc != MDB_PAGESIZE) {
2053 DPRINTF("read: %s", strerror(err));
2057 p = (MDB_page *)page;
2059 if (!F_ISSET(p->mp_flags, P_META)) {
2060 DPRINTF("page %zu not a meta page", p->mp_pgno);
2065 if (m->mm_magic != MDB_MAGIC) {
2066 DPUTS("meta has invalid magic");
2070 if (m->mm_version != MDB_VERSION) {
2071 DPRINTF("database is version %u, expected version %u",
2072 m->mm_version, MDB_VERSION);
2073 return MDB_VERSION_MISMATCH;
2076 memcpy(meta, m, sizeof(*m));
2080 /** Write the environment parameters of a freshly created DB environment.
2081 * @param[in] env the environment handle
2082 * @param[out] meta address of where to store the meta information
2083 * @return 0 on success, non-zero on failure.
2086 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2093 DPUTS("writing new meta page");
2095 GET_PAGESIZE(psize);
2097 meta->mm_magic = MDB_MAGIC;
2098 meta->mm_version = MDB_VERSION;
2099 meta->mm_psize = psize;
2100 meta->mm_last_pg = 1;
2101 meta->mm_flags = env->me_flags & 0xffff;
2102 meta->mm_flags |= MDB_INTEGERKEY;
2103 meta->mm_dbs[0].md_root = P_INVALID;
2104 meta->mm_dbs[1].md_root = P_INVALID;
2106 p = calloc(2, psize);
2108 p->mp_flags = P_META;
2111 memcpy(m, meta, sizeof(*meta));
2113 q = (MDB_page *)((char *)p + psize);
2116 q->mp_flags = P_META;
2119 memcpy(m, meta, sizeof(*meta));
2124 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2125 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2128 rc = write(env->me_fd, p, psize * 2);
2129 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2135 /** Update the environment info to commit a transaction.
2136 * @param[in] txn the transaction that's being committed
2137 * @return 0 on success, non-zero on failure.
2140 mdb_env_write_meta(MDB_txn *txn)
2143 MDB_meta meta, metab;
2145 int rc, len, toggle;
2151 assert(txn != NULL);
2152 assert(txn->mt_env != NULL);
2154 toggle = !txn->mt_toggle;
2155 DPRINTF("writing meta page %d for root page %zu",
2156 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2160 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2161 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2163 ptr = (char *)&meta;
2164 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2165 len = sizeof(MDB_meta) - off;
2168 meta.mm_dbs[0] = txn->mt_dbs[0];
2169 meta.mm_dbs[1] = txn->mt_dbs[1];
2170 meta.mm_last_pg = txn->mt_next_pgno - 1;
2171 meta.mm_txnid = txn->mt_txnid;
2174 off += env->me_psize;
2177 /* Write to the SYNC fd */
2180 memset(&ov, 0, sizeof(ov));
2182 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2185 rc = pwrite(env->me_mfd, ptr, len, off);
2190 DPUTS("write failed, disk error?");
2191 /* On a failure, the pagecache still contains the new data.
2192 * Write some old data back, to prevent it from being used.
2193 * Use the non-SYNC fd; we know it will fail anyway.
2195 meta.mm_last_pg = metab.mm_last_pg;
2196 meta.mm_txnid = metab.mm_txnid;
2198 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2200 r2 = pwrite(env->me_fd, ptr, len, off);
2202 env->me_flags |= MDB_FATAL_ERROR;
2205 /* Memory ordering issues are irrelevant; since the entire writer
2206 * is wrapped by wmutex, all of these changes will become visible
2207 * after the wmutex is unlocked. Since the DB is multi-version,
2208 * readers will get consistent data regardless of how fresh or
2209 * how stale their view of these values is.
2211 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
2212 txn->mt_env->me_txns->mti_me_toggle = toggle;
2213 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2214 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
2219 /** Check both meta pages to see which one is newer.
2220 * @param[in] env the environment handle
2221 * @param[out] which address of where to store the meta toggle ID
2222 * @return 0 on success, non-zero on failure.
2225 mdb_env_read_meta(MDB_env *env, int *which)
2229 assert(env != NULL);
2231 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2234 DPRINTF("Using meta page %d", toggle);
2241 mdb_env_create(MDB_env **env)
2245 e = calloc(1, sizeof(MDB_env));
2249 e->me_free_pgs = mdb_midl_alloc();
2250 if (!e->me_free_pgs) {
2254 e->me_maxreaders = DEFAULT_READERS;
2256 e->me_fd = INVALID_HANDLE_VALUE;
2257 e->me_lfd = INVALID_HANDLE_VALUE;
2258 e->me_mfd = INVALID_HANDLE_VALUE;
2259 VGMEMP_CREATE(e,0,0);
2265 mdb_env_set_mapsize(MDB_env *env, size_t size)
2269 env->me_mapsize = size;
2271 env->me_maxpg = env->me_mapsize / env->me_psize;
2276 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2280 env->me_maxdbs = dbs;
2285 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2287 if (env->me_map || readers < 1)
2289 env->me_maxreaders = readers;
2294 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2296 if (!env || !readers)
2298 *readers = env->me_maxreaders;
2302 /** Further setup required for opening an MDB environment
2305 mdb_env_open2(MDB_env *env, unsigned int flags)
2307 int i, newenv = 0, toggle;
2311 env->me_flags = flags;
2313 memset(&meta, 0, sizeof(meta));
2315 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2318 DPUTS("new mdbenv");
2322 if (!env->me_mapsize) {
2323 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2329 LONG sizelo, sizehi;
2330 sizelo = env->me_mapsize & 0xffffffff;
2331 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2333 /* Windows won't create mappings for zero length files.
2334 * Just allocate the maxsize right now.
2337 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2338 if (!SetEndOfFile(env->me_fd))
2340 SetFilePointer(env->me_fd, 0, NULL, 0);
2342 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2343 sizehi, sizelo, NULL);
2346 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2354 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2356 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2358 if (env->me_map == MAP_FAILED)
2363 meta.mm_mapsize = env->me_mapsize;
2364 if (flags & MDB_FIXEDMAP)
2365 meta.mm_address = env->me_map;
2366 i = mdb_env_init_meta(env, &meta);
2367 if (i != MDB_SUCCESS) {
2368 munmap(env->me_map, env->me_mapsize);
2372 env->me_psize = meta.mm_psize;
2374 env->me_maxpg = env->me_mapsize / env->me_psize;
2376 p = (MDB_page *)env->me_map;
2377 env->me_metas[0] = METADATA(p);
2378 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2380 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
2383 DPRINTF("opened database version %u, pagesize %u",
2384 env->me_metas[toggle]->mm_version, env->me_psize);
2385 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
2386 DPRINTF("entries: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
2387 DPRINTF("branch pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
2388 DPRINTF("leaf pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
2389 DPRINTF("overflow pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
2390 DPRINTF("root: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
2396 /** Release a reader thread's slot in the reader lock table.
2397 * This function is called automatically when a thread exits.
2398 * Windows doesn't support destructor callbacks for thread-specific storage,
2399 * so this function is not compiled there.
2400 * @param[in] ptr This points to the slot in the reader lock table.
2403 mdb_env_reader_dest(void *ptr)
2405 MDB_reader *reader = ptr;
2407 reader->mr_txnid = 0;
2413 /** Downgrade the exclusive lock on the region back to shared */
2415 mdb_env_share_locks(MDB_env *env)
2419 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2421 env->me_txns->mti_me_toggle = toggle;
2422 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2427 /* First acquire a shared lock. The Unlock will
2428 * then release the existing exclusive lock.
2430 memset(&ov, 0, sizeof(ov));
2431 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2432 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2436 struct flock lock_info;
2437 /* The shared lock replaces the existing lock */
2438 memset((void *)&lock_info, 0, sizeof(lock_info));
2439 lock_info.l_type = F_RDLCK;
2440 lock_info.l_whence = SEEK_SET;
2441 lock_info.l_start = 0;
2442 lock_info.l_len = 1;
2443 fcntl(env->me_lfd, F_SETLK, &lock_info);
2447 #if defined(_WIN32) || defined(__APPLE__)
2449 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2451 * @(#) $Revision: 5.1 $
2452 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2453 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2455 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2459 * Please do not copyright this code. This code is in the public domain.
2461 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2462 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2463 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2464 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2465 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2466 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2467 * PERFORMANCE OF THIS SOFTWARE.
2470 * chongo <Landon Curt Noll> /\oo/\
2471 * http://www.isthe.com/chongo/
2473 * Share and Enjoy! :-)
2476 typedef unsigned long long mdb_hash_t;
2477 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2479 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2480 * @param[in] str string to hash
2481 * @param[in] hval initial value for hash
2482 * @return 64 bit hash
2484 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2485 * hval arg on the first call.
2488 mdb_hash_str(char *str, mdb_hash_t hval)
2490 unsigned char *s = (unsigned char *)str; /* unsigned string */
2492 * FNV-1a hash each octet of the string
2495 /* xor the bottom with the current octet */
2496 hval ^= (mdb_hash_t)*s++;
2498 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2499 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2500 (hval << 7) + (hval << 8) + (hval << 40);
2502 /* return our new hash value */
2506 /** Hash the string and output the hash in hex.
2507 * @param[in] str string to hash
2508 * @param[out] hexbuf an array of 17 chars to hold the hash
2511 mdb_hash_hex(char *str, char *hexbuf)
2514 mdb_hash_t h = mdb_hash_str(str, MDB_HASH_INIT);
2515 for (i=0; i<8; i++) {
2516 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2522 /** Open and/or initialize the lock region for the environment.
2523 * @param[in] env The MDB environment.
2524 * @param[in] lpath The pathname of the file used for the lock region.
2525 * @param[in] mode The Unix permissions for the file, if we create it.
2526 * @param[out] excl Set to true if we got an exclusive lock on the region.
2527 * @return 0 on success, non-zero on failure.
2530 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2538 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2539 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2540 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2544 /* Try to get exclusive lock. If we succeed, then
2545 * nobody is using the lock region and we should initialize it.
2548 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2552 memset(&ov, 0, sizeof(ov));
2553 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2559 size = GetFileSize(env->me_lfd, NULL);
2561 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2565 /* Try to get exclusive lock. If we succeed, then
2566 * nobody is using the lock region and we should initialize it.
