2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <netinet/in.h>
63 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
66 #if defined(__APPLE__) || defined (BSD)
67 # define MDB_USE_POSIX_SEM 1
68 # define MDB_FDATASYNC fsync
69 #elif defined(ANDROID)
70 # define MDB_FDATASYNC fsync
75 #ifdef MDB_USE_POSIX_SEM
76 #include <semaphore.h>
81 #include <valgrind/memcheck.h>
82 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
83 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
84 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
85 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
86 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
88 #define VGMEMP_CREATE(h,r,z)
89 #define VGMEMP_ALLOC(h,a,s)
90 #define VGMEMP_FREE(h,a)
91 #define VGMEMP_DESTROY(h)
92 #define VGMEMP_DEFINED(a,s)
96 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
97 /* Solaris just defines one or the other */
98 # define LITTLE_ENDIAN 1234
99 # define BIG_ENDIAN 4321
100 # ifdef _LITTLE_ENDIAN
101 # define BYTE_ORDER LITTLE_ENDIAN
103 # define BYTE_ORDER BIG_ENDIAN
106 # define BYTE_ORDER __BYTE_ORDER
110 #ifndef LITTLE_ENDIAN
111 #define LITTLE_ENDIAN __LITTLE_ENDIAN
114 #define BIG_ENDIAN __BIG_ENDIAN
117 #if defined(__i386) || defined(__x86_64)
118 #define MISALIGNED_OK 1
124 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
125 # error "Unknown or unsupported endianness (BYTE_ORDER)"
126 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
127 # error "Two's complement, reasonably sized integer types, please"
130 /** @defgroup internal MDB Internals
133 /** @defgroup compat Windows Compatibility Macros
134 * A bunch of macros to minimize the amount of platform-specific ifdefs
135 * needed throughout the rest of the code. When the features this library
136 * needs are similar enough to POSIX to be hidden in a one-or-two line
137 * replacement, this macro approach is used.
141 #define pthread_t DWORD
142 #define pthread_mutex_t HANDLE
143 #define pthread_key_t DWORD
144 #define pthread_self() GetCurrentThreadId()
145 #define pthread_key_create(x,y) \
146 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
147 #define pthread_key_delete(x) TlsFree(x)
148 #define pthread_getspecific(x) TlsGetValue(x)
149 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
150 #define pthread_mutex_unlock(x) ReleaseMutex(x)
151 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
152 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
153 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
154 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
155 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
156 #define getpid() GetCurrentProcessId()
157 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
158 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
159 #define ErrCode() GetLastError()
160 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
161 #define close(fd) CloseHandle(fd)
162 #define munmap(ptr,len) UnmapViewOfFile(ptr)
165 #ifdef MDB_USE_POSIX_SEM
167 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
168 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
169 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
170 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
173 mdb_sem_wait(sem_t *sem)
176 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
181 /** Lock the reader mutex.
183 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
184 /** Unlock the reader mutex.
186 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
188 /** Lock the writer mutex.
189 * Only a single write transaction is allowed at a time. Other writers
190 * will block waiting for this mutex.
192 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
193 /** Unlock the writer mutex.
195 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
196 #endif /* MDB_USE_POSIX_SEM */
198 /** Get the error code for the last failed system function.
200 #define ErrCode() errno
202 /** An abstraction for a file handle.
203 * On POSIX systems file handles are small integers. On Windows
204 * they're opaque pointers.
208 /** A value for an invalid file handle.
209 * Mainly used to initialize file variables and signify that they are
212 #define INVALID_HANDLE_VALUE (-1)
214 /** Get the size of a memory page for the system.
215 * This is the basic size that the platform's memory manager uses, and is
216 * fundamental to the use of memory-mapped files.
218 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
221 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
224 #define MNAME_LEN (sizeof(pthread_mutex_t))
230 /** A flag for opening a file and requesting synchronous data writes.
231 * This is only used when writing a meta page. It's not strictly needed;
232 * we could just do a normal write and then immediately perform a flush.
233 * But if this flag is available it saves us an extra system call.
235 * @note If O_DSYNC is undefined but exists in /usr/include,
236 * preferably set some compiler flag to get the definition.
237 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
240 # define MDB_DSYNC O_DSYNC
244 /** Function for flushing the data of a file. Define this to fsync
245 * if fdatasync() is not supported.
247 #ifndef MDB_FDATASYNC
248 # define MDB_FDATASYNC fdatasync
252 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
263 /** A page number in the database.
264 * Note that 64 bit page numbers are overkill, since pages themselves
265 * already represent 12-13 bits of addressable memory, and the OS will
266 * always limit applications to a maximum of 63 bits of address space.
268 * @note In the #MDB_node structure, we only store 48 bits of this value,
269 * which thus limits us to only 60 bits of addressable data.
271 typedef MDB_ID pgno_t;
273 /** A transaction ID.
274 * See struct MDB_txn.mt_txnid for details.
276 typedef MDB_ID txnid_t;
278 /** @defgroup debug Debug Macros
282 /** Enable debug output.
283 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
284 * read from and written to the database (used for free space management).
289 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
292 # define DPRINTF (void) /* Vararg macros may be unsupported */
294 static int mdb_debug;
295 static txnid_t mdb_debug_start;
297 /** Print a debug message with printf formatting. */
298 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
299 ((void) ((mdb_debug) && \
300 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
302 # define DPRINTF(fmt, ...) ((void) 0)
303 # define MDB_DEBUG_SKIP
305 /** Print a debug string.
306 * The string is printed literally, with no format processing.
308 #define DPUTS(arg) DPRINTF("%s", arg)
311 /** A default memory page size.
312 * The actual size is platform-dependent, but we use this for
313 * boot-strapping. We probably should not be using this any more.
314 * The #GET_PAGESIZE() macro is used to get the actual size.
316 * Note that we don't currently support Huge pages. On Linux,
317 * regular data files cannot use Huge pages, and in general
318 * Huge pages aren't actually pageable. We rely on the OS
319 * demand-pager to read our data and page it out when memory
320 * pressure from other processes is high. So until OSs have
321 * actual paging support for Huge pages, they're not viable.
323 #define MDB_PAGESIZE 4096
325 /** The minimum number of keys required in a database page.
326 * Setting this to a larger value will place a smaller bound on the
327 * maximum size of a data item. Data items larger than this size will
328 * be pushed into overflow pages instead of being stored directly in
329 * the B-tree node. This value used to default to 4. With a page size
330 * of 4096 bytes that meant that any item larger than 1024 bytes would
331 * go into an overflow page. That also meant that on average 2-3KB of
332 * each overflow page was wasted space. The value cannot be lower than
333 * 2 because then there would no longer be a tree structure. With this
334 * value, items larger than 2KB will go into overflow pages, and on
335 * average only 1KB will be wasted.
337 #define MDB_MINKEYS 2
339 /** A stamp that identifies a file as an MDB file.
340 * There's nothing special about this value other than that it is easily
341 * recognizable, and it will reflect any byte order mismatches.
343 #define MDB_MAGIC 0xBEEFC0DE
345 /** The version number for a database's file format. */
346 #define MDB_VERSION 1
348 /** @brief The maximum size of a key in the database.
350 * We require that keys all fit onto a regular page. This limit
351 * could be raised a bit further if needed; to something just
352 * under #MDB_PAGESIZE / #MDB_MINKEYS.
354 * Note that data items in an #MDB_DUPSORT database are actually keys
355 * of a subDB, so they're also limited to this size.
357 #ifndef MDB_MAXKEYSIZE
358 #define MDB_MAXKEYSIZE 511
361 /** @brief The maximum size of a data item.
363 * We only store a 32 bit value for node sizes.
365 #define MAXDATASIZE 0xffffffffUL
370 * This is used for printing a hex dump of a key's contents.
372 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
373 /** Display a key in hex.
375 * Invoke a function to display a key in hex.
377 #define DKEY(x) mdb_dkey(x, kbuf)
379 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
383 /** An invalid page number.
384 * Mainly used to denote an empty tree.
386 #define P_INVALID (~(pgno_t)0)
388 /** Test if the flags \b f are set in a flag word \b w. */
389 #define F_ISSET(w, f) (((w) & (f)) == (f))
391 /** Used for offsets within a single page.
392 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
395 typedef uint16_t indx_t;
397 /** Default size of memory map.
398 * This is certainly too small for any actual applications. Apps should always set
399 * the size explicitly using #mdb_env_set_mapsize().
401 #define DEFAULT_MAPSIZE 1048576
403 /** @defgroup readers Reader Lock Table
404 * Readers don't acquire any locks for their data access. Instead, they
405 * simply record their transaction ID in the reader table. The reader
406 * mutex is needed just to find an empty slot in the reader table. The
407 * slot's address is saved in thread-specific data so that subsequent read
408 * transactions started by the same thread need no further locking to proceed.
410 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
412 * No reader table is used if the database is on a read-only filesystem.
414 * Since the database uses multi-version concurrency control, readers don't
415 * actually need any locking. This table is used to keep track of which
416 * readers are using data from which old transactions, so that we'll know
417 * when a particular old transaction is no longer in use. Old transactions
418 * that have discarded any data pages can then have those pages reclaimed
419 * for use by a later write transaction.
421 * The lock table is constructed such that reader slots are aligned with the
422 * processor's cache line size. Any slot is only ever used by one thread.
423 * This alignment guarantees that there will be no contention or cache
424 * thrashing as threads update their own slot info, and also eliminates
425 * any need for locking when accessing a slot.
427 * A writer thread will scan every slot in the table to determine the oldest
428 * outstanding reader transaction. Any freed pages older than this will be
429 * reclaimed by the writer. The writer doesn't use any locks when scanning
430 * this table. This means that there's no guarantee that the writer will
431 * see the most up-to-date reader info, but that's not required for correct
432 * operation - all we need is to know the upper bound on the oldest reader,
433 * we don't care at all about the newest reader. So the only consequence of
434 * reading stale information here is that old pages might hang around a
435 * while longer before being reclaimed. That's actually good anyway, because
436 * the longer we delay reclaiming old pages, the more likely it is that a
437 * string of contiguous pages can be found after coalescing old pages from
438 * many old transactions together.
441 /** Number of slots in the reader table.
442 * This value was chosen somewhat arbitrarily. 126 readers plus a
443 * couple mutexes fit exactly into 8KB on my development machine.
444 * Applications should set the table size using #mdb_env_set_maxreaders().
446 #define DEFAULT_READERS 126
448 /** The size of a CPU cache line in bytes. We want our lock structures
449 * aligned to this size to avoid false cache line sharing in the
451 * This value works for most CPUs. For Itanium this should be 128.
457 /** The information we store in a single slot of the reader table.
458 * In addition to a transaction ID, we also record the process and
459 * thread ID that owns a slot, so that we can detect stale information,
460 * e.g. threads or processes that went away without cleaning up.
461 * @note We currently don't check for stale records. We simply re-init
462 * the table when we know that we're the only process opening the
465 typedef struct MDB_rxbody {
466 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
467 * Multiple readers that start at the same time will probably have the
468 * same ID here. Again, it's not important to exclude them from
469 * anything; all we need to know is which version of the DB they
470 * started from so we can avoid overwriting any data used in that
471 * particular version.
474 /** The process ID of the process owning this reader txn. */
476 /** The thread ID of the thread owning this txn. */
480 /** The actual reader record, with cacheline padding. */
481 typedef struct MDB_reader {
484 /** shorthand for mrb_txnid */
485 #define mr_txnid mru.mrx.mrb_txnid
486 #define mr_pid mru.mrx.mrb_pid
487 #define mr_tid mru.mrx.mrb_tid
488 /** cache line alignment */
489 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
493 /** The header for the reader table.
494 * The table resides in a memory-mapped file. (This is a different file
495 * than is used for the main database.)
497 * For POSIX the actual mutexes reside in the shared memory of this
498 * mapped file. On Windows, mutexes are named objects allocated by the
499 * kernel; we store the mutex names in this mapped file so that other
500 * processes can grab them. This same approach is also used on
501 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
502 * process-shared POSIX mutexes. For these cases where a named object
503 * is used, the object name is derived from a 64 bit FNV hash of the
504 * environment pathname. As such, naming collisions are extremely
505 * unlikely. If a collision occurs, the results are unpredictable.
507 typedef struct MDB_txbody {
508 /** Stamp identifying this as an MDB file. It must be set
511 /** Version number of this lock file. Must be set to #MDB_VERSION. */
512 uint32_t mtb_version;
513 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
514 char mtb_rmname[MNAME_LEN];
516 /** Mutex protecting access to this table.
517 * This is the reader lock that #LOCK_MUTEX_R acquires.
519 pthread_mutex_t mtb_mutex;
521 /** The ID of the last transaction committed to the database.
522 * This is recorded here only for convenience; the value can always
523 * be determined by reading the main database meta pages.
526 /** The number of slots that have been used in the reader table.
527 * This always records the maximum count, it is not decremented
528 * when readers release their slots.
530 unsigned mtb_numreaders;
533 /** The actual reader table definition. */
534 typedef struct MDB_txninfo {
537 #define mti_magic mt1.mtb.mtb_magic
538 #define mti_version mt1.mtb.mtb_version
539 #define mti_mutex mt1.mtb.mtb_mutex
540 #define mti_rmname mt1.mtb.mtb_rmname
541 #define mti_txnid mt1.mtb.mtb_txnid
542 #define mti_numreaders mt1.mtb.mtb_numreaders
543 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
546 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
547 char mt2_wmname[MNAME_LEN];
548 #define mti_wmname mt2.mt2_wmname
550 pthread_mutex_t mt2_wmutex;
551 #define mti_wmutex mt2.mt2_wmutex
553 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
555 MDB_reader mti_readers[1];
559 /** Common header for all page types.
560 * Overflow records occupy a number of contiguous pages with no
561 * headers on any page after the first.
563 typedef struct MDB_page {
564 #define mp_pgno mp_p.p_pgno
565 #define mp_next mp_p.p_next
567 pgno_t p_pgno; /**< page number */
568 void * p_next; /**< for in-memory list of freed structs */
571 /** @defgroup mdb_page Page Flags
573 * Flags for the page headers.
576 #define P_BRANCH 0x01 /**< branch page */
577 #define P_LEAF 0x02 /**< leaf page */
578 #define P_OVERFLOW 0x04 /**< overflow page */
579 #define P_META 0x08 /**< meta page */
580 #define P_DIRTY 0x10 /**< dirty page */
581 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
582 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
584 uint16_t mp_flags; /**< @ref mdb_page */
585 #define mp_lower mp_pb.pb.pb_lower
586 #define mp_upper mp_pb.pb.pb_upper
587 #define mp_pages mp_pb.pb_pages
590 indx_t pb_lower; /**< lower bound of free space */
591 indx_t pb_upper; /**< upper bound of free space */
593 uint32_t pb_pages; /**< number of overflow pages */
595 indx_t mp_ptrs[1]; /**< dynamic size */
598 /** Size of the page header, excluding dynamic data at the end */
599 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
601 /** Address of first usable data byte in a page, after the header */
602 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
604 /** Number of nodes on a page */
605 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
607 /** The amount of space remaining in the page */
608 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
610 /** The percentage of space used in the page, in tenths of a percent. */
611 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
612 ((env)->me_psize - PAGEHDRSZ))
613 /** The minimum page fill factor, in tenths of a percent.
614 * Pages emptier than this are candidates for merging.
616 #define FILL_THRESHOLD 250
618 /** Test if a page is a leaf page */
619 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
620 /** Test if a page is a LEAF2 page */
621 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
622 /** Test if a page is a branch page */
623 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
624 /** Test if a page is an overflow page */
625 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
626 /** Test if a page is a sub page */
627 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
629 /** The number of overflow pages needed to store the given size. */
630 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
632 /** Header for a single key/data pair within a page.
633 * We guarantee 2-byte alignment for nodes.
635 typedef struct MDB_node {
636 /** lo and hi are used for data size on leaf nodes and for
637 * child pgno on branch nodes. On 64 bit platforms, flags
638 * is also used for pgno. (Branch nodes have no flags).
639 * They are in host byte order in case that lets some
640 * accesses be optimized into a 32-bit word access.
642 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
643 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
644 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
645 /** @defgroup mdb_node Node Flags
647 * Flags for node headers.
650 #define F_BIGDATA 0x01 /**< data put on overflow page */
651 #define F_SUBDATA 0x02 /**< data is a sub-database */
652 #define F_DUPDATA 0x04 /**< data has duplicates */
654 /** valid flags for #mdb_node_add() */
655 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
658 unsigned short mn_flags; /**< @ref mdb_node */
659 unsigned short mn_ksize; /**< key size */
660 char mn_data[1]; /**< key and data are appended here */
663 /** Size of the node header, excluding dynamic data at the end */
664 #define NODESIZE offsetof(MDB_node, mn_data)
666 /** Bit position of top word in page number, for shifting mn_flags */
667 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
669 /** Size of a node in a branch page with a given key.
670 * This is just the node header plus the key, there is no data.
672 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
674 /** Size of a node in a leaf page with a given key and data.
675 * This is node header plus key plus data size.
677 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
679 /** Address of node \b i in page \b p */
680 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
682 /** Address of the key for the node */
683 #define NODEKEY(node) (void *)((node)->mn_data)
685 /** Address of the data for a node */
686 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
688 /** Get the page number pointed to by a branch node */
689 #define NODEPGNO(node) \
690 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
691 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
692 /** Set the page number in a branch node */
693 #define SETPGNO(node,pgno) do { \
694 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
695 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
697 /** Get the size of the data in a leaf node */
698 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
699 /** Set the size of the data for a leaf node */
700 #define SETDSZ(node,size) do { \
701 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
702 /** The size of a key in a node */
703 #define NODEKSZ(node) ((node)->mn_ksize)
705 /** Copy a page number from src to dst */
707 #define COPY_PGNO(dst,src) dst = src
709 #if SIZE_MAX > 4294967295UL
710 #define COPY_PGNO(dst,src) do { \
711 unsigned short *s, *d; \
712 s = (unsigned short *)&(src); \
713 d = (unsigned short *)&(dst); \
720 #define COPY_PGNO(dst,src) do { \
721 unsigned short *s, *d; \
722 s = (unsigned short *)&(src); \
723 d = (unsigned short *)&(dst); \
729 /** The address of a key in a LEAF2 page.
730 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
731 * There are no node headers, keys are stored contiguously.
733 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
735 /** Set the \b node's key into \b key, if requested. */
736 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
737 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
739 /** Information about a single database in the environment. */
740 typedef struct MDB_db {
741 uint32_t md_pad; /**< also ksize for LEAF2 pages */
742 uint16_t md_flags; /**< @ref mdb_dbi_open */
743 uint16_t md_depth; /**< depth of this tree */
744 pgno_t md_branch_pages; /**< number of internal pages */
745 pgno_t md_leaf_pages; /**< number of leaf pages */
746 pgno_t md_overflow_pages; /**< number of overflow pages */
747 size_t md_entries; /**< number of data items */
748 pgno_t md_root; /**< the root page of this tree */
751 /** mdb_dbi_open flags */
752 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
753 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
754 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
755 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
757 /** Handle for the DB used to track free pages. */
759 /** Handle for the default DB. */
762 /** Meta page content. */
763 typedef struct MDB_meta {
764 /** Stamp identifying this as an MDB file. It must be set
767 /** Version number of this lock file. Must be set to #MDB_VERSION. */
769 void *mm_address; /**< address for fixed mapping */
770 size_t mm_mapsize; /**< size of mmap region */
771 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
772 /** The size of pages used in this DB */
773 #define mm_psize mm_dbs[0].md_pad
774 /** Any persistent environment flags. @ref mdb_env */
775 #define mm_flags mm_dbs[0].md_flags
776 pgno_t mm_last_pg; /**< last used page in file */
777 txnid_t mm_txnid; /**< txnid that committed this page */
780 /** Buffer for a stack-allocated dirty page.
781 * The members define size and alignment, and silence type
782 * aliasing warnings. They are not used directly; that could
783 * mean incorrectly using several union members in parallel.
785 typedef union MDB_pagebuf {
786 char mb_raw[MDB_PAGESIZE];
789 char mm_pad[PAGEHDRSZ];
794 /** Auxiliary DB info.
795 * The information here is mostly static/read-only. There is
796 * only a single copy of this record in the environment.
798 typedef struct MDB_dbx {
799 MDB_val md_name; /**< name of the database */
800 MDB_cmp_func *md_cmp; /**< function for comparing keys */
801 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
802 MDB_rel_func *md_rel; /**< user relocate function */
803 void *md_relctx; /**< user-provided context for md_rel */
806 /** A database transaction.
807 * Every operation requires a transaction handle.
810 MDB_txn *mt_parent; /**< parent of a nested txn */
811 MDB_txn *mt_child; /**< nested txn under this txn */
812 pgno_t mt_next_pgno; /**< next unallocated page */
813 /** The ID of this transaction. IDs are integers incrementing from 1.
814 * Only committed write transactions increment the ID. If a transaction
815 * aborts, the ID may be re-used by the next writer.
818 MDB_env *mt_env; /**< the DB environment */
819 /** The list of pages that became unused during this transaction.
823 MDB_ID2L dirty_list; /**< for write txns: modified pages */
824 MDB_reader *reader; /**< this thread's reader table slot or NULL */
826 /** Array of records for each DB known in the environment. */
828 /** Array of MDB_db records for each known DB */
830 /** @defgroup mt_dbflag Transaction DB Flags
834 #define DB_DIRTY 0x01 /**< DB was written in this txn */
835 #define DB_STALE 0x02 /**< DB record is older than txnID */
836 #define DB_NEW 0x04 /**< DB handle opened in this txn */
837 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
839 /** In write txns, array of cursors for each DB */
840 MDB_cursor **mt_cursors;
841 /** Array of flags for each DB */
842 unsigned char *mt_dbflags;
843 /** Number of DB records in use. This number only ever increments;
844 * we don't decrement it when individual DB handles are closed.
848 /** @defgroup mdb_txn Transaction Flags
852 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
853 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
854 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
856 unsigned int mt_flags; /**< @ref mdb_txn */
857 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
858 unsigned int mt_dirty_room;
859 /** Tracks which of the two meta pages was used at the start
860 * of this transaction.
862 unsigned int mt_toggle;
865 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
866 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
867 * raise this on a 64 bit machine.
869 #define CURSOR_STACK 32
873 /** Cursors are used for all DB operations */
875 /** Next cursor on this DB in this txn */
877 /** Original cursor if this is a shadow */
879 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
880 struct MDB_xcursor *mc_xcursor;
881 /** The transaction that owns this cursor */
883 /** The database handle this cursor operates on */
885 /** The database record for this cursor */
887 /** The database auxiliary record for this cursor */
889 /** The @ref mt_dbflag for this database */
890 unsigned char *mc_dbflag;
891 unsigned short mc_snum; /**< number of pushed pages */
892 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
893 /** @defgroup mdb_cursor Cursor Flags
895 * Cursor state flags.
898 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
899 #define C_EOF 0x02 /**< No more data */
900 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
901 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
902 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
903 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
904 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
906 unsigned int mc_flags; /**< @ref mdb_cursor */
907 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
908 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
911 /** Context for sorted-dup records.
912 * We could have gone to a fully recursive design, with arbitrarily
913 * deep nesting of sub-databases. But for now we only handle these
914 * levels - main DB, optional sub-DB, sorted-duplicate DB.
916 typedef struct MDB_xcursor {
917 /** A sub-cursor for traversing the Dup DB */
918 MDB_cursor mx_cursor;
919 /** The database record for this Dup DB */
921 /** The auxiliary DB record for this Dup DB */
923 /** The @ref mt_dbflag for this Dup DB */
924 unsigned char mx_dbflag;
927 /** State of FreeDB old pages, stored in the MDB_env */
928 typedef struct MDB_pgstate {
929 txnid_t mf_pglast; /**< ID of last old page record we used */
930 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
931 pgno_t *mf_pgfree; /**< memory to free when dropping me_pghead */
934 /** The database environment. */
936 HANDLE me_fd; /**< The main data file */
937 HANDLE me_lfd; /**< The lock file */
938 HANDLE me_mfd; /**< just for writing the meta pages */
939 /** Failed to update the meta page. Probably an I/O error. */
940 #define MDB_FATAL_ERROR 0x80000000U
941 /** Some fields are initialized. */
942 #define MDB_ENV_ACTIVE 0x20000000U
943 /** me_txkey is set */
944 #define MDB_ENV_TXKEY 0x10000000U
945 uint32_t me_flags; /**< @ref mdb_env */
946 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
947 unsigned int me_maxreaders; /**< size of the reader table */
948 unsigned int me_numreaders; /**< max numreaders set by this env */
949 MDB_dbi me_numdbs; /**< number of DBs opened */
950 MDB_dbi me_maxdbs; /**< size of the DB table */
951 pid_t me_pid; /**< process ID of this env */
952 char *me_path; /**< path to the DB files */
953 char *me_map; /**< the memory map of the data file */
954 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
955 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
956 MDB_txn *me_txn; /**< current write transaction */
957 size_t me_mapsize; /**< size of the data memory map */
958 off_t me_size; /**< current file size */
959 pgno_t me_maxpg; /**< me_mapsize / me_psize */
960 MDB_dbx *me_dbxs; /**< array of static DB info */
961 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
962 pthread_key_t me_txkey; /**< thread-key for readers */
963 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
964 # define me_pglast me_pgstate.mf_pglast
965 # define me_pghead me_pgstate.mf_pghead
966 # define me_pgfree me_pgstate.mf_pgfree
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 */
973 unsigned int me_maxfree_1pg;
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 = malloc(MDB_IDL_SIZEOF(idl));
1410 txn->mt_env->me_pglast = last;
1411 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = 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=0; i<idl[0]; i++) {
1420 DPRINTF("IDL %zu", idl[i+1]);
1427 if (txn->mt_env->me_pghead) {
1428 pgno_t *mop = txn->mt_env->me_pghead;
1431 int retry = 1, readit = 0, n2 = num-1;
1432 unsigned int i, j, k;
1434 /* If current list is too short, must fetch more and coalesce */
1435 if (mop[0] < (unsigned)num)
1438 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1440 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1441 /* If on freelist, don't try to read more. If what we have
1442 * right now isn't enough just use new pages.
1443 * TODO: get all of this working. Many circular dependencies...
1445 if (mc->mc_dbi == FREE_DBI) {
1454 last = txn->mt_env->me_pglast + 1;
1456 /* We haven't hit the readers list yet? */
1458 oldest = mdb_find_oldest(txn);
1461 /* There's nothing we can use on the freelist */
1462 if (oldest - last < 1)
1465 key.mv_data = &last;
1466 key.mv_size = sizeof(last);
1467 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1469 if (rc == MDB_NOTFOUND)
1473 last = *(txnid_t*)key.mv_data;
1476 idl = (MDB_ID *) data.mv_data;
1477 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1480 /* merge in sorted order */
1481 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1483 while (i>0 || j>0) {
1484 if (i && idl[i] < mop[j])
1485 mop2[k--] = idl[i--];
1487 mop2[k--] = mop[j--];
1489 txn->mt_env->me_pglast = last;
1490 free(txn->mt_env->me_pgfree);
1491 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1493 /* Keep trying to read until we have enough */
1494 if (mop[0] < (unsigned)num) {
1499 /* current list has enough pages, but are they contiguous? */
1500 for (i=mop[0]; i>=(unsigned)num; i--) {
1501 if (mop[i-n2] == mop[i] + n2) {
1504 /* move any stragglers down */
1505 for (j=i+num; j<=mop[0]; j++)
1512 /* Stop if we succeeded, or no retries */
1513 if (!retry || pgno != P_INVALID)
1519 /* peel pages off tail, so we only have to truncate the list */
1520 pgno = MDB_IDL_LAST(mop);
1523 if (MDB_IDL_IS_ZERO(mop)) {
1524 free(txn->mt_env->me_pgfree);
1525 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1530 if (pgno == P_INVALID) {
1531 /* DB size is maxed out */
1532 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1533 DPUTS("DB size maxed out");
1534 return MDB_MAP_FULL;
1537 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1538 if (pgno == P_INVALID) {
1539 pgno = txn->mt_next_pgno;
1540 txn->mt_next_pgno += num;
1542 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1545 if (!(np = mdb_page_malloc(mc, num)))
1547 if (pgno == P_INVALID) {
1548 np->mp_pgno = txn->mt_next_pgno;
1549 txn->mt_next_pgno += num;
1554 mid.mid = np->mp_pgno;
1556 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1557 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1559 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1561 txn->mt_dirty_room--;
1567 /** Copy a page: avoid copying unused portions of the page.
1568 * @param[in] dst page to copy into
1569 * @param[in] src page to copy from
1570 * @param[in] psize size of a page
1573 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1575 dst->mp_flags = src->mp_flags | P_DIRTY;
1576 dst->mp_pages = src->mp_pages;
1578 if (IS_LEAF2(src)) {
1579 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1581 unsigned int i, nkeys = NUMKEYS(src);
1582 for (i=0; i<nkeys; i++)
1583 dst->mp_ptrs[i] = src->mp_ptrs[i];
1584 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1585 psize - src->mp_upper);
1589 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1590 * @param[in] mc cursor pointing to the page to be touched
1591 * @return 0 on success, non-zero on failure.
1594 mdb_page_touch(MDB_cursor *mc)
1596 MDB_page *mp = mc->mc_pg[mc->mc_top];
1600 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1602 if ((rc = mdb_page_alloc(mc, 1, &np)))
1604 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1605 assert(mp->mp_pgno != np->mp_pgno);
1606 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1608 /* If page isn't full, just copy the used portion */
1609 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1612 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1614 np->mp_flags |= P_DIRTY;
1619 /* Adjust other cursors pointing to mp */
1620 if (mc->mc_flags & C_SUB) {
1621 MDB_cursor *m2, *m3;
1622 MDB_dbi dbi = mc->mc_dbi-1;
1624 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1625 if (m2 == mc) continue;
1626 m3 = &m2->mc_xcursor->mx_cursor;
1627 if (m3->mc_snum < mc->mc_snum) continue;
1628 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1629 m3->mc_pg[mc->mc_top] = mp;
1635 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1636 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1637 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1638 m2->mc_pg[mc->mc_top] = mp;
1642 mc->mc_pg[mc->mc_top] = mp;
1643 /** If this page has a parent, update the parent to point to
1647 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1649 mc->mc_db->md_root = mp->mp_pgno;
1650 } else if (mc->mc_txn->mt_parent && !(mp->mp_flags & P_SUBP)) {
1652 MDB_ID2 mid, *dl = mc->mc_txn->mt_u.dirty_list;
1653 /* If txn has a parent, make sure the page is in our
1657 unsigned x = mdb_mid2l_search(dl, mp->mp_pgno);
1658 if (x <= dl[0].mid && dl[x].mid == mp->mp_pgno) {
1661 mc->mc_pg[mc->mc_top] = np;
1665 assert(dl[0].mid < MDB_IDL_UM_MAX);
1667 np = mdb_page_malloc(mc, 1);
1670 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1671 mid.mid = np->mp_pgno;
1673 mdb_mid2l_insert(dl, &mid);
1681 mdb_env_sync(MDB_env *env, int force)
1684 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1685 if (env->me_flags & MDB_WRITEMAP) {
1686 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1687 ? MS_ASYNC : MS_SYNC;
1688 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1691 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1695 if (MDB_FDATASYNC(env->me_fd))
1702 /** Make shadow copies of all of parent txn's cursors */
1704 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1706 MDB_cursor *mc, *m2;
1707 unsigned int i, j, size;
1709 for (i=0;i<src->mt_numdbs; i++) {
1710 if (src->mt_cursors[i]) {
1711 size = sizeof(MDB_cursor);
1712 if (src->mt_cursors[i]->mc_xcursor)
1713 size += sizeof(MDB_xcursor);
1714 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1721 mc->mc_db = &dst->mt_dbs[i];
1722 mc->mc_dbx = m2->mc_dbx;
1723 mc->mc_dbflag = &dst->mt_dbflags[i];
1724 mc->mc_snum = m2->mc_snum;
1725 mc->mc_top = m2->mc_top;
1726 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1727 for (j=0; j<mc->mc_snum; j++) {
1728 mc->mc_pg[j] = m2->mc_pg[j];
1729 mc->mc_ki[j] = m2->mc_ki[j];
1731 if (m2->mc_xcursor) {
1732 MDB_xcursor *mx, *mx2;
1733 mx = (MDB_xcursor *)(mc+1);
1734 mc->mc_xcursor = mx;
1735 mx2 = m2->mc_xcursor;
1736 mx->mx_db = mx2->mx_db;
1737 mx->mx_dbx = mx2->mx_dbx;
1738 mx->mx_dbflag = mx2->mx_dbflag;
1739 mx->mx_cursor.mc_txn = dst;
1740 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1741 mx->mx_cursor.mc_db = &mx->mx_db;
1742 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1743 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1744 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1745 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1746 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1747 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1748 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1749 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1752 mc->mc_xcursor = NULL;
1754 mc->mc_next = dst->mt_cursors[i];
1755 dst->mt_cursors[i] = mc;
1762 /** Close this write txn's cursors, after optionally merging its shadow
1763 * cursors back into parent's.
1764 * @param[in] txn the transaction handle.
1765 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1766 * @return 0 on success, non-zero on failure.
1769 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1771 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1774 for (i = txn->mt_numdbs; --i >= 0; ) {
1775 for (mc = cursors[i]; mc; mc = next) {
1777 if (mc->mc_flags & merge) {
1778 MDB_cursor *m2 = mc->mc_orig;
1779 m2->mc_snum = mc->mc_snum;
1780 m2->mc_top = mc->mc_top;
1781 for (j = mc->mc_snum; --j >= 0; ) {
1782 m2->mc_pg[j] = mc->mc_pg[j];
1783 m2->mc_ki[j] = mc->mc_ki[j];
1786 if (mc->mc_flags & C_ALLOCD)
1794 mdb_txn_reset0(MDB_txn *txn);
1796 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1797 * @param[in] txn the transaction handle to initialize
1798 * @return 0 on success, non-zero on failure.
1801 mdb_txn_renew0(MDB_txn *txn)
1803 MDB_env *env = txn->mt_env;
1806 int rc, new_notls = 0;
1809 txn->mt_numdbs = env->me_numdbs;
1810 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1812 if (txn->mt_flags & MDB_TXN_RDONLY) {
1813 if (!env->me_txns) {
1814 i = mdb_env_pick_meta(env);
1815 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1816 txn->mt_u.reader = NULL;
1818 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1819 pthread_getspecific(env->me_txkey);
1821 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1822 return MDB_BAD_RSLOT;
1824 pid_t pid = env->me_pid;
1825 pthread_t tid = pthread_self();
1828 for (i=0; i<env->me_txns->mti_numreaders; i++)
1829 if (env->me_txns->mti_readers[i].mr_pid == 0)
1831 if (i == env->me_maxreaders) {
1832 UNLOCK_MUTEX_R(env);
1833 return MDB_READERS_FULL;
1835 env->me_txns->mti_readers[i].mr_pid = pid;
1836 env->me_txns->mti_readers[i].mr_tid = tid;
1837 if (i >= env->me_txns->mti_numreaders)
1838 env->me_txns->mti_numreaders = i+1;
1839 /* Save numreaders for un-mutexed mdb_env_close() */
1840 env->me_numreaders = env->me_txns->mti_numreaders;
1841 UNLOCK_MUTEX_R(env);
1842 r = &env->me_txns->mti_readers[i];
1843 new_notls = (env->me_flags & MDB_NOTLS);
1844 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1849 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1850 txn->mt_u.reader = r;
1852 txn->mt_toggle = txn->mt_txnid & 1;
1853 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1857 txn->mt_txnid = env->me_txns->mti_txnid;
1858 txn->mt_toggle = txn->mt_txnid & 1;
1859 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1862 if (txn->mt_txnid == mdb_debug_start)
1865 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1866 txn->mt_u.dirty_list = env->me_dirty_list;
1867 txn->mt_u.dirty_list[0].mid = 0;
1868 txn->mt_free_pgs = env->me_free_pgs;
1869 txn->mt_free_pgs[0] = 0;
1873 /* Copy the DB info and flags */
1874 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1875 for (i=2; i<txn->mt_numdbs; i++) {
1876 x = env->me_dbflags[i];
1877 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1878 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1880 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1882 if (env->me_maxpg < txn->mt_next_pgno) {
1883 mdb_txn_reset0(txn);
1885 txn->mt_u.reader->mr_pid = 0;
1886 txn->mt_u.reader = NULL;
1888 return MDB_MAP_RESIZED;
1895 mdb_txn_renew(MDB_txn *txn)
1899 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1902 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1903 DPUTS("environment had fatal error, must shutdown!");
1907 rc = mdb_txn_renew0(txn);
1908 if (rc == MDB_SUCCESS) {
1909 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1910 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1911 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1917 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1921 int rc, size, tsize = sizeof(MDB_txn);
1923 if (env->me_flags & MDB_FATAL_ERROR) {
1924 DPUTS("environment had fatal error, must shutdown!");
1927 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1930 /* Nested transactions: Max 1 child, write txns only, no writemap */
1931 if (parent->mt_child ||
1932 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1933 (env->me_flags & MDB_WRITEMAP))
1937 tsize = sizeof(MDB_ntxn);
1939 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1940 if (!(flags & MDB_RDONLY))
1941 size += env->me_maxdbs * sizeof(MDB_cursor *);
1943 if ((txn = calloc(1, size)) == NULL) {
1944 DPRINTF("calloc: %s", strerror(ErrCode()));
1947 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1948 if (flags & MDB_RDONLY) {
1949 txn->mt_flags |= MDB_TXN_RDONLY;
1950 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1952 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1953 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1959 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1960 if (!txn->mt_u.dirty_list ||
1961 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1963 free(txn->mt_u.dirty_list);
1967 txn->mt_txnid = parent->mt_txnid;
1968 txn->mt_toggle = parent->mt_toggle;
1969 txn->mt_dirty_room = parent->mt_dirty_room;
1970 txn->mt_u.dirty_list[0].mid = 0;
1971 txn->mt_free_pgs[0] = 0;
1972 txn->mt_next_pgno = parent->mt_next_pgno;
1973 parent->mt_child = txn;
1974 txn->mt_parent = parent;
1975 txn->mt_numdbs = parent->mt_numdbs;
1976 txn->mt_dbxs = parent->mt_dbxs;
1977 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1978 /* Copy parent's mt_dbflags, but clear DB_NEW */
1979 for (i=0; i<txn->mt_numdbs; i++)
1980 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1982 ntxn = (MDB_ntxn *)txn;
1983 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1984 if (env->me_pghead) {
1985 size = MDB_IDL_SIZEOF(env->me_pghead);
1986 env->me_pghead = malloc(size);
1988 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1992 env->me_pgfree = env->me_pghead;
1994 rc = mdb_cursor_shadow(parent, txn);
1996 mdb_txn_reset0(txn);
1998 rc = mdb_txn_renew0(txn);
2004 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2005 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2006 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2012 /** Export or close DBI handles opened in this txn. */
2014 mdb_dbis_update(MDB_txn *txn, int keep)
2017 MDB_dbi n = txn->mt_numdbs;
2018 MDB_env *env = txn->mt_env;
2019 unsigned char *tdbflags = txn->mt_dbflags;
2021 for (i = n; --i >= 2;) {
2022 if (tdbflags[i] & DB_NEW) {
2024 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2026 char *ptr = env->me_dbxs[i].md_name.mv_data;
2027 env->me_dbxs[i].md_name.mv_data = NULL;
2028 env->me_dbxs[i].md_name.mv_size = 0;
2029 env->me_dbflags[i] = 0;
2034 if (keep && env->me_numdbs < n)
2038 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2039 * May be called twice for readonly txns: First reset it, then abort.
2040 * @param[in] txn the transaction handle to reset
2043 mdb_txn_reset0(MDB_txn *txn)
2045 MDB_env *env = txn->mt_env;
2047 /* Close any DBI handles opened in this txn */
2048 mdb_dbis_update(txn, 0);
2050 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2051 if (txn->mt_u.reader) {
2052 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2053 if (!(env->me_flags & MDB_NOTLS))
2054 txn->mt_u.reader = NULL; /* txn does not own reader */
2056 txn->mt_numdbs = 0; /* close nothing if called again */
2057 txn->mt_dbxs = NULL; /* mark txn as reset */
2059 mdb_cursors_close(txn, 0);
2061 if (!(env->me_flags & MDB_WRITEMAP)) {
2062 mdb_dlist_free(txn);
2064 free(env->me_pgfree);
2066 if (txn->mt_parent) {
2067 txn->mt_parent->mt_child = NULL;
2068 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2069 mdb_midl_free(txn->mt_free_pgs);
2070 free(txn->mt_u.dirty_list);
2073 if (mdb_midl_shrink(&txn->mt_free_pgs))
2074 env->me_free_pgs = txn->mt_free_pgs;
2077 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2078 txn->mt_env->me_pglast = 0;
2081 /* The writer mutex was locked in mdb_txn_begin. */
2082 UNLOCK_MUTEX_W(env);
2087 mdb_txn_reset(MDB_txn *txn)
2092 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2093 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2094 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2096 /* This call is only valid for read-only txns */
2097 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2100 mdb_txn_reset0(txn);
2104 mdb_txn_abort(MDB_txn *txn)
2109 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2110 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2111 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2114 mdb_txn_abort(txn->mt_child);
2116 mdb_txn_reset0(txn);
2117 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2118 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2119 txn->mt_u.reader->mr_pid = 0;
2125 mdb_txn_commit(MDB_txn *txn)
2133 pgno_t next, freecnt;
2134 txnid_t oldpg_txnid, id;
2137 assert(txn != NULL);
2138 assert(txn->mt_env != NULL);
2140 if (txn->mt_child) {
2141 rc = mdb_txn_commit(txn->mt_child);
2142 txn->mt_child = NULL;
2151 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2152 mdb_dbis_update(txn, 1);
2153 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2158 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2159 DPUTS("error flag is set, can't commit");
2161 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2166 if (txn->mt_parent) {
2167 MDB_txn *parent = txn->mt_parent;
2171 /* Append our free list to parent's */
2172 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2176 mdb_midl_free(txn->mt_free_pgs);
2178 parent->mt_next_pgno = txn->mt_next_pgno;
2179 parent->mt_flags = txn->mt_flags;
2181 /* Merge our cursors into parent's and close them */
2182 mdb_cursors_close(txn, C_SHADOW);
2184 /* Update parent's DB table. */
2185 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2186 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2187 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2188 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2189 for (i=2; i<txn->mt_numdbs; i++) {
2190 /* preserve parent's DB_NEW status */
2191 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2192 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2195 dst = txn->mt_parent->mt_u.dirty_list;
2196 src = txn->mt_u.dirty_list;
2197 /* Find len = length of merging our dirty list with parent's */
2199 dst[0].mid = 0; /* simplify loops */
2200 if (parent->mt_parent) {
2201 len = x + src[0].mid;
2202 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2203 for (i = x; y && i; y--) {
2204 pgno_t yp = src[y].mid;
2205 while (yp < dst[i].mid)
2207 if (yp == dst[i].mid) {
2212 } else { /* Simplify the above for single-ancestor case */
2213 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2215 /* Merge our dirty list with parent's */
2217 for (i = len; y; dst[i--] = src[y--]) {
2218 pgno_t yp = src[y].mid;
2219 while (yp < dst[x].mid)
2220 dst[i--] = dst[x--];
2221 if (yp == dst[x].mid)
2222 free(dst[x--].mptr);
2226 free(txn->mt_u.dirty_list);
2227 parent->mt_dirty_room = txn->mt_dirty_room;
2229 txn->mt_parent->mt_child = NULL;
2230 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2235 if (txn != env->me_txn) {
2236 DPUTS("attempt to commit unknown transaction");
2241 mdb_cursors_close(txn, 0);
2243 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2246 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2247 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2249 /* Update DB root pointers */
2250 if (txn->mt_numdbs > 2) {
2253 data.mv_size = sizeof(MDB_db);
2255 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2256 for (i = 2; i < txn->mt_numdbs; i++) {
2257 if (txn->mt_dbflags[i] & DB_DIRTY) {
2258 data.mv_data = &txn->mt_dbs[i];
2259 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2266 /* Save the freelist as of this transaction to the freeDB. This
2267 * can change the freelist, so keep trying until it stabilizes.
2269 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2270 * except the code below can decrease env->me_pglast to split pghead.
2271 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2272 * can only appear in env->me_pghead when env->me_pglast increases.
2273 * Until then, the me_pghead pointer won't move but can become NULL.
2276 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2277 oldpg_txnid = id = 0;
2280 /* should only be one record now */
2281 if (env->me_pghead || env->me_pglast) {
2282 /* make sure first page of freeDB is touched and on freelist */
2283 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2284 if (rc && rc != MDB_NOTFOUND) {
2291 /* Delete IDLs we used from the free list */
2292 if (env->me_pglast) {
2297 rc = mdb_cursor_first(&mc, &key, NULL);
2300 oldpg_txnid = *(txnid_t *)key.mv_data;
2302 assert(oldpg_txnid <= env->me_pglast);
2304 rc = mdb_cursor_del(&mc, 0);
2307 } while (oldpg_txnid < env->me_pglast);
2310 /* Save IDL of pages freed by this txn, to freeDB */
2312 if (freecnt != txn->mt_free_pgs[0]) {
2315 /* make sure last page of freeDB is touched and on freelist */
2316 key.mv_size = MDB_MAXKEYSIZE+1;
2318 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2319 if (rc && rc != MDB_NOTFOUND)
2325 MDB_IDL idl = txn->mt_free_pgs;
2326 mdb_midl_sort(txn->mt_free_pgs);
2327 DPRINTF("IDL write txn %zu root %zu num %zu",
2328 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2329 for (i=1; i<=idl[0]; i++) {
2330 DPRINTF("IDL %zu", idl[i]);
2334 /* write to last page of freeDB */
2335 key.mv_size = sizeof(pgno_t);
2336 key.mv_data = &txn->mt_txnid;
2337 /* The free list can still grow during this call,
2338 * despite the pre-emptive touches above. So retry
2339 * until the reserved space remains big enough.
2342 assert(freecnt < txn->mt_free_pgs[0]);
2343 freecnt = txn->mt_free_pgs[0];
2344 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2345 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2348 } while (freecnt != txn->mt_free_pgs[0]);
2349 mdb_midl_sort(txn->mt_free_pgs);
2350 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2351 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2352 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2355 /* Put back page numbers we took from freeDB but did not use */
2356 if (env->me_pghead) {
2361 mop = env->me_pghead;
2362 id = env->me_pglast;
2363 key.mv_size = sizeof(id);
2365 /* These steps may grow the freelist again
2366 * due to freed overflow pages...
2371 if (orig > env->me_maxfree_1pg && id > 4)
2372 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2373 data.mv_size = (orig + 1) * sizeof(pgno_t);
2374 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2377 assert(!env->me_pghead || env->me_pglast);
2378 /* mop could have been used again here */
2379 if (id != env->me_pglast || env->me_pghead == NULL)
2380 goto again; /* was completely used up */
2381 assert(mop == env->me_pghead);
2382 } while (mop[0] < orig && --i);
2383 memcpy(data.mv_data, mop, data.mv_size);
2386 *(pgno_t *)data.mv_data = orig;
2387 mop[orig] = mop[0] - orig;
2388 env->me_pghead = mop += orig;
2389 /* Save more oldpages at the previous txnid. */
2390 assert(env->me_pglast == id && id == oldpg_txnid);
2391 env->me_pglast = --oldpg_txnid;
2395 /* Check for growth of freelist again */
2396 if (freecnt != txn->mt_free_pgs[0])
2399 free(env->me_pgfree);
2400 env->me_pghead = env->me_pgfree = NULL;
2402 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2403 if (mdb_midl_shrink(&txn->mt_free_pgs))
2404 env->me_free_pgs = txn->mt_free_pgs;
2411 if (env->me_flags & MDB_WRITEMAP) {
2412 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2413 dp = txn->mt_u.dirty_list[i].mptr;
2414 /* clear dirty flag */
2415 dp->mp_flags &= ~P_DIRTY;
2417 txn->mt_u.dirty_list[0].mid = 0;
2421 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2427 /* Windows actually supports scatter/gather I/O, but only on
2428 * unbuffered file handles. Since we're relying on the OS page
2429 * cache for all our data, that's self-defeating. So we just
2430 * write pages one at a time. We use the ov structure to set
2431 * the write offset, to at least save the overhead of a Seek
2435 memset(&ov, 0, sizeof(ov));
2436 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2438 dp = txn->mt_u.dirty_list[i].mptr;
2439 DPRINTF("committing page %zu", dp->mp_pgno);
2440 size = dp->mp_pgno * env->me_psize;
2441 ov.Offset = size & 0xffffffff;
2442 ov.OffsetHigh = size >> 16;
2443 ov.OffsetHigh >>= 16;
2444 /* clear dirty flag */
2445 dp->mp_flags &= ~P_DIRTY;
2446 wsize = env->me_psize;
2447 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2448 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2451 DPRINTF("WriteFile: %d", n);
2458 struct iovec iov[MDB_COMMIT_PAGES];
2462 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2463 dp = txn->mt_u.dirty_list[i].mptr;
2464 if (dp->mp_pgno != next) {
2466 rc = writev(env->me_fd, iov, n);
2470 DPUTS("short write, filesystem full?");
2472 DPRINTF("writev: %s", strerror(n));
2479 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2482 DPRINTF("committing page %zu", dp->mp_pgno);
2483 iov[n].iov_len = env->me_psize;
2484 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2485 iov[n].iov_base = (char *)dp;
2486 size += iov[n].iov_len;
2487 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2488 /* clear dirty flag */
2489 dp->mp_flags &= ~P_DIRTY;
2490 if (++n >= MDB_COMMIT_PAGES) {
2500 rc = writev(env->me_fd, iov, n);
2504 DPUTS("short write, filesystem full?");
2506 DPRINTF("writev: %s", strerror(n));
2513 mdb_dlist_free(txn);
2516 if ((n = mdb_env_sync(env, 0)) != 0 ||
2517 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2525 mdb_dbis_update(txn, 1);
2527 UNLOCK_MUTEX_W(env);
2533 /** Read the environment parameters of a DB environment before
2534 * mapping it into memory.
2535 * @param[in] env the environment handle
2536 * @param[out] meta address of where to store the meta information
2537 * @return 0 on success, non-zero on failure.
2540 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2547 /* We don't know the page size yet, so use a minimum value.
2548 * Read both meta pages so we can use the latest one.
2551 for (i=0; i<2; i++) {
2553 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2555 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2560 else if (rc != MDB_PAGESIZE) {
2564 DPRINTF("read: %s", strerror(err));
2568 p = (MDB_page *)&pbuf;
2570 if (!F_ISSET(p->mp_flags, P_META)) {
2571 DPRINTF("page %zu not a meta page", p->mp_pgno);
2576 if (m->mm_magic != MDB_MAGIC) {
2577 DPUTS("meta has invalid magic");
2581 if (m->mm_version != MDB_VERSION) {
2582 DPRINTF("database is version %u, expected version %u",
2583 m->mm_version, MDB_VERSION);
2584 return MDB_VERSION_MISMATCH;
2588 if (m->mm_txnid > meta->mm_txnid)
2589 memcpy(meta, m, sizeof(*m));
2591 memcpy(meta, m, sizeof(*m));
2593 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2595 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2603 /** Write the environment parameters of a freshly created DB environment.
2604 * @param[in] env the environment handle
2605 * @param[out] meta address of where to store the meta information
2606 * @return 0 on success, non-zero on failure.
2609 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2616 DPUTS("writing new meta page");
2618 GET_PAGESIZE(psize);
2620 meta->mm_magic = MDB_MAGIC;
2621 meta->mm_version = MDB_VERSION;
2622 meta->mm_mapsize = env->me_mapsize;
2623 meta->mm_psize = psize;
2624 meta->mm_last_pg = 1;
2625 meta->mm_flags = env->me_flags & 0xffff;
2626 meta->mm_flags |= MDB_INTEGERKEY;
2627 meta->mm_dbs[0].md_root = P_INVALID;
2628 meta->mm_dbs[1].md_root = P_INVALID;
2630 p = calloc(2, psize);
2632 p->mp_flags = P_META;
2635 memcpy(m, meta, sizeof(*meta));
2637 q = (MDB_page *)((char *)p + psize);
2640 q->mp_flags = P_META;
2643 memcpy(m, meta, sizeof(*meta));
2648 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2649 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2650 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2653 lseek(env->me_fd, 0, SEEK_SET);
2654 rc = write(env->me_fd, p, psize * 2);
2655 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2661 /** Update the environment info to commit a transaction.
2662 * @param[in] txn the transaction that's being committed
2663 * @return 0 on success, non-zero on failure.
2666 mdb_env_write_meta(MDB_txn *txn)
2669 MDB_meta meta, metab, *mp;
2671 int rc, len, toggle;
2678 assert(txn != NULL);
2679 assert(txn->mt_env != NULL);
2681 toggle = !txn->mt_toggle;
2682 DPRINTF("writing meta page %d for root page %zu",
2683 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2686 mp = env->me_metas[toggle];
2688 if (env->me_flags & MDB_WRITEMAP) {
2689 /* Persist any increases of mapsize config */
2690 if (env->me_mapsize > mp->mm_mapsize)
2691 mp->mm_mapsize = env->me_mapsize;
2692 mp->mm_dbs[0] = txn->mt_dbs[0];
2693 mp->mm_dbs[1] = txn->mt_dbs[1];
2694 mp->mm_last_pg = txn->mt_next_pgno - 1;
2695 mp->mm_txnid = txn->mt_txnid;
2696 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2697 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2700 ptr += env->me_psize;
2701 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2708 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2709 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2711 ptr = (char *)&meta;
2712 if (env->me_mapsize > mp->mm_mapsize) {
2713 /* Persist any increases of mapsize config */
2714 meta.mm_mapsize = env->me_mapsize;
2715 off = offsetof(MDB_meta, mm_mapsize);
2717 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2719 len = sizeof(MDB_meta) - off;
2722 meta.mm_dbs[0] = txn->mt_dbs[0];
2723 meta.mm_dbs[1] = txn->mt_dbs[1];
2724 meta.mm_last_pg = txn->mt_next_pgno - 1;
2725 meta.mm_txnid = txn->mt_txnid;
2728 off += env->me_psize;
2731 /* Write to the SYNC fd */
2732 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2733 env->me_fd : env->me_mfd;
2736 memset(&ov, 0, sizeof(ov));
2738 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2741 rc = pwrite(mfd, ptr, len, off);
2746 DPUTS("write failed, disk error?");
2747 /* On a failure, the pagecache still contains the new data.
2748 * Write some old data back, to prevent it from being used.
2749 * Use the non-SYNC fd; we know it will fail anyway.
2751 meta.mm_last_pg = metab.mm_last_pg;
2752 meta.mm_txnid = metab.mm_txnid;
2754 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2756 r2 = pwrite(env->me_fd, ptr, len, off);
2759 env->me_flags |= MDB_FATAL_ERROR;
2763 /* Memory ordering issues are irrelevant; since the entire writer
2764 * is wrapped by wmutex, all of these changes will become visible
2765 * after the wmutex is unlocked. Since the DB is multi-version,
2766 * readers will get consistent data regardless of how fresh or
2767 * how stale their view of these values is.
2769 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2774 /** Check both meta pages to see which one is newer.
2775 * @param[in] env the environment handle
2776 * @return meta toggle (0 or 1).
2779 mdb_env_pick_meta(const MDB_env *env)
2781 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2785 mdb_env_create(MDB_env **env)
2789 e = calloc(1, sizeof(MDB_env));
2793 e->me_maxreaders = DEFAULT_READERS;
2794 e->me_maxdbs = e->me_numdbs = 2;
2795 e->me_fd = INVALID_HANDLE_VALUE;
2796 e->me_lfd = INVALID_HANDLE_VALUE;
2797 e->me_mfd = INVALID_HANDLE_VALUE;
2798 #ifdef MDB_USE_POSIX_SEM
2799 e->me_rmutex = SEM_FAILED;
2800 e->me_wmutex = SEM_FAILED;
2802 e->me_pid = getpid();
2803 VGMEMP_CREATE(e,0,0);
2809 mdb_env_set_mapsize(MDB_env *env, size_t size)
2813 env->me_mapsize = size;
2815 env->me_maxpg = env->me_mapsize / env->me_psize;
2820 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2824 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2829 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2831 if (env->me_map || readers < 1)
2833 env->me_maxreaders = readers;
2838 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2840 if (!env || !readers)
2842 *readers = env->me_maxreaders;
2846 /** Further setup required for opening an MDB environment
2849 mdb_env_open2(MDB_env *env)
2851 unsigned int flags = env->me_flags;
2852 int i, newenv = 0, prot;
2856 memset(&meta, 0, sizeof(meta));
2858 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2861 DPUTS("new mdbenv");
2865 /* Was a mapsize configured? */
2866 if (!env->me_mapsize) {
2867 /* If this is a new environment, take the default,
2868 * else use the size recorded in the existing env.
2870 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2871 } else if (env->me_mapsize < meta.mm_mapsize) {
2872 /* If the configured size is smaller, make sure it's
2873 * still big enough. Silently round up to minimum if not.
2875 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2876 if (env->me_mapsize < minsize)
2877 env->me_mapsize = minsize;
2883 LONG sizelo, sizehi;
2884 sizelo = env->me_mapsize & 0xffffffff;
2885 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2887 /* Windows won't create mappings for zero length files.
2888 * Just allocate the maxsize right now.
2891 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2892 if (!SetEndOfFile(env->me_fd))
2894 SetFilePointer(env->me_fd, 0, NULL, 0);
2896 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2897 PAGE_READWRITE : PAGE_READONLY,
2898 sizehi, sizelo, NULL);
2901 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2902 FILE_MAP_WRITE : FILE_MAP_READ,
2903 0, 0, env->me_mapsize, meta.mm_address);
2911 if (flags & MDB_WRITEMAP) {
2913 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2916 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2918 if (env->me_map == MAP_FAILED) {
2922 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2924 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2926 #ifdef POSIX_MADV_RANDOM
2927 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2928 #endif /* POSIX_MADV_RANDOM */
2929 #endif /* MADV_RANDOM */
2933 if (flags & MDB_FIXEDMAP)
2934 meta.mm_address = env->me_map;
2935 i = mdb_env_init_meta(env, &meta);
2936 if (i != MDB_SUCCESS) {
2939 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2940 /* Can happen because the address argument to mmap() is just a
2941 * hint. mmap() can pick another, e.g. if the range is in use.
2942 * The MAP_FIXED flag would prevent that, but then mmap could
2943 * instead unmap existing pages to make room for the new map.
2945 return EBUSY; /* TODO: Make a new MDB_* error code? */
2947 env->me_psize = meta.mm_psize;
2948 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2949 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2951 env->me_maxpg = env->me_mapsize / env->me_psize;
2953 p = (MDB_page *)env->me_map;
2954 env->me_metas[0] = METADATA(p);
2955 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2959 int toggle = mdb_env_pick_meta(env);
2960 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2962 DPRINTF("opened database version %u, pagesize %u",
2963 env->me_metas[0]->mm_version, env->me_psize);
2964 DPRINTF("using meta page %d", toggle);
2965 DPRINTF("depth: %u", db->md_depth);
2966 DPRINTF("entries: %zu", db->md_entries);
2967 DPRINTF("branch pages: %zu", db->md_branch_pages);
2968 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2969 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2970 DPRINTF("root: %zu", db->md_root);
2978 /** Release a reader thread's slot in the reader lock table.
2979 * This function is called automatically when a thread exits.
2980 * @param[in] ptr This points to the slot in the reader lock table.
2983 mdb_env_reader_dest(void *ptr)
2985 MDB_reader *reader = ptr;
2991 /** Junk for arranging thread-specific callbacks on Windows. This is
2992 * necessarily platform and compiler-specific. Windows supports up
2993 * to 1088 keys. Let's assume nobody opens more than 64 environments
2994 * in a single process, for now. They can override this if needed.
2996 #ifndef MAX_TLS_KEYS
2997 #define MAX_TLS_KEYS 64
2999 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3000 static int mdb_tls_nkeys;
3002 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3006 case DLL_PROCESS_ATTACH: break;
3007 case DLL_THREAD_ATTACH: break;
3008 case DLL_THREAD_DETACH:
3009 for (i=0; i<mdb_tls_nkeys; i++) {
3010 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3011 mdb_env_reader_dest(r);
3014 case DLL_PROCESS_DETACH: break;
3019 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3021 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3025 /* Force some symbol references.
3026 * _tls_used forces the linker to create the TLS directory if not already done
3027 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3029 #pragma comment(linker, "/INCLUDE:_tls_used")
3030 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3031 #pragma const_seg(".CRT$XLB")
3032 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3033 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3036 #pragma comment(linker, "/INCLUDE:__tls_used")
3037 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3038 #pragma data_seg(".CRT$XLB")
3039 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3041 #endif /* WIN 32/64 */
3042 #endif /* !__GNUC__ */
3045 /** Downgrade the exclusive lock on the region back to shared */
3047 mdb_env_share_locks(MDB_env *env, int *excl)
3049 int rc = 0, toggle = mdb_env_pick_meta(env);
3051 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3056 /* First acquire a shared lock. The Unlock will
3057 * then release the existing exclusive lock.
3059 memset(&ov, 0, sizeof(ov));
3060 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3063 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3069 struct flock lock_info;
3070 /* The shared lock replaces the existing lock */
3071 memset((void *)&lock_info, 0, sizeof(lock_info));
3072 lock_info.l_type = F_RDLCK;
3073 lock_info.l_whence = SEEK_SET;
3074 lock_info.l_start = 0;
3075 lock_info.l_len = 1;
3076 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3077 (rc = ErrCode()) == EINTR) ;
3078 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3085 /** Try to get exlusive lock, otherwise shared.
3086 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3089 mdb_env_excl_lock(MDB_env *env, int *excl)
3093 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3097 memset(&ov, 0, sizeof(ov));
3098 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3105 struct flock lock_info;
3106 memset((void *)&lock_info, 0, sizeof(lock_info));
3107 lock_info.l_type = F_WRLCK;
3108 lock_info.l_whence = SEEK_SET;
3109 lock_info.l_start = 0;
3110 lock_info.l_len = 1;
3111 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3112 (rc = ErrCode()) == EINTR) ;
3116 # ifdef MDB_USE_POSIX_SEM
3117 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3120 lock_info.l_type = F_RDLCK;
3121 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3122 (rc = ErrCode()) == EINTR) ;
3130 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3132 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3134 * @(#) $Revision: 5.1 $
3135 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3136 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3138 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3142 * Please do not copyright this code. This code is in the public domain.
3144 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3145 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3146 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3147 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3148 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3149 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3150 * PERFORMANCE OF THIS SOFTWARE.
3153 * chongo <Landon Curt Noll> /\oo/\
3154 * http://www.isthe.com/chongo/
3156 * Share and Enjoy! :-)
3159 typedef unsigned long long mdb_hash_t;
3160 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3162 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3163 * @param[in] str string to hash
3164 * @param[in] hval initial value for hash
3165 * @return 64 bit hash
3167 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3168 * hval arg on the first call.
3171 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3173 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3174 unsigned char *end = s + val->mv_size;
3176 * FNV-1a hash each octet of the string
3179 /* xor the bottom with the current octet */
3180 hval ^= (mdb_hash_t)*s++;
3182 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3183 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3184 (hval << 7) + (hval << 8) + (hval << 40);
3186 /* return our new hash value */
3190 /** Hash the string and output the hash in hex.
3191 * @param[in] str string to hash
3192 * @param[out] hexbuf an array of 17 chars to hold the hash
3195 mdb_hash_hex(MDB_val *val, char *hexbuf)
3198 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3199 for (i=0; i<8; i++) {
3200 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3206 /** Open and/or initialize the lock region for the environment.
3207 * @param[in] env The MDB environment.
3208 * @param[in] lpath The pathname of the file used for the lock region.
3209 * @param[in] mode The Unix permissions for the file, if we create it.
3210 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3211 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3212 * @return 0 on success, non-zero on failure.
3215 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3218 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3220 # define MDB_ERRCODE_ROFS EROFS
3221 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3222 # define MDB_CLOEXEC O_CLOEXEC
3225 # define MDB_CLOEXEC 0
3232 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3233 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3234 FILE_ATTRIBUTE_NORMAL, NULL);
3236 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3238 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3240 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3245 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3246 /* Lose record locks when exec*() */
3247 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3248 fcntl(env->me_lfd, F_SETFD, fdflags);
3251 if (!(env->me_flags & MDB_NOTLS)) {
3252 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3255 env->me_flags |= MDB_ENV_TXKEY;
3257 /* Windows TLS callbacks need help finding their TLS info. */
3258 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3262 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3266 /* Try to get exclusive lock. If we succeed, then
3267 * nobody is using the lock region and we should initialize it.
3269 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3272 size = GetFileSize(env->me_lfd, NULL);
3274 size = lseek(env->me_lfd, 0, SEEK_END);
3276 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3277 if (size < rsize && *excl > 0) {
3279 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3280 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3282 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3286 size = rsize - sizeof(MDB_txninfo);
3287 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3292 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3294 if (!mh) goto fail_errno;
3295 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3297 if (!env->me_txns) goto fail_errno;
3299 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3301 if (m == MAP_FAILED) goto fail_errno;
3307 BY_HANDLE_FILE_INFORMATION stbuf;
3316 if (!mdb_sec_inited) {
3317 InitializeSecurityDescriptor(&mdb_null_sd,
3318 SECURITY_DESCRIPTOR_REVISION);
3319 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3320 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3321 mdb_all_sa.bInheritHandle = FALSE;
3322 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3325 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3326 idbuf.volume = stbuf.dwVolumeSerialNumber;
3327 idbuf.nhigh = stbuf.nFileIndexHigh;
3328 idbuf.nlow = stbuf.nFileIndexLow;
3329 val.mv_data = &idbuf;
3330 val.mv_size = sizeof(idbuf);
3331 mdb_hash_hex(&val, hexbuf);
3332 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3333 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3334 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3335 if (!env->me_rmutex) goto fail_errno;
3336 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3337 if (!env->me_wmutex) goto fail_errno;
3338 #elif defined(MDB_USE_POSIX_SEM)
3347 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3348 idbuf.dev = stbuf.st_dev;
3349 idbuf.ino = stbuf.st_ino;
3350 val.mv_data = &idbuf;
3351 val.mv_size = sizeof(idbuf);
3352 mdb_hash_hex(&val, hexbuf);
3353 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3354 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3355 /* Clean up after a previous run, if needed: Try to
3356 * remove both semaphores before doing anything else.
3358 sem_unlink(env->me_txns->mti_rmname);
3359 sem_unlink(env->me_txns->mti_wmname);
3360 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3361 O_CREAT|O_EXCL, mode, 1);
3362 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3363 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3364 O_CREAT|O_EXCL, mode, 1);
3365 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3366 #else /* MDB_USE_POSIX_SEM */
3367 pthread_mutexattr_t mattr;
3369 if ((rc = pthread_mutexattr_init(&mattr))
3370 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3371 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3372 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3374 pthread_mutexattr_destroy(&mattr);
3375 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3377 env->me_txns->mti_version = MDB_VERSION;
3378 env->me_txns->mti_magic = MDB_MAGIC;
3379 env->me_txns->mti_txnid = 0;
3380 env->me_txns->mti_numreaders = 0;
3383 if (env->me_txns->mti_magic != MDB_MAGIC) {
3384 DPUTS("lock region has invalid magic");
3388 if (env->me_txns->mti_version != MDB_VERSION) {
3389 DPRINTF("lock region is version %u, expected version %u",
3390 env->me_txns->mti_version, MDB_VERSION);
3391 rc = MDB_VERSION_MISMATCH;
3395 if (rc != EACCES && rc != EAGAIN) {
3399 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3400 if (!env->me_rmutex) goto fail_errno;
3401 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3402 if (!env->me_wmutex) goto fail_errno;
3403 #elif defined(MDB_USE_POSIX_SEM)
3404 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3405 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3406 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3407 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3418 /** The name of the lock file in the DB environment */
3419 #define LOCKNAME "/lock.mdb"
3420 /** The name of the data file in the DB environment */
3421 #define DATANAME "/data.mdb"
3422 /** The suffix of the lock file when no subdir is used */
3423 #define LOCKSUFF "-lock"
3424 /** Only a subset of the @ref mdb_env flags can be changed
3425 * at runtime. Changing other flags requires closing the
3426 * environment and re-opening it with the new flags.
3428 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3429 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3432 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3434 int oflags, rc, len, excl = -1;
3435 char *lpath, *dpath;
3437 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3441 if (flags & MDB_NOSUBDIR) {
3442 rc = len + sizeof(LOCKSUFF) + len + 1;
3444 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3449 if (flags & MDB_NOSUBDIR) {
3450 dpath = lpath + len + sizeof(LOCKSUFF);
3451 sprintf(lpath, "%s" LOCKSUFF, path);
3452 strcpy(dpath, path);
3454 dpath = lpath + len + sizeof(LOCKNAME);
3455 sprintf(lpath, "%s" LOCKNAME, path);
3456 sprintf(dpath, "%s" DATANAME, path);
3460 flags |= env->me_flags;
3461 if (flags & MDB_RDONLY) {
3462 /* silently ignore WRITEMAP when we're only getting read access */
3463 flags &= ~MDB_WRITEMAP;
3465 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3466 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3469 env->me_flags = flags |= MDB_ENV_ACTIVE;
3473 env->me_path = strdup(path);
3474 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3475 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3476 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3481 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3486 if (F_ISSET(flags, MDB_RDONLY)) {
3487 oflags = GENERIC_READ;
3488 len = OPEN_EXISTING;
3490 oflags = GENERIC_READ|GENERIC_WRITE;
3493 mode = FILE_ATTRIBUTE_NORMAL;
3494 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3495 NULL, len, mode, NULL);
3497 if (F_ISSET(flags, MDB_RDONLY))
3500 oflags = O_RDWR | O_CREAT;
3502 env->me_fd = open(dpath, oflags, mode);
3504 if (env->me_fd == INVALID_HANDLE_VALUE) {
3509 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3510 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3511 env->me_mfd = env->me_fd;
3513 /* Synchronous fd for meta writes. Needed even with
3514 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3517 env->me_mfd = CreateFile(dpath, oflags,
3518 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3519 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3521 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3523 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3528 DPRINTF("opened dbenv %p", (void *) env);
3530 rc = mdb_env_share_locks(env, &excl);
3536 mdb_env_close0(env, excl);
3542 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3544 mdb_env_close0(MDB_env *env, int excl)
3548 if (!(env->me_flags & MDB_ENV_ACTIVE))
3551 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3552 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3553 free(env->me_dbxs[i].md_name.mv_data);
3555 free(env->me_dbflags);
3558 free(env->me_dirty_list);
3559 if (env->me_free_pgs)
3560 mdb_midl_free(env->me_free_pgs);
3562 if (env->me_flags & MDB_ENV_TXKEY) {
3563 pthread_key_delete(env->me_txkey);
3565 /* Delete our key from the global list */
3566 for (i=0; i<mdb_tls_nkeys; i++)
3567 if (mdb_tls_keys[i] == env->me_txkey) {
3568 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3576 munmap(env->me_map, env->me_mapsize);
3578 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3580 if (env->me_fd != INVALID_HANDLE_VALUE)
3583 pid_t pid = env->me_pid;
3584 /* Clearing readers is done in this function because
3585 * me_txkey with its destructor must be disabled first.
3587 for (i = env->me_numreaders; --i >= 0; )
3588 if (env->me_txns->mti_readers[i].mr_pid == pid)
3589 env->me_txns->mti_readers[i].mr_pid = 0;
3591 if (env->me_rmutex) {
3592 CloseHandle(env->me_rmutex);
3593 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3595 /* Windows automatically destroys the mutexes when
3596 * the last handle closes.
3598 #elif defined(MDB_USE_POSIX_SEM)
3599 if (env->me_rmutex != SEM_FAILED) {
3600 sem_close(env->me_rmutex);
3601 if (env->me_wmutex != SEM_FAILED)
3602 sem_close(env->me_wmutex);
3603 /* If we have the filelock: If we are the
3604 * only remaining user, clean up semaphores.
3607 mdb_env_excl_lock(env, &excl);
3609 sem_unlink(env->me_txns->mti_rmname);
3610 sem_unlink(env->me_txns->mti_wmname);
3614 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3616 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3619 /* Unlock the lockfile. Windows would have unlocked it
3620 * after closing anyway, but not necessarily at once.
3622 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3628 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3632 mdb_env_copy(MDB_env *env, const char *path)
3634 MDB_txn *txn = NULL;
3638 HANDLE newfd = INVALID_HANDLE_VALUE;
3640 if (env->me_flags & MDB_NOSUBDIR) {
3641 lpath = (char *)path;
3644 len += sizeof(DATANAME);
3645 lpath = malloc(len);
3648 sprintf(lpath, "%s" DATANAME, path);
3651 /* The destination path must exist, but the destination file must not.
3652 * We don't want the OS to cache the writes, since the source data is
3653 * already in the OS cache.
3656 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3657 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3659 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3665 if (!(env->me_flags & MDB_NOSUBDIR))
3667 if (newfd == INVALID_HANDLE_VALUE) {
3672 #ifdef F_NOCACHE /* __APPLE__ */
3673 rc = fcntl(newfd, F_NOCACHE, 1);
3680 /* Do the lock/unlock of the reader mutex before starting the
3681 * write txn. Otherwise other read txns could block writers.
3683 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3688 /* We must start the actual read txn after blocking writers */
3689 mdb_txn_reset0(txn);
3691 /* Temporarily block writers until we snapshot the meta pages */
3694 rc = mdb_txn_renew0(txn);
3696 UNLOCK_MUTEX_W(env);
3701 wsize = env->me_psize * 2;
3705 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3706 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3709 rc = write(newfd, env->me_map, wsize);
3710 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3713 UNLOCK_MUTEX_W(env);
3718 ptr = env->me_map + wsize;
3719 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3720 #define MAX_WRITE 2147483648U
3724 if (wsize > MAX_WRITE)
3728 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3729 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3738 if (wsize > MAX_WRITE)
3742 wres = write(newfd, ptr, w2);
3743 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3752 if (newfd != INVALID_HANDLE_VALUE)
3759 mdb_env_close(MDB_env *env)
3766 VGMEMP_DESTROY(env);
3767 while ((dp = env->me_dpages) != NULL) {
3768 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3769 env->me_dpages = dp->mp_next;
3773 mdb_env_close0(env, 0);
3777 /** Compare two items pointing at aligned size_t's */
3779 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3781 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3782 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3785 /** Compare two items pointing at aligned int's */
3787 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3789 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3790 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3793 /** Compare two items pointing at ints of unknown alignment.
3794 * Nodes and keys are guaranteed to be 2-byte aligned.
3797 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3799 #if BYTE_ORDER == LITTLE_ENDIAN
3800 unsigned short *u, *c;
3803 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3804 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3807 } while(!x && u > (unsigned short *)a->mv_data);
3810 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3814 /** Compare two items lexically */
3816 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3823 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3829 diff = memcmp(a->mv_data, b->mv_data, len);
3830 return diff ? diff : len_diff<0 ? -1 : len_diff;
3833 /** Compare two items in reverse byte order */
3835 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3837 const unsigned char *p1, *p2, *p1_lim;
3841 p1_lim = (const unsigned char *)a->mv_data;
3842 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3843 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3845 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3851 while (p1 > p1_lim) {
3852 diff = *--p1 - *--p2;
3856 return len_diff<0 ? -1 : len_diff;
3859 /** Search for key within a page, using binary search.
3860 * Returns the smallest entry larger or equal to the key.
3861 * If exactp is non-null, stores whether the found entry was an exact match
3862 * in *exactp (1 or 0).
3863 * Updates the cursor index with the index of the found entry.
3864 * If no entry larger or equal to the key is found, returns NULL.
3867 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3869 unsigned int i = 0, nkeys;
3872 MDB_page *mp = mc->mc_pg[mc->mc_top];
3873 MDB_node *node = NULL;
3878 nkeys = NUMKEYS(mp);
3883 COPY_PGNO(pgno, mp->mp_pgno);
3884 DPRINTF("searching %u keys in %s %spage %zu",
3885 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3892 low = IS_LEAF(mp) ? 0 : 1;
3894 cmp = mc->mc_dbx->md_cmp;
3896 /* Branch pages have no data, so if using integer keys,
3897 * alignment is guaranteed. Use faster mdb_cmp_int.
3899 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3900 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3907 nodekey.mv_size = mc->mc_db->md_pad;
3908 node = NODEPTR(mp, 0); /* fake */
3909 while (low <= high) {
3910 i = (low + high) >> 1;
3911 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3912 rc = cmp(key, &nodekey);
3913 DPRINTF("found leaf index %u [%s], rc = %i",
3914 i, DKEY(&nodekey), rc);
3923 while (low <= high) {
3924 i = (low + high) >> 1;
3926 node = NODEPTR(mp, i);
3927 nodekey.mv_size = NODEKSZ(node);
3928 nodekey.mv_data = NODEKEY(node);
3930 rc = cmp(key, &nodekey);
3933 DPRINTF("found leaf index %u [%s], rc = %i",
3934 i, DKEY(&nodekey), rc);
3936 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3937 i, DKEY(&nodekey), NODEPGNO(node), rc);
3948 if (rc > 0) { /* Found entry is less than the key. */
3949 i++; /* Skip to get the smallest entry larger than key. */
3951 node = NODEPTR(mp, i);
3954 *exactp = (rc == 0);
3955 /* store the key index */
3956 mc->mc_ki[mc->mc_top] = i;
3958 /* There is no entry larger or equal to the key. */
3961 /* nodeptr is fake for LEAF2 */
3967 mdb_cursor_adjust(MDB_cursor *mc, func)
3971 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3972 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3979 /** Pop a page off the top of the cursor's stack. */
3981 mdb_cursor_pop(MDB_cursor *mc)
3984 #ifndef MDB_DEBUG_SKIP
3985 MDB_page *top = mc->mc_pg[mc->mc_top];
3991 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3992 mc->mc_dbi, (void *) mc);
3996 /** Push a page onto the top of the cursor's stack. */
3998 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4000 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4001 mc->mc_dbi, (void *) mc);
4003 if (mc->mc_snum >= CURSOR_STACK) {
4004 assert(mc->mc_snum < CURSOR_STACK);
4005 return MDB_CURSOR_FULL;
4008 mc->mc_top = mc->mc_snum++;
4009 mc->mc_pg[mc->mc_top] = mp;
4010 mc->mc_ki[mc->mc_top] = 0;
4015 /** Find the address of the page corresponding to a given page number.
4016 * @param[in] txn the transaction for this access.
4017 * @param[in] pgno the page number for the page to retrieve.
4018 * @param[out] ret address of a pointer where the page's address will be stored.
4019 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4020 * @return 0 on success, non-zero on failure.
4023 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4028 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4029 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4034 MDB_ID2L dl = tx2->mt_u.dirty_list;
4036 unsigned x = mdb_mid2l_search(dl, pgno);
4037 if (x <= dl[0].mid && dl[x].mid == pgno) {
4043 } while ((tx2 = tx2->mt_parent) != NULL);
4046 if (pgno < txn->mt_next_pgno) {
4048 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4050 DPRINTF("page %zu not found", pgno);
4052 return MDB_PAGE_NOTFOUND;
4062 /** Search for the page a given key should be in.
4063 * Pushes parent pages on the cursor stack. This function continues a
4064 * search on a cursor that has already been initialized. (Usually by
4065 * #mdb_page_search() but also by #mdb_node_move().)
4066 * @param[in,out] mc the cursor for this operation.
4067 * @param[in] key the key to search for. If NULL, search for the lowest
4068 * page. (This is used by #mdb_cursor_first().)
4069 * @param[in] modify If true, visited pages are updated with new page numbers.
4070 * @return 0 on success, non-zero on failure.
4073 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4075 MDB_page *mp = mc->mc_pg[mc->mc_top];
4080 while (IS_BRANCH(mp)) {
4084 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4085 assert(NUMKEYS(mp) > 1);
4086 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4088 if (key == NULL) /* Initialize cursor to first page. */
4090 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4091 /* cursor to last page */
4095 node = mdb_node_search(mc, key, &exact);
4097 i = NUMKEYS(mp) - 1;
4099 i = mc->mc_ki[mc->mc_top];
4108 DPRINTF("following index %u for key [%s]",
4110 assert(i < NUMKEYS(mp));
4111 node = NODEPTR(mp, i);
4113 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4116 mc->mc_ki[mc->mc_top] = i;
4117 if ((rc = mdb_cursor_push(mc, mp)))
4121 if ((rc = mdb_page_touch(mc)) != 0)
4123 mp = mc->mc_pg[mc->mc_top];
4128 DPRINTF("internal error, index points to a %02X page!?",
4130 return MDB_CORRUPTED;
4133 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4134 key ? DKEY(key) : NULL);
4139 /** Search for the lowest key under the current branch page.
4140 * This just bypasses a NUMKEYS check in the current page
4141 * before calling mdb_page_search_root(), because the callers
4142 * are all in situations where the current page is known to
4146 mdb_page_search_lowest(MDB_cursor *mc)
4148 MDB_page *mp = mc->mc_pg[mc->mc_top];
4149 MDB_node *node = NODEPTR(mp, 0);
4152 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4155 mc->mc_ki[mc->mc_top] = 0;
4156 if ((rc = mdb_cursor_push(mc, mp)))
4158 return mdb_page_search_root(mc, NULL, 0);
4161 /** Search for the page a given key should be in.
4162 * Pushes parent pages on the cursor stack. This function just sets up
4163 * the search; it finds the root page for \b mc's database and sets this
4164 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4165 * called to complete the search.
4166 * @param[in,out] mc the cursor for this operation.
4167 * @param[in] key the key to search for. If NULL, search for the lowest
4168 * page. (This is used by #mdb_cursor_first().)
4169 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4170 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4171 * @return 0 on success, non-zero on failure.
4174 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4179 /* Make sure the txn is still viable, then find the root from
4180 * the txn's db table.
4182 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4183 DPUTS("transaction has failed, must abort");
4186 /* Make sure we're using an up-to-date root */
4187 if (mc->mc_dbi > MAIN_DBI) {
4188 if ((*mc->mc_dbflag & DB_STALE) ||
4189 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4191 unsigned char dbflag = 0;
4192 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4193 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4196 if (*mc->mc_dbflag & DB_STALE) {
4200 MDB_node *leaf = mdb_node_search(&mc2,
4201 &mc->mc_dbx->md_name, &exact);
4203 return MDB_NOTFOUND;
4204 mdb_node_read(mc->mc_txn, leaf, &data);
4205 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4207 /* The txn may not know this DBI, or another process may
4208 * have dropped and recreated the DB with other flags.
4210 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4211 return MDB_INCOMPATIBLE;
4212 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4214 if (flags & MDB_PS_MODIFY)
4216 *mc->mc_dbflag &= ~DB_STALE;
4217 *mc->mc_dbflag |= dbflag;
4220 root = mc->mc_db->md_root;
4222 if (root == P_INVALID) { /* Tree is empty. */
4223 DPUTS("tree is empty");
4224 return MDB_NOTFOUND;
4229 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4230 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4236 DPRINTF("db %u root page %zu has flags 0x%X",
4237 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4239 if (flags & MDB_PS_MODIFY) {
4240 if ((rc = mdb_page_touch(mc)))
4244 if (flags & MDB_PS_ROOTONLY)
4247 return mdb_page_search_root(mc, key, flags);
4250 /** Return the data associated with a given node.
4251 * @param[in] txn The transaction for this operation.
4252 * @param[in] leaf The node being read.
4253 * @param[out] data Updated to point to the node's data.
4254 * @return 0 on success, non-zero on failure.
4257 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4259 MDB_page *omp; /* overflow page */
4263 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4264 data->mv_size = NODEDSZ(leaf);
4265 data->mv_data = NODEDATA(leaf);
4269 /* Read overflow data.
4271 data->mv_size = NODEDSZ(leaf);
4272 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4273 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4274 DPRINTF("read overflow page %zu failed", pgno);
4277 data->mv_data = METADATA(omp);
4283 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4284 MDB_val *key, MDB_val *data)
4293 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4295 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4298 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4302 mdb_cursor_init(&mc, txn, dbi, &mx);
4303 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4306 /** Find a sibling for a page.
4307 * Replaces the page at the top of the cursor's stack with the
4308 * specified sibling, if one exists.
4309 * @param[in] mc The cursor for this operation.
4310 * @param[in] move_right Non-zero if the right sibling is requested,
4311 * otherwise the left sibling.
4312 * @return 0 on success, non-zero on failure.
4315 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4321 if (mc->mc_snum < 2) {
4322 return MDB_NOTFOUND; /* root has no siblings */
4326 DPRINTF("parent page is page %zu, index %u",
4327 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4329 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4330 : (mc->mc_ki[mc->mc_top] == 0)) {
4331 DPRINTF("no more keys left, moving to %s sibling",
4332 move_right ? "right" : "left");
4333 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4334 /* undo cursor_pop before returning */
4341 mc->mc_ki[mc->mc_top]++;
4343 mc->mc_ki[mc->mc_top]--;
4344 DPRINTF("just moving to %s index key %u",
4345 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4347 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4349 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4350 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4353 mdb_cursor_push(mc, mp);
4355 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4360 /** Move the cursor to the next data item. */
4362 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4368 if (mc->mc_flags & C_EOF) {
4369 return MDB_NOTFOUND;
4372 assert(mc->mc_flags & C_INITIALIZED);
4374 mp = mc->mc_pg[mc->mc_top];
4376 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4377 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4378 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4379 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4380 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4381 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4385 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4386 if (op == MDB_NEXT_DUP)
4387 return MDB_NOTFOUND;
4391 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4393 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4394 DPUTS("=====> move to next sibling page");
4395 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4396 mc->mc_flags |= C_EOF;
4397 mc->mc_flags &= ~C_INITIALIZED;
4398 return MDB_NOTFOUND;
4400 mp = mc->mc_pg[mc->mc_top];
4401 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4403 mc->mc_ki[mc->mc_top]++;
4405 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4406 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4409 key->mv_size = mc->mc_db->md_pad;
4410 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4414 assert(IS_LEAF(mp));
4415 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4417 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4418 mdb_xcursor_init1(mc, leaf);
4421 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4424 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4425 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4426 if (rc != MDB_SUCCESS)
4431 MDB_GET_KEY(leaf, key);
4435 /** Move the cursor to the previous data item. */
4437 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4443 assert(mc->mc_flags & C_INITIALIZED);
4445 mp = mc->mc_pg[mc->mc_top];
4447 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4448 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4449 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4450 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4451 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4452 if (op != MDB_PREV || rc == MDB_SUCCESS)
4455 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4456 if (op == MDB_PREV_DUP)
4457 return MDB_NOTFOUND;
4462 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4464 if (mc->mc_ki[mc->mc_top] == 0) {
4465 DPUTS("=====> move to prev sibling page");
4466 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4467 mc->mc_flags &= ~C_INITIALIZED;
4468 return MDB_NOTFOUND;
4470 mp = mc->mc_pg[mc->mc_top];
4471 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4472 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4474 mc->mc_ki[mc->mc_top]--;
4476 mc->mc_flags &= ~C_EOF;
4478 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4479 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4482 key->mv_size = mc->mc_db->md_pad;
4483 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4487 assert(IS_LEAF(mp));
4488 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4490 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4491 mdb_xcursor_init1(mc, leaf);
4494 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4497 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4498 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4499 if (rc != MDB_SUCCESS)
4504 MDB_GET_KEY(leaf, key);
4508 /** Set the cursor on a specific data item. */
4510 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4511 MDB_cursor_op op, int *exactp)
4515 MDB_node *leaf = NULL;
4520 assert(key->mv_size > 0);
4522 /* See if we're already on the right page */
4523 if (mc->mc_flags & C_INITIALIZED) {
4526 mp = mc->mc_pg[mc->mc_top];
4528 mc->mc_ki[mc->mc_top] = 0;
4529 return MDB_NOTFOUND;
4531 if (mp->mp_flags & P_LEAF2) {
4532 nodekey.mv_size = mc->mc_db->md_pad;
4533 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4535 leaf = NODEPTR(mp, 0);
4536 MDB_GET_KEY(leaf, &nodekey);
4538 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4540 /* Probably happens rarely, but first node on the page
4541 * was the one we wanted.
4543 mc->mc_ki[mc->mc_top] = 0;
4550 unsigned int nkeys = NUMKEYS(mp);
4552 if (mp->mp_flags & P_LEAF2) {
4553 nodekey.mv_data = LEAF2KEY(mp,
4554 nkeys-1, nodekey.mv_size);
4556 leaf = NODEPTR(mp, nkeys-1);
4557 MDB_GET_KEY(leaf, &nodekey);
4559 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4561 /* last node was the one we wanted */
4562 mc->mc_ki[mc->mc_top] = nkeys-1;
4568 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4569 /* This is definitely the right page, skip search_page */
4570 if (mp->mp_flags & P_LEAF2) {
4571 nodekey.mv_data = LEAF2KEY(mp,
4572 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4574 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4575 MDB_GET_KEY(leaf, &nodekey);
4577 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4579 /* current node was the one we wanted */
4589 /* If any parents have right-sibs, search.
4590 * Otherwise, there's nothing further.
4592 for (i=0; i<mc->mc_top; i++)
4594 NUMKEYS(mc->mc_pg[i])-1)
4596 if (i == mc->mc_top) {
4597 /* There are no other pages */
4598 mc->mc_ki[mc->mc_top] = nkeys;
4599 return MDB_NOTFOUND;
4603 /* There are no other pages */
4604 mc->mc_ki[mc->mc_top] = 0;
4605 return MDB_NOTFOUND;
4609 rc = mdb_page_search(mc, key, 0);
4610 if (rc != MDB_SUCCESS)
4613 mp = mc->mc_pg[mc->mc_top];
4614 assert(IS_LEAF(mp));
4617 leaf = mdb_node_search(mc, key, exactp);
4618 if (exactp != NULL && !*exactp) {
4619 /* MDB_SET specified and not an exact match. */
4620 return MDB_NOTFOUND;
4624 DPUTS("===> inexact leaf not found, goto sibling");
4625 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4626 return rc; /* no entries matched */
4627 mp = mc->mc_pg[mc->mc_top];
4628 assert(IS_LEAF(mp));
4629 leaf = NODEPTR(mp, 0);
4633 mc->mc_flags |= C_INITIALIZED;
4634 mc->mc_flags &= ~C_EOF;
4637 key->mv_size = mc->mc_db->md_pad;
4638 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4642 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4643 mdb_xcursor_init1(mc, leaf);
4646 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4647 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4648 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4651 if (op == MDB_GET_BOTH) {
4657 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4658 if (rc != MDB_SUCCESS)
4661 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4663 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4665 rc = mc->mc_dbx->md_dcmp(data, &d2);
4667 if (op == MDB_GET_BOTH || rc > 0)
4668 return MDB_NOTFOUND;
4673 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4674 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4679 /* The key already matches in all other cases */
4680 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4681 MDB_GET_KEY(leaf, key);
4682 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4687 /** Move the cursor to the first item in the database. */
4689 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4694 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4695 rc = mdb_page_search(mc, NULL, 0);
4696 if (rc != MDB_SUCCESS)
4699 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4701 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4702 mc->mc_flags |= C_INITIALIZED;
4703 mc->mc_flags &= ~C_EOF;
4705 mc->mc_ki[mc->mc_top] = 0;
4707 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4708 key->mv_size = mc->mc_db->md_pad;
4709 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4714 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4715 mdb_xcursor_init1(mc, leaf);
4716 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4721 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4722 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4726 MDB_GET_KEY(leaf, key);
4730 /** Move the cursor to the last item in the database. */
4732 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4737 if (!(mc->mc_flags & C_EOF)) {
4739 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4742 lkey.mv_size = MDB_MAXKEYSIZE+1;
4743 lkey.mv_data = NULL;
4744 rc = mdb_page_search(mc, &lkey, 0);
4745 if (rc != MDB_SUCCESS)
4748 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4751 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4752 mc->mc_flags |= C_INITIALIZED|C_EOF;
4753 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4755 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4756 key->mv_size = mc->mc_db->md_pad;
4757 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4762 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4763 mdb_xcursor_init1(mc, leaf);
4764 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4769 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4770 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4775 MDB_GET_KEY(leaf, key);
4780 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4789 case MDB_GET_CURRENT:
4790 if (!(mc->mc_flags & C_INITIALIZED)) {
4793 MDB_page *mp = mc->mc_pg[mc->mc_top];
4795 mc->mc_ki[mc->mc_top] = 0;
4801 key->mv_size = mc->mc_db->md_pad;
4802 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4804 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4805 MDB_GET_KEY(leaf, key);
4807 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4808 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4810 rc = mdb_node_read(mc->mc_txn, leaf, data);
4817 case MDB_GET_BOTH_RANGE:
4818 if (data == NULL || mc->mc_xcursor == NULL) {
4826 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4828 } else if (op == MDB_SET_RANGE)
4829 rc = mdb_cursor_set(mc, key, data, op, NULL);
4831 rc = mdb_cursor_set(mc, key, data, op, &exact);
4833 case MDB_GET_MULTIPLE:
4835 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4836 !(mc->mc_flags & C_INITIALIZED)) {
4841 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4842 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4845 case MDB_NEXT_MULTIPLE:
4847 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4851 if (!(mc->mc_flags & C_INITIALIZED))
4852 rc = mdb_cursor_first(mc, key, data);
4854 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4855 if (rc == MDB_SUCCESS) {
4856 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4859 mx = &mc->mc_xcursor->mx_cursor;
4860 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4862 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4863 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4871 case MDB_NEXT_NODUP:
4872 if (!(mc->mc_flags & C_INITIALIZED))
4873 rc = mdb_cursor_first(mc, key, data);
4875 rc = mdb_cursor_next(mc, key, data, op);
4879 case MDB_PREV_NODUP:
4880 if (!(mc->mc_flags & C_INITIALIZED)) {
4881 rc = mdb_cursor_last(mc, key, data);
4882 mc->mc_flags |= C_INITIALIZED;
4883 mc->mc_ki[mc->mc_top]++;
4885 rc = mdb_cursor_prev(mc, key, data, op);
4888 rc = mdb_cursor_first(mc, key, data);
4892 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4893 !(mc->mc_flags & C_INITIALIZED) ||
4894 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4898 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4901 rc = mdb_cursor_last(mc, key, data);
4905 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4906 !(mc->mc_flags & C_INITIALIZED) ||
4907 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4911 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4914 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4922 /** Touch all the pages in the cursor stack.
4923 * Makes sure all the pages are writable, before attempting a write operation.
4924 * @param[in] mc The cursor to operate on.
4927 mdb_cursor_touch(MDB_cursor *mc)
4931 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4934 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
4935 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4938 *mc->mc_dbflag |= DB_DIRTY;
4940 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4941 rc = mdb_page_touch(mc);
4945 mc->mc_top = mc->mc_snum-1;
4950 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4953 MDB_node *leaf = NULL;
4954 MDB_val xdata, *rdata, dkey;
4957 int do_sub = 0, insert = 0;
4958 unsigned int mcount = 0;
4962 char dbuf[MDB_MAXKEYSIZE+1];
4963 unsigned int nflags;
4966 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4969 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4972 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4975 #if SIZE_MAX > MAXDATASIZE
4976 if (data->mv_size > MAXDATASIZE)
4980 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4981 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4985 if (flags == MDB_CURRENT) {
4986 if (!(mc->mc_flags & C_INITIALIZED))
4989 } else if (mc->mc_db->md_root == P_INVALID) {
4991 /* new database, write a root leaf page */
4992 DPUTS("allocating new root leaf page");
4993 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4997 mdb_cursor_push(mc, np);
4998 mc->mc_db->md_root = np->mp_pgno;
4999 mc->mc_db->md_depth++;
5000 *mc->mc_dbflag |= DB_DIRTY;
5001 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5003 np->mp_flags |= P_LEAF2;
5004 mc->mc_flags |= C_INITIALIZED;
5010 if (flags & MDB_APPEND) {
5012 rc = mdb_cursor_last(mc, &k2, &d2);
5014 rc = mc->mc_dbx->md_cmp(key, &k2);
5017 mc->mc_ki[mc->mc_top]++;
5019 /* new key is <= last key */
5024 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5026 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5027 DPRINTF("duplicate key [%s]", DKEY(key));
5029 return MDB_KEYEXIST;
5031 if (rc && rc != MDB_NOTFOUND)
5035 /* Cursor is positioned, now make sure all pages are writable */
5036 rc2 = mdb_cursor_touch(mc);
5041 /* The key already exists */
5042 if (rc == MDB_SUCCESS) {
5043 /* there's only a key anyway, so this is a no-op */
5044 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5045 unsigned int ksize = mc->mc_db->md_pad;
5046 if (key->mv_size != ksize)
5048 if (flags == MDB_CURRENT) {
5049 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5050 memcpy(ptr, key->mv_data, ksize);
5055 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5058 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5059 /* Was a single item before, must convert now */
5061 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5062 /* Just overwrite the current item */
5063 if (flags == MDB_CURRENT)
5066 dkey.mv_size = NODEDSZ(leaf);
5067 dkey.mv_data = NODEDATA(leaf);
5068 #if UINT_MAX < SIZE_MAX
5069 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5070 #ifdef MISALIGNED_OK
5071 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5073 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5076 /* if data matches, ignore it */
5077 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5078 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5080 /* create a fake page for the dup items */
5081 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5082 dkey.mv_data = dbuf;
5083 fp = (MDB_page *)&pbuf;
5084 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5085 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5086 fp->mp_lower = PAGEHDRSZ;
5087 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5088 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5089 fp->mp_flags |= P_LEAF2;
5090 fp->mp_pad = data->mv_size;
5091 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5093 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5094 (dkey.mv_size & 1) + (data->mv_size & 1);
5096 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5099 xdata.mv_size = fp->mp_upper;
5104 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5105 /* See if we need to convert from fake page to subDB */
5107 unsigned int offset;
5110 fp = NODEDATA(leaf);
5111 if (flags == MDB_CURRENT) {
5113 fp->mp_flags |= P_DIRTY;
5114 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5115 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5119 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5120 offset = fp->mp_pad;
5121 if (SIZELEFT(fp) >= offset)
5123 offset *= 4; /* space for 4 more */
5125 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5127 offset += offset & 1;
5128 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5129 offset >= mc->mc_txn->mt_env->me_nodemax) {
5130 /* yes, convert it */
5132 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5133 dummy.md_pad = fp->mp_pad;
5134 dummy.md_flags = MDB_DUPFIXED;
5135 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5136 dummy.md_flags |= MDB_INTEGERKEY;
5139 dummy.md_branch_pages = 0;
5140 dummy.md_leaf_pages = 1;
5141 dummy.md_overflow_pages = 0;
5142 dummy.md_entries = NUMKEYS(fp);
5144 xdata.mv_size = sizeof(MDB_db);
5145 xdata.mv_data = &dummy;
5146 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5148 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5149 flags |= F_DUPDATA|F_SUBDATA;
5150 dummy.md_root = mp->mp_pgno;
5152 /* no, just grow it */
5154 xdata.mv_size = NODEDSZ(leaf) + offset;
5155 xdata.mv_data = &pbuf;
5156 mp = (MDB_page *)&pbuf;
5157 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5160 mp->mp_flags = fp->mp_flags | P_DIRTY;
5161 mp->mp_pad = fp->mp_pad;
5162 mp->mp_lower = fp->mp_lower;
5163 mp->mp_upper = fp->mp_upper + offset;
5165 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5167 nsize = NODEDSZ(leaf) - fp->mp_upper;
5168 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5169 for (i=0; i<NUMKEYS(fp); i++)
5170 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5172 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5176 /* data is on sub-DB, just store it */
5177 flags |= F_DUPDATA|F_SUBDATA;
5181 /* overflow page overwrites need special handling */
5182 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5185 unsigned psize = mc->mc_txn->mt_env->me_psize;
5186 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5188 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5189 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5191 ovpages = omp->mp_pages;
5193 /* Is the ov page writable and large enough? */
5194 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5195 /* yes, overwrite it. Note in this case we don't
5196 * bother to try shrinking the page if the new data
5197 * is smaller than the overflow threshold.
5200 /* It is writable only in a parent txn */
5201 size_t sz = (size_t) psize * ovpages, off;
5202 MDB_page *np = mdb_page_malloc(mc, ovpages);
5208 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5209 if (!(flags & MDB_RESERVE)) {
5210 /* Copy end of page, adjusting alignment so
5211 * compiler may copy words instead of bytes.
5213 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5214 memcpy((size_t *)((char *)np + off),
5215 (size_t *)((char *)omp + off), sz - off);
5218 memcpy(np, omp, sz); /* Copy beginning of page */
5221 SETDSZ(leaf, data->mv_size);
5222 if (F_ISSET(flags, MDB_RESERVE))
5223 data->mv_data = METADATA(omp);
5225 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5228 /* no, free ovpages */
5230 mc->mc_db->md_overflow_pages -= ovpages;
5231 for (i=0; i<ovpages; i++) {
5232 DPRINTF("freed ov page %zu", pg);
5233 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5237 } else if (NODEDSZ(leaf) == data->mv_size) {
5238 /* same size, just replace it. Note that we could
5239 * also reuse this node if the new data is smaller,
5240 * but instead we opt to shrink the node in that case.
5242 if (F_ISSET(flags, MDB_RESERVE))
5243 data->mv_data = NODEDATA(leaf);
5244 else if (data->mv_size)
5245 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5247 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5250 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5251 mc->mc_db->md_entries--;
5253 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5260 nflags = flags & NODE_ADD_FLAGS;
5261 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5262 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5263 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5264 nflags &= ~MDB_APPEND;
5266 nflags |= MDB_SPLIT_REPLACE;
5267 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5269 /* There is room already in this leaf page. */
5270 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5271 if (rc == 0 && !do_sub && insert) {
5272 /* Adjust other cursors pointing to mp */
5273 MDB_cursor *m2, *m3;
5274 MDB_dbi dbi = mc->mc_dbi;
5275 unsigned i = mc->mc_top;
5276 MDB_page *mp = mc->mc_pg[i];
5278 if (mc->mc_flags & C_SUB)
5281 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5282 if (mc->mc_flags & C_SUB)
5283 m3 = &m2->mc_xcursor->mx_cursor;
5286 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5287 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5294 if (rc != MDB_SUCCESS)
5295 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5297 /* Now store the actual data in the child DB. Note that we're
5298 * storing the user data in the keys field, so there are strict
5299 * size limits on dupdata. The actual data fields of the child
5300 * DB are all zero size.
5307 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5308 if (flags & MDB_CURRENT) {
5309 xflags = MDB_CURRENT;
5311 mdb_xcursor_init1(mc, leaf);
5312 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5314 /* converted, write the original data first */
5316 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5320 /* Adjust other cursors pointing to mp */
5322 unsigned i = mc->mc_top;
5323 MDB_page *mp = mc->mc_pg[i];
5325 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5326 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5327 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5328 mdb_xcursor_init1(m2, leaf);
5332 /* we've done our job */
5335 if (flags & MDB_APPENDDUP)
5336 xflags |= MDB_APPEND;
5337 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5338 if (flags & F_SUBDATA) {
5339 void *db = NODEDATA(leaf);
5340 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5343 /* sub-writes might have failed so check rc again.
5344 * Don't increment count if we just replaced an existing item.
5346 if (!rc && !(flags & MDB_CURRENT))
5347 mc->mc_db->md_entries++;
5348 if (flags & MDB_MULTIPLE) {
5350 if (mcount < data[1].mv_size) {
5351 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5352 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5358 /* If we succeeded and the key didn't exist before, make sure
5359 * the cursor is marked valid.
5362 mc->mc_flags |= C_INITIALIZED;
5367 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5372 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5375 if (!(mc->mc_flags & C_INITIALIZED))
5378 rc = mdb_cursor_touch(mc);
5382 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5384 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5385 if (flags != MDB_NODUPDATA) {
5386 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5387 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5389 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5390 /* If sub-DB still has entries, we're done */
5391 if (mc->mc_xcursor->mx_db.md_entries) {
5392 if (leaf->mn_flags & F_SUBDATA) {
5393 /* update subDB info */
5394 void *db = NODEDATA(leaf);
5395 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5397 /* shrink fake page */
5398 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5400 mc->mc_db->md_entries--;
5403 /* otherwise fall thru and delete the sub-DB */
5406 if (leaf->mn_flags & F_SUBDATA) {
5407 /* add all the child DB's pages to the free list */
5408 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5409 if (rc == MDB_SUCCESS) {
5410 mc->mc_db->md_entries -=
5411 mc->mc_xcursor->mx_db.md_entries;
5416 return mdb_cursor_del0(mc, leaf);
5419 /** Allocate and initialize new pages for a database.
5420 * @param[in] mc a cursor on the database being added to.
5421 * @param[in] flags flags defining what type of page is being allocated.
5422 * @param[in] num the number of pages to allocate. This is usually 1,
5423 * unless allocating overflow pages for a large record.
5424 * @param[out] mp Address of a page, or NULL on failure.
5425 * @return 0 on success, non-zero on failure.
5428 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5433 if ((rc = mdb_page_alloc(mc, num, &np)))
5435 DPRINTF("allocated new mpage %zu, page size %u",
5436 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5437 np->mp_flags = flags | P_DIRTY;
5438 np->mp_lower = PAGEHDRSZ;
5439 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5442 mc->mc_db->md_branch_pages++;
5443 else if (IS_LEAF(np))
5444 mc->mc_db->md_leaf_pages++;
5445 else if (IS_OVERFLOW(np)) {
5446 mc->mc_db->md_overflow_pages += num;
5454 /** Calculate the size of a leaf node.
5455 * The size depends on the environment's page size; if a data item
5456 * is too large it will be put onto an overflow page and the node
5457 * size will only include the key and not the data. Sizes are always
5458 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5459 * of the #MDB_node headers.
5460 * @param[in] env The environment handle.
5461 * @param[in] key The key for the node.
5462 * @param[in] data The data for the node.
5463 * @return The number of bytes needed to store the node.
5466 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5470 sz = LEAFSIZE(key, data);
5471 if (sz >= env->me_nodemax) {
5472 /* put on overflow page */
5473 sz -= data->mv_size - sizeof(pgno_t);
5477 return sz + sizeof(indx_t);
5480 /** Calculate the size of a branch node.
5481 * The size should depend on the environment's page size but since
5482 * we currently don't support spilling large keys onto overflow
5483 * pages, it's simply the size of the #MDB_node header plus the
5484 * size of the key. Sizes are always rounded up to an even number
5485 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5486 * @param[in] env The environment handle.
5487 * @param[in] key The key for the node.
5488 * @return The number of bytes needed to store the node.
5491 mdb_branch_size(MDB_env *env, MDB_val *key)
5496 if (sz >= env->me_nodemax) {
5497 /* put on overflow page */
5498 /* not implemented */
5499 /* sz -= key->size - sizeof(pgno_t); */
5502 return sz + sizeof(indx_t);
5505 /** Add a node to the page pointed to by the cursor.
5506 * @param[in] mc The cursor for this operation.
5507 * @param[in] indx The index on the page where the new node should be added.
5508 * @param[in] key The key for the new node.
5509 * @param[in] data The data for the new node, if any.
5510 * @param[in] pgno The page number, if adding a branch node.
5511 * @param[in] flags Flags for the node.
5512 * @return 0 on success, non-zero on failure. Possible errors are:
5514 * <li>ENOMEM - failed to allocate overflow pages for the node.
5515 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5516 * should never happen since all callers already calculate the
5517 * page's free space before calling this function.
5521 mdb_node_add(MDB_cursor *mc, indx_t indx,
5522 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5525 size_t node_size = NODESIZE;
5528 MDB_page *mp = mc->mc_pg[mc->mc_top];
5529 MDB_page *ofp = NULL; /* overflow page */
5532 assert(mp->mp_upper >= mp->mp_lower);
5534 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5535 IS_LEAF(mp) ? "leaf" : "branch",
5536 IS_SUBP(mp) ? "sub-" : "",
5537 mp->mp_pgno, indx, data ? data->mv_size : 0,
5538 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5541 /* Move higher keys up one slot. */
5542 int ksize = mc->mc_db->md_pad, dif;
5543 char *ptr = LEAF2KEY(mp, indx, ksize);
5544 dif = NUMKEYS(mp) - indx;
5546 memmove(ptr+ksize, ptr, dif*ksize);
5547 /* insert new key */
5548 memcpy(ptr, key->mv_data, ksize);
5550 /* Just using these for counting */
5551 mp->mp_lower += sizeof(indx_t);
5552 mp->mp_upper -= ksize - sizeof(indx_t);
5557 node_size += key->mv_size;
5561 if (F_ISSET(flags, F_BIGDATA)) {
5562 /* Data already on overflow page. */
5563 node_size += sizeof(pgno_t);
5564 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5565 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5567 /* Put data on overflow page. */
5568 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5569 data->mv_size, node_size+data->mv_size);
5570 node_size += sizeof(pgno_t);
5571 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5573 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5576 node_size += data->mv_size;
5579 node_size += node_size & 1;
5581 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5582 DPRINTF("not enough room in page %zu, got %u ptrs",
5583 mp->mp_pgno, NUMKEYS(mp));
5584 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5585 mp->mp_upper - mp->mp_lower);
5586 DPRINTF("node size = %zu", node_size);
5587 return MDB_PAGE_FULL;
5590 /* Move higher pointers up one slot. */
5591 for (i = NUMKEYS(mp); i > indx; i--)
5592 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5594 /* Adjust free space offsets. */
5595 ofs = mp->mp_upper - node_size;
5596 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5597 mp->mp_ptrs[indx] = ofs;
5599 mp->mp_lower += sizeof(indx_t);
5601 /* Write the node data. */
5602 node = NODEPTR(mp, indx);
5603 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5604 node->mn_flags = flags;
5606 SETDSZ(node,data->mv_size);
5611 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5616 if (F_ISSET(flags, F_BIGDATA))
5617 memcpy(node->mn_data + key->mv_size, data->mv_data,
5619 else if (F_ISSET(flags, MDB_RESERVE))
5620 data->mv_data = node->mn_data + key->mv_size;
5622 memcpy(node->mn_data + key->mv_size, data->mv_data,
5625 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5627 if (F_ISSET(flags, MDB_RESERVE))
5628 data->mv_data = METADATA(ofp);
5630 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5637 /** Delete the specified node from a page.
5638 * @param[in] mp The page to operate on.
5639 * @param[in] indx The index of the node to delete.
5640 * @param[in] ksize The size of a node. Only used if the page is
5641 * part of a #MDB_DUPFIXED database.
5644 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5647 indx_t i, j, numkeys, ptr;
5654 COPY_PGNO(pgno, mp->mp_pgno);
5655 DPRINTF("delete node %u on %s page %zu", indx,
5656 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5659 assert(indx < NUMKEYS(mp));
5662 int x = NUMKEYS(mp) - 1 - indx;
5663 base = LEAF2KEY(mp, indx, ksize);
5665 memmove(base, base + ksize, x * ksize);
5666 mp->mp_lower -= sizeof(indx_t);
5667 mp->mp_upper += ksize - sizeof(indx_t);
5671 node = NODEPTR(mp, indx);
5672 sz = NODESIