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)
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 * We require that keys all fit onto a regular page. This limit
353 * could be raised a bit further if needed; to something just
354 * under #MDB_PAGESIZE / #MDB_MINKEYS.
356 * Note that data items in an #MDB_DUPSORT database are actually keys
357 * of a subDB, so they're also limited to this size.
359 #ifndef MDB_MAXKEYSIZE
360 #define MDB_MAXKEYSIZE 511
363 /** @brief The maximum size of a data item.
365 * We only store a 32 bit value for node sizes.
367 #define MAXDATASIZE 0xffffffffUL
372 * This is used for printing a hex dump of a key's contents.
374 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
375 /** Display a key in hex.
377 * Invoke a function to display a key in hex.
379 #define DKEY(x) mdb_dkey(x, kbuf)
381 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
385 /** An invalid page number.
386 * Mainly used to denote an empty tree.
388 #define P_INVALID (~(pgno_t)0)
390 /** Test if the flags \b f are set in a flag word \b w. */
391 #define F_ISSET(w, f) (((w) & (f)) == (f))
393 /** Used for offsets within a single page.
394 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
397 typedef uint16_t indx_t;
399 /** Default size of memory map.
400 * This is certainly too small for any actual applications. Apps should always set
401 * the size explicitly using #mdb_env_set_mapsize().
403 #define DEFAULT_MAPSIZE 1048576
405 /** @defgroup readers Reader Lock Table
406 * Readers don't acquire any locks for their data access. Instead, they
407 * simply record their transaction ID in the reader table. The reader
408 * mutex is needed just to find an empty slot in the reader table. The
409 * slot's address is saved in thread-specific data so that subsequent read
410 * transactions started by the same thread need no further locking to proceed.
412 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
414 * No reader table is used if the database is on a read-only filesystem.
416 * Since the database uses multi-version concurrency control, readers don't
417 * actually need any locking. This table is used to keep track of which
418 * readers are using data from which old transactions, so that we'll know
419 * when a particular old transaction is no longer in use. Old transactions
420 * that have discarded any data pages can then have those pages reclaimed
421 * for use by a later write transaction.
423 * The lock table is constructed such that reader slots are aligned with the
424 * processor's cache line size. Any slot is only ever used by one thread.
425 * This alignment guarantees that there will be no contention or cache
426 * thrashing as threads update their own slot info, and also eliminates
427 * any need for locking when accessing a slot.
429 * A writer thread will scan every slot in the table to determine the oldest
430 * outstanding reader transaction. Any freed pages older than this will be
431 * reclaimed by the writer. The writer doesn't use any locks when scanning
432 * this table. This means that there's no guarantee that the writer will
433 * see the most up-to-date reader info, but that's not required for correct
434 * operation - all we need is to know the upper bound on the oldest reader,
435 * we don't care at all about the newest reader. So the only consequence of
436 * reading stale information here is that old pages might hang around a
437 * while longer before being reclaimed. That's actually good anyway, because
438 * the longer we delay reclaiming old pages, the more likely it is that a
439 * string of contiguous pages can be found after coalescing old pages from
440 * many old transactions together.
443 /** Number of slots in the reader table.
444 * This value was chosen somewhat arbitrarily. 126 readers plus a
445 * couple mutexes fit exactly into 8KB on my development machine.
446 * Applications should set the table size using #mdb_env_set_maxreaders().
448 #define DEFAULT_READERS 126
450 /** The size of a CPU cache line in bytes. We want our lock structures
451 * aligned to this size to avoid false cache line sharing in the
453 * This value works for most CPUs. For Itanium this should be 128.
459 /** The information we store in a single slot of the reader table.
460 * In addition to a transaction ID, we also record the process and
461 * thread ID that owns a slot, so that we can detect stale information,
462 * e.g. threads or processes that went away without cleaning up.
463 * @note We currently don't check for stale records. We simply re-init
464 * the table when we know that we're the only process opening the
467 typedef struct MDB_rxbody {
468 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
469 * Multiple readers that start at the same time will probably have the
470 * same ID here. Again, it's not important to exclude them from
471 * anything; all we need to know is which version of the DB they
472 * started from so we can avoid overwriting any data used in that
473 * particular version.
476 /** The process ID of the process owning this reader txn. */
478 /** The thread ID of the thread owning this txn. */
482 /** The actual reader record, with cacheline padding. */
483 typedef struct MDB_reader {
486 /** shorthand for mrb_txnid */
487 #define mr_txnid mru.mrx.mrb_txnid
488 #define mr_pid mru.mrx.mrb_pid
489 #define mr_tid mru.mrx.mrb_tid
490 /** cache line alignment */
491 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
495 /** The header for the reader table.
496 * The table resides in a memory-mapped file. (This is a different file
497 * than is used for the main database.)
499 * For POSIX the actual mutexes reside in the shared memory of this
500 * mapped file. On Windows, mutexes are named objects allocated by the
501 * kernel; we store the mutex names in this mapped file so that other
502 * processes can grab them. This same approach is also used on
503 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
504 * process-shared POSIX mutexes. For these cases where a named object
505 * is used, the object name is derived from a 64 bit FNV hash of the
506 * environment pathname. As such, naming collisions are extremely
507 * unlikely. If a collision occurs, the results are unpredictable.
509 typedef struct MDB_txbody {
510 /** Stamp identifying this as an MDB file. It must be set
513 /** Version number of this lock file. Must be set to #MDB_VERSION. */
514 uint32_t mtb_version;
515 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
516 char mtb_rmname[MNAME_LEN];
518 /** Mutex protecting access to this table.
519 * This is the reader lock that #LOCK_MUTEX_R acquires.
521 pthread_mutex_t mtb_mutex;
523 /** The ID of the last transaction committed to the database.
524 * This is recorded here only for convenience; the value can always
525 * be determined by reading the main database meta pages.
528 /** The number of slots that have been used in the reader table.
529 * This always records the maximum count, it is not decremented
530 * when readers release their slots.
532 unsigned mtb_numreaders;
535 /** The actual reader table definition. */
536 typedef struct MDB_txninfo {
539 #define mti_magic mt1.mtb.mtb_magic
540 #define mti_version mt1.mtb.mtb_version
541 #define mti_mutex mt1.mtb.mtb_mutex
542 #define mti_rmname mt1.mtb.mtb_rmname
543 #define mti_txnid mt1.mtb.mtb_txnid
544 #define mti_numreaders mt1.mtb.mtb_numreaders
545 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
548 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
549 char mt2_wmname[MNAME_LEN];
550 #define mti_wmname mt2.mt2_wmname
552 pthread_mutex_t mt2_wmutex;
553 #define mti_wmutex mt2.mt2_wmutex
555 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
557 MDB_reader mti_readers[1];
561 /** Common header for all page types.
562 * Overflow records occupy a number of contiguous pages with no
563 * headers on any page after the first.
565 typedef struct MDB_page {
566 #define mp_pgno mp_p.p_pgno
567 #define mp_next mp_p.p_next
569 pgno_t p_pgno; /**< page number */
570 void * p_next; /**< for in-memory list of freed structs */
573 /** @defgroup mdb_page Page Flags
575 * Flags for the page headers.
578 #define P_BRANCH 0x01 /**< branch page */
579 #define P_LEAF 0x02 /**< leaf page */
580 #define P_OVERFLOW 0x04 /**< overflow page */
581 #define P_META 0x08 /**< meta page */
582 #define P_DIRTY 0x10 /**< dirty page */
583 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
584 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
586 uint16_t mp_flags; /**< @ref mdb_page */
587 #define mp_lower mp_pb.pb.pb_lower
588 #define mp_upper mp_pb.pb.pb_upper
589 #define mp_pages mp_pb.pb_pages
592 indx_t pb_lower; /**< lower bound of free space */
593 indx_t pb_upper; /**< upper bound of free space */
595 uint32_t pb_pages; /**< number of overflow pages */
597 indx_t mp_ptrs[1]; /**< dynamic size */
600 /** Size of the page header, excluding dynamic data at the end */
601 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
603 /** Address of first usable data byte in a page, after the header */
604 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
606 /** Number of nodes on a page */
607 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
609 /** The amount of space remaining in the page */
610 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
612 /** The percentage of space used in the page, in tenths of a percent. */
613 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
614 ((env)->me_psize - PAGEHDRSZ))
615 /** The minimum page fill factor, in tenths of a percent.
616 * Pages emptier than this are candidates for merging.
618 #define FILL_THRESHOLD 250
620 /** Test if a page is a leaf page */
621 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
622 /** Test if a page is a LEAF2 page */
623 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
624 /** Test if a page is a branch page */
625 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
626 /** Test if a page is an overflow page */
627 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
628 /** Test if a page is a sub page */
629 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
631 /** The number of overflow pages needed to store the given size. */
632 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
634 /** Header for a single key/data pair within a page.
635 * We guarantee 2-byte alignment for nodes.
637 typedef struct MDB_node {
638 /** lo and hi are used for data size on leaf nodes and for
639 * child pgno on branch nodes. On 64 bit platforms, flags
640 * is also used for pgno. (Branch nodes have no flags).
641 * They are in host byte order in case that lets some
642 * accesses be optimized into a 32-bit word access.
644 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
645 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
646 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
647 /** @defgroup mdb_node Node Flags
649 * Flags for node headers.
652 #define F_BIGDATA 0x01 /**< data put on overflow page */
653 #define F_SUBDATA 0x02 /**< data is a sub-database */
654 #define F_DUPDATA 0x04 /**< data has duplicates */
656 /** valid flags for #mdb_node_add() */
657 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
660 unsigned short mn_flags; /**< @ref mdb_node */
661 unsigned short mn_ksize; /**< key size */
662 char mn_data[1]; /**< key and data are appended here */
665 /** Size of the node header, excluding dynamic data at the end */
666 #define NODESIZE offsetof(MDB_node, mn_data)
668 /** Bit position of top word in page number, for shifting mn_flags */
669 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
671 /** Size of a node in a branch page with a given key.
672 * This is just the node header plus the key, there is no data.
674 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
676 /** Size of a node in a leaf page with a given key and data.
677 * This is node header plus key plus data size.
679 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
681 /** Address of node \b i in page \b p */
682 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
684 /** Address of the key for the node */
685 #define NODEKEY(node) (void *)((node)->mn_data)
687 /** Address of the data for a node */
688 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
690 /** Get the page number pointed to by a branch node */
691 #define NODEPGNO(node) \
692 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
693 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
694 /** Set the page number in a branch node */
695 #define SETPGNO(node,pgno) do { \
696 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
697 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
699 /** Get the size of the data in a leaf node */
700 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
701 /** Set the size of the data for a leaf node */
702 #define SETDSZ(node,size) do { \
703 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
704 /** The size of a key in a node */
705 #define NODEKSZ(node) ((node)->mn_ksize)
707 /** Copy a page number from src to dst */
709 #define COPY_PGNO(dst,src) dst = src
711 #if SIZE_MAX > 4294967295UL
712 #define COPY_PGNO(dst,src) do { \
713 unsigned short *s, *d; \
714 s = (unsigned short *)&(src); \
715 d = (unsigned short *)&(dst); \
722 #define COPY_PGNO(dst,src) do { \
723 unsigned short *s, *d; \
724 s = (unsigned short *)&(src); \
725 d = (unsigned short *)&(dst); \
731 /** The address of a key in a LEAF2 page.
732 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
733 * There are no node headers, keys are stored contiguously.
735 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
737 /** Set the \b node's key into \b key, if requested. */
738 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
739 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
741 /** Information about a single database in the environment. */
742 typedef struct MDB_db {
743 uint32_t md_pad; /**< also ksize for LEAF2 pages */
744 uint16_t md_flags; /**< @ref mdb_dbi_open */
745 uint16_t md_depth; /**< depth of this tree */
746 pgno_t md_branch_pages; /**< number of internal pages */
747 pgno_t md_leaf_pages; /**< number of leaf pages */
748 pgno_t md_overflow_pages; /**< number of overflow pages */
749 size_t md_entries; /**< number of data items */
750 pgno_t md_root; /**< the root page of this tree */
753 /** mdb_dbi_open flags */
754 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
755 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
756 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
757 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
759 /** Handle for the DB used to track free pages. */
761 /** Handle for the default DB. */
764 /** Meta page content. */
765 typedef struct MDB_meta {
766 /** Stamp identifying this as an MDB file. It must be set
769 /** Version number of this lock file. Must be set to #MDB_VERSION. */
771 void *mm_address; /**< address for fixed mapping */
772 size_t mm_mapsize; /**< size of mmap region */
773 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
774 /** The size of pages used in this DB */
775 #define mm_psize mm_dbs[0].md_pad
776 /** Any persistent environment flags. @ref mdb_env */
777 #define mm_flags mm_dbs[0].md_flags
778 pgno_t mm_last_pg; /**< last used page in file */
779 txnid_t mm_txnid; /**< txnid that committed this page */
782 /** Buffer for a stack-allocated dirty page.
783 * The members define size and alignment, and silence type
784 * aliasing warnings. They are not used directly; that could
785 * mean incorrectly using several union members in parallel.
787 typedef union MDB_pagebuf {
788 char mb_raw[MDB_PAGESIZE];
791 char mm_pad[PAGEHDRSZ];
796 /** Auxiliary DB info.
797 * The information here is mostly static/read-only. There is
798 * only a single copy of this record in the environment.
800 typedef struct MDB_dbx {
801 MDB_val md_name; /**< name of the database */
802 MDB_cmp_func *md_cmp; /**< function for comparing keys */
803 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
804 MDB_rel_func *md_rel; /**< user relocate function */
805 void *md_relctx; /**< user-provided context for md_rel */
808 /** A database transaction.
809 * Every operation requires a transaction handle.
812 MDB_txn *mt_parent; /**< parent of a nested txn */
813 MDB_txn *mt_child; /**< nested txn under this txn */
814 pgno_t mt_next_pgno; /**< next unallocated page */
815 /** The ID of this transaction. IDs are integers incrementing from 1.
816 * Only committed write transactions increment the ID. If a transaction
817 * aborts, the ID may be re-used by the next writer.
820 MDB_env *mt_env; /**< the DB environment */
821 /** The list of pages that became unused during this transaction.
825 MDB_ID2L dirty_list; /**< for write txns: modified pages */
826 MDB_reader *reader; /**< this thread's reader table slot or NULL */
828 /** Array of records for each DB known in the environment. */
830 /** Array of MDB_db records for each known DB */
832 /** @defgroup mt_dbflag Transaction DB Flags
836 #define DB_DIRTY 0x01 /**< DB was written in this txn */
837 #define DB_STALE 0x02 /**< DB record is older than txnID */
838 #define DB_NEW 0x04 /**< DB handle opened in this txn */
839 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
841 /** In write txns, array of cursors for each DB */
842 MDB_cursor **mt_cursors;
843 /** Array of flags for each DB */
844 unsigned char *mt_dbflags;
845 /** Number of DB records in use. This number only ever increments;
846 * we don't decrement it when individual DB handles are closed.
850 /** @defgroup mdb_txn Transaction Flags
854 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
855 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
856 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
858 unsigned int mt_flags; /**< @ref mdb_txn */
859 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
860 unsigned int mt_dirty_room;
861 /** Tracks which of the two meta pages was used at the start
862 * of this transaction.
864 unsigned int mt_toggle;
867 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
868 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
869 * raise this on a 64 bit machine.
871 #define CURSOR_STACK 32
875 /** Cursors are used for all DB operations */
877 /** Next cursor on this DB in this txn */
879 /** Original cursor if this is a shadow */
881 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
882 struct MDB_xcursor *mc_xcursor;
883 /** The transaction that owns this cursor */
885 /** The database handle this cursor operates on */
887 /** The database record for this cursor */
889 /** The database auxiliary record for this cursor */
891 /** The @ref mt_dbflag for this database */
892 unsigned char *mc_dbflag;
893 unsigned short mc_snum; /**< number of pushed pages */
894 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
895 /** @defgroup mdb_cursor Cursor Flags
897 * Cursor state flags.
900 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
901 #define C_EOF 0x02 /**< No more data */
902 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
903 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
904 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
905 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
906 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
908 unsigned int mc_flags; /**< @ref mdb_cursor */
909 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
910 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
913 /** Context for sorted-dup records.
914 * We could have gone to a fully recursive design, with arbitrarily
915 * deep nesting of sub-databases. But for now we only handle these
916 * levels - main DB, optional sub-DB, sorted-duplicate DB.
918 typedef struct MDB_xcursor {
919 /** A sub-cursor for traversing the Dup DB */
920 MDB_cursor mx_cursor;
921 /** The database record for this Dup DB */
923 /** The auxiliary DB record for this Dup DB */
925 /** The @ref mt_dbflag for this Dup DB */
926 unsigned char mx_dbflag;
929 /** State of FreeDB old pages, stored in the MDB_env */
930 typedef struct MDB_pgstate {
931 txnid_t mf_pglast; /**< ID of last old page record we used */
932 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
935 /** The database environment. */
937 HANDLE me_fd; /**< The main data file */
938 HANDLE me_lfd; /**< The lock file */
939 HANDLE me_mfd; /**< just for writing the meta pages */
940 /** Failed to update the meta page. Probably an I/O error. */
941 #define MDB_FATAL_ERROR 0x80000000U
942 /** Some fields are initialized. */
943 #define MDB_ENV_ACTIVE 0x20000000U
944 /** me_txkey is set */
945 #define MDB_ENV_TXKEY 0x10000000U
946 uint32_t me_flags; /**< @ref mdb_env */
947 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
948 unsigned int me_maxreaders; /**< size of the reader table */
949 unsigned int me_numreaders; /**< max numreaders set by this env */
950 MDB_dbi me_numdbs; /**< number of DBs opened */
951 MDB_dbi me_maxdbs; /**< size of the DB table */
952 pid_t me_pid; /**< process ID of this env */
953 char *me_path; /**< path to the DB files */
954 char *me_map; /**< the memory map of the data file */
955 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
956 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
957 MDB_txn *me_txn; /**< current write transaction */
958 size_t me_mapsize; /**< size of the data memory map */
959 off_t me_size; /**< current file size */
960 pgno_t me_maxpg; /**< me_mapsize / me_psize */
961 MDB_dbx *me_dbxs; /**< array of static DB info */
962 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
963 pthread_key_t me_txkey; /**< thread-key for readers */
964 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
965 # define me_pglast me_pgstate.mf_pglast
966 # define me_pghead me_pgstate.mf_pghead
967 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
968 /** IDL of pages that became unused in a write txn */
970 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
971 MDB_ID2L me_dirty_list;
972 /** Max number of freelist items that can fit in a single overflow page */
974 /** Max size of a node on a page */
975 unsigned int me_nodemax;
977 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
979 #elif defined(MDB_USE_POSIX_SEM)
980 sem_t *me_rmutex; /* Shared mutexes are not supported */
985 /** Nested transaction */
986 typedef struct MDB_ntxn {
987 MDB_txn mnt_txn; /* the transaction */
988 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
991 /** max number of pages to commit in one writev() call */
992 #define MDB_COMMIT_PAGES 64
993 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
994 #undef MDB_COMMIT_PAGES
995 #define MDB_COMMIT_PAGES IOV_MAX
998 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
999 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1000 static int mdb_page_touch(MDB_cursor *mc);
1002 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1003 static int mdb_page_search_root(MDB_cursor *mc,
1004 MDB_val *key, int modify);
1005 #define MDB_PS_MODIFY 1
1006 #define MDB_PS_ROOTONLY 2
1007 static int mdb_page_search(MDB_cursor *mc,
1008 MDB_val *key, int flags);
1009 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1011 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1012 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1013 pgno_t newpgno, unsigned int nflags);
1015 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1016 static int mdb_env_pick_meta(const MDB_env *env);
1017 static int mdb_env_write_meta(MDB_txn *txn);
1018 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1019 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1021 static void mdb_env_close0(MDB_env *env, int excl);
1023 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1024 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1025 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1026 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1027 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1028 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1029 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1030 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1031 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1033 static int mdb_rebalance(MDB_cursor *mc);
1034 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1036 static void mdb_cursor_pop(MDB_cursor *mc);
1037 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1039 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1040 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1041 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1042 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1043 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1045 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1046 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1048 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1049 static void mdb_xcursor_init0(MDB_cursor *mc);
1050 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1052 static int mdb_drop0(MDB_cursor *mc, int subs);
1053 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1056 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1060 static SECURITY_DESCRIPTOR mdb_null_sd;
1061 static SECURITY_ATTRIBUTES mdb_all_sa;
1062 static int mdb_sec_inited;
1065 /** Return the library version info. */
1067 mdb_version(int *major, int *minor, int *patch)
1069 if (major) *major = MDB_VERSION_MAJOR;
1070 if (minor) *minor = MDB_VERSION_MINOR;
1071 if (patch) *patch = MDB_VERSION_PATCH;
1072 return MDB_VERSION_STRING;
1075 /** Table of descriptions for MDB @ref errors */
1076 static char *const mdb_errstr[] = {
1077 "MDB_KEYEXIST: Key/data pair already exists",
1078 "MDB_NOTFOUND: No matching key/data pair found",
1079 "MDB_PAGE_NOTFOUND: Requested page not found",
1080 "MDB_CORRUPTED: Located page was wrong type",
1081 "MDB_PANIC: Update of meta page failed",
1082 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1083 "MDB_INVALID: File is not an MDB file",
1084 "MDB_MAP_FULL: Environment mapsize limit reached",
1085 "MDB_DBS_FULL: Environment maxdbs limit reached",
1086 "MDB_READERS_FULL: Environment maxreaders limit reached",
1087 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1088 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1089 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1090 "MDB_PAGE_FULL: Internal error - page has no more space",
1091 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1092 "MDB_INCOMPATIBLE: Database flags changed or would change",
1093 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1097 mdb_strerror(int err)
1101 return ("Successful return: 0");
1103 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1104 i = err - MDB_KEYEXIST;
1105 return mdb_errstr[i];
1108 return strerror(err);
1112 /** Display a key in hexadecimal and return the address of the result.
1113 * @param[in] key the key to display
1114 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1115 * @return The key in hexadecimal form.
1118 mdb_dkey(MDB_val *key, char *buf)
1121 unsigned char *c = key->mv_data;
1127 if (key->mv_size > MDB_MAXKEYSIZE)
1128 return "MDB_MAXKEYSIZE";
1129 /* may want to make this a dynamic check: if the key is mostly
1130 * printable characters, print it as-is instead of converting to hex.
1134 for (i=0; i<key->mv_size; i++)
1135 ptr += sprintf(ptr, "%02x", *c++);
1137 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1142 /** Display all the keys in the page. */
1144 mdb_page_list(MDB_page *mp)
1147 unsigned int i, nkeys, nsize;
1151 nkeys = NUMKEYS(mp);
1152 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1153 for (i=0; i<nkeys; i++) {
1154 node = NODEPTR(mp, i);
1155 key.mv_size = node->mn_ksize;
1156 key.mv_data = node->mn_data;
1157 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1158 if (IS_BRANCH(mp)) {
1159 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1162 if (F_ISSET(node->mn_flags, F_BIGDATA))
1163 nsize += sizeof(pgno_t);
1165 nsize += NODEDSZ(node);
1166 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1172 mdb_cursor_chk(MDB_cursor *mc)
1178 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1179 for (i=0; i<mc->mc_top; i++) {
1181 node = NODEPTR(mp, mc->mc_ki[i]);
1182 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1185 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1191 /** Count all the pages in each DB and in the freelist
1192 * and make sure it matches the actual number of pages
1195 static void mdb_audit(MDB_txn *txn)
1199 MDB_ID freecount, count;
1204 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1205 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1206 freecount += *(MDB_ID *)data.mv_data;
1209 for (i = 0; i<txn->mt_numdbs; i++) {
1211 mdb_cursor_init(&mc, txn, i, &mx);
1212 if (txn->mt_dbs[i].md_root == P_INVALID)
1214 count += txn->mt_dbs[i].md_branch_pages +
1215 txn->mt_dbs[i].md_leaf_pages +
1216 txn->mt_dbs[i].md_overflow_pages;
1217 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1218 mdb_page_search(&mc, NULL, 0);
1222 mp = mc.mc_pg[mc.mc_top];
1223 for (j=0; j<NUMKEYS(mp); j++) {
1224 MDB_node *leaf = NODEPTR(mp, j);
1225 if (leaf->mn_flags & F_SUBDATA) {
1227 memcpy(&db, NODEDATA(leaf), sizeof(db));
1228 count += db.md_branch_pages + db.md_leaf_pages +
1229 db.md_overflow_pages;
1233 while (mdb_cursor_sibling(&mc, 1) == 0);
1236 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1237 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1238 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1244 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1246 return txn->mt_dbxs[dbi].md_cmp(a, b);
1250 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1252 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1255 /** Allocate a page.
1256 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1259 mdb_page_malloc(MDB_cursor *mc, unsigned num)
1261 MDB_env *env = mc->mc_txn->mt_env;
1262 MDB_page *ret = env->me_dpages;
1263 size_t sz = env->me_psize;
1266 VGMEMP_ALLOC(env, ret, sz);
1267 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1268 env->me_dpages = ret->mp_next;
1274 if ((ret = malloc(sz)) != NULL) {
1275 VGMEMP_ALLOC(env, ret, sz);
1280 /** Free a single page.
1281 * Saves single pages to a list, for future reuse.
1282 * (This is not used for multi-page overflow pages.)
1285 mdb_page_free(MDB_env *env, MDB_page *mp)
1287 mp->mp_next = env->me_dpages;
1288 VGMEMP_FREE(env, mp);
1289 env->me_dpages = mp;
1292 /* Return all dirty pages to dpage list */
1294 mdb_dlist_free(MDB_txn *txn)
1296 MDB_env *env = txn->mt_env;
1297 MDB_ID2L dl = txn->mt_u.dirty_list;
1298 unsigned i, n = dl[0].mid;
1300 for (i = 1; i <= n; i++) {
1301 MDB_page *dp = dl[i].mptr;
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);
1313 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1315 mdb_find_oldest(MDB_txn *txn)
1318 txnid_t mr, oldest = txn->mt_txnid - 1;
1319 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1320 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1330 /** Allocate pages for writing.
1331 * If there are free pages available from older transactions, they
1332 * will be re-used first. Otherwise a new page will be allocated.
1333 * @param[in] mc cursor A cursor handle identifying the transaction and
1334 * database for which we are allocating.
1335 * @param[in] num the number of pages to allocate.
1336 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1337 * will always be satisfied by a single contiguous chunk of memory.
1338 * @return 0 on success, non-zero on failure.
1341 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1343 MDB_txn *txn = mc->mc_txn;
1345 pgno_t pgno = P_INVALID;
1347 txnid_t oldest = 0, last;
1352 /* If our dirty list is already full, we can't do anything */
1353 if (txn->mt_dirty_room == 0)
1354 return MDB_TXN_FULL;
1356 /* The free list won't have any content at all until txn 2 has
1357 * committed. The pages freed by txn 2 will be unreferenced
1358 * after txn 3 commits, and so will be safe to re-use in txn 4.
1360 if (txn->mt_txnid > 3) {
1361 if (!txn->mt_env->me_pghead &&
1362 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1363 /* See if there's anything in the free DB */
1369 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1370 if (!txn->mt_env->me_pglast) {
1371 mdb_page_search(&m2, NULL, 0);
1372 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1373 kptr = (txnid_t *)NODEKEY(leaf);
1378 last = txn->mt_env->me_pglast + 1;
1380 key.mv_data = &last;
1381 key.mv_size = sizeof(last);
1382 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1385 last = *(txnid_t *)key.mv_data;
1389 oldest = mdb_find_oldest(txn);
1391 if (oldest > last) {
1392 /* It's usable, grab it.
1396 if (!txn->mt_env->me_pglast) {
1397 mdb_node_read(txn, leaf, &data);
1399 idl = (MDB_ID *) data.mv_data;
1400 /* We might have a zero-length IDL due to freelist growth
1401 * during a prior commit
1404 txn->mt_env->me_pglast = last;
1407 mop = mdb_midl_alloc(idl[0]);
1410 txn->mt_env->me_pglast = last;
1411 txn->mt_env->me_pghead = mop;
1412 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1417 DPRINTF("IDL read txn %zu root %zu num %zu",
1418 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1419 for (i = idl[0]; i; i--)
1420 DPRINTF("IDL %zu", idl[i]);
1426 if (txn->mt_env->me_pghead) {
1427 pgno_t *mop = txn->mt_env->me_pghead;
1430 int retry = 1, readit = 0, n2 = num-1;
1431 unsigned int i, j, k;
1433 /* If current list is too short, must fetch more and coalesce */
1434 if (mop[0] < (unsigned)num)
1437 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1439 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1440 /* If on freelist, don't try to read more. If what we have
1441 * right now isn't enough just use new pages.
1442 * TODO: get all of this working. Many circular dependencies...
1444 if (mc->mc_dbi == FREE_DBI) {
1453 last = txn->mt_env->me_pglast + 1;
1455 /* We haven't hit the readers list yet? */
1457 oldest = mdb_find_oldest(txn);
1460 /* There's nothing we can use on the freelist */
1461 if (oldest - last < 1)
1464 key.mv_data = &last;
1465 key.mv_size = sizeof(last);
1466 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1468 if (rc == MDB_NOTFOUND)
1472 last = *(txnid_t*)key.mv_data;
1475 idl = (MDB_ID *) data.mv_data;
1476 mop2 = mdb_midl_alloc(idl[0] + mop[0]);
1479 /* merge in sorted order */
1480 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1482 while (i>0 || j>0) {
1483 if (i && idl[i] < mop[j])
1484 mop2[k--] = idl[i--];
1486 mop2[k--] = mop[j--];
1488 txn->mt_env->me_pglast = last;
1489 mdb_midl_free(txn->mt_env->me_pghead);
1490 txn->mt_env->me_pghead = mop2;
1492 /* Keep trying to read until we have enough */
1493 if (mop[0] < (unsigned)num) {
1498 /* current list has enough pages, but are they contiguous? */
1499 for (i=mop[0]; i>=(unsigned)num; i--) {
1500 if (mop[i-n2] == mop[i] + n2) {
1503 /* move any stragglers down */
1504 for (j=i+num; j<=mop[0]; j++)
1511 /* Stop if we succeeded, or no retries */
1512 if (!retry || pgno != P_INVALID)
1518 /* peel pages off tail, so we only have to truncate the list */
1519 pgno = MDB_IDL_LAST(mop);
1522 if (MDB_IDL_IS_ZERO(mop)) {
1523 mdb_midl_free(txn->mt_env->me_pghead);
1524 txn->mt_env->me_pghead = NULL;
1529 if (pgno == P_INVALID) {
1530 /* DB size is maxed out */
1531 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1532 DPUTS("DB size maxed out");
1533 return MDB_MAP_FULL;
1536 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1537 if (pgno == P_INVALID) {
1538 pgno = txn->mt_next_pgno;
1539 txn->mt_next_pgno += num;
1541 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1544 if (!(np = mdb_page_malloc(mc, num)))
1546 if (pgno == P_INVALID) {
1547 np->mp_pgno = txn->mt_next_pgno;
1548 txn->mt_next_pgno += num;
1553 mid.mid = np->mp_pgno;
1555 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1556 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1558 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1560 txn->mt_dirty_room--;
1566 /** Copy a page: avoid copying unused portions of the page.
1567 * @param[in] dst page to copy into
1568 * @param[in] src page to copy from
1569 * @param[in] psize size of a page
1572 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1574 dst->mp_flags = src->mp_flags | P_DIRTY;
1575 dst->mp_pages = src->mp_pages;
1577 if (IS_LEAF2(src)) {
1578 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1580 unsigned int i, nkeys = NUMKEYS(src);
1581 for (i=0; i<nkeys; i++)
1582 dst->mp_ptrs[i] = src->mp_ptrs[i];
1583 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1584 psize - src->mp_upper);
1588 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1589 * @param[in] mc cursor pointing to the page to be touched
1590 * @return 0 on success, non-zero on failure.
1593 mdb_page_touch(MDB_cursor *mc)
1595 MDB_page *mp = mc->mc_pg[mc->mc_top];
1599 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1601 if ((rc = mdb_page_alloc(mc, 1, &np)))
1603 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1604 assert(mp->mp_pgno != np->mp_pgno);
1605 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1607 /* If page isn't full, just copy the used portion */
1608 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1611 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1613 np->mp_flags |= P_DIRTY;
1618 /* Adjust other cursors pointing to mp */
1619 if (mc->mc_flags & C_SUB) {
1620 MDB_cursor *m2, *m3;
1621 MDB_dbi dbi = mc->mc_dbi-1;
1623 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1624 if (m2 == mc) continue;
1625 m3 = &m2->mc_xcursor->mx_cursor;
1626 if (m3->mc_snum < mc->mc_snum) continue;
1627 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1628 m3->mc_pg[mc->mc_top] = mp;
1634 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1635 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1636 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1637 m2->mc_pg[mc->mc_top] = mp;
1638 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1639 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
1640 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
1641 if (!(leaf->mn_flags & F_SUBDATA)) {
1642 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1648 mc->mc_pg[mc->mc_top] = mp;
1649 /** If this page has a parent, update the parent to point to
1653 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1655 mc->mc_db->md_root = mp->mp_pgno;
1656 } else if (mc->mc_txn->mt_parent && !(mp->mp_flags & P_SUBP)) {
1658 MDB_ID2 mid, *dl = mc->mc_txn->mt_u.dirty_list;
1659 /* If txn has a parent, make sure the page is in our
1663 unsigned x = mdb_mid2l_search(dl, mp->mp_pgno);
1664 if (x <= dl[0].mid && dl[x].mid == mp->mp_pgno) {
1667 mc->mc_pg[mc->mc_top] = np;
1671 assert(dl[0].mid < MDB_IDL_UM_MAX);
1673 np = mdb_page_malloc(mc, 1);
1676 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1677 mid.mid = np->mp_pgno;
1679 mdb_mid2l_insert(dl, &mid);
1687 mdb_env_sync(MDB_env *env, int force)
1690 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1691 if (env->me_flags & MDB_WRITEMAP) {
1692 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1693 ? MS_ASYNC : MS_SYNC;
1694 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1697 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1701 if (MDB_FDATASYNC(env->me_fd))
1708 /** Make shadow copies of all of parent txn's cursors */
1710 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1712 MDB_cursor *mc, *m2;
1713 unsigned int i, j, size;
1715 for (i=0;i<src->mt_numdbs; i++) {
1716 if (src->mt_cursors[i]) {
1717 size = sizeof(MDB_cursor);
1718 if (src->mt_cursors[i]->mc_xcursor)
1719 size += sizeof(MDB_xcursor);
1720 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1727 mc->mc_db = &dst->mt_dbs[i];
1728 mc->mc_dbx = m2->mc_dbx;
1729 mc->mc_dbflag = &dst->mt_dbflags[i];
1730 mc->mc_snum = m2->mc_snum;
1731 mc->mc_top = m2->mc_top;
1732 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1733 for (j=0; j<mc->mc_snum; j++) {
1734 mc->mc_pg[j] = m2->mc_pg[j];
1735 mc->mc_ki[j] = m2->mc_ki[j];
1737 if (m2->mc_xcursor) {
1738 MDB_xcursor *mx, *mx2;
1739 mx = (MDB_xcursor *)(mc+1);
1740 mc->mc_xcursor = mx;
1741 mx2 = m2->mc_xcursor;
1742 mx->mx_db = mx2->mx_db;
1743 mx->mx_dbx = mx2->mx_dbx;
1744 mx->mx_dbflag = mx2->mx_dbflag;
1745 mx->mx_cursor.mc_txn = dst;
1746 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1747 mx->mx_cursor.mc_db = &mx->mx_db;
1748 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1749 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1750 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1751 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1752 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1753 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1754 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1755 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1758 mc->mc_xcursor = NULL;
1760 mc->mc_next = dst->mt_cursors[i];
1761 dst->mt_cursors[i] = mc;
1768 /** Close this write txn's cursors, after optionally merging its shadow
1769 * cursors back into parent's.
1770 * @param[in] txn the transaction handle.
1771 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1772 * @return 0 on success, non-zero on failure.
1775 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1777 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1780 for (i = txn->mt_numdbs; --i >= 0; ) {
1781 for (mc = cursors[i]; mc; mc = next) {
1783 if (mc->mc_flags & merge) {
1784 MDB_cursor *m2 = mc->mc_orig;
1785 m2->mc_snum = mc->mc_snum;
1786 m2->mc_top = mc->mc_top;
1787 for (j = mc->mc_snum; --j >= 0; ) {
1788 m2->mc_pg[j] = mc->mc_pg[j];
1789 m2->mc_ki[j] = mc->mc_ki[j];
1792 if (mc->mc_flags & C_ALLOCD)
1800 mdb_txn_reset0(MDB_txn *txn);
1802 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1803 * @param[in] txn the transaction handle to initialize
1804 * @return 0 on success, non-zero on failure.
1807 mdb_txn_renew0(MDB_txn *txn)
1809 MDB_env *env = txn->mt_env;
1812 int rc, new_notls = 0;
1815 txn->mt_numdbs = env->me_numdbs;
1816 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1818 if (txn->mt_flags & MDB_TXN_RDONLY) {
1819 if (!env->me_txns) {
1820 i = mdb_env_pick_meta(env);
1821 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1822 txn->mt_u.reader = NULL;
1824 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1825 pthread_getspecific(env->me_txkey);
1827 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1828 return MDB_BAD_RSLOT;
1830 pid_t pid = env->me_pid;
1831 pthread_t tid = pthread_self();
1834 for (i=0; i<env->me_txns->mti_numreaders; i++)
1835 if (env->me_txns->mti_readers[i].mr_pid == 0)
1837 if (i == env->me_maxreaders) {
1838 UNLOCK_MUTEX_R(env);
1839 return MDB_READERS_FULL;
1841 env->me_txns->mti_readers[i].mr_pid = pid;
1842 env->me_txns->mti_readers[i].mr_tid = tid;
1843 if (i >= env->me_txns->mti_numreaders)
1844 env->me_txns->mti_numreaders = i+1;
1845 /* Save numreaders for un-mutexed mdb_env_close() */
1846 env->me_numreaders = env->me_txns->mti_numreaders;
1847 UNLOCK_MUTEX_R(env);
1848 r = &env->me_txns->mti_readers[i];
1849 new_notls = (env->me_flags & MDB_NOTLS);
1850 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1855 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1856 txn->mt_u.reader = r;
1858 txn->mt_toggle = txn->mt_txnid & 1;
1859 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1863 txn->mt_txnid = env->me_txns->mti_txnid;
1864 txn->mt_toggle = txn->mt_txnid & 1;
1865 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1868 if (txn->mt_txnid == mdb_debug_start)
1871 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1872 txn->mt_u.dirty_list = env->me_dirty_list;
1873 txn->mt_u.dirty_list[0].mid = 0;
1874 txn->mt_free_pgs = env->me_free_pgs;
1875 txn->mt_free_pgs[0] = 0;
1879 /* Copy the DB info and flags */
1880 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1881 for (i=2; i<txn->mt_numdbs; i++) {
1882 x = env->me_dbflags[i];
1883 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1884 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1886 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1888 if (env->me_maxpg < txn->mt_next_pgno) {
1889 mdb_txn_reset0(txn);
1891 txn->mt_u.reader->mr_pid = 0;
1892 txn->mt_u.reader = NULL;
1894 return MDB_MAP_RESIZED;
1901 mdb_txn_renew(MDB_txn *txn)
1905 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1908 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1909 DPUTS("environment had fatal error, must shutdown!");
1913 rc = mdb_txn_renew0(txn);
1914 if (rc == MDB_SUCCESS) {
1915 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1916 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1917 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1923 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1927 int rc, size, tsize = sizeof(MDB_txn);
1929 if (env->me_flags & MDB_FATAL_ERROR) {
1930 DPUTS("environment had fatal error, must shutdown!");
1933 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1936 /* Nested transactions: Max 1 child, write txns only, no writemap */
1937 if (parent->mt_child ||
1938 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1939 (env->me_flags & MDB_WRITEMAP))
1943 tsize = sizeof(MDB_ntxn);
1945 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1946 if (!(flags & MDB_RDONLY))
1947 size += env->me_maxdbs * sizeof(MDB_cursor *);
1949 if ((txn = calloc(1, size)) == NULL) {
1950 DPRINTF("calloc: %s", strerror(ErrCode()));
1953 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1954 if (flags & MDB_RDONLY) {
1955 txn->mt_flags |= MDB_TXN_RDONLY;
1956 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1958 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1959 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1965 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1966 if (!txn->mt_u.dirty_list ||
1967 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1969 free(txn->mt_u.dirty_list);
1973 txn->mt_txnid = parent->mt_txnid;
1974 txn->mt_toggle = parent->mt_toggle;
1975 txn->mt_dirty_room = parent->mt_dirty_room;
1976 txn->mt_u.dirty_list[0].mid = 0;
1977 txn->mt_free_pgs[0] = 0;
1978 txn->mt_next_pgno = parent->mt_next_pgno;
1979 parent->mt_child = txn;
1980 txn->mt_parent = parent;
1981 txn->mt_numdbs = parent->mt_numdbs;
1982 txn->mt_dbxs = parent->mt_dbxs;
1983 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1984 /* Copy parent's mt_dbflags, but clear DB_NEW */
1985 for (i=0; i<txn->mt_numdbs; i++)
1986 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1988 ntxn = (MDB_ntxn *)txn;
1989 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1990 if (env->me_pghead) {
1991 size = MDB_IDL_SIZEOF(env->me_pghead);
1992 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1994 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1999 rc = mdb_cursor_shadow(parent, txn);
2001 mdb_txn_reset0(txn);
2003 rc = mdb_txn_renew0(txn);
2009 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2010 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2011 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2017 /** Export or close DBI handles opened in this txn. */
2019 mdb_dbis_update(MDB_txn *txn, int keep)
2022 MDB_dbi n = txn->mt_numdbs;
2023 MDB_env *env = txn->mt_env;
2024 unsigned char *tdbflags = txn->mt_dbflags;
2026 for (i = n; --i >= 2;) {
2027 if (tdbflags[i] & DB_NEW) {
2029 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2031 char *ptr = env->me_dbxs[i].md_name.mv_data;
2032 env->me_dbxs[i].md_name.mv_data = NULL;
2033 env->me_dbxs[i].md_name.mv_size = 0;
2034 env->me_dbflags[i] = 0;
2039 if (keep && env->me_numdbs < n)
2043 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2044 * May be called twice for readonly txns: First reset it, then abort.
2045 * @param[in] txn the transaction handle to reset
2048 mdb_txn_reset0(MDB_txn *txn)
2050 MDB_env *env = txn->mt_env;
2052 /* Close any DBI handles opened in this txn */
2053 mdb_dbis_update(txn, 0);
2055 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2056 if (txn->mt_u.reader) {
2057 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2058 if (!(env->me_flags & MDB_NOTLS))
2059 txn->mt_u.reader = NULL; /* txn does not own reader */
2061 txn->mt_numdbs = 0; /* close nothing if called again */
2062 txn->mt_dbxs = NULL; /* mark txn as reset */
2064 mdb_cursors_close(txn, 0);
2066 if (!(env->me_flags & MDB_WRITEMAP)) {
2067 mdb_dlist_free(txn);
2069 mdb_midl_free(env->me_pghead);
2071 if (txn->mt_parent) {
2072 txn->mt_parent->mt_child = NULL;
2073 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2074 mdb_midl_free(txn->mt_free_pgs);
2075 free(txn->mt_u.dirty_list);
2078 if (mdb_midl_shrink(&txn->mt_free_pgs))
2079 env->me_free_pgs = txn->mt_free_pgs;
2082 txn->mt_env->me_pghead = NULL;
2083 txn->mt_env->me_pglast = 0;
2086 /* The writer mutex was locked in mdb_txn_begin. */
2087 UNLOCK_MUTEX_W(env);
2092 mdb_txn_reset(MDB_txn *txn)
2097 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2098 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2099 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2101 /* This call is only valid for read-only txns */
2102 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2105 mdb_txn_reset0(txn);
2109 mdb_txn_abort(MDB_txn *txn)
2114 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2115 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2116 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2119 mdb_txn_abort(txn->mt_child);
2121 mdb_txn_reset0(txn);
2122 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2123 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2124 txn->mt_u.reader->mr_pid = 0;
2129 /** Save the freelist as of this transaction to the freeDB.
2130 * This changes the freelist. Keep trying until it stabilizes.
2133 mdb_freelist_save(MDB_txn *txn)
2135 /* env->me_pghead[] can grow and shrink during this call.
2136 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2137 * Page numbers cannot disappear from txn->mt_free_pgs[].
2140 MDB_env *env = txn->mt_env;
2141 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2142 txnid_t pglast = 0, head_id = 0;
2143 pgno_t freecnt = 0, *free_pgs, *mop;
2144 ssize_t head_room = 0, total_room = 0, mop_len;
2146 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2148 if (env->me_pghead || env->me_pglast) {
2149 /* Make sure first page of freeDB is touched and on freelist */
2150 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2151 if (rc && rc != MDB_NOTFOUND)
2156 /* Come back here after each Put() in case freelist changed */
2159 /* If using records from freeDB which we have not yet
2160 * deleted, delete them and any we reserved for me_pghead.
2162 while (pglast < env->me_pglast) {
2163 rc = mdb_cursor_first(&mc, &key, NULL);
2166 pglast = head_id = *(txnid_t *)key.mv_data;
2167 total_room = head_room = 0;
2168 assert(pglast <= env->me_pglast);
2169 rc = mdb_cursor_del(&mc, 0);
2174 /* Save the IDL of pages freed by this txn, to a single record */
2175 if (freecnt < txn->mt_free_pgs[0]) {
2177 /* Make sure last page of freeDB is touched and on freelist */
2178 key.mv_size = MDB_MAXKEYSIZE+1;
2180 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2181 if (rc && rc != MDB_NOTFOUND)
2184 free_pgs = txn->mt_free_pgs;
2185 /* Write to last page of freeDB */
2186 key.mv_size = sizeof(txn->mt_txnid);
2187 key.mv_data = &txn->mt_txnid;
2189 freecnt = free_pgs[0];
2190 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2191 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2194 /* Retry if mt_free_pgs[] grew during the Put() */
2195 free_pgs = txn->mt_free_pgs;
2196 } while (freecnt < free_pgs[0]);
2197 mdb_midl_sort(free_pgs);
2198 memcpy(data.mv_data, free_pgs, data.mv_size);
2201 unsigned int i = free_pgs[0];
2202 DPRINTF("IDL write txn %zu root %zu num %u",
2203 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2205 DPRINTF("IDL %zu", free_pgs[i]);
2211 mop = env->me_pghead;
2212 mop_len = mop ? mop[0] : 0;
2214 /* Reserve records for me_pghead[]. Split it if multi-page,
2215 * to avoid searching freeDB for a page range. Use keys in
2216 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2218 if (total_room >= mop_len) {
2219 if (total_room == mop_len || --more < 0)
2221 } else if (head_room >= maxfree_1pg && head_id > 1) {
2222 /* Keep current record (overflow page), add a new one */
2226 /* (Re)write {key = head_id, IDL length = head_room} */
2227 total_room -= head_room;
2228 head_room = mop_len - total_room;
2229 if (head_room > maxfree_1pg && head_id > 1) {
2230 /* Overflow multi-page for part of me_pghead */
2231 head_room /= head_id; /* amortize page sizes */
2232 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2233 } else if (head_room < 0) {
2234 /* Rare case, not bothering to delete this record */
2237 key.mv_size = sizeof(head_id);
2238 key.mv_data = &head_id;
2239 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2240 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2243 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2244 total_room += head_room;
2247 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2248 * so they cannot rearrange anything, just read the destinations.
2255 rc = mdb_cursor_first(&mc, &key, &data);
2256 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2257 MDB_IDL dest = data.mv_data;
2258 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2260 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2264 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2265 if (! (mop_len -= len))
2273 mdb_txn_commit(MDB_txn *txn)
2283 assert(txn != NULL);
2284 assert(txn->mt_env != NULL);
2286 if (txn->mt_child) {
2287 rc = mdb_txn_commit(txn->mt_child);
2288 txn->mt_child = NULL;
2297 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2298 mdb_dbis_update(txn, 1);
2299 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2304 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2305 DPUTS("error flag is set, can't commit");
2307 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2312 if (txn->mt_parent) {
2313 MDB_txn *parent = txn->mt_parent;
2317 /* Append our free list to parent's */
2318 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2322 mdb_midl_free(txn->mt_free_pgs);
2324 parent->mt_next_pgno = txn->mt_next_pgno;
2325 parent->mt_flags = txn->mt_flags;
2327 /* Merge our cursors into parent's and close them */
2328 mdb_cursors_close(txn, C_SHADOW);
2330 /* Update parent's DB table. */
2331 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2332 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2333 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2334 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2335 for (i=2; i<txn->mt_numdbs; i++) {
2336 /* preserve parent's DB_NEW status */
2337 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2338 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2341 dst = txn->mt_parent->mt_u.dirty_list;
2342 src = txn->mt_u.dirty_list;
2343 /* Find len = length of merging our dirty list with parent's */
2345 dst[0].mid = 0; /* simplify loops */
2346 if (parent->mt_parent) {
2347 len = x + src[0].mid;
2348 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2349 for (i = x; y && i; y--) {
2350 pgno_t yp = src[y].mid;
2351 while (yp < dst[i].mid)
2353 if (yp == dst[i].mid) {
2358 } else { /* Simplify the above for single-ancestor case */
2359 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2361 /* Merge our dirty list with parent's */
2363 for (i = len; y; dst[i--] = src[y--]) {
2364 pgno_t yp = src[y].mid;
2365 while (yp < dst[x].mid)
2366 dst[i--] = dst[x--];
2367 if (yp == dst[x].mid)
2368 free(dst[x--].mptr);
2372 free(txn->mt_u.dirty_list);
2373 parent->mt_dirty_room = txn->mt_dirty_room;
2375 txn->mt_parent->mt_child = NULL;
2376 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2381 if (txn != env->me_txn) {
2382 DPUTS("attempt to commit unknown transaction");
2387 mdb_cursors_close(txn, 0);
2389 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2392 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2393 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2395 /* Update DB root pointers */
2396 if (txn->mt_numdbs > 2) {
2400 data.mv_size = sizeof(MDB_db);
2402 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2403 for (i = 2; i < txn->mt_numdbs; i++) {
2404 if (txn->mt_dbflags[i] & DB_DIRTY) {
2405 data.mv_data = &txn->mt_dbs[i];
2406 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2413 rc = mdb_freelist_save(txn);
2417 mdb_midl_free(env->me_pghead);
2418 env->me_pghead = NULL;
2419 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2420 if (mdb_midl_shrink(&txn->mt_free_pgs))
2421 env->me_free_pgs = txn->mt_free_pgs;
2428 if (env->me_flags & MDB_WRITEMAP) {
2429 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2430 dp = txn->mt_u.dirty_list[i].mptr;
2431 /* clear dirty flag */
2432 dp->mp_flags &= ~P_DIRTY;
2434 txn->mt_u.dirty_list[0].mid = 0;
2438 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2444 /* Windows actually supports scatter/gather I/O, but only on
2445 * unbuffered file handles. Since we're relying on the OS page
2446 * cache for all our data, that's self-defeating. So we just
2447 * write pages one at a time. We use the ov structure to set
2448 * the write offset, to at least save the overhead of a Seek
2452 memset(&ov, 0, sizeof(ov));
2453 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2455 dp = txn->mt_u.dirty_list[i].mptr;
2456 DPRINTF("committing page %zu", dp->mp_pgno);
2457 size = dp->mp_pgno * env->me_psize;
2458 ov.Offset = size & 0xffffffff;
2459 ov.OffsetHigh = size >> 16;
2460 ov.OffsetHigh >>= 16;
2461 /* clear dirty flag */
2462 dp->mp_flags &= ~P_DIRTY;
2463 wsize = env->me_psize;
2464 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2465 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2468 DPRINTF("WriteFile: %d", n);
2475 struct iovec iov[MDB_COMMIT_PAGES];
2479 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2480 dp = txn->mt_u.dirty_list[i].mptr;
2481 if (dp->mp_pgno != next) {
2483 rc = writev(env->me_fd, iov, n);
2487 DPUTS("short write, filesystem full?");
2489 DPRINTF("writev: %s", strerror(n));
2496 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2499 DPRINTF("committing page %zu", dp->mp_pgno);
2500 iov[n].iov_len = env->me_psize;
2501 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2502 iov[n].iov_base = (char *)dp;
2503 size += iov[n].iov_len;
2504 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2505 /* clear dirty flag */
2506 dp->mp_flags &= ~P_DIRTY;
2507 if (++n >= MDB_COMMIT_PAGES) {
2517 rc = writev(env->me_fd, iov, n);
2521 DPUTS("short write, filesystem full?");
2523 DPRINTF("writev: %s", strerror(n));
2530 mdb_dlist_free(txn);
2533 if ((n = mdb_env_sync(env, 0)) != 0 ||
2534 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2542 mdb_dbis_update(txn, 1);
2544 UNLOCK_MUTEX_W(env);
2554 /** Read the environment parameters of a DB environment before
2555 * mapping it into memory.
2556 * @param[in] env the environment handle
2557 * @param[out] meta address of where to store the meta information
2558 * @return 0 on success, non-zero on failure.
2561 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2568 /* We don't know the page size yet, so use a minimum value.
2569 * Read both meta pages so we can use the latest one.
2572 for (i=0; i<2; i++) {
2574 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2576 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2581 else if (rc != MDB_PAGESIZE) {
2585 DPRINTF("read: %s", strerror(err));
2589 p = (MDB_page *)&pbuf;
2591 if (!F_ISSET(p->mp_flags, P_META)) {
2592 DPRINTF("page %zu not a meta page", p->mp_pgno);
2597 if (m->mm_magic != MDB_MAGIC) {
2598 DPUTS("meta has invalid magic");
2602 if (m->mm_version != MDB_VERSION) {
2603 DPRINTF("database is version %u, expected version %u",
2604 m->mm_version, MDB_VERSION);
2605 return MDB_VERSION_MISMATCH;
2609 if (m->mm_txnid > meta->mm_txnid)
2610 memcpy(meta, m, sizeof(*m));
2612 memcpy(meta, m, sizeof(*m));
2614 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2616 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2624 /** Write the environment parameters of a freshly created DB environment.
2625 * @param[in] env the environment handle
2626 * @param[out] meta address of where to store the meta information
2627 * @return 0 on success, non-zero on failure.
2630 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2637 DPUTS("writing new meta page");
2639 GET_PAGESIZE(psize);
2641 meta->mm_magic = MDB_MAGIC;
2642 meta->mm_version = MDB_VERSION;
2643 meta->mm_mapsize = env->me_mapsize;
2644 meta->mm_psize = psize;
2645 meta->mm_last_pg = 1;
2646 meta->mm_flags = env->me_flags & 0xffff;
2647 meta->mm_flags |= MDB_INTEGERKEY;
2648 meta->mm_dbs[0].md_root = P_INVALID;
2649 meta->mm_dbs[1].md_root = P_INVALID;
2651 p = calloc(2, psize);
2653 p->mp_flags = P_META;
2656 memcpy(m, meta, sizeof(*meta));
2658 q = (MDB_page *)((char *)p + psize);
2661 q->mp_flags = P_META;
2664 memcpy(m, meta, sizeof(*meta));
2669 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2670 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2671 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2674 lseek(env->me_fd, 0, SEEK_SET);
2675 rc = write(env->me_fd, p, psize * 2);
2676 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2682 /** Update the environment info to commit a transaction.
2683 * @param[in] txn the transaction that's being committed
2684 * @return 0 on success, non-zero on failure.
2687 mdb_env_write_meta(MDB_txn *txn)
2690 MDB_meta meta, metab, *mp;
2692 int rc, len, toggle;
2699 assert(txn != NULL);
2700 assert(txn->mt_env != NULL);
2702 toggle = !txn->mt_toggle;
2703 DPRINTF("writing meta page %d for root page %zu",
2704 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2707 mp = env->me_metas[toggle];
2709 if (env->me_flags & MDB_WRITEMAP) {
2710 /* Persist any increases of mapsize config */
2711 if (env->me_mapsize > mp->mm_mapsize)
2712 mp->mm_mapsize = env->me_mapsize;
2713 mp->mm_dbs[0] = txn->mt_dbs[0];
2714 mp->mm_dbs[1] = txn->mt_dbs[1];
2715 mp->mm_last_pg = txn->mt_next_pgno - 1;
2716 mp->mm_txnid = txn->mt_txnid;
2717 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2718 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2721 ptr += env->me_psize;
2722 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2729 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2730 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2732 ptr = (char *)&meta;
2733 if (env->me_mapsize > mp->mm_mapsize) {
2734 /* Persist any increases of mapsize config */
2735 meta.mm_mapsize = env->me_mapsize;
2736 off = offsetof(MDB_meta, mm_mapsize);
2738 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2740 len = sizeof(MDB_meta) - off;
2743 meta.mm_dbs[0] = txn->mt_dbs[0];
2744 meta.mm_dbs[1] = txn->mt_dbs[1];
2745 meta.mm_last_pg = txn->mt_next_pgno - 1;
2746 meta.mm_txnid = txn->mt_txnid;
2749 off += env->me_psize;
2752 /* Write to the SYNC fd */
2753 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2754 env->me_fd : env->me_mfd;
2757 memset(&ov, 0, sizeof(ov));
2759 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2762 rc = pwrite(mfd, ptr, len, off);
2767 DPUTS("write failed, disk error?");
2768 /* On a failure, the pagecache still contains the new data.
2769 * Write some old data back, to prevent it from being used.
2770 * Use the non-SYNC fd; we know it will fail anyway.
2772 meta.mm_last_pg = metab.mm_last_pg;
2773 meta.mm_txnid = metab.mm_txnid;
2775 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2777 r2 = pwrite(env->me_fd, ptr, len, off);
2780 env->me_flags |= MDB_FATAL_ERROR;
2784 /* Memory ordering issues are irrelevant; since the entire writer
2785 * is wrapped by wmutex, all of these changes will become visible
2786 * after the wmutex is unlocked. Since the DB is multi-version,
2787 * readers will get consistent data regardless of how fresh or
2788 * how stale their view of these values is.
2790 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2795 /** Check both meta pages to see which one is newer.
2796 * @param[in] env the environment handle
2797 * @return meta toggle (0 or 1).
2800 mdb_env_pick_meta(const MDB_env *env)
2802 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2806 mdb_env_create(MDB_env **env)
2810 e = calloc(1, sizeof(MDB_env));
2814 e->me_maxreaders = DEFAULT_READERS;
2815 e->me_maxdbs = e->me_numdbs = 2;
2816 e->me_fd = INVALID_HANDLE_VALUE;
2817 e->me_lfd = INVALID_HANDLE_VALUE;
2818 e->me_mfd = INVALID_HANDLE_VALUE;
2819 #ifdef MDB_USE_POSIX_SEM
2820 e->me_rmutex = SEM_FAILED;
2821 e->me_wmutex = SEM_FAILED;
2823 e->me_pid = getpid();
2824 VGMEMP_CREATE(e,0,0);
2830 mdb_env_set_mapsize(MDB_env *env, size_t size)
2834 env->me_mapsize = size;
2836 env->me_maxpg = env->me_mapsize / env->me_psize;
2841 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2845 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2850 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2852 if (env->me_map || readers < 1)
2854 env->me_maxreaders = readers;
2859 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2861 if (!env || !readers)
2863 *readers = env->me_maxreaders;
2867 /** Further setup required for opening an MDB environment
2870 mdb_env_open2(MDB_env *env)
2872 unsigned int flags = env->me_flags;
2873 int i, newenv = 0, prot;
2877 memset(&meta, 0, sizeof(meta));
2879 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2882 DPUTS("new mdbenv");
2886 /* Was a mapsize configured? */
2887 if (!env->me_mapsize) {
2888 /* If this is a new environment, take the default,
2889 * else use the size recorded in the existing env.
2891 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2892 } else if (env->me_mapsize < meta.mm_mapsize) {
2893 /* If the configured size is smaller, make sure it's
2894 * still big enough. Silently round up to minimum if not.
2896 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2897 if (env->me_mapsize < minsize)
2898 env->me_mapsize = minsize;
2904 LONG sizelo, sizehi;
2905 sizelo = env->me_mapsize & 0xffffffff;
2906 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2908 /* Windows won't create mappings for zero length files.
2909 * Just allocate the maxsize right now.
2912 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2913 if (!SetEndOfFile(env->me_fd))
2915 SetFilePointer(env->me_fd, 0, NULL, 0);
2917 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2918 PAGE_READWRITE : PAGE_READONLY,
2919 sizehi, sizelo, NULL);
2922 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2923 FILE_MAP_WRITE : FILE_MAP_READ,
2924 0, 0, env->me_mapsize, meta.mm_address);
2932 if (flags & MDB_WRITEMAP) {
2934 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2937 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2939 if (env->me_map == MAP_FAILED) {
2943 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2945 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2947 #ifdef POSIX_MADV_RANDOM
2948 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2949 #endif /* POSIX_MADV_RANDOM */
2950 #endif /* MADV_RANDOM */
2954 if (flags & MDB_FIXEDMAP)
2955 meta.mm_address = env->me_map;
2956 i = mdb_env_init_meta(env, &meta);
2957 if (i != MDB_SUCCESS) {
2960 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2961 /* Can happen because the address argument to mmap() is just a
2962 * hint. mmap() can pick another, e.g. if the range is in use.
2963 * The MAP_FIXED flag would prevent that, but then mmap could
2964 * instead unmap existing pages to make room for the new map.
2966 return EBUSY; /* TODO: Make a new MDB_* error code? */
2968 env->me_psize = meta.mm_psize;
2969 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2970 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2972 env->me_maxpg = env->me_mapsize / env->me_psize;
2974 p = (MDB_page *)env->me_map;
2975 env->me_metas[0] = METADATA(p);
2976 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2980 int toggle = mdb_env_pick_meta(env);
2981 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2983 DPRINTF("opened database version %u, pagesize %u",
2984 env->me_metas[0]->mm_version, env->me_psize);
2985 DPRINTF("using meta page %d", toggle);
2986 DPRINTF("depth: %u", db->md_depth);
2987 DPRINTF("entries: %zu", db->md_entries);
2988 DPRINTF("branch pages: %zu", db->md_branch_pages);
2989 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2990 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2991 DPRINTF("root: %zu", db->md_root);
2999 /** Release a reader thread's slot in the reader lock table.
3000 * This function is called automatically when a thread exits.
3001 * @param[in] ptr This points to the slot in the reader lock table.
3004 mdb_env_reader_dest(void *ptr)
3006 MDB_reader *reader = ptr;
3012 /** Junk for arranging thread-specific callbacks on Windows. This is
3013 * necessarily platform and compiler-specific. Windows supports up
3014 * to 1088 keys. Let's assume nobody opens more than 64 environments
3015 * in a single process, for now. They can override this if needed.
3017 #ifndef MAX_TLS_KEYS
3018 #define MAX_TLS_KEYS 64
3020 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3021 static int mdb_tls_nkeys;
3023 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3027 case DLL_PROCESS_ATTACH: break;
3028 case DLL_THREAD_ATTACH: break;
3029 case DLL_THREAD_DETACH:
3030 for (i=0; i<mdb_tls_nkeys; i++) {
3031 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3032 mdb_env_reader_dest(r);
3035 case DLL_PROCESS_DETACH: break;
3040 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3042 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3046 /* Force some symbol references.
3047 * _tls_used forces the linker to create the TLS directory if not already done
3048 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3050 #pragma comment(linker, "/INCLUDE:_tls_used")
3051 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3052 #pragma const_seg(".CRT$XLB")
3053 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3054 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3057 #pragma comment(linker, "/INCLUDE:__tls_used")
3058 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3059 #pragma data_seg(".CRT$XLB")
3060 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3062 #endif /* WIN 32/64 */
3063 #endif /* !__GNUC__ */
3066 /** Downgrade the exclusive lock on the region back to shared */
3068 mdb_env_share_locks(MDB_env *env, int *excl)
3070 int rc = 0, toggle = mdb_env_pick_meta(env);
3072 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3077 /* First acquire a shared lock. The Unlock will
3078 * then release the existing exclusive lock.
3080 memset(&ov, 0, sizeof(ov));
3081 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3084 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3090 struct flock lock_info;
3091 /* The shared lock replaces the existing lock */
3092 memset((void *)&lock_info, 0, sizeof(lock_info));
3093 lock_info.l_type = F_RDLCK;
3094 lock_info.l_whence = SEEK_SET;
3095 lock_info.l_start = 0;
3096 lock_info.l_len = 1;
3097 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3098 (rc = ErrCode()) == EINTR) ;
3099 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3106 /** Try to get exlusive lock, otherwise shared.
3107 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3110 mdb_env_excl_lock(MDB_env *env, int *excl)
3114 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3118 memset(&ov, 0, sizeof(ov));
3119 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3126 struct flock lock_info;
3127 memset((void *)&lock_info, 0, sizeof(lock_info));
3128 lock_info.l_type = F_WRLCK;
3129 lock_info.l_whence = SEEK_SET;
3130 lock_info.l_start = 0;
3131 lock_info.l_len = 1;
3132 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3133 (rc = ErrCode()) == EINTR) ;
3137 # ifdef MDB_USE_POSIX_SEM
3138 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3141 lock_info.l_type = F_RDLCK;
3142 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3143 (rc = ErrCode()) == EINTR) ;
3151 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3153 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3155 * @(#) $Revision: 5.1 $
3156 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3157 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3159 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3163 * Please do not copyright this code. This code is in the public domain.
3165 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3166 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3167 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3168 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3169 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3170 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3171 * PERFORMANCE OF THIS SOFTWARE.
3174 * chongo <Landon Curt Noll> /\oo/\
3175 * http://www.isthe.com/chongo/
3177 * Share and Enjoy! :-)
3180 typedef unsigned long long mdb_hash_t;
3181 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3183 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3184 * @param[in] str string to hash
3185 * @param[in] hval initial value for hash
3186 * @return 64 bit hash
3188 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3189 * hval arg on the first call.
3192 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3194 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3195 unsigned char *end = s + val->mv_size;
3197 * FNV-1a hash each octet of the string
3200 /* xor the bottom with the current octet */
3201 hval ^= (mdb_hash_t)*s++;
3203 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3204 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3205 (hval << 7) + (hval << 8) + (hval << 40);
3207 /* return our new hash value */
3211 /** Hash the string and output the hash in hex.
3212 * @param[in] str string to hash
3213 * @param[out] hexbuf an array of 17 chars to hold the hash
3216 mdb_hash_hex(MDB_val *val, char *hexbuf)
3219 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3220 for (i=0; i<8; i++) {
3221 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3227 /** Open and/or initialize the lock region for the environment.
3228 * @param[in] env The MDB environment.
3229 * @param[in] lpath The pathname of the file used for the lock region.
3230 * @param[in] mode The Unix permissions for the file, if we create it.
3231 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3232 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3233 * @return 0 on success, non-zero on failure.
3236 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3239 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3241 # define MDB_ERRCODE_ROFS EROFS
3242 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3243 # define MDB_CLOEXEC O_CLOEXEC
3246 # define MDB_CLOEXEC 0
3253 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3254 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3255 FILE_ATTRIBUTE_NORMAL, NULL);
3257 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3259 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3261 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3266 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3267 /* Lose record locks when exec*() */
3268 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3269 fcntl(env->me_lfd, F_SETFD, fdflags);
3272 if (!(env->me_flags & MDB_NOTLS)) {
3273 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3276 env->me_flags |= MDB_ENV_TXKEY;
3278 /* Windows TLS callbacks need help finding their TLS info. */
3279 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3283 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3287 /* Try to get exclusive lock. If we succeed, then
3288 * nobody is using the lock region and we should initialize it.
3290 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3293 size = GetFileSize(env->me_lfd, NULL);
3295 size = lseek(env->me_lfd, 0, SEEK_END);
3297 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3298 if (size < rsize && *excl > 0) {
3300 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3301 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3303 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3307 size = rsize - sizeof(MDB_txninfo);
3308 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3313 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3315 if (!mh) goto fail_errno;
3316 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3318 if (!env->me_txns) goto fail_errno;
3320 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3322 if (m == MAP_FAILED) goto fail_errno;
3328 BY_HANDLE_FILE_INFORMATION stbuf;
3337 if (!mdb_sec_inited) {
3338 InitializeSecurityDescriptor(&mdb_null_sd,
3339 SECURITY_DESCRIPTOR_REVISION);
3340 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3341 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3342 mdb_all_sa.bInheritHandle = FALSE;
3343 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3346 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3347 idbuf.volume = stbuf.dwVolumeSerialNumber;
3348 idbuf.nhigh = stbuf.nFileIndexHigh;
3349 idbuf.nlow = stbuf.nFileIndexLow;
3350 val.mv_data = &idbuf;
3351 val.mv_size = sizeof(idbuf);
3352 mdb_hash_hex(&val, hexbuf);
3353 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3354 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3355 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3356 if (!env->me_rmutex) goto fail_errno;
3357 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3358 if (!env->me_wmutex) goto fail_errno;
3359 #elif defined(MDB_USE_POSIX_SEM)
3368 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3369 idbuf.dev = stbuf.st_dev;
3370 idbuf.ino = stbuf.st_ino;
3371 val.mv_data = &idbuf;
3372 val.mv_size = sizeof(idbuf);
3373 mdb_hash_hex(&val, hexbuf);
3374 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3375 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3376 /* Clean up after a previous run, if needed: Try to
3377 * remove both semaphores before doing anything else.
3379 sem_unlink(env->me_txns->mti_rmname);
3380 sem_unlink(env->me_txns->mti_wmname);
3381 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3382 O_CREAT|O_EXCL, mode, 1);
3383 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3384 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3385 O_CREAT|O_EXCL, mode, 1);
3386 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3387 #else /* MDB_USE_POSIX_SEM */
3388 pthread_mutexattr_t mattr;
3390 if ((rc = pthread_mutexattr_init(&mattr))
3391 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3392 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3393 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3395 pthread_mutexattr_destroy(&mattr);
3396 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3398 env->me_txns->mti_version = MDB_VERSION;
3399 env->me_txns->mti_magic = MDB_MAGIC;
3400 env->me_txns->mti_txnid = 0;
3401 env->me_txns->mti_numreaders = 0;
3404 if (env->me_txns->mti_magic != MDB_MAGIC) {
3405 DPUTS("lock region has invalid magic");
3409 if (env->me_txns->mti_version != MDB_VERSION) {
3410 DPRINTF("lock region is version %u, expected version %u",
3411 env->me_txns->mti_version, MDB_VERSION);
3412 rc = MDB_VERSION_MISMATCH;
3416 if (rc != EACCES && rc != EAGAIN) {
3420 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3421 if (!env->me_rmutex) goto fail_errno;
3422 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3423 if (!env->me_wmutex) goto fail_errno;
3424 #elif defined(MDB_USE_POSIX_SEM)
3425 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3426 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3427 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3428 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3439 /** The name of the lock file in the DB environment */
3440 #define LOCKNAME "/lock.mdb"
3441 /** The name of the data file in the DB environment */
3442 #define DATANAME "/data.mdb"
3443 /** The suffix of the lock file when no subdir is used */
3444 #define LOCKSUFF "-lock"
3445 /** Only a subset of the @ref mdb_env flags can be changed
3446 * at runtime. Changing other flags requires closing the
3447 * environment and re-opening it with the new flags.
3449 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3450 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3453 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3455 int oflags, rc, len, excl = -1;
3456 char *lpath, *dpath;
3458 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3462 if (flags & MDB_NOSUBDIR) {
3463 rc = len + sizeof(LOCKSUFF) + len + 1;
3465 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3470 if (flags & MDB_NOSUBDIR) {
3471 dpath = lpath + len + sizeof(LOCKSUFF);
3472 sprintf(lpath, "%s" LOCKSUFF, path);
3473 strcpy(dpath, path);
3475 dpath = lpath + len + sizeof(LOCKNAME);
3476 sprintf(lpath, "%s" LOCKNAME, path);
3477 sprintf(dpath, "%s" DATANAME, path);
3481 flags |= env->me_flags;
3482 if (flags & MDB_RDONLY) {
3483 /* silently ignore WRITEMAP when we're only getting read access */
3484 flags &= ~MDB_WRITEMAP;
3486 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3487 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3490 env->me_flags = flags |= MDB_ENV_ACTIVE;
3494 env->me_path = strdup(path);
3495 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3496 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3497 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3502 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3507 if (F_ISSET(flags, MDB_RDONLY)) {
3508 oflags = GENERIC_READ;
3509 len = OPEN_EXISTING;
3511 oflags = GENERIC_READ|GENERIC_WRITE;
3514 mode = FILE_ATTRIBUTE_NORMAL;
3515 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3516 NULL, len, mode, NULL);
3518 if (F_ISSET(flags, MDB_RDONLY))
3521 oflags = O_RDWR | O_CREAT;
3523 env->me_fd = open(dpath, oflags, mode);
3525 if (env->me_fd == INVALID_HANDLE_VALUE) {
3530 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3531 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3532 env->me_mfd = env->me_fd;
3534 /* Synchronous fd for meta writes. Needed even with
3535 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3538 env->me_mfd = CreateFile(dpath, oflags,
3539 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3540 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3542 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3544 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3549 DPRINTF("opened dbenv %p", (void *) env);
3551 rc = mdb_env_share_locks(env, &excl);
3557 mdb_env_close0(env, excl);
3563 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3565 mdb_env_close0(MDB_env *env, int excl)
3569 if (!(env->me_flags & MDB_ENV_ACTIVE))
3572 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3573 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3574 free(env->me_dbxs[i].md_name.mv_data);
3576 free(env->me_dbflags);
3579 free(env->me_dirty_list);
3580 if (env->me_free_pgs)
3581 mdb_midl_free(env->me_free_pgs);
3583 if (env->me_flags & MDB_ENV_TXKEY) {
3584 pthread_key_delete(env->me_txkey);
3586 /* Delete our key from the global list */
3587 for (i=0; i<mdb_tls_nkeys; i++)
3588 if (mdb_tls_keys[i] == env->me_txkey) {
3589 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3597 munmap(env->me_map, env->me_mapsize);
3599 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3601 if (env->me_fd != INVALID_HANDLE_VALUE)
3604 pid_t pid = env->me_pid;
3605 /* Clearing readers is done in this function because
3606 * me_txkey with its destructor must be disabled first.
3608 for (i = env->me_numreaders; --i >= 0; )
3609 if (env->me_txns->mti_readers[i].mr_pid == pid)
3610 env->me_txns->mti_readers[i].mr_pid = 0;
3612 if (env->me_rmutex) {
3613 CloseHandle(env->me_rmutex);
3614 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3616 /* Windows automatically destroys the mutexes when
3617 * the last handle closes.
3619 #elif defined(MDB_USE_POSIX_SEM)
3620 if (env->me_rmutex != SEM_FAILED) {
3621 sem_close(env->me_rmutex);
3622 if (env->me_wmutex != SEM_FAILED)
3623 sem_close(env->me_wmutex);
3624 /* If we have the filelock: If we are the
3625 * only remaining user, clean up semaphores.
3628 mdb_env_excl_lock(env, &excl);
3630 sem_unlink(env->me_txns->mti_rmname);
3631 sem_unlink(env->me_txns->mti_wmname);
3635 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3637 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3640 /* Unlock the lockfile. Windows would have unlocked it
3641 * after closing anyway, but not necessarily at once.
3643 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3649 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3653 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3655 MDB_txn *txn = NULL;
3660 /* Do the lock/unlock of the reader mutex before starting the
3661 * write txn. Otherwise other read txns could block writers.
3663 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3668 /* We must start the actual read txn after blocking writers */
3669 mdb_txn_reset0(txn);
3671 /* Temporarily block writers until we snapshot the meta pages */
3674 rc = mdb_txn_renew0(txn);
3676 UNLOCK_MUTEX_W(env);
3681 wsize = env->me_psize * 2;
3685 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3686 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3689 rc = write(fd, env->me_map, wsize);
3690 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3693 UNLOCK_MUTEX_W(env);
3698 ptr = env->me_map + wsize;
3699 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3700 #define MAX_WRITE 2147483648U
3704 if (wsize > MAX_WRITE)
3708 rc = WriteFile(fd, ptr, w2, &len, NULL);
3709 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3718 if (wsize > MAX_WRITE)
3722 wres = write(fd, ptr, w2);
3723 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3736 mdb_env_copy(MDB_env *env, const char *path)
3740 HANDLE newfd = INVALID_HANDLE_VALUE;
3742 if (env->me_flags & MDB_NOSUBDIR) {
3743 lpath = (char *)path;
3746 len += sizeof(DATANAME);
3747 lpath = malloc(len);
3750 sprintf(lpath, "%s" DATANAME, path);
3753 /* The destination path must exist, but the destination file must not.
3754 * We don't want the OS to cache the writes, since the source data is
3755 * already in the OS cache.
3758 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3759 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3761 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3767 if (!(env->me_flags & MDB_NOSUBDIR))
3769 if (newfd == INVALID_HANDLE_VALUE) {
3774 #ifdef F_NOCACHE /* __APPLE__ */
3775 rc = fcntl(newfd, F_NOCACHE, 1);
3782 rc = mdb_env_copyfd(env, newfd);
3785 if (newfd != INVALID_HANDLE_VALUE)
3792 mdb_env_close(MDB_env *env)
3799 VGMEMP_DESTROY(env);
3800 while ((dp = env->me_dpages) != NULL) {
3801 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3802 env->me_dpages = dp->mp_next;
3806 mdb_env_close0(env, 0);
3810 /** Compare two items pointing at aligned size_t's */
3812 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3814 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3815 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3818 /** Compare two items pointing at aligned int's */
3820 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3822 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3823 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3826 /** Compare two items pointing at ints of unknown alignment.
3827 * Nodes and keys are guaranteed to be 2-byte aligned.
3830 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3832 #if BYTE_ORDER == LITTLE_ENDIAN
3833 unsigned short *u, *c;
3836 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3837 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3840 } while(!x && u > (unsigned short *)a->mv_data);
3843 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3847 /** Compare two items lexically */
3849 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3856 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3862 diff = memcmp(a->mv_data, b->mv_data, len);
3863 return diff ? diff : len_diff<0 ? -1 : len_diff;
3866 /** Compare two items in reverse byte order */
3868 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3870 const unsigned char *p1, *p2, *p1_lim;
3874 p1_lim = (const unsigned char *)a->mv_data;
3875 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3876 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3878 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3884 while (p1 > p1_lim) {
3885 diff = *--p1 - *--p2;
3889 return len_diff<0 ? -1 : len_diff;
3892 /** Search for key within a page, using binary search.
3893 * Returns the smallest entry larger or equal to the key.
3894 * If exactp is non-null, stores whether the found entry was an exact match
3895 * in *exactp (1 or 0).
3896 * Updates the cursor index with the index of the found entry.
3897 * If no entry larger or equal to the key is found, returns NULL.
3900 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3902 unsigned int i = 0, nkeys;
3905 MDB_page *mp = mc->mc_pg[mc->mc_top];
3906 MDB_node *node = NULL;
3911 nkeys = NUMKEYS(mp);
3916 COPY_PGNO(pgno, mp->mp_pgno);
3917 DPRINTF("searching %u keys in %s %spage %zu",
3918 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3925 low = IS_LEAF(mp) ? 0 : 1;
3927 cmp = mc->mc_dbx->md_cmp;
3929 /* Branch pages have no data, so if using integer keys,
3930 * alignment is guaranteed. Use faster mdb_cmp_int.
3932 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3933 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3940 nodekey.mv_size = mc->mc_db->md_pad;
3941 node = NODEPTR(mp, 0); /* fake */
3942 while (low <= high) {
3943 i = (low + high) >> 1;
3944 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3945 rc = cmp(key, &nodekey);
3946 DPRINTF("found leaf index %u [%s], rc = %i",
3947 i, DKEY(&nodekey), rc);
3956 while (low <= high) {
3957 i = (low + high) >> 1;
3959 node = NODEPTR(mp, i);
3960 nodekey.mv_size = NODEKSZ(node);
3961 nodekey.mv_data = NODEKEY(node);
3963 rc = cmp(key, &nodekey);
3966 DPRINTF("found leaf index %u [%s], rc = %i",
3967 i, DKEY(&nodekey), rc);
3969 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3970 i, DKEY(&nodekey), NODEPGNO(node), rc);
3981 if (rc > 0) { /* Found entry is less than the key. */
3982 i++; /* Skip to get the smallest entry larger than key. */
3984 node = NODEPTR(mp, i);
3987 *exactp = (rc == 0);
3988 /* store the key index */
3989 mc->mc_ki[mc->mc_top] = i;
3991 /* There is no entry larger or equal to the key. */
3994 /* nodeptr is fake for LEAF2 */
4000 mdb_cursor_adjust(MDB_cursor *mc, func)
4004 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4005 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4012 /** Pop a page off the top of the cursor's stack. */
4014 mdb_cursor_pop(MDB_cursor *mc)
4017 #ifndef MDB_DEBUG_SKIP
4018 MDB_page *top = mc->mc_pg[mc->mc_top];
4024 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4025 mc->mc_dbi, (void *) mc);
4029 /** Push a page onto the top of the cursor's stack. */
4031 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4033 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4034 mc->mc_dbi, (void *) mc);
4036 if (mc->mc_snum >= CURSOR_STACK) {
4037 assert(mc->mc_snum < CURSOR_STACK);
4038 return MDB_CURSOR_FULL;
4041 mc->mc_top = mc->mc_snum++;
4042 mc->mc_pg[mc->mc_top] = mp;
4043 mc->mc_ki[mc->mc_top] = 0;
4048 /** Find the address of the page corresponding to a given page number.
4049 * @param[in] txn the transaction for this access.
4050 * @param[in] pgno the page number for the page to retrieve.
4051 * @param[out] ret address of a pointer where the page's address will be stored.
4052 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4053 * @return 0 on success, non-zero on failure.
4056 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4061 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4062 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4067 MDB_ID2L dl = tx2->mt_u.dirty_list;
4069 unsigned x = mdb_mid2l_search(dl, pgno);
4070 if (x <= dl[0].mid && dl[x].mid == pgno) {
4076 } while ((tx2 = tx2->mt_parent) != NULL);
4079 if (pgno < txn->mt_next_pgno) {
4081 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4083 DPRINTF("page %zu not found", pgno);
4085 return MDB_PAGE_NOTFOUND;
4095 /** Search for the page a given key should be in.
4096 * Pushes parent pages on the cursor stack. This function continues a
4097 * search on a cursor that has already been initialized. (Usually by
4098 * #mdb_page_search() but also by #mdb_node_move().)
4099 * @param[in,out] mc the cursor for this operation.
4100 * @param[in] key the key to search for. If NULL, search for the lowest
4101 * page. (This is used by #mdb_cursor_first().)
4102 * @param[in] modify If true, visited pages are updated with new page numbers.
4103 * @return 0 on success, non-zero on failure.
4106 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4108 MDB_page *mp = mc->mc_pg[mc->mc_top];
4113 while (IS_BRANCH(mp)) {
4117 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4118 assert(NUMKEYS(mp) > 1);
4119 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4121 if (key == NULL) /* Initialize cursor to first page. */
4123 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4124 /* cursor to last page */
4128 node = mdb_node_search(mc, key, &exact);
4130 i = NUMKEYS(mp) - 1;
4132 i = mc->mc_ki[mc->mc_top];
4141 DPRINTF("following index %u for key [%s]",
4143 assert(i < NUMKEYS(mp));
4144 node = NODEPTR(mp, i);
4146 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4149 mc->mc_ki[mc->mc_top] = i;
4150 if ((rc = mdb_cursor_push(mc, mp)))
4154 if ((rc = mdb_page_touch(mc)) != 0)
4156 mp = mc->mc_pg[mc->mc_top];
4161 DPRINTF("internal error, index points to a %02X page!?",
4163 return MDB_CORRUPTED;
4166 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4167 key ? DKEY(key) : NULL);
4172 /** Search for the lowest key under the current branch page.
4173 * This just bypasses a NUMKEYS check in the current page
4174 * before calling mdb_page_search_root(), because the callers
4175 * are all in situations where the current page is known to
4179 mdb_page_search_lowest(MDB_cursor *mc)
4181 MDB_page *mp = mc->mc_pg[mc->mc_top];
4182 MDB_node *node = NODEPTR(mp, 0);
4185 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4188 mc->mc_ki[mc->mc_top] = 0;
4189 if ((rc = mdb_cursor_push(mc, mp)))
4191 return mdb_page_search_root(mc, NULL, 0);
4194 /** Search for the page a given key should be in.
4195 * Pushes parent pages on the cursor stack. This function just sets up
4196 * the search; it finds the root page for \b mc's database and sets this
4197 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4198 * called to complete the search.
4199 * @param[in,out] mc the cursor for this operation.
4200 * @param[in] key the key to search for. If NULL, search for the lowest
4201 * page. (This is used by #mdb_cursor_first().)
4202 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4203 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4204 * @return 0 on success, non-zero on failure.
4207 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4212 /* Make sure the txn is still viable, then find the root from
4213 * the txn's db table.
4215 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4216 DPUTS("transaction has failed, must abort");
4219 /* Make sure we're using an up-to-date root */
4220 if (mc->mc_dbi > MAIN_DBI) {
4221 if ((*mc->mc_dbflag & DB_STALE) ||
4222 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4224 unsigned char dbflag = 0;
4225 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4226 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4229 if (*mc->mc_dbflag & DB_STALE) {
4233 MDB_node *leaf = mdb_node_search(&mc2,
4234 &mc->mc_dbx->md_name, &exact);
4236 return MDB_NOTFOUND;
4237 mdb_node_read(mc->mc_txn, leaf, &data);
4238 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4240 /* The txn may not know this DBI, or another process may
4241 * have dropped and recreated the DB with other flags.
4243 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4244 return MDB_INCOMPATIBLE;
4245 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4247 if (flags & MDB_PS_MODIFY)
4249 *mc->mc_dbflag &= ~DB_STALE;
4250 *mc->mc_dbflag |= dbflag;
4253 root = mc->mc_db->md_root;
4255 if (root == P_INVALID) { /* Tree is empty. */
4256 DPUTS("tree is empty");
4257 return MDB_NOTFOUND;
4262 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4263 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4269 DPRINTF("db %u root page %zu has flags 0x%X",
4270 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4272 if (flags & MDB_PS_MODIFY) {
4273 if ((rc = mdb_page_touch(mc)))
4277 if (flags & MDB_PS_ROOTONLY)
4280 return mdb_page_search_root(mc, key, flags);
4284 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4286 MDB_txn *txn = mc->mc_txn;
4287 pgno_t pg = mp->mp_pgno;
4288 unsigned i, ovpages = mp->mp_pages;
4291 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4292 /* If the page is dirty we just acquired it, so we should
4293 * give it back to our current free list, if any.
4294 * Not currently supported in nested txns.
4295 * Otherwise put it onto the list of pages we freed in this txn.
4297 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && txn->mt_env->me_pghead) {
4299 pgno_t *mop = txn->mt_env->me_pghead;
4300 MDB_ID2 *dl, ix, iy;
4301 /* Prepare to insert pg */
4302 j = mop[0] + ovpages;
4304 rc = mdb_midl_grow(&mop, ovpages);
4307 txn->mt_env->me_pghead = mop;
4309 /* Remove from dirty list */
4310 dl = txn->mt_u.dirty_list;
4312 for (ix = dl[x]; ix.mid != pg; ix = iy) {
4319 return MDB_CORRUPTED;
4322 /* Insert in me_pghead */
4323 for (i = mop[0]; i && mop[i] < pg; i--)
4329 for (i=0; i<ovpages; i++) {
4330 mdb_midl_append(&txn->mt_free_pgs, pg);
4334 mc->mc_db->md_overflow_pages -= ovpages;
4338 /** Return the data associated with a given node.
4339 * @param[in] txn The transaction for this operation.
4340 * @param[in] leaf The node being read.
4341 * @param[out] data Updated to point to the node's data.
4342 * @return 0 on success, non-zero on failure.
4345 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4347 MDB_page *omp; /* overflow page */
4351 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4352 data->mv_size = NODEDSZ(leaf);
4353 data->mv_data = NODEDATA(leaf);
4357 /* Read overflow data.
4359 data->mv_size = NODEDSZ(leaf);
4360 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4361 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4362 DPRINTF("read overflow page %zu failed", pgno);
4365 data->mv_data = METADATA(omp);
4371 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4372 MDB_val *key, MDB_val *data)
4381 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4383 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4386 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4390 mdb_cursor_init(&mc, txn, dbi, &mx);
4391 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4394 /** Find a sibling for a page.
4395 * Replaces the page at the top of the cursor's stack with the
4396 * specified sibling, if one exists.
4397 * @param[in] mc The cursor for this operation.
4398 * @param[in] move_right Non-zero if the right sibling is requested,
4399 * otherwise the left sibling.
4400 * @return 0 on success, non-zero on failure.
4403 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4409 if (mc->mc_snum < 2) {
4410 return MDB_NOTFOUND; /* root has no siblings */
4414 DPRINTF("parent page is page %zu, index %u",
4415 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4417 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4418 : (mc->mc_ki[mc->mc_top] == 0)) {
4419 DPRINTF("no more keys left, moving to %s sibling",
4420 move_right ? "right" : "left");
4421 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4422 /* undo cursor_pop before returning */
4429 mc->mc_ki[mc->mc_top]++;
4431 mc->mc_ki[mc->mc_top]--;
4432 DPRINTF("just moving to %s index key %u",
4433 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4435 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4437 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4438 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4441 mdb_cursor_push(mc, mp);
4443 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4448 /** Move the cursor to the next data item. */
4450 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4456 if (mc->mc_flags & C_EOF) {
4457 return MDB_NOTFOUND;
4460 assert(mc->mc_flags & C_INITIALIZED);
4462 mp = mc->mc_pg[mc->mc_top];
4464 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4465 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4466 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4467 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4468 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4469 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4473 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4474 if (op == MDB_NEXT_DUP)
4475 return MDB_NOTFOUND;
4479 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4481 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4482 DPUTS("=====> move to next sibling page");
4483 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4484 mc->mc_flags |= C_EOF;
4485 mc->mc_flags &= ~C_INITIALIZED;
4486 return MDB_NOTFOUND;
4488 mp = mc->mc_pg[mc->mc_top];
4489 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4491 mc->mc_ki[mc->mc_top]++;
4493 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4494 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4497 key->mv_size = mc->mc_db->md_pad;
4498 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4502 assert(IS_LEAF(mp));
4503 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4505 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4506 mdb_xcursor_init1(mc, leaf);
4509 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4512 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4513 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4514 if (rc != MDB_SUCCESS)
4519 MDB_GET_KEY(leaf, key);
4523 /** Move the cursor to the previous data item. */
4525 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4531 assert(mc->mc_flags & C_INITIALIZED);
4533 mp = mc->mc_pg[mc->mc_top];
4535 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4536 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4537 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4538 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4539 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4540 if (op != MDB_PREV || rc == MDB_SUCCESS)
4543 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4544 if (op == MDB_PREV_DUP)
4545 return MDB_NOTFOUND;
4550 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4552 if (mc->mc_ki[mc->mc_top] == 0) {
4553 DPUTS("=====> move to prev sibling page");
4554 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4555 mc->mc_flags &= ~C_INITIALIZED;
4556 return MDB_NOTFOUND;
4558 mp = mc->mc_pg[mc->mc_top];
4559 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4560 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4562 mc->mc_ki[mc->mc_top]--;
4564 mc->mc_flags &= ~C_EOF;
4566 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4567 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4570 key->mv_size = mc->mc_db->md_pad;
4571 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4575 assert(IS_LEAF(mp));
4576 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4578 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4579 mdb_xcursor_init1(mc, leaf);
4582 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4585 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4586 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4587 if (rc != MDB_SUCCESS)
4592 MDB_GET_KEY(leaf, key);
4596 /** Set the cursor on a specific data item. */
4598 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4599 MDB_cursor_op op, int *exactp)
4603 MDB_node *leaf = NULL;
4608 assert(key->mv_size > 0);
4610 /* See if we're already on the right page */
4611 if (mc->mc_flags & C_INITIALIZED) {
4614 mp = mc->mc_pg[mc->mc_top];
4616 mc->mc_ki[mc->mc_top] = 0;
4617 return MDB_NOTFOUND;
4619 if (mp->mp_flags & P_LEAF2) {
4620 nodekey.mv_size = mc->mc_db->md_pad;
4621 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4623 leaf = NODEPTR(mp, 0);
4624 MDB_GET_KEY(leaf, &nodekey);
4626 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4628 /* Probably happens rarely, but first node on the page
4629 * was the one we wanted.
4631 mc->mc_ki[mc->mc_top] = 0;
4638 unsigned int nkeys = NUMKEYS(mp);
4640 if (mp->mp_flags & P_LEAF2) {
4641 nodekey.mv_data = LEAF2KEY(mp,
4642 nkeys-1, nodekey.mv_size);
4644 leaf = NODEPTR(mp, nkeys-1);
4645 MDB_GET_KEY(leaf, &nodekey);
4647 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4649 /* last node was the one we wanted */
4650 mc->mc_ki[mc->mc_top] = nkeys-1;
4656 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4657 /* This is definitely the right page, skip search_page */
4658 if (mp->mp_flags & P_LEAF2) {
4659 nodekey.mv_data = LEAF2KEY(mp,
4660 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4662 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4663 MDB_GET_KEY(leaf, &nodekey);
4665 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4667 /* current node was the one we wanted */
4677 /* If any parents have right-sibs, search.
4678 * Otherwise, there's nothing further.
4680 for (i=0; i<mc->mc_top; i++)
4682 NUMKEYS(mc->mc_pg[i])-1)
4684 if (i == mc->mc_top) {
4685 /* There are no other pages */
4686 mc->mc_ki[mc->mc_top] = nkeys;
4687 return MDB_NOTFOUND;
4691 /* There are no other pages */
4692 mc->mc_ki[mc->mc_top] = 0;
4693 return MDB_NOTFOUND;
4697 rc = mdb_page_search(mc, key, 0);
4698 if (rc != MDB_SUCCESS)
4701 mp = mc->mc_pg[mc->mc_top];
4702 assert(IS_LEAF(mp));
4705 leaf = mdb_node_search(mc, key, exactp);
4706 if (exactp != NULL && !*exactp) {
4707 /* MDB_SET specified and not an exact match. */
4708 return MDB_NOTFOUND;
4712 DPUTS("===> inexact leaf not found, goto sibling");
4713 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4714 return rc; /* no entries matched */
4715 mp = mc->mc_pg[mc->mc_top];
4716 assert(IS_LEAF(mp));
4717 leaf = NODEPTR(mp, 0);
4721 mc->mc_flags |= C_INITIALIZED;
4722 mc->mc_flags &= ~C_EOF;
4725 key->mv_size = mc->mc_db->md_pad;
4726 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4730 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4731 mdb_xcursor_init1(mc, leaf);
4734 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4735 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4736 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4739 if (op == MDB_GET_BOTH) {
4745 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4746 if (rc != MDB_SUCCESS)
4749 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4751 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4753 rc = mc->mc_dbx->md_dcmp(data, &d2);
4755 if (op == MDB_GET_BOTH || rc > 0)
4756 return MDB_NOTFOUND;
4761 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4762 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4767 /* The key already matches in all other cases */
4768 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4769 MDB_GET_KEY(leaf, key);
4770 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4775 /** Move the cursor to the first item in the database. */
4777 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4782 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4783 rc = mdb_page_search(mc, NULL, 0);
4784 if (rc != MDB_SUCCESS)
4787 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4789 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4790 mc->mc_flags |= C_INITIALIZED;
4791 mc->mc_flags &= ~C_EOF;
4793 mc->mc_ki[mc->mc_top] = 0;
4795 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4796 key->mv_size = mc->mc_db->md_pad;
4797 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4802 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4803 mdb_xcursor_init1(mc, leaf);
4804 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4809 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4810 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4814 MDB_GET_KEY(leaf, key);
4818 /** Move the cursor to the last item in the database. */
4820 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4825 if (!(mc->mc_flags & C_EOF)) {
4827 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4830 lkey.mv_size = MDB_MAXKEYSIZE+1;
4831 lkey.mv_data = NULL;
4832 rc = mdb_page_search(mc, &lkey, 0);
4833 if (rc != MDB_SUCCESS)
4836 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4839 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4840 mc->mc_flags |= C_INITIALIZED|C_EOF;
4841 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4843 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4844 key->mv_size = mc->mc_db->md_pad;
4845 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4850 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4851 mdb_xcursor_init1(mc, leaf);
4852 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4857 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4858 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4863 MDB_GET_KEY(leaf, key);
4868 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4877 case MDB_GET_CURRENT:
4878 if (!(mc->mc_flags & C_INITIALIZED)) {
4881 MDB_page *mp = mc->mc_pg[mc->mc_top];
4883 mc->mc_ki[mc->mc_top] = 0;
4889 key->mv_size = mc->mc_db->md_pad;
4890 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4892 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4893 MDB_GET_KEY(leaf, key);
4895 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4896 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4898 rc = mdb_node_read(mc->mc_txn, leaf, data);
4905 case MDB_GET_BOTH_RANGE:
4906 if (data == NULL || mc->mc_xcursor == NULL) {
4914 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4916 } else if (op == MDB_SET_RANGE)
4917 rc = mdb_cursor_set(mc, key, data, op, NULL);
4919 rc = mdb_cursor_set(mc, key, data, op, &exact);
4921 case MDB_GET_MULTIPLE:
4923 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4924 !(mc->mc_flags & C_INITIALIZED)) {
4929 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4930 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4933 case MDB_NEXT_MULTIPLE:
4935 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4939 if (!(mc->mc_flags & C_INITIALIZED))
4940 rc = mdb_cursor_first(mc, key, data);
4942 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4943 if (rc == MDB_SUCCESS) {
4944 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4947 mx = &mc->mc_xcursor->mx_cursor;
4948 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4950 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4951 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4959 case MDB_NEXT_NODUP:
4960 if (!(mc->mc_flags & C_INITIALIZED))
4961 rc = mdb_cursor_first(mc, key, data);
4963 rc = mdb_cursor_next(mc, key, data, op);
4967 case MDB_PREV_NODUP:
4968 if (!(mc->mc_flags & C_INITIALIZED)) {
4969 rc = mdb_cursor_last(mc, key, data);
4972 mc->mc_flags |= C_INITIALIZED;
4973 mc->mc_ki[mc->mc_top]++;
4975 rc = mdb_cursor_prev(mc, key, data, op);
4978 rc = mdb_cursor_first(mc, key, data);
4982 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4983 !(mc->mc_flags & C_INITIALIZED) ||
4984 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4988 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4991 rc = mdb_cursor_last(mc, key, data);
4995 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4996 !(mc->mc_flags & C_INITIALIZED) ||
4997 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5001 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5004 DPRINTF("unhandled/unimplemented cursor operation %u", op);
5012 /** Touch all the pages in the cursor stack.
5013 * Makes sure all the pages are writable, before attempting a write operation.
5014 * @param[in] mc The cursor to operate on.
5017 mdb_cursor_touch(MDB_cursor *mc)
5021 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5024 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5025 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5028 *mc->mc_dbflag |= DB_DIRTY;
5030 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5031 rc = mdb_page_touch(mc);
5035 mc->mc_top = mc->mc_snum-1;
5040 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5043 MDB_node *leaf = NULL;
5044 MDB_val xdata, *rdata, dkey;
5047 int do_sub = 0, insert = 0;
5048 unsigned int mcount = 0;
5052 char dbuf[MDB_MAXKEYSIZE+1];
5053 unsigned int nflags;
5056 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5059 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5062 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5065 #if SIZE_MAX > MAXDATASIZE
5066 if (data->mv_size > MAXDATASIZE)
5070 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5071 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5075 if (flags == MDB_CURRENT) {
5076 if (!(mc->mc_flags & C_INITIALIZED))
5079 } else if (mc->mc_db->md_root == P_INVALID) {
5081 /* new database, write a root leaf page */
5082 DPUTS("allocating new root leaf page");
5083 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5087 mdb_cursor_push(mc, np);
5088 mc->mc_db->md_root = np->mp_pgno;
5089 mc->mc_db->md_depth++;
5090 *mc->mc_dbflag |= DB_DIRTY;
5091 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5093 np->mp_flags |= P_LEAF2;
5094 mc->mc_flags |= C_INITIALIZED;
5100 if (flags & MDB_APPEND) {
5102 rc = mdb_cursor_last(mc, &k2, &d2);
5104 rc = mc->mc_dbx->md_cmp(key, &k2);
5107 mc->mc_ki[mc->mc_top]++;
5109 /* new key is <= last key */
5114 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5116 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5117 DPRINTF("duplicate key [%s]", DKEY(key));
5119 return MDB_KEYEXIST;
5121 if (rc && rc != MDB_NOTFOUND)
5125 /* Cursor is positioned, now make sure all pages are writable */
5126 rc2 = mdb_cursor_touch(mc);
5131 /* The key already exists */
5132 if (rc == MDB_SUCCESS) {
5133 /* there's only a key anyway, so this is a no-op */
5134 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5135 unsigned int ksize = mc->mc_db->md_pad;
5136 if (key->mv_size != ksize)
5138 if (flags == MDB_CURRENT) {
5139 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5140 memcpy(ptr, key->mv_data, ksize);
5145 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5148 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5149 /* Was a single item before, must convert now */
5151 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5152 /* Just overwrite the current item */
5153 if (flags == MDB_CURRENT)
5156 dkey.mv_size = NODEDSZ(leaf);
5157 dkey.mv_data = NODEDATA(leaf);
5158 #if UINT_MAX < SIZE_MAX
5159 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5160 #ifdef MISALIGNED_OK
5161 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5163 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5166 /* if data matches, ignore it */
5167 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5168 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5170 /* create a fake page for the dup items */
5171 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5172 dkey.mv_data = dbuf;
5173 fp = (MDB_page *)&pbuf;
5174 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5175 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5176 fp->mp_lower = PAGEHDRSZ;
5177 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5178 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5179 fp->mp_flags |= P_LEAF2;
5180 fp->mp_pad = data->mv_size;
5181 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5183 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5184 (dkey.mv_size & 1) + (data->mv_size & 1);
5186 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5189 xdata.mv_size = fp->mp_upper;
5194 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5195 /* See if we need to convert from fake page to subDB */
5197 unsigned int offset;
5200 fp = NODEDATA(leaf);
5201 if (flags == MDB_CURRENT) {
5203 fp->mp_flags |= P_DIRTY;
5204 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5205 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5209 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5210 offset = fp->mp_pad;
5211 if (SIZELEFT(fp) >= offset)
5213 offset *= 4; /* space for 4 more */
5215 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5217 offset += offset & 1;
5218 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5219 offset >= mc->mc_txn->mt_env->me_nodemax) {
5220 /* yes, convert it */
5222 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5223 dummy.md_pad = fp->mp_pad;
5224 dummy.md_flags = MDB_DUPFIXED;
5225 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5226 dummy.md_flags |= MDB_INTEGERKEY;
5229 dummy.md_branch_pages = 0;
5230 dummy.md_leaf_pages = 1;
5231 dummy.md_overflow_pages = 0;
5232 dummy.md_entries = NUMKEYS(fp);
5234 xdata.mv_size = sizeof(MDB_db);
5235 xdata.mv_data = &dummy;
5236 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5238 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5239 flags |= F_DUPDATA|F_SUBDATA;
5240 dummy.md_root = mp->mp_pgno;
5242 /* no, just grow it */
5244 xdata.mv_size = NODEDSZ(leaf) + offset;
5245 xdata.mv_data = &pbuf;
5246 mp = (MDB_page *)&pbuf;
5247 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5250 mp->mp_flags = fp->mp_flags | P_DIRTY;
5251 mp->mp_pad = fp->mp_pad;
5252 mp->mp_lower = fp->mp_lower;
5253 mp->mp_upper = fp->mp_upper + offset;
5255 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5257 nsize = NODEDSZ(leaf) - fp->mp_upper;
5258 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5259 for (i=0; i<NUMKEYS(fp); i++)
5260 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5262 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5266 /* data is on sub-DB, just store it */
5267 flags |= F_DUPDATA|F_SUBDATA;
5271 /* overflow page overwrites need special handling */
5272 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5275 unsigned psize = mc->mc_txn->mt_env->me_psize;
5276 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5278 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5279 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5281 ovpages = omp->mp_pages;
5283 /* Is the ov page writable and large enough? */
5284 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5285 /* yes, overwrite it. Note in this case we don't
5286 * bother to try shrinking the page if the new data
5287 * is smaller than the overflow threshold.
5290 /* It is writable only in a parent txn */
5291 size_t sz = (size_t) psize * ovpages, off;
5292 MDB_page *np = mdb_page_malloc(mc, ovpages);
5298 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5299 if (!(flags & MDB_RESERVE)) {
5300 /* Copy end of page, adjusting alignment so
5301 * compiler may copy words instead of bytes.
5303 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5304 memcpy((size_t *)((char *)np + off),
5305 (size_t *)((char *)omp + off), sz - off);
5308 memcpy(np, omp, sz); /* Copy beginning of page */
5311 SETDSZ(leaf, data->mv_size);
5312 if (F_ISSET(flags, MDB_RESERVE))
5313 data->mv_data = METADATA(omp);
5315 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5318 mdb_ovpage_free(mc, omp);
5320 } else if (NODEDSZ(leaf) == data->mv_size) {
5321 /* same size, just replace it. Note that we could
5322 * also reuse this node if the new data is smaller,
5323 * but instead we opt to shrink the node in that case.
5325 if (F_ISSET(flags, MDB_RESERVE))
5326 data->mv_data = NODEDATA(leaf);
5327 else if (data->mv_size)
5328 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5330 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5333 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5334 mc->mc_db->md_entries--;
5336 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5343 nflags = flags & NODE_ADD_FLAGS;
5344 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5345 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5346 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5347 nflags &= ~MDB_APPEND;
5349 nflags |= MDB_SPLIT_REPLACE;
5350 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5352 /* There is room already in this leaf page. */
5353 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5354 if (rc == 0 && !do_sub && insert) {
5355 /* Adjust other cursors pointing to mp */
5356 MDB_cursor *m2, *m3;
5357 MDB_dbi dbi = mc->mc_dbi;
5358 unsigned i = mc->mc_top;
5359 MDB_page *mp = mc->mc_pg[i];
5361 if (mc->mc_flags & C_SUB)
5364 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5365 if (mc->mc_flags & C_SUB)
5366 m3 = &m2->mc_xcursor->mx_cursor;
5369 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5370 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5377 if (rc != MDB_SUCCESS)
5378 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5380 /* Now store the actual data in the child DB. Note that we're
5381 * storing the user data in the keys field, so there are strict
5382 * size limits on dupdata. The actual data fields of the child
5383 * DB are all zero size.
5390 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5391 if (flags & MDB_CURRENT) {
5392 xflags = MDB_CURRENT;
5394 mdb_xcursor_init1(mc, leaf);
5395 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5397 /* converted, write the original data first */
5399 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5403 /* Adjust other cursors pointing to mp */
5405 unsigned i = mc->mc_top;
5406 MDB_page *mp = mc->mc_pg[i];
5408 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5409 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5410 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5411 mdb_xcursor_init1(m2, leaf);
5415 /* we've done our job */
5418 if (flags & MDB_APPENDDUP)
5419 xflags |= MDB_APPEND;
5420 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5421 if (flags & F_SUBDATA) {
5422 void *db = NODEDATA(leaf);
5423 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5426 /* sub-writes might have failed so check rc again.
5427 * Don't increment count if we just replaced an existing item.
5429 if (!rc && !(flags & MDB_CURRENT))
5430 mc->mc_db->md_entries++;
5431 if (flags & MDB_MULTIPLE) {
5433 if (mcount < data[1].mv_size) {
5434 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5435 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5441 /* If we succeeded and the key didn't exist before, make sure
5442 * the cursor is marked valid.
5445 mc->mc_flags |= C_INITIALIZED;
5450 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5455 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5458 if (!(mc->mc_flags & C_INITIALIZED))
5461 rc = mdb_cursor_touch(mc);
5465 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5467 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5468 if (flags != MDB_NODUPDATA) {
5469 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5470 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5472 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5473 /* If sub-DB still has entries, we're done */
5474 if (mc->mc_xcursor->mx_db.md_entries) {
5475 if (leaf->mn_flags & F_SUBDATA) {
5476 /* update subDB info */
5477 void *db = NODEDATA(leaf);
5478 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5481 /* shrink fake page */
5482 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5483 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5484 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5485 /* fix other sub-DB cursors pointed at this fake page */
5486 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5487 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5488 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5489 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5490 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5493 mc->mc_db->md_entries--;
5496 /* otherwise fall thru and delete the sub-DB */
5499 if (leaf->mn_flags & F_SUBDATA) {
5500 /* add all the child DB's pages to the free list */
5501 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5502 if (rc == MDB_SUCCESS) {
5503 mc->mc_db->md_entries -=
5504 mc->mc_xcursor->mx_db.md_entries;
5509 return mdb_cursor_del0(mc, leaf);
5512 /** Allocate and initialize new pages for a database.
5513 * @param[in] mc a cursor on the database being added to.
5514 * @param[in] flags flags defining what type of page is being allocated.
5515 * @param[in] num the number of pages to allocate. This is usually 1,
5516 * unless allocating overflow pages for a large record.
5517 * @param[out] mp Address of a page, or NULL on failure.
5518 * @return 0 on success, non-zero on failure.
5521 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5526 if ((rc = mdb_page_alloc(mc, num, &np)))
5528 DPRINTF("allocated new mpage %zu, page size %u",
5529 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5530 np->mp_flags = flags | P_DIRTY;
5531 np->mp_lower = PAGEHDRSZ;
5532 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5535 mc->mc_db->md_branch_pages++;
5536 else if (IS_LEAF(np))
5537 mc->mc_db->md_leaf_pages++;
5538 else if (IS_OVERFLOW(np)) {
5539 mc->mc_db->md_overflow_pages += num;
5547 /** Calculate the size of a leaf node.
5548 * The size depends on the environment's page size; if a data item
5549 * is too large it will be put onto an overflow page and the node
5550 * size will only include the key and not the data. Sizes are always
5551 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5552 * of the #MDB_node headers.
5553 * @param[in] env The environment handle.
5554 * @param[in] key The key for the node.
5555 * @param[in] data The data for the node.
5556 * @return The number of bytes needed to store the node.
5559 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5563 sz = LEAFSIZE(key, data);
5564 if (sz >= env->me_nodemax) {
5565 /* put on overflow page */
5566 sz -= data->mv_size - sizeof(pgno_t);
5570 return sz + sizeof(indx_t);
5573 /** Calculate the size of a branch node.
5574 * The size should depend on the environment's page size but since
5575 * we currently don't support spilling large keys onto overflow
5576 * pages, it's simply the size of the #MDB_node header plus the
5577 * size of the key. Sizes are always rounded up to an even number
5578 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5579 * @param[in] env The environment handle.
5580 * @param[in] key The key for the node.
5581 * @return The number of bytes needed to store the node.
5584 mdb_branch_size(MDB_env *env, MDB_val *key)
5589 if (sz >= env->me_nodemax) {
5590 /* put on overflow page */
5591 /* not implemented */
5592 /* sz -= key->size - sizeof(pgno_t); */
5595 return sz + sizeof(indx_t);
5598 /** Add a node to the page pointed to by the cursor.
5599 * @param[in] mc The cursor for this operation.
5600 * @param[in] indx The index on the page where the new node should be added.
5601 * @param[in] key The key for the new node.
5602 * @param[in] data The data for the new node, if any.
5603 * @param[in] pgno The page number, if adding a branch node.
5604 * @param[in] flags Flags for the node.
5605 * @return 0 on success, non-zero on failure. Possible errors are:
5607 * <li>ENOMEM - failed to allocate overflow pages for the node.
5608 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5609 * should never happen since all callers already calculate the
5610 * page's free space before calling this function.
5614 mdb_node_add(MDB_cursor *mc, indx_t indx,
5615 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5618 size_t node_size = NODESIZE;
5621 MDB_page *mp = mc->mc_pg[mc->mc_top];
5622 MDB_page *ofp = NULL; /* overflow page */
5625 assert(mp->mp_upper >= mp->mp_lower);
5627 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5628 IS_LEAF(mp) ? "leaf" : "branch",
5629 IS_SUBP(mp) ? "sub-" : "",
5630 mp->mp_pgno, indx, data ? data->mv_size : 0,
5631 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5634 /* Move higher keys up one slot. */
5635 int ksize = mc->mc_db->md_pad, dif;
5636 char *ptr = LEAF2KEY(mp, indx, ksize);
5637 dif = NUMKEYS(mp) - indx;
5639 memmove(ptr+ksize, ptr, dif*ksize);
5640 /* insert new key */
5641 memcpy(ptr, key->mv_data, ksize);
5643 /* Just using these for counting */
5644 mp->mp_lower += sizeof(indx_t);
5645 mp->mp_upper -= ksize - sizeof(indx_t);
5650 node_size += key->mv_size;
5654 if (F_ISSET(flags, F_BIGDATA)) {
5655 /* Data already on overflow page. */
5656 node_size += sizeof(pgno_t);
5657 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5658 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5660 /* Put data on overflow page. */
5661 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5662 data->mv_size, node_size+data->mv_size);
5663 node_size += sizeof(pgno_t);
5664 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5666 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5669 node_size += data->mv_size;
5672 node_size += node_size & 1;
5674 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5675 DPRINTF("not enough room in page %zu, got %u ptrs",
5676 mp->mp_pgno, NUMKEYS(mp));
5677 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5678 mp->mp_upper - mp->mp_lower);
5679 DPRINTF("node size = %zu", node_size);
5680 return MDB_PAGE_FULL;
5683 /* Move higher pointers up one slot. */
5684 for (i = NUMKEYS(mp); i > indx; i--)
5685 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5687 /* Adjust free space offsets. */
5688 ofs = mp->mp_upper - node_size;
5689 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5690 mp->mp_ptrs[indx] = ofs;
5692 mp->mp_lower += sizeof(indx_t);
5694 /* Write the node data. */
5695 node = NODEPTR(mp, indx);
5696 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5697 node->mn_flags = flags;
5699 SETDSZ(node,data->mv_size);
5704 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5709 if (F_ISSET(flags, F_BIGDATA))
5710 memcpy(node->mn_data + key->mv_size, data->mv_data,
5712 else if (F_ISSET(flags, MDB_RESERVE))
5713 data->mv_data = node->mn_data + key->mv_size;
5715 memcpy(node->mn_data + key->mv_size, data->mv_data,
5718 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5720 if (F_ISSET(flags, MDB_RESERVE))
5721 data->mv_data = METADATA(ofp);
5723 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5730 /** Delete the specified node from a page.
5731 * @param[in] mp The page to operate on.
5732 * @param[in] indx The index of the node to delete.
5733 * @param[in] ksize The size of a node. Only used if the page is
5734 * part of a #MDB_DUPFIXED database.
5737 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5740 indx_t i, j, numkeys, ptr;
5747 COPY_PGNO(pgno, mp->mp_pgno);
5748 DPRINTF("delete node %u on %s page %zu", indx,
5749 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5752 assert(indx < NUMKEYS(mp));
5755 int x = NUMKEYS(mp) - 1 - indx;
5756 base = LEAF2KEY(mp, indx, ksize);
5758 memmove(base, base + ksize, x * ksize);
5759 mp->mp_lower -= sizeof(indx_t);
5760 mp->mp_upper += ksize - sizeof(indx_t);
5764 node = NODEPTR(mp, indx);
5765 sz = NODESIZE + node->mn_ksize;
5767 if (F_ISSET(node->mn_flags, F_BIGDATA))
5768 sz += sizeof(pgno_t);
5770 sz += NODEDSZ(node);
5774 ptr = mp->mp_ptrs[indx];
5775 numkeys = NUMKEYS(mp);
5776 for (i = j = 0; i < numkeys; i++) {
5778 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5779 if (mp->mp_ptrs[i] < ptr)
5780 mp->mp_ptrs[j] += sz;
5785 base = (char *)mp + mp->mp_upper;
5786 memmove(base + sz, base, ptr - mp->mp_upper);
5788 mp->mp_lower -= sizeof(indx_t);
5792 /** Compact the main page after deleting a node on a subpage.
5793 * @param[in] mp The main page to operate on.
5794 * @param[in] indx The index of the subpage on the main page.
5797 mdb_node_shrink(MDB_page *mp, indx_t indx)
5804 indx_t i, numkeys, ptr;
5806 node = NODEPTR(mp, indx);
5807 sp = (MDB_page *)NODEDATA(node);
5808 osize = NODEDSZ(node);
5810 delta = sp->mp_upper - sp->mp_lower;
5811 SETDSZ(node, osize - delta);
5812 xp = (MDB_page *)((char *)sp + delta);
5814 /* shift subpage upward */
5816 nsize = NUMKEYS(sp) * sp->mp_pad;
5817 memmove(METADATA(xp), METADATA(sp), nsize);
5820 nsize = osize - sp->mp_upper;
5821 numkeys = NUMKEYS(sp);
5822 for (i=numkeys-1; i>=0; i--)
5823 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5825 xp->mp_upper = sp->mp_lower;
5826 xp->mp_lower = sp->mp_lower;
5827 xp->mp_flags = sp->mp_flags;
5828 xp->mp_pad = sp->mp_pad;
5829 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5831 /* shift lower nodes upward */
5832 ptr = mp->mp_ptrs[indx];
5833 numkeys = NUMKEYS(mp);
5834 for (i = 0; i < numkeys; i++) {
5835 if (mp->mp_ptrs[i] <= ptr)
5836 mp->mp_ptrs[i] += delta;
5839 base = (char *)mp + mp->mp_upper;
5840 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5841 mp->mp_upper += delta;
5844 /** Initial setup of a sorted-dups cursor.
5845 * Sorted duplicates are implemented as a sub-database for the given key.
5846 * The duplicate data items are actually keys of the sub-database.
5847 * Operations on the duplicate data items are performed using a sub-cursor
5848 * initialized when the sub-database is first accessed. This function does
5849 * the preliminary setup of the sub-cursor, filling in the fields that
5850 * depend only on the parent DB.
5851 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5854 mdb_xcursor_init0(MDB_cursor *mc)
5856 MDB_xcursor *mx = mc->mc_xcursor;
5858 mx->mx_cursor.mc_xcursor = NULL;
5859 mx->mx_cursor.mc_txn = mc->mc_txn;
5860 mx->mx_cursor.mc_db = &mx->mx_db;
5861 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5862 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5863 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5864 mx->mx_cursor.mc_snum = 0;
5865 mx->mx_cursor.mc_top = 0;
5866 mx->mx_cursor.mc_flags = C_SUB;
5867 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5868 mx->mx_dbx.md_dcmp = NULL;
5869 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5872 /** Final setup of a sorted-dups cursor.
5873 * Sets up the fields that depend on the data from the main cursor.
5874 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5875 * @param[in] node The data containing the #MDB_db record for the
5876 * sorted-dup database.
5879 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5881 MDB_xcursor *mx = mc->mc_xcursor;
5883 if (node->mn_flags & F_SUBDATA) {
5884 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5885 mx->mx_cursor.mc_pg[0] = 0;
5886 mx->mx_cursor.mc_snum = 0;
5887 mx->mx_cursor.mc_flags = C_SUB;
5889 MDB_page *fp = NODEDATA(node);
5890 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5891 mx->mx_db.md_flags = 0;
5892 mx->mx_db.md_depth = 1;
5893 mx->mx_db.md_branch_pages = 0;
5894 mx->mx_db.md_leaf_pages = 1;
5895 mx->mx_db.md_overflow_pages = 0;
5896 mx->mx_db.md_entries = NUMKEYS(fp);
5897 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5898 mx->mx_cursor.mc_snum = 1;
5899 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5900 mx->mx_cursor.mc_top = 0;
5901 mx->mx_cursor.mc_pg[0] = fp;
5902 mx->mx_cursor.mc_ki[0] = 0;
5903 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5904 mx->mx_db.md_flags = MDB_DUPFIXED;
5905 mx->mx_db.md_pad = fp->mp_pad;
5906 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5907 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5910 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5912 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5914 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5915 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5916 #if UINT_MAX < SIZE_MAX
5917 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5918 #ifdef MISALIGNED_OK
5919 mx->mx_dbx.md_cmp = mdb_cmp_long;
5921 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5926 /** Initialize a cursor for a given transaction and database. */
5928 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5933 mc->mc_db = &txn->mt_dbs[dbi];
5934 mc->mc_dbx = &txn->mt_dbxs[dbi];
5935 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5940 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5942 mc->mc_xcursor = mx;
5943 mdb_xcursor_init0(mc);
5945 mc->mc_xcursor = NULL;
5947 if (*mc->mc_dbflag & DB_STALE) {
5948 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5953 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5956 size_t size = sizeof(MDB_cursor);
5958 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5961 /* Allow read access to the freelist */
5962 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5965 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5966 size += sizeof(MDB_xcursor);
5968 if ((mc = malloc(size)) != NULL) {
5969 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5970 if (txn->mt_cursors) {
5971 mc->mc_next = txn->mt_cursors[dbi];
5972 txn->mt_cursors[dbi] = mc;
5973 mc->mc_flags |= C_UNTRACK;
5975 mc->mc_flags |= C_ALLOCD;
5986 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5990 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5993 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5996 flags = mc->mc_flags;
5998 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6000 mc->mc_flags |= (flags & C_ALLOCD);
6004 /* Return the count of duplicate data items for the current key */
6006 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6010 if (mc == NULL || countp == NULL)
6013 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
6016 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6017 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6020 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6023 *countp = mc->mc_xcursor->mx_db.md_entries;
6029 mdb_cursor_close(MDB_cursor *mc)
6032 /* remove from txn, if tracked */
6033 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6034 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6035 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6037 *prev = mc->mc_next;
6039 if (mc->mc_flags & C_ALLOCD)
6045 mdb_cursor_txn(MDB_cursor *mc)
6047 if (!mc) return NULL;
6052 mdb_cursor_dbi(MDB_cursor *mc)
6058 /** Replace the key for a node with a new key.
6059 * @param[in] mc Cursor pointing to the node to operate on.
6060 * @param[in] key The new key to use.
6061 * @return 0 on success, non-zero on failure.
6064 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6071 indx_t ptr, i, numkeys, indx;
6074 indx = mc->mc_ki[mc->mc_top];
6075 mp = mc->mc_pg[mc->mc_top];
6076 node = NODEPTR(mp, indx);
6077 ptr = mp->mp_ptrs[indx];
6081 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6082 k2.mv_data = NODEKEY(node);
6083 k2.mv_size = node->mn_ksize;
6084 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6086 mdb_dkey(&k2, kbuf2),
6092 delta0 = delta = key->mv_size - node->mn_ksize;
6094 /* Must be 2-byte aligned. If new key is
6095 * shorter by 1, the shift will be skipped.
6097 delta += (delta & 1);
6099 if (delta > 0 && SIZELEFT(mp) < delta) {
6101 /* not enough space left, do a delete and split */
6102 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6103 pgno = NODEPGNO(node);
6104 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6105 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6108 numkeys = NUMKEYS(mp);
6109 for (i = 0; i < numkeys; i++) {
6110 if (mp->mp_ptrs[i] <= ptr)
6111 mp->mp_ptrs[i] -= delta;
6114 base = (char *)mp + mp->mp_upper;
6115 len = ptr - mp->mp_upper + NODESIZE;
6116 memmove(base - delta, base, len);
6117 mp->mp_upper -= delta;
6119 node = NODEPTR(mp, indx);
6122 /* But even if no shift was needed, update ksize */
6124 node->mn_ksize = key->mv_size;
6127 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6133 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6135 /** Move a node from csrc to cdst.
6138 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6145 unsigned short flags;
6149 /* Mark src and dst as dirty. */
6150 if ((rc = mdb_page_touch(csrc)) ||
6151 (rc = mdb_page_touch(cdst)))
6154 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6155 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6156 key.mv_size = csrc->mc_db->md_pad;
6157 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6159 data.mv_data = NULL;
6163 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6164 assert(!((long)srcnode&1));
6165 srcpg = NODEPGNO(srcnode);
6166 flags = srcnode->mn_flags;
6167 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6168 unsigned int snum = csrc->mc_snum;
6170 /* must find the lowest key below src */
6171 mdb_page_search_lowest(csrc);
6172 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6173 key.mv_size = csrc->mc_db->md_pad;
6174 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6176 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6177 key.mv_size = NODEKSZ(s2);
6178 key.mv_data = NODEKEY(s2);
6180 csrc->mc_snum = snum--;
6181 csrc->mc_top = snum;
6183 key.mv_size = NODEKSZ(srcnode);
6184 key.mv_data = NODEKEY(srcnode);
6186 data.mv_size = NODEDSZ(srcnode);
6187 data.mv_data = NODEDATA(srcnode);
6189 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6190 unsigned int snum = cdst->mc_snum;
6193 /* must find the lowest key below dst */
6194 mdb_page_search_lowest(cdst);
6195 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6196 bkey.mv_size = cdst->mc_db->md_pad;
6197 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6199 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6200 bkey.mv_size = NODEKSZ(s2);
6201 bkey.mv_data = NODEKEY(s2);
6203 cdst->mc_snum = snum--;
6204 cdst->mc_top = snum;
6205 mdb_cursor_copy(cdst, &mn);
6207 rc = mdb_update_key(&mn, &bkey);
6212 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6213 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6214 csrc->mc_ki[csrc->mc_top],
6216 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6217 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6219 /* Add the node to the destination page.
6221 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6222 if (rc != MDB_SUCCESS)
6225 /* Delete the node from the source page.
6227 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6230 /* Adjust other cursors pointing to mp */
6231 MDB_cursor *m2, *m3;
6232 MDB_dbi dbi = csrc->mc_dbi;
6233 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6235 if (csrc->mc_flags & C_SUB)
6238 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6239 if (m2 == csrc) continue;
6240 if (csrc->mc_flags & C_SUB)
6241 m3 = &m2->mc_xcursor->mx_cursor;
6244 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6245 csrc->mc_ki[csrc->mc_top]) {
6246 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6247 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6252 /* Update the parent separators.
6254 if (csrc->mc_ki[csrc->mc_top] == 0) {
6255 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6256 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6257 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6259 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6260 key.mv_size = NODEKSZ(srcnode);
6261 key.mv_data = NODEKEY(srcnode);
6263 DPRINTF("update separator for source page %zu to [%s]",
6264 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6265 mdb_cursor_copy(csrc, &mn);
6268 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6271 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6273 indx_t ix = csrc->mc_ki[csrc->mc_top];
6274 nullkey.mv_size = 0;
6275 csrc->mc_ki[csrc->mc_top] = 0;
6276 rc = mdb_update_key(csrc, &nullkey);
6277 csrc->mc_ki[csrc->mc_top] = ix;
6278 assert(rc == MDB_SUCCESS);
6282 if (cdst->mc_ki[cdst->mc_top] == 0) {
6283 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6284 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6285 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6287 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6288 key.mv_size = NODEKSZ(srcnode);
6289 key.mv_data = NODEKEY(srcnode);
6291 DPRINTF("update separator for destination page %zu to [%s]",
6292 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6293 mdb_cursor_copy(cdst, &mn);
6296 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6299 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6301 indx_t ix = cdst->mc_ki[cdst->mc_top];
6302 nullkey.mv_size = 0;
6303 cdst->mc_ki[cdst->mc_top] = 0;
6304 rc = mdb_update_key(cdst, &nullkey);
6305 cdst->mc_ki[cdst->mc_top] = ix;
6306 assert(rc == MDB_SUCCESS);
6313 /** Merge one page into another.
6314 * The nodes from the page pointed to by \b csrc will
6315 * be copied to the page pointed to by \b cdst and then
6316 * the \b csrc page will be freed.
6317 * @param[in] csrc Cursor pointing to the source page.
6318 * @param[in] cdst Cursor pointing to the destination page.
6321 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6329 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6330 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6332 assert(csrc->mc_snum > 1); /* can't merge root page */
6333 assert(cdst->mc_snum > 1);
6335 /* Mark dst as dirty. */
6336 if ((rc = mdb_page_touch(cdst)))
6339 /* Move all nodes from src to dst.
6341 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6342 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6343 key.mv_size = csrc->mc_db->md_pad;
6344 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6345 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6346 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6347 if (rc != MDB_SUCCESS)
6349 key.mv_data = (char *)key.mv_data + key.mv_size;
6352 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6353 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6354 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6355 unsigned int snum = csrc->mc_snum;
6357 /* must find the lowest key below src */
6358 mdb_page_search_lowest(csrc);
6359 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6360 key.mv_size = csrc->mc_db->md_pad;
6361 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6363 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6364 key.mv_size = NODEKSZ(s2);
6365 key.mv_data = NODEKEY(s2);
6367 csrc->mc_snum = snum--;
6368 csrc->mc_top = snum;
6370 key.mv_size = srcnode->mn_ksize;
6371 key.mv_data = NODEKEY(srcnode);
6374 data.mv_size = NODEDSZ(srcnode);
6375 data.mv_data = NODEDATA(srcnode);
6376 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6377 if (rc != MDB_SUCCESS)
6382 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6383 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);
6385 /* Unlink the src page from parent and add to free list.
6387 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6388 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6391 rc = mdb_update_key(csrc, &key);
6397 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6398 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6399 csrc->mc_db->md_leaf_pages--;
6401 csrc->mc_db->md_branch_pages--;
6403 /* Adjust other cursors pointing to mp */
6404 MDB_cursor *m2, *m3;
6405 MDB_dbi dbi = csrc->mc_dbi;
6406 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6408 if (csrc->mc_flags & C_SUB)
6411 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6412 if (csrc->mc_flags & C_SUB)
6413 m3 = &m2->mc_xcursor->mx_cursor;
6416 if (m3 == csrc) continue;
6417 if (m3->mc_snum < csrc->mc_snum) continue;
6418 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6419 m3->mc_pg[csrc->mc_top] = mp;
6420 m3->mc_ki[csrc->mc_top] += nkeys;
6424 mdb_cursor_pop(csrc);
6426 return mdb_rebalance(csrc);
6429 /** Copy the contents of a cursor.
6430 * @param[in] csrc The cursor to copy from.
6431 * @param[out] cdst The cursor to copy to.
6434 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6438 cdst->mc_txn = csrc->mc_txn;
6439 cdst->mc_dbi = csrc->mc_dbi;
6440 cdst->mc_db = csrc->mc_db;
6441 cdst->mc_dbx = csrc->mc_dbx;
6442 cdst->mc_snum = csrc->mc_snum;
6443 cdst->mc_top = csrc->mc_top;
6444 cdst->mc_flags = csrc->mc_flags;
6446 for (i=0; i<csrc->mc_snum; i++) {
6447 cdst->mc_pg[i] = csrc->mc_pg[i];
6448 cdst->mc_ki[i] = csrc->mc_ki[i];
6452 /** Rebalance the tree after a delete operation.
6453 * @param[in] mc Cursor pointing to the page where rebalancing
6455 * @return 0 on success, non-zero on failure.
6458 mdb_rebalance(MDB_cursor *mc)
6462 unsigned int ptop, minkeys;
6465 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6469 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6470 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6471 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6472 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6476 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6477 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6480 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6481 DPRINTF("no need to rebalance page %zu, above fill threshold",
6487 if (mc->mc_snum < 2) {
6488 MDB_page *mp = mc->mc_pg[0];
6490 DPUTS("Can't rebalance a subpage, ignoring");
6493 if (NUMKEYS(mp) == 0) {
6494 DPUTS("tree is completely empty");
6495 mc->mc_db->md_root = P_INVALID;
6496 mc->mc_db->md_depth = 0;
6497 mc->mc_db->md_leaf_pages = 0;
6498 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6502 /* Adjust other cursors pointing to mp */
6503 MDB_cursor *m2, *m3;
6504 MDB_dbi dbi = mc->mc_dbi;
6506 if (mc->mc_flags & C_SUB)
6509 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6510 if (m2 == mc) continue;
6511 if (mc->mc_flags & C_SUB)
6512 m3 = &m2->mc_xcursor->mx_cursor;
6515 if (m3->mc_snum < mc->mc_snum) continue;
6516 if (m3->mc_pg[0] == mp) {
6522 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6523 DPUTS("collapsing root page!");
6524 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6525 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6526 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6529 mc->mc_db->md_depth--;
6530 mc->mc_db->md_branch_pages--;
6532 /* Adjust other cursors pointing to mp */
6533 MDB_cursor *m2, *m3;
6534 MDB_dbi dbi = mc->mc_dbi;
6536 if (mc->mc_flags & C_SUB)
6539 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6540 if (m2 == mc) continue;
6541 if (mc->mc_flags & C_SUB)
6542 m3 = &m2->mc_xcursor->mx_cursor;
6545 if (m3->mc_snum < mc->mc_snum) continue;
6546 if (m3->mc_pg[0] == mp) {
6547 m3->mc_pg[0] = mc->mc_pg[0];
6554 DPUTS("root page doesn't need rebalancing");
6558 /* The parent (branch page) must have at least 2 pointers,
6559 * otherwise the tree is invalid.
6561 ptop = mc->mc_top-1;
6562 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6564 /* Leaf page fill factor is below the threshold.
6565 * Try to move keys from left or right neighbor, or
6566 * merge with a neighbor page.
6571 mdb_cursor_copy(mc, &mn);
6572 mn.mc_xcursor = NULL;
6574 if (mc->mc_ki[ptop] == 0) {
6575 /* We're the leftmost leaf in our parent.
6577 DPUTS("reading right neighbor");
6579 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6580 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6583 mn.mc_ki[mn.mc_top] = 0;
6584 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6586 /* There is at least one neighbor to the left.
6588 DPUTS("reading left neighbor");
6590 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6591 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6594 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6595 mc->mc_ki[mc->mc_top] = 0;
6598 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6599 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);
6601 /* If the neighbor page is above threshold and has enough keys,
6602 * move one key from it. Otherwise we should try to merge them.
6603 * (A branch page must never have less than 2 keys.)
6605 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6606 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6607 return mdb_node_move(&mn, mc);
6609 if (mc->mc_ki[ptop] == 0)
6610 rc = mdb_page_merge(&mn, mc);
6612 rc = mdb_page_merge(mc, &mn);
6613 mc->mc_flags &= ~C_INITIALIZED;
6618 /** Complete a delete operation started by #mdb_cursor_del(). */
6620 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6624 /* add overflow pages to free list */
6625 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6629 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6630 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6632 assert(IS_OVERFLOW(omp));
6633 mdb_ovpage_free(mc, omp);
6635 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6636 mc->mc_db->md_entries--;
6637 rc = mdb_rebalance(mc);
6638 if (rc != MDB_SUCCESS)
6639 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6640 /* if mc points past last node in page, invalidate */
6641 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6642 mc->mc_flags &= ~C_INITIALIZED;
6648 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6649 MDB_val *key, MDB_val *data)
6654 MDB_val rdata, *xdata;
6658 assert(key != NULL);
6660 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6662 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6665 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6669 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6673 mdb_cursor_init(&mc, txn, dbi, &mx);
6684 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6686 /* let mdb_page_split know about this cursor if needed:
6687 * delete will trigger a rebalance; if it needs to move
6688 * a node from one page to another, it will have to
6689 * update the parent's separator key(s). If the new sepkey
6690 * is larger than the current one, the parent page may
6691 * run out of space, triggering a split. We need this
6692 * cursor to be consistent until the end of the rebalance.
6694 mc.mc_next = txn->mt_cursors[dbi];
6695 txn->mt_cursors[dbi] = &mc;
6696 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6697 txn->mt_cursors[dbi] = mc.mc_next;
6702 /** Split a page and insert a new node.
6703 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6704 * The cursor will be updated to point to the actual page and index where
6705 * the node got inserted after the split.
6706 * @param[in] newkey The key for the newly inserted node.
6707 * @param[in] newdata The data for the newly inserted node.
6708 * @param[in] newpgno The page number, if the new node is a branch node.
6709 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6710 * @return 0 on success, non-zero on failure.
6713 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6714 unsigned int nflags)
6717 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6720 unsigned int i, j, split_indx, nkeys, pmax;
6722 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6724 MDB_page *mp, *rp, *pp;
6729 mp = mc->mc_pg[mc->mc_top];
6730 newindx = mc->mc_ki[mc->mc_top];
6732 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6733 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6734 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6736 /* Create a right sibling. */
6737 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6739 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6741 if (mc->mc_snum < 2) {
6742 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6744 /* shift current top to make room for new parent */
6745 mc->mc_pg[1] = mc->mc_pg[0];
6746 mc->mc_ki[1] = mc->mc_ki[0];
6749 mc->mc_db->md_root = pp->mp_pgno;
6750 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6751 mc->mc_db->md_depth++;
6754 /* Add left (implicit) pointer. */
6755 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6756 /* undo the pre-push */
6757 mc->mc_pg[0] = mc->mc_pg[1];
6758 mc->mc_ki[0] = mc->mc_ki[1];
6759 mc->mc_db->md_root = mp->mp_pgno;
6760 mc->mc_db->md_depth--;
6767 ptop = mc->mc_top-1;
6768 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6771 mc->mc_flags |= C_SPLITTING;
6772 mdb_cursor_copy(mc, &mn);
6773 mn.mc_pg[mn.mc_top] = rp;
6774 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6776 if (nflags & MDB_APPEND) {
6777 mn.mc_ki[mn.mc_top] = 0;
6779 split_indx = newindx;
6784 nkeys = NUMKEYS(mp);
6785 split_indx = nkeys / 2;
6786 if (newindx < split_indx)
6792 unsigned int lsize, rsize, ksize;
6793 /* Move half of the keys to the right sibling */
6795 x = mc->mc_ki[mc->mc_top] - split_indx;
6796 ksize = mc->mc_db->md_pad;
6797 split = LEAF2KEY(mp, split_indx, ksize);
6798 rsize = (nkeys - split_indx) * ksize;
6799 lsize = (nkeys - split_indx) * sizeof(indx_t);
6800 mp->mp_lower -= lsize;
6801 rp->mp_lower += lsize;
6802 mp->mp_upper += rsize - lsize;
6803 rp->mp_upper -= rsize - lsize;
6804 sepkey.mv_size = ksize;
6805 if (newindx == split_indx) {
6806 sepkey.mv_data = newkey->mv_data;
6808 sepkey.mv_data = split;
6811 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6812 memcpy(rp->mp_ptrs, split, rsize);
6813 sepkey.mv_data = rp->mp_ptrs;
6814 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6815 memcpy(ins, newkey->mv_data, ksize);
6816 mp->mp_lower += sizeof(indx_t);
6817 mp->mp_upper -= ksize - sizeof(indx_t);
6820 memcpy(rp->mp_ptrs, split, x * ksize);
6821 ins = LEAF2KEY(rp, x, ksize);
6822 memcpy(ins, newkey->mv_data, ksize);
6823 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6824 rp->mp_lower += sizeof(indx_t);
6825 rp->mp_upper -= ksize - sizeof(indx_t);
6826 mc->mc_ki[mc->mc_top] = x;
6827 mc->mc_pg[mc->mc_top] = rp;
6832 /* For leaf pages, check the split point based on what
6833 * fits where, since otherwise mdb_node_add can fail.
6835 * This check is only needed when the data items are
6836 * relatively large, such that being off by one will
6837 * make the difference between success or failure.
6839 * It's also relevant if a page happens to be laid out
6840 * such that one half of its nodes are all "small" and
6841 * the other half of its nodes are "large." If the new
6842 * item is also "large" and falls on the half with
6843 * "large" nodes, it also may not fit.
6846 unsigned int psize, nsize;
6847 /* Maximum free space in an empty page */
6848 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6849 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6850 if ((nkeys < 20) || (nsize > pmax/16)) {
6851 if (newindx <= split_indx) {
6854 for (i=0; i<split_indx; i++) {
6855 node = NODEPTR(mp, i);
6856 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6857 if (F_ISSET(node->mn_flags, F_BIGDATA))
6858 psize += sizeof(pgno_t);
6860 psize += NODEDSZ(node);
6864 split_indx = newindx;
6875 for (i=nkeys-1; i>=split_indx; i--) {
6876 node = NODEPTR(mp, i);
6877 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6878 if (F_ISSET(node->mn_flags, F_BIGDATA))
6879 psize += sizeof(pgno_t);
6881 psize += NODEDSZ(node);
6885 split_indx = newindx;
6896 /* First find the separating key between the split pages.
6897 * The case where newindx == split_indx is ambiguous; the
6898 * new item could go to the new page or stay on the original
6899 * page. If newpos == 1 it goes to the new page.
6901 if (newindx == split_indx && newpos) {
6902 sepkey.mv_size = newkey->mv_size;
6903 sepkey.mv_data = newkey->mv_data;
6905 node = NODEPTR(mp, split_indx);
6906 sepkey.mv_size = node->mn_ksize;
6907 sepkey.mv_data = NODEKEY(node);
6911 DPRINTF("separator is [%s]", DKEY(&sepkey));
6913 /* Copy separator key to the parent.
6915 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6919 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6922 if (mn.mc_snum == mc->mc_snum) {
6923 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6924 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6925 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6926 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6931 /* Right page might now have changed parent.
6932 * Check if left page also changed parent.
6934 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6935 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6936 for (i=0; i<ptop; i++) {
6937 mc->mc_pg[i] = mn.mc_pg[i];
6938 mc->mc_ki[i] = mn.mc_ki[i];
6940 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6941 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6945 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6948 mc->mc_flags ^= C_SPLITTING;
6949 if (rc != MDB_SUCCESS) {
6952 if (nflags & MDB_APPEND) {
6953 mc->mc_pg[mc->mc_top] = rp;
6954 mc->mc_ki[mc->mc_top] = 0;
6955 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6958 for (i=0; i<mc->mc_top; i++)
6959 mc->mc_ki[i] = mn.mc_ki[i];
6966 /* Move half of the keys to the right sibling. */
6968 /* grab a page to hold a temporary copy */
6969 copy = mdb_page_malloc(mc, 1);
6973 copy->mp_pgno = mp->mp_pgno;
6974 copy->mp_flags = mp->mp_flags;
6975 copy->mp_lower = PAGEHDRSZ;
6976 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6977 mc->mc_pg[mc->mc_top] = copy;
6978 for (i = j = 0; i <= nkeys; j++) {
6979 if (i == split_indx) {
6980 /* Insert in right sibling. */
6981 /* Reset insert index for right sibling. */
6982 if (i != newindx || (newpos ^ ins_new)) {
6984 mc->mc_pg[mc->mc_top] = rp;
6988 if (i == newindx && !ins_new) {
6989 /* Insert the original entry that caused the split. */
6990 rkey.mv_data = newkey->mv_data;
6991 rkey.mv_size = newkey->mv_size;
7000 /* Update index for the new key. */
7001 mc->mc_ki[mc->mc_top] = j;
7002 } else if (i == nkeys) {
7005 node = NODEPTR(mp, i);
7006 rkey.mv_data = NODEKEY(node);
7007 rkey.mv_size = node->mn_ksize;
7009 xdata.mv_data = NODEDATA(node);
7010 xdata.mv_size = NODEDSZ(node);
7013 pgno = NODEPGNO(node);
7014 flags = node->mn_flags;
7019 if (!IS_LEAF(mp) && j == 0) {
7020 /* First branch index doesn't need key data. */
7024 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7028 nkeys = NUMKEYS(copy);
7029 for (i=0; i<nkeys; i++)
7030 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7031 mp->mp_lower = copy->mp_lower;
7032 mp->mp_upper = copy->mp_upper;
7033 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7034 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7036 /* reset back to original page */
7037 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7038 mc->mc_pg[mc->mc_top] = mp;
7039 if (nflags & MDB_RESERVE) {
7040 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7041 if (!(node->mn_flags & F_BIGDATA))
7042 newdata->mv_data = NODEDATA(node);
7046 /* Make sure mc_ki is still valid.
7048 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7049 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7050 for (i=0; i<ptop; i++) {
7051 mc->mc_pg[i] = mn.mc_pg[i];
7052 mc->mc_ki[i] = mn.mc_ki[i];
7054 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7055 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7059 /* return tmp page to freelist */
7060 mdb_page_free(mc->mc_txn->mt_env, copy);
7063 /* Adjust other cursors pointing to mp */
7064 MDB_cursor *m2, *m3;
7065 MDB_dbi dbi = mc->mc_dbi;
7066 int fixup = NUMKEYS(mp);
7068 if (mc->mc_flags & C_SUB)
7071 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7072 if (m2 == mc) continue;
7073 if (mc->mc_flags & C_SUB)
7074 m3 = &m2->mc_xcursor->mx_cursor;
7077 if (!(m3->mc_flags & C_INITIALIZED))
7079 if (m3->mc_flags & C_SPLITTING)
7084 for (k=m3->mc_top; k>=0; k--) {
7085 m3->mc_ki[k+1] = m3->mc_ki[k];
7086 m3->mc_pg[k+1] = m3->mc_pg[k];
7088 if (m3->mc_ki[0] >= split_indx) {
7093 m3->mc_pg[0] = mc->mc_pg[0];
7097 if (m3->mc_pg[mc->mc_top] == mp) {
7098 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7099 m3->mc_ki[mc->mc_top]++;
7100 if (m3->mc_ki[mc->mc_top] >= fixup) {
7101 m3->mc_pg[mc->mc_top] = rp;
7102 m3->mc_ki[mc->mc_top] -= fixup;
7103 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7105 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7106 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7115 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7116 MDB_val *key, MDB_val *data, unsigned int flags)
7121 assert(key != NULL);
7122 assert(data != NULL);
7124 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7127 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7131 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7135 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7138 mdb_cursor_init(&mc, txn, dbi, &mx);
7139 return mdb_cursor_put(&mc, key, data, flags);
7143 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7145 if ((flag & CHANGEABLE) != flag)
7148 env->me_flags |= flag;
7150 env->me_flags &= ~flag;
7155 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7160 *arg = env->me_flags;
7165 mdb_env_get_path(MDB_env *env, const char **arg)
7170 *arg = env->me_path;
7174 /** Common code for #mdb_stat() and #mdb_env_stat().
7175 * @param[in] env the environment to operate in.
7176 * @param[in] db the #MDB_db record containing the stats to return.
7177 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7178 * @return 0, this function always succeeds.
7181 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7183 arg->ms_psize = env->me_psize;
7184 arg->ms_depth = db->md_depth;
7185 arg->ms_branch_pages = db->md_branch_pages;
7186 arg->ms_leaf_pages = db->md_leaf_pages;
7187 arg->ms_overflow_pages = db->md_overflow_pages;
7188 arg->ms_entries = db->md_entries;
7193 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7197 if (env == NULL || arg == NULL)
7200 toggle = mdb_env_pick_meta(env);
7202 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7206 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7210 if (env == NULL || arg == NULL)
7213 toggle = mdb_env_pick_meta(env);
7214 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7215 arg->me_mapsize = env->me_mapsize;
7216 arg->me_maxreaders = env->me_maxreaders;
7217 arg->me_numreaders = env->me_numreaders;
7218 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7219 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7223 /** Set the default comparison functions for a database.
7224 * Called immediately after a database is opened to set the defaults.
7225 * The user can then override them with #mdb_set_compare() or
7226 * #mdb_set_dupsort().
7227 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7228 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7231 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7233 uint16_t f = txn->mt_dbs[dbi].md_flags;
7235 txn->mt_dbxs[dbi].md_cmp =
7236 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7237 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7239 txn->mt_dbxs[dbi].md_dcmp =
7240 !(f & MDB_DUPSORT) ? 0 :
7241 ((f & MDB_INTEGERDUP)
7242 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7243 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7246 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7251 int rc, dbflag, exact;
7252 unsigned int unused = 0;
7255 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7256 mdb_default_cmp(txn, FREE_DBI);
7259 if ((flags & VALID_FLAGS) != flags)
7265 if (flags & PERSISTENT_FLAGS) {
7266 uint16_t f2 = flags & PERSISTENT_FLAGS;
7267 /* make sure flag changes get committed */
7268 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7269 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7270 txn->mt_flags |= MDB_TXN_DIRTY;
7273 mdb_default_cmp(txn, MAIN_DBI);
7277 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7278 mdb_default_cmp(txn, MAIN_DBI);
7281 /* Is the DB already open? */
7283 for (i=2; i<txn->mt_numdbs; i++) {
7284 if (!txn->mt_dbxs[i].md_name.mv_size) {
7285 /* Remember this free slot */
7286 if (!unused) unused = i;
7289 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7290 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7296 /* If no free slot and max hit, fail */
7297 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7298 return MDB_DBS_FULL;
7300 /* Cannot mix named databases with some mainDB flags */
7301 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7302 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7304 /* Find the DB info */
7305 dbflag = DB_NEW|DB_VALID;
7308 key.mv_data = (void *)name;
7309 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7310 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7311 if (rc == MDB_SUCCESS) {
7312 /* make sure this is actually a DB */
7313 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7314 if (!(node->mn_flags & F_SUBDATA))
7316 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7317 /* Create if requested */
7319 data.mv_size = sizeof(MDB_db);
7320 data.mv_data = &dummy;
7321 memset(&dummy, 0, sizeof(dummy));
7322 dummy.md_root = P_INVALID;
7323 dummy.md_flags = flags & PERSISTENT_FLAGS;
7324 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7328 /* OK, got info, add to table */
7329 if (rc == MDB_SUCCESS) {
7330 unsigned int slot = unused ? unused : txn->mt_numdbs;
7331 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7332 txn->mt_dbxs[slot].md_name.mv_size = len;
7333 txn->mt_dbxs[slot].md_rel = NULL;
7334 txn->mt_dbflags[slot] = dbflag;
7335 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7337 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7338 mdb_default_cmp(txn, slot);
7347 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7349 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7352 if (txn->mt_dbflags[dbi] & DB_STALE) {
7355 /* Stale, must read the DB's root. cursor_init does it for us. */
7356 mdb_cursor_init(&mc, txn, dbi, &mx);
7358 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7361 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7364 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7366 ptr = env->me_dbxs[dbi].md_name.mv_data;
7367 env->me_dbxs[dbi].md_name.mv_data = NULL;
7368 env->me_dbxs[dbi].md_name.mv_size = 0;
7369 env->me_dbflags[dbi] = 0;
7373 /** Add all the DB's pages to the free list.
7374 * @param[in] mc Cursor on the DB to free.
7375 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7376 * @return 0 on success, non-zero on failure.
7379 mdb_drop0(MDB_cursor *mc, int subs)
7383 rc = mdb_page_search(mc, NULL, 0);
7384 if (rc == MDB_SUCCESS) {
7389 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7390 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7393 mdb_cursor_copy(mc, &mx);
7394 while (mc->mc_snum > 0) {
7395 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7396 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7397 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7398 if (ni->mn_flags & F_BIGDATA) {
7402 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7403 rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL);
7406 assert(IS_OVERFLOW(omp));
7407 ovpages = omp->mp_pages;
7408 for (j=0; j<ovpages; j++) {
7409 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7412 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7413 mdb_xcursor_init1(mc, ni);
7414 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7420 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7422 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7425 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7430 mc->mc_ki[mc->mc_top] = i;
7431 rc = mdb_cursor_sibling(mc, 1);
7433 /* no more siblings, go back to beginning
7434 * of previous level.
7438 for (i=1; i<mc->mc_snum; i++) {
7440 mc->mc_pg[i] = mx.mc_pg[i];
7445 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7446 mc->mc_db->md_root);
7451 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7456 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7459 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7462 rc = mdb_cursor_open(txn, dbi, &mc);
7466 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7470 /* Can't delete the main DB */
7471 if (del && dbi > MAIN_DBI) {
7472 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7474 txn->mt_dbflags[dbi] = DB_STALE;
7475 mdb_dbi_close(txn->mt_env, dbi);
7478 /* reset the DB record, mark it dirty */
7479 txn->mt_dbflags[dbi] |= DB_DIRTY;
7480 txn->mt_dbs[dbi].md_depth = 0;
7481 txn->mt_dbs[dbi].md_branch_pages = 0;
7482 txn->mt_dbs[dbi].md_leaf_pages = 0;
7483 txn->mt_dbs[dbi].md_overflow_pages = 0;
7484 txn->mt_dbs[dbi].md_entries = 0;
7485 txn->mt_dbs[dbi].md_root = P_INVALID;
7487 txn->mt_flags |= MDB_TXN_DIRTY;
7490 mdb_cursor_close(mc);
7494 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7496 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7499 txn->mt_dbxs[dbi].md_cmp = cmp;
7503 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7505 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7508 txn->mt_dbxs[dbi].md_dcmp = cmp;
7512 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7514 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7517 txn->mt_dbxs[dbi].md_rel = rel;
7521 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7523 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7526 txn->mt_dbxs[dbi].md_relctx = ctx;