2569 struct flock lock_info;
2570 memset((void *)&lock_info, 0, sizeof(lock_info));
2571 lock_info.l_type = F_WRLCK;
2572 lock_info.l_whence = SEEK_SET;
2573 lock_info.l_start = 0;
2574 lock_info.l_len = 1;
2575 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2579 lock_info.l_type = F_RDLCK;
2580 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2587 size = lseek(env->me_lfd, 0, SEEK_END);
2589 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2590 if (size < rsize && *excl) {
2592 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2593 if (!SetEndOfFile(env->me_lfd)) {
2598 if (ftruncate(env->me_lfd, rsize) != 0) {
2605 size = rsize - sizeof(MDB_txninfo);
2606 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2611 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2617 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2619 if (!env->me_txns) {
2625 env->me_txns = (MDB_txninfo *)mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2627 if (env->me_txns == MAP_FAILED) {
2635 if (!mdb_sec_inited) {
2636 InitializeSecurityDescriptor(&mdb_null_sd,
2637 SECURITY_DESCRIPTOR_REVISION);
2638 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2639 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2640 mdb_all_sa.bInheritHandle = FALSE;
2641 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2644 mdb_hash_hex(lpath, hexbuf);
2645 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2646 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2647 if (!env->me_rmutex) {
2651 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2652 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2653 if (!env->me_wmutex) {
2660 mdb_hash_hex(lpath, hexbuf);
2661 sprintf(env->me_txns->mti_rmname, "MDBr%s", hexbuf);
2662 if (sem_unlink(env->me_txns->mti_rmname)) {
2664 if (rc != ENOENT && rc != EINVAL)
2667 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2668 if (!env->me_rmutex) {
2672 sprintf(env->me_txns->mti_wmname, "MDBw%s", hexbuf);
2673 if (sem_unlink(env->me_txns->mti_wmname)) {
2675 if (rc != ENOENT && rc != EINVAL)
2678 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2679 if (!env->me_wmutex) {
2683 #else /* __APPLE__ */
2684 pthread_mutexattr_t mattr;
2686 pthread_mutexattr_init(&mattr);
2687 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2691 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2692 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2693 #endif /* __APPLE__ */
2695 env->me_txns->mti_version = MDB_VERSION;
2696 env->me_txns->mti_magic = MDB_MAGIC;
2697 env->me_txns->mti_txnid = 0;
2698 env->me_txns->mti_numreaders = 0;
2699 env->me_txns->mti_me_toggle = 0;
2702 if (env->me_txns->mti_magic != MDB_MAGIC) {
2703 DPUTS("lock region has invalid magic");
2707 if (env->me_txns->mti_version != MDB_VERSION) {
2708 DPRINTF("lock region is version %u, expected version %u",
2709 env->me_txns->mti_version, MDB_VERSION);
2710 rc = MDB_VERSION_MISMATCH;
2714 if (rc != EACCES && rc != EAGAIN) {
2718 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2719 if (!env->me_rmutex) {
2723 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2724 if (!env->me_wmutex) {
2730 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2731 if (!env->me_rmutex) {
2735 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2736 if (!env->me_wmutex) {
2746 env->me_lfd = INVALID_HANDLE_VALUE;
2751 /** The name of the lock file in the DB environment */
2752 #define LOCKNAME "/lock.mdb"
2753 /** The name of the data file in the DB environment */
2754 #define DATANAME "/data.mdb"
2755 /** The suffix of the lock file when no subdir is used */
2756 #define LOCKSUFF "-lock"
2759 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2761 int oflags, rc, len, excl;
2762 char *lpath, *dpath;
2765 if (flags & MDB_NOSUBDIR) {
2766 rc = len + sizeof(LOCKSUFF) + len + 1;
2768 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2773 if (flags & MDB_NOSUBDIR) {
2774 dpath = lpath + len + sizeof(LOCKSUFF);
2775 sprintf(lpath, "%s" LOCKSUFF, path);
2776 strcpy(dpath, path);
2778 dpath = lpath + len + sizeof(LOCKNAME);
2779 sprintf(lpath, "%s" LOCKNAME, path);
2780 sprintf(dpath, "%s" DATANAME, path);
2783 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2788 if (F_ISSET(flags, MDB_RDONLY)) {
2789 oflags = GENERIC_READ;
2790 len = OPEN_EXISTING;
2792 oflags = GENERIC_READ|GENERIC_WRITE;
2795 mode = FILE_ATTRIBUTE_NORMAL;
2796 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2797 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2802 if (F_ISSET(flags, MDB_RDONLY))
2805 oflags = O_RDWR | O_CREAT;
2807 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2813 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2814 /* synchronous fd for meta writes */
2816 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2817 mode |= FILE_FLAG_WRITE_THROUGH;
2818 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2819 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2824 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2825 oflags |= MDB_DSYNC;
2826 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2831 env->me_path = strdup(path);
2832 DPRINTF("opened dbenv %p", (void *) env);
2833 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2834 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2836 mdb_env_share_locks(env);
2837 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
2838 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
2839 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
2845 if (env->me_fd != INVALID_HANDLE_VALUE) {
2847 env->me_fd = INVALID_HANDLE_VALUE;
2849 if (env->me_lfd != INVALID_HANDLE_VALUE) {
2851 env->me_lfd = INVALID_HANDLE_VALUE;
2859 mdb_env_close(MDB_env *env)
2866 VGMEMP_DESTROY(env);
2867 while (env->me_dpages) {
2868 dp = env->me_dpages;
2869 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
2870 env->me_dpages = dp->mp_next;
2874 free(env->me_dbs[1]);
2875 free(env->me_dbs[0]);
2879 LAZY_RWLOCK_DESTROY(&env->me_dblock);
2880 pthread_key_delete(env->me_txkey);
2883 munmap(env->me_map, env->me_mapsize);
2888 pid_t pid = getpid();
2890 for (i=0; i<env->me_txns->mti_numreaders; i++)
2891 if (env->me_txns->mti_readers[i].mr_pid == pid)
2892 env->me_txns->mti_readers[i].mr_pid = 0;
2893 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
2896 mdb_midl_free(env->me_free_pgs);
2900 /** Compare two items pointing at aligned size_t's */
2902 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
2904 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
2905 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
2908 /** Compare two items pointing at aligned int's */
2910 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
2912 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
2913 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
2916 /** Compare two items pointing at ints of unknown alignment.
2917 * Nodes and keys are guaranteed to be 2-byte aligned.
2920 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
2922 #if BYTE_ORDER == LITTLE_ENDIAN
2923 unsigned short *u, *c;
2926 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
2927 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
2930 } while(!x && u > (unsigned short *)a->mv_data);
2933 return memcmp(a->mv_data, b->mv_data, a->mv_size);
2937 /** Compare two items lexically */
2939 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
2946 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
2952 diff = memcmp(a->mv_data, b->mv_data, len);
2953 return diff ? diff : len_diff<0 ? -1 : len_diff;
2956 /** Compare two items in reverse byte order */
2958 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
2960 const unsigned char *p1, *p2, *p1_lim;
2964 p1_lim = (const unsigned char *)a->mv_data;
2965 p1 = (const unsigned char *)a->mv_data + a->mv_size;
2966 p2 = (const unsigned char *)b->mv_data + b->mv_size;
2968 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
2974 while (p1 > p1_lim) {
2975 diff = *--p1 - *--p2;
2979 return len_diff<0 ? -1 : len_diff;
2982 /** Search for key within a page, using binary search.
2983 * Returns the smallest entry larger or equal to the key.
2984 * If exactp is non-null, stores whether the found entry was an exact match
2985 * in *exactp (1 or 0).
2986 * Updates the cursor index with the index of the found entry.
2987 * If no entry larger or equal to the key is found, returns NULL.
2990 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
2992 unsigned int i = 0, nkeys;
2995 MDB_page *mp = mc->mc_pg[mc->mc_top];
2996 MDB_node *node = NULL;
3001 nkeys = NUMKEYS(mp);
3006 COPY_PGNO(pgno, mp->mp_pgno);
3007 DPRINTF("searching %u keys in %s %spage %zu",
3008 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3015 low = IS_LEAF(mp) ? 0 : 1;
3017 cmp = mc->mc_dbx->md_cmp;
3019 /* Branch pages have no data, so if using integer keys,
3020 * alignment is guaranteed. Use faster mdb_cmp_int.
3022 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3023 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3030 nodekey.mv_size = mc->mc_db->md_pad;
3031 node = NODEPTR(mp, 0); /* fake */
3032 while (low <= high) {
3033 i = (low + high) >> 1;
3034 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3035 rc = cmp(key, &nodekey);
3036 DPRINTF("found leaf index %u [%s], rc = %i",
3037 i, DKEY(&nodekey), rc);
3046 while (low <= high) {
3047 i = (low + high) >> 1;
3049 node = NODEPTR(mp, i);
3050 nodekey.mv_size = NODEKSZ(node);
3051 nodekey.mv_data = NODEKEY(node);
3053 rc = cmp(key, &nodekey);
3056 DPRINTF("found leaf index %u [%s], rc = %i",
3057 i, DKEY(&nodekey), rc);
3059 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3060 i, DKEY(&nodekey), NODEPGNO(node), rc);
3071 if (rc > 0) { /* Found entry is less than the key. */
3072 i++; /* Skip to get the smallest entry larger than key. */
3074 node = NODEPTR(mp, i);
3077 *exactp = (rc == 0);
3078 /* store the key index */
3079 mc->mc_ki[mc->mc_top] = i;
3081 /* There is no entry larger or equal to the key. */
3084 /* nodeptr is fake for LEAF2 */
3090 mdb_cursor_adjust(MDB_cursor *mc, func)
3094 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3095 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3102 /** Pop a page off the top of the cursor's stack. */
3104 mdb_cursor_pop(MDB_cursor *mc)
3109 top = mc->mc_pg[mc->mc_top];
3114 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3115 mc->mc_dbi, (void *) mc);
3119 /** Push a page onto the top of the cursor's stack. */
3121 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3123 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3124 mc->mc_dbi, (void *) mc);
3126 if (mc->mc_snum >= CURSOR_STACK) {
3127 assert(mc->mc_snum < CURSOR_STACK);
3131 mc->mc_top = mc->mc_snum++;
3132 mc->mc_pg[mc->mc_top] = mp;
3133 mc->mc_ki[mc->mc_top] = 0;
3138 /** Find the address of the page corresponding to a given page number.
3139 * @param[in] txn the transaction for this access.
3140 * @param[in] pgno the page number for the page to retrieve.
3141 * @param[out] ret address of a pointer where the page's address will be stored.
3142 * @return 0 on success, non-zero on failure.
3145 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3149 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3151 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3152 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3153 p = txn->mt_u.dirty_list[x].mptr;
3157 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
3158 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3162 DPRINTF("page %zu not found", pgno);
3165 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3168 /** Search for the page a given key should be in.
3169 * Pushes parent pages on the cursor stack. This function continues a
3170 * search on a cursor that has already been initialized. (Usually by
3171 * #mdb_page_search() but also by #mdb_node_move().)
3172 * @param[in,out] mc the cursor for this operation.
3173 * @param[in] key the key to search for. If NULL, search for the lowest
3174 * page. (This is used by #mdb_cursor_first().)
3175 * @param[in] modify If true, visited pages are updated with new page numbers.
3176 * @return 0 on success, non-zero on failure.
3179 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3181 MDB_page *mp = mc->mc_pg[mc->mc_top];
3186 while (IS_BRANCH(mp)) {
3190 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3191 assert(NUMKEYS(mp) > 1);
3192 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3194 if (key == NULL) /* Initialize cursor to first page. */
3196 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3197 /* cursor to last page */
3201 node = mdb_node_search(mc, key, &exact);
3203 i = NUMKEYS(mp) - 1;
3205 i = mc->mc_ki[mc->mc_top];
3214 DPRINTF("following index %u for key [%s]",
3216 assert(i < NUMKEYS(mp));
3217 node = NODEPTR(mp, i);
3219 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3222 mc->mc_ki[mc->mc_top] = i;
3223 if ((rc = mdb_cursor_push(mc, mp)))
3227 if ((rc = mdb_page_touch(mc)) != 0)
3229 mp = mc->mc_pg[mc->mc_top];
3234 DPRINTF("internal error, index points to a %02X page!?",
3236 return MDB_CORRUPTED;
3239 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3240 key ? DKEY(key) : NULL);
3245 /** Search for the page a given key should be in.
3246 * Pushes parent pages on the cursor stack. This function just sets up
3247 * the search; it finds the root page for \b mc's database and sets this
3248 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3249 * called to complete the search.
3250 * @param[in,out] mc the cursor for this operation.
3251 * @param[in] key the key to search for. If NULL, search for the lowest
3252 * page. (This is used by #mdb_cursor_first().)
3253 * @param[in] modify If true, visited pages are updated with new page numbers.
3254 * @return 0 on success, non-zero on failure.
3257 mdb_page_search(MDB_cursor *mc, MDB_val *key, int modify)
3262 /* Make sure the txn is still viable, then find the root from
3263 * the txn's db table.
3265 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3266 DPUTS("transaction has failed, must abort");
3269 /* Make sure we're using an up-to-date root */
3270 if (mc->mc_dbi > MAIN_DBI) {
3271 if ((*mc->mc_dbflag & DB_STALE) ||
3272 (modify && !(*mc->mc_dbflag & DB_DIRTY))) {
3274 unsigned char dbflag = 0;
3275 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3276 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, modify);
3279 if (*mc->mc_dbflag & DB_STALE) {
3282 MDB_node *leaf = mdb_node_search(&mc2,
3283 &mc->mc_dbx->md_name, &exact);
3285 return MDB_NOTFOUND;
3286 mdb_node_read(mc->mc_txn, leaf, &data);
3287 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3291 *mc->mc_dbflag = dbflag;
3294 root = mc->mc_db->md_root;
3296 if (root == P_INVALID) { /* Tree is empty. */
3297 DPUTS("tree is empty");
3298 return MDB_NOTFOUND;
3303 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3309 DPRINTF("db %u root page %zu has flags 0x%X",
3310 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3313 if ((rc = mdb_page_touch(mc)))
3317 return mdb_page_search_root(mc, key, modify);
3320 /** Return the data associated with a given node.
3321 * @param[in] txn The transaction for this operation.
3322 * @param[in] leaf The node being read.
3323 * @param[out] data Updated to point to the node's data.
3324 * @return 0 on success, non-zero on failure.
3327 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3329 MDB_page *omp; /* overflow page */
3333 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3334 data->mv_size = NODEDSZ(leaf);
3335 data->mv_data = NODEDATA(leaf);
3339 /* Read overflow data.
3341 data->mv_size = NODEDSZ(leaf);
3342 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3343 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3344 DPRINTF("read overflow page %zu failed", pgno);
3347 data->mv_data = METADATA(omp);
3353 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3354 MDB_val *key, MDB_val *data)
3363 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3365 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3368 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3372 mdb_cursor_init(&mc, txn, dbi, &mx);
3373 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3376 /** Find a sibling for a page.
3377 * Replaces the page at the top of the cursor's stack with the
3378 * specified sibling, if one exists.
3379 * @param[in] mc The cursor for this operation.
3380 * @param[in] move_right Non-zero if the right sibling is requested,
3381 * otherwise the left sibling.
3382 * @return 0 on success, non-zero on failure.
3385 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3391 if (mc->mc_snum < 2) {
3392 return MDB_NOTFOUND; /* root has no siblings */
3396 DPRINTF("parent page is page %zu, index %u",
3397 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3399 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3400 : (mc->mc_ki[mc->mc_top] == 0)) {
3401 DPRINTF("no more keys left, moving to %s sibling",
3402 move_right ? "right" : "left");
3403 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3407 mc->mc_ki[mc->mc_top]++;
3409 mc->mc_ki[mc->mc_top]--;
3410 DPRINTF("just moving to %s index key %u",
3411 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3413 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3415 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3416 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3419 mdb_cursor_push(mc, mp);
3424 /** Move the cursor to the next data item. */
3426 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3432 if (mc->mc_flags & C_EOF) {
3433 return MDB_NOTFOUND;
3436 assert(mc->mc_flags & C_INITIALIZED);
3438 mp = mc->mc_pg[mc->mc_top];
3440 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3441 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3442 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3443 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3444 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3445 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3449 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3450 if (op == MDB_NEXT_DUP)
3451 return MDB_NOTFOUND;
3455 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3457 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3458 DPUTS("=====> move to next sibling page");
3459 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3460 mc->mc_flags |= C_EOF;
3461 mc->mc_flags &= ~C_INITIALIZED;
3462 return MDB_NOTFOUND;
3464 mp = mc->mc_pg[mc->mc_top];
3465 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3467 mc->mc_ki[mc->mc_top]++;
3469 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3470 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3473 key->mv_size = mc->mc_db->md_pad;
3474 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3478 assert(IS_LEAF(mp));
3479 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3481 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3482 mdb_xcursor_init1(mc, leaf);
3485 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3488 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3489 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3490 if (rc != MDB_SUCCESS)
3495 MDB_SET_KEY(leaf, key);
3499 /** Move the cursor to the previous data item. */
3501 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3507 assert(mc->mc_flags & C_INITIALIZED);
3509 mp = mc->mc_pg[mc->mc_top];
3511 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3512 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3513 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3514 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3515 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3516 if (op != MDB_PREV || rc == MDB_SUCCESS)
3519 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3520 if (op == MDB_PREV_DUP)
3521 return MDB_NOTFOUND;
3526 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3528 if (mc->mc_ki[mc->mc_top] == 0) {
3529 DPUTS("=====> move to prev sibling page");
3530 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3531 mc->mc_flags &= ~C_INITIALIZED;
3532 return MDB_NOTFOUND;
3534 mp = mc->mc_pg[mc->mc_top];
3535 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3536 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3538 mc->mc_ki[mc->mc_top]--;
3540 mc->mc_flags &= ~C_EOF;
3542 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3543 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3546 key->mv_size = mc->mc_db->md_pad;
3547 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3551 assert(IS_LEAF(mp));
3552 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3554 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3555 mdb_xcursor_init1(mc, leaf);
3558 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3561 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3562 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3563 if (rc != MDB_SUCCESS)
3568 MDB_SET_KEY(leaf, key);
3572 /** Set the cursor on a specific data item. */
3574 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3575 MDB_cursor_op op, int *exactp)
3584 assert(key->mv_size > 0);
3586 /* See if we're already on the right page */
3587 if (mc->mc_flags & C_INITIALIZED) {
3590 mp = mc->mc_pg[mc->mc_top];
3592 mc->mc_ki[mc->mc_top] = 0;
3593 return MDB_NOTFOUND;
3595 if (mp->mp_flags & P_LEAF2) {
3596 nodekey.mv_size = mc->mc_db->md_pad;
3597 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3599 leaf = NODEPTR(mp, 0);
3600 MDB_SET_KEY(leaf, &nodekey);
3602 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3604 /* Probably happens rarely, but first node on the page
3605 * was the one we wanted.
3607 mc->mc_ki[mc->mc_top] = 0;
3608 leaf = NODEPTR(mp, 0);
3615 unsigned int nkeys = NUMKEYS(mp);
3617 if (mp->mp_flags & P_LEAF2) {
3618 nodekey.mv_data = LEAF2KEY(mp,
3619 nkeys-1, nodekey.mv_size);
3621 leaf = NODEPTR(mp, nkeys-1);
3622 MDB_SET_KEY(leaf, &nodekey);
3624 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3626 /* last node was the one we wanted */
3627 mc->mc_ki[mc->mc_top] = nkeys-1;
3628 leaf = NODEPTR(mp, nkeys-1);
3634 /* This is definitely the right page, skip search_page */
3639 /* If any parents have right-sibs, search.
3640 * Otherwise, there's nothing further.
3642 for (i=0; i<mc->mc_top; i++)
3644 NUMKEYS(mc->mc_pg[i])-1)
3646 if (i == mc->mc_top) {
3647 /* There are no other pages */
3648 mc->mc_ki[mc->mc_top] = nkeys;
3649 return MDB_NOTFOUND;
3653 /* There are no other pages */
3654 mc->mc_ki[mc->mc_top] = 0;
3655 return MDB_NOTFOUND;
3659 rc = mdb_page_search(mc, key, 0);
3660 if (rc != MDB_SUCCESS)
3663 mp = mc->mc_pg[mc->mc_top];
3664 assert(IS_LEAF(mp));
3667 leaf = mdb_node_search(mc, key, exactp);
3668 if (exactp != NULL && !*exactp) {
3669 /* MDB_SET specified and not an exact match. */
3670 return MDB_NOTFOUND;
3674 DPUTS("===> inexact leaf not found, goto sibling");
3675 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3676 return rc; /* no entries matched */
3677 mp = mc->mc_pg[mc->mc_top];
3678 assert(IS_LEAF(mp));
3679 leaf = NODEPTR(mp, 0);
3683 mc->mc_flags |= C_INITIALIZED;
3684 mc->mc_flags &= ~C_EOF;
3687 key->mv_size = mc->mc_db->md_pad;
3688 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3692 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3693 mdb_xcursor_init1(mc, leaf);
3696 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3697 if (op == MDB_SET || op == MDB_SET_RANGE) {
3698 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3701 if (op == MDB_GET_BOTH) {
3707 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3708 if (rc != MDB_SUCCESS)
3711 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3713 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3715 rc = mc->mc_dbx->md_dcmp(data, &d2);
3717 if (op == MDB_GET_BOTH || rc > 0)
3718 return MDB_NOTFOUND;
3723 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3724 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3729 /* The key already matches in all other cases */
3730 if (op == MDB_SET_RANGE)
3731 MDB_SET_KEY(leaf, key);
3732 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3737 /** Move the cursor to the first item in the database. */
3739 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3744 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3745 rc = mdb_page_search(mc, NULL, 0);
3746 if (rc != MDB_SUCCESS)
3749 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3751 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3752 mc->mc_flags |= C_INITIALIZED;
3753 mc->mc_flags &= ~C_EOF;
3755 mc->mc_ki[mc->mc_top] = 0;
3757 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3758 key->mv_size = mc->mc_db->md_pad;
3759 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3764 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3765 mdb_xcursor_init1(mc, leaf);
3766 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3771 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3772 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3776 MDB_SET_KEY(leaf, key);
3780 /** Move the cursor to the last item in the database. */
3782 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3788 lkey.mv_size = MAXKEYSIZE+1;
3789 lkey.mv_data = NULL;
3791 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3792 rc = mdb_page_search(mc, &lkey, 0);
3793 if (rc != MDB_SUCCESS)
3796 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3798 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3799 mc->mc_flags |= C_INITIALIZED;
3800 mc->mc_flags &= ~C_EOF;
3802 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3804 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3805 key->mv_size = mc->mc_db->md_pad;
3806 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3811 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3812 mdb_xcursor_init1(mc, leaf);
3813 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3818 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3819 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3824 MDB_SET_KEY(leaf, key);
3829 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3839 case MDB_GET_BOTH_RANGE:
3840 if (data == NULL || mc->mc_xcursor == NULL) {
3847 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3849 } else if (op == MDB_SET_RANGE)
3850 rc = mdb_cursor_set(mc, key, data, op, NULL);
3852 rc = mdb_cursor_set(mc, key, data, op, &exact);
3854 case MDB_GET_MULTIPLE:
3856 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
3857 !(mc->mc_flags & C_INITIALIZED)) {
3862 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
3863 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
3866 case MDB_NEXT_MULTIPLE:
3868 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
3872 if (!(mc->mc_flags & C_INITIALIZED))
3873 rc = mdb_cursor_first(mc, key, data);
3875 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
3876 if (rc == MDB_SUCCESS) {
3877 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
3880 mx = &mc->mc_xcursor->mx_cursor;
3881 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
3883 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
3884 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
3892 case MDB_NEXT_NODUP:
3893 if (!(mc->mc_flags & C_INITIALIZED))
3894 rc = mdb_cursor_first(mc, key, data);
3896 rc = mdb_cursor_next(mc, key, data, op);
3900 case MDB_PREV_NODUP:
3901 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
3902 rc = mdb_cursor_last(mc, key, data);
3904 rc = mdb_cursor_prev(mc, key, data, op);
3907 rc = mdb_cursor_first(mc, key, data);
3911 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3912 !(mc->mc_flags & C_INITIALIZED) ||
3913 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3917 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3920 rc = mdb_cursor_last(mc, key, data);
3924 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3925 !(mc->mc_flags & C_INITIALIZED) ||
3926 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3930 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3933 DPRINTF("unhandled/unimplemented cursor operation %u", op);
3941 /** Touch all the pages in the cursor stack.
3942 * Makes sure all the pages are writable, before attempting a write operation.
3943 * @param[in] mc The cursor to operate on.
3946 mdb_cursor_touch(MDB_cursor *mc)
3950 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
3952 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3953 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 1);
3956 *mc->mc_dbflag = DB_DIRTY;
3958 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
3959 rc = mdb_page_touch(mc);
3963 mc->mc_top = mc->mc_snum-1;
3968 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3971 MDB_node *leaf = NULL;
3972 MDB_val xdata, *rdata, dkey;
3976 unsigned int mcount = 0;
3979 char pbuf[MDB_PAGESIZE];
3980 char dbuf[MAXKEYSIZE+1];
3981 unsigned int nflags;
3984 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3987 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
3988 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
3992 if (flags == MDB_CURRENT) {
3993 if (!(mc->mc_flags & C_INITIALIZED))
3996 } else if (mc->mc_db->md_root == P_INVALID) {
3998 /* new database, write a root leaf page */
3999 DPUTS("allocating new root leaf page");
4000 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4004 mdb_cursor_push(mc, np);
4005 mc->mc_db->md_root = np->mp_pgno;
4006 mc->mc_db->md_depth++;
4007 *mc->mc_dbflag = DB_DIRTY;
4008 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4010 np->mp_flags |= P_LEAF2;
4011 mc->mc_flags |= C_INITIALIZED;
4017 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4018 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4019 DPRINTF("duplicate key [%s]", DKEY(key));
4021 return MDB_KEYEXIST;
4023 if (rc && rc != MDB_NOTFOUND)
4027 /* Cursor is positioned, now make sure all pages are writable */
4028 rc2 = mdb_cursor_touch(mc);
4033 /* The key already exists */
4034 if (rc == MDB_SUCCESS) {
4035 /* there's only a key anyway, so this is a no-op */
4036 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4037 unsigned int ksize = mc->mc_db->md_pad;
4038 if (key->mv_size != ksize)
4040 if (flags == MDB_CURRENT) {
4041 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4042 memcpy(ptr, key->mv_data, ksize);
4047 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4050 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4051 /* Was a single item before, must convert now */
4053 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4054 /* Just overwrite the current item */
4055 if (flags == MDB_CURRENT)
4058 dkey.mv_size = NODEDSZ(leaf);
4059 dkey.mv_data = NODEDATA(leaf);
4060 #if UINT_MAX < SIZE_MAX
4061 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4062 #ifdef MISALIGNED_OK
4063 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4065 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4068 /* if data matches, ignore it */
4069 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4070 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4072 /* create a fake page for the dup items */
4073 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4074 dkey.mv_data = dbuf;
4075 fp = (MDB_page *)pbuf;
4076 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4077 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4078 fp->mp_lower = PAGEHDRSZ;
4079 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4080 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4081 fp->mp_flags |= P_LEAF2;
4082 fp->mp_pad = data->mv_size;
4084 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4085 (dkey.mv_size & 1) + (data->mv_size & 1);
4087 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4090 xdata.mv_size = fp->mp_upper;
4091 xdata.mv_data = pbuf;
4095 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4096 /* See if we need to convert from fake page to subDB */
4098 unsigned int offset;
4101 fp = NODEDATA(leaf);
4102 if (flags == MDB_CURRENT) {
4103 fp->mp_flags |= P_DIRTY;
4104 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4105 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4109 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4110 offset = fp->mp_pad;
4112 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4114 offset += offset & 1;
4115 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4116 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4118 /* yes, convert it */
4120 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4121 dummy.md_pad = fp->mp_pad;
4122 dummy.md_flags = MDB_DUPFIXED;
4123 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4124 dummy.md_flags |= MDB_INTEGERKEY;
4127 dummy.md_branch_pages = 0;
4128 dummy.md_leaf_pages = 1;
4129 dummy.md_overflow_pages = 0;
4130 dummy.md_entries = NUMKEYS(fp);
4132 xdata.mv_size = sizeof(MDB_db);
4133 xdata.mv_data = &dummy;
4134 mp = mdb_page_alloc(mc, 1);
4137 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4138 flags |= F_DUPDATA|F_SUBDATA;
4139 dummy.md_root = mp->mp_pgno;
4141 /* no, just grow it */
4143 xdata.mv_size = NODEDSZ(leaf) + offset;
4144 xdata.mv_data = pbuf;
4145 mp = (MDB_page *)pbuf;
4146 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4149 mp->mp_flags = fp->mp_flags | P_DIRTY;
4150 mp->mp_pad = fp->mp_pad;
4151 mp->mp_lower = fp->mp_lower;
4152 mp->mp_upper = fp->mp_upper + offset;
4154 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4156 nsize = NODEDSZ(leaf) - fp->mp_upper;
4157 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4158 for (i=0; i<NUMKEYS(fp); i++)
4159 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4161 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4165 /* data is on sub-DB, just store it */
4166 flags |= F_DUPDATA|F_SUBDATA;
4170 /* same size, just replace it */
4171 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
4172 NODEDSZ(leaf) == data->mv_size) {
4173 if (F_ISSET(flags, MDB_RESERVE))
4174 data->mv_data = NODEDATA(leaf);
4176 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4179 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4181 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4187 nflags = flags & NODE_ADD_FLAGS;
4188 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4189 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4190 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4191 nflags &= ~MDB_APPEND;
4192 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4194 /* There is room already in this leaf page. */
4195 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4196 if (rc == 0 && !do_sub) {
4197 /* Adjust other cursors pointing to mp */
4198 MDB_cursor *m2, *m3;
4199 MDB_dbi dbi = mc->mc_dbi;
4200 unsigned i = mc->mc_top;
4201 MDB_page *mp = mc->mc_pg[i];
4203 if (mc->mc_flags & C_SUB)
4206 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4207 if (mc->mc_flags & C_SUB)
4208 m3 = &m2->mc_xcursor->mx_cursor;
4211 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4212 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4219 if (rc != MDB_SUCCESS)
4220 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4222 /* Now store the actual data in the child DB. Note that we're
4223 * storing the user data in the keys field, so there are strict
4224 * size limits on dupdata. The actual data fields of the child
4225 * DB are all zero size.
4233 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4234 if (flags & MDB_CURRENT) {
4235 xflags = MDB_CURRENT;
4237 mdb_xcursor_init1(mc, leaf);
4238 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4240 /* converted, write the original data first */
4242 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4246 /* Adjust other cursors pointing to mp */
4248 unsigned i = mc->mc_top;
4249 MDB_page *mp = mc->mc_pg[i];
4251 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4252 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4253 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4254 mdb_xcursor_init1(m2, leaf);
4259 xflags |= (flags & MDB_APPEND);
4260 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4261 if (flags & F_SUBDATA) {
4262 db = NODEDATA(leaf);
4263 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4266 /* sub-writes might have failed so check rc again.
4267 * Don't increment count if we just replaced an existing item.
4269 if (!rc && !(flags & MDB_CURRENT))
4270 mc->mc_db->md_entries++;
4271 if (flags & MDB_MULTIPLE) {
4273 if (mcount < data[1].mv_size) {
4274 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4275 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4285 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4290 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4293 if (!mc->mc_flags & C_INITIALIZED)
4296 rc = mdb_cursor_touch(mc);
4300 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4302 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4303 if (flags != MDB_NODUPDATA) {
4304 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4305 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4307 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4308 /* If sub-DB still has entries, we're done */
4309 if (mc->mc_xcursor->mx_db.md_entries) {
4310 if (leaf->mn_flags & F_SUBDATA) {
4311 /* update subDB info */
4312 MDB_db *db = NODEDATA(leaf);
4313 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4315 /* shrink fake page */
4316 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4318 mc->mc_db->md_entries--;
4321 /* otherwise fall thru and delete the sub-DB */
4324 if (leaf->mn_flags & F_SUBDATA) {
4325 /* add all the child DB's pages to the free list */
4326 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4327 if (rc == MDB_SUCCESS) {
4328 mc->mc_db->md_entries -=
4329 mc->mc_xcursor->mx_db.md_entries;
4334 return mdb_cursor_del0(mc, leaf);
4337 /** Allocate and initialize new pages for a database.
4338 * @param[in] mc a cursor on the database being added to.
4339 * @param[in] flags flags defining what type of page is being allocated.
4340 * @param[in] num the number of pages to allocate. This is usually 1,
4341 * unless allocating overflow pages for a large record.
4342 * @return Address of a page, or NULL on failure.
4345 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4349 if ((np = mdb_page_alloc(mc, num)) == NULL)
4351 DPRINTF("allocated new mpage %zu, page size %u",
4352 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4353 np->mp_flags = flags | P_DIRTY;
4354 np->mp_lower = PAGEHDRSZ;
4355 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4358 mc->mc_db->md_branch_pages++;
4359 else if (IS_LEAF(np))
4360 mc->mc_db->md_leaf_pages++;
4361 else if (IS_OVERFLOW(np)) {
4362 mc->mc_db->md_overflow_pages += num;
4369 /** Calculate the size of a leaf node.
4370 * The size depends on the environment's page size; if a data item
4371 * is too large it will be put onto an overflow page and the node
4372 * size will only include the key and not the data. Sizes are always
4373 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4374 * of the #MDB_node headers.
4375 * @param[in] env The environment handle.
4376 * @param[in] key The key for the node.
4377 * @param[in] data The data for the node.
4378 * @return The number of bytes needed to store the node.
4381 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4385 sz = LEAFSIZE(key, data);
4386 if (sz >= env->me_psize / MDB_MINKEYS) {
4387 /* put on overflow page */
4388 sz -= data->mv_size - sizeof(pgno_t);
4392 return sz + sizeof(indx_t);
4395 /** Calculate the size of a branch node.
4396 * The size should depend on the environment's page size but since
4397 * we currently don't support spilling large keys onto overflow
4398 * pages, it's simply the size of the #MDB_node header plus the
4399 * size of the key. Sizes are always rounded up to an even number
4400 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4401 * @param[in] env The environment handle.
4402 * @param[in] key The key for the node.
4403 * @return The number of bytes needed to store the node.
4406 mdb_branch_size(MDB_env *env, MDB_val *key)
4411 if (sz >= env->me_psize / MDB_MINKEYS) {
4412 /* put on overflow page */
4413 /* not implemented */
4414 /* sz -= key->size - sizeof(pgno_t); */
4417 return sz + sizeof(indx_t);
4420 /** Add a node to the page pointed to by the cursor.
4421 * @param[in] mc The cursor for this operation.
4422 * @param[in] indx The index on the page where the new node should be added.
4423 * @param[in] key The key for the new node.
4424 * @param[in] data The data for the new node, if any.
4425 * @param[in] pgno The page number, if adding a branch node.
4426 * @param[in] flags Flags for the node.
4427 * @return 0 on success, non-zero on failure. Possible errors are:
4429 * <li>ENOMEM - failed to allocate overflow pages for the node.
4430 * <li>ENOSPC - there is insufficient room in the page. This error
4431 * should never happen since all callers already calculate the
4432 * page's free space before calling this function.
4436 mdb_node_add(MDB_cursor *mc, indx_t indx,
4437 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4440 size_t node_size = NODESIZE;
4443 MDB_page *mp = mc->mc_pg[mc->mc_top];
4444 MDB_page *ofp = NULL; /* overflow page */
4447 assert(mp->mp_upper >= mp->mp_lower);
4449 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4450 IS_LEAF(mp) ? "leaf" : "branch",
4451 IS_SUBP(mp) ? "sub-" : "",
4452 mp->mp_pgno, indx, data ? data->mv_size : 0,
4453 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4456 /* Move higher keys up one slot. */
4457 int ksize = mc->mc_db->md_pad, dif;
4458 char *ptr = LEAF2KEY(mp, indx, ksize);
4459 dif = NUMKEYS(mp) - indx;
4461 memmove(ptr+ksize, ptr, dif*ksize);
4462 /* insert new key */
4463 memcpy(ptr, key->mv_data, ksize);
4465 /* Just using these for counting */
4466 mp->mp_lower += sizeof(indx_t);
4467 mp->mp_upper -= ksize - sizeof(indx_t);
4472 node_size += key->mv_size;
4476 if (F_ISSET(flags, F_BIGDATA)) {
4477 /* Data already on overflow page. */
4478 node_size += sizeof(pgno_t);
4479 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4480 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4481 /* Put data on overflow page. */
4482 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4483 data->mv_size, node_size+data->mv_size);
4484 node_size += sizeof(pgno_t);
4485 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4487 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4490 node_size += data->mv_size;
4493 node_size += node_size & 1;
4495 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4496 DPRINTF("not enough room in page %zu, got %u ptrs",
4497 mp->mp_pgno, NUMKEYS(mp));
4498 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4499 mp->mp_upper - mp->mp_lower);
4500 DPRINTF("node size = %zu", node_size);
4504 /* Move higher pointers up one slot. */
4505 for (i = NUMKEYS(mp); i > indx; i--)
4506 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4508 /* Adjust free space offsets. */
4509 ofs = mp->mp_upper - node_size;
4510 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4511 mp->mp_ptrs[indx] = ofs;
4513 mp->mp_lower += sizeof(indx_t);
4515 /* Write the node data. */
4516 node = NODEPTR(mp, indx);
4517 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4518 node->mn_flags = flags;
4520 SETDSZ(node,data->mv_size);
4525 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4530 if (F_ISSET(flags, F_BIGDATA))
4531 memcpy(node->mn_data + key->mv_size, data->mv_data,
4533 else if (F_ISSET(flags, MDB_RESERVE))
4534 data->mv_data = node->mn_data + key->mv_size;
4536 memcpy(node->mn_data + key->mv_size, data->mv_data,
4539 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4541 if (F_ISSET(flags, MDB_RESERVE))
4542 data->mv_data = METADATA(ofp);
4544 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4551 /** Delete the specified node from a page.
4552 * @param[in] mp The page to operate on.
4553 * @param[in] indx The index of the node to delete.
4554 * @param[in] ksize The size of a node. Only used if the page is
4555 * part of a #MDB_DUPFIXED database.
4558 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4561 indx_t i, j, numkeys, ptr;
4568 COPY_PGNO(pgno, mp->mp_pgno);
4569 DPRINTF("delete node %u on %s page %zu", indx,
4570 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4573 assert(indx < NUMKEYS(mp));
4576 int x = NUMKEYS(mp) - 1 - indx;
4577 base = LEAF2KEY(mp, indx, ksize);
4579 memmove(base, base + ksize, x * ksize);
4580 mp->mp_lower -= sizeof(indx_t);
4581 mp->mp_upper += ksize - sizeof(indx_t);
4585 node = NODEPTR(mp, indx);
4586 sz = NODESIZE + node->mn_ksize;
4588 if (F_ISSET(node->mn_flags, F_BIGDATA))
4589 sz += sizeof(pgno_t);
4591 sz += NODEDSZ(node);
4595 ptr = mp->mp_ptrs[indx];
4596 numkeys = NUMKEYS(mp);
4597 for (i = j = 0; i < numkeys; i++) {
4599 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4600 if (mp->mp_ptrs[i] < ptr)
4601 mp->mp_ptrs[j] += sz;
4606 base = (char *)mp + mp->mp_upper;
4607 memmove(base + sz, base, ptr - mp->mp_upper);
4609 mp->mp_lower -= sizeof(indx_t);
4613 /** Compact the main page after deleting a node on a subpage.
4614 * @param[in] mp The main page to operate on.
4615 * @param[in] indx The index of the subpage on the main page.
4618 mdb_node_shrink(MDB_page *mp, indx_t indx)
4625 indx_t i, numkeys, ptr;
4627 node = NODEPTR(mp, indx);
4628 sp = (MDB_page *)NODEDATA(node);
4629 osize = NODEDSZ(node);
4631 delta = sp->mp_upper - sp->mp_lower;
4632 SETDSZ(node, osize - delta);
4633 xp = (MDB_page *)((char *)sp + delta);
4635 /* shift subpage upward */
4637 nsize = NUMKEYS(sp) * sp->mp_pad;
4638 memmove(METADATA(xp), METADATA(sp), nsize);
4641 nsize = osize - sp->mp_upper;
4642 numkeys = NUMKEYS(sp);
4643 for (i=numkeys-1; i>=0; i--)
4644 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4646 xp->mp_upper = sp->mp_lower;
4647 xp->mp_lower = sp->mp_lower;
4648 xp->mp_flags = sp->mp_flags;
4649 xp->mp_pad = sp->mp_pad;
4650 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4652 /* shift lower nodes upward */
4653 ptr = mp->mp_ptrs[indx];
4654 numkeys = NUMKEYS(mp);
4655 for (i = 0; i < numkeys; i++) {
4656 if (mp->mp_ptrs[i] <= ptr)
4657 mp->mp_ptrs[i] += delta;
4660 base = (char *)mp + mp->mp_upper;
4661 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4662 mp->mp_upper += delta;
4665 /** Initial setup of a sorted-dups cursor.
4666 * Sorted duplicates are implemented as a sub-database for the given key.
4667 * The duplicate data items are actually keys of the sub-database.
4668 * Operations on the duplicate data items are performed using a sub-cursor
4669 * initialized when the sub-database is first accessed. This function does
4670 * the preliminary setup of the sub-cursor, filling in the fields that
4671 * depend only on the parent DB.
4672 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4675 mdb_xcursor_init0(MDB_cursor *mc)
4677 MDB_xcursor *mx = mc->mc_xcursor;
4679 mx->mx_cursor.mc_xcursor = NULL;
4680 mx->mx_cursor.mc_txn = mc->mc_txn;
4681 mx->mx_cursor.mc_db = &mx->mx_db;
4682 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4683 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4684 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4685 mx->mx_cursor.mc_snum = 0;
4686 mx->mx_cursor.mc_flags = C_SUB;
4687 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4688 mx->mx_dbx.md_dcmp = NULL;
4689 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4692 /** Final setup of a sorted-dups cursor.
4693 * Sets up the fields that depend on the data from the main cursor.
4694 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4695 * @param[in] node The data containing the #MDB_db record for the
4696 * sorted-dup database.
4699 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
4701 MDB_xcursor *mx = mc->mc_xcursor;
4703 if (node->mn_flags & F_SUBDATA) {
4704 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
4705 mx->mx_cursor.mc_snum = 0;
4706 mx->mx_cursor.mc_flags = C_SUB;
4708 MDB_page *fp = NODEDATA(node);
4709 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
4710 mx->mx_db.md_flags = 0;
4711 mx->mx_db.md_depth = 1;
4712 mx->mx_db.md_branch_pages = 0;
4713 mx->mx_db.md_leaf_pages = 1;
4714 mx->mx_db.md_overflow_pages = 0;
4715 mx->mx_db.md_entries = NUMKEYS(fp);
4716 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
4717 mx->mx_cursor.mc_snum = 1;
4718 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
4719 mx->mx_cursor.mc_top = 0;
4720 mx->mx_cursor.mc_pg[0] = fp;
4721 mx->mx_cursor.mc_ki[0] = 0;
4722 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4723 mx->mx_db.md_flags = MDB_DUPFIXED;
4724 mx->mx_db.md_pad = fp->mp_pad;
4725 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4726 mx->mx_db.md_flags |= MDB_INTEGERKEY;
4729 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
4731 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
4733 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
4734 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
4735 #if UINT_MAX < SIZE_MAX
4736 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
4737 #ifdef MISALIGNED_OK
4738 mx->mx_dbx.md_cmp = mdb_cmp_long;
4740 mx->mx_dbx.md_cmp = mdb_cmp_cint;
4745 /** Initialize a cursor for a given transaction and database. */
4747 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
4752 mc->mc_db = &txn->mt_dbs[dbi];
4753 mc->mc_dbx = &txn->mt_dbxs[dbi];
4754 mc->mc_dbflag = &txn->mt_dbflags[dbi];
4758 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4760 mc->mc_xcursor = mx;
4761 mdb_xcursor_init0(mc);
4763 mc->mc_xcursor = NULL;
4768 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
4771 MDB_xcursor *mx = NULL;
4772 size_t size = sizeof(MDB_cursor);
4774 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
4777 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
4778 size += sizeof(MDB_xcursor);
4780 if ((mc = malloc(size)) != NULL) {
4781 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4782 mx = (MDB_xcursor *)(mc + 1);
4784 mdb_cursor_init(mc, txn, dbi, mx);
4785 if (txn->mt_cursors) {
4786 mc->mc_next = txn->mt_cursors[dbi];
4787 txn->mt_cursors[dbi] = mc;
4789 mc->mc_flags |= C_ALLOCD;
4799 /* Return the count of duplicate data items for the current key */
4801 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
4805 if (mc == NULL || countp == NULL)
4808 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
4811 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4812 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4815 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
4818 *countp = mc->mc_xcursor->mx_db.md_entries;
4824 mdb_cursor_close(MDB_cursor *mc)
4827 /* remove from txn, if tracked */
4828 if (mc->mc_txn->mt_cursors) {
4829 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
4830 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
4832 *prev = mc->mc_next;
4834 if (mc->mc_flags & C_ALLOCD)
4840 mdb_cursor_txn(MDB_cursor *mc)
4842 if (!mc) return NULL;
4847 mdb_cursor_dbi(MDB_cursor *mc)
4853 /** Replace the key for a node with a new key.
4854 * @param[in] mp The page containing the node to operate on.
4855 * @param[in] indx The index of the node to operate on.
4856 * @param[in] key The new key to use.
4857 * @return 0 on success, non-zero on failure.
4860 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
4862 indx_t ptr, i, numkeys;
4869 node = NODEPTR(mp, indx);
4870 ptr = mp->mp_ptrs[indx];
4871 DPRINTF("update key %u (ofs %u) [%.*s] to [%s] on page %zu",
4873 (int)node->mn_ksize, (char *)NODEKEY(node),
4877 delta = key->mv_size - node->mn_ksize;
4879 if (delta > 0 && SIZELEFT(mp) < delta) {
4880 DPRINTF("OUCH! Not enough room, delta = %d", delta);
4884 numkeys = NUMKEYS(mp);
4885 for (i = 0; i < numkeys; i++) {
4886 if (mp->mp_ptrs[i] <= ptr)
4887 mp->mp_ptrs[i] -= delta;
4890 base = (char *)mp + mp->mp_upper;
4891 len = ptr - mp->mp_upper + NODESIZE;
4892 memmove(base - delta, base, len);
4893 mp->mp_upper -= delta;
4895 node = NODEPTR(mp, indx);
4896 node->mn_ksize = key->mv_size;
4899 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4904 /** Move a node from csrc to cdst.
4907 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
4914 /* Mark src and dst as dirty. */
4915 if ((rc = mdb_page_touch(csrc)) ||
4916 (rc = mdb_page_touch(cdst)))
4919 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
4920 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
4921 key.mv_size = csrc->mc_db->md_pad;
4922 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
4924 data.mv_data = NULL;
4926 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
4927 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
4928 unsigned int snum = csrc->mc_snum;
4930 /* must find the lowest key below src */
4931 mdb_page_search_root(csrc, NULL, 0);
4932 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
4933 key.mv_size = NODEKSZ(s2);
4934 key.mv_data = NODEKEY(s2);
4935 csrc->mc_snum = snum--;
4936 csrc->mc_top = snum;
4938 key.mv_size = NODEKSZ(srcnode);
4939 key.mv_data = NODEKEY(srcnode);
4941 data.mv_size = NODEDSZ(srcnode);
4942 data.mv_data = NODEDATA(srcnode);
4944 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
4945 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
4946 csrc->mc_ki[csrc->mc_top],
4948 csrc->mc_pg[csrc->mc_top]->mp_pgno,
4949 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
4951 /* Add the node to the destination page.
4953 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, NODEPGNO(srcnode),
4955 if (rc != MDB_SUCCESS)
4958 /* Delete the node from the source page.
4960 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
4963 /* Adjust other cursors pointing to mp */
4964 MDB_cursor *m2, *m3;
4965 MDB_dbi dbi = csrc->mc_dbi;
4966 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
4968 if (csrc->mc_flags & C_SUB)
4971 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4972 if (m2 == csrc) continue;
4973 if (csrc->mc_flags & C_SUB)
4974 m3 = &m2->mc_xcursor->mx_cursor;
4977 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
4978 csrc->mc_ki[csrc->mc_top]) {
4979 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
4980 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
4985 /* Update the parent separators.
4987 if (csrc->mc_ki[csrc->mc_top] == 0) {
4988 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
4989 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
4990 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
4992 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
4993 key.mv_size = NODEKSZ(srcnode);
4994 key.mv_data = NODEKEY(srcnode);
4996 DPRINTF("update separator for source page %zu to [%s]",
4997 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
4998 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
4999 &key)) != MDB_SUCCESS)
5002 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5004 nullkey.mv_size = 0;
5005 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5006 assert(rc == MDB_SUCCESS);
5010 if (cdst->mc_ki[cdst->mc_top] == 0) {
5011 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5012 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5013 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5015 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5016 key.mv_size = NODEKSZ(srcnode);
5017 key.mv_data = NODEKEY(srcnode);
5019 DPRINTF("update separator for destination page %zu to [%s]",
5020 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5021 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5022 &key)) != MDB_SUCCESS)
5025 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5027 nullkey.mv_size = 0;
5028 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5029 assert(rc == MDB_SUCCESS);
5036 /** Merge one page into another.
5037 * The nodes from the page pointed to by \b csrc will
5038 * be copied to the page pointed to by \b cdst and then
5039 * the \b csrc page will be freed.
5040 * @param[in] csrc Cursor pointing to the source page.
5041 * @param[in] cdst Cursor pointing to the destination page.
5044 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5052 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5053 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5055 assert(csrc->mc_snum > 1); /* can't merge root page */
5056 assert(cdst->mc_snum > 1);
5058 /* Mark dst as dirty. */
5059 if ((rc = mdb_page_touch(cdst)))
5062 /* Move all nodes from src to dst.
5064 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5065 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5066 key.mv_size = csrc->mc_db->md_pad;
5067 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5068 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5069 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5070 if (rc != MDB_SUCCESS)
5072 key.mv_data = (char *)key.mv_data + key.mv_size;
5075 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5076 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5078 key.mv_size = srcnode->mn_ksize;
5079 key.mv_data = NODEKEY(srcnode);
5080 data.mv_size = NODEDSZ(srcnode);
5081 data.mv_data = NODEDATA(srcnode);
5082 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5083 if (rc != MDB_SUCCESS)
5088 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5089 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);
5091 /* Unlink the src page from parent and add to free list.
5093 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5094 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5096 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5100 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5101 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5102 csrc->mc_db->md_leaf_pages--;
5104 csrc->mc_db->md_branch_pages--;
5106 /* Adjust other cursors pointing to mp */
5107 MDB_cursor *m2, *m3;
5108 MDB_dbi dbi = csrc->mc_dbi;
5109 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5111 if (csrc->mc_flags & C_SUB)
5114 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5115 if (m2 == csrc) continue;
5116 if (csrc->mc_flags & C_SUB)
5117 m3 = &m2->mc_xcursor->mx_cursor;
5120 if (m3->mc_snum < csrc->mc_snum) continue;
5121 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5122 m3->mc_pg[csrc->mc_top] = mp;
5123 m3->mc_ki[csrc->mc_top] += nkeys;
5127 mdb_cursor_pop(csrc);
5129 return mdb_rebalance(csrc);
5132 /** Copy the contents of a cursor.
5133 * @param[in] csrc The cursor to copy from.
5134 * @param[out] cdst The cursor to copy to.
5137 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5141 cdst->mc_txn = csrc->mc_txn;
5142 cdst->mc_dbi = csrc->mc_dbi;
5143 cdst->mc_db = csrc->mc_db;
5144 cdst->mc_dbx = csrc->mc_dbx;
5145 cdst->mc_snum = csrc->mc_snum;
5146 cdst->mc_top = csrc->mc_top;
5147 cdst->mc_flags = csrc->mc_flags;
5149 for (i=0; i<csrc->mc_snum; i++) {
5150 cdst->mc_pg[i] = csrc->mc_pg[i];
5151 cdst->mc_ki[i] = csrc->mc_ki[i];
5155 /** Rebalance the tree after a delete operation.
5156 * @param[in] mc Cursor pointing to the page where rebalancing
5158 * @return 0 on success, non-zero on failure.
5161 mdb_rebalance(MDB_cursor *mc)
5171 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5172 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5173 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5174 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5178 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5181 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5182 DPRINTF("no need to rebalance page %zu, above fill threshold",
5188 if (mc->mc_snum < 2) {
5189 MDB_page *mp = mc->mc_pg[0];
5190 if (NUMKEYS(mp) == 0) {
5191 DPUTS("tree is completely empty");
5192 mc->mc_db->md_root = P_INVALID;
5193 mc->mc_db->md_depth = 0;
5194 mc->mc_db->md_leaf_pages = 0;
5195 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5199 /* Adjust other cursors pointing to mp */
5200 MDB_cursor *m2, *m3;
5201 MDB_dbi dbi = mc->mc_dbi;
5203 if (mc->mc_flags & C_SUB)
5206 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5207 if (m2 == mc) continue;
5208 if (mc->mc_flags & C_SUB)
5209 m3 = &m2->mc_xcursor->mx_cursor;
5212 if (m3->mc_snum < mc->mc_snum) continue;
5213 if (m3->mc_pg[0] == mp) {
5219 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5220 DPUTS("collapsing root page!");
5221 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5222 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5223 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5226 mc->mc_db->md_depth--;
5227 mc->mc_db->md_branch_pages--;
5229 /* Adjust other cursors pointing to mp */
5230 MDB_cursor *m2, *m3;
5231 MDB_dbi dbi = mc->mc_dbi;
5233 if (mc->mc_flags & C_SUB)
5236 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5237 if (m2 == mc) continue;
5238 if (mc->mc_flags & C_SUB)
5239 m3 = &m2->mc_xcursor->mx_cursor;
5242 if (m3->mc_snum < mc->mc_snum) continue;
5243 if (m3->mc_pg[0] == mp) {
5244 m3->mc_pg[0] = mc->mc_pg[0];
5249 DPUTS("root page doesn't need rebalancing");
5253 /* The parent (branch page) must have at least 2 pointers,
5254 * otherwise the tree is invalid.
5256 ptop = mc->mc_top-1;
5257 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5259 /* Leaf page fill factor is below the threshold.
5260 * Try to move keys from left or right neighbor, or
5261 * merge with a neighbor page.
5266 mdb_cursor_copy(mc, &mn);
5267 mn.mc_xcursor = NULL;
5269 if (mc->mc_ki[ptop] == 0) {
5270 /* We're the leftmost leaf in our parent.
5272 DPUTS("reading right neighbor");
5274 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5275 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5277 mn.mc_ki[mn.mc_top] = 0;
5278 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5280 /* There is at least one neighbor to the left.
5282 DPUTS("reading left neighbor");
5284 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5285 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5287 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5288 mc->mc_ki[mc->mc_top] = 0;
5291 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5292 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);
5294 /* If the neighbor page is above threshold and has at least two
5295 * keys, move one key from it.
5297 * Otherwise we should try to merge them.
5299 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5300 return mdb_node_move(&mn, mc);
5301 else { /* FIXME: if (has_enough_room()) */
5302 mc->mc_flags &= ~C_INITIALIZED;
5303 if (mc->mc_ki[ptop] == 0)
5304 return mdb_page_merge(&mn, mc);
5306 return mdb_page_merge(mc, &mn);
5310 /** Complete a delete operation started by #mdb_cursor_del(). */
5312 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5316 /* add overflow pages to free list */
5317 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5321 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5322 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5323 for (i=0; i<ovpages; i++) {
5324 DPRINTF("freed ov page %zu", pg);
5325 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5329 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5330 mc->mc_db->md_entries--;
5331 rc = mdb_rebalance(mc);
5332 if (rc != MDB_SUCCESS)
5333 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5339 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5340 MDB_val *key, MDB_val *data)
5345 MDB_val rdata, *xdata;
5349 assert(key != NULL);
5351 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5353 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5356 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5360 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5364 mdb_cursor_init(&mc, txn, dbi, &mx);
5375 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5377 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5381 /** Split a page and insert a new node.
5382 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5383 * The cursor will be updated to point to the actual page and index where
5384 * the node got inserted after the split.
5385 * @param[in] newkey The key for the newly inserted node.
5386 * @param[in] newdata The data for the newly inserted node.
5387 * @param[in] newpgno The page number, if the new node is a branch node.
5388 * @return 0 on success, non-zero on failure.
5391 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5392 unsigned int nflags)
5395 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0;
5398 unsigned int i, j, split_indx, nkeys, pmax;
5400 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5402 MDB_page *mp, *rp, *pp;
5407 mp = mc->mc_pg[mc->mc_top];
5408 newindx = mc->mc_ki[mc->mc_top];
5410 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5411 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5412 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5414 if (mc->mc_snum < 2) {
5415 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5417 /* shift current top to make room for new parent */
5418 mc->mc_pg[1] = mc->mc_pg[0];
5419 mc->mc_ki[1] = mc->mc_ki[0];
5422 mc->mc_db->md_root = pp->mp_pgno;
5423 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5424 mc->mc_db->md_depth++;
5427 /* Add left (implicit) pointer. */
5428 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5429 /* undo the pre-push */
5430 mc->mc_pg[0] = mc->mc_pg[1];
5431 mc->mc_ki[0] = mc->mc_ki[1];
5432 mc->mc_db->md_root = mp->mp_pgno;
5433 mc->mc_db->md_depth--;
5440 ptop = mc->mc_top-1;
5441 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5444 /* Create a right sibling. */
5445 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5447 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5449 mdb_cursor_copy(mc, &mn);
5450 mn.mc_pg[mn.mc_top] = rp;
5451 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5453 if (nflags & MDB_APPEND) {
5454 mn.mc_ki[mn.mc_top] = 0;
5461 nkeys = NUMKEYS(mp);
5462 split_indx = nkeys / 2 + 1;
5467 unsigned int lsize, rsize, ksize;
5468 /* Move half of the keys to the right sibling */
5470 x = mc->mc_ki[mc->mc_top] - split_indx;
5471 ksize = mc->mc_db->md_pad;
5472 split = LEAF2KEY(mp, split_indx, ksize);
5473 rsize = (nkeys - split_indx) * ksize;
5474 lsize = (nkeys - split_indx) * sizeof(indx_t);
5475 mp->mp_lower -= lsize;
5476 rp->mp_lower += lsize;
5477 mp->mp_upper += rsize - lsize;
5478 rp->mp_upper -= rsize - lsize;
5479 sepkey.mv_size = ksize;
5480 if (newindx == split_indx) {
5481 sepkey.mv_data = newkey->mv_data;
5483 sepkey.mv_data = split;
5486 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5487 memcpy(rp->mp_ptrs, split, rsize);
5488 sepkey.mv_data = rp->mp_ptrs;
5489 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5490 memcpy(ins, newkey->mv_data, ksize);
5491 mp->mp_lower += sizeof(indx_t);
5492 mp->mp_upper -= ksize - sizeof(indx_t);
5495 memcpy(rp->mp_ptrs, split, x * ksize);
5496 ins = LEAF2KEY(rp, x, ksize);
5497 memcpy(ins, newkey->mv_data, ksize);
5498 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5499 rp->mp_lower += sizeof(indx_t);
5500 rp->mp_upper -= ksize - sizeof(indx_t);
5501 mc->mc_ki[mc->mc_top] = x;
5502 mc->mc_pg[mc->mc_top] = rp;
5507 /* For leaf pages, check the split point based on what
5508 * fits where, since otherwise add_node can fail.
5511 unsigned int psize, nsize;
5512 /* Maximum free space in an empty page */
5513 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5514 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5515 if (newindx < split_indx) {
5517 for (i=0; i<split_indx; i++) {
5518 node = NODEPTR(mp, i);
5519 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5520 if (F_ISSET(node->mn_flags, F_BIGDATA))
5521 psize += sizeof(pgno_t);
5523 psize += NODEDSZ(node);
5532 for (i=nkeys-1; i>=split_indx; i--) {
5533 node = NODEPTR(mp, i);
5534 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5535 if (F_ISSET(node->mn_flags, F_BIGDATA))
5536 psize += sizeof(pgno_t);
5538 psize += NODEDSZ(node);
5548 /* First find the separating key between the split pages.
5550 if (newindx == split_indx) {
5551 sepkey.mv_size = newkey->mv_size;
5552 sepkey.mv_data = newkey->mv_data;
5554 node = NODEPTR(mp, split_indx);
5555 sepkey.mv_size = node->mn_ksize;
5556 sepkey.mv_data = NODEKEY(node);
5560 DPRINTF("separator is [%s]", DKEY(&sepkey));
5562 /* Copy separator key to the parent.
5564 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5567 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5569 /* Right page might now have changed parent.
5570 * Check if left page also changed parent.
5572 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5573 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5574 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5575 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5579 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5582 if (rc != MDB_SUCCESS) {
5585 if (nflags & MDB_APPEND) {
5586 mc->mc_pg[mc->mc_top] = rp;
5587 mc->mc_ki[mc->mc_top] = 0;
5588 return mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
5594 /* Move half of the keys to the right sibling. */
5596 /* grab a page to hold a temporary copy */
5597 copy = mdb_page_malloc(mc);
5601 copy->mp_pgno = mp->mp_pgno;
5602 copy->mp_flags = mp->mp_flags;
5603 copy->mp_lower = PAGEHDRSZ;
5604 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
5605 mc->mc_pg[mc->mc_top] = copy;
5606 for (i = j = 0; i <= nkeys; j++) {
5607 if (i == split_indx) {
5608 /* Insert in right sibling. */
5609 /* Reset insert index for right sibling. */
5610 j = (i == newindx && ins_new);
5611 mc->mc_pg[mc->mc_top] = rp;
5614 if (i == newindx && !ins_new) {
5615 /* Insert the original entry that caused the split. */
5616 rkey.mv_data = newkey->mv_data;
5617 rkey.mv_size = newkey->mv_size;
5626 /* Update page and index for the new key. */
5628 mc->mc_pg[mc->mc_top] = copy;
5629 mc->mc_ki[mc->mc_top] = j;
5630 } else if (i == nkeys) {
5633 node = NODEPTR(mp, i);
5634 rkey.mv_data = NODEKEY(node);
5635 rkey.mv_size = node->mn_ksize;
5637 xdata.mv_data = NODEDATA(node);
5638 xdata.mv_size = NODEDSZ(node);
5641 pgno = NODEPGNO(node);
5642 flags = node->mn_flags;
5647 if (!IS_LEAF(mp) && j == 0) {
5648 /* First branch index doesn't need key data. */
5652 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
5655 nkeys = NUMKEYS(copy);
5656 for (i=0; i<nkeys; i++)
5657 mp->mp_ptrs[i] = copy->mp_ptrs[i];
5658 mp->mp_lower = copy->mp_lower;
5659 mp->mp_upper = copy->mp_upper;
5660 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
5661 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
5663 /* reset back to original page */
5664 if (!newindx || (newindx < split_indx)) {
5665 mc->mc_pg[mc->mc_top] = mp;
5666 if (nflags & MDB_RESERVE) {
5667 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5668 if (!(node->mn_flags & F_BIGDATA))
5669 newdata->mv_data = NODEDATA(node);
5673 /* return tmp page to freelist */
5674 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
5675 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
5676 mc->mc_txn->mt_env->me_dpages = copy;
5679 /* Adjust other cursors pointing to mp */
5680 MDB_cursor *m2, *m3;
5681 MDB_dbi dbi = mc->mc_dbi;
5683 if (mc->mc_flags & C_SUB)
5686 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5687 if (m2 == mc) continue;
5688 if (mc->mc_flags & C_SUB)
5689 m3 = &m2->mc_xcursor->mx_cursor;
5692 if (!(m3->mc_flags & C_INITIALIZED))
5696 for (i=m3->mc_top; i>0; i--) {
5697 m3->mc_ki[i+1] = m3->mc_ki[i];
5698 m3->mc_pg[i+1] = m3->mc_pg[i];
5700 m3->mc_ki[0] = mc->mc_ki[0];
5701 m3->mc_pg[0] = mc->mc_pg[0];
5705 if (m3->mc_pg[mc->mc_top] == mp) {
5706 if (m3->mc_ki[m3->mc_top] >= split_indx) {
5707 m3->mc_pg[m3->mc_top] = rp;
5708 m3->mc_ki[m3->mc_top] -= split_indx;
5717 mdb_put(MDB_txn *txn, MDB_dbi dbi,
5718 MDB_val *key, MDB_val *data, unsigned int flags)
5723 assert(key != NULL);
5724 assert(data != NULL);
5726 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5729 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5733 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5737 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
5740 mdb_cursor_init(&mc, txn, dbi, &mx);
5741 return mdb_cursor_put(&mc, key, data, flags);
5744 /** Only a subset of the @ref mdb_env flags can be changed
5745 * at runtime. Changing other flags requires closing the environment
5746 * and re-opening it with the new flags.
5748 #define CHANGEABLE (MDB_NOSYNC)
5750 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
5752 if ((flag & CHANGEABLE) != flag)
5755 env->me_flags |= flag;
5757 env->me_flags &= ~flag;
5762 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
5767 *arg = env->me_flags;
5772 mdb_env_get_path(MDB_env *env, const char **arg)
5777 *arg = env->me_path;
5781 /** Common code for #mdb_stat() and #mdb_env_stat().
5782 * @param[in] env the environment to operate in.
5783 * @param[in] db the #MDB_db record containing the stats to return.
5784 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
5785 * @return 0, this function always succeeds.
5788 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
5790 arg->ms_psize = env->me_psize;
5791 arg->ms_depth = db->md_depth;
5792 arg->ms_branch_pages = db->md_branch_pages;
5793 arg->ms_leaf_pages = db->md_leaf_pages;
5794 arg->ms_overflow_pages = db->md_overflow_pages;
5795 arg->ms_entries = db->md_entries;
5800 mdb_env_stat(MDB_env *env, MDB_stat *arg)
5804 if (env == NULL || arg == NULL)
5807 mdb_env_read_meta(env, &toggle);
5809 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
5812 /** Set the default comparison functions for a database.
5813 * Called immediately after a database is opened to set the defaults.
5814 * The user can then override them with #mdb_set_compare() or
5815 * #mdb_set_dupsort().
5816 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
5817 * @param[in] dbi A database handle returned by #mdb_open()
5820 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
5822 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
5823 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memnr;
5824 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
5825 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_cint;
5827 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memn;
5829 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5830 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
5831 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
5832 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_int;
5834 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_cint;
5835 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
5836 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memnr;
5838 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memn;
5841 txn->mt_dbxs[dbi].md_dcmp = NULL;
5845 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
5850 int rc, dbflag, exact;
5853 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
5854 mdb_default_cmp(txn, FREE_DBI);
5860 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
5861 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
5862 mdb_default_cmp(txn, MAIN_DBI);
5866 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
5867 mdb_default_cmp(txn, MAIN_DBI);
5870 /* Is the DB already open? */
5872 for (i=2; i<txn->mt_numdbs; i++) {
5873 if (len == txn->mt_dbxs[i].md_name.mv_size &&
5874 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
5880 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
5883 /* Find the DB info */
5887 key.mv_data = (void *)name;
5888 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
5889 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
5890 if (rc == MDB_SUCCESS) {
5891 /* make sure this is actually a DB */
5892 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
5893 if (!(node->mn_flags & F_SUBDATA))
5895 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
5896 /* Create if requested */
5898 data.mv_size = sizeof(MDB_db);
5899 data.mv_data = &dummy;
5900 memset(&dummy, 0, sizeof(dummy));
5901 dummy.md_root = P_INVALID;
5902 dummy.md_flags = flags & 0xffff;
5903 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
5907 /* OK, got info, add to table */
5908 if (rc == MDB_SUCCESS) {
5909 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
5910 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
5911 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
5912 txn->mt_dbflags[txn->mt_numdbs] = dbflag;
5913 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
5914 *dbi = txn->mt_numdbs;
5915 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
5916 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
5917 mdb_default_cmp(txn, txn->mt_numdbs);
5924 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
5926 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
5929 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
5932 void mdb_close(MDB_env *env, MDB_dbi dbi)
5935 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
5937 ptr = env->me_dbxs[dbi].md_name.mv_data;
5938 env->me_dbxs[dbi].md_name.mv_data = NULL;
5939 env->me_dbxs[dbi].md_name.mv_size = 0;
5943 /** Add all the DB's pages to the free list.
5944 * @param[in] mc Cursor on the DB to free.
5945 * @param[in] subs non-Zero to check for sub-DBs in this DB.
5946 * @return 0 on success, non-zero on failure.
5949 mdb_drop0(MDB_cursor *mc, int subs)
5953 rc = mdb_page_search(mc, NULL, 0);
5954 if (rc == MDB_SUCCESS) {
5959 /* LEAF2 pages have no nodes, cannot have sub-DBs */
5960 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
5963 mdb_cursor_copy(mc, &mx);
5964 while (mc->mc_snum > 0) {
5965 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
5966 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
5967 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
5968 if (ni->mn_flags & F_SUBDATA) {
5969 mdb_xcursor_init1(mc, ni);
5970 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5976 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
5978 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
5981 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5986 rc = mdb_cursor_sibling(mc, 1);
5988 /* no more siblings, go back to beginning
5989 * of previous level. (stack was already popped
5990 * by mdb_cursor_sibling)
5992 for (i=1; i<mc->mc_top; i++)
5993 mc->mc_pg[i] = mx.mc_pg[i];
5997 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
5998 mc->mc_db->md_root);
6003 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6008 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6011 rc = mdb_cursor_open(txn, dbi, &mc);
6015 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6019 /* Can't delete the main DB */
6020 if (del && dbi > MAIN_DBI) {
6021 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6023 mdb_close(txn->mt_env, dbi);
6025 txn->mt_dbflags[dbi] |= DB_DIRTY;
6026 txn->mt_dbs[dbi].md_depth = 0;
6027 txn->mt_dbs[dbi].md_branch_pages = 0;
6028 txn->mt_dbs[dbi].md_leaf_pages = 0;
6029 txn->mt_dbs[dbi].md_overflow_pages = 0;
6030 txn->mt_dbs[dbi].md_entries = 0;
6031 txn->mt_dbs[dbi].md_root = P_INVALID;
6034 mdb_cursor_close(mc);
6038 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6040 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6043 txn->mt_dbxs[dbi].md_cmp = cmp;
6047 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6049 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6052 txn->mt_dbxs[dbi].md_dcmp = cmp;
6056 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6058 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6061 txn->mt_dbxs[dbi].md_rel = rel;
6065 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6067 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6070 txn->mt_dbxs[dbi].md_relctx = ctx;