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__)
206 #define MNAME_LEN (sizeof(pthread_mutex_t))
212 /** A flag for opening a file and requesting synchronous data writes.
213 * This is only used when writing a meta page. It's not strictly needed;
214 * we could just do a normal write and then immediately perform a flush.
215 * But if this flag is available it saves us an extra system call.
217 * @note If O_DSYNC is undefined but exists in /usr/include,
218 * preferably set some compiler flag to get the definition.
219 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
222 # define MDB_DSYNC O_DSYNC
226 /** Function for flushing the data of a file. Define this to fsync
227 * if fdatasync() is not supported.
229 #ifndef MDB_FDATASYNC
230 # define MDB_FDATASYNC fdatasync
233 /** A page number in the database.
234 * Note that 64 bit page numbers are overkill, since pages themselves
235 * already represent 12-13 bits of addressable memory, and the OS will
236 * always limit applications to a maximum of 63 bits of address space.
238 * @note In the #MDB_node structure, we only store 48 bits of this value,
239 * which thus limits us to only 60 bits of addressable data.
243 /** A transaction ID.
244 * See struct MDB_txn.mt_txnid for details.
248 /** @defgroup debug Debug Macros
252 /** Enable debug output.
253 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
254 * read from and written to the database (used for free space management).
259 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
260 # define DPRINTF (void) /* Vararg macros may be unsupported */
262 /** Print a debug message with printf formatting. */
263 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
264 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
266 # define DPRINTF(fmt, ...) ((void) 0)
268 /** Print a debug string.
269 * The string is printed literally, with no format processing.
271 #define DPUTS(arg) DPRINTF("%s", arg)
274 /** A default memory page size.
275 * The actual size is platform-dependent, but we use this for
276 * boot-strapping. We probably should not be using this any more.
277 * The #GET_PAGESIZE() macro is used to get the actual size.
279 * Note that we don't currently support Huge pages. On Linux,
280 * regular data files cannot use Huge pages, and in general
281 * Huge pages aren't actually pageable. We rely on the OS
282 * demand-pager to read our data and page it out when memory
283 * pressure from other processes is high. So until OSs have
284 * actual paging support for Huge pages, they're not viable.
286 #define MDB_PAGESIZE 4096
288 /** The minimum number of keys required in a database page.
289 * Setting this to a larger value will place a smaller bound on the
290 * maximum size of a data item. Data items larger than this size will
291 * be pushed into overflow pages instead of being stored directly in
292 * the B-tree node. This value used to default to 4. With a page size
293 * of 4096 bytes that meant that any item larger than 1024 bytes would
294 * go into an overflow page. That also meant that on average 2-3KB of
295 * each overflow page was wasted space. The value cannot be lower than
296 * 2 because then there would no longer be a tree structure. With this
297 * value, items larger than 2KB will go into overflow pages, and on
298 * average only 1KB will be wasted.
300 #define MDB_MINKEYS 2
302 /** A stamp that identifies a file as an MDB file.
303 * There's nothing special about this value other than that it is easily
304 * recognizable, and it will reflect any byte order mismatches.
306 #define MDB_MAGIC 0xBEEFC0DE
308 /** The version number for a database's file format. */
309 #define MDB_VERSION 1
311 /** The maximum size of a key in the database.
312 * While data items have essentially unbounded size, we require that
313 * keys all fit onto a regular page. This limit could be raised a bit
314 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
316 #define MAXKEYSIZE 511
321 * This is used for printing a hex dump of a key's contents.
323 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
324 /** Display a key in hex.
326 * Invoke a function to display a key in hex.
328 #define DKEY(x) mdb_dkey(x, kbuf)
330 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
334 /** @defgroup lazylock Lazy Locking
335 * Macros for locks that aren't actually needed.
336 * The DB view is always consistent because all writes are wrapped in
337 * the wmutex. Finer-grained locks aren't necessary.
341 /** Use lazy locking. I.e., don't lock these accesses at all. */
345 /** Grab the reader lock */
346 #define LAZY_MUTEX_LOCK(x)
347 /** Release the reader lock */
348 #define LAZY_MUTEX_UNLOCK(x)
349 /** Release the DB table reader/writer lock */
350 #define LAZY_RWLOCK_UNLOCK(x)
351 /** Grab the DB table write lock */
352 #define LAZY_RWLOCK_WRLOCK(x)
353 /** Grab the DB table read lock */
354 #define LAZY_RWLOCK_RDLOCK(x)
355 /** Declare the DB table rwlock. Should not be followed by ';'. */
356 #define LAZY_RWLOCK_DEF(x)
357 /** Initialize the DB table rwlock */
358 #define LAZY_RWLOCK_INIT(x,y)
359 /** Destroy the DB table rwlock */
360 #define LAZY_RWLOCK_DESTROY(x)
362 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
363 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
364 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
365 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
366 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
367 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x;
368 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
369 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
373 /** An invalid page number.
374 * Mainly used to denote an empty tree.
376 #define P_INVALID (~0UL)
378 /** Test if a flag \b f is set in a flag word \b w. */
379 #define F_ISSET(w, f) (((w) & (f)) == (f))
381 /** Used for offsets within a single page.
382 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
385 typedef uint16_t indx_t;
387 /** Default size of memory map.
388 * This is certainly too small for any actual applications. Apps should always set
389 * the size explicitly using #mdb_env_set_mapsize().
391 #define DEFAULT_MAPSIZE 1048576
393 /** @defgroup readers Reader Lock Table
394 * Readers don't acquire any locks for their data access. Instead, they
395 * simply record their transaction ID in the reader table. The reader
396 * mutex is needed just to find an empty slot in the reader table. The
397 * slot's address is saved in thread-specific data so that subsequent read
398 * transactions started by the same thread need no further locking to proceed.
400 * Since the database uses multi-version concurrency control, readers don't
401 * actually need any locking. This table is used to keep track of which
402 * readers are using data from which old transactions, so that we'll know
403 * when a particular old transaction is no longer in use. Old transactions
404 * that have discarded any data pages can then have those pages reclaimed
405 * for use by a later write transaction.
407 * The lock table is constructed such that reader slots are aligned with the
408 * processor's cache line size. Any slot is only ever used by one thread.
409 * This alignment guarantees that there will be no contention or cache
410 * thrashing as threads update their own slot info, and also eliminates
411 * any need for locking when accessing a slot.
413 * A writer thread will scan every slot in the table to determine the oldest
414 * outstanding reader transaction. Any freed pages older than this will be
415 * reclaimed by the writer. The writer doesn't use any locks when scanning
416 * this table. This means that there's no guarantee that the writer will
417 * see the most up-to-date reader info, but that's not required for correct
418 * operation - all we need is to know the upper bound on the oldest reader,
419 * we don't care at all about the newest reader. So the only consequence of
420 * reading stale information here is that old pages might hang around a
421 * while longer before being reclaimed. That's actually good anyway, because
422 * the longer we delay reclaiming old pages, the more likely it is that a
423 * string of contiguous pages can be found after coalescing old pages from
424 * many old transactions together.
426 * @todo We don't actually do such coalescing yet, we grab pages from one
427 * old transaction at a time.
430 /** Number of slots in the reader table.
431 * This value was chosen somewhat arbitrarily. 126 readers plus a
432 * couple mutexes fit exactly into 8KB on my development machine.
433 * Applications should set the table size using #mdb_env_set_maxreaders().
435 #define DEFAULT_READERS 126
437 /** The size of a CPU cache line in bytes. We want our lock structures
438 * aligned to this size to avoid false cache line sharing in the
440 * This value works for most CPUs. For Itanium this should be 128.
446 /** The information we store in a single slot of the reader table.
447 * In addition to a transaction ID, we also record the process and
448 * thread ID that owns a slot, so that we can detect stale information,
449 * e.g. threads or processes that went away without cleaning up.
450 * @note We currently don't check for stale records. We simply re-init
451 * the table when we know that we're the only process opening the
454 typedef struct MDB_rxbody {
455 /** The current Transaction ID when this transaction began.
456 * Multiple readers that start at the same time will probably have the
457 * same ID here. Again, it's not important to exclude them from
458 * anything; all we need to know is which version of the DB they
459 * started from so we can avoid overwriting any data used in that
460 * particular version.
463 /** The process ID of the process owning this reader txn. */
465 /** The thread ID of the thread owning this txn. */
469 /** The actual reader record, with cacheline padding. */
470 typedef struct MDB_reader {
473 /** shorthand for mrb_txnid */
474 #define mr_txnid mru.mrx.mrb_txnid
475 #define mr_pid mru.mrx.mrb_pid
476 #define mr_tid mru.mrx.mrb_tid
477 /** cache line alignment */
478 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
482 /** The header for the reader table.
483 * The table resides in a memory-mapped file. (This is a different file
484 * than is used for the main database.)
486 * For POSIX the actual mutexes reside in the shared memory of this
487 * mapped file. On Windows, mutexes are named objects allocated by the
488 * kernel; we store the mutex names in this mapped file so that other
489 * processes can grab them. This same approach is also used on
490 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
491 * process-shared POSIX mutexes. For these cases where a named object
492 * is used, the object name is derived from a 64 bit FNV hash of the
493 * environment pathname. As such, naming collisions are extremely
494 * unlikely. If a collision occurs, the results are unpredictable.
496 typedef struct MDB_txbody {
497 /** Stamp identifying this as an MDB file. It must be set
500 /** Version number of this lock file. Must be set to #MDB_VERSION. */
501 uint32_t mtb_version;
502 #if defined(_WIN32) || defined(__APPLE__)
503 char mtb_rmname[MNAME_LEN];
505 /** Mutex protecting access to this table.
506 * This is the reader lock that #LOCK_MUTEX_R acquires.
508 pthread_mutex_t mtb_mutex;
510 /** The ID of the last transaction committed to the database.
511 * This is recorded here only for convenience; the value can always
512 * be determined by reading the main database meta pages.
515 /** The number of slots that have been used in the reader table.
516 * This always records the maximum count, it is not decremented
517 * when readers release their slots.
519 unsigned mtb_numreaders;
520 /** The ID of the most recent meta page in the database.
521 * This is recorded here only for convenience; the value can always
522 * be determined by reading the main database meta pages.
524 uint32_t mtb_me_toggle;
527 /** The actual reader table definition. */
528 typedef struct MDB_txninfo {
531 #define mti_magic mt1.mtb.mtb_magic
532 #define mti_version mt1.mtb.mtb_version
533 #define mti_mutex mt1.mtb.mtb_mutex
534 #define mti_rmname mt1.mtb.mtb_rmname
535 #define mti_txnid mt1.mtb.mtb_txnid
536 #define mti_numreaders mt1.mtb.mtb_numreaders
537 #define mti_me_toggle mt1.mtb.mtb_me_toggle
538 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
541 #if defined(_WIN32) || defined(__APPLE__)
542 char mt2_wmname[MNAME_LEN];
543 #define mti_wmname mt2.mt2_wmname
545 pthread_mutex_t mt2_wmutex;
546 #define mti_wmutex mt2.mt2_wmutex
548 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
550 MDB_reader mti_readers[1];
554 /** Common header for all page types.
555 * Overflow records occupy a number of contiguous pages with no
556 * headers on any page after the first.
558 typedef struct MDB_page {
559 #define mp_pgno mp_p.p_pgno
560 #define mp_next mp_p.p_next
562 pgno_t p_pgno; /**< page number */
563 void * p_next; /**< for in-memory list of freed structs */
566 /** @defgroup mdb_page Page Flags
568 * Flags for the page headers.
571 #define P_BRANCH 0x01 /**< branch page */
572 #define P_LEAF 0x02 /**< leaf page */
573 #define P_OVERFLOW 0x04 /**< overflow page */
574 #define P_META 0x08 /**< meta page */
575 #define P_DIRTY 0x10 /**< dirty page */
576 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
577 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
579 uint16_t mp_flags; /**< @ref mdb_page */
580 #define mp_lower mp_pb.pb.pb_lower
581 #define mp_upper mp_pb.pb.pb_upper
582 #define mp_pages mp_pb.pb_pages
585 indx_t pb_lower; /**< lower bound of free space */
586 indx_t pb_upper; /**< upper bound of free space */
588 uint32_t pb_pages; /**< number of overflow pages */
590 indx_t mp_ptrs[1]; /**< dynamic size */
593 /** Size of the page header, excluding dynamic data at the end */
594 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
596 /** Address of first usable data byte in a page, after the header */
597 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
599 /** Number of nodes on a page */
600 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
602 /** The amount of space remaining in the page */
603 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
605 /** The percentage of space used in the page, in tenths of a percent. */
606 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
607 ((env)->me_psize - PAGEHDRSZ))
608 /** The minimum page fill factor, in tenths of a percent.
609 * Pages emptier than this are candidates for merging.
611 #define FILL_THRESHOLD 250
613 /** Test if a page is a leaf page */
614 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
615 /** Test if a page is a LEAF2 page */
616 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
617 /** Test if a page is a branch page */
618 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
619 /** Test if a page is an overflow page */
620 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
621 /** Test if a page is a sub page */
622 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
624 /** The number of overflow pages needed to store the given size. */
625 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
627 /** Header for a single key/data pair within a page.
628 * We guarantee 2-byte alignment for nodes.
630 typedef struct MDB_node {
631 /** lo and hi are used for data size on leaf nodes and for
632 * child pgno on branch nodes. On 64 bit platforms, flags
633 * is also used for pgno. (Branch nodes have no flags).
634 * They are in host byte order in case that lets some
635 * accesses be optimized into a 32-bit word access.
637 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
638 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
639 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
640 /** @defgroup mdb_node Node Flags
642 * Flags for node headers.
645 #define F_BIGDATA 0x01 /**< data put on overflow page */
646 #define F_SUBDATA 0x02 /**< data is a sub-database */
647 #define F_DUPDATA 0x04 /**< data has duplicates */
649 /** valid flags for #mdb_node_add() */
650 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
653 unsigned short mn_flags; /**< @ref mdb_node */
654 unsigned short mn_ksize; /**< key size */
655 char mn_data[1]; /**< key and data are appended here */
658 /** Size of the node header, excluding dynamic data at the end */
659 #define NODESIZE offsetof(MDB_node, mn_data)
661 /** Bit position of top word in page number, for shifting mn_flags */
662 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
664 /** Size of a node in a branch page with a given key.
665 * This is just the node header plus the key, there is no data.
667 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
669 /** Size of a node in a leaf page with a given key and data.
670 * This is node header plus key plus data size.
672 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
674 /** Address of node \b i in page \b p */
675 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
677 /** Address of the key for the node */
678 #define NODEKEY(node) (void *)((node)->mn_data)
680 /** Address of the data for a node */
681 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
683 /** Get the page number pointed to by a branch node */
684 #define NODEPGNO(node) \
685 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
686 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
687 /** Set the page number in a branch node */
688 #define SETPGNO(node,pgno) do { \
689 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
690 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
692 /** Get the size of the data in a leaf node */
693 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
694 /** Set the size of the data for a leaf node */
695 #define SETDSZ(node,size) do { \
696 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
697 /** The size of a key in a node */
698 #define NODEKSZ(node) ((node)->mn_ksize)
700 /** Copy a page number from src to dst */
702 #define COPY_PGNO(dst,src) dst = src
704 #if SIZE_MAX > 4294967295UL
705 #define COPY_PGNO(dst,src) do { \
706 unsigned short *s, *d; \
707 s = (unsigned short *)&(src); \
708 d = (unsigned short *)&(dst); \
715 #define COPY_PGNO(dst,src) do { \
716 unsigned short *s, *d; \
717 s = (unsigned short *)&(src); \
718 d = (unsigned short *)&(dst); \
724 /** The address of a key in a LEAF2 page.
725 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
726 * There are no node headers, keys are stored contiguously.
728 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
730 /** Set the \b node's key into \b key, if requested. */
731 #define MDB_SET_KEY(node, key) { if ((key) != NULL) { \
732 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
734 /** Information about a single database in the environment. */
735 typedef struct MDB_db {
736 uint32_t md_pad; /**< also ksize for LEAF2 pages */
737 uint16_t md_flags; /**< @ref mdb_open */
738 uint16_t md_depth; /**< depth of this tree */
739 pgno_t md_branch_pages; /**< number of internal pages */
740 pgno_t md_leaf_pages; /**< number of leaf pages */
741 pgno_t md_overflow_pages; /**< number of overflow pages */
742 size_t md_entries; /**< number of data items */
743 pgno_t md_root; /**< the root page of this tree */
746 /** Handle for the DB used to track free pages. */
748 /** Handle for the default DB. */
751 /** Meta page content. */
752 typedef struct MDB_meta {
753 /** Stamp identifying this as an MDB file. It must be set
756 /** Version number of this lock file. Must be set to #MDB_VERSION. */
758 void *mm_address; /**< address for fixed mapping */
759 size_t mm_mapsize; /**< size of mmap region */
760 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
761 /** The size of pages used in this DB */
762 #define mm_psize mm_dbs[0].md_pad
763 /** Any persistent environment flags. @ref mdb_env */
764 #define mm_flags mm_dbs[0].md_flags
765 pgno_t mm_last_pg; /**< last used page in file */
766 txnid_t mm_txnid; /**< txnid that committed this page */
769 /** Buffer for a stack-allocated dirty page.
770 * The members define size and alignment, and silence type
771 * aliasing warnings. They are not used directly; that could
772 * mean incorrectly using several union members in parallel.
774 typedef union MDB_pagebuf {
775 char mb_raw[MDB_PAGESIZE];
778 char mm_pad[PAGEHDRSZ];
783 /** Auxiliary DB info.
784 * The information here is mostly static/read-only. There is
785 * only a single copy of this record in the environment.
787 typedef struct MDB_dbx {
788 MDB_val md_name; /**< name of the database */
789 MDB_cmp_func *md_cmp; /**< function for comparing keys */
790 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
791 MDB_rel_func *md_rel; /**< user relocate function */
792 void *md_relctx; /**< user-provided context for md_rel */
795 /** A database transaction.
796 * Every operation requires a transaction handle.
799 MDB_txn *mt_parent; /**< parent of a nested txn */
800 MDB_txn *mt_child; /**< nested txn under this txn */
801 pgno_t mt_next_pgno; /**< next unallocated page */
802 /** The ID of this transaction. IDs are integers incrementing from 1.
803 * Only committed write transactions increment the ID. If a transaction
804 * aborts, the ID may be re-used by the next writer.
807 MDB_env *mt_env; /**< the DB environment */
808 /** The list of pages that became unused during this transaction.
812 ID2L dirty_list; /**< modified pages */
813 MDB_reader *reader; /**< this thread's slot in the reader table */
815 /** Array of records for each DB known in the environment. */
817 /** Array of MDB_db records for each known DB */
819 /** @defgroup mt_dbflag Transaction DB Flags
823 #define DB_DIRTY 0x01 /**< DB was written in this txn */
824 #define DB_STALE 0x02 /**< DB record is older than txnID */
826 /** Array of cursors for each DB */
827 MDB_cursor **mt_cursors;
828 /** Array of flags for each DB */
829 unsigned char *mt_dbflags;
830 /** Number of DB records in use. This number only ever increments;
831 * we don't decrement it when individual DB handles are closed.
835 /** @defgroup mdb_txn Transaction Flags
839 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
840 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
842 unsigned int mt_flags; /**< @ref mdb_txn */
843 /** Tracks which of the two meta pages was used at the start
844 * of this transaction.
846 unsigned int mt_toggle;
849 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
850 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
851 * raise this on a 64 bit machine.
853 #define CURSOR_STACK 32
857 /** Cursors are used for all DB operations */
859 /** Next cursor on this DB in this txn */
861 /** Original cursor if this is a shadow */
863 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
864 struct MDB_xcursor *mc_xcursor;
865 /** The transaction that owns this cursor */
867 /** The database handle this cursor operates on */
869 /** The database record for this cursor */
871 /** The database auxiliary record for this cursor */
873 /** The @ref mt_dbflag for this database */
874 unsigned char *mc_dbflag;
875 unsigned short mc_snum; /**< number of pushed pages */
876 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
877 /** @defgroup mdb_cursor Cursor Flags
879 * Cursor state flags.
882 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
883 #define C_EOF 0x02 /**< No more data */
884 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
885 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
886 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
888 unsigned int mc_flags; /**< @ref mdb_cursor */
889 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
890 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
893 /** Context for sorted-dup records.
894 * We could have gone to a fully recursive design, with arbitrarily
895 * deep nesting of sub-databases. But for now we only handle these
896 * levels - main DB, optional sub-DB, sorted-duplicate DB.
898 typedef struct MDB_xcursor {
899 /** A sub-cursor for traversing the Dup DB */
900 MDB_cursor mx_cursor;
901 /** The database record for this Dup DB */
903 /** The auxiliary DB record for this Dup DB */
905 /** The @ref mt_dbflag for this Dup DB */
906 unsigned char mx_dbflag;
909 /** A set of pages freed by an earlier transaction. */
910 typedef struct MDB_oldpages {
911 /** Usually we only read one record from the FREEDB at a time, but
912 * in case we read more, this will chain them together.
914 struct MDB_oldpages *mo_next;
915 /** The ID of the transaction in which these pages were freed. */
917 /** An #IDL of the pages */
918 pgno_t mo_pages[1]; /* dynamic */
921 /** The database environment. */
923 HANDLE me_fd; /**< The main data file */
924 HANDLE me_lfd; /**< The lock file */
925 HANDLE me_mfd; /**< just for writing the meta pages */
926 /** Failed to update the meta page. Probably an I/O error. */
927 #define MDB_FATAL_ERROR 0x80000000U
928 uint32_t me_flags; /**< @ref mdb_env */
929 uint32_t me_extrapad; /**< unused for now */
930 unsigned int me_maxreaders; /**< size of the reader table */
931 MDB_dbi me_numdbs; /**< number of DBs opened */
932 MDB_dbi me_maxdbs; /**< size of the DB table */
933 char *me_path; /**< path to the DB files */
934 char *me_map; /**< the memory map of the data file */
935 MDB_txninfo *me_txns; /**< the memory map of the lock file */
936 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
937 MDB_txn *me_txn; /**< current write transaction */
938 size_t me_mapsize; /**< size of the data memory map */
939 off_t me_size; /**< current file size */
940 pgno_t me_maxpg; /**< me_mapsize / me_psize */
941 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
942 unsigned int me_db_toggle; /**< which DB table is current */
943 txnid_t me_wtxnid; /**< ID of last txn we committed */
944 MDB_dbx *me_dbxs; /**< array of static DB info */
945 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
946 MDB_oldpages *me_pghead; /**< list of old page records */
947 pthread_key_t me_txkey; /**< thread-key for readers */
948 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
949 /** IDL of pages that became unused in a write txn */
951 /** ID2L of pages that were written during a write txn */
952 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
953 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
954 LAZY_RWLOCK_DEF(me_dblock)
956 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
960 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
964 /** max number of pages to commit in one writev() call */
965 #define MDB_COMMIT_PAGES 64
966 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
967 #undef MDB_COMMIT_PAGES
968 #define MDB_COMMIT_PAGES IOV_MAX
971 static MDB_page *mdb_page_alloc(MDB_cursor *mc, int num);
972 static MDB_page *mdb_page_new(MDB_cursor *mc, uint32_t flags, int num);
973 static int mdb_page_touch(MDB_cursor *mc);
975 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
976 static int mdb_page_search_root(MDB_cursor *mc,
977 MDB_val *key, int modify);
978 static int mdb_page_search(MDB_cursor *mc,
979 MDB_val *key, int modify);
980 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
981 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
982 pgno_t newpgno, unsigned int nflags);
984 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
985 static int mdb_env_read_meta(MDB_env *env, int *which);
986 static int mdb_env_write_meta(MDB_txn *txn);
988 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
989 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
990 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
991 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
992 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
993 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
994 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
995 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
996 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
998 static int mdb_rebalance(MDB_cursor *mc);
999 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
1001 static void mdb_cursor_pop(MDB_cursor *mc);
1002 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1004 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1005 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1006 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1007 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1008 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1010 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1011 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1013 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1014 static void mdb_xcursor_init0(MDB_cursor *mc);
1015 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1017 static int mdb_drop0(MDB_cursor *mc, int subs);
1018 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1021 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1025 static SECURITY_DESCRIPTOR mdb_null_sd;
1026 static SECURITY_ATTRIBUTES mdb_all_sa;
1027 static int mdb_sec_inited;
1030 /** Return the library version info. */
1032 mdb_version(int *major, int *minor, int *patch)
1034 if (major) *major = MDB_VERSION_MAJOR;
1035 if (minor) *minor = MDB_VERSION_MINOR;
1036 if (patch) *patch = MDB_VERSION_PATCH;
1037 return MDB_VERSION_STRING;
1040 /** Table of descriptions for MDB @ref errors */
1041 static char *const mdb_errstr[] = {
1042 "MDB_KEYEXIST: Key/data pair already exists",
1043 "MDB_NOTFOUND: No matching key/data pair found",
1044 "MDB_PAGE_NOTFOUND: Requested page not found",
1045 "MDB_CORRUPTED: Located page was wrong type",
1046 "MDB_PANIC: Update of meta page failed",
1047 "MDB_VERSION_MISMATCH: Database environment version mismatch"
1051 mdb_strerror(int err)
1054 return ("Successful return: 0");
1056 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
1057 return mdb_errstr[err - MDB_KEYEXIST];
1059 return strerror(err);
1063 /** Display a key in hexadecimal and return the address of the result.
1064 * @param[in] key the key to display
1065 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1066 * @return The key in hexadecimal form.
1069 mdb_dkey(MDB_val *key, char *buf)
1072 unsigned char *c = key->mv_data;
1074 if (key->mv_size > MAXKEYSIZE)
1075 return "MAXKEYSIZE";
1076 /* may want to make this a dynamic check: if the key is mostly
1077 * printable characters, print it as-is instead of converting to hex.
1081 for (i=0; i<key->mv_size; i++)
1082 ptr += sprintf(ptr, "%02x", *c++);
1084 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1089 /** Display all the keys in the page. */
1091 mdb_page_keys(MDB_page *mp)
1094 unsigned int i, nkeys;
1098 nkeys = NUMKEYS(mp);
1099 DPRINTF("numkeys %d", nkeys);
1100 for (i=0; i<nkeys; i++) {
1101 node = NODEPTR(mp, i);
1102 key.mv_size = node->mn_ksize;
1103 key.mv_data = node->mn_data;
1104 DPRINTF("key %d: %s", i, DKEY(&key));
1110 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1112 return txn->mt_dbxs[dbi].md_cmp(a, b);
1116 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1118 if (txn->mt_dbxs[dbi].md_dcmp)
1119 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1121 return EINVAL; /* too bad you can't distinguish this from a valid result */
1124 /** Allocate a single page.
1125 * Re-use old malloc'd pages first, otherwise just malloc.
1128 mdb_page_malloc(MDB_cursor *mc) {
1130 size_t sz = mc->mc_txn->mt_env->me_psize;
1131 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1132 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1133 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1134 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1135 } else if ((ret = malloc(sz)) != NULL) {
1136 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1141 /** Allocate pages for writing.
1142 * If there are free pages available from older transactions, they
1143 * will be re-used first. Otherwise a new page will be allocated.
1144 * @param[in] mc cursor A cursor handle identifying the transaction and
1145 * database for which we are allocating.
1146 * @param[in] num the number of pages to allocate.
1147 * @return Address of the allocated page(s). Requests for multiple pages
1148 * will always be satisfied by a single contiguous chunk of memory.
1151 mdb_page_alloc(MDB_cursor *mc, int num)
1153 MDB_txn *txn = mc->mc_txn;
1155 pgno_t pgno = P_INVALID;
1158 if (txn->mt_txnid > 2) {
1160 if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
1161 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1162 /* See if there's anything in the free DB */
1165 txnid_t *kptr, oldest;
1167 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1168 mdb_page_search(&m2, NULL, 0);
1169 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1170 kptr = (txnid_t *)NODEKEY(leaf);
1174 oldest = txn->mt_txnid - 1;
1175 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1176 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1177 if (mr && mr < oldest)
1182 if (oldest > *kptr) {
1183 /* It's usable, grab it.
1189 mdb_node_read(txn, leaf, &data);
1190 idl = (ID *) data.mv_data;
1191 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1192 mop->mo_next = txn->mt_env->me_pghead;
1193 mop->mo_txnid = *kptr;
1194 txn->mt_env->me_pghead = mop;
1195 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1200 DPRINTF("IDL read txn %zu root %zu num %zu",
1201 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1202 for (i=0; i<idl[0]; i++) {
1203 DPRINTF("IDL %zu", idl[i+1]);
1207 /* drop this IDL from the DB */
1208 m2.mc_ki[m2.mc_top] = 0;
1209 m2.mc_flags = C_INITIALIZED;
1210 mdb_cursor_del(&m2, 0);
1213 if (txn->mt_env->me_pghead) {
1214 MDB_oldpages *mop = txn->mt_env->me_pghead;
1216 /* FIXME: For now, always use fresh pages. We
1217 * really ought to search the free list for a
1222 /* peel pages off tail, so we only have to truncate the list */
1223 pgno = MDB_IDL_LAST(mop->mo_pages);
1224 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1226 if (mop->mo_pages[2] > mop->mo_pages[1])
1227 mop->mo_pages[0] = 0;
1231 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1232 txn->mt_env->me_pghead = mop->mo_next;
1239 if (pgno == P_INVALID) {
1240 /* DB size is maxed out */
1241 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1242 DPUTS("DB size maxed out");
1246 if (txn->mt_env->me_dpages && num == 1) {
1247 np = txn->mt_env->me_dpages;
1248 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1249 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1250 txn->mt_env->me_dpages = np->mp_next;
1252 size_t sz = txn->mt_env->me_psize * num;
1253 if ((np = malloc(sz)) == NULL)
1255 VGMEMP_ALLOC(txn->mt_env, np, sz);
1257 if (pgno == P_INVALID) {
1258 np->mp_pgno = txn->mt_next_pgno;
1259 txn->mt_next_pgno += num;
1263 mid.mid = np->mp_pgno;
1265 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1270 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1271 * @param[in] mc cursor pointing to the page to be touched
1272 * @return 0 on success, non-zero on failure.
1275 mdb_page_touch(MDB_cursor *mc)
1277 MDB_page *mp = mc->mc_pg[mc->mc_top];
1280 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1282 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1284 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1285 assert(mp->mp_pgno != np->mp_pgno);
1286 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1288 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1291 mp->mp_flags |= P_DIRTY;
1294 /* Adjust other cursors pointing to mp */
1295 if (mc->mc_flags & C_SUB) {
1296 MDB_cursor *m2, *m3;
1297 MDB_dbi dbi = mc->mc_dbi-1;
1299 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1300 if (m2 == mc) continue;
1301 m3 = &m2->mc_xcursor->mx_cursor;
1302 if (m3->mc_snum < mc->mc_snum) continue;
1303 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1304 m3->mc_pg[mc->mc_top] = mp;
1310 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1311 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1312 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1313 m2->mc_pg[mc->mc_top] = mp;
1317 mc->mc_pg[mc->mc_top] = mp;
1318 /** If this page has a parent, update the parent to point to
1322 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1324 mc->mc_db->md_root = mp->mp_pgno;
1325 } else if (mc->mc_txn->mt_parent) {
1328 /* If txn has a parent, make sure the page is in our
1331 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1332 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1333 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1334 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1335 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1336 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1337 mc->mc_pg[mc->mc_top] = mp;
1343 np = mdb_page_malloc(mc);
1344 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1345 mid.mid = np->mp_pgno;
1347 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1355 mdb_env_sync(MDB_env *env, int force)
1358 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1359 if (MDB_FDATASYNC(env->me_fd))
1365 /** Make shadow copies of all of parent txn's cursors */
1367 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1369 MDB_cursor *mc, *m2;
1370 unsigned int i, j, size;
1372 for (i=0;i<src->mt_numdbs; i++) {
1373 if (src->mt_cursors[i]) {
1374 size = sizeof(MDB_cursor);
1375 if (src->mt_cursors[i]->mc_xcursor)
1376 size += sizeof(MDB_xcursor);
1377 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1384 mc->mc_db = &dst->mt_dbs[i];
1385 mc->mc_dbx = m2->mc_dbx;
1386 mc->mc_dbflag = &dst->mt_dbflags[i];
1387 mc->mc_snum = m2->mc_snum;
1388 mc->mc_top = m2->mc_top;
1389 mc->mc_flags = m2->mc_flags | C_SHADOW;
1390 for (j=0; j<mc->mc_snum; j++) {
1391 mc->mc_pg[j] = m2->mc_pg[j];
1392 mc->mc_ki[j] = m2->mc_ki[j];
1394 if (m2->mc_xcursor) {
1395 MDB_xcursor *mx, *mx2;
1396 mx = (MDB_xcursor *)(mc+1);
1397 mc->mc_xcursor = mx;
1398 mx2 = m2->mc_xcursor;
1399 mx->mx_db = mx2->mx_db;
1400 mx->mx_dbx = mx2->mx_dbx;
1401 mx->mx_dbflag = mx2->mx_dbflag;
1402 mx->mx_cursor.mc_txn = dst;
1403 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1404 mx->mx_cursor.mc_db = &mx->mx_db;
1405 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1406 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1407 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1408 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1409 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1410 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1411 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1412 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1415 mc->mc_xcursor = NULL;
1417 mc->mc_next = dst->mt_cursors[i];
1418 dst->mt_cursors[i] = mc;
1425 /** Merge shadow cursors back into parent's */
1427 mdb_cursor_merge(MDB_txn *txn)
1430 for (i=0; i<txn->mt_numdbs; i++) {
1431 if (txn->mt_cursors[i]) {
1433 while ((mc = txn->mt_cursors[i])) {
1434 txn->mt_cursors[i] = mc->mc_next;
1435 if (mc->mc_flags & C_SHADOW) {
1436 MDB_cursor *m2 = mc->mc_orig;
1438 m2->mc_snum = mc->mc_snum;
1439 m2->mc_top = mc->mc_top;
1440 for (j=0; j<mc->mc_snum; j++) {
1441 m2->mc_pg[j] = mc->mc_pg[j];
1442 m2->mc_ki[j] = mc->mc_ki[j];
1445 if (mc->mc_flags & C_ALLOCD)
1453 mdb_txn_reset0(MDB_txn *txn);
1455 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1456 * @param[in] txn the transaction handle to initialize
1457 * @return 0 on success, non-zero on failure. This can only
1458 * fail for read-only transactions, and then only if the
1459 * reader table is full.
1462 mdb_txn_renew0(MDB_txn *txn)
1464 MDB_env *env = txn->mt_env;
1467 if (txn->mt_flags & MDB_TXN_RDONLY) {
1468 MDB_reader *r = pthread_getspecific(env->me_txkey);
1471 pid_t pid = getpid();
1472 pthread_t tid = pthread_self();
1475 for (i=0; i<env->me_txns->mti_numreaders; i++)
1476 if (env->me_txns->mti_readers[i].mr_pid == 0)
1478 if (i == env->me_maxreaders) {
1479 UNLOCK_MUTEX_R(env);
1482 env->me_txns->mti_readers[i].mr_pid = pid;
1483 env->me_txns->mti_readers[i].mr_tid = tid;
1484 if (i >= env->me_txns->mti_numreaders)
1485 env->me_txns->mti_numreaders = i+1;
1486 UNLOCK_MUTEX_R(env);
1487 r = &env->me_txns->mti_readers[i];
1488 pthread_setspecific(env->me_txkey, r);
1490 txn->mt_toggle = env->me_txns->mti_me_toggle;
1491 txn->mt_txnid = env->me_txns->mti_txnid;
1492 /* This happens if a different process was the
1493 * last writer to the DB.
1495 if (env->me_wtxnid < txn->mt_txnid)
1496 mt_dbflag = DB_STALE;
1497 r->mr_txnid = txn->mt_txnid;
1498 txn->mt_u.reader = r;
1502 txn->mt_txnid = env->me_txns->mti_txnid;
1503 if (env->me_wtxnid < txn->mt_txnid)
1504 mt_dbflag = DB_STALE;
1506 txn->mt_toggle = env->me_txns->mti_me_toggle;
1507 txn->mt_u.dirty_list = env->me_dirty_list;
1508 txn->mt_u.dirty_list[0].mid = 0;
1509 txn->mt_free_pgs = env->me_free_pgs;
1510 txn->mt_free_pgs[0] = 0;
1511 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1515 /* Copy the DB arrays */
1516 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1517 txn->mt_numdbs = env->me_numdbs;
1518 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1519 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1520 if (txn->mt_numdbs > 2)
1521 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1522 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1523 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1525 memset(txn->mt_dbflags, mt_dbflag, env->me_numdbs);
1531 mdb_txn_renew(MDB_txn *txn)
1538 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1539 DPUTS("environment had fatal error, must shutdown!");
1543 rc = mdb_txn_renew0(txn);
1544 if (rc == MDB_SUCCESS) {
1545 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1546 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1547 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1553 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1558 if (env->me_flags & MDB_FATAL_ERROR) {
1559 DPUTS("environment had fatal error, must shutdown!");
1563 /* parent already has an active child txn */
1564 if (parent->mt_child) {
1568 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1569 if (!(flags & MDB_RDONLY))
1570 size += env->me_maxdbs * sizeof(MDB_cursor *);
1572 if ((txn = calloc(1, size)) == NULL) {
1573 DPRINTF("calloc: %s", strerror(ErrCode()));
1576 txn->mt_dbs = (MDB_db *)(txn+1);
1577 if (flags & MDB_RDONLY) {
1578 txn->mt_flags |= MDB_TXN_RDONLY;
1579 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1581 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1582 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1587 txn->mt_free_pgs = mdb_midl_alloc();
1588 if (!txn->mt_free_pgs) {
1592 txn->mt_u.dirty_list = malloc(sizeof(ID2)*MDB_IDL_UM_SIZE);
1593 if (!txn->mt_u.dirty_list) {
1594 free(txn->mt_free_pgs);
1598 txn->mt_txnid = parent->mt_txnid;
1599 txn->mt_toggle = parent->mt_toggle;
1600 txn->mt_u.dirty_list[0].mid = 0;
1601 txn->mt_free_pgs[0] = 0;
1602 txn->mt_next_pgno = parent->mt_next_pgno;
1603 parent->mt_child = txn;
1604 txn->mt_parent = parent;
1605 txn->mt_numdbs = parent->mt_numdbs;
1606 txn->mt_dbxs = parent->mt_dbxs;
1607 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1608 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1609 mdb_cursor_shadow(parent, txn);
1612 rc = mdb_txn_renew0(txn);
1618 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1619 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1620 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1626 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1627 * @param[in] txn the transaction handle to reset
1630 mdb_txn_reset0(MDB_txn *txn)
1632 MDB_env *env = txn->mt_env;
1634 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1635 txn->mt_u.reader->mr_txnid = 0;
1641 /* close(free) all cursors */
1642 for (i=0; i<txn->mt_numdbs; i++) {
1643 if (txn->mt_cursors[i]) {
1645 while ((mc = txn->mt_cursors[i])) {
1646 txn->mt_cursors[i] = mc->mc_next;
1647 if (mc->mc_flags & C_ALLOCD)
1653 /* return all dirty pages to dpage list */
1654 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1655 dp = txn->mt_u.dirty_list[i].mptr;
1656 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1657 dp->mp_next = txn->mt_env->me_dpages;
1658 VGMEMP_FREE(txn->mt_env, dp);
1659 txn->mt_env->me_dpages = dp;
1661 /* large pages just get freed directly */
1662 VGMEMP_FREE(txn->mt_env, dp);
1667 if (txn->mt_parent) {
1668 txn->mt_parent->mt_child = NULL;
1669 free(txn->mt_free_pgs);
1670 free(txn->mt_u.dirty_list);
1673 if (mdb_midl_shrink(&txn->mt_free_pgs))
1674 env->me_free_pgs = txn->mt_free_pgs;
1677 while ((mop = txn->mt_env->me_pghead)) {
1678 txn->mt_env->me_pghead = mop->mo_next;
1683 /* The writer mutex was locked in mdb_txn_begin. */
1684 UNLOCK_MUTEX_W(env);
1689 mdb_txn_reset(MDB_txn *txn)
1694 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1695 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1696 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1698 mdb_txn_reset0(txn);
1702 mdb_txn_abort(MDB_txn *txn)
1707 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1708 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1709 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1712 mdb_txn_abort(txn->mt_child);
1714 mdb_txn_reset0(txn);
1719 mdb_txn_commit(MDB_txn *txn)
1730 assert(txn != NULL);
1731 assert(txn->mt_env != NULL);
1733 if (txn->mt_child) {
1734 mdb_txn_commit(txn->mt_child);
1735 txn->mt_child = NULL;
1740 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1741 if (txn->mt_numdbs > env->me_numdbs) {
1742 /* update the DB tables */
1743 int toggle = !env->me_db_toggle;
1747 ip = &env->me_dbs[toggle][env->me_numdbs];
1748 jp = &txn->mt_dbs[env->me_numdbs];
1749 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1750 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
1754 env->me_db_toggle = toggle;
1755 env->me_numdbs = txn->mt_numdbs;
1756 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1762 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1763 DPUTS("error flag is set, can't commit");
1765 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1770 /* Merge (and close) our cursors with parent's */
1771 mdb_cursor_merge(txn);
1773 if (txn->mt_parent) {
1779 /* Update parent's DB table */
1780 ip = &txn->mt_parent->mt_dbs[2];
1781 jp = &txn->mt_dbs[2];
1782 for (i = 2; i < txn->mt_numdbs; i++) {
1783 if (ip->md_root != jp->md_root)
1787 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1789 /* Append our free list to parent's */
1790 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1792 mdb_midl_free(txn->mt_free_pgs);
1794 /* Merge our dirty list with parent's */
1795 dst = txn->mt_parent->mt_u.dirty_list;
1796 src = txn->mt_u.dirty_list;
1797 x = mdb_mid2l_search(dst, src[1].mid);
1798 for (y=1; y<=src[0].mid; y++) {
1799 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1803 dst[x].mptr = src[y].mptr;
1806 for (; y<=src[0].mid; y++) {
1807 if (++x >= MDB_IDL_UM_MAX) {
1814 free(txn->mt_u.dirty_list);
1815 txn->mt_parent->mt_child = NULL;
1820 if (txn != env->me_txn) {
1821 DPUTS("attempt to commit unknown transaction");
1826 if (!txn->mt_u.dirty_list[0].mid)
1829 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1830 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1832 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1834 /* should only be one record now */
1835 if (env->me_pghead) {
1836 /* make sure first page of freeDB is touched and on freelist */
1837 mdb_page_search(&mc, NULL, 1);
1839 /* save to free list */
1840 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1844 /* make sure last page of freeDB is touched and on freelist */
1845 key.mv_size = MAXKEYSIZE+1;
1847 mdb_page_search(&mc, &key, 1);
1849 mdb_midl_sort(txn->mt_free_pgs);
1853 ID *idl = txn->mt_free_pgs;
1854 DPRINTF("IDL write txn %zu root %zu num %zu",
1855 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1856 for (i=0; i<idl[0]; i++) {
1857 DPRINTF("IDL %zu", idl[i+1]);
1861 /* write to last page of freeDB */
1862 key.mv_size = sizeof(pgno_t);
1863 key.mv_data = &txn->mt_txnid;
1864 data.mv_data = txn->mt_free_pgs;
1865 /* The free list can still grow during this call,
1866 * despite the pre-emptive touches above. So check
1867 * and make sure the entire thing got written.
1870 i = txn->mt_free_pgs[0];
1871 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1872 rc = mdb_cursor_put(&mc, &key, &data, 0);
1877 } while (i != txn->mt_free_pgs[0]);
1878 if (mdb_midl_shrink(&txn->mt_free_pgs))
1879 env->me_free_pgs = txn->mt_free_pgs;
1881 /* should only be one record now */
1882 if (env->me_pghead) {
1886 mop = env->me_pghead;
1887 env->me_pghead = NULL;
1888 key.mv_size = sizeof(pgno_t);
1889 key.mv_data = &mop->mo_txnid;
1890 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1891 data.mv_data = mop->mo_pages;
1892 mdb_cursor_put(&mc, &key, &data, 0);
1896 /* Update DB root pointers. Their pages have already been
1897 * touched so this is all in-place and cannot fail.
1902 data.mv_size = sizeof(MDB_db);
1904 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
1905 for (i = 2; i < txn->mt_numdbs; i++) {
1906 if (txn->mt_dbflags[i] & DB_DIRTY) {
1907 data.mv_data = &txn->mt_dbs[i];
1908 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1913 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1919 /* Windows actually supports scatter/gather I/O, but only on
1920 * unbuffered file handles. Since we're relying on the OS page
1921 * cache for all our data, that's self-defeating. So we just
1922 * write pages one at a time. We use the ov structure to set
1923 * the write offset, to at least save the overhead of a Seek
1927 memset(&ov, 0, sizeof(ov));
1928 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1930 dp = txn->mt_u.dirty_list[i].mptr;
1931 DPRINTF("committing page %zu", dp->mp_pgno);
1932 size = dp->mp_pgno * env->me_psize;
1933 ov.Offset = size & 0xffffffff;
1934 ov.OffsetHigh = size >> 16;
1935 ov.OffsetHigh >>= 16;
1936 /* clear dirty flag */
1937 dp->mp_flags &= ~P_DIRTY;
1938 wsize = env->me_psize;
1939 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
1940 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
1943 DPRINTF("WriteFile: %d", n);
1950 struct iovec iov[MDB_COMMIT_PAGES];
1954 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1955 dp = txn->mt_u.dirty_list[i].mptr;
1956 if (dp->mp_pgno != next) {
1958 DPRINTF("committing %u dirty pages", n);
1959 rc = writev(env->me_fd, iov, n);
1963 DPUTS("short write, filesystem full?");
1965 DPRINTF("writev: %s", strerror(n));
1972 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
1975 DPRINTF("committing page %zu", dp->mp_pgno);
1976 iov[n].iov_len = env->me_psize;
1977 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
1978 iov[n].iov_base = (char *)dp;
1979 size += iov[n].iov_len;
1980 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
1981 /* clear dirty flag */
1982 dp->mp_flags &= ~P_DIRTY;
1983 if (++n >= MDB_COMMIT_PAGES) {
1993 DPRINTF("committing %u dirty pages", n);
1994 rc = writev(env->me_fd, iov, n);
1998 DPUTS("short write, filesystem full?");
2000 DPRINTF("writev: %s", strerror(n));
2007 /* Drop the dirty pages.
2009 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2010 dp = txn->mt_u.dirty_list[i].mptr;
2011 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2012 dp->mp_next = txn->mt_env->me_dpages;
2013 VGMEMP_FREE(txn->mt_env, dp);
2014 txn->mt_env->me_dpages = dp;
2016 VGMEMP_FREE(txn->mt_env, dp);
2019 txn->mt_u.dirty_list[i].mid = 0;
2021 txn->mt_u.dirty_list[0].mid = 0;
2023 if ((n = mdb_env_sync(env, 0)) != 0 ||
2024 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2028 env->me_wtxnid = txn->mt_txnid;
2032 /* update the DB tables */
2034 int toggle = !env->me_db_toggle;
2038 ip = &env->me_dbs[toggle][2];
2039 jp = &txn->mt_dbs[2];
2040 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
2041 for (i = 2; i < txn->mt_numdbs; i++) {
2042 if (ip->md_root != jp->md_root)
2047 env->me_db_toggle = toggle;
2048 env->me_numdbs = txn->mt_numdbs;
2049 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
2052 UNLOCK_MUTEX_W(env);
2058 /** Read the environment parameters of a DB environment before
2059 * mapping it into memory.
2060 * @param[in] env the environment handle
2061 * @param[out] meta address of where to store the meta information
2062 * @return 0 on success, non-zero on failure.
2065 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2072 /* We don't know the page size yet, so use a minimum value.
2076 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2078 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2083 else if (rc != MDB_PAGESIZE) {
2087 DPRINTF("read: %s", strerror(err));
2091 p = (MDB_page *)&pbuf;
2093 if (!F_ISSET(p->mp_flags, P_META)) {
2094 DPRINTF("page %zu not a meta page", p->mp_pgno);
2099 if (m->mm_magic != MDB_MAGIC) {
2100 DPUTS("meta has invalid magic");
2104 if (m->mm_version != MDB_VERSION) {
2105 DPRINTF("database is version %u, expected version %u",
2106 m->mm_version, MDB_VERSION);
2107 return MDB_VERSION_MISMATCH;
2110 memcpy(meta, m, sizeof(*m));
2114 /** Write the environment parameters of a freshly created DB environment.
2115 * @param[in] env the environment handle
2116 * @param[out] meta address of where to store the meta information
2117 * @return 0 on success, non-zero on failure.
2120 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2127 DPUTS("writing new meta page");
2129 GET_PAGESIZE(psize);
2131 meta->mm_magic = MDB_MAGIC;
2132 meta->mm_version = MDB_VERSION;
2133 meta->mm_psize = psize;
2134 meta->mm_last_pg = 1;
2135 meta->mm_flags = env->me_flags & 0xffff;
2136 meta->mm_flags |= MDB_INTEGERKEY;
2137 meta->mm_dbs[0].md_root = P_INVALID;
2138 meta->mm_dbs[1].md_root = P_INVALID;
2140 p = calloc(2, psize);
2142 p->mp_flags = P_META;
2145 memcpy(m, meta, sizeof(*meta));
2147 q = (MDB_page *)((char *)p + psize);
2150 q->mp_flags = P_META;
2153 memcpy(m, meta, sizeof(*meta));
2158 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2159 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2162 rc = write(env->me_fd, p, psize * 2);
2163 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2169 /** Update the environment info to commit a transaction.
2170 * @param[in] txn the transaction that's being committed
2171 * @return 0 on success, non-zero on failure.
2174 mdb_env_write_meta(MDB_txn *txn)
2177 MDB_meta meta, metab;
2179 int rc, len, toggle;
2185 assert(txn != NULL);
2186 assert(txn->mt_env != NULL);
2188 toggle = !txn->mt_toggle;
2189 DPRINTF("writing meta page %d for root page %zu",
2190 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2194 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2195 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2197 ptr = (char *)&meta;
2198 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2199 len = sizeof(MDB_meta) - off;
2202 meta.mm_dbs[0] = txn->mt_dbs[0];
2203 meta.mm_dbs[1] = txn->mt_dbs[1];
2204 meta.mm_last_pg = txn->mt_next_pgno - 1;
2205 meta.mm_txnid = txn->mt_txnid;
2208 off += env->me_psize;
2211 /* Write to the SYNC fd */
2214 memset(&ov, 0, sizeof(ov));
2216 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2219 rc = pwrite(env->me_mfd, ptr, len, off);
2224 DPUTS("write failed, disk error?");
2225 /* On a failure, the pagecache still contains the new data.
2226 * Write some old data back, to prevent it from being used.
2227 * Use the non-SYNC fd; we know it will fail anyway.
2229 meta.mm_last_pg = metab.mm_last_pg;
2230 meta.mm_txnid = metab.mm_txnid;
2232 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2234 r2 = pwrite(env->me_fd, ptr, len, off);
2236 env->me_flags |= MDB_FATAL_ERROR;
2239 /* Memory ordering issues are irrelevant; since the entire writer
2240 * is wrapped by wmutex, all of these changes will become visible
2241 * after the wmutex is unlocked. Since the DB is multi-version,
2242 * readers will get consistent data regardless of how fresh or
2243 * how stale their view of these values is.
2245 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
2246 txn->mt_env->me_txns->mti_me_toggle = toggle;
2247 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2248 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
2253 /** Check both meta pages to see which one is newer.
2254 * @param[in] env the environment handle
2255 * @param[out] which address of where to store the meta toggle ID
2256 * @return 0 on success, non-zero on failure.
2259 mdb_env_read_meta(MDB_env *env, int *which)
2263 assert(env != NULL);
2265 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2268 DPRINTF("Using meta page %d", toggle);
2275 mdb_env_create(MDB_env **env)
2279 e = calloc(1, sizeof(MDB_env));
2283 e->me_free_pgs = mdb_midl_alloc();
2284 if (!e->me_free_pgs) {
2288 e->me_maxreaders = DEFAULT_READERS;
2290 e->me_fd = INVALID_HANDLE_VALUE;
2291 e->me_lfd = INVALID_HANDLE_VALUE;
2292 e->me_mfd = INVALID_HANDLE_VALUE;
2293 VGMEMP_CREATE(e,0,0);
2299 mdb_env_set_mapsize(MDB_env *env, size_t size)
2303 env->me_mapsize = size;
2305 env->me_maxpg = env->me_mapsize / env->me_psize;
2310 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2314 env->me_maxdbs = dbs;
2319 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2321 if (env->me_map || readers < 1)
2323 env->me_maxreaders = readers;
2328 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2330 if (!env || !readers)
2332 *readers = env->me_maxreaders;
2336 /** Further setup required for opening an MDB environment
2339 mdb_env_open2(MDB_env *env, unsigned int flags)
2341 int i, newenv = 0, toggle;
2345 env->me_flags = flags;
2347 memset(&meta, 0, sizeof(meta));
2349 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2352 DPUTS("new mdbenv");
2356 if (!env->me_mapsize) {
2357 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2363 LONG sizelo, sizehi;
2364 sizelo = env->me_mapsize & 0xffffffff;
2365 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2367 /* Windows won't create mappings for zero length files.
2368 * Just allocate the maxsize right now.
2371 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2372 if (!SetEndOfFile(env->me_fd))
2374 SetFilePointer(env->me_fd, 0, NULL, 0);
2376 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2377 sizehi, sizelo, NULL);
2380 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2388 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2390 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2392 if (env->me_map == MAP_FAILED) {
2399 meta.mm_mapsize = env->me_mapsize;
2400 if (flags & MDB_FIXEDMAP)
2401 meta.mm_address = env->me_map;
2402 i = mdb_env_init_meta(env, &meta);
2403 if (i != MDB_SUCCESS) {
2404 munmap(env->me_map, env->me_mapsize);
2408 env->me_psize = meta.mm_psize;
2410 env->me_maxpg = env->me_mapsize / env->me_psize;
2412 p = (MDB_page *)env->me_map;
2413 env->me_metas[0] = METADATA(p);
2414 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2416 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
2419 DPRINTF("opened database version %u, pagesize %u",
2420 env->me_metas[toggle]->mm_version, env->me_psize);
2421 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
2422 DPRINTF("entries: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
2423 DPRINTF("branch pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
2424 DPRINTF("leaf pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
2425 DPRINTF("overflow pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
2426 DPRINTF("root: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
2432 /** Release a reader thread's slot in the reader lock table.
2433 * This function is called automatically when a thread exits.
2434 * Windows doesn't support destructor callbacks for thread-specific storage,
2435 * so this function is not compiled there.
2436 * @param[in] ptr This points to the slot in the reader lock table.
2439 mdb_env_reader_dest(void *ptr)
2441 MDB_reader *reader = ptr;
2443 reader->mr_txnid = 0;
2449 /** Downgrade the exclusive lock on the region back to shared */
2451 mdb_env_share_locks(MDB_env *env)
2455 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2457 env->me_txns->mti_me_toggle = toggle;
2458 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2463 /* First acquire a shared lock. The Unlock will
2464 * then release the existing exclusive lock.
2466 memset(&ov, 0, sizeof(ov));
2467 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2468 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2472 struct flock lock_info;
2473 /* The shared lock replaces the existing lock */
2474 memset((void *)&lock_info, 0, sizeof(lock_info));
2475 lock_info.l_type = F_RDLCK;
2476 lock_info.l_whence = SEEK_SET;
2477 lock_info.l_start = 0;
2478 lock_info.l_len = 1;
2479 fcntl(env->me_lfd, F_SETLK, &lock_info);
2483 #if defined(_WIN32) || defined(__APPLE__)
2485 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2487 * @(#) $Revision: 5.1 $
2488 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2489 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2491 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2495 * Please do not copyright this code. This code is in the public domain.
2497 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2498 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2499 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2500 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2501 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2502 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2503 * PERFORMANCE OF THIS SOFTWARE.
2506 * chongo <Landon Curt Noll> /\oo/\
2507 * http://www.isthe.com/chongo/
2509 * Share and Enjoy! :-)
2512 typedef unsigned long long mdb_hash_t;
2513 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2515 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2516 * @param[in] str string to hash
2517 * @param[in] hval initial value for hash
2518 * @return 64 bit hash
2520 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2521 * hval arg on the first call.
2524 mdb_hash_str(char *str, mdb_hash_t hval)
2526 unsigned char *s = (unsigned char *)str; /* unsigned string */
2528 * FNV-1a hash each octet of the string
2531 /* xor the bottom with the current octet */
2532 hval ^= (mdb_hash_t)*s++;
2534 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2535 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2536 (hval << 7) + (hval << 8) + (hval << 40);
2538 /* return our new hash value */
2542 /** Hash the string and output the hash in hex.
2543 * @param[in] str string to hash
2544 * @param[out] hexbuf an array of 17 chars to hold the hash
2547 mdb_hash_hex(char *str, char *hexbuf)
2550 mdb_hash_t h = mdb_hash_str(str, MDB_HASH_INIT);
2551 for (i=0; i<8; i++) {
2552 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2558 /** Open and/or initialize the lock region for the environment.
2559 * @param[in] env The MDB environment.
2560 * @param[in] lpath The pathname of the file used for the lock region.
2561 * @param[in] mode The Unix permissions for the file, if we create it.
2562 * @param[out] excl Set to true if we got an exclusive lock on the region.
2563 * @return 0 on success, non-zero on failure.
2566 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2574 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2575 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2576 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2580 /* Try to get exclusive lock. If we succeed, then
2581 * nobody is using the lock region and we should initialize it.
2584 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2588 memset(&ov, 0, sizeof(ov));
2589 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2595 size = GetFileSize(env->me_lfd, NULL);
2597 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2601 /* Try to get exclusive lock. If we succeed, then
2602 * nobody is using the lock region and we should initialize it.
2605 struct flock lock_info;
2606 memset((void *)&lock_info, 0, sizeof(lock_info));
2607 lock_info.l_type = F_WRLCK;
2608 lock_info.l_whence = SEEK_SET;
2609 lock_info.l_start = 0;
2610 lock_info.l_len = 1;
2611 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2615 lock_info.l_type = F_RDLCK;
2616 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2623 size = lseek(env->me_lfd, 0, SEEK_END);
2625 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2626 if (size < rsize && *excl) {
2628 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2629 if (!SetEndOfFile(env->me_lfd)) {
2634 if (ftruncate(env->me_lfd, rsize) != 0) {
2641 size = rsize - sizeof(MDB_txninfo);
2642 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2647 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2653 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2655 if (!env->me_txns) {
2660 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2662 if (m == MAP_FAILED) {
2663 env->me_txns = NULL;
2673 if (!mdb_sec_inited) {
2674 InitializeSecurityDescriptor(&mdb_null_sd,
2675 SECURITY_DESCRIPTOR_REVISION);
2676 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2677 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2678 mdb_all_sa.bInheritHandle = FALSE;
2679 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2682 mdb_hash_hex(lpath, hexbuf);
2683 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2684 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2685 if (!env->me_rmutex) {
2689 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2690 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2691 if (!env->me_wmutex) {
2698 mdb_hash_hex(lpath, hexbuf);
2699 sprintf(env->me_txns->mti_rmname, "MDBr%s", hexbuf);
2700 if (sem_unlink(env->me_txns->mti_rmname)) {
2702 if (rc != ENOENT && rc != EINVAL)
2705 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2706 if (!env->me_rmutex) {
2710 sprintf(env->me_txns->mti_wmname, "MDBw%s", hexbuf);
2711 if (sem_unlink(env->me_txns->mti_wmname)) {
2713 if (rc != ENOENT && rc != EINVAL)
2716 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2717 if (!env->me_wmutex) {
2721 #else /* __APPLE__ */
2722 pthread_mutexattr_t mattr;
2724 pthread_mutexattr_init(&mattr);
2725 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2729 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2730 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2731 #endif /* __APPLE__ */
2733 env->me_txns->mti_version = MDB_VERSION;
2734 env->me_txns->mti_magic = MDB_MAGIC;
2735 env->me_txns->mti_txnid = 0;
2736 env->me_txns->mti_numreaders = 0;
2737 env->me_txns->mti_me_toggle = 0;
2740 if (env->me_txns->mti_magic != MDB_MAGIC) {
2741 DPUTS("lock region has invalid magic");
2745 if (env->me_txns->mti_version != MDB_VERSION) {
2746 DPRINTF("lock region is version %u, expected version %u",
2747 env->me_txns->mti_version, MDB_VERSION);
2748 rc = MDB_VERSION_MISMATCH;
2752 if (rc != EACCES && rc != EAGAIN) {
2756 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2757 if (!env->me_rmutex) {
2761 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2762 if (!env->me_wmutex) {
2768 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2769 if (!env->me_rmutex) {
2773 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2774 if (!env->me_wmutex) {
2784 env->me_lfd = INVALID_HANDLE_VALUE;
2789 /** The name of the lock file in the DB environment */
2790 #define LOCKNAME "/lock.mdb"
2791 /** The name of the data file in the DB environment */
2792 #define DATANAME "/data.mdb"
2793 /** The suffix of the lock file when no subdir is used */
2794 #define LOCKSUFF "-lock"
2797 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2799 int oflags, rc, len, excl;
2800 char *lpath, *dpath;
2803 if (flags & MDB_NOSUBDIR) {
2804 rc = len + sizeof(LOCKSUFF) + len + 1;
2806 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2811 if (flags & MDB_NOSUBDIR) {
2812 dpath = lpath + len + sizeof(LOCKSUFF);
2813 sprintf(lpath, "%s" LOCKSUFF, path);
2814 strcpy(dpath, path);
2816 dpath = lpath + len + sizeof(LOCKNAME);
2817 sprintf(lpath, "%s" LOCKNAME, path);
2818 sprintf(dpath, "%s" DATANAME, path);
2821 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2826 if (F_ISSET(flags, MDB_RDONLY)) {
2827 oflags = GENERIC_READ;
2828 len = OPEN_EXISTING;
2830 oflags = GENERIC_READ|GENERIC_WRITE;
2833 mode = FILE_ATTRIBUTE_NORMAL;
2834 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2835 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2840 if (F_ISSET(flags, MDB_RDONLY))
2843 oflags = O_RDWR | O_CREAT;
2845 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2851 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2852 /* synchronous fd for meta writes */
2854 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2855 mode |= FILE_FLAG_WRITE_THROUGH;
2856 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2857 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2862 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2863 oflags |= MDB_DSYNC;
2864 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2869 env->me_path = strdup(path);
2870 DPRINTF("opened dbenv %p", (void *) env);
2871 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2872 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2874 mdb_env_share_locks(env);
2875 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
2876 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
2877 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
2883 if (env->me_fd != INVALID_HANDLE_VALUE) {
2885 env->me_fd = INVALID_HANDLE_VALUE;
2887 if (env->me_lfd != INVALID_HANDLE_VALUE) {
2889 env->me_lfd = INVALID_HANDLE_VALUE;
2897 mdb_env_close(MDB_env *env)
2904 VGMEMP_DESTROY(env);
2905 while (env->me_dpages) {
2906 dp = env->me_dpages;
2907 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
2908 env->me_dpages = dp->mp_next;
2912 free(env->me_dbs[1]);
2913 free(env->me_dbs[0]);
2917 LAZY_RWLOCK_DESTROY(&env->me_dblock);
2918 pthread_key_delete(env->me_txkey);
2921 munmap(env->me_map, env->me_mapsize);
2926 pid_t pid = getpid();
2928 for (i=0; i<env->me_txns->mti_numreaders; i++)
2929 if (env->me_txns->mti_readers[i].mr_pid == pid)
2930 env->me_txns->mti_readers[i].mr_pid = 0;
2931 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
2934 mdb_midl_free(env->me_free_pgs);
2938 /** Compare two items pointing at aligned size_t's */
2940 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
2942 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
2943 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
2946 /** Compare two items pointing at aligned int's */
2948 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
2950 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
2951 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
2954 /** Compare two items pointing at ints of unknown alignment.
2955 * Nodes and keys are guaranteed to be 2-byte aligned.
2958 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
2960 #if BYTE_ORDER == LITTLE_ENDIAN
2961 unsigned short *u, *c;
2964 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
2965 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
2968 } while(!x && u > (unsigned short *)a->mv_data);
2971 return memcmp(a->mv_data, b->mv_data, a->mv_size);
2975 /** Compare two items lexically */
2977 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
2984 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
2990 diff = memcmp(a->mv_data, b->mv_data, len);
2991 return diff ? diff : len_diff<0 ? -1 : len_diff;
2994 /** Compare two items in reverse byte order */
2996 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
2998 const unsigned char *p1, *p2, *p1_lim;
3002 p1_lim = (const unsigned char *)a->mv_data;
3003 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3004 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3006 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3012 while (p1 > p1_lim) {
3013 diff = *--p1 - *--p2;
3017 return len_diff<0 ? -1 : len_diff;
3020 /** Search for key within a page, using binary search.
3021 * Returns the smallest entry larger or equal to the key.
3022 * If exactp is non-null, stores whether the found entry was an exact match
3023 * in *exactp (1 or 0).
3024 * Updates the cursor index with the index of the found entry.
3025 * If no entry larger or equal to the key is found, returns NULL.
3028 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3030 unsigned int i = 0, nkeys;
3033 MDB_page *mp = mc->mc_pg[mc->mc_top];
3034 MDB_node *node = NULL;
3039 nkeys = NUMKEYS(mp);
3044 COPY_PGNO(pgno, mp->mp_pgno);
3045 DPRINTF("searching %u keys in %s %spage %zu",
3046 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3053 low = IS_LEAF(mp) ? 0 : 1;
3055 cmp = mc->mc_dbx->md_cmp;
3057 /* Branch pages have no data, so if using integer keys,
3058 * alignment is guaranteed. Use faster mdb_cmp_int.
3060 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3061 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3068 nodekey.mv_size = mc->mc_db->md_pad;
3069 node = NODEPTR(mp, 0); /* fake */
3070 while (low <= high) {
3071 i = (low + high) >> 1;
3072 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3073 rc = cmp(key, &nodekey);
3074 DPRINTF("found leaf index %u [%s], rc = %i",
3075 i, DKEY(&nodekey), rc);
3084 while (low <= high) {
3085 i = (low + high) >> 1;
3087 node = NODEPTR(mp, i);
3088 nodekey.mv_size = NODEKSZ(node);
3089 nodekey.mv_data = NODEKEY(node);
3091 rc = cmp(key, &nodekey);
3094 DPRINTF("found leaf index %u [%s], rc = %i",
3095 i, DKEY(&nodekey), rc);
3097 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3098 i, DKEY(&nodekey), NODEPGNO(node), rc);
3109 if (rc > 0) { /* Found entry is less than the key. */
3110 i++; /* Skip to get the smallest entry larger than key. */
3112 node = NODEPTR(mp, i);
3115 *exactp = (rc == 0);
3116 /* store the key index */
3117 mc->mc_ki[mc->mc_top] = i;
3119 /* There is no entry larger or equal to the key. */
3122 /* nodeptr is fake for LEAF2 */
3128 mdb_cursor_adjust(MDB_cursor *mc, func)
3132 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3133 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3140 /** Pop a page off the top of the cursor's stack. */
3142 mdb_cursor_pop(MDB_cursor *mc)
3147 top = mc->mc_pg[mc->mc_top];
3152 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3153 mc->mc_dbi, (void *) mc);
3157 /** Push a page onto the top of the cursor's stack. */
3159 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3161 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3162 mc->mc_dbi, (void *) mc);
3164 if (mc->mc_snum >= CURSOR_STACK) {
3165 assert(mc->mc_snum < CURSOR_STACK);
3169 mc->mc_top = mc->mc_snum++;
3170 mc->mc_pg[mc->mc_top] = mp;
3171 mc->mc_ki[mc->mc_top] = 0;
3176 /** Find the address of the page corresponding to a given page number.
3177 * @param[in] txn the transaction for this access.
3178 * @param[in] pgno the page number for the page to retrieve.
3179 * @param[out] ret address of a pointer where the page's address will be stored.
3180 * @return 0 on success, non-zero on failure.
3183 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3187 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3189 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3190 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3191 p = txn->mt_u.dirty_list[x].mptr;
3195 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
3196 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3200 DPRINTF("page %zu not found", pgno);
3203 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3206 /** Search for the page a given key should be in.
3207 * Pushes parent pages on the cursor stack. This function continues a
3208 * search on a cursor that has already been initialized. (Usually by
3209 * #mdb_page_search() but also by #mdb_node_move().)
3210 * @param[in,out] mc the cursor for this operation.
3211 * @param[in] key the key to search for. If NULL, search for the lowest
3212 * page. (This is used by #mdb_cursor_first().)
3213 * @param[in] modify If true, visited pages are updated with new page numbers.
3214 * @return 0 on success, non-zero on failure.
3217 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3219 MDB_page *mp = mc->mc_pg[mc->mc_top];
3224 while (IS_BRANCH(mp)) {
3228 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3229 assert(NUMKEYS(mp) > 1);
3230 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3232 if (key == NULL) /* Initialize cursor to first page. */
3234 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3235 /* cursor to last page */
3239 node = mdb_node_search(mc, key, &exact);
3241 i = NUMKEYS(mp) - 1;
3243 i = mc->mc_ki[mc->mc_top];
3252 DPRINTF("following index %u for key [%s]",
3254 assert(i < NUMKEYS(mp));
3255 node = NODEPTR(mp, i);
3257 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3260 mc->mc_ki[mc->mc_top] = i;
3261 if ((rc = mdb_cursor_push(mc, mp)))
3265 if ((rc = mdb_page_touch(mc)) != 0)
3267 mp = mc->mc_pg[mc->mc_top];
3272 DPRINTF("internal error, index points to a %02X page!?",
3274 return MDB_CORRUPTED;
3277 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3278 key ? DKEY(key) : NULL);
3283 /** Search for the page a given key should be in.
3284 * Pushes parent pages on the cursor stack. This function just sets up
3285 * the search; it finds the root page for \b mc's database and sets this
3286 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3287 * called to complete the search.
3288 * @param[in,out] mc the cursor for this operation.
3289 * @param[in] key the key to search for. If NULL, search for the lowest
3290 * page. (This is used by #mdb_cursor_first().)
3291 * @param[in] modify If true, visited pages are updated with new page numbers.
3292 * @return 0 on success, non-zero on failure.
3295 mdb_page_search(MDB_cursor *mc, MDB_val *key, int modify)
3300 /* Make sure the txn is still viable, then find the root from
3301 * the txn's db table.
3303 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3304 DPUTS("transaction has failed, must abort");
3307 /* Make sure we're using an up-to-date root */
3308 if (mc->mc_dbi > MAIN_DBI) {
3309 if ((*mc->mc_dbflag & DB_STALE) ||
3310 (modify && !(*mc->mc_dbflag & DB_DIRTY))) {
3312 unsigned char dbflag = 0;
3313 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3314 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, modify);
3317 if (*mc->mc_dbflag & DB_STALE) {
3320 MDB_node *leaf = mdb_node_search(&mc2,
3321 &mc->mc_dbx->md_name, &exact);
3323 return MDB_NOTFOUND;
3324 mdb_node_read(mc->mc_txn, leaf, &data);
3325 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3329 *mc->mc_dbflag = dbflag;
3332 root = mc->mc_db->md_root;
3334 if (root == P_INVALID) { /* Tree is empty. */
3335 DPUTS("tree is empty");
3336 return MDB_NOTFOUND;
3341 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3347 DPRINTF("db %u root page %zu has flags 0x%X",
3348 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3351 if ((rc = mdb_page_touch(mc)))
3355 return mdb_page_search_root(mc, key, modify);
3358 /** Return the data associated with a given node.
3359 * @param[in] txn The transaction for this operation.
3360 * @param[in] leaf The node being read.
3361 * @param[out] data Updated to point to the node's data.
3362 * @return 0 on success, non-zero on failure.
3365 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3367 MDB_page *omp; /* overflow page */
3371 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3372 data->mv_size = NODEDSZ(leaf);
3373 data->mv_data = NODEDATA(leaf);
3377 /* Read overflow data.
3379 data->mv_size = NODEDSZ(leaf);
3380 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3381 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3382 DPRINTF("read overflow page %zu failed", pgno);
3385 data->mv_data = METADATA(omp);
3391 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3392 MDB_val *key, MDB_val *data)
3401 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3403 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3406 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3410 mdb_cursor_init(&mc, txn, dbi, &mx);
3411 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3414 /** Find a sibling for a page.
3415 * Replaces the page at the top of the cursor's stack with the
3416 * specified sibling, if one exists.
3417 * @param[in] mc The cursor for this operation.
3418 * @param[in] move_right Non-zero if the right sibling is requested,
3419 * otherwise the left sibling.
3420 * @return 0 on success, non-zero on failure.
3423 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3429 if (mc->mc_snum < 2) {
3430 return MDB_NOTFOUND; /* root has no siblings */
3434 DPRINTF("parent page is page %zu, index %u",
3435 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3437 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3438 : (mc->mc_ki[mc->mc_top] == 0)) {
3439 DPRINTF("no more keys left, moving to %s sibling",
3440 move_right ? "right" : "left");
3441 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3445 mc->mc_ki[mc->mc_top]++;
3447 mc->mc_ki[mc->mc_top]--;
3448 DPRINTF("just moving to %s index key %u",
3449 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3451 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3453 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3454 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3457 mdb_cursor_push(mc, mp);
3462 /** Move the cursor to the next data item. */
3464 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3470 if (mc->mc_flags & C_EOF) {
3471 return MDB_NOTFOUND;
3474 assert(mc->mc_flags & C_INITIALIZED);
3476 mp = mc->mc_pg[mc->mc_top];
3478 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3479 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3480 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3481 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3482 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3483 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3487 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3488 if (op == MDB_NEXT_DUP)
3489 return MDB_NOTFOUND;
3493 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3495 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3496 DPUTS("=====> move to next sibling page");
3497 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3498 mc->mc_flags |= C_EOF;
3499 mc->mc_flags &= ~C_INITIALIZED;
3500 return MDB_NOTFOUND;
3502 mp = mc->mc_pg[mc->mc_top];
3503 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3505 mc->mc_ki[mc->mc_top]++;
3507 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3508 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3511 key->mv_size = mc->mc_db->md_pad;
3512 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3516 assert(IS_LEAF(mp));
3517 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3519 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3520 mdb_xcursor_init1(mc, leaf);
3523 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3526 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3527 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3528 if (rc != MDB_SUCCESS)
3533 MDB_SET_KEY(leaf, key);
3537 /** Move the cursor to the previous data item. */
3539 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3545 assert(mc->mc_flags & C_INITIALIZED);
3547 mp = mc->mc_pg[mc->mc_top];
3549 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3550 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3551 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3552 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3553 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3554 if (op != MDB_PREV || rc == MDB_SUCCESS)
3557 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3558 if (op == MDB_PREV_DUP)
3559 return MDB_NOTFOUND;
3564 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3566 if (mc->mc_ki[mc->mc_top] == 0) {
3567 DPUTS("=====> move to prev sibling page");
3568 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3569 mc->mc_flags &= ~C_INITIALIZED;
3570 return MDB_NOTFOUND;
3572 mp = mc->mc_pg[mc->mc_top];
3573 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3574 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3576 mc->mc_ki[mc->mc_top]--;
3578 mc->mc_flags &= ~C_EOF;
3580 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3581 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3584 key->mv_size = mc->mc_db->md_pad;
3585 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3589 assert(IS_LEAF(mp));
3590 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3592 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3593 mdb_xcursor_init1(mc, leaf);
3596 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3599 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3600 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3601 if (rc != MDB_SUCCESS)
3606 MDB_SET_KEY(leaf, key);
3610 /** Set the cursor on a specific data item. */
3612 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3613 MDB_cursor_op op, int *exactp)
3622 assert(key->mv_size > 0);
3624 /* See if we're already on the right page */
3625 if (mc->mc_flags & C_INITIALIZED) {
3628 mp = mc->mc_pg[mc->mc_top];
3630 mc->mc_ki[mc->mc_top] = 0;
3631 return MDB_NOTFOUND;
3633 if (mp->mp_flags & P_LEAF2) {
3634 nodekey.mv_size = mc->mc_db->md_pad;
3635 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3637 leaf = NODEPTR(mp, 0);
3638 MDB_SET_KEY(leaf, &nodekey);
3640 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3642 /* Probably happens rarely, but first node on the page
3643 * was the one we wanted.
3645 mc->mc_ki[mc->mc_top] = 0;
3646 leaf = NODEPTR(mp, 0);
3653 unsigned int nkeys = NUMKEYS(mp);
3655 if (mp->mp_flags & P_LEAF2) {
3656 nodekey.mv_data = LEAF2KEY(mp,
3657 nkeys-1, nodekey.mv_size);
3659 leaf = NODEPTR(mp, nkeys-1);
3660 MDB_SET_KEY(leaf, &nodekey);
3662 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3664 /* last node was the one we wanted */
3665 mc->mc_ki[mc->mc_top] = nkeys-1;
3666 leaf = NODEPTR(mp, nkeys-1);
3672 /* This is definitely the right page, skip search_page */
3677 /* If any parents have right-sibs, search.
3678 * Otherwise, there's nothing further.
3680 for (i=0; i<mc->mc_top; i++)
3682 NUMKEYS(mc->mc_pg[i])-1)
3684 if (i == mc->mc_top) {
3685 /* There are no other pages */
3686 mc->mc_ki[mc->mc_top] = nkeys;
3687 return MDB_NOTFOUND;
3691 /* There are no other pages */
3692 mc->mc_ki[mc->mc_top] = 0;
3693 return MDB_NOTFOUND;
3697 rc = mdb_page_search(mc, key, 0);
3698 if (rc != MDB_SUCCESS)
3701 mp = mc->mc_pg[mc->mc_top];
3702 assert(IS_LEAF(mp));
3705 leaf = mdb_node_search(mc, key, exactp);
3706 if (exactp != NULL && !*exactp) {
3707 /* MDB_SET specified and not an exact match. */
3708 return MDB_NOTFOUND;
3712 DPUTS("===> inexact leaf not found, goto sibling");
3713 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3714 return rc; /* no entries matched */
3715 mp = mc->mc_pg[mc->mc_top];
3716 assert(IS_LEAF(mp));
3717 leaf = NODEPTR(mp, 0);
3721 mc->mc_flags |= C_INITIALIZED;
3722 mc->mc_flags &= ~C_EOF;
3725 key->mv_size = mc->mc_db->md_pad;
3726 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3730 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3731 mdb_xcursor_init1(mc, leaf);
3734 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3735 if (op == MDB_SET || op == MDB_SET_RANGE) {
3736 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3739 if (op == MDB_GET_BOTH) {
3745 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3746 if (rc != MDB_SUCCESS)
3749 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3751 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3753 rc = mc->mc_dbx->md_dcmp(data, &d2);
3755 if (op == MDB_GET_BOTH || rc > 0)
3756 return MDB_NOTFOUND;
3761 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3762 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3767 /* The key already matches in all other cases */
3768 if (op == MDB_SET_RANGE)
3769 MDB_SET_KEY(leaf, key);
3770 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3775 /** Move the cursor to the first item in the database. */
3777 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3782 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3783 rc = mdb_page_search(mc, NULL, 0);
3784 if (rc != MDB_SUCCESS)
3787 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3789 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3790 mc->mc_flags |= C_INITIALIZED;
3791 mc->mc_flags &= ~C_EOF;
3793 mc->mc_ki[mc->mc_top] = 0;
3795 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3796 key->mv_size = mc->mc_db->md_pad;
3797 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3802 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3803 mdb_xcursor_init1(mc, leaf);
3804 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3809 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3810 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3814 MDB_SET_KEY(leaf, key);
3818 /** Move the cursor to the last item in the database. */
3820 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3826 lkey.mv_size = MAXKEYSIZE+1;
3827 lkey.mv_data = NULL;
3829 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3830 rc = mdb_page_search(mc, &lkey, 0);
3831 if (rc != MDB_SUCCESS)
3834 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3836 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3837 mc->mc_flags |= C_INITIALIZED;
3838 mc->mc_flags &= ~C_EOF;
3840 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3842 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3843 key->mv_size = mc->mc_db->md_pad;
3844 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3849 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3850 mdb_xcursor_init1(mc, leaf);
3851 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3856 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3857 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3862 MDB_SET_KEY(leaf, key);
3867 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3877 case MDB_GET_BOTH_RANGE:
3878 if (data == NULL || mc->mc_xcursor == NULL) {
3885 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3887 } else if (op == MDB_SET_RANGE)
3888 rc = mdb_cursor_set(mc, key, data, op, NULL);
3890 rc = mdb_cursor_set(mc, key, data, op, &exact);
3892 case MDB_GET_MULTIPLE:
3894 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
3895 !(mc->mc_flags & C_INITIALIZED)) {
3900 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
3901 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
3904 case MDB_NEXT_MULTIPLE:
3906 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
3910 if (!(mc->mc_flags & C_INITIALIZED))
3911 rc = mdb_cursor_first(mc, key, data);
3913 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
3914 if (rc == MDB_SUCCESS) {
3915 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
3918 mx = &mc->mc_xcursor->mx_cursor;
3919 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
3921 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
3922 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
3930 case MDB_NEXT_NODUP:
3931 if (!(mc->mc_flags & C_INITIALIZED))
3932 rc = mdb_cursor_first(mc, key, data);
3934 rc = mdb_cursor_next(mc, key, data, op);
3938 case MDB_PREV_NODUP:
3939 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
3940 rc = mdb_cursor_last(mc, key, data);
3942 rc = mdb_cursor_prev(mc, key, data, op);
3945 rc = mdb_cursor_first(mc, key, data);
3949 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3950 !(mc->mc_flags & C_INITIALIZED) ||
3951 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3955 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3958 rc = mdb_cursor_last(mc, key, data);
3962 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3963 !(mc->mc_flags & C_INITIALIZED) ||
3964 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3968 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3971 DPRINTF("unhandled/unimplemented cursor operation %u", op);
3979 /** Touch all the pages in the cursor stack.
3980 * Makes sure all the pages are writable, before attempting a write operation.
3981 * @param[in] mc The cursor to operate on.
3984 mdb_cursor_touch(MDB_cursor *mc)
3988 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
3990 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3991 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 1);
3994 *mc->mc_dbflag = DB_DIRTY;
3996 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
3997 rc = mdb_page_touch(mc);
4001 mc->mc_top = mc->mc_snum-1;
4006 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4009 MDB_node *leaf = NULL;
4010 MDB_val xdata, *rdata, dkey;
4014 unsigned int mcount = 0;
4018 char dbuf[MAXKEYSIZE+1];
4019 unsigned int nflags;
4022 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4025 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4026 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4030 if (flags == MDB_CURRENT) {
4031 if (!(mc->mc_flags & C_INITIALIZED))
4034 } else if (mc->mc_db->md_root == P_INVALID) {
4036 /* new database, write a root leaf page */
4037 DPUTS("allocating new root leaf page");
4038 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4042 mdb_cursor_push(mc, np);
4043 mc->mc_db->md_root = np->mp_pgno;
4044 mc->mc_db->md_depth++;
4045 *mc->mc_dbflag = DB_DIRTY;
4046 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4048 np->mp_flags |= P_LEAF2;
4049 mc->mc_flags |= C_INITIALIZED;
4055 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4056 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4057 DPRINTF("duplicate key [%s]", DKEY(key));
4059 return MDB_KEYEXIST;
4061 if (rc && rc != MDB_NOTFOUND)
4065 /* Cursor is positioned, now make sure all pages are writable */
4066 rc2 = mdb_cursor_touch(mc);
4071 /* The key already exists */
4072 if (rc == MDB_SUCCESS) {
4073 /* there's only a key anyway, so this is a no-op */
4074 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4075 unsigned int ksize = mc->mc_db->md_pad;
4076 if (key->mv_size != ksize)
4078 if (flags == MDB_CURRENT) {
4079 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4080 memcpy(ptr, key->mv_data, ksize);
4085 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4088 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4089 /* Was a single item before, must convert now */
4091 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4092 /* Just overwrite the current item */
4093 if (flags == MDB_CURRENT)
4096 dkey.mv_size = NODEDSZ(leaf);
4097 dkey.mv_data = NODEDATA(leaf);
4098 #if UINT_MAX < SIZE_MAX
4099 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4100 #ifdef MISALIGNED_OK
4101 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4103 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4106 /* if data matches, ignore it */
4107 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4108 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4110 /* create a fake page for the dup items */
4111 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4112 dkey.mv_data = dbuf;
4113 fp = (MDB_page *)&pbuf;
4114 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4115 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4116 fp->mp_lower = PAGEHDRSZ;
4117 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4118 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4119 fp->mp_flags |= P_LEAF2;
4120 fp->mp_pad = data->mv_size;
4122 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4123 (dkey.mv_size & 1) + (data->mv_size & 1);
4125 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4128 xdata.mv_size = fp->mp_upper;
4133 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4134 /* See if we need to convert from fake page to subDB */
4136 unsigned int offset;
4139 fp = NODEDATA(leaf);
4140 if (flags == MDB_CURRENT) {
4141 fp->mp_flags |= P_DIRTY;
4142 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4143 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4147 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4148 offset = fp->mp_pad;
4150 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4152 offset += offset & 1;
4153 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4154 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4156 /* yes, convert it */
4158 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4159 dummy.md_pad = fp->mp_pad;
4160 dummy.md_flags = MDB_DUPFIXED;
4161 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4162 dummy.md_flags |= MDB_INTEGERKEY;
4165 dummy.md_branch_pages = 0;
4166 dummy.md_leaf_pages = 1;
4167 dummy.md_overflow_pages = 0;
4168 dummy.md_entries = NUMKEYS(fp);
4170 xdata.mv_size = sizeof(MDB_db);
4171 xdata.mv_data = &dummy;
4172 mp = mdb_page_alloc(mc, 1);
4175 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4176 flags |= F_DUPDATA|F_SUBDATA;
4177 dummy.md_root = mp->mp_pgno;
4179 /* no, just grow it */
4181 xdata.mv_size = NODEDSZ(leaf) + offset;
4182 xdata.mv_data = &pbuf;
4183 mp = (MDB_page *)&pbuf;
4184 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4187 mp->mp_flags = fp->mp_flags | P_DIRTY;
4188 mp->mp_pad = fp->mp_pad;
4189 mp->mp_lower = fp->mp_lower;
4190 mp->mp_upper = fp->mp_upper + offset;
4192 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4194 nsize = NODEDSZ(leaf) - fp->mp_upper;
4195 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4196 for (i=0; i<NUMKEYS(fp); i++)
4197 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4199 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4203 /* data is on sub-DB, just store it */
4204 flags |= F_DUPDATA|F_SUBDATA;
4208 /* same size, just replace it */
4209 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
4210 NODEDSZ(leaf) == data->mv_size) {
4211 if (F_ISSET(flags, MDB_RESERVE))
4212 data->mv_data = NODEDATA(leaf);
4214 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4217 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4219 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4225 nflags = flags & NODE_ADD_FLAGS;
4226 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4227 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4228 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4229 nflags &= ~MDB_APPEND;
4230 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4232 /* There is room already in this leaf page. */
4233 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4234 if (rc == 0 && !do_sub) {
4235 /* Adjust other cursors pointing to mp */
4236 MDB_cursor *m2, *m3;
4237 MDB_dbi dbi = mc->mc_dbi;
4238 unsigned i = mc->mc_top;
4239 MDB_page *mp = mc->mc_pg[i];
4241 if (mc->mc_flags & C_SUB)
4244 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4245 if (mc->mc_flags & C_SUB)
4246 m3 = &m2->mc_xcursor->mx_cursor;
4249 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4250 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4257 if (rc != MDB_SUCCESS)
4258 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4260 /* Now store the actual data in the child DB. Note that we're
4261 * storing the user data in the keys field, so there are strict
4262 * size limits on dupdata. The actual data fields of the child
4263 * DB are all zero size.
4270 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4271 if (flags & MDB_CURRENT) {
4272 xflags = MDB_CURRENT;
4274 mdb_xcursor_init1(mc, leaf);
4275 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4277 /* converted, write the original data first */
4279 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4283 /* Adjust other cursors pointing to mp */
4285 unsigned i = mc->mc_top;
4286 MDB_page *mp = mc->mc_pg[i];
4288 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4289 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4290 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4291 mdb_xcursor_init1(m2, leaf);
4296 xflags |= (flags & MDB_APPEND);
4297 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4298 if (flags & F_SUBDATA) {
4299 void *db = NODEDATA(leaf);
4300 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4303 /* sub-writes might have failed so check rc again.
4304 * Don't increment count if we just replaced an existing item.
4306 if (!rc && !(flags & MDB_CURRENT))
4307 mc->mc_db->md_entries++;
4308 if (flags & MDB_MULTIPLE) {
4310 if (mcount < data[1].mv_size) {
4311 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4312 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4322 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4327 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4330 if (!mc->mc_flags & C_INITIALIZED)
4333 rc = mdb_cursor_touch(mc);
4337 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4339 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4340 if (flags != MDB_NODUPDATA) {
4341 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4342 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4344 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4345 /* If sub-DB still has entries, we're done */
4346 if (mc->mc_xcursor->mx_db.md_entries) {
4347 if (leaf->mn_flags & F_SUBDATA) {
4348 /* update subDB info */
4349 void *db = NODEDATA(leaf);
4350 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4352 /* shrink fake page */
4353 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4355 mc->mc_db->md_entries--;
4358 /* otherwise fall thru and delete the sub-DB */
4361 if (leaf->mn_flags & F_SUBDATA) {
4362 /* add all the child DB's pages to the free list */
4363 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4364 if (rc == MDB_SUCCESS) {
4365 mc->mc_db->md_entries -=
4366 mc->mc_xcursor->mx_db.md_entries;
4371 return mdb_cursor_del0(mc, leaf);
4374 /** Allocate and initialize new pages for a database.
4375 * @param[in] mc a cursor on the database being added to.
4376 * @param[in] flags flags defining what type of page is being allocated.
4377 * @param[in] num the number of pages to allocate. This is usually 1,
4378 * unless allocating overflow pages for a large record.
4379 * @return Address of a page, or NULL on failure.
4382 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4386 if ((np = mdb_page_alloc(mc, num)) == NULL)
4388 DPRINTF("allocated new mpage %zu, page size %u",
4389 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4390 np->mp_flags = flags | P_DIRTY;
4391 np->mp_lower = PAGEHDRSZ;
4392 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4395 mc->mc_db->md_branch_pages++;
4396 else if (IS_LEAF(np))
4397 mc->mc_db->md_leaf_pages++;
4398 else if (IS_OVERFLOW(np)) {
4399 mc->mc_db->md_overflow_pages += num;
4406 /** Calculate the size of a leaf node.
4407 * The size depends on the environment's page size; if a data item
4408 * is too large it will be put onto an overflow page and the node
4409 * size will only include the key and not the data. Sizes are always
4410 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4411 * of the #MDB_node headers.
4412 * @param[in] env The environment handle.
4413 * @param[in] key The key for the node.
4414 * @param[in] data The data for the node.
4415 * @return The number of bytes needed to store the node.
4418 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4422 sz = LEAFSIZE(key, data);
4423 if (sz >= env->me_psize / MDB_MINKEYS) {
4424 /* put on overflow page */
4425 sz -= data->mv_size - sizeof(pgno_t);
4429 return sz + sizeof(indx_t);
4432 /** Calculate the size of a branch node.
4433 * The size should depend on the environment's page size but since
4434 * we currently don't support spilling large keys onto overflow
4435 * pages, it's simply the size of the #MDB_node header plus the
4436 * size of the key. Sizes are always rounded up to an even number
4437 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4438 * @param[in] env The environment handle.
4439 * @param[in] key The key for the node.
4440 * @return The number of bytes needed to store the node.
4443 mdb_branch_size(MDB_env *env, MDB_val *key)
4448 if (sz >= env->me_psize / MDB_MINKEYS) {
4449 /* put on overflow page */
4450 /* not implemented */
4451 /* sz -= key->size - sizeof(pgno_t); */
4454 return sz + sizeof(indx_t);
4457 /** Add a node to the page pointed to by the cursor.
4458 * @param[in] mc The cursor for this operation.
4459 * @param[in] indx The index on the page where the new node should be added.
4460 * @param[in] key The key for the new node.
4461 * @param[in] data The data for the new node, if any.
4462 * @param[in] pgno The page number, if adding a branch node.
4463 * @param[in] flags Flags for the node.
4464 * @return 0 on success, non-zero on failure. Possible errors are:
4466 * <li>ENOMEM - failed to allocate overflow pages for the node.
4467 * <li>ENOSPC - there is insufficient room in the page. This error
4468 * should never happen since all callers already calculate the
4469 * page's free space before calling this function.
4473 mdb_node_add(MDB_cursor *mc, indx_t indx,
4474 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4477 size_t node_size = NODESIZE;
4480 MDB_page *mp = mc->mc_pg[mc->mc_top];
4481 MDB_page *ofp = NULL; /* overflow page */
4484 assert(mp->mp_upper >= mp->mp_lower);
4486 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4487 IS_LEAF(mp) ? "leaf" : "branch",
4488 IS_SUBP(mp) ? "sub-" : "",
4489 mp->mp_pgno, indx, data ? data->mv_size : 0,
4490 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4493 /* Move higher keys up one slot. */
4494 int ksize = mc->mc_db->md_pad, dif;
4495 char *ptr = LEAF2KEY(mp, indx, ksize);
4496 dif = NUMKEYS(mp) - indx;
4498 memmove(ptr+ksize, ptr, dif*ksize);
4499 /* insert new key */
4500 memcpy(ptr, key->mv_data, ksize);
4502 /* Just using these for counting */
4503 mp->mp_lower += sizeof(indx_t);
4504 mp->mp_upper -= ksize - sizeof(indx_t);
4509 node_size += key->mv_size;
4513 if (F_ISSET(flags, F_BIGDATA)) {
4514 /* Data already on overflow page. */
4515 node_size += sizeof(pgno_t);
4516 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4517 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4518 /* Put data on overflow page. */
4519 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4520 data->mv_size, node_size+data->mv_size);
4521 node_size += sizeof(pgno_t);
4522 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4524 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4527 node_size += data->mv_size;
4530 node_size += node_size & 1;
4532 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4533 DPRINTF("not enough room in page %zu, got %u ptrs",
4534 mp->mp_pgno, NUMKEYS(mp));
4535 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4536 mp->mp_upper - mp->mp_lower);
4537 DPRINTF("node size = %zu", node_size);
4541 /* Move higher pointers up one slot. */
4542 for (i = NUMKEYS(mp); i > indx; i--)
4543 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4545 /* Adjust free space offsets. */
4546 ofs = mp->mp_upper - node_size;
4547 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4548 mp->mp_ptrs[indx] = ofs;
4550 mp->mp_lower += sizeof(indx_t);
4552 /* Write the node data. */
4553 node = NODEPTR(mp, indx);
4554 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4555 node->mn_flags = flags;
4557 SETDSZ(node,data->mv_size);
4562 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4567 if (F_ISSET(flags, F_BIGDATA))
4568 memcpy(node->mn_data + key->mv_size, data->mv_data,
4570 else if (F_ISSET(flags, MDB_RESERVE))
4571 data->mv_data = node->mn_data + key->mv_size;
4573 memcpy(node->mn_data + key->mv_size, data->mv_data,
4576 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4578 if (F_ISSET(flags, MDB_RESERVE))
4579 data->mv_data = METADATA(ofp);
4581 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4588 /** Delete the specified node from a page.
4589 * @param[in] mp The page to operate on.
4590 * @param[in] indx The index of the node to delete.
4591 * @param[in] ksize The size of a node. Only used if the page is
4592 * part of a #MDB_DUPFIXED database.
4595 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4598 indx_t i, j, numkeys, ptr;
4605 COPY_PGNO(pgno, mp->mp_pgno);
4606 DPRINTF("delete node %u on %s page %zu", indx,
4607 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4610 assert(indx < NUMKEYS(mp));
4613 int x = NUMKEYS(mp) - 1 - indx;
4614 base = LEAF2KEY(mp, indx, ksize);
4616 memmove(base, base + ksize, x * ksize);
4617 mp->mp_lower -= sizeof(indx_t);
4618 mp->mp_upper += ksize - sizeof(indx_t);
4622 node = NODEPTR(mp, indx);
4623 sz = NODESIZE + node->mn_ksize;
4625 if (F_ISSET(node->mn_flags, F_BIGDATA))
4626 sz += sizeof(pgno_t);
4628 sz += NODEDSZ(node);
4632 ptr = mp->mp_ptrs[indx];
4633 numkeys = NUMKEYS(mp);
4634 for (i = j = 0; i < numkeys; i++) {
4636 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4637 if (mp->mp_ptrs[i] < ptr)
4638 mp->mp_ptrs[j] += sz;
4643 base = (char *)mp + mp->mp_upper;
4644 memmove(base + sz, base, ptr - mp->mp_upper);
4646 mp->mp_lower -= sizeof(indx_t);
4650 /** Compact the main page after deleting a node on a subpage.
4651 * @param[in] mp The main page to operate on.
4652 * @param[in] indx The index of the subpage on the main page.
4655 mdb_node_shrink(MDB_page *mp, indx_t indx)
4662 indx_t i, numkeys, ptr;
4664 node = NODEPTR(mp, indx);
4665 sp = (MDB_page *)NODEDATA(node);
4666 osize = NODEDSZ(node);
4668 delta = sp->mp_upper - sp->mp_lower;
4669 SETDSZ(node, osize - delta);
4670 xp = (MDB_page *)((char *)sp + delta);
4672 /* shift subpage upward */
4674 nsize = NUMKEYS(sp) * sp->mp_pad;
4675 memmove(METADATA(xp), METADATA(sp), nsize);
4678 nsize = osize - sp->mp_upper;
4679 numkeys = NUMKEYS(sp);
4680 for (i=numkeys-1; i>=0; i--)
4681 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4683 xp->mp_upper = sp->mp_lower;
4684 xp->mp_lower = sp->mp_lower;
4685 xp->mp_flags = sp->mp_flags;
4686 xp->mp_pad = sp->mp_pad;
4687 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4689 /* shift lower nodes upward */
4690 ptr = mp->mp_ptrs[indx];
4691 numkeys = NUMKEYS(mp);
4692 for (i = 0; i < numkeys; i++) {
4693 if (mp->mp_ptrs[i] <= ptr)
4694 mp->mp_ptrs[i] += delta;
4697 base = (char *)mp + mp->mp_upper;
4698 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4699 mp->mp_upper += delta;
4702 /** Initial setup of a sorted-dups cursor.
4703 * Sorted duplicates are implemented as a sub-database for the given key.
4704 * The duplicate data items are actually keys of the sub-database.
4705 * Operations on the duplicate data items are performed using a sub-cursor
4706 * initialized when the sub-database is first accessed. This function does
4707 * the preliminary setup of the sub-cursor, filling in the fields that
4708 * depend only on the parent DB.
4709 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4712 mdb_xcursor_init0(MDB_cursor *mc)
4714 MDB_xcursor *mx = mc->mc_xcursor;
4716 mx->mx_cursor.mc_xcursor = NULL;
4717 mx->mx_cursor.mc_txn = mc->mc_txn;
4718 mx->mx_cursor.mc_db = &mx->mx_db;
4719 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4720 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4721 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4722 mx->mx_cursor.mc_snum = 0;
4723 mx->mx_cursor.mc_flags = C_SUB;
4724 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4725 mx->mx_dbx.md_dcmp = NULL;
4726 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4729 /** Final setup of a sorted-dups cursor.
4730 * Sets up the fields that depend on the data from the main cursor.
4731 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4732 * @param[in] node The data containing the #MDB_db record for the
4733 * sorted-dup database.
4736 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
4738 MDB_xcursor *mx = mc->mc_xcursor;
4740 if (node->mn_flags & F_SUBDATA) {
4741 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
4742 mx->mx_cursor.mc_snum = 0;
4743 mx->mx_cursor.mc_flags = C_SUB;
4745 MDB_page *fp = NODEDATA(node);
4746 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
4747 mx->mx_db.md_flags = 0;
4748 mx->mx_db.md_depth = 1;
4749 mx->mx_db.md_branch_pages = 0;
4750 mx->mx_db.md_leaf_pages = 1;
4751 mx->mx_db.md_overflow_pages = 0;
4752 mx->mx_db.md_entries = NUMKEYS(fp);
4753 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
4754 mx->mx_cursor.mc_snum = 1;
4755 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
4756 mx->mx_cursor.mc_top = 0;
4757 mx->mx_cursor.mc_pg[0] = fp;
4758 mx->mx_cursor.mc_ki[0] = 0;
4759 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4760 mx->mx_db.md_flags = MDB_DUPFIXED;
4761 mx->mx_db.md_pad = fp->mp_pad;
4762 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4763 mx->mx_db.md_flags |= MDB_INTEGERKEY;
4766 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
4768 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
4770 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
4771 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
4772 #if UINT_MAX < SIZE_MAX
4773 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
4774 #ifdef MISALIGNED_OK
4775 mx->mx_dbx.md_cmp = mdb_cmp_long;
4777 mx->mx_dbx.md_cmp = mdb_cmp_cint;
4782 /** Initialize a cursor for a given transaction and database. */
4784 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
4789 mc->mc_db = &txn->mt_dbs[dbi];
4790 mc->mc_dbx = &txn->mt_dbxs[dbi];
4791 mc->mc_dbflag = &txn->mt_dbflags[dbi];
4795 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4797 mc->mc_xcursor = mx;
4798 mdb_xcursor_init0(mc);
4800 mc->mc_xcursor = NULL;
4805 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
4808 MDB_xcursor *mx = NULL;
4809 size_t size = sizeof(MDB_cursor);
4811 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
4814 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
4815 size += sizeof(MDB_xcursor);
4817 if ((mc = malloc(size)) != NULL) {
4818 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4819 mx = (MDB_xcursor *)(mc + 1);
4821 mdb_cursor_init(mc, txn, dbi, mx);
4822 if (txn->mt_cursors) {
4823 mc->mc_next = txn->mt_cursors[dbi];
4824 txn->mt_cursors[dbi] = mc;
4826 mc->mc_flags |= C_ALLOCD;
4836 /* Return the count of duplicate data items for the current key */
4838 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
4842 if (mc == NULL || countp == NULL)
4845 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
4848 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4849 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4852 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
4855 *countp = mc->mc_xcursor->mx_db.md_entries;
4861 mdb_cursor_close(MDB_cursor *mc)
4864 /* remove from txn, if tracked */
4865 if (mc->mc_txn->mt_cursors) {
4866 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
4867 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
4869 *prev = mc->mc_next;
4871 if (mc->mc_flags & C_ALLOCD)
4877 mdb_cursor_txn(MDB_cursor *mc)
4879 if (!mc) return NULL;
4884 mdb_cursor_dbi(MDB_cursor *mc)
4890 /** Replace the key for a node with a new key.
4891 * @param[in] mp The page containing the node to operate on.
4892 * @param[in] indx The index of the node to operate on.
4893 * @param[in] key The new key to use.
4894 * @return 0 on success, non-zero on failure.
4897 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
4899 indx_t ptr, i, numkeys;
4906 node = NODEPTR(mp, indx);
4907 ptr = mp->mp_ptrs[indx];
4911 char kbuf2[(MAXKEYSIZE*2+1)];
4912 k2.mv_data = NODEKEY(node);
4913 k2.mv_size = node->mn_ksize;
4914 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
4916 mdb_dkey(&k2, kbuf2),
4922 delta = key->mv_size - node->mn_ksize;
4924 if (delta > 0 && SIZELEFT(mp) < delta) {
4925 DPRINTF("OUCH! Not enough room, delta = %d", delta);
4929 numkeys = NUMKEYS(mp);
4930 for (i = 0; i < numkeys; i++) {
4931 if (mp->mp_ptrs[i] <= ptr)
4932 mp->mp_ptrs[i] -= delta;
4935 base = (char *)mp + mp->mp_upper;
4936 len = ptr - mp->mp_upper + NODESIZE;
4937 memmove(base - delta, base, len);
4938 mp->mp_upper -= delta;
4940 node = NODEPTR(mp, indx);
4941 node->mn_ksize = key->mv_size;
4945 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4950 /** Move a node from csrc to cdst.
4953 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
4959 unsigned short flags;
4963 /* Mark src and dst as dirty. */
4964 if ((rc = mdb_page_touch(csrc)) ||
4965 (rc = mdb_page_touch(cdst)))
4968 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
4969 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
4970 key.mv_size = csrc->mc_db->md_pad;
4971 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
4973 data.mv_data = NULL;
4977 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
4978 assert(!((long)srcnode&1));
4979 srcpg = NODEPGNO(srcnode);
4980 flags = srcnode->mn_flags;
4981 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
4982 unsigned int snum = csrc->mc_snum;
4984 /* must find the lowest key below src */
4985 mdb_page_search_root(csrc, NULL, 0);
4986 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
4987 key.mv_size = NODEKSZ(s2);
4988 key.mv_data = NODEKEY(s2);
4989 csrc->mc_snum = snum--;
4990 csrc->mc_top = snum;
4992 key.mv_size = NODEKSZ(srcnode);
4993 key.mv_data = NODEKEY(srcnode);
4995 data.mv_size = NODEDSZ(srcnode);
4996 data.mv_data = NODEDATA(srcnode);
4998 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
4999 unsigned int snum = cdst->mc_snum;
5002 /* must find the lowest key below dst */
5003 mdb_page_search_root(cdst, NULL, 0);
5004 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5005 bkey.mv_size = NODEKSZ(s2);
5006 bkey.mv_data = NODEKEY(s2);
5007 cdst->mc_snum = snum--;
5008 cdst->mc_top = snum;
5009 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5012 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5013 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5014 csrc->mc_ki[csrc->mc_top],
5016 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5017 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5019 /* Add the node to the destination page.
5021 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5022 if (rc != MDB_SUCCESS)
5025 /* Delete the node from the source page.
5027 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5030 /* Adjust other cursors pointing to mp */
5031 MDB_cursor *m2, *m3;
5032 MDB_dbi dbi = csrc->mc_dbi;
5033 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5035 if (csrc->mc_flags & C_SUB)
5038 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5039 if (m2 == csrc) continue;
5040 if (csrc->mc_flags & C_SUB)
5041 m3 = &m2->mc_xcursor->mx_cursor;
5044 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5045 csrc->mc_ki[csrc->mc_top]) {
5046 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5047 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5052 /* Update the parent separators.
5054 if (csrc->mc_ki[csrc->mc_top] == 0) {
5055 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5056 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5057 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5059 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5060 key.mv_size = NODEKSZ(srcnode);
5061 key.mv_data = NODEKEY(srcnode);
5063 DPRINTF("update separator for source page %zu to [%s]",
5064 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5065 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5066 &key)) != MDB_SUCCESS)
5069 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5071 nullkey.mv_size = 0;
5072 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5073 assert(rc == MDB_SUCCESS);
5077 if (cdst->mc_ki[cdst->mc_top] == 0) {
5078 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5079 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5080 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5082 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5083 key.mv_size = NODEKSZ(srcnode);
5084 key.mv_data = NODEKEY(srcnode);
5086 DPRINTF("update separator for destination page %zu to [%s]",
5087 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5088 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5089 &key)) != MDB_SUCCESS)
5092 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5094 nullkey.mv_size = 0;
5095 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5096 assert(rc == MDB_SUCCESS);
5103 /** Merge one page into another.
5104 * The nodes from the page pointed to by \b csrc will
5105 * be copied to the page pointed to by \b cdst and then
5106 * the \b csrc page will be freed.
5107 * @param[in] csrc Cursor pointing to the source page.
5108 * @param[in] cdst Cursor pointing to the destination page.
5111 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5119 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5120 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5122 assert(csrc->mc_snum > 1); /* can't merge root page */
5123 assert(cdst->mc_snum > 1);
5125 /* Mark dst as dirty. */
5126 if ((rc = mdb_page_touch(cdst)))
5129 /* Move all nodes from src to dst.
5131 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5132 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5133 key.mv_size = csrc->mc_db->md_pad;
5134 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5135 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5136 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5137 if (rc != MDB_SUCCESS)
5139 key.mv_data = (char *)key.mv_data + key.mv_size;
5142 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5143 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5144 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5145 unsigned int snum = csrc->mc_snum;
5147 /* must find the lowest key below src */
5148 mdb_page_search_root(csrc, NULL, 0);
5149 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5150 key.mv_size = NODEKSZ(s2);
5151 key.mv_data = NODEKEY(s2);
5152 csrc->mc_snum = snum--;
5153 csrc->mc_top = snum;
5155 key.mv_size = srcnode->mn_ksize;
5156 key.mv_data = NODEKEY(srcnode);
5159 data.mv_size = NODEDSZ(srcnode);
5160 data.mv_data = NODEDATA(srcnode);
5161 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5162 if (rc != MDB_SUCCESS)
5167 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5168 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);
5170 /* Unlink the src page from parent and add to free list.
5172 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5173 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5175 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5179 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5180 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5181 csrc->mc_db->md_leaf_pages--;
5183 csrc->mc_db->md_branch_pages--;
5185 /* Adjust other cursors pointing to mp */
5186 MDB_cursor *m2, *m3;
5187 MDB_dbi dbi = csrc->mc_dbi;
5188 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5190 if (csrc->mc_flags & C_SUB)
5193 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5194 if (csrc->mc_flags & C_SUB)
5195 m3 = &m2->mc_xcursor->mx_cursor;
5198 if (m3 == csrc) continue;
5199 if (m3->mc_snum < csrc->mc_snum) continue;
5200 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5201 m3->mc_pg[csrc->mc_top] = mp;
5202 m3->mc_ki[csrc->mc_top] += nkeys;
5206 mdb_cursor_pop(csrc);
5208 return mdb_rebalance(csrc);
5211 /** Copy the contents of a cursor.
5212 * @param[in] csrc The cursor to copy from.
5213 * @param[out] cdst The cursor to copy to.
5216 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5220 cdst->mc_txn = csrc->mc_txn;
5221 cdst->mc_dbi = csrc->mc_dbi;
5222 cdst->mc_db = csrc->mc_db;
5223 cdst->mc_dbx = csrc->mc_dbx;
5224 cdst->mc_snum = csrc->mc_snum;
5225 cdst->mc_top = csrc->mc_top;
5226 cdst->mc_flags = csrc->mc_flags;
5228 for (i=0; i<csrc->mc_snum; i++) {
5229 cdst->mc_pg[i] = csrc->mc_pg[i];
5230 cdst->mc_ki[i] = csrc->mc_ki[i];
5234 /** Rebalance the tree after a delete operation.
5235 * @param[in] mc Cursor pointing to the page where rebalancing
5237 * @return 0 on success, non-zero on failure.
5240 mdb_rebalance(MDB_cursor *mc)
5250 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5251 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5252 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5253 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5257 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5260 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5261 DPRINTF("no need to rebalance page %zu, above fill threshold",
5267 if (mc->mc_snum < 2) {
5268 MDB_page *mp = mc->mc_pg[0];
5269 if (NUMKEYS(mp) == 0) {
5270 DPUTS("tree is completely empty");
5271 mc->mc_db->md_root = P_INVALID;
5272 mc->mc_db->md_depth = 0;
5273 mc->mc_db->md_leaf_pages = 0;
5274 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5278 /* Adjust other cursors pointing to mp */
5279 MDB_cursor *m2, *m3;
5280 MDB_dbi dbi = mc->mc_dbi;
5282 if (mc->mc_flags & C_SUB)
5285 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5286 if (m2 == mc) continue;
5287 if (mc->mc_flags & C_SUB)
5288 m3 = &m2->mc_xcursor->mx_cursor;
5291 if (m3->mc_snum < mc->mc_snum) continue;
5292 if (m3->mc_pg[0] == mp) {
5298 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5299 DPUTS("collapsing root page!");
5300 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5301 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5302 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5305 mc->mc_db->md_depth--;
5306 mc->mc_db->md_branch_pages--;
5308 /* Adjust other cursors pointing to mp */
5309 MDB_cursor *m2, *m3;
5310 MDB_dbi dbi = mc->mc_dbi;
5312 if (mc->mc_flags & C_SUB)
5315 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5316 if (m2 == mc) continue;
5317 if (mc->mc_flags & C_SUB)
5318 m3 = &m2->mc_xcursor->mx_cursor;
5321 if (m3->mc_snum < mc->mc_snum) continue;
5322 if (m3->mc_pg[0] == mp) {
5323 m3->mc_pg[0] = mc->mc_pg[0];
5328 DPUTS("root page doesn't need rebalancing");
5332 /* The parent (branch page) must have at least 2 pointers,
5333 * otherwise the tree is invalid.
5335 ptop = mc->mc_top-1;
5336 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5338 /* Leaf page fill factor is below the threshold.
5339 * Try to move keys from left or right neighbor, or
5340 * merge with a neighbor page.
5345 mdb_cursor_copy(mc, &mn);
5346 mn.mc_xcursor = NULL;
5348 if (mc->mc_ki[ptop] == 0) {
5349 /* We're the leftmost leaf in our parent.
5351 DPUTS("reading right neighbor");
5353 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5354 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5356 mn.mc_ki[mn.mc_top] = 0;
5357 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5359 /* There is at least one neighbor to the left.
5361 DPUTS("reading left neighbor");
5363 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5364 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5366 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5367 mc->mc_ki[mc->mc_top] = 0;
5370 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5371 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);
5373 /* If the neighbor page is above threshold and has at least two
5374 * keys, move one key from it.
5376 * Otherwise we should try to merge them.
5378 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5379 return mdb_node_move(&mn, mc);
5380 else { /* FIXME: if (has_enough_room()) */
5381 mc->mc_flags &= ~C_INITIALIZED;
5382 if (mc->mc_ki[ptop] == 0)
5383 return mdb_page_merge(&mn, mc);
5385 return mdb_page_merge(mc, &mn);
5389 /** Complete a delete operation started by #mdb_cursor_del(). */
5391 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5395 /* add overflow pages to free list */
5396 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5400 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5401 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5402 mc->mc_db->md_overflow_pages -= ovpages;
5403 for (i=0; i<ovpages; i++) {
5404 DPRINTF("freed ov page %zu", pg);
5405 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5409 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5410 mc->mc_db->md_entries--;
5411 rc = mdb_rebalance(mc);
5412 if (rc != MDB_SUCCESS)
5413 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5419 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5420 MDB_val *key, MDB_val *data)
5425 MDB_val rdata, *xdata;
5429 assert(key != NULL);
5431 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5433 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5436 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5440 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5444 mdb_cursor_init(&mc, txn, dbi, &mx);
5455 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5457 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5461 /** Split a page and insert a new node.
5462 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5463 * The cursor will be updated to point to the actual page and index where
5464 * the node got inserted after the split.
5465 * @param[in] newkey The key for the newly inserted node.
5466 * @param[in] newdata The data for the newly inserted node.
5467 * @param[in] newpgno The page number, if the new node is a branch node.
5468 * @return 0 on success, non-zero on failure.
5471 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5472 unsigned int nflags)
5475 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0;
5478 unsigned int i, j, split_indx, nkeys, pmax;
5480 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5482 MDB_page *mp, *rp, *pp;
5487 mp = mc->mc_pg[mc->mc_top];
5488 newindx = mc->mc_ki[mc->mc_top];
5490 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5491 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5492 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5494 if (mc->mc_snum < 2) {
5495 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5497 /* shift current top to make room for new parent */
5498 mc->mc_pg[1] = mc->mc_pg[0];
5499 mc->mc_ki[1] = mc->mc_ki[0];
5502 mc->mc_db->md_root = pp->mp_pgno;
5503 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5504 mc->mc_db->md_depth++;
5507 /* Add left (implicit) pointer. */
5508 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5509 /* undo the pre-push */
5510 mc->mc_pg[0] = mc->mc_pg[1];
5511 mc->mc_ki[0] = mc->mc_ki[1];
5512 mc->mc_db->md_root = mp->mp_pgno;
5513 mc->mc_db->md_depth--;
5520 ptop = mc->mc_top-1;
5521 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5524 /* Create a right sibling. */
5525 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5527 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5529 mdb_cursor_copy(mc, &mn);
5530 mn.mc_pg[mn.mc_top] = rp;
5531 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5533 if (nflags & MDB_APPEND) {
5534 mn.mc_ki[mn.mc_top] = 0;
5541 nkeys = NUMKEYS(mp);
5542 split_indx = nkeys / 2 + 1;
5547 unsigned int lsize, rsize, ksize;
5548 /* Move half of the keys to the right sibling */
5550 x = mc->mc_ki[mc->mc_top] - split_indx;
5551 ksize = mc->mc_db->md_pad;
5552 split = LEAF2KEY(mp, split_indx, ksize);
5553 rsize = (nkeys - split_indx) * ksize;
5554 lsize = (nkeys - split_indx) * sizeof(indx_t);
5555 mp->mp_lower -= lsize;
5556 rp->mp_lower += lsize;
5557 mp->mp_upper += rsize - lsize;
5558 rp->mp_upper -= rsize - lsize;
5559 sepkey.mv_size = ksize;
5560 if (newindx == split_indx) {
5561 sepkey.mv_data = newkey->mv_data;
5563 sepkey.mv_data = split;
5566 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5567 memcpy(rp->mp_ptrs, split, rsize);
5568 sepkey.mv_data = rp->mp_ptrs;
5569 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5570 memcpy(ins, newkey->mv_data, ksize);
5571 mp->mp_lower += sizeof(indx_t);
5572 mp->mp_upper -= ksize - sizeof(indx_t);
5575 memcpy(rp->mp_ptrs, split, x * ksize);
5576 ins = LEAF2KEY(rp, x, ksize);
5577 memcpy(ins, newkey->mv_data, ksize);
5578 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5579 rp->mp_lower += sizeof(indx_t);
5580 rp->mp_upper -= ksize - sizeof(indx_t);
5581 mc->mc_ki[mc->mc_top] = x;
5582 mc->mc_pg[mc->mc_top] = rp;
5587 /* For leaf pages, check the split point based on what
5588 * fits where, since otherwise add_node can fail.
5591 unsigned int psize, nsize;
5592 /* Maximum free space in an empty page */
5593 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5594 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5595 if (newindx < split_indx) {
5597 for (i=0; i<split_indx; i++) {
5598 node = NODEPTR(mp, i);
5599 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5600 if (F_ISSET(node->mn_flags, F_BIGDATA))
5601 psize += sizeof(pgno_t);
5603 psize += NODEDSZ(node);
5612 for (i=nkeys-1; i>=split_indx; i--) {
5613 node = NODEPTR(mp, i);
5614 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5615 if (F_ISSET(node->mn_flags, F_BIGDATA))
5616 psize += sizeof(pgno_t);
5618 psize += NODEDSZ(node);
5628 /* First find the separating key between the split pages.
5630 if (newindx == split_indx) {
5631 sepkey.mv_size = newkey->mv_size;
5632 sepkey.mv_data = newkey->mv_data;
5634 node = NODEPTR(mp, split_indx);
5635 sepkey.mv_size = node->mn_ksize;
5636 sepkey.mv_data = NODEKEY(node);
5640 DPRINTF("separator is [%s]", DKEY(&sepkey));
5642 /* Copy separator key to the parent.
5644 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5647 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5649 /* Right page might now have changed parent.
5650 * Check if left page also changed parent.
5652 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5653 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5654 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5655 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5659 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5662 if (rc != MDB_SUCCESS) {
5665 if (nflags & MDB_APPEND) {
5666 mc->mc_pg[mc->mc_top] = rp;
5667 mc->mc_ki[mc->mc_top] = 0;
5668 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
5677 /* Move half of the keys to the right sibling. */
5679 /* grab a page to hold a temporary copy */
5680 copy = mdb_page_malloc(mc);
5684 copy->mp_pgno = mp->mp_pgno;
5685 copy->mp_flags = mp->mp_flags;
5686 copy->mp_lower = PAGEHDRSZ;
5687 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
5688 mc->mc_pg[mc->mc_top] = copy;
5689 for (i = j = 0; i <= nkeys; j++) {
5690 if (i == split_indx) {
5691 /* Insert in right sibling. */
5692 /* Reset insert index for right sibling. */
5693 j = (i == newindx && ins_new);
5694 mc->mc_pg[mc->mc_top] = rp;
5697 if (i == newindx && !ins_new) {
5698 /* Insert the original entry that caused the split. */
5699 rkey.mv_data = newkey->mv_data;
5700 rkey.mv_size = newkey->mv_size;
5709 /* Update page and index for the new key. */
5711 mc->mc_pg[mc->mc_top] = copy;
5712 mc->mc_ki[mc->mc_top] = j;
5713 } else if (i == nkeys) {
5716 node = NODEPTR(mp, i);
5717 rkey.mv_data = NODEKEY(node);
5718 rkey.mv_size = node->mn_ksize;
5720 xdata.mv_data = NODEDATA(node);
5721 xdata.mv_size = NODEDSZ(node);
5724 pgno = NODEPGNO(node);
5725 flags = node->mn_flags;
5730 if (!IS_LEAF(mp) && j == 0) {
5731 /* First branch index doesn't need key data. */
5735 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
5738 nkeys = NUMKEYS(copy);
5739 for (i=0; i<nkeys; i++)
5740 mp->mp_ptrs[i] = copy->mp_ptrs[i];
5741 mp->mp_lower = copy->mp_lower;
5742 mp->mp_upper = copy->mp_upper;
5743 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
5744 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
5746 /* reset back to original page */
5747 if (!newindx || (newindx < split_indx)) {
5748 mc->mc_pg[mc->mc_top] = mp;
5749 if (nflags & MDB_RESERVE) {
5750 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5751 if (!(node->mn_flags & F_BIGDATA))
5752 newdata->mv_data = NODEDATA(node);
5756 /* return tmp page to freelist */
5757 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
5758 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
5759 mc->mc_txn->mt_env->me_dpages = copy;
5762 /* Adjust other cursors pointing to mp */
5763 MDB_cursor *m2, *m3;
5764 MDB_dbi dbi = mc->mc_dbi;
5766 if (mc->mc_flags & C_SUB)
5769 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5770 if (m2 == mc) continue;
5771 if (mc->mc_flags & C_SUB)
5772 m3 = &m2->mc_xcursor->mx_cursor;
5775 if (!(m3->mc_flags & C_INITIALIZED))
5780 for (k=m3->mc_top; k>=0; k--) {
5781 m3->mc_ki[k+1] = m3->mc_ki[k];
5782 m3->mc_pg[k+1] = m3->mc_pg[k];
5784 m3->mc_ki[0] = mc->mc_ki[0];
5785 m3->mc_pg[0] = mc->mc_pg[0];
5789 if (m3->mc_pg[mc->mc_top] == mp) {
5790 if (m3->mc_ki[m3->mc_top] >= split_indx) {
5791 m3->mc_pg[m3->mc_top] = rp;
5792 m3->mc_ki[m3->mc_top] -= split_indx;
5801 mdb_put(MDB_txn *txn, MDB_dbi dbi,
5802 MDB_val *key, MDB_val *data, unsigned int flags)
5807 assert(key != NULL);
5808 assert(data != NULL);
5810 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5813 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5817 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5821 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
5824 mdb_cursor_init(&mc, txn, dbi, &mx);
5825 return mdb_cursor_put(&mc, key, data, flags);
5828 /** Only a subset of the @ref mdb_env flags can be changed
5829 * at runtime. Changing other flags requires closing the environment
5830 * and re-opening it with the new flags.
5832 #define CHANGEABLE (MDB_NOSYNC)
5834 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
5836 if ((flag & CHANGEABLE) != flag)
5839 env->me_flags |= flag;
5841 env->me_flags &= ~flag;
5846 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
5851 *arg = env->me_flags;
5856 mdb_env_get_path(MDB_env *env, const char **arg)
5861 *arg = env->me_path;
5865 /** Common code for #mdb_stat() and #mdb_env_stat().
5866 * @param[in] env the environment to operate in.
5867 * @param[in] db the #MDB_db record containing the stats to return.
5868 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
5869 * @return 0, this function always succeeds.
5872 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
5874 arg->ms_psize = env->me_psize;
5875 arg->ms_depth = db->md_depth;
5876 arg->ms_branch_pages = db->md_branch_pages;
5877 arg->ms_leaf_pages = db->md_leaf_pages;
5878 arg->ms_overflow_pages = db->md_overflow_pages;
5879 arg->ms_entries = db->md_entries;
5884 mdb_env_stat(MDB_env *env, MDB_stat *arg)
5888 if (env == NULL || arg == NULL)
5891 mdb_env_read_meta(env, &toggle);
5893 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
5896 /** Set the default comparison functions for a database.
5897 * Called immediately after a database is opened to set the defaults.
5898 * The user can then override them with #mdb_set_compare() or
5899 * #mdb_set_dupsort().
5900 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
5901 * @param[in] dbi A database handle returned by #mdb_open()
5904 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
5906 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
5907 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memnr;
5908 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
5909 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_cint;
5911 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memn;
5913 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5914 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
5915 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
5916 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_int;
5918 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_cint;
5919 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
5920 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memnr;
5922 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memn;
5925 txn->mt_dbxs[dbi].md_dcmp = NULL;
5929 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
5934 int rc, dbflag, exact;
5937 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
5938 mdb_default_cmp(txn, FREE_DBI);
5944 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
5945 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
5946 mdb_default_cmp(txn, MAIN_DBI);
5950 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
5951 mdb_default_cmp(txn, MAIN_DBI);
5954 /* Is the DB already open? */
5956 for (i=2; i<txn->mt_numdbs; i++) {
5957 if (len == txn->mt_dbxs[i].md_name.mv_size &&
5958 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
5964 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
5967 /* Find the DB info */
5971 key.mv_data = (void *)name;
5972 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
5973 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
5974 if (rc == MDB_SUCCESS) {
5975 /* make sure this is actually a DB */
5976 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
5977 if (!(node->mn_flags & F_SUBDATA))
5979 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
5980 /* Create if requested */
5982 data.mv_size = sizeof(MDB_db);
5983 data.mv_data = &dummy;
5984 memset(&dummy, 0, sizeof(dummy));
5985 dummy.md_root = P_INVALID;
5986 dummy.md_flags = flags & 0xffff;
5987 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
5991 /* OK, got info, add to table */
5992 if (rc == MDB_SUCCESS) {
5993 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
5994 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
5995 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
5996 txn->mt_dbflags[txn->mt_numdbs] = dbflag;
5997 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
5998 *dbi = txn->mt_numdbs;
5999 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6000 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6001 mdb_default_cmp(txn, txn->mt_numdbs);
6008 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6010 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6013 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6016 void mdb_close(MDB_env *env, MDB_dbi dbi)
6019 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6021 ptr = env->me_dbxs[dbi].md_name.mv_data;
6022 env->me_dbxs[dbi].md_name.mv_data = NULL;
6023 env->me_dbxs[dbi].md_name.mv_size = 0;
6027 /** Add all the DB's pages to the free list.
6028 * @param[in] mc Cursor on the DB to free.
6029 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6030 * @return 0 on success, non-zero on failure.
6033 mdb_drop0(MDB_cursor *mc, int subs)
6037 rc = mdb_page_search(mc, NULL, 0);
6038 if (rc == MDB_SUCCESS) {
6043 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6044 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6047 mdb_cursor_copy(mc, &mx);
6048 while (mc->mc_snum > 0) {
6049 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6050 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6051 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6052 if (ni->mn_flags & F_SUBDATA) {
6053 mdb_xcursor_init1(mc, ni);
6054 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6060 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6062 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6065 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6070 rc = mdb_cursor_sibling(mc, 1);
6072 /* no more siblings, go back to beginning
6073 * of previous level. (stack was already popped
6074 * by mdb_cursor_sibling)
6076 for (i=1; i<mc->mc_top; i++)
6077 mc->mc_pg[i] = mx.mc_pg[i];
6081 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6082 mc->mc_db->md_root);
6087 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6092 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6095 rc = mdb_cursor_open(txn, dbi, &mc);
6099 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6103 /* Can't delete the main DB */
6104 if (del && dbi > MAIN_DBI) {
6105 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6107 mdb_close(txn->mt_env, dbi);
6109 txn->mt_dbflags[dbi] |= DB_DIRTY;
6110 txn->mt_dbs[dbi].md_depth = 0;
6111 txn->mt_dbs[dbi].md_branch_pages = 0;
6112 txn->mt_dbs[dbi].md_leaf_pages = 0;
6113 txn->mt_dbs[dbi].md_overflow_pages = 0;
6114 txn->mt_dbs[dbi].md_entries = 0;
6115 txn->mt_dbs[dbi].md_root = P_INVALID;
6118 mdb_cursor_close(mc);
6122 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6124 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6127 txn->mt_dbxs[dbi].md_cmp = cmp;
6131 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6133 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6136 txn->mt_dbxs[dbi].md_dcmp = cmp;
6140 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6142 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6145 txn->mt_dbxs[dbi].md_rel = rel;
6149 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6151 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6154 txn->mt_dbxs[dbi].md_relctx = ctx;