2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define pthread_t DWORD
144 #define pthread_mutex_t HANDLE
145 #define pthread_key_t DWORD
146 #define pthread_self() GetCurrentThreadId()
147 #define pthread_key_create(x,y) \
148 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
149 #define pthread_key_delete(x) TlsFree(x)
150 #define pthread_getspecific(x) TlsGetValue(x)
151 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
152 #define pthread_mutex_unlock(x) ReleaseMutex(x)
153 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
154 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
158 #define getpid() GetCurrentProcessId()
159 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
160 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
161 #define ErrCode() GetLastError()
162 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
163 #define close(fd) (CloseHandle(fd) ? 0 : -1)
164 #define munmap(ptr,len) UnmapViewOfFile(ptr)
167 #ifdef MDB_USE_POSIX_SEM
169 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
170 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
171 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
172 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
175 mdb_sem_wait(sem_t *sem)
178 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
183 /** Lock the reader mutex.
185 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
186 /** Unlock the reader mutex.
188 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
190 /** Lock the writer mutex.
191 * Only a single write transaction is allowed at a time. Other writers
192 * will block waiting for this mutex.
194 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
195 /** Unlock the writer mutex.
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
198 #endif /* MDB_USE_POSIX_SEM */
200 /** Get the error code for the last failed system function.
202 #define ErrCode() errno
204 /** An abstraction for a file handle.
205 * On POSIX systems file handles are small integers. On Windows
206 * they're opaque pointers.
210 /** A value for an invalid file handle.
211 * Mainly used to initialize file variables and signify that they are
214 #define INVALID_HANDLE_VALUE (-1)
216 /** Get the size of a memory page for the system.
217 * This is the basic size that the platform's memory manager uses, and is
218 * fundamental to the use of memory-mapped files.
220 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
223 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
226 #define MNAME_LEN (sizeof(pthread_mutex_t))
232 /** A flag for opening a file and requesting synchronous data writes.
233 * This is only used when writing a meta page. It's not strictly needed;
234 * we could just do a normal write and then immediately perform a flush.
235 * But if this flag is available it saves us an extra system call.
237 * @note If O_DSYNC is undefined but exists in /usr/include,
238 * preferably set some compiler flag to get the definition.
239 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
242 # define MDB_DSYNC O_DSYNC
246 /** Function for flushing the data of a file. Define this to fsync
247 * if fdatasync() is not supported.
249 #ifndef MDB_FDATASYNC
250 # define MDB_FDATASYNC fdatasync
254 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
265 /** A page number in the database.
266 * Note that 64 bit page numbers are overkill, since pages themselves
267 * already represent 12-13 bits of addressable memory, and the OS will
268 * always limit applications to a maximum of 63 bits of address space.
270 * @note In the #MDB_node structure, we only store 48 bits of this value,
271 * which thus limits us to only 60 bits of addressable data.
273 typedef MDB_ID pgno_t;
275 /** A transaction ID.
276 * See struct MDB_txn.mt_txnid for details.
278 typedef MDB_ID txnid_t;
280 /** @defgroup debug Debug Macros
284 /** Enable debug output.
285 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
286 * read from and written to the database (used for free space management).
291 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
294 # define DPRINTF (void) /* Vararg macros may be unsupported */
296 static int mdb_debug;
297 static txnid_t mdb_debug_start;
299 /** Print a debug message with printf formatting. */
300 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
301 ((void) ((mdb_debug) && \
302 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
304 # define DPRINTF(fmt, ...) ((void) 0)
305 # define MDB_DEBUG_SKIP
307 /** Print a debug string.
308 * The string is printed literally, with no format processing.
310 #define DPUTS(arg) DPRINTF("%s", arg)
313 /** A default memory page size.
314 * The actual size is platform-dependent, but we use this for
315 * boot-strapping. We probably should not be using this any more.
316 * The #GET_PAGESIZE() macro is used to get the actual size.
318 * Note that we don't currently support Huge pages. On Linux,
319 * regular data files cannot use Huge pages, and in general
320 * Huge pages aren't actually pageable. We rely on the OS
321 * demand-pager to read our data and page it out when memory
322 * pressure from other processes is high. So until OSs have
323 * actual paging support for Huge pages, they're not viable.
325 #define MDB_PAGESIZE 4096
327 /** The minimum number of keys required in a database page.
328 * Setting this to a larger value will place a smaller bound on the
329 * maximum size of a data item. Data items larger than this size will
330 * be pushed into overflow pages instead of being stored directly in
331 * the B-tree node. This value used to default to 4. With a page size
332 * of 4096 bytes that meant that any item larger than 1024 bytes would
333 * go into an overflow page. That also meant that on average 2-3KB of
334 * each overflow page was wasted space. The value cannot be lower than
335 * 2 because then there would no longer be a tree structure. With this
336 * value, items larger than 2KB will go into overflow pages, and on
337 * average only 1KB will be wasted.
339 #define MDB_MINKEYS 2
341 /** A stamp that identifies a file as an MDB file.
342 * There's nothing special about this value other than that it is easily
343 * recognizable, and it will reflect any byte order mismatches.
345 #define MDB_MAGIC 0xBEEFC0DE
347 /** The version number for a database's file format. */
348 #define MDB_VERSION 1
350 /** @brief The maximum size of a key in the database.
352 * The library rejects bigger keys, and cannot deal with records
353 * with bigger keys stored by a library with bigger max keysize.
355 * We require that keys all fit onto a regular page. This limit
356 * could be raised a bit further if needed; to something just
357 * under #MDB_PAGESIZE / #MDB_MINKEYS.
359 * Note that data items in an #MDB_DUPSORT database are actually keys
360 * of a subDB, so they're also limited to this size.
362 #ifndef MDB_MAXKEYSIZE
363 #define MDB_MAXKEYSIZE 511
366 /** @brief The maximum size of a data item.
368 * We only store a 32 bit value for node sizes.
370 #define MAXDATASIZE 0xffffffffUL
375 * This is used for printing a hex dump of a key's contents.
377 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
378 /** Display a key in hex.
380 * Invoke a function to display a key in hex.
382 #define DKEY(x) mdb_dkey(x, kbuf)
384 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
388 /** An invalid page number.
389 * Mainly used to denote an empty tree.
391 #define P_INVALID (~(pgno_t)0)
393 /** Test if the flags \b f are set in a flag word \b w. */
394 #define F_ISSET(w, f) (((w) & (f)) == (f))
396 /** Used for offsets within a single page.
397 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
400 typedef uint16_t indx_t;
402 /** Default size of memory map.
403 * This is certainly too small for any actual applications. Apps should always set
404 * the size explicitly using #mdb_env_set_mapsize().
406 #define DEFAULT_MAPSIZE 1048576
408 /** @defgroup readers Reader Lock Table
409 * Readers don't acquire any locks for their data access. Instead, they
410 * simply record their transaction ID in the reader table. The reader
411 * mutex is needed just to find an empty slot in the reader table. The
412 * slot's address is saved in thread-specific data so that subsequent read
413 * transactions started by the same thread need no further locking to proceed.
415 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
417 * No reader table is used if the database is on a read-only filesystem.
419 * Since the database uses multi-version concurrency control, readers don't
420 * actually need any locking. This table is used to keep track of which
421 * readers are using data from which old transactions, so that we'll know
422 * when a particular old transaction is no longer in use. Old transactions
423 * that have discarded any data pages can then have those pages reclaimed
424 * for use by a later write transaction.
426 * The lock table is constructed such that reader slots are aligned with the
427 * processor's cache line size. Any slot is only ever used by one thread.
428 * This alignment guarantees that there will be no contention or cache
429 * thrashing as threads update their own slot info, and also eliminates
430 * any need for locking when accessing a slot.
432 * A writer thread will scan every slot in the table to determine the oldest
433 * outstanding reader transaction. Any freed pages older than this will be
434 * reclaimed by the writer. The writer doesn't use any locks when scanning
435 * this table. This means that there's no guarantee that the writer will
436 * see the most up-to-date reader info, but that's not required for correct
437 * operation - all we need is to know the upper bound on the oldest reader,
438 * we don't care at all about the newest reader. So the only consequence of
439 * reading stale information here is that old pages might hang around a
440 * while longer before being reclaimed. That's actually good anyway, because
441 * the longer we delay reclaiming old pages, the more likely it is that a
442 * string of contiguous pages can be found after coalescing old pages from
443 * many old transactions together.
446 /** Number of slots in the reader table.
447 * This value was chosen somewhat arbitrarily. 126 readers plus a
448 * couple mutexes fit exactly into 8KB on my development machine.
449 * Applications should set the table size using #mdb_env_set_maxreaders().
451 #define DEFAULT_READERS 126
453 /** The size of a CPU cache line in bytes. We want our lock structures
454 * aligned to this size to avoid false cache line sharing in the
456 * This value works for most CPUs. For Itanium this should be 128.
462 /** The information we store in a single slot of the reader table.
463 * In addition to a transaction ID, we also record the process and
464 * thread ID that owns a slot, so that we can detect stale information,
465 * e.g. threads or processes that went away without cleaning up.
466 * @note We currently don't check for stale records. We simply re-init
467 * the table when we know that we're the only process opening the
470 typedef struct MDB_rxbody {
471 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
472 * Multiple readers that start at the same time will probably have the
473 * same ID here. Again, it's not important to exclude them from
474 * anything; all we need to know is which version of the DB they
475 * started from so we can avoid overwriting any data used in that
476 * particular version.
479 /** The process ID of the process owning this reader txn. */
481 /** The thread ID of the thread owning this txn. */
485 /** The actual reader record, with cacheline padding. */
486 typedef struct MDB_reader {
489 /** shorthand for mrb_txnid */
490 #define mr_txnid mru.mrx.mrb_txnid
491 #define mr_pid mru.mrx.mrb_pid
492 #define mr_tid mru.mrx.mrb_tid
493 /** cache line alignment */
494 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
498 /** The header for the reader table.
499 * The table resides in a memory-mapped file. (This is a different file
500 * than is used for the main database.)
502 * For POSIX the actual mutexes reside in the shared memory of this
503 * mapped file. On Windows, mutexes are named objects allocated by the
504 * kernel; we store the mutex names in this mapped file so that other
505 * processes can grab them. This same approach is also used on
506 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
507 * process-shared POSIX mutexes. For these cases where a named object
508 * is used, the object name is derived from a 64 bit FNV hash of the
509 * environment pathname. As such, naming collisions are extremely
510 * unlikely. If a collision occurs, the results are unpredictable.
512 typedef struct MDB_txbody {
513 /** Stamp identifying this as an MDB file. It must be set
516 /** Version number of this lock file. Must be set to #MDB_VERSION. */
517 uint32_t mtb_version;
518 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
519 char mtb_rmname[MNAME_LEN];
521 /** Mutex protecting access to this table.
522 * This is the reader lock that #LOCK_MUTEX_R acquires.
524 pthread_mutex_t mtb_mutex;
526 /** The ID of the last transaction committed to the database.
527 * This is recorded here only for convenience; the value can always
528 * be determined by reading the main database meta pages.
531 /** The number of slots that have been used in the reader table.
532 * This always records the maximum count, it is not decremented
533 * when readers release their slots.
535 unsigned mtb_numreaders;
538 /** The actual reader table definition. */
539 typedef struct MDB_txninfo {
542 #define mti_magic mt1.mtb.mtb_magic
543 #define mti_version mt1.mtb.mtb_version
544 #define mti_mutex mt1.mtb.mtb_mutex
545 #define mti_rmname mt1.mtb.mtb_rmname
546 #define mti_txnid mt1.mtb.mtb_txnid
547 #define mti_numreaders mt1.mtb.mtb_numreaders
548 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
551 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
552 char mt2_wmname[MNAME_LEN];
553 #define mti_wmname mt2.mt2_wmname
555 pthread_mutex_t mt2_wmutex;
556 #define mti_wmutex mt2.mt2_wmutex
558 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
560 MDB_reader mti_readers[1];
564 /** Common header for all page types.
565 * Overflow records occupy a number of contiguous pages with no
566 * headers on any page after the first.
568 typedef struct MDB_page {
569 #define mp_pgno mp_p.p_pgno
570 #define mp_next mp_p.p_next
572 pgno_t p_pgno; /**< page number */
573 void * p_next; /**< for in-memory list of freed structs */
576 /** @defgroup mdb_page Page Flags
578 * Flags for the page headers.
581 #define P_BRANCH 0x01 /**< branch page */
582 #define P_LEAF 0x02 /**< leaf page */
583 #define P_OVERFLOW 0x04 /**< overflow page */
584 #define P_META 0x08 /**< meta page */
585 #define P_DIRTY 0x10 /**< dirty page */
586 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
587 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
589 uint16_t mp_flags; /**< @ref mdb_page */
590 #define mp_lower mp_pb.pb.pb_lower
591 #define mp_upper mp_pb.pb.pb_upper
592 #define mp_pages mp_pb.pb_pages
595 indx_t pb_lower; /**< lower bound of free space */
596 indx_t pb_upper; /**< upper bound of free space */
598 uint32_t pb_pages; /**< number of overflow pages */
600 indx_t mp_ptrs[1]; /**< dynamic size */
603 /** Size of the page header, excluding dynamic data at the end */
604 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
606 /** Address of first usable data byte in a page, after the header */
607 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
609 /** Number of nodes on a page */
610 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
612 /** The amount of space remaining in the page */
613 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
615 /** The percentage of space used in the page, in tenths of a percent. */
616 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
617 ((env)->me_psize - PAGEHDRSZ))
618 /** The minimum page fill factor, in tenths of a percent.
619 * Pages emptier than this are candidates for merging.
621 #define FILL_THRESHOLD 250
623 /** Test if a page is a leaf page */
624 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
625 /** Test if a page is a LEAF2 page */
626 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
627 /** Test if a page is a branch page */
628 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
629 /** Test if a page is an overflow page */
630 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
631 /** Test if a page is a sub page */
632 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
634 /** The number of overflow pages needed to store the given size. */
635 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
637 /** Header for a single key/data pair within a page.
638 * We guarantee 2-byte alignment for nodes.
640 typedef struct MDB_node {
641 /** lo and hi are used for data size on leaf nodes and for
642 * child pgno on branch nodes. On 64 bit platforms, flags
643 * is also used for pgno. (Branch nodes have no flags).
644 * They are in host byte order in case that lets some
645 * accesses be optimized into a 32-bit word access.
647 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
648 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
649 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
650 /** @defgroup mdb_node Node Flags
652 * Flags for node headers.
655 #define F_BIGDATA 0x01 /**< data put on overflow page */
656 #define F_SUBDATA 0x02 /**< data is a sub-database */
657 #define F_DUPDATA 0x04 /**< data has duplicates */
659 /** valid flags for #mdb_node_add() */
660 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
663 unsigned short mn_flags; /**< @ref mdb_node */
664 unsigned short mn_ksize; /**< key size */
665 char mn_data[1]; /**< key and data are appended here */
668 /** Size of the node header, excluding dynamic data at the end */
669 #define NODESIZE offsetof(MDB_node, mn_data)
671 /** Bit position of top word in page number, for shifting mn_flags */
672 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
674 /** Size of a node in a branch page with a given key.
675 * This is just the node header plus the key, there is no data.
677 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
679 /** Size of a node in a leaf page with a given key and data.
680 * This is node header plus key plus data size.
682 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
684 /** Address of node \b i in page \b p */
685 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
687 /** Address of the key for the node */
688 #define NODEKEY(node) (void *)((node)->mn_data)
690 /** Address of the data for a node */
691 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
693 /** Get the page number pointed to by a branch node */
694 #define NODEPGNO(node) \
695 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
696 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
697 /** Set the page number in a branch node */
698 #define SETPGNO(node,pgno) do { \
699 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
700 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
702 /** Get the size of the data in a leaf node */
703 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
704 /** Set the size of the data for a leaf node */
705 #define SETDSZ(node,size) do { \
706 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
707 /** The size of a key in a node */
708 #define NODEKSZ(node) ((node)->mn_ksize)
710 /** Copy a page number from src to dst */
712 #define COPY_PGNO(dst,src) dst = src
714 #if SIZE_MAX > 4294967295UL
715 #define COPY_PGNO(dst,src) do { \
716 unsigned short *s, *d; \
717 s = (unsigned short *)&(src); \
718 d = (unsigned short *)&(dst); \
725 #define COPY_PGNO(dst,src) do { \
726 unsigned short *s, *d; \
727 s = (unsigned short *)&(src); \
728 d = (unsigned short *)&(dst); \
734 /** The address of a key in a LEAF2 page.
735 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
736 * There are no node headers, keys are stored contiguously.
738 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
740 /** Set the \b node's key into \b key, if requested. */
741 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
742 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
744 /** Information about a single database in the environment. */
745 typedef struct MDB_db {
746 uint32_t md_pad; /**< also ksize for LEAF2 pages */
747 uint16_t md_flags; /**< @ref mdb_dbi_open */
748 uint16_t md_depth; /**< depth of this tree */
749 pgno_t md_branch_pages; /**< number of internal pages */
750 pgno_t md_leaf_pages; /**< number of leaf pages */
751 pgno_t md_overflow_pages; /**< number of overflow pages */
752 size_t md_entries; /**< number of data items */
753 pgno_t md_root; /**< the root page of this tree */
756 /** mdb_dbi_open flags */
757 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
758 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
759 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
760 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
762 /** Handle for the DB used to track free pages. */
764 /** Handle for the default DB. */
767 /** Meta page content. */
768 typedef struct MDB_meta {
769 /** Stamp identifying this as an MDB file. It must be set
772 /** Version number of this lock file. Must be set to #MDB_VERSION. */
774 void *mm_address; /**< address for fixed mapping */
775 size_t mm_mapsize; /**< size of mmap region */
776 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
777 /** The size of pages used in this DB */
778 #define mm_psize mm_dbs[0].md_pad
779 /** Any persistent environment flags. @ref mdb_env */
780 #define mm_flags mm_dbs[0].md_flags
781 pgno_t mm_last_pg; /**< last used page in file */
782 txnid_t mm_txnid; /**< txnid that committed this page */
785 /** Buffer for a stack-allocated dirty page.
786 * The members define size and alignment, and silence type
787 * aliasing warnings. They are not used directly; that could
788 * mean incorrectly using several union members in parallel.
790 typedef union MDB_pagebuf {
791 char mb_raw[MDB_PAGESIZE];
794 char mm_pad[PAGEHDRSZ];
799 /** Auxiliary DB info.
800 * The information here is mostly static/read-only. There is
801 * only a single copy of this record in the environment.
803 typedef struct MDB_dbx {
804 MDB_val md_name; /**< name of the database */
805 MDB_cmp_func *md_cmp; /**< function for comparing keys */
806 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
807 MDB_rel_func *md_rel; /**< user relocate function */
808 void *md_relctx; /**< user-provided context for md_rel */
811 /** A database transaction.
812 * Every operation requires a transaction handle.
815 MDB_txn *mt_parent; /**< parent of a nested txn */
816 MDB_txn *mt_child; /**< nested txn under this txn */
817 pgno_t mt_next_pgno; /**< next unallocated page */
818 /** The ID of this transaction. IDs are integers incrementing from 1.
819 * Only committed write transactions increment the ID. If a transaction
820 * aborts, the ID may be re-used by the next writer.
823 MDB_env *mt_env; /**< the DB environment */
824 /** The list of pages that became unused during this transaction.
828 MDB_ID2L dirty_list; /**< for write txns: modified pages */
829 MDB_reader *reader; /**< this thread's reader table slot or NULL */
831 /** Array of records for each DB known in the environment. */
833 /** Array of MDB_db records for each known DB */
835 /** @defgroup mt_dbflag Transaction DB Flags
839 #define DB_DIRTY 0x01 /**< DB was written in this txn */
840 #define DB_STALE 0x02 /**< DB record is older than txnID */
841 #define DB_NEW 0x04 /**< DB handle opened in this txn */
842 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
844 /** In write txns, array of cursors for each DB */
845 MDB_cursor **mt_cursors;
846 /** Array of flags for each DB */
847 unsigned char *mt_dbflags;
848 /** Number of DB records in use. This number only ever increments;
849 * we don't decrement it when individual DB handles are closed.
853 /** @defgroup mdb_txn Transaction Flags
857 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
858 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
859 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
861 unsigned int mt_flags; /**< @ref mdb_txn */
862 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
863 unsigned int mt_dirty_room;
864 /** Tracks which of the two meta pages was used at the start
865 * of this transaction.
867 unsigned int mt_toggle;
870 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
871 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
872 * raise this on a 64 bit machine.
874 #define CURSOR_STACK 32
878 /** Cursors are used for all DB operations */
880 /** Next cursor on this DB in this txn */
882 /** Original cursor if this is a shadow */
884 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
885 struct MDB_xcursor *mc_xcursor;
886 /** The transaction that owns this cursor */
888 /** The database handle this cursor operates on */
890 /** The database record for this cursor */
892 /** The database auxiliary record for this cursor */
894 /** The @ref mt_dbflag for this database */
895 unsigned char *mc_dbflag;
896 unsigned short mc_snum; /**< number of pushed pages */
897 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
898 /** @defgroup mdb_cursor Cursor Flags
900 * Cursor state flags.
903 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
904 #define C_EOF 0x02 /**< No more data */
905 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
906 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
907 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
908 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
909 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
911 unsigned int mc_flags; /**< @ref mdb_cursor */
912 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
913 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
916 /** Context for sorted-dup records.
917 * We could have gone to a fully recursive design, with arbitrarily
918 * deep nesting of sub-databases. But for now we only handle these
919 * levels - main DB, optional sub-DB, sorted-duplicate DB.
921 typedef struct MDB_xcursor {
922 /** A sub-cursor for traversing the Dup DB */
923 MDB_cursor mx_cursor;
924 /** The database record for this Dup DB */
926 /** The auxiliary DB record for this Dup DB */
928 /** The @ref mt_dbflag for this Dup DB */
929 unsigned char mx_dbflag;
932 /** State of FreeDB old pages, stored in the MDB_env */
933 typedef struct MDB_pgstate {
934 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
935 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
938 /** The database environment. */
940 HANDLE me_fd; /**< The main data file */
941 HANDLE me_lfd; /**< The lock file */
942 HANDLE me_mfd; /**< just for writing the meta pages */
943 /** Failed to update the meta page. Probably an I/O error. */
944 #define MDB_FATAL_ERROR 0x80000000U
945 /** Some fields are initialized. */
946 #define MDB_ENV_ACTIVE 0x20000000U
947 /** me_txkey is set */
948 #define MDB_ENV_TXKEY 0x10000000U
949 uint32_t me_flags; /**< @ref mdb_env */
950 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
951 unsigned int me_maxreaders; /**< size of the reader table */
952 unsigned int me_numreaders; /**< max numreaders set by this env */
953 MDB_dbi me_numdbs; /**< number of DBs opened */
954 MDB_dbi me_maxdbs; /**< size of the DB table */
955 pid_t me_pid; /**< process ID of this env */
956 char *me_path; /**< path to the DB files */
957 char *me_map; /**< the memory map of the data file */
958 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
959 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
960 MDB_txn *me_txn; /**< current write transaction */
961 size_t me_mapsize; /**< size of the data memory map */
962 off_t me_size; /**< current file size */
963 pgno_t me_maxpg; /**< me_mapsize / me_psize */
964 MDB_dbx *me_dbxs; /**< array of static DB info */
965 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
966 pthread_key_t me_txkey; /**< thread-key for readers */
967 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
968 # define me_pglast me_pgstate.mf_pglast
969 # define me_pghead me_pgstate.mf_pghead
970 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
971 /** IDL of pages that became unused in a write txn */
973 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
974 MDB_ID2L me_dirty_list;
975 /** Max number of freelist items that can fit in a single overflow page */
977 /** Max size of a node on a page */
978 unsigned int me_nodemax;
980 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
982 #elif defined(MDB_USE_POSIX_SEM)
983 sem_t *me_rmutex; /* Shared mutexes are not supported */
988 /** Nested transaction */
989 typedef struct MDB_ntxn {
990 MDB_txn mnt_txn; /* the transaction */
991 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
994 /** max number of pages to commit in one writev() call */
995 #define MDB_COMMIT_PAGES 64
996 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
997 #undef MDB_COMMIT_PAGES
998 #define MDB_COMMIT_PAGES IOV_MAX
1001 /* max bytes to write in one call */
1002 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1004 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1005 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1006 static int mdb_page_touch(MDB_cursor *mc);
1008 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1009 static int mdb_page_search_root(MDB_cursor *mc,
1010 MDB_val *key, int modify);
1011 #define MDB_PS_MODIFY 1
1012 #define MDB_PS_ROOTONLY 2
1013 static int mdb_page_search(MDB_cursor *mc,
1014 MDB_val *key, int flags);
1015 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1017 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1018 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1019 pgno_t newpgno, unsigned int nflags);
1021 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1022 static int mdb_env_pick_meta(const MDB_env *env);
1023 static int mdb_env_write_meta(MDB_txn *txn);
1024 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1025 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1027 static void mdb_env_close0(MDB_env *env, int excl);
1029 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1030 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1031 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1032 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1033 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1034 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1035 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1036 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1037 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1039 static int mdb_rebalance(MDB_cursor *mc);
1040 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1042 static void mdb_cursor_pop(MDB_cursor *mc);
1043 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1045 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1046 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1047 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1048 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1049 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1051 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1052 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1054 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1055 static void mdb_xcursor_init0(MDB_cursor *mc);
1056 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1058 static int mdb_drop0(MDB_cursor *mc, int subs);
1059 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1062 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1066 static SECURITY_DESCRIPTOR mdb_null_sd;
1067 static SECURITY_ATTRIBUTES mdb_all_sa;
1068 static int mdb_sec_inited;
1071 /** Return the library version info. */
1073 mdb_version(int *major, int *minor, int *patch)
1075 if (major) *major = MDB_VERSION_MAJOR;
1076 if (minor) *minor = MDB_VERSION_MINOR;
1077 if (patch) *patch = MDB_VERSION_PATCH;
1078 return MDB_VERSION_STRING;
1081 /** Table of descriptions for MDB @ref errors */
1082 static char *const mdb_errstr[] = {
1083 "MDB_KEYEXIST: Key/data pair already exists",
1084 "MDB_NOTFOUND: No matching key/data pair found",
1085 "MDB_PAGE_NOTFOUND: Requested page not found",
1086 "MDB_CORRUPTED: Located page was wrong type",
1087 "MDB_PANIC: Update of meta page failed",
1088 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1089 "MDB_INVALID: File is not an MDB file",
1090 "MDB_MAP_FULL: Environment mapsize limit reached",
1091 "MDB_DBS_FULL: Environment maxdbs limit reached",
1092 "MDB_READERS_FULL: Environment maxreaders limit reached",
1093 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1094 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1095 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1096 "MDB_PAGE_FULL: Internal error - page has no more space",
1097 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1098 "MDB_INCOMPATIBLE: Database flags changed or would change",
1099 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1103 mdb_strerror(int err)
1107 return ("Successful return: 0");
1109 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1110 i = err - MDB_KEYEXIST;
1111 return mdb_errstr[i];
1114 return strerror(err);
1118 /** Display a key in hexadecimal and return the address of the result.
1119 * @param[in] key the key to display
1120 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1121 * @return The key in hexadecimal form.
1124 mdb_dkey(MDB_val *key, char *buf)
1127 unsigned char *c = key->mv_data;
1133 if (key->mv_size > MDB_MAXKEYSIZE)
1134 return "MDB_MAXKEYSIZE";
1135 /* may want to make this a dynamic check: if the key is mostly
1136 * printable characters, print it as-is instead of converting to hex.
1140 for (i=0; i<key->mv_size; i++)
1141 ptr += sprintf(ptr, "%02x", *c++);
1143 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1148 /** Display all the keys in the page. */
1150 mdb_page_list(MDB_page *mp)
1153 unsigned int i, nkeys, nsize;
1157 nkeys = NUMKEYS(mp);
1158 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1159 for (i=0; i<nkeys; i++) {
1160 node = NODEPTR(mp, i);
1161 key.mv_size = node->mn_ksize;
1162 key.mv_data = node->mn_data;
1163 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1164 if (IS_BRANCH(mp)) {
1165 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1168 if (F_ISSET(node->mn_flags, F_BIGDATA))
1169 nsize += sizeof(pgno_t);
1171 nsize += NODEDSZ(node);
1172 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1178 mdb_cursor_chk(MDB_cursor *mc)
1184 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1185 for (i=0; i<mc->mc_top; i++) {
1187 node = NODEPTR(mp, mc->mc_ki[i]);
1188 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1191 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1197 /** Count all the pages in each DB and in the freelist
1198 * and make sure it matches the actual number of pages
1201 static void mdb_audit(MDB_txn *txn)
1205 MDB_ID freecount, count;
1210 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1211 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1212 freecount += *(MDB_ID *)data.mv_data;
1215 for (i = 0; i<txn->mt_numdbs; i++) {
1217 mdb_cursor_init(&mc, txn, i, &mx);
1218 if (txn->mt_dbs[i].md_root == P_INVALID)
1220 count += txn->mt_dbs[i].md_branch_pages +
1221 txn->mt_dbs[i].md_leaf_pages +
1222 txn->mt_dbs[i].md_overflow_pages;
1223 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1224 mdb_page_search(&mc, NULL, 0);
1228 mp = mc.mc_pg[mc.mc_top];
1229 for (j=0; j<NUMKEYS(mp); j++) {
1230 MDB_node *leaf = NODEPTR(mp, j);
1231 if (leaf->mn_flags & F_SUBDATA) {
1233 memcpy(&db, NODEDATA(leaf), sizeof(db));
1234 count += db.md_branch_pages + db.md_leaf_pages +
1235 db.md_overflow_pages;
1239 while (mdb_cursor_sibling(&mc, 1) == 0);
1242 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1243 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1244 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1250 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1252 return txn->mt_dbxs[dbi].md_cmp(a, b);
1256 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1258 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1261 /** Allocate a page.
1262 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1265 mdb_page_malloc(MDB_txn *txn, unsigned num)
1267 MDB_env *env = txn->mt_env;
1268 MDB_page *ret = env->me_dpages;
1269 size_t sz = env->me_psize;
1272 VGMEMP_ALLOC(env, ret, sz);
1273 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1274 env->me_dpages = ret->mp_next;
1280 if ((ret = malloc(sz)) != NULL) {
1281 VGMEMP_ALLOC(env, ret, sz);
1286 /** Free a single page.
1287 * Saves single pages to a list, for future reuse.
1288 * (This is not used for multi-page overflow pages.)
1291 mdb_page_free(MDB_env *env, MDB_page *mp)
1293 mp->mp_next = env->me_dpages;
1294 VGMEMP_FREE(env, mp);
1295 env->me_dpages = mp;
1298 /* Free a dirty page */
1300 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1302 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1303 mdb_page_free(env, dp);
1305 /* large pages just get freed directly */
1306 VGMEMP_FREE(env, dp);
1311 /* Return all dirty pages to dpage list */
1313 mdb_dlist_free(MDB_txn *txn)
1315 MDB_env *env = txn->mt_env;
1316 MDB_ID2L dl = txn->mt_u.dirty_list;
1317 unsigned i, n = dl[0].mid;
1319 for (i = 1; i <= n; i++) {
1320 mdb_dpage_free(env, dl[i].mptr);
1325 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1327 mdb_find_oldest(MDB_txn *txn)
1330 txnid_t mr, oldest = txn->mt_txnid - 1;
1331 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1332 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1342 /** Allocate pages for writing.
1343 * If there are free pages available from older transactions, they
1344 * will be re-used first. Otherwise a new page will be allocated.
1345 * @param[in] mc cursor A cursor handle identifying the transaction and
1346 * database for which we are allocating.
1347 * @param[in] num the number of pages to allocate.
1348 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1349 * will always be satisfied by a single contiguous chunk of memory.
1350 * @return 0 on success, non-zero on failure.
1353 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1355 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1356 /* Get at most <Max_retries> more freeDB records once me_pghead
1357 * has enough pages. If not enough, use new pages from the map.
1358 * If <Paranoid> and mc is updating the freeDB, only get new
1359 * records if me_pghead is empty. Then the freelist cannot play
1360 * catch-up with itself by growing while trying to save it.
1362 enum { Paranoid = 1, Max_retries = 500 };
1364 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1366 int rc, n2 = num-1, retry = Max_retries;
1367 MDB_txn *txn = mc->mc_txn;
1368 MDB_env *env = txn->mt_env;
1369 pgno_t pgno, *mop = env->me_pghead;
1370 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1373 txnid_t oldest = 0, last;
1376 int (*insert)(MDB_ID2L, MDB_ID2 *);
1380 /* If our dirty list is already full, we can't do anything */
1381 if (txn->mt_dirty_room == 0)
1382 return MDB_TXN_FULL;
1384 for (op = MDB_FIRST;; op = MDB_NEXT) {
1387 pgno_t *idl, old_id, new_id;
1389 /* Seek a big enough contiguous page range. Prefer
1390 * pages at the tail, just truncating the list.
1392 if (mop_len >= (unsigned)num) {
1396 if (mop[i-n2] == pgno+n2)
1398 } while (--i >= (unsigned)num);
1399 if (Max_retries < INT_MAX && --retry < 0)
1403 if (op == MDB_FIRST) { /* 1st iteration */
1404 /* Prepare to fetch more and coalesce */
1405 oldest = mdb_find_oldest(txn);
1406 last = env->me_pglast;
1407 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1410 key.mv_data = &last; /* will loop up last+1 */
1411 key.mv_size = sizeof(last);
1413 if (Paranoid && mc->mc_dbi == FREE_DBI)
1416 if (Paranoid && retry < 0 && mop_len)
1420 /* Do not fetch more if the record will be too recent */
1423 rc = mdb_cursor_get(&m2, &key, NULL, op);
1425 if (rc == MDB_NOTFOUND)
1429 last = *(txnid_t*)key.mv_data;
1432 np = m2.mc_pg[m2.mc_top];
1433 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1434 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1437 idl = (MDB_ID *) data.mv_data;
1440 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1443 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1445 mop = env->me_pghead;
1447 env->me_pglast = last;
1449 DPRINTF("IDL read txn %zu root %zu num %u",
1450 last, txn->mt_dbs[FREE_DBI].md_root, i);
1452 DPRINTF("IDL %zu", idl[k]);
1454 /* Merge in descending sorted order */
1457 mop[0] = (pgno_t)-1;
1461 for (; old_id < new_id; old_id = mop[--j])
1468 /* Use new pages from the map when nothing suitable in the freeDB */
1470 pgno = txn->mt_next_pgno;
1471 if (pgno + num >= env->me_maxpg) {
1472 DPUTS("DB size maxed out");
1473 return MDB_MAP_FULL;
1477 if (env->me_flags & MDB_WRITEMAP) {
1478 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1479 insert = mdb_mid2l_append;
1481 if (!(np = mdb_page_malloc(txn, num)))
1483 insert = mdb_mid2l_insert;
1486 mop[0] = mop_len -= num;
1487 /* Move any stragglers down */
1488 for (j = i-num; j < mop_len; )
1489 mop[++j] = mop[++i];
1491 txn->mt_next_pgno = pgno + num;
1493 mid.mid = np->mp_pgno = pgno;
1495 insert(txn->mt_u.dirty_list, &mid);
1496 txn->mt_dirty_room--;
1502 /** Copy the used portions of a non-overflow page.
1503 * @param[in] dst page to copy into
1504 * @param[in] src page to copy from
1505 * @param[in] psize size of a page
1508 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1510 enum { Align = sizeof(pgno_t) };
1511 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1513 /* If page isn't full, just copy the used portion. Adjust
1514 * alignment so memcpy may copy words instead of bytes.
1516 if ((unused &= -Align) && !IS_LEAF2(src)) {
1518 memcpy(dst, src, (lower + (Align-1)) & -Align);
1519 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1522 memcpy(dst, src, psize - unused);
1526 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1527 * @param[in] mc cursor pointing to the page to be touched
1528 * @return 0 on success, non-zero on failure.
1531 mdb_page_touch(MDB_cursor *mc)
1533 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1534 MDB_txn *txn = mc->mc_txn;
1535 MDB_cursor *m2, *m3;
1540 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1541 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1542 (rc = mdb_page_alloc(mc, 1, &np)))
1545 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi,mp->mp_pgno,pgno);
1546 assert(mp->mp_pgno != pgno);
1547 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1548 /* Update the parent page, if any, to point to the new page */
1550 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1551 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1552 SETPGNO(node, pgno);
1554 mc->mc_db->md_root = pgno;
1556 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1557 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1559 /* If txn has a parent, make sure the page is in our
1563 unsigned x = mdb_mid2l_search(dl, pgno);
1564 if (x <= dl[0].mid && dl[x].mid == pgno) {
1567 mc->mc_pg[mc->mc_top] = np;
1571 assert(dl[0].mid < MDB_IDL_UM_MAX);
1573 np = mdb_page_malloc(txn, 1);
1578 mdb_mid2l_insert(dl, &mid);
1583 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1585 np->mp_flags |= P_DIRTY;
1587 /* Adjust cursors pointing to mp */
1588 mc->mc_pg[mc->mc_top] = np;
1590 if (mc->mc_flags & C_SUB) {
1592 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1593 m3 = &m2->mc_xcursor->mx_cursor;
1594 if (m3->mc_snum < mc->mc_snum) continue;
1595 if (m3->mc_pg[mc->mc_top] == mp)
1596 m3->mc_pg[mc->mc_top] = np;
1599 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1600 if (m2->mc_snum < mc->mc_snum) continue;
1601 if (m2->mc_pg[mc->mc_top] == mp) {
1602 m2->mc_pg[mc->mc_top] = np;
1603 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1604 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1606 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1607 if (!(leaf->mn_flags & F_SUBDATA))
1608 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1617 mdb_env_sync(MDB_env *env, int force)
1620 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1621 if (env->me_flags & MDB_WRITEMAP) {
1622 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1623 ? MS_ASYNC : MS_SYNC;
1624 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1627 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1631 if (MDB_FDATASYNC(env->me_fd))
1638 /** Make shadow copies of all of parent txn's cursors */
1640 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1642 MDB_cursor *mc, *m2;
1643 unsigned int i, j, size;
1645 for (i=0;i<src->mt_numdbs; i++) {
1646 if (src->mt_cursors[i]) {
1647 size = sizeof(MDB_cursor);
1648 if (src->mt_cursors[i]->mc_xcursor)
1649 size += sizeof(MDB_xcursor);
1650 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1657 mc->mc_db = &dst->mt_dbs[i];
1658 mc->mc_dbx = m2->mc_dbx;
1659 mc->mc_dbflag = &dst->mt_dbflags[i];
1660 mc->mc_snum = m2->mc_snum;
1661 mc->mc_top = m2->mc_top;
1662 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1663 for (j=0; j<mc->mc_snum; j++) {
1664 mc->mc_pg[j] = m2->mc_pg[j];
1665 mc->mc_ki[j] = m2->mc_ki[j];
1667 if (m2->mc_xcursor) {
1668 MDB_xcursor *mx, *mx2;
1669 mx = (MDB_xcursor *)(mc+1);
1670 mc->mc_xcursor = mx;
1671 mx2 = m2->mc_xcursor;
1672 mx->mx_db = mx2->mx_db;
1673 mx->mx_dbx = mx2->mx_dbx;
1674 mx->mx_dbflag = mx2->mx_dbflag;
1675 mx->mx_cursor.mc_txn = dst;
1676 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1677 mx->mx_cursor.mc_db = &mx->mx_db;
1678 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1679 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1680 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1681 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1682 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1683 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1684 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1685 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1688 mc->mc_xcursor = NULL;
1690 mc->mc_next = dst->mt_cursors[i];
1691 dst->mt_cursors[i] = mc;
1698 /** Close this write txn's cursors, after optionally merging its shadow
1699 * cursors back into parent's.
1700 * @param[in] txn the transaction handle.
1701 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1702 * @return 0 on success, non-zero on failure.
1705 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1707 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1710 for (i = txn->mt_numdbs; --i >= 0; ) {
1711 for (mc = cursors[i]; mc; mc = next) {
1713 if (mc->mc_flags & merge) {
1714 MDB_cursor *m2 = mc->mc_orig;
1715 m2->mc_snum = mc->mc_snum;
1716 m2->mc_top = mc->mc_top;
1717 for (j = mc->mc_snum; --j >= 0; ) {
1718 m2->mc_pg[j] = mc->mc_pg[j];
1719 m2->mc_ki[j] = mc->mc_ki[j];
1722 if (mc->mc_flags & C_ALLOCD)
1729 #ifdef MDB_DEBUG_SKIP
1730 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
1733 mdb_txn_reset0(MDB_txn *txn, const char *act);
1735 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1736 * @param[in] txn the transaction handle to initialize
1737 * @return 0 on success, non-zero on failure.
1740 mdb_txn_renew0(MDB_txn *txn)
1742 MDB_env *env = txn->mt_env;
1745 int rc, new_notls = 0;
1749 txn->mt_numdbs = env->me_numdbs;
1750 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1752 if (txn->mt_flags & MDB_TXN_RDONLY) {
1753 if (!env->me_txns) {
1754 i = mdb_env_pick_meta(env);
1755 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1756 txn->mt_u.reader = NULL;
1758 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1759 pthread_getspecific(env->me_txkey);
1761 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1762 return MDB_BAD_RSLOT;
1764 pid_t pid = env->me_pid;
1765 pthread_t tid = pthread_self();
1768 for (i=0; i<env->me_txns->mti_numreaders; i++)
1769 if (env->me_txns->mti_readers[i].mr_pid == 0)
1771 if (i == env->me_maxreaders) {
1772 UNLOCK_MUTEX_R(env);
1773 return MDB_READERS_FULL;
1775 env->me_txns->mti_readers[i].mr_pid = pid;
1776 env->me_txns->mti_readers[i].mr_tid = tid;
1777 if (i >= env->me_txns->mti_numreaders)
1778 env->me_txns->mti_numreaders = i+1;
1779 /* Save numreaders for un-mutexed mdb_env_close() */
1780 env->me_numreaders = env->me_txns->mti_numreaders;
1781 UNLOCK_MUTEX_R(env);
1782 r = &env->me_txns->mti_readers[i];
1783 new_notls = (env->me_flags & MDB_NOTLS);
1784 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1789 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1790 txn->mt_u.reader = r;
1792 txn->mt_toggle = txn->mt_txnid & 1;
1793 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1797 txn->mt_txnid = env->me_txns->mti_txnid;
1798 txn->mt_toggle = txn->mt_txnid & 1;
1799 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1802 if (txn->mt_txnid == mdb_debug_start)
1805 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1806 txn->mt_u.dirty_list = env->me_dirty_list;
1807 txn->mt_u.dirty_list[0].mid = 0;
1808 txn->mt_free_pgs = env->me_free_pgs;
1809 txn->mt_free_pgs[0] = 0;
1813 /* Copy the DB info and flags */
1814 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1815 /* In a read txn, there is a data race here. Make sure our
1816 * last_pg/next_pg are up to date.
1818 lastpg2 = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1819 if (lastpg2 != txn->mt_next_pgno) {
1820 txn->mt_next_pgno = lastpg2;
1821 /* When this situation occurs, the txnid will certainly also
1822 * be out of date. But as noted before, we don't care about having
1823 * up to date read txn IDs.
1826 for (i=2; i<txn->mt_numdbs; i++) {
1827 x = env->me_dbflags[i];
1828 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1829 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1831 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1833 if (env->me_maxpg < txn->mt_next_pgno) {
1834 mdb_txn_reset0(txn, "renew0-mapfail");
1836 txn->mt_u.reader->mr_pid = 0;
1837 txn->mt_u.reader = NULL;
1839 return MDB_MAP_RESIZED;
1846 mdb_txn_renew(MDB_txn *txn)
1850 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1853 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1854 DPUTS("environment had fatal error, must shutdown!");
1858 rc = mdb_txn_renew0(txn);
1859 if (rc == MDB_SUCCESS) {
1860 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1861 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1862 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1868 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1872 int rc, size, tsize = sizeof(MDB_txn);
1874 if (env->me_flags & MDB_FATAL_ERROR) {
1875 DPUTS("environment had fatal error, must shutdown!");
1878 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1881 /* Nested transactions: Max 1 child, write txns only, no writemap */
1882 if (parent->mt_child ||
1883 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1884 (env->me_flags & MDB_WRITEMAP))
1888 tsize = sizeof(MDB_ntxn);
1890 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1891 if (!(flags & MDB_RDONLY))
1892 size += env->me_maxdbs * sizeof(MDB_cursor *);
1894 if ((txn = calloc(1, size)) == NULL) {
1895 DPRINTF("calloc: %s", strerror(ErrCode()));
1898 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1899 if (flags & MDB_RDONLY) {
1900 txn->mt_flags |= MDB_TXN_RDONLY;
1901 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1903 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1904 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1910 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1911 if (!txn->mt_u.dirty_list ||
1912 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1914 free(txn->mt_u.dirty_list);
1918 txn->mt_txnid = parent->mt_txnid;
1919 txn->mt_toggle = parent->mt_toggle;
1920 txn->mt_dirty_room = parent->mt_dirty_room;
1921 txn->mt_u.dirty_list[0].mid = 0;
1922 txn->mt_next_pgno = parent->mt_next_pgno;
1923 parent->mt_child = txn;
1924 txn->mt_parent = parent;
1925 txn->mt_numdbs = parent->mt_numdbs;
1926 txn->mt_flags = parent->mt_flags;
1927 txn->mt_dbxs = parent->mt_dbxs;
1928 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1929 /* Copy parent's mt_dbflags, but clear DB_NEW */
1930 for (i=0; i<txn->mt_numdbs; i++)
1931 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1933 ntxn = (MDB_ntxn *)txn;
1934 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1935 if (env->me_pghead) {
1936 size = MDB_IDL_SIZEOF(env->me_pghead);
1937 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1939 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1944 rc = mdb_cursor_shadow(parent, txn);
1946 mdb_txn_reset0(txn, "beginchild-fail");
1948 rc = mdb_txn_renew0(txn);
1954 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1955 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1956 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1962 /** Export or close DBI handles opened in this txn. */
1964 mdb_dbis_update(MDB_txn *txn, int keep)
1967 MDB_dbi n = txn->mt_numdbs;
1968 MDB_env *env = txn->mt_env;
1969 unsigned char *tdbflags = txn->mt_dbflags;
1971 for (i = n; --i >= 2;) {
1972 if (tdbflags[i] & DB_NEW) {
1974 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
1976 char *ptr = env->me_dbxs[i].md_name.mv_data;
1977 env->me_dbxs[i].md_name.mv_data = NULL;
1978 env->me_dbxs[i].md_name.mv_size = 0;
1979 env->me_dbflags[i] = 0;
1984 if (keep && env->me_numdbs < n)
1988 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1989 * May be called twice for readonly txns: First reset it, then abort.
1990 * @param[in] txn the transaction handle to reset
1993 mdb_txn_reset0(MDB_txn *txn, const char *act)
1995 MDB_env *env = txn->mt_env;
1997 /* Close any DBI handles opened in this txn */
1998 mdb_dbis_update(txn, 0);
2000 DPRINTF("%s txn %zu%c %p on mdbenv %p, root page %zu",
2001 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2002 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2004 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2005 if (txn->mt_u.reader) {
2006 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2007 if (!(env->me_flags & MDB_NOTLS))
2008 txn->mt_u.reader = NULL; /* txn does not own reader */
2010 txn->mt_numdbs = 0; /* close nothing if called again */
2011 txn->mt_dbxs = NULL; /* mark txn as reset */
2013 mdb_cursors_close(txn, 0);
2015 if (!(env->me_flags & MDB_WRITEMAP)) {
2016 mdb_dlist_free(txn);
2018 mdb_midl_free(env->me_pghead);
2020 if (txn->mt_parent) {
2021 txn->mt_parent->mt_child = NULL;
2022 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2023 mdb_midl_free(txn->mt_free_pgs);
2024 free(txn->mt_u.dirty_list);
2028 if (mdb_midl_shrink(&txn->mt_free_pgs))
2029 env->me_free_pgs = txn->mt_free_pgs;
2030 env->me_pghead = NULL;
2034 /* The writer mutex was locked in mdb_txn_begin. */
2035 UNLOCK_MUTEX_W(env);
2040 mdb_txn_reset(MDB_txn *txn)
2045 /* This call is only valid for read-only txns */
2046 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2049 mdb_txn_reset0(txn, "reset");
2053 mdb_txn_abort(MDB_txn *txn)
2059 mdb_txn_abort(txn->mt_child);
2061 mdb_txn_reset0(txn, "abort");
2062 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2063 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2064 txn->mt_u.reader->mr_pid = 0;
2069 /** Save the freelist as of this transaction to the freeDB.
2070 * This changes the freelist. Keep trying until it stabilizes.
2073 mdb_freelist_save(MDB_txn *txn)
2075 /* env->me_pghead[] can grow and shrink during this call.
2076 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2077 * Page numbers cannot disappear from txn->mt_free_pgs[].
2080 MDB_env *env = txn->mt_env;
2081 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2082 txnid_t pglast = 0, head_id = 0;
2083 pgno_t freecnt = 0, *free_pgs, *mop;
2084 ssize_t head_room = 0, total_room = 0, mop_len;
2086 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2088 if (env->me_pghead) {
2089 /* Make sure first page of freeDB is touched and on freelist */
2090 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2091 if (rc && rc != MDB_NOTFOUND)
2096 /* Come back here after each Put() in case freelist changed */
2099 /* If using records from freeDB which we have not yet
2100 * deleted, delete them and any we reserved for me_pghead.
2102 while (pglast < env->me_pglast) {
2103 rc = mdb_cursor_first(&mc, &key, NULL);
2106 pglast = head_id = *(txnid_t *)key.mv_data;
2107 total_room = head_room = 0;
2108 assert(pglast <= env->me_pglast);
2109 rc = mdb_cursor_del(&mc, 0);
2114 /* Save the IDL of pages freed by this txn, to a single record */
2115 if (freecnt < txn->mt_free_pgs[0]) {
2117 /* Make sure last page of freeDB is touched and on freelist */
2118 key.mv_size = MDB_MAXKEYSIZE+1;
2120 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2121 if (rc && rc != MDB_NOTFOUND)
2124 free_pgs = txn->mt_free_pgs;
2125 /* Write to last page of freeDB */
2126 key.mv_size = sizeof(txn->mt_txnid);
2127 key.mv_data = &txn->mt_txnid;
2129 freecnt = free_pgs[0];
2130 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2131 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2134 /* Retry if mt_free_pgs[] grew during the Put() */
2135 free_pgs = txn->mt_free_pgs;
2136 } while (freecnt < free_pgs[0]);
2137 mdb_midl_sort(free_pgs);
2138 memcpy(data.mv_data, free_pgs, data.mv_size);
2141 unsigned int i = free_pgs[0];
2142 DPRINTF("IDL write txn %zu root %zu num %u",
2143 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2145 DPRINTF("IDL %zu", free_pgs[i]);
2151 mop = env->me_pghead;
2152 mop_len = mop ? mop[0] : 0;
2154 /* Reserve records for me_pghead[]. Split it if multi-page,
2155 * to avoid searching freeDB for a page range. Use keys in
2156 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2158 if (total_room >= mop_len) {
2159 if (total_room == mop_len || --more < 0)
2161 } else if (head_room >= maxfree_1pg && head_id > 1) {
2162 /* Keep current record (overflow page), add a new one */
2166 /* (Re)write {key = head_id, IDL length = head_room} */
2167 total_room -= head_room;
2168 head_room = mop_len - total_room;
2169 if (head_room > maxfree_1pg && head_id > 1) {
2170 /* Overflow multi-page for part of me_pghead */
2171 head_room /= head_id; /* amortize page sizes */
2172 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2173 } else if (head_room < 0) {
2174 /* Rare case, not bothering to delete this record */
2177 key.mv_size = sizeof(head_id);
2178 key.mv_data = &head_id;
2179 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2180 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2183 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2184 total_room += head_room;
2187 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2188 * so they cannot rearrange anything, just read the destinations.
2195 rc = mdb_cursor_first(&mc, &key, &data);
2196 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2197 MDB_IDL dest = data.mv_data;
2198 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2200 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2204 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2205 if (! (mop_len -= len))
2212 /** Flush dirty pages to the map, after clearing their dirty flag.
2215 mdb_page_flush(MDB_txn *txn)
2217 MDB_env *env = txn->mt_env;
2218 MDB_ID2L dl = txn->mt_u.dirty_list;
2219 unsigned psize = env->me_psize;
2220 int i, pagecount = dl[0].mid, rc;
2221 size_t size = 0, pos = 0;
2223 MDB_page *dp = NULL;
2227 struct iovec iov[MDB_COMMIT_PAGES];
2228 ssize_t wpos, wsize = 0, wres;
2229 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2233 if (env->me_flags & MDB_WRITEMAP) {
2234 /* Clear dirty flags */
2235 for (i = pagecount; i; i--) {
2237 dp->mp_flags &= ~P_DIRTY;
2243 /* Write the pages */
2245 if (i <= pagecount) {
2248 /* clear dirty flag */
2249 dp->mp_flags &= ~P_DIRTY;
2252 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2257 /* Windows actually supports scatter/gather I/O, but only on
2258 * unbuffered file handles. Since we're relying on the OS page
2259 * cache for all our data, that's self-defeating. So we just
2260 * write pages one at a time. We use the ov structure to set
2261 * the write offset, to at least save the overhead of a Seek
2264 DPRINTF("committing page %zu", pgno);
2265 memset(&ov, 0, sizeof(ov));
2266 ov.Offset = pos & 0xffffffff;
2267 ov.OffsetHigh = pos >> 16 >> 16;
2268 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2270 DPRINTF("WriteFile: %d", rc);
2274 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2275 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2277 /* Write previous page(s) */
2278 #ifdef MDB_USE_PWRITEV
2279 wres = pwritev(env->me_fd, iov, n, wpos);
2282 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2284 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2286 DPRINTF("lseek: %s", strerror(rc));
2289 wres = writev(env->me_fd, iov, n);
2292 if (wres != wsize) {
2295 DPRINTF("Write error: %s", strerror(rc));
2297 rc = EIO; /* TODO: Use which error code? */
2298 DPUTS("short write, filesystem full?");
2309 DPRINTF("committing page %zu", pgno);
2310 next_pos = pos + size;
2311 iov[n].iov_len = size;
2312 iov[n].iov_base = (char *)dp;
2318 mdb_dlist_free(txn);
2324 mdb_txn_commit(MDB_txn *txn)
2330 assert(txn != NULL);
2331 assert(txn->mt_env != NULL);
2333 if (txn->mt_child) {
2334 rc = mdb_txn_commit(txn->mt_child);
2335 txn->mt_child = NULL;
2342 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2343 mdb_dbis_update(txn, 1);
2344 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2349 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2350 DPUTS("error flag is set, can't commit");
2352 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2357 if (txn->mt_parent) {
2358 MDB_txn *parent = txn->mt_parent;
2362 /* Append our free list to parent's */
2363 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2366 mdb_midl_free(txn->mt_free_pgs);
2368 parent->mt_next_pgno = txn->mt_next_pgno;
2369 parent->mt_flags = txn->mt_flags;
2371 /* Merge our cursors into parent's and close them */
2372 mdb_cursors_close(txn, C_SHADOW);
2374 /* Update parent's DB table. */
2375 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2376 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2377 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2378 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2379 for (i=2; i<txn->mt_numdbs; i++) {
2380 /* preserve parent's DB_NEW status */
2381 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2382 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2385 dst = txn->mt_parent->mt_u.dirty_list;
2386 src = txn->mt_u.dirty_list;
2387 /* Find len = length of merging our dirty list with parent's */
2389 dst[0].mid = 0; /* simplify loops */
2390 if (parent->mt_parent) {
2391 len = x + src[0].mid;
2392 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2393 for (i = x; y && i; y--) {
2394 pgno_t yp = src[y].mid;
2395 while (yp < dst[i].mid)
2397 if (yp == dst[i].mid) {
2402 } else { /* Simplify the above for single-ancestor case */
2403 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2405 /* Merge our dirty list with parent's */
2407 for (i = len; y; dst[i--] = src[y--]) {
2408 pgno_t yp = src[y].mid;
2409 while (yp < dst[x].mid)
2410 dst[i--] = dst[x--];
2411 if (yp == dst[x].mid)
2412 free(dst[x--].mptr);
2416 free(txn->mt_u.dirty_list);
2417 parent->mt_dirty_room = txn->mt_dirty_room;
2419 txn->mt_parent->mt_child = NULL;
2420 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2425 if (txn != env->me_txn) {
2426 DPUTS("attempt to commit unknown transaction");
2431 mdb_cursors_close(txn, 0);
2433 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2436 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2437 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2439 /* Update DB root pointers */
2440 if (txn->mt_numdbs > 2) {
2444 data.mv_size = sizeof(MDB_db);
2446 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2447 for (i = 2; i < txn->mt_numdbs; i++) {
2448 if (txn->mt_dbflags[i] & DB_DIRTY) {
2449 data.mv_data = &txn->mt_dbs[i];
2450 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2457 rc = mdb_freelist_save(txn);
2461 mdb_midl_free(env->me_pghead);
2462 env->me_pghead = NULL;
2463 if (mdb_midl_shrink(&txn->mt_free_pgs))
2464 env->me_free_pgs = txn->mt_free_pgs;
2470 if ((rc = mdb_page_flush(txn)) ||
2471 (rc = mdb_env_sync(env, 0)) ||
2472 (rc = mdb_env_write_meta(txn)))
2478 mdb_dbis_update(txn, 1);
2480 UNLOCK_MUTEX_W(env);
2490 /** Read the environment parameters of a DB environment before
2491 * mapping it into memory.
2492 * @param[in] env the environment handle
2493 * @param[out] meta address of where to store the meta information
2494 * @return 0 on success, non-zero on failure.
2497 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2504 /* We don't know the page size yet, so use a minimum value.
2505 * Read both meta pages so we can use the latest one.
2508 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2512 memset(&ov, 0, sizeof(ov));
2514 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2516 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2518 if (rc != MDB_PAGESIZE) {
2519 if (rc == 0 && off == 0)
2521 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2522 DPRINTF("read: %s", mdb_strerror(rc));
2526 p = (MDB_page *)&pbuf;
2528 if (!F_ISSET(p->mp_flags, P_META)) {
2529 DPRINTF("page %zu not a meta page", p->mp_pgno);
2534 if (m->mm_magic != MDB_MAGIC) {
2535 DPUTS("meta has invalid magic");
2539 if (m->mm_version != MDB_VERSION) {
2540 DPRINTF("database is version %u, expected version %u",
2541 m->mm_version, MDB_VERSION);
2542 return MDB_VERSION_MISMATCH;
2545 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2551 /** Write the environment parameters of a freshly created DB environment.
2552 * @param[in] env the environment handle
2553 * @param[out] meta address of where to store the meta information
2554 * @return 0 on success, non-zero on failure.
2557 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2563 DPUTS("writing new meta page");
2565 GET_PAGESIZE(psize);
2567 meta->mm_magic = MDB_MAGIC;
2568 meta->mm_version = MDB_VERSION;
2569 meta->mm_mapsize = env->me_mapsize;
2570 meta->mm_psize = psize;
2571 meta->mm_last_pg = 1;
2572 meta->mm_flags = env->me_flags & 0xffff;
2573 meta->mm_flags |= MDB_INTEGERKEY;
2574 meta->mm_dbs[0].md_root = P_INVALID;
2575 meta->mm_dbs[1].md_root = P_INVALID;
2577 p = calloc(2, psize);
2579 p->mp_flags = P_META;
2580 *(MDB_meta *)METADATA(p) = *meta;
2582 q = (MDB_page *)((char *)p + psize);
2584 q->mp_flags = P_META;
2585 *(MDB_meta *)METADATA(q) = *meta;
2591 memset(&ov, 0, sizeof(ov));
2592 rc = WriteFile(env->me_fd, p, psize * 2, &len, &ov);
2593 rc = rc ? (len == psize * 2 ? MDB_SUCCESS : EIO) : ErrCode();
2596 rc = pwrite(env->me_fd, p, psize * 2, 0);
2597 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
2603 /** Update the environment info to commit a transaction.
2604 * @param[in] txn the transaction that's being committed
2605 * @return 0 on success, non-zero on failure.
2608 mdb_env_write_meta(MDB_txn *txn)
2611 MDB_meta meta, metab, *mp;
2613 int rc, len, toggle;
2622 assert(txn != NULL);
2623 assert(txn->mt_env != NULL);
2625 toggle = !txn->mt_toggle;
2626 DPRINTF("writing meta page %d for root page %zu",
2627 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2630 mp = env->me_metas[toggle];
2632 if (env->me_flags & MDB_WRITEMAP) {
2633 /* Persist any increases of mapsize config */
2634 if (env->me_mapsize > mp->mm_mapsize)
2635 mp->mm_mapsize = env->me_mapsize;
2636 mp->mm_dbs[0] = txn->mt_dbs[0];
2637 mp->mm_dbs[1] = txn->mt_dbs[1];
2638 mp->mm_last_pg = txn->mt_next_pgno - 1;
2639 mp->mm_txnid = txn->mt_txnid;
2640 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2641 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2644 ptr += env->me_psize;
2645 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2652 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2653 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2655 ptr = (char *)&meta;
2656 if (env->me_mapsize > mp->mm_mapsize) {
2657 /* Persist any increases of mapsize config */
2658 meta.mm_mapsize = env->me_mapsize;
2659 off = offsetof(MDB_meta, mm_mapsize);
2661 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2663 len = sizeof(MDB_meta) - off;
2666 meta.mm_dbs[0] = txn->mt_dbs[0];
2667 meta.mm_dbs[1] = txn->mt_dbs[1];
2668 meta.mm_last_pg = txn->mt_next_pgno - 1;
2669 meta.mm_txnid = txn->mt_txnid;
2672 off += env->me_psize;
2675 /* Write to the SYNC fd */
2676 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2677 env->me_fd : env->me_mfd;
2680 memset(&ov, 0, sizeof(ov));
2682 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
2686 rc = pwrite(mfd, ptr, len, off);
2689 rc = rc < 0 ? ErrCode() : EIO;
2690 DPUTS("write failed, disk error?");
2691 /* On a failure, the pagecache still contains the new data.
2692 * Write some old data back, to prevent it from being used.
2693 * Use the non-SYNC fd; we know it will fail anyway.
2695 meta.mm_last_pg = metab.mm_last_pg;
2696 meta.mm_txnid = metab.mm_txnid;
2698 memset(&ov, 0, sizeof(ov));
2700 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2702 r2 = pwrite(env->me_fd, ptr, len, off);
2705 env->me_flags |= MDB_FATAL_ERROR;
2709 /* Memory ordering issues are irrelevant; since the entire writer
2710 * is wrapped by wmutex, all of these changes will become visible
2711 * after the wmutex is unlocked. Since the DB is multi-version,
2712 * readers will get consistent data regardless of how fresh or
2713 * how stale their view of these values is.
2715 env->me_txns->mti_txnid = txn->mt_txnid;
2720 /** Check both meta pages to see which one is newer.
2721 * @param[in] env the environment handle
2722 * @return meta toggle (0 or 1).
2725 mdb_env_pick_meta(const MDB_env *env)
2727 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2731 mdb_env_create(MDB_env **env)
2735 e = calloc(1, sizeof(MDB_env));
2739 e->me_maxreaders = DEFAULT_READERS;
2740 e->me_maxdbs = e->me_numdbs = 2;
2741 e->me_fd = INVALID_HANDLE_VALUE;
2742 e->me_lfd = INVALID_HANDLE_VALUE;
2743 e->me_mfd = INVALID_HANDLE_VALUE;
2744 #ifdef MDB_USE_POSIX_SEM
2745 e->me_rmutex = SEM_FAILED;
2746 e->me_wmutex = SEM_FAILED;
2748 e->me_pid = getpid();
2749 VGMEMP_CREATE(e,0,0);
2755 mdb_env_set_mapsize(MDB_env *env, size_t size)
2759 env->me_mapsize = size;
2761 env->me_maxpg = env->me_mapsize / env->me_psize;
2766 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2770 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2775 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2777 if (env->me_map || readers < 1)
2779 env->me_maxreaders = readers;
2784 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2786 if (!env || !readers)
2788 *readers = env->me_maxreaders;
2792 /** Further setup required for opening an MDB environment
2795 mdb_env_open2(MDB_env *env)
2797 unsigned int flags = env->me_flags;
2805 memset(&meta, 0, sizeof(meta));
2807 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2810 DPUTS("new mdbenv");
2814 /* Was a mapsize configured? */
2815 if (!env->me_mapsize) {
2816 /* If this is a new environment, take the default,
2817 * else use the size recorded in the existing env.
2819 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2820 } else if (env->me_mapsize < meta.mm_mapsize) {
2821 /* If the configured size is smaller, make sure it's
2822 * still big enough. Silently round up to minimum if not.
2824 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2825 if (env->me_mapsize < minsize)
2826 env->me_mapsize = minsize;
2833 LONG sizelo, sizehi;
2834 sizelo = env->me_mapsize & 0xffffffff;
2835 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
2836 /* Windows won't create mappings for zero length files.
2837 * Just allocate the maxsize right now.
2840 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
2841 || !SetEndOfFile(env->me_fd)
2842 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
2845 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2846 PAGE_READWRITE : PAGE_READONLY,
2847 sizehi, sizelo, NULL);
2850 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2851 FILE_MAP_WRITE : FILE_MAP_READ,
2852 0, 0, env->me_mapsize, meta.mm_address);
2853 rc = env->me_map ? 0 : ErrCode();
2861 if (flags & MDB_WRITEMAP) {
2863 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2866 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2868 if (env->me_map == MAP_FAILED) {
2872 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2874 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2876 #ifdef POSIX_MADV_RANDOM
2877 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2878 #endif /* POSIX_MADV_RANDOM */
2879 #endif /* MADV_RANDOM */
2883 if (flags & MDB_FIXEDMAP)
2884 meta.mm_address = env->me_map;
2885 i = mdb_env_init_meta(env, &meta);
2886 if (i != MDB_SUCCESS) {
2889 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2890 /* Can happen because the address argument to mmap() is just a
2891 * hint. mmap() can pick another, e.g. if the range is in use.
2892 * The MAP_FIXED flag would prevent that, but then mmap could
2893 * instead unmap existing pages to make room for the new map.
2895 return EBUSY; /* TODO: Make a new MDB_* error code? */
2897 env->me_psize = meta.mm_psize;
2898 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2899 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2901 env->me_maxpg = env->me_mapsize / env->me_psize;
2903 p = (MDB_page *)env->me_map;
2904 env->me_metas[0] = METADATA(p);
2905 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2909 int toggle = mdb_env_pick_meta(env);
2910 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2912 DPRINTF("opened database version %u, pagesize %u",
2913 env->me_metas[0]->mm_version, env->me_psize);
2914 DPRINTF("using meta page %d", toggle);
2915 DPRINTF("depth: %u", db->md_depth);
2916 DPRINTF("entries: %zu", db->md_entries);
2917 DPRINTF("branch pages: %zu", db->md_branch_pages);
2918 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2919 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2920 DPRINTF("root: %zu", db->md_root);
2928 /** Release a reader thread's slot in the reader lock table.
2929 * This function is called automatically when a thread exits.
2930 * @param[in] ptr This points to the slot in the reader lock table.
2933 mdb_env_reader_dest(void *ptr)
2935 MDB_reader *reader = ptr;
2941 /** Junk for arranging thread-specific callbacks on Windows. This is
2942 * necessarily platform and compiler-specific. Windows supports up
2943 * to 1088 keys. Let's assume nobody opens more than 64 environments
2944 * in a single process, for now. They can override this if needed.
2946 #ifndef MAX_TLS_KEYS
2947 #define MAX_TLS_KEYS 64
2949 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2950 static int mdb_tls_nkeys;
2952 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2956 case DLL_PROCESS_ATTACH: break;
2957 case DLL_THREAD_ATTACH: break;
2958 case DLL_THREAD_DETACH:
2959 for (i=0; i<mdb_tls_nkeys; i++) {
2960 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2961 mdb_env_reader_dest(r);
2964 case DLL_PROCESS_DETACH: break;
2969 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2971 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2975 /* Force some symbol references.
2976 * _tls_used forces the linker to create the TLS directory if not already done
2977 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2979 #pragma comment(linker, "/INCLUDE:_tls_used")
2980 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2981 #pragma const_seg(".CRT$XLB")
2982 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2983 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2986 #pragma comment(linker, "/INCLUDE:__tls_used")
2987 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2988 #pragma data_seg(".CRT$XLB")
2989 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2991 #endif /* WIN 32/64 */
2992 #endif /* !__GNUC__ */
2995 /** Downgrade the exclusive lock on the region back to shared */
2997 mdb_env_share_locks(MDB_env *env, int *excl)
2999 int rc = 0, toggle = mdb_env_pick_meta(env);
3001 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3006 /* First acquire a shared lock. The Unlock will
3007 * then release the existing exclusive lock.
3009 memset(&ov, 0, sizeof(ov));
3010 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3013 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3019 struct flock lock_info;
3020 /* The shared lock replaces the existing lock */
3021 memset((void *)&lock_info, 0, sizeof(lock_info));
3022 lock_info.l_type = F_RDLCK;
3023 lock_info.l_whence = SEEK_SET;
3024 lock_info.l_start = 0;
3025 lock_info.l_len = 1;
3026 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3027 (rc = ErrCode()) == EINTR) ;
3028 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3035 /** Try to get exlusive lock, otherwise shared.
3036 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3039 mdb_env_excl_lock(MDB_env *env, int *excl)
3043 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3047 memset(&ov, 0, sizeof(ov));
3048 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3055 struct flock lock_info;
3056 memset((void *)&lock_info, 0, sizeof(lock_info));
3057 lock_info.l_type = F_WRLCK;
3058 lock_info.l_whence = SEEK_SET;
3059 lock_info.l_start = 0;
3060 lock_info.l_len = 1;
3061 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3062 (rc = ErrCode()) == EINTR) ;
3066 # ifdef MDB_USE_POSIX_SEM
3067 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3070 lock_info.l_type = F_RDLCK;
3071 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3072 (rc = ErrCode()) == EINTR) ;
3080 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3082 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3084 * @(#) $Revision: 5.1 $
3085 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3086 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3088 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3092 * Please do not copyright this code. This code is in the public domain.
3094 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3095 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3096 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3097 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3098 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3099 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3100 * PERFORMANCE OF THIS SOFTWARE.
3103 * chongo <Landon Curt Noll> /\oo/\
3104 * http://www.isthe.com/chongo/
3106 * Share and Enjoy! :-)
3109 typedef unsigned long long mdb_hash_t;
3110 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3112 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3113 * @param[in] str string to hash
3114 * @param[in] hval initial value for hash
3115 * @return 64 bit hash
3117 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3118 * hval arg on the first call.
3121 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3123 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3124 unsigned char *end = s + val->mv_size;
3126 * FNV-1a hash each octet of the string
3129 /* xor the bottom with the current octet */
3130 hval ^= (mdb_hash_t)*s++;
3132 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3133 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3134 (hval << 7) + (hval << 8) + (hval << 40);
3136 /* return our new hash value */
3140 /** Hash the string and output the hash in hex.
3141 * @param[in] str string to hash
3142 * @param[out] hexbuf an array of 17 chars to hold the hash
3145 mdb_hash_hex(MDB_val *val, char *hexbuf)
3148 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3149 for (i=0; i<8; i++) {
3150 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3156 /** Open and/or initialize the lock region for the environment.
3157 * @param[in] env The MDB environment.
3158 * @param[in] lpath The pathname of the file used for the lock region.
3159 * @param[in] mode The Unix permissions for the file, if we create it.
3160 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3161 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3162 * @return 0 on success, non-zero on failure.
3165 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3168 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3170 # define MDB_ERRCODE_ROFS EROFS
3171 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3172 # define MDB_CLOEXEC O_CLOEXEC
3175 # define MDB_CLOEXEC 0
3182 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3183 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3184 FILE_ATTRIBUTE_NORMAL, NULL);
3186 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3188 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3190 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3195 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3196 /* Lose record locks when exec*() */
3197 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3198 fcntl(env->me_lfd, F_SETFD, fdflags);
3201 if (!(env->me_flags & MDB_NOTLS)) {
3202 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3205 env->me_flags |= MDB_ENV_TXKEY;
3207 /* Windows TLS callbacks need help finding their TLS info. */
3208 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3212 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3216 /* Try to get exclusive lock. If we succeed, then
3217 * nobody is using the lock region and we should initialize it.
3219 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3222 size = GetFileSize(env->me_lfd, NULL);
3224 size = lseek(env->me_lfd, 0, SEEK_END);
3225 if (size == -1) goto fail_errno;
3227 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3228 if (size < rsize && *excl > 0) {
3230 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3231 || !SetEndOfFile(env->me_lfd))
3234 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3238 size = rsize - sizeof(MDB_txninfo);
3239 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3244 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3246 if (!mh) goto fail_errno;
3247 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3249 if (!env->me_txns) goto fail_errno;
3251 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3253 if (m == MAP_FAILED) goto fail_errno;
3259 BY_HANDLE_FILE_INFORMATION stbuf;
3268 if (!mdb_sec_inited) {
3269 InitializeSecurityDescriptor(&mdb_null_sd,
3270 SECURITY_DESCRIPTOR_REVISION);
3271 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3272 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3273 mdb_all_sa.bInheritHandle = FALSE;
3274 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3277 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3278 idbuf.volume = stbuf.dwVolumeSerialNumber;
3279 idbuf.nhigh = stbuf.nFileIndexHigh;
3280 idbuf.nlow = stbuf.nFileIndexLow;
3281 val.mv_data = &idbuf;
3282 val.mv_size = sizeof(idbuf);
3283 mdb_hash_hex(&val, hexbuf);
3284 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3285 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3286 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3287 if (!env->me_rmutex) goto fail_errno;
3288 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3289 if (!env->me_wmutex) goto fail_errno;
3290 #elif defined(MDB_USE_POSIX_SEM)
3299 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3300 idbuf.dev = stbuf.st_dev;
3301 idbuf.ino = stbuf.st_ino;
3302 val.mv_data = &idbuf;
3303 val.mv_size = sizeof(idbuf);
3304 mdb_hash_hex(&val, hexbuf);
3305 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3306 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3307 /* Clean up after a previous run, if needed: Try to
3308 * remove both semaphores before doing anything else.
3310 sem_unlink(env->me_txns->mti_rmname);
3311 sem_unlink(env->me_txns->mti_wmname);
3312 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3313 O_CREAT|O_EXCL, mode, 1);
3314 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3315 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3316 O_CREAT|O_EXCL, mode, 1);
3317 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3318 #else /* MDB_USE_POSIX_SEM */
3319 pthread_mutexattr_t mattr;
3321 if ((rc = pthread_mutexattr_init(&mattr))
3322 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3323 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3324 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3326 pthread_mutexattr_destroy(&mattr);
3327 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3329 env->me_txns->mti_version = MDB_VERSION;
3330 env->me_txns->mti_magic = MDB_MAGIC;
3331 env->me_txns->mti_txnid = 0;
3332 env->me_txns->mti_numreaders = 0;
3335 if (env->me_txns->mti_magic != MDB_MAGIC) {
3336 DPUTS("lock region has invalid magic");
3340 if (env->me_txns->mti_version != MDB_VERSION) {
3341 DPRINTF("lock region is version %u, expected version %u",
3342 env->me_txns->mti_version, MDB_VERSION);
3343 rc = MDB_VERSION_MISMATCH;
3347 if (rc && rc != EACCES && rc != EAGAIN) {
3351 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3352 if (!env->me_rmutex) goto fail_errno;
3353 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3354 if (!env->me_wmutex) goto fail_errno;
3355 #elif defined(MDB_USE_POSIX_SEM)
3356 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3357 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3358 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3359 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3370 /** The name of the lock file in the DB environment */
3371 #define LOCKNAME "/lock.mdb"
3372 /** The name of the data file in the DB environment */
3373 #define DATANAME "/data.mdb"
3374 /** The suffix of the lock file when no subdir is used */
3375 #define LOCKSUFF "-lock"
3376 /** Only a subset of the @ref mdb_env flags can be changed
3377 * at runtime. Changing other flags requires closing the
3378 * environment and re-opening it with the new flags.
3380 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3381 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3384 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3386 int oflags, rc, len, excl = -1;
3387 char *lpath, *dpath;
3389 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3393 if (flags & MDB_NOSUBDIR) {
3394 rc = len + sizeof(LOCKSUFF) + len + 1;
3396 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3401 if (flags & MDB_NOSUBDIR) {
3402 dpath = lpath + len + sizeof(LOCKSUFF);
3403 sprintf(lpath, "%s" LOCKSUFF, path);
3404 strcpy(dpath, path);
3406 dpath = lpath + len + sizeof(LOCKNAME);
3407 sprintf(lpath, "%s" LOCKNAME, path);
3408 sprintf(dpath, "%s" DATANAME, path);
3412 flags |= env->me_flags;
3413 if (flags & MDB_RDONLY) {
3414 /* silently ignore WRITEMAP when we're only getting read access */
3415 flags &= ~MDB_WRITEMAP;
3417 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3418 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3421 env->me_flags = flags |= MDB_ENV_ACTIVE;
3425 env->me_path = strdup(path);
3426 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3427 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3428 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3433 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3438 if (F_ISSET(flags, MDB_RDONLY)) {
3439 oflags = GENERIC_READ;
3440 len = OPEN_EXISTING;
3442 oflags = GENERIC_READ|GENERIC_WRITE;
3445 mode = FILE_ATTRIBUTE_NORMAL;
3446 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3447 NULL, len, mode, NULL);
3449 if (F_ISSET(flags, MDB_RDONLY))
3452 oflags = O_RDWR | O_CREAT;
3454 env->me_fd = open(dpath, oflags, mode);
3456 if (env->me_fd == INVALID_HANDLE_VALUE) {
3461 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3462 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3463 env->me_mfd = env->me_fd;
3465 /* Synchronous fd for meta writes. Needed even with
3466 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3469 env->me_mfd = CreateFile(dpath, oflags,
3470 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3471 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3473 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3475 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3480 DPRINTF("opened dbenv %p", (void *) env);
3482 rc = mdb_env_share_locks(env, &excl);
3488 mdb_env_close0(env, excl);
3494 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3496 mdb_env_close0(MDB_env *env, int excl)
3500 if (!(env->me_flags & MDB_ENV_ACTIVE))
3503 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3504 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3505 free(env->me_dbxs[i].md_name.mv_data);
3507 free(env->me_dbflags);
3510 free(env->me_dirty_list);
3511 mdb_midl_free(env->me_free_pgs);
3513 if (env->me_flags & MDB_ENV_TXKEY) {
3514 pthread_key_delete(env->me_txkey);
3516 /* Delete our key from the global list */
3517 for (i=0; i<mdb_tls_nkeys; i++)
3518 if (mdb_tls_keys[i] == env->me_txkey) {
3519 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3527 munmap(env->me_map, env->me_mapsize);
3529 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3530 (void) close(env->me_mfd);
3531 if (env->me_fd != INVALID_HANDLE_VALUE)
3532 (void) close(env->me_fd);
3534 pid_t pid = env->me_pid;
3535 /* Clearing readers is done in this function because
3536 * me_txkey with its destructor must be disabled first.
3538 for (i = env->me_numreaders; --i >= 0; )
3539 if (env->me_txns->mti_readers[i].mr_pid == pid)
3540 env->me_txns->mti_readers[i].mr_pid = 0;
3542 if (env->me_rmutex) {
3543 CloseHandle(env->me_rmutex);
3544 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3546 /* Windows automatically destroys the mutexes when
3547 * the last handle closes.
3549 #elif defined(MDB_USE_POSIX_SEM)
3550 if (env->me_rmutex != SEM_FAILED) {
3551 sem_close(env->me_rmutex);
3552 if (env->me_wmutex != SEM_FAILED)
3553 sem_close(env->me_wmutex);
3554 /* If we have the filelock: If we are the
3555 * only remaining user, clean up semaphores.
3558 mdb_env_excl_lock(env, &excl);
3560 sem_unlink(env->me_txns->mti_rmname);
3561 sem_unlink(env->me_txns->mti_wmname);
3565 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3567 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3570 /* Unlock the lockfile. Windows would have unlocked it
3571 * after closing anyway, but not necessarily at once.
3573 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3576 (void) close(env->me_lfd);
3579 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3583 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3585 MDB_txn *txn = NULL;
3590 /* Do the lock/unlock of the reader mutex before starting the
3591 * write txn. Otherwise other read txns could block writers.
3593 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3598 /* We must start the actual read txn after blocking writers */
3599 mdb_txn_reset0(txn, "reset-stage1");
3601 /* Temporarily block writers until we snapshot the meta pages */
3604 rc = mdb_txn_renew0(txn);
3606 UNLOCK_MUTEX_W(env);
3611 wsize = env->me_psize * 2;
3615 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3616 rc = rc ? (len == wsize ? MDB_SUCCESS : EIO) : ErrCode();
3619 rc = write(fd, env->me_map, wsize);
3620 rc = rc == (int)wsize ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
3623 UNLOCK_MUTEX_W(env);
3628 ptr = env->me_map + wsize;
3629 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3633 if (wsize > MAX_WRITE)
3637 rc = WriteFile(fd, ptr, w2, &len, NULL);
3638 rc = rc ? (len == w2 ? MDB_SUCCESS : EIO) : ErrCode();
3647 if (wsize > MAX_WRITE)
3651 wres = write(fd, ptr, w2);
3652 rc = wres == (ssize_t)w2 ? MDB_SUCCESS : wres < 0 ? ErrCode() : EIO;
3665 mdb_env_copy(MDB_env *env, const char *path)
3669 HANDLE newfd = INVALID_HANDLE_VALUE;
3671 if (env->me_flags & MDB_NOSUBDIR) {
3672 lpath = (char *)path;
3675 len += sizeof(DATANAME);
3676 lpath = malloc(len);
3679 sprintf(lpath, "%s" DATANAME, path);
3682 /* The destination path must exist, but the destination file must not.
3683 * We don't want the OS to cache the writes, since the source data is
3684 * already in the OS cache.
3687 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3688 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3690 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3696 if (newfd == INVALID_HANDLE_VALUE) {
3701 #ifdef F_NOCACHE /* __APPLE__ */
3702 rc = fcntl(newfd, F_NOCACHE, 1);
3709 rc = mdb_env_copyfd(env, newfd);
3712 if (!(env->me_flags & MDB_NOSUBDIR))
3714 if (newfd != INVALID_HANDLE_VALUE)
3715 if (close(newfd) < 0 && rc == MDB_SUCCESS)
3722 mdb_env_close(MDB_env *env)
3729 VGMEMP_DESTROY(env);
3730 while ((dp = env->me_dpages) != NULL) {
3731 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3732 env->me_dpages = dp->mp_next;
3736 mdb_env_close0(env, 0);
3740 /** Compare two items pointing at aligned size_t's */
3742 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3744 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3745 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3748 /** Compare two items pointing at aligned int's */
3750 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3752 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3753 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3756 /** Compare two items pointing at ints of unknown alignment.
3757 * Nodes and keys are guaranteed to be 2-byte aligned.
3760 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3762 #if BYTE_ORDER == LITTLE_ENDIAN
3763 unsigned short *u, *c;
3766 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3767 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3770 } while(!x && u > (unsigned short *)a->mv_data);
3773 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3777 /** Compare two items lexically */
3779 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3786 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3792 diff = memcmp(a->mv_data, b->mv_data, len);
3793 return diff ? diff : len_diff<0 ? -1 : len_diff;
3796 /** Compare two items in reverse byte order */
3798 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3800 const unsigned char *p1, *p2, *p1_lim;
3804 p1_lim = (const unsigned char *)a->mv_data;
3805 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3806 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3808 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3814 while (p1 > p1_lim) {
3815 diff = *--p1 - *--p2;
3819 return len_diff<0 ? -1 : len_diff;
3822 /** Search for key within a page, using binary search.
3823 * Returns the smallest entry larger or equal to the key.
3824 * If exactp is non-null, stores whether the found entry was an exact match
3825 * in *exactp (1 or 0).
3826 * Updates the cursor index with the index of the found entry.
3827 * If no entry larger or equal to the key is found, returns NULL.
3830 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3832 unsigned int i = 0, nkeys;
3835 MDB_page *mp = mc->mc_pg[mc->mc_top];
3836 MDB_node *node = NULL;
3841 nkeys = NUMKEYS(mp);
3846 COPY_PGNO(pgno, mp->mp_pgno);
3847 DPRINTF("searching %u keys in %s %spage %zu",
3848 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3855 low = IS_LEAF(mp) ? 0 : 1;
3857 cmp = mc->mc_dbx->md_cmp;
3859 /* Branch pages have no data, so if using integer keys,
3860 * alignment is guaranteed. Use faster mdb_cmp_int.
3862 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3863 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3870 nodekey.mv_size = mc->mc_db->md_pad;
3871 node = NODEPTR(mp, 0); /* fake */
3872 while (low <= high) {
3873 i = (low + high) >> 1;
3874 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3875 rc = cmp(key, &nodekey);
3876 DPRINTF("found leaf index %u [%s], rc = %i",
3877 i, DKEY(&nodekey), rc);
3886 while (low <= high) {
3887 i = (low + high) >> 1;
3889 node = NODEPTR(mp, i);
3890 nodekey.mv_size = NODEKSZ(node);
3891 nodekey.mv_data = NODEKEY(node);
3893 rc = cmp(key, &nodekey);
3896 DPRINTF("found leaf index %u [%s], rc = %i",
3897 i, DKEY(&nodekey), rc);
3899 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3900 i, DKEY(&nodekey), NODEPGNO(node), rc);
3911 if (rc > 0) { /* Found entry is less than the key. */
3912 i++; /* Skip to get the smallest entry larger than key. */
3914 node = NODEPTR(mp, i);
3917 *exactp = (rc == 0);
3918 /* store the key index */
3919 mc->mc_ki[mc->mc_top] = i;
3921 /* There is no entry larger or equal to the key. */
3924 /* nodeptr is fake for LEAF2 */
3930 mdb_cursor_adjust(MDB_cursor *mc, func)
3934 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3935 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3942 /** Pop a page off the top of the cursor's stack. */
3944 mdb_cursor_pop(MDB_cursor *mc)
3947 #ifndef MDB_DEBUG_SKIP
3948 MDB_page *top = mc->mc_pg[mc->mc_top];
3954 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3955 mc->mc_dbi, (void *) mc);
3959 /** Push a page onto the top of the cursor's stack. */
3961 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3963 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3964 mc->mc_dbi, (void *) mc);
3966 if (mc->mc_snum >= CURSOR_STACK) {
3967 assert(mc->mc_snum < CURSOR_STACK);
3968 return MDB_CURSOR_FULL;
3971 mc->mc_top = mc->mc_snum++;
3972 mc->mc_pg[mc->mc_top] = mp;
3973 mc->mc_ki[mc->mc_top] = 0;
3978 /** Find the address of the page corresponding to a given page number.
3979 * @param[in] txn the transaction for this access.
3980 * @param[in] pgno the page number for the page to retrieve.
3981 * @param[out] ret address of a pointer where the page's address will be stored.
3982 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
3983 * @return 0 on success, non-zero on failure.
3986 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
3991 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3992 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3997 MDB_ID2L dl = tx2->mt_u.dirty_list;
3999 unsigned x = mdb_mid2l_search(dl, pgno);
4000 if (x <= dl[0].mid && dl[x].mid == pgno) {
4006 } while ((tx2 = tx2->mt_parent) != NULL);
4009 if (pgno < txn->mt_next_pgno) {
4011 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4013 DPRINTF("page %zu not found", pgno);
4015 return MDB_PAGE_NOTFOUND;
4025 /** Search for the page a given key should be in.
4026 * Pushes parent pages on the cursor stack. This function continues a
4027 * search on a cursor that has already been initialized. (Usually by
4028 * #mdb_page_search() but also by #mdb_node_move().)
4029 * @param[in,out] mc the cursor for this operation.
4030 * @param[in] key the key to search for. If NULL, search for the lowest
4031 * page. (This is used by #mdb_cursor_first().)
4032 * @param[in] modify If true, visited pages are updated with new page numbers.
4033 * @return 0 on success, non-zero on failure.
4036 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4038 MDB_page *mp = mc->mc_pg[mc->mc_top];
4043 while (IS_BRANCH(mp)) {
4047 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4048 assert(NUMKEYS(mp) > 1);
4049 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4051 if (key == NULL) /* Initialize cursor to first page. */
4053 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4054 /* cursor to last page */
4058 node = mdb_node_search(mc, key, &exact);
4060 i = NUMKEYS(mp) - 1;
4062 i = mc->mc_ki[mc->mc_top];
4071 DPRINTF("following index %u for key [%s]",
4073 assert(i < NUMKEYS(mp));
4074 node = NODEPTR(mp, i);
4076 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4079 mc->mc_ki[mc->mc_top] = i;
4080 if ((rc = mdb_cursor_push(mc, mp)))
4084 if ((rc = mdb_page_touch(mc)) != 0)
4086 mp = mc->mc_pg[mc->mc_top];
4091 DPRINTF("internal error, index points to a %02X page!?",
4093 return MDB_CORRUPTED;
4096 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4097 key ? DKEY(key) : NULL);
4102 /** Search for the lowest key under the current branch page.
4103 * This just bypasses a NUMKEYS check in the current page
4104 * before calling mdb_page_search_root(), because the callers
4105 * are all in situations where the current page is known to
4109 mdb_page_search_lowest(MDB_cursor *mc)
4111 MDB_page *mp = mc->mc_pg[mc->mc_top];
4112 MDB_node *node = NODEPTR(mp, 0);
4115 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4118 mc->mc_ki[mc->mc_top] = 0;
4119 if ((rc = mdb_cursor_push(mc, mp)))
4121 return mdb_page_search_root(mc, NULL, 0);
4124 /** Search for the page a given key should be in.
4125 * Pushes parent pages on the cursor stack. This function just sets up
4126 * the search; it finds the root page for \b mc's database and sets this
4127 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4128 * called to complete the search.
4129 * @param[in,out] mc the cursor for this operation.
4130 * @param[in] key the key to search for. If NULL, search for the lowest
4131 * page. (This is used by #mdb_cursor_first().)
4132 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4133 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4134 * @return 0 on success, non-zero on failure.
4137 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4142 /* Make sure the txn is still viable, then find the root from
4143 * the txn's db table.
4145 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4146 DPUTS("transaction has failed, must abort");
4149 /* Make sure we're using an up-to-date root */
4150 if (mc->mc_dbi > MAIN_DBI) {
4151 if ((*mc->mc_dbflag & DB_STALE) ||
4152 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4154 unsigned char dbflag = 0;
4155 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4156 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4159 if (*mc->mc_dbflag & DB_STALE) {
4163 MDB_node *leaf = mdb_node_search(&mc2,
4164 &mc->mc_dbx->md_name, &exact);
4166 return MDB_NOTFOUND;
4167 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4170 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4172 /* The txn may not know this DBI, or another process may
4173 * have dropped and recreated the DB with other flags.
4175 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4176 return MDB_INCOMPATIBLE;
4177 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4179 if (flags & MDB_PS_MODIFY)
4181 *mc->mc_dbflag &= ~DB_STALE;
4182 *mc->mc_dbflag |= dbflag;
4185 root = mc->mc_db->md_root;
4187 if (root == P_INVALID) { /* Tree is empty. */
4188 DPUTS("tree is empty");
4189 return MDB_NOTFOUND;
4194 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4195 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4201 DPRINTF("db %u root page %zu has flags 0x%X",
4202 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4204 if (flags & MDB_PS_MODIFY) {
4205 if ((rc = mdb_page_touch(mc)))
4209 if (flags & MDB_PS_ROOTONLY)
4212 return mdb_page_search_root(mc, key, flags);
4216 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4218 MDB_txn *txn = mc->mc_txn;
4219 pgno_t pg = mp->mp_pgno;
4220 unsigned i, ovpages = mp->mp_pages;
4221 MDB_env *env = txn->mt_env;
4224 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4225 /* If the page is dirty we just acquired it, so we should
4226 * give it back to our current free list, if any.
4227 * Not currently supported in nested txns.
4228 * Otherwise put it onto the list of pages we freed in this txn.
4230 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && env->me_pghead) {
4233 MDB_ID2 *dl, ix, iy;
4234 rc = mdb_midl_need(&env->me_pghead, ovpages);
4237 /* Remove from dirty list */
4238 dl = txn->mt_u.dirty_list;
4240 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4248 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4249 txn->mt_flags |= MDB_TXN_ERROR;
4250 return MDB_CORRUPTED;
4253 if (!(env->me_flags & MDB_WRITEMAP))
4254 mdb_dpage_free(env, mp);
4255 /* Insert in me_pghead */
4256 mop = env->me_pghead;
4257 j = mop[0] + ovpages;
4258 for (i = mop[0]; i && mop[i] < pg; i--)
4264 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4268 mc->mc_db->md_overflow_pages -= ovpages;
4272 /** Return the data associated with a given node.
4273 * @param[in] txn The transaction for this operation.
4274 * @param[in] leaf The node being read.
4275 * @param[out] data Updated to point to the node's data.
4276 * @return 0 on success, non-zero on failure.
4279 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4281 MDB_page *omp; /* overflow page */
4285 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4286 data->mv_size = NODEDSZ(leaf);
4287 data->mv_data = NODEDATA(leaf);
4291 /* Read overflow data.
4293 data->mv_size = NODEDSZ(leaf);
4294 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4295 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4296 DPRINTF("read overflow page %zu failed", pgno);
4299 data->mv_data = METADATA(omp);
4305 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4306 MDB_val *key, MDB_val *data)
4315 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4317 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4320 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4324 mdb_cursor_init(&mc, txn, dbi, &mx);
4325 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4328 /** Find a sibling for a page.
4329 * Replaces the page at the top of the cursor's stack with the
4330 * specified sibling, if one exists.
4331 * @param[in] mc The cursor for this operation.
4332 * @param[in] move_right Non-zero if the right sibling is requested,
4333 * otherwise the left sibling.
4334 * @return 0 on success, non-zero on failure.
4337 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4343 if (mc->mc_snum < 2) {
4344 return MDB_NOTFOUND; /* root has no siblings */
4348 DPRINTF("parent page is page %zu, index %u",
4349 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4351 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4352 : (mc->mc_ki[mc->mc_top] == 0)) {
4353 DPRINTF("no more keys left, moving to %s sibling",
4354 move_right ? "right" : "left");
4355 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4356 /* undo cursor_pop before returning */
4363 mc->mc_ki[mc->mc_top]++;
4365 mc->mc_ki[mc->mc_top]--;
4366 DPRINTF("just moving to %s index key %u",
4367 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4369 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4371 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4372 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4375 mdb_cursor_push(mc, mp);
4377 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4382 /** Move the cursor to the next data item. */
4384 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4390 if (mc->mc_flags & C_EOF) {
4391 return MDB_NOTFOUND;
4394 assert(mc->mc_flags & C_INITIALIZED);
4396 mp = mc->mc_pg[mc->mc_top];
4398 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4399 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4400 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4401 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4402 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4403 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4407 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4408 if (op == MDB_NEXT_DUP)
4409 return MDB_NOTFOUND;
4413 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4415 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4416 DPUTS("=====> move to next sibling page");
4417 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4418 mc->mc_flags |= C_EOF;
4419 mc->mc_flags &= ~C_INITIALIZED;
4422 mp = mc->mc_pg[mc->mc_top];
4423 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4425 mc->mc_ki[mc->mc_top]++;
4427 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4428 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4431 key->mv_size = mc->mc_db->md_pad;
4432 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4436 assert(IS_LEAF(mp));
4437 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4439 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4440 mdb_xcursor_init1(mc, leaf);
4443 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4446 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4447 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4448 if (rc != MDB_SUCCESS)
4453 MDB_GET_KEY(leaf, key);
4457 /** Move the cursor to the previous data item. */
4459 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4465 assert(mc->mc_flags & C_INITIALIZED);
4467 mp = mc->mc_pg[mc->mc_top];
4469 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4470 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4471 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4472 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4473 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4474 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4477 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4478 if (op == MDB_PREV_DUP)
4479 return MDB_NOTFOUND;
4484 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4486 if (mc->mc_ki[mc->mc_top] == 0) {
4487 DPUTS("=====> move to prev sibling page");
4488 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
4489 mc->mc_flags &= ~C_INITIALIZED;
4492 mp = mc->mc_pg[mc->mc_top];
4493 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4494 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4496 mc->mc_ki[mc->mc_top]--;
4498 mc->mc_flags &= ~C_EOF;
4500 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4501 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4504 key->mv_size = mc->mc_db->md_pad;
4505 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4509 assert(IS_LEAF(mp));
4510 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4512 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4513 mdb_xcursor_init1(mc, leaf);
4516 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4519 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4520 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4521 if (rc != MDB_SUCCESS)
4526 MDB_GET_KEY(leaf, key);
4530 /** Set the cursor on a specific data item. */
4532 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4533 MDB_cursor_op op, int *exactp)
4537 MDB_node *leaf = NULL;
4542 assert(key->mv_size > 0);
4544 /* See if we're already on the right page */
4545 if (mc->mc_flags & C_INITIALIZED) {
4548 mp = mc->mc_pg[mc->mc_top];
4550 mc->mc_ki[mc->mc_top] = 0;
4551 return MDB_NOTFOUND;
4553 if (mp->mp_flags & P_LEAF2) {
4554 nodekey.mv_size = mc->mc_db->md_pad;
4555 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4557 leaf = NODEPTR(mp, 0);
4558 MDB_GET_KEY(leaf, &nodekey);
4560 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4562 /* Probably happens rarely, but first node on the page
4563 * was the one we wanted.
4565 mc->mc_ki[mc->mc_top] = 0;
4572 unsigned int nkeys = NUMKEYS(mp);
4574 if (mp->mp_flags & P_LEAF2) {
4575 nodekey.mv_data = LEAF2KEY(mp,
4576 nkeys-1, nodekey.mv_size);
4578 leaf = NODEPTR(mp, nkeys-1);
4579 MDB_GET_KEY(leaf, &nodekey);
4581 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4583 /* last node was the one we wanted */
4584 mc->mc_ki[mc->mc_top] = nkeys-1;
4590 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4591 /* This is definitely the right page, skip search_page */
4592 if (mp->mp_flags & P_LEAF2) {
4593 nodekey.mv_data = LEAF2KEY(mp,
4594 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4596 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4597 MDB_GET_KEY(leaf, &nodekey);
4599 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4601 /* current node was the one we wanted */
4611 /* If any parents have right-sibs, search.
4612 * Otherwise, there's nothing further.
4614 for (i=0; i<mc->mc_top; i++)
4616 NUMKEYS(mc->mc_pg[i])-1)
4618 if (i == mc->mc_top) {
4619 /* There are no other pages */
4620 mc->mc_ki[mc->mc_top] = nkeys;
4621 return MDB_NOTFOUND;
4625 /* There are no other pages */
4626 mc->mc_ki[mc->mc_top] = 0;
4627 return MDB_NOTFOUND;
4631 rc = mdb_page_search(mc, key, 0);
4632 if (rc != MDB_SUCCESS)
4635 mp = mc->mc_pg[mc->mc_top];
4636 assert(IS_LEAF(mp));
4639 leaf = mdb_node_search(mc, key, exactp);
4640 if (exactp != NULL && !*exactp) {
4641 /* MDB_SET specified and not an exact match. */
4642 return MDB_NOTFOUND;
4646 DPUTS("===> inexact leaf not found, goto sibling");
4647 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4648 return rc; /* no entries matched */
4649 mp = mc->mc_pg[mc->mc_top];
4650 assert(IS_LEAF(mp));
4651 leaf = NODEPTR(mp, 0);
4655 mc->mc_flags |= C_INITIALIZED;
4656 mc->mc_flags &= ~C_EOF;
4659 key->mv_size = mc->mc_db->md_pad;
4660 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4664 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4665 mdb_xcursor_init1(mc, leaf);
4668 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4669 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4670 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4673 if (op == MDB_GET_BOTH) {
4679 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4680 if (rc != MDB_SUCCESS)
4683 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4685 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4687 rc = mc->mc_dbx->md_dcmp(data, &d2);
4689 if (op == MDB_GET_BOTH || rc > 0)
4690 return MDB_NOTFOUND;
4695 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4696 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4701 /* The key already matches in all other cases */
4702 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4703 MDB_GET_KEY(leaf, key);
4704 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4709 /** Move the cursor to the first item in the database. */
4711 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4716 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4717 rc = mdb_page_search(mc, NULL, 0);
4718 if (rc != MDB_SUCCESS)
4721 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4723 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4724 mc->mc_flags |= C_INITIALIZED;
4725 mc->mc_flags &= ~C_EOF;
4727 mc->mc_ki[mc->mc_top] = 0;
4729 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4730 key->mv_size = mc->mc_db->md_pad;
4731 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4736 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4737 mdb_xcursor_init1(mc, leaf);
4738 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4743 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4744 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4748 MDB_GET_KEY(leaf, key);
4752 /** Move the cursor to the last item in the database. */
4754 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4759 if (!(mc->mc_flags & C_EOF)) {
4761 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4764 lkey.mv_size = MDB_MAXKEYSIZE+1;
4765 lkey.mv_data = NULL;
4766 rc = mdb_page_search(mc, &lkey, 0);
4767 if (rc != MDB_SUCCESS)
4770 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4773 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4774 mc->mc_flags |= C_INITIALIZED|C_EOF;
4775 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4777 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4778 key->mv_size = mc->mc_db->md_pad;
4779 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4784 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4785 mdb_xcursor_init1(mc, leaf);
4786 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4791 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4792 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4797 MDB_GET_KEY(leaf, key);
4802 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4811 case MDB_GET_CURRENT:
4812 if (!(mc->mc_flags & C_INITIALIZED)) {
4815 MDB_page *mp = mc->mc_pg[mc->mc_top];
4817 mc->mc_ki[mc->mc_top] = 0;
4823 key->mv_size = mc->mc_db->md_pad;
4824 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4826 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4827 MDB_GET_KEY(leaf, key);
4829 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4830 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4832 rc = mdb_node_read(mc->mc_txn, leaf, data);
4839 case MDB_GET_BOTH_RANGE:
4840 if (data == NULL || mc->mc_xcursor == NULL) {
4848 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4850 } else if (op == MDB_SET_RANGE)
4851 rc = mdb_cursor_set(mc, key, data, op, NULL);
4853 rc = mdb_cursor_set(mc, key, data, op, &exact);
4855 case MDB_GET_MULTIPLE:
4857 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4858 !(mc->mc_flags & C_INITIALIZED)) {
4863 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4864 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4867 case MDB_NEXT_MULTIPLE:
4869 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4873 if (!(mc->mc_flags & C_INITIALIZED))
4874 rc = mdb_cursor_first(mc, key, data);
4876 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4877 if (rc == MDB_SUCCESS) {
4878 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4881 mx = &mc->mc_xcursor->mx_cursor;
4882 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4884 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4885 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4893 case MDB_NEXT_NODUP:
4894 if (!(mc->mc_flags & C_INITIALIZED))
4895 rc = mdb_cursor_first(mc, key, data);
4897 rc = mdb_cursor_next(mc, key, data, op);
4901 case MDB_PREV_NODUP:
4902 if (!(mc->mc_flags & C_INITIALIZED)) {
4903 rc = mdb_cursor_last(mc, key, data);
4906 mc->mc_flags |= C_INITIALIZED;
4907 mc->mc_ki[mc->mc_top]++;
4909 rc = mdb_cursor_prev(mc, key, data, op);
4912 rc = mdb_cursor_first(mc, key, data);
4916 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4917 !(mc->mc_flags & C_INITIALIZED) ||
4918 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4922 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4925 rc = mdb_cursor_last(mc, key, data);
4929 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4930 !(mc->mc_flags & C_INITIALIZED) ||
4931 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4935 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4938 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4946 /** Touch all the pages in the cursor stack.
4947 * Makes sure all the pages are writable, before attempting a write operation.
4948 * @param[in] mc The cursor to operate on.
4951 mdb_cursor_touch(MDB_cursor *mc)
4955 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4958 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
4959 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4962 *mc->mc_dbflag |= DB_DIRTY;
4964 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4965 rc = mdb_page_touch(mc);
4969 mc->mc_top = mc->mc_snum-1;
4974 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4977 MDB_node *leaf = NULL;
4978 MDB_val xdata, *rdata, dkey;
4981 int do_sub = 0, insert = 0;
4982 unsigned int mcount = 0;
4986 char dbuf[MDB_MAXKEYSIZE+1];
4987 unsigned int nflags;
4990 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4993 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4996 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4999 #if SIZE_MAX > MAXDATASIZE
5000 if (data->mv_size > MAXDATASIZE)
5004 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5005 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5009 if (flags == MDB_CURRENT) {
5010 if (!(mc->mc_flags & C_INITIALIZED))
5013 } else if (mc->mc_db->md_root == P_INVALID) {
5015 /* new database, write a root leaf page */
5016 DPUTS("allocating new root leaf page");
5017 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5021 mdb_cursor_push(mc, np);
5022 mc->mc_db->md_root = np->mp_pgno;
5023 mc->mc_db->md_depth++;
5024 *mc->mc_dbflag |= DB_DIRTY;
5025 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5027 np->mp_flags |= P_LEAF2;
5028 mc->mc_flags |= C_INITIALIZED;
5034 if (flags & MDB_APPEND) {
5036 rc = mdb_cursor_last(mc, &k2, &d2);
5038 rc = mc->mc_dbx->md_cmp(key, &k2);
5041 mc->mc_ki[mc->mc_top]++;
5043 /* new key is <= last key */
5048 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5050 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5051 DPRINTF("duplicate key [%s]", DKEY(key));
5053 return MDB_KEYEXIST;
5055 if (rc && rc != MDB_NOTFOUND)
5059 /* Cursor is positioned, now make sure all pages are writable */
5060 rc2 = mdb_cursor_touch(mc);
5065 /* The key already exists */
5066 if (rc == MDB_SUCCESS) {
5067 /* there's only a key anyway, so this is a no-op */
5068 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5069 unsigned int ksize = mc->mc_db->md_pad;
5070 if (key->mv_size != ksize)
5072 if (flags == MDB_CURRENT) {
5073 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5074 memcpy(ptr, key->mv_data, ksize);
5079 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5082 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5083 /* Was a single item before, must convert now */
5085 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5086 /* Just overwrite the current item */
5087 if (flags == MDB_CURRENT)
5090 dkey.mv_size = NODEDSZ(leaf);
5091 dkey.mv_data = NODEDATA(leaf);
5092 #if UINT_MAX < SIZE_MAX
5093 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5094 #ifdef MISALIGNED_OK
5095 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5097 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5100 /* if data matches, ignore it */
5101 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5102 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5104 /* create a fake page for the dup items */
5105 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5106 dkey.mv_data = dbuf;
5107 fp = (MDB_page *)&pbuf;
5108 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5109 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5110 fp->mp_lower = PAGEHDRSZ;
5111 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5112 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5113 fp->mp_flags |= P_LEAF2;
5114 fp->mp_pad = data->mv_size;
5115 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5117 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5118 (dkey.mv_size & 1) + (data->mv_size & 1);
5120 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5123 xdata.mv_size = fp->mp_upper;
5128 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5129 /* See if we need to convert from fake page to subDB */
5131 unsigned int offset;
5135 fp = NODEDATA(leaf);
5136 if (flags == MDB_CURRENT) {
5138 fp->mp_flags |= P_DIRTY;
5139 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5140 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5144 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5145 offset = fp->mp_pad;
5146 if (SIZELEFT(fp) >= offset)
5148 offset *= 4; /* space for 4 more */
5150 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5152 offset += offset & 1;
5153 fp_flags = fp->mp_flags;
5154 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5155 offset >= mc->mc_txn->mt_env->me_nodemax) {
5156 /* yes, convert it */
5158 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5159 dummy.md_pad = fp->mp_pad;
5160 dummy.md_flags = MDB_DUPFIXED;
5161 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5162 dummy.md_flags |= MDB_INTEGERKEY;
5165 dummy.md_branch_pages = 0;
5166 dummy.md_leaf_pages = 1;
5167 dummy.md_overflow_pages = 0;
5168 dummy.md_entries = NUMKEYS(fp);
5170 xdata.mv_size = sizeof(MDB_db);
5171 xdata.mv_data = &dummy;
5172 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5174 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5175 flags |= F_DUPDATA|F_SUBDATA;
5176 dummy.md_root = mp->mp_pgno;
5177 fp_flags &= ~P_SUBP;
5179 /* no, just grow it */
5181 xdata.mv_size = NODEDSZ(leaf) + offset;
5182 xdata.mv_data = &pbuf;
5183 mp = (MDB_page *)&pbuf;
5184 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5187 mp->mp_flags = fp_flags | P_DIRTY;
5188 mp->mp_pad = fp->mp_pad;
5189 mp->mp_lower = fp->mp_lower;
5190 mp->mp_upper = fp->mp_upper + offset;
5192 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5194 nsize = NODEDSZ(leaf) - fp->mp_upper;
5195 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5196 for (i=0; i<NUMKEYS(fp); i++)
5197 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5199 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5203 /* data is on sub-DB, just store it */
5204 flags |= F_DUPDATA|F_SUBDATA;
5208 /* overflow page overwrites need special handling */
5209 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5212 unsigned psize = mc->mc_txn->mt_env->me_psize;
5213 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5215 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5216 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5218 ovpages = omp->mp_pages;
5220 /* Is the ov page writable and large enough? */
5221 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5222 /* yes, overwrite it. Note in this case we don't
5223 * bother to try shrinking the page if the new data
5224 * is smaller than the overflow threshold.
5227 /* It is writable only in a parent txn */
5228 size_t sz = (size_t) psize * ovpages, off;
5229 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5235 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5236 if (!(flags & MDB_RESERVE)) {
5237 /* Copy end of page, adjusting alignment so
5238 * compiler may copy words instead of bytes.
5240 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5241 memcpy((size_t *)((char *)np + off),
5242 (size_t *)((char *)omp + off), sz - off);
5245 memcpy(np, omp, sz); /* Copy beginning of page */
5248 SETDSZ(leaf, data->mv_size);
5249 if (F_ISSET(flags, MDB_RESERVE))
5250 data->mv_data = METADATA(omp);
5252 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5255 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5258 } else if (NODEDSZ(leaf) == data->mv_size) {
5259 /* same size, just replace it. Note that we could
5260 * also reuse this node if the new data is smaller,
5261 * but instead we opt to shrink the node in that case.
5263 if (F_ISSET(flags, MDB_RESERVE))
5264 data->mv_data = NODEDATA(leaf);
5265 else if (data->mv_size)
5266 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5268 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5271 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5272 mc->mc_db->md_entries--;
5274 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5281 nflags = flags & NODE_ADD_FLAGS;
5282 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5283 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5284 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5285 nflags &= ~MDB_APPEND;
5287 nflags |= MDB_SPLIT_REPLACE;
5288 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5290 /* There is room already in this leaf page. */
5291 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5292 if (rc == 0 && !do_sub && insert) {
5293 /* Adjust other cursors pointing to mp */
5294 MDB_cursor *m2, *m3;
5295 MDB_dbi dbi = mc->mc_dbi;
5296 unsigned i = mc->mc_top;
5297 MDB_page *mp = mc->mc_pg[i];
5299 if (mc->mc_flags & C_SUB)
5302 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5303 if (mc->mc_flags & C_SUB)
5304 m3 = &m2->mc_xcursor->mx_cursor;
5307 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5308 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5315 if (rc != MDB_SUCCESS)
5316 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5318 /* Now store the actual data in the child DB. Note that we're
5319 * storing the user data in the keys field, so there are strict
5320 * size limits on dupdata. The actual data fields of the child
5321 * DB are all zero size.
5328 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5329 if (flags & MDB_CURRENT) {
5330 xflags = MDB_CURRENT;
5332 mdb_xcursor_init1(mc, leaf);
5333 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5335 /* converted, write the original data first */
5337 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5341 /* Adjust other cursors pointing to mp */
5343 unsigned i = mc->mc_top;
5344 MDB_page *mp = mc->mc_pg[i];
5346 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5347 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5348 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5349 mdb_xcursor_init1(m2, leaf);
5353 /* we've done our job */
5356 if (flags & MDB_APPENDDUP)
5357 xflags |= MDB_APPEND;
5358 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5359 if (flags & F_SUBDATA) {
5360 void *db = NODEDATA(leaf);
5361 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5364 /* sub-writes might have failed so check rc again.
5365 * Don't increment count if we just replaced an existing item.
5367 if (!rc && !(flags & MDB_CURRENT))
5368 mc->mc_db->md_entries++;
5369 if (flags & MDB_MULTIPLE) {
5371 if (mcount < data[1].mv_size) {
5372 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5373 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5379 /* If we succeeded and the key didn't exist before, make sure
5380 * the cursor is marked valid.
5383 mc->mc_flags |= C_INITIALIZED;
5388 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5393 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5396 if (!(mc->mc_flags & C_INITIALIZED))
5399 rc = mdb_cursor_touch(mc);
5403 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5405 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5406 if (flags != MDB_NODUPDATA) {
5407 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5408 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5410 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5411 /* If sub-DB still has entries, we're done */
5412 if (mc->mc_xcursor->mx_db.md_entries) {
5413 if (leaf->mn_flags & F_SUBDATA) {
5414 /* update subDB info */
5415 void *db = NODEDATA(leaf);
5416 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5419 /* shrink fake page */
5420 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5421 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5422 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5423 /* fix other sub-DB cursors pointed at this fake page */
5424 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5425 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5426 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5427 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5428 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5431 mc->mc_db->md_entries--;
5434 /* otherwise fall thru and delete the sub-DB */
5437 if (leaf->mn_flags & F_SUBDATA) {
5438 /* add all the child DB's pages to the free list */
5439 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5440 if (rc == MDB_SUCCESS) {
5441 mc->mc_db->md_entries -=
5442 mc->mc_xcursor->mx_db.md_entries;
5447 return mdb_cursor_del0(mc, leaf);
5450 /** Allocate and initialize new pages for a database.
5451 * @param[in] mc a cursor on the database being added to.
5452 * @param[in] flags flags defining what type of page is being allocated.
5453 * @param[in] num the number of pages to allocate. This is usually 1,
5454 * unless allocating overflow pages for a large record.
5455 * @param[out] mp Address of a page, or NULL on failure.
5456 * @return 0 on success, non-zero on failure.
5459 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5464 if ((rc = mdb_page_alloc(mc, num, &np)))
5466 DPRINTF("allocated new mpage %zu, page size %u",
5467 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5468 np->mp_flags = flags | P_DIRTY;
5469 np->mp_lower = PAGEHDRSZ;
5470 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5473 mc->mc_db->md_branch_pages++;
5474 else if (IS_LEAF(np))
5475 mc->mc_db->md_leaf_pages++;
5476 else if (IS_OVERFLOW(np)) {
5477 mc->mc_db->md_overflow_pages += num;
5485 /** Calculate the size of a leaf node.
5486 * The size depends on the environment's page size; if a data item
5487 * is too large it will be put onto an overflow page and the node
5488 * size will only include the key and not the data. Sizes are always
5489 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5490 * of the #MDB_node headers.
5491 * @param[in] env The environment handle.
5492 * @param[in] key The key for the node.
5493 * @param[in] data The data for the node.
5494 * @return The number of bytes needed to store the node.
5497 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5501 sz = LEAFSIZE(key, data);
5502 if (sz >= env->me_nodemax) {
5503 /* put on overflow page */
5504 sz -= data->mv_size - sizeof(pgno_t);
5508 return sz + sizeof(indx_t);
5511 /** Calculate the size of a branch node.
5512 * The size should depend on the environment's page size but since
5513 * we currently don't support spilling large keys onto overflow
5514 * pages, it's simply the size of the #MDB_node header plus the
5515 * size of the key. Sizes are always rounded up to an even number
5516 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5517 * @param[in] env The environment handle.
5518 * @param[in] key The key for the node.
5519 * @return The number of bytes needed to store the node.
5522 mdb_branch_size(MDB_env *env, MDB_val *key)
5527 if (sz >= env->me_nodemax) {
5528 /* put on overflow page */
5529 /* not implemented */
5530 /* sz -= key->size - sizeof(pgno_t); */
5533 return sz + sizeof(indx_t);
5536 /** Add a node to the page pointed to by the cursor.
5537 * @param[in] mc The cursor for this operation.
5538 * @param[in] indx The index on the page where the new node should be added.
5539 * @param[in] key The key for the new node.
5540 * @param[in] data The data for the new node, if any.
5541 * @param[in] pgno The page number, if adding a branch node.
5542 * @param[in] flags Flags for the node.
5543 * @return 0 on success, non-zero on failure. Possible errors are:
5545 * <li>ENOMEM - failed to allocate overflow pages for the node.
5546 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5547 * should never happen since all callers already calculate the
5548 * page's free space before calling this function.
5552 mdb_node_add(MDB_cursor *mc, indx_t indx,
5553 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5556 size_t node_size = NODESIZE;
5559 MDB_page *mp = mc->mc_pg[mc->mc_top];
5560 MDB_page *ofp = NULL; /* overflow page */
5563 assert(mp->mp_upper >= mp->mp_lower);
5565 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5566 IS_LEAF(mp) ? "leaf" : "branch",
5567 IS_SUBP(mp) ? "sub-" : "",
5568 mp->mp_pgno, indx, data ? data->mv_size : 0,
5569 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5572 /* Move higher keys up one slot. */
5573 int ksize = mc->mc_db->md_pad, dif;
5574 char *ptr = LEAF2KEY(mp, indx, ksize);
5575 dif = NUMKEYS(mp) - indx;
5577 memmove(ptr+ksize, ptr, dif*ksize);
5578 /* insert new key */
5579 memcpy(ptr, key->mv_data, ksize);
5581 /* Just using these for counting */
5582 mp->mp_lower += sizeof(indx_t);
5583 mp->mp_upper -= ksize - sizeof(indx_t);
5588 node_size += key->mv_size;
5592 if (F_ISSET(flags, F_BIGDATA)) {
5593 /* Data already on overflow page. */
5594 node_size += sizeof(pgno_t);
5595 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5596 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5598 /* Put data on overflow page. */
5599 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5600 data->mv_size, node_size+data->mv_size);
5601 node_size += sizeof(pgno_t);
5602 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5604 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5607 node_size += data->mv_size;
5610 node_size += node_size & 1;
5612 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5613 DPRINTF("not enough room in page %zu, got %u ptrs",
5614 mp->mp_pgno, NUMKEYS(mp));
5615 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5616 mp->mp_upper - mp->mp_lower);
5617 DPRINTF("node size = %zu", node_size);
5618 return MDB_PAGE_FULL;
5621 /* Move higher pointers up one slot. */
5622 for (i = NUMKEYS(mp); i > indx; i--)
5623 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5625 /* Adjust free space offsets. */
5626 ofs = mp->mp_upper - node_size;
5627 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5628 mp->mp_ptrs[indx] = ofs;
5630 mp->mp_lower += sizeof(indx_t);
5632 /* Write the node data. */
5633 node = NODEPTR(mp, indx);
5634 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5635 node->mn_flags = flags;
5637 SETDSZ(node,data->mv_size);
5642 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5647 if (F_ISSET(flags, F_BIGDATA))
5648 memcpy(node->mn_data + key->mv_size, data->mv_data,
5650 else if (F_ISSET(flags, MDB_RESERVE))
5651 data->mv_data = node->mn_data + key->mv_size;
5653 memcpy(node->mn_data + key->mv_size, data->mv_data,
5656 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5658 if (F_ISSET(flags, MDB_RESERVE))
5659 data->mv_data = METADATA(ofp);
5661 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5668 /** Delete the specified node from a page.
5669 * @param[in] mp The page to operate on.
5670 * @param[in] indx The index of the node to delete.
5671 * @param[in] ksize The size of a node. Only used if the page is
5672 * part of a #MDB_DUPFIXED database.
5675 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5678 indx_t i, j, numkeys, ptr;
5685 COPY_PGNO(pgno, mp->mp_pgno);
5686 DPRINTF("delete node %u on %s page %zu", indx,
5687 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5690 assert(indx < NUMKEYS(mp));
5693 int x = NUMKEYS(mp) - 1 - indx;
5694 base = LEAF2KEY(mp, indx, ksize);
5696 memmove(base, base + ksize, x * ksize);
5697 mp->mp_lower -= sizeof(indx_t);
5698 mp->mp_upper += ksize - sizeof(indx_t);
5702 node = NODEPTR(mp, indx);
5703 sz = NODESIZE + node->mn_ksize;
5705 if (F_ISSET(node->mn_flags, F_BIGDATA))
5706 sz += sizeof(pgno_t);
5708 sz += NODEDSZ(node);
5712 ptr = mp->mp_ptrs[indx];
5713 numkeys = NUMKEYS(mp);
5714 for (i = j = 0; i < numkeys; i++) {
5716 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5717 if (mp->mp_ptrs[i] < ptr)
5718 mp->mp_ptrs[j] += sz;
5723 base = (char *)mp + mp->mp_upper;
5724 memmove(base + sz, base, ptr - mp->mp_upper);
5726 mp->mp_lower -= sizeof(indx_t);
5730 /** Compact the main page after deleting a node on a subpage.
5731 * @param[in] mp The main page to operate on.
5732 * @param[in] indx The index of the subpage on the main page.
5735 mdb_node_shrink(MDB_page *mp, indx_t indx)
5742 indx_t i, numkeys, ptr;
5744 node = NODEPTR(mp, indx);
5745 sp = (MDB_page *)NODEDATA(node);
5746 osize = NODEDSZ(node);
5748 delta = sp->mp_upper - sp->mp_lower;
5749 SETDSZ(node, osize - delta);
5750 xp = (MDB_page *)((char *)sp + delta);
5752 /* shift subpage upward */
5754 nsize = NUMKEYS(sp) * sp->mp_pad;
5755 memmove(METADATA(xp), METADATA(sp), nsize);
5758 nsize = osize - sp->mp_upper;
5759 numkeys = NUMKEYS(sp);
5760 for (i=numkeys-1; i>=0; i--)
5761 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5763 xp->mp_upper = sp->mp_lower;
5764 xp->mp_lower = sp->mp_lower;
5765 xp->mp_flags = sp->mp_flags;
5766 xp->mp_pad = sp->mp_pad;
5767 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5769 /* shift lower nodes upward */
5770 ptr = mp->mp_ptrs[indx];
5771 numkeys = NUMKEYS(mp);
5772 for (i = 0; i < numkeys; i++) {
5773 if (mp->mp_ptrs[i] <= ptr)
5774 mp->mp_ptrs[i] += delta;
5777 base = (char *)mp + mp->mp_upper;
5778 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5779 mp->mp_upper += delta;
5782 /** Initial setup of a sorted-dups cursor.
5783 * Sorted duplicates are implemented as a sub-database for the given key.
5784 * The duplicate data items are actually keys of the sub-database.
5785 * Operations on the duplicate data items are performed using a sub-cursor
5786 * initialized when the sub-database is first accessed. This function does
5787 * the preliminary setup of the sub-cursor, filling in the fields that
5788 * depend only on the parent DB.
5789 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5792 mdb_xcursor_init0(MDB_cursor *mc)
5794 MDB_xcursor *mx = mc->mc_xcursor;
5796 mx->mx_cursor.mc_xcursor = NULL;
5797 mx->mx_cursor.mc_txn = mc->mc_txn;
5798 mx->mx_cursor.mc_db = &mx->mx_db;
5799 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5800 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5801 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5802 mx->mx_cursor.mc_snum = 0;
5803 mx->mx_cursor.mc_top = 0;
5804 mx->mx_cursor.mc_flags = C_SUB;
5805 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5806 mx->mx_dbx.md_dcmp = NULL;
5807 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5810 /** Final setup of a sorted-dups cursor.
5811 * Sets up the fields that depend on the data from the main cursor.
5812 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5813 * @param[in] node The data containing the #MDB_db record for the
5814 * sorted-dup database.
5817 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5819 MDB_xcursor *mx = mc->mc_xcursor;
5821 if (node->mn_flags & F_SUBDATA) {
5822 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5823 mx->mx_cursor.mc_pg[0] = 0;
5824 mx->mx_cursor.mc_snum = 0;
5825 mx->mx_cursor.mc_flags = C_SUB;
5827 MDB_page *fp = NODEDATA(node);
5828 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5829 mx->mx_db.md_flags = 0;
5830 mx->mx_db.md_depth = 1;
5831 mx->mx_db.md_branch_pages = 0;
5832 mx->mx_db.md_leaf_pages = 1;
5833 mx->mx_db.md_overflow_pages = 0;
5834 mx->mx_db.md_entries = NUMKEYS(fp);
5835 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5836 mx->mx_cursor.mc_snum = 1;
5837 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5838 mx->mx_cursor.mc_top = 0;
5839 mx->mx_cursor.mc_pg[0] = fp;
5840 mx->mx_cursor.mc_ki[0] = 0;
5841 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5842 mx->mx_db.md_flags = MDB_DUPFIXED;
5843 mx->mx_db.md_pad = fp->mp_pad;
5844 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5845 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5848 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5850 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5852 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5853 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5854 #if UINT_MAX < SIZE_MAX
5855 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5856 #ifdef MISALIGNED_OK
5857 mx->mx_dbx.md_cmp = mdb_cmp_long;
5859 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5864 /** Initialize a cursor for a given transaction and database. */
5866 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5871 mc->mc_db = &txn->mt_dbs[dbi];
5872 mc->mc_dbx = &txn->mt_dbxs[dbi];
5873 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5878 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5880 mc->mc_xcursor = mx;
5881 mdb_xcursor_init0(mc);
5883 mc->mc_xcursor = NULL;
5885 if (*mc->mc_dbflag & DB_STALE) {
5886 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5891 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5894 size_t size = sizeof(MDB_cursor);
5896 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5899 /* Allow read access to the freelist */
5900 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5903 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5904 size += sizeof(MDB_xcursor);
5906 if ((mc = malloc(size)) != NULL) {
5907 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5908 if (txn->mt_cursors) {
5909 mc->mc_next = txn->mt_cursors[dbi];
5910 txn->mt_cursors[dbi] = mc;
5911 mc->mc_flags |= C_UNTRACK;
5913 mc->mc_flags |= C_ALLOCD;
5924 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5928 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5931 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5934 flags = mc->mc_flags;
5936 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5938 mc->mc_flags |= (flags & C_ALLOCD);
5942 /* Return the count of duplicate data items for the current key */
5944 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5948 if (mc == NULL || countp == NULL)
5951 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5954 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5955 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5958 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5961 *countp = mc->mc_xcursor->mx_db.md_entries;
5967 mdb_cursor_close(MDB_cursor *mc)
5970 /* remove from txn, if tracked */
5971 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
5972 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5973 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5975 *prev = mc->mc_next;
5977 if (mc->mc_flags & C_ALLOCD)
5983 mdb_cursor_txn(MDB_cursor *mc)
5985 if (!mc) return NULL;
5990 mdb_cursor_dbi(MDB_cursor *mc)
5996 /** Replace the key for a node with a new key.
5997 * @param[in] mc Cursor pointing to the node to operate on.
5998 * @param[in] key The new key to use.
5999 * @return 0 on success, non-zero on failure.
6002 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6009 indx_t ptr, i, numkeys, indx;
6012 indx = mc->mc_ki[mc->mc_top];
6013 mp = mc->mc_pg[mc->mc_top];
6014 node = NODEPTR(mp, indx);
6015 ptr = mp->mp_ptrs[indx];
6019 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6020 k2.mv_data = NODEKEY(node);
6021 k2.mv_size = node->mn_ksize;
6022 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6024 mdb_dkey(&k2, kbuf2),
6030 delta0 = delta = key->mv_size - node->mn_ksize;
6032 /* Must be 2-byte aligned. If new key is
6033 * shorter by 1, the shift will be skipped.
6035 delta += (delta & 1);
6037 if (delta > 0 && SIZELEFT(mp) < delta) {
6039 /* not enough space left, do a delete and split */
6040 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6041 pgno = NODEPGNO(node);
6042 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6043 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6046 numkeys = NUMKEYS(mp);
6047 for (i = 0; i < numkeys; i++) {
6048 if (mp->mp_ptrs[i] <= ptr)
6049 mp->mp_ptrs[i] -= delta;
6052 base = (char *)mp + mp->mp_upper;
6053 len = ptr - mp->mp_upper + NODESIZE;
6054 memmove(base - delta, base, len);
6055 mp->mp_upper -= delta;
6057 node = NODEPTR(mp, indx);
6060 /* But even if no shift was needed, update ksize */
6062 node->mn_ksize = key->mv_size;
6065 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6071 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6073 /** Move a node from csrc to cdst.
6076 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6083 unsigned short flags;
6087 /* Mark src and dst as dirty. */
6088 if ((rc = mdb_page_touch(csrc)) ||
6089 (rc = mdb_page_touch(cdst)))
6092 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6093 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6094 key.mv_size = csrc->mc_db->md_pad;
6095 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6097 data.mv_data = NULL;
6101 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6102 assert(!((long)srcnode&1));
6103 srcpg = NODEPGNO(srcnode);
6104 flags = srcnode->mn_flags;
6105 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6106 unsigned int snum = csrc->mc_snum;
6108 /* must find the lowest key below src */
6109 mdb_page_search_lowest(csrc);
6110 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6111 key.mv_size = csrc->mc_db->md_pad;
6112 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6114 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6115 key.mv_size = NODEKSZ(s2);
6116 key.mv_data = NODEKEY(s2);
6118 csrc->mc_snum = snum--;
6119 csrc->mc_top = snum;
6121 key.mv_size = NODEKSZ(srcnode);
6122 key.mv_data = NODEKEY(srcnode);
6124 data.mv_size = NODEDSZ(srcnode);
6125 data.mv_data = NODEDATA(srcnode);
6127 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6128 unsigned int snum = cdst->mc_snum;
6131 /* must find the lowest key below dst */
6132 mdb_page_search_lowest(cdst);
6133 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6134 bkey.mv_size = cdst->mc_db->md_pad;
6135 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6137 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6138 bkey.mv_size = NODEKSZ(s2);
6139 bkey.mv_data = NODEKEY(s2);
6141 cdst->mc_snum = snum--;
6142 cdst->mc_top = snum;
6143 mdb_cursor_copy(cdst, &mn);
6145 rc = mdb_update_key(&mn, &bkey);
6150 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6151 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6152 csrc->mc_ki[csrc->mc_top],
6154 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6155 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6157 /* Add the node to the destination page.
6159 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6160 if (rc != MDB_SUCCESS)
6163 /* Delete the node from the source page.
6165 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6168 /* Adjust other cursors pointing to mp */
6169 MDB_cursor *m2, *m3;
6170 MDB_dbi dbi = csrc->mc_dbi;
6171 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6173 if (csrc->mc_flags & C_SUB)
6176 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6177 if (csrc->mc_flags & C_SUB)
6178 m3 = &m2->mc_xcursor->mx_cursor;
6181 if (m3 == csrc) continue;
6182 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6183 csrc->mc_ki[csrc->mc_top]) {
6184 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6185 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6190 /* Update the parent separators.
6192 if (csrc->mc_ki[csrc->mc_top] == 0) {
6193 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6194 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6195 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6197 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6198 key.mv_size = NODEKSZ(srcnode);
6199 key.mv_data = NODEKEY(srcnode);
6201 DPRINTF("update separator for source page %zu to [%s]",
6202 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6203 mdb_cursor_copy(csrc, &mn);
6206 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6209 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6211 indx_t ix = csrc->mc_ki[csrc->mc_top];
6212 nullkey.mv_size = 0;
6213 csrc->mc_ki[csrc->mc_top] = 0;
6214 rc = mdb_update_key(csrc, &nullkey);
6215 csrc->mc_ki[csrc->mc_top] = ix;
6216 assert(rc == MDB_SUCCESS);
6220 if (cdst->mc_ki[cdst->mc_top] == 0) {
6221 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6222 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6223 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6225 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6226 key.mv_size = NODEKSZ(srcnode);
6227 key.mv_data = NODEKEY(srcnode);
6229 DPRINTF("update separator for destination page %zu to [%s]",
6230 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6231 mdb_cursor_copy(cdst, &mn);
6234 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6237 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6239 indx_t ix = cdst->mc_ki[cdst->mc_top];
6240 nullkey.mv_size = 0;
6241 cdst->mc_ki[cdst->mc_top] = 0;
6242 rc = mdb_update_key(cdst, &nullkey);
6243 cdst->mc_ki[cdst->mc_top] = ix;
6244 assert(rc == MDB_SUCCESS);
6251 /** Merge one page into another.
6252 * The nodes from the page pointed to by \b csrc will
6253 * be copied to the page pointed to by \b cdst and then
6254 * the \b csrc page will be freed.
6255 * @param[in] csrc Cursor pointing to the source page.
6256 * @param[in] cdst Cursor pointing to the destination page.
6259 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6267 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6268 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6270 assert(csrc->mc_snum > 1); /* can't merge root page */
6271 assert(cdst->mc_snum > 1);
6273 /* Mark dst as dirty. */
6274 if ((rc = mdb_page_touch(cdst)))
6277 /* Move all nodes from src to dst.
6279 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6280 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6281 key.mv_size = csrc->mc_db->md_pad;
6282 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6283 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6284 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6285 if (rc != MDB_SUCCESS)
6287 key.mv_data = (char *)key.mv_data + key.mv_size;
6290 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6291 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6292 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6293 unsigned int snum = csrc->mc_snum;
6295 /* must find the lowest key below src */
6296 mdb_page_search_lowest(csrc);
6297 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6298 key.mv_size = csrc->mc_db->md_pad;
6299 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6301 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6302 key.mv_size = NODEKSZ(s2);
6303 key.mv_data = NODEKEY(s2);
6305 csrc->mc_snum = snum--;
6306 csrc->mc_top = snum;
6308 key.mv_size = srcnode->mn_ksize;
6309 key.mv_data = NODEKEY(srcnode);
6312 data.mv_size = NODEDSZ(srcnode);
6313 data.mv_data = NODEDATA(srcnode);
6314 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6315 if (rc != MDB_SUCCESS)
6320 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6321 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);
6323 /* Unlink the src page from parent and add to free list.
6325 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6326 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6329 rc = mdb_update_key(csrc, &key);
6335 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6336 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6339 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6340 csrc->mc_db->md_leaf_pages--;
6342 csrc->mc_db->md_branch_pages--;
6344 /* Adjust other cursors pointing to mp */
6345 MDB_cursor *m2, *m3;
6346 MDB_dbi dbi = csrc->mc_dbi;
6347 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6349 if (csrc->mc_flags & C_SUB)
6352 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6353 if (csrc->mc_flags & C_SUB)
6354 m3 = &m2->mc_xcursor->mx_cursor;
6357 if (m3 == csrc) continue;
6358 if (m3->mc_snum < csrc->mc_snum) continue;
6359 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6360 m3->mc_pg[csrc->mc_top] = mp;
6361 m3->mc_ki[csrc->mc_top] += nkeys;
6365 mdb_cursor_pop(csrc);
6367 return mdb_rebalance(csrc);
6370 /** Copy the contents of a cursor.
6371 * @param[in] csrc The cursor to copy from.
6372 * @param[out] cdst The cursor to copy to.
6375 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6379 cdst->mc_txn = csrc->mc_txn;
6380 cdst->mc_dbi = csrc->mc_dbi;
6381 cdst->mc_db = csrc->mc_db;
6382 cdst->mc_dbx = csrc->mc_dbx;
6383 cdst->mc_snum = csrc->mc_snum;
6384 cdst->mc_top = csrc->mc_top;
6385 cdst->mc_flags = csrc->mc_flags;
6387 for (i=0; i<csrc->mc_snum; i++) {
6388 cdst->mc_pg[i] = csrc->mc_pg[i];
6389 cdst->mc_ki[i] = csrc->mc_ki[i];
6393 /** Rebalance the tree after a delete operation.
6394 * @param[in] mc Cursor pointing to the page where rebalancing
6396 * @return 0 on success, non-zero on failure.
6399 mdb_rebalance(MDB_cursor *mc)
6403 unsigned int ptop, minkeys;
6406 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6410 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6411 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6412 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6413 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6417 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6418 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6421 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6422 DPRINTF("no need to rebalance page %zu, above fill threshold",
6428 if (mc->mc_snum < 2) {
6429 MDB_page *mp = mc->mc_pg[0];
6431 DPUTS("Can't rebalance a subpage, ignoring");
6434 if (NUMKEYS(mp) == 0) {
6435 DPUTS("tree is completely empty");
6436 mc->mc_db->md_root = P_INVALID;
6437 mc->mc_db->md_depth = 0;
6438 mc->mc_db->md_leaf_pages = 0;
6439 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6442 /* Adjust cursors pointing to mp */
6446 MDB_cursor *m2, *m3;
6447 MDB_dbi dbi = mc->mc_dbi;
6449 if (mc->mc_flags & C_SUB)
6452 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6453 if (mc->mc_flags & C_SUB)
6454 m3 = &m2->mc_xcursor->mx_cursor;
6457 if (m3->mc_snum < mc->mc_snum) continue;
6458 if (m3->mc_pg[0] == mp) {
6464 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6465 DPUTS("collapsing root page!");
6466 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6469 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6470 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6473 mc->mc_db->md_depth--;
6474 mc->mc_db->md_branch_pages--;
6476 /* Adjust other cursors pointing to mp */
6477 MDB_cursor *m2, *m3;
6478 MDB_dbi dbi = mc->mc_dbi;
6480 if (mc->mc_flags & C_SUB)
6483 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6484 if (mc->mc_flags & C_SUB)
6485 m3 = &m2->mc_xcursor->mx_cursor;
6488 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6489 if (m3->mc_pg[0] == mp) {
6490 m3->mc_pg[0] = mc->mc_pg[0];
6497 DPUTS("root page doesn't need rebalancing");
6501 /* The parent (branch page) must have at least 2 pointers,
6502 * otherwise the tree is invalid.
6504 ptop = mc->mc_top-1;
6505 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6507 /* Leaf page fill factor is below the threshold.
6508 * Try to move keys from left or right neighbor, or
6509 * merge with a neighbor page.
6514 mdb_cursor_copy(mc, &mn);
6515 mn.mc_xcursor = NULL;
6517 if (mc->mc_ki[ptop] == 0) {
6518 /* We're the leftmost leaf in our parent.
6520 DPUTS("reading right neighbor");
6522 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6523 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6526 mn.mc_ki[mn.mc_top] = 0;
6527 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6529 /* There is at least one neighbor to the left.
6531 DPUTS("reading left neighbor");
6533 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6534 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6537 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6538 mc->mc_ki[mc->mc_top] = 0;
6541 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6542 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);
6544 /* If the neighbor page is above threshold and has enough keys,
6545 * move one key from it. Otherwise we should try to merge them.
6546 * (A branch page must never have less than 2 keys.)
6548 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6549 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6550 return mdb_node_move(&mn, mc);
6552 if (mc->mc_ki[ptop] == 0)
6553 rc = mdb_page_merge(&mn, mc);
6555 rc = mdb_page_merge(mc, &mn);
6556 mc->mc_flags &= ~C_INITIALIZED;
6561 /** Complete a delete operation started by #mdb_cursor_del(). */
6563 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6567 /* add overflow pages to free list */
6568 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6572 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6573 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6574 (rc = mdb_ovpage_free(mc, omp)))
6577 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6578 mc->mc_db->md_entries--;
6579 rc = mdb_rebalance(mc);
6580 if (rc != MDB_SUCCESS)
6581 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6582 /* if mc points past last node in page, invalidate */
6583 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6584 mc->mc_flags &= ~C_INITIALIZED;
6590 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6591 MDB_val *key, MDB_val *data)
6596 MDB_val rdata, *xdata;
6600 assert(key != NULL);
6602 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6604 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6607 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6611 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6615 mdb_cursor_init(&mc, txn, dbi, &mx);
6626 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6628 /* let mdb_page_split know about this cursor if needed:
6629 * delete will trigger a rebalance; if it needs to move
6630 * a node from one page to another, it will have to
6631 * update the parent's separator key(s). If the new sepkey
6632 * is larger than the current one, the parent page may
6633 * run out of space, triggering a split. We need this
6634 * cursor to be consistent until the end of the rebalance.
6636 mc.mc_next = txn->mt_cursors[dbi];
6637 txn->mt_cursors[dbi] = &mc;
6638 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6639 txn->mt_cursors[dbi] = mc.mc_next;
6644 /** Split a page and insert a new node.
6645 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6646 * The cursor will be updated to point to the actual page and index where
6647 * the node got inserted after the split.
6648 * @param[in] newkey The key for the newly inserted node.
6649 * @param[in] newdata The data for the newly inserted node.
6650 * @param[in] newpgno The page number, if the new node is a branch node.
6651 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6652 * @return 0 on success, non-zero on failure.
6655 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6656 unsigned int nflags)
6659 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6662 unsigned int i, j, split_indx, nkeys, pmax;
6664 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6666 MDB_page *mp, *rp, *pp;
6671 mp = mc->mc_pg[mc->mc_top];
6672 newindx = mc->mc_ki[mc->mc_top];
6674 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6675 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6676 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6678 /* Create a right sibling. */
6679 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6681 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6683 if (mc->mc_snum < 2) {
6684 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6686 /* shift current top to make room for new parent */
6687 mc->mc_pg[1] = mc->mc_pg[0];
6688 mc->mc_ki[1] = mc->mc_ki[0];
6691 mc->mc_db->md_root = pp->mp_pgno;
6692 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6693 mc->mc_db->md_depth++;
6696 /* Add left (implicit) pointer. */
6697 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6698 /* undo the pre-push */
6699 mc->mc_pg[0] = mc->mc_pg[1];
6700 mc->mc_ki[0] = mc->mc_ki[1];
6701 mc->mc_db->md_root = mp->mp_pgno;
6702 mc->mc_db->md_depth--;
6709 ptop = mc->mc_top-1;
6710 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6713 mc->mc_flags |= C_SPLITTING;
6714 mdb_cursor_copy(mc, &mn);
6715 mn.mc_pg[mn.mc_top] = rp;
6716 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6718 if (nflags & MDB_APPEND) {
6719 mn.mc_ki[mn.mc_top] = 0;
6721 split_indx = newindx;
6726 nkeys = NUMKEYS(mp);
6727 split_indx = nkeys / 2;
6728 if (newindx < split_indx)
6734 unsigned int lsize, rsize, ksize;
6735 /* Move half of the keys to the right sibling */
6737 x = mc->mc_ki[mc->mc_top] - split_indx;
6738 ksize = mc->mc_db->md_pad;
6739 split = LEAF2KEY(mp, split_indx, ksize);
6740 rsize = (nkeys - split_indx) * ksize;
6741 lsize = (nkeys - split_indx) * sizeof(indx_t);
6742 mp->mp_lower -= lsize;
6743 rp->mp_lower += lsize;
6744 mp->mp_upper += rsize - lsize;
6745 rp->mp_upper -= rsize - lsize;
6746 sepkey.mv_size = ksize;
6747 if (newindx == split_indx) {
6748 sepkey.mv_data = newkey->mv_data;
6750 sepkey.mv_data = split;
6753 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6754 memcpy(rp->mp_ptrs, split, rsize);
6755 sepkey.mv_data = rp->mp_ptrs;
6756 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6757 memcpy(ins, newkey->mv_data, ksize);
6758 mp->mp_lower += sizeof(indx_t);
6759 mp->mp_upper -= ksize - sizeof(indx_t);
6762 memcpy(rp->mp_ptrs, split, x * ksize);
6763 ins = LEAF2KEY(rp, x, ksize);
6764 memcpy(ins, newkey->mv_data, ksize);
6765 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6766 rp->mp_lower += sizeof(indx_t);
6767 rp->mp_upper -= ksize - sizeof(indx_t);
6768 mc->mc_ki[mc->mc_top] = x;
6769 mc->mc_pg[mc->mc_top] = rp;
6774 /* For leaf pages, check the split point based on what
6775 * fits where, since otherwise mdb_node_add can fail.
6777 * This check is only needed when the data items are
6778 * relatively large, such that being off by one will
6779 * make the difference between success or failure.
6781 * It's also relevant if a page happens to be laid out
6782 * such that one half of its nodes are all "small" and
6783 * the other half of its nodes are "large." If the new
6784 * item is also "large" and falls on the half with
6785 * "large" nodes, it also may not fit.
6788 unsigned int psize, nsize;
6789 /* Maximum free space in an empty page */
6790 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6791 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6792 if ((nkeys < 20) || (nsize > pmax/16)) {
6793 if (newindx <= split_indx) {
6796 for (i=0; i<split_indx; i++) {
6797 node = NODEPTR(mp, i);
6798 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6799 if (F_ISSET(node->mn_flags, F_BIGDATA))
6800 psize += sizeof(pgno_t);
6802 psize += NODEDSZ(node);
6806 split_indx = newindx;
6817 for (i=nkeys-1; i>=split_indx; i--) {
6818 node = NODEPTR(mp, i);
6819 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6820 if (F_ISSET(node->mn_flags, F_BIGDATA))
6821 psize += sizeof(pgno_t);
6823 psize += NODEDSZ(node);
6827 split_indx = newindx;
6838 /* First find the separating key between the split pages.
6839 * The case where newindx == split_indx is ambiguous; the
6840 * new item could go to the new page or stay on the original
6841 * page. If newpos == 1 it goes to the new page.
6843 if (newindx == split_indx && newpos) {
6844 sepkey.mv_size = newkey->mv_size;
6845 sepkey.mv_data = newkey->mv_data;
6847 node = NODEPTR(mp, split_indx);
6848 sepkey.mv_size = node->mn_ksize;
6849 sepkey.mv_data = NODEKEY(node);
6853 DPRINTF("separator is [%s]", DKEY(&sepkey));
6855 /* Copy separator key to the parent.
6857 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6861 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6864 if (mn.mc_snum == mc->mc_snum) {
6865 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6866 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6867 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6868 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6873 /* Right page might now have changed parent.
6874 * Check if left page also changed parent.
6876 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6877 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6878 for (i=0; i<ptop; i++) {
6879 mc->mc_pg[i] = mn.mc_pg[i];
6880 mc->mc_ki[i] = mn.mc_ki[i];
6882 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6883 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6887 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6890 mc->mc_flags ^= C_SPLITTING;
6891 if (rc != MDB_SUCCESS) {
6894 if (nflags & MDB_APPEND) {
6895 mc->mc_pg[mc->mc_top] = rp;
6896 mc->mc_ki[mc->mc_top] = 0;
6897 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6900 for (i=0; i<mc->mc_top; i++)
6901 mc->mc_ki[i] = mn.mc_ki[i];
6908 /* Move half of the keys to the right sibling. */
6910 /* grab a page to hold a temporary copy */
6911 copy = mdb_page_malloc(mc->mc_txn, 1);
6915 copy->mp_pgno = mp->mp_pgno;
6916 copy->mp_flags = mp->mp_flags;
6917 copy->mp_lower = PAGEHDRSZ;
6918 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6919 mc->mc_pg[mc->mc_top] = copy;
6920 for (i = j = 0; i <= nkeys; j++) {
6921 if (i == split_indx) {
6922 /* Insert in right sibling. */
6923 /* Reset insert index for right sibling. */
6924 if (i != newindx || (newpos ^ ins_new)) {
6926 mc->mc_pg[mc->mc_top] = rp;
6930 if (i == newindx && !ins_new) {
6931 /* Insert the original entry that caused the split. */
6932 rkey.mv_data = newkey->mv_data;
6933 rkey.mv_size = newkey->mv_size;
6942 /* Update index for the new key. */
6943 mc->mc_ki[mc->mc_top] = j;
6944 } else if (i == nkeys) {
6947 node = NODEPTR(mp, i);
6948 rkey.mv_data = NODEKEY(node);
6949 rkey.mv_size = node->mn_ksize;
6951 xdata.mv_data = NODEDATA(node);
6952 xdata.mv_size = NODEDSZ(node);
6955 pgno = NODEPGNO(node);
6956 flags = node->mn_flags;
6961 if (!IS_LEAF(mp) && j == 0) {
6962 /* First branch index doesn't need key data. */
6966 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6970 nkeys = NUMKEYS(copy);
6971 for (i=0; i<nkeys; i++)
6972 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6973 mp->mp_lower = copy->mp_lower;
6974 mp->mp_upper = copy->mp_upper;
6975 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6976 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6978 /* reset back to original page */
6979 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6980 mc->mc_pg[mc->mc_top] = mp;
6981 if (nflags & MDB_RESERVE) {
6982 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6983 if (!(node->mn_flags & F_BIGDATA))
6984 newdata->mv_data = NODEDATA(node);
6988 /* Make sure mc_ki is still valid.
6990 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6991 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6992 for (i=0; i<ptop; i++) {
6993 mc->mc_pg[i] = mn.mc_pg[i];
6994 mc->mc_ki[i] = mn.mc_ki[i];
6996 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6997 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7001 /* return tmp page to freelist */
7002 mdb_page_free(mc->mc_txn->mt_env, copy);
7005 /* Adjust other cursors pointing to mp */
7006 MDB_cursor *m2, *m3;
7007 MDB_dbi dbi = mc->mc_dbi;
7008 int fixup = NUMKEYS(mp);
7010 if (mc->mc_flags & C_SUB)
7013 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7014 if (mc->mc_flags & C_SUB)
7015 m3 = &m2->mc_xcursor->mx_cursor;
7020 if (!(m3->mc_flags & C_INITIALIZED))
7022 if (m3->mc_flags & C_SPLITTING)
7027 for (k=m3->mc_top; k>=0; k--) {
7028 m3->mc_ki[k+1] = m3->mc_ki[k];
7029 m3->mc_pg[k+1] = m3->mc_pg[k];
7031 if (m3->mc_ki[0] >= split_indx) {
7036 m3->mc_pg[0] = mc->mc_pg[0];
7040 if (m3->mc_pg[mc->mc_top] == mp) {
7041 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7042 m3->mc_ki[mc->mc_top]++;
7043 if (m3->mc_ki[mc->mc_top] >= fixup) {
7044 m3->mc_pg[mc->mc_top] = rp;
7045 m3->mc_ki[mc->mc_top] -= fixup;
7046 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7048 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7049 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7058 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7059 MDB_val *key, MDB_val *data, unsigned int flags)
7064 assert(key != NULL);
7065 assert(data != NULL);
7067 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7070 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7074 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7078 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7081 mdb_cursor_init(&mc, txn, dbi, &mx);
7082 return mdb_cursor_put(&mc, key, data, flags);
7086 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7088 if ((flag & CHANGEABLE) != flag)
7091 env->me_flags |= flag;
7093 env->me_flags &= ~flag;
7098 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7103 *arg = env->me_flags;
7108 mdb_env_get_path(MDB_env *env, const char **arg)
7113 *arg = env->me_path;
7117 /** Common code for #mdb_stat() and #mdb_env_stat().
7118 * @param[in] env the environment to operate in.
7119 * @param[in] db the #MDB_db record containing the stats to return.
7120 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7121 * @return 0, this function always succeeds.
7124 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7126 arg->ms_psize = env->me_psize;
7127 arg->ms_depth = db->md_depth;
7128 arg->ms_branch_pages = db->md_branch_pages;
7129 arg->ms_leaf_pages = db->md_leaf_pages;
7130 arg->ms_overflow_pages = db->md_overflow_pages;
7131 arg->ms_entries = db->md_entries;
7136 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7140 if (env == NULL || arg == NULL)
7143 toggle = mdb_env_pick_meta(env);
7145 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7149 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7153 if (env == NULL || arg == NULL)
7156 toggle = mdb_env_pick_meta(env);
7157 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7158 arg->me_mapsize = env->me_mapsize;
7159 arg->me_maxreaders = env->me_maxreaders;
7160 arg->me_numreaders = env->me_numreaders;
7161 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7162 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7166 /** Set the default comparison functions for a database.
7167 * Called immediately after a database is opened to set the defaults.
7168 * The user can then override them with #mdb_set_compare() or
7169 * #mdb_set_dupsort().
7170 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7171 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7174 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7176 uint16_t f = txn->mt_dbs[dbi].md_flags;
7178 txn->mt_dbxs[dbi].md_cmp =
7179 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7180 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7182 txn->mt_dbxs[dbi].md_dcmp =
7183 !(f & MDB_DUPSORT) ? 0 :
7184 ((f & MDB_INTEGERDUP)
7185 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7186 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7189 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7194 int rc, dbflag, exact;
7195 unsigned int unused = 0;
7198 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7199 mdb_default_cmp(txn, FREE_DBI);
7202 if ((flags & VALID_FLAGS) != flags)
7208 if (flags & PERSISTENT_FLAGS) {
7209 uint16_t f2 = flags & PERSISTENT_FLAGS;
7210 /* make sure flag changes get committed */
7211 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7212 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7213 txn->mt_flags |= MDB_TXN_DIRTY;
7216 mdb_default_cmp(txn, MAIN_DBI);
7220 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7221 mdb_default_cmp(txn, MAIN_DBI);
7224 /* Is the DB already open? */
7226 for (i=2; i<txn->mt_numdbs; i++) {
7227 if (!txn->mt_dbxs[i].md_name.mv_size) {
7228 /* Remember this free slot */
7229 if (!unused) unused = i;
7232 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7233 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7239 /* If no free slot and max hit, fail */
7240 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7241 return MDB_DBS_FULL;
7243 /* Cannot mix named databases with some mainDB flags */
7244 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7245 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7247 /* Find the DB info */
7248 dbflag = DB_NEW|DB_VALID;
7251 key.mv_data = (void *)name;
7252 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7253 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7254 if (rc == MDB_SUCCESS) {
7255 /* make sure this is actually a DB */
7256 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7257 if (!(node->mn_flags & F_SUBDATA))
7259 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7260 /* Create if requested */
7262 data.mv_size = sizeof(MDB_db);
7263 data.mv_data = &dummy;
7264 memset(&dummy, 0, sizeof(dummy));
7265 dummy.md_root = P_INVALID;
7266 dummy.md_flags = flags & PERSISTENT_FLAGS;
7267 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7271 /* OK, got info, add to table */
7272 if (rc == MDB_SUCCESS) {
7273 unsigned int slot = unused ? unused : txn->mt_numdbs;
7274 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7275 txn->mt_dbxs[slot].md_name.mv_size = len;
7276 txn->mt_dbxs[slot].md_rel = NULL;
7277 txn->mt_dbflags[slot] = dbflag;
7278 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7280 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7281 mdb_default_cmp(txn, slot);
7290 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7292 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7295 if (txn->mt_dbflags[dbi] & DB_STALE) {
7298 /* Stale, must read the DB's root. cursor_init does it for us. */
7299 mdb_cursor_init(&mc, txn, dbi, &mx);
7301 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7304 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7307 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7309 ptr = env->me_dbxs[dbi].md_name.mv_data;
7310 env->me_dbxs[dbi].md_name.mv_data = NULL;
7311 env->me_dbxs[dbi].md_name.mv_size = 0;
7312 env->me_dbflags[dbi] = 0;
7316 /** Add all the DB's pages to the free list.
7317 * @param[in] mc Cursor on the DB to free.
7318 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7319 * @return 0 on success, non-zero on failure.
7322 mdb_drop0(MDB_cursor *mc, int subs)
7326 rc = mdb_page_search(mc, NULL, 0);
7327 if (rc == MDB_SUCCESS) {
7328 MDB_txn *txn = mc->mc_txn;
7333 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7334 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7337 mdb_cursor_copy(mc, &mx);
7338 while (mc->mc_snum > 0) {
7339 MDB_page *mp = mc->mc_pg[mc->mc_top];
7340 unsigned n = NUMKEYS(mp);
7342 for (i=0; i<n; i++) {
7343 ni = NODEPTR(mp, i);
7344 if (ni->mn_flags & F_BIGDATA) {
7347 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7348 rc = mdb_page_get(txn, pg, &omp, NULL);
7351 assert(IS_OVERFLOW(omp));
7352 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7356 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7357 mdb_xcursor_init1(mc, ni);
7358 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7364 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7366 for (i=0; i<n; i++) {
7368 ni = NODEPTR(mp, i);
7371 mdb_midl_xappend(txn->mt_free_pgs, pg);
7376 mc->mc_ki[mc->mc_top] = i;
7377 rc = mdb_cursor_sibling(mc, 1);
7379 /* no more siblings, go back to beginning
7380 * of previous level.
7384 for (i=1; i<mc->mc_snum; i++) {
7386 mc->mc_pg[i] = mx.mc_pg[i];
7391 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
7392 } else if (rc == MDB_NOTFOUND) {
7398 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7400 MDB_cursor *mc, *m2;
7403 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7406 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7409 rc = mdb_cursor_open(txn, dbi, &mc);
7413 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7414 /* Invalidate the dropped DB's cursors */
7415 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7416 m2->mc_flags &= ~C_INITIALIZED;
7420 /* Can't delete the main DB */
7421 if (del && dbi > MAIN_DBI) {
7422 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7424 txn->mt_dbflags[dbi] = DB_STALE;
7425 mdb_dbi_close(txn->mt_env, dbi);
7428 /* reset the DB record, mark it dirty */
7429 txn->mt_dbflags[dbi] |= DB_DIRTY;
7430 txn->mt_dbs[dbi].md_depth = 0;
7431 txn->mt_dbs[dbi].md_branch_pages = 0;
7432 txn->mt_dbs[dbi].md_leaf_pages = 0;
7433 txn->mt_dbs[dbi].md_overflow_pages = 0;
7434 txn->mt_dbs[dbi].md_entries = 0;
7435 txn->mt_dbs[dbi].md_root = P_INVALID;
7437 txn->mt_flags |= MDB_TXN_DIRTY;
7440 mdb_cursor_close(mc);
7444 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7446 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7449 txn->mt_dbxs[dbi].md_cmp = cmp;
7453 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7455 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7458 txn->mt_dbxs[dbi].md_dcmp = cmp;
7462 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7464 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7467 txn->mt_dbxs[dbi].md_rel = rel;
7471 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7473 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7476 txn->mt_dbxs[dbi].md_relctx = ctx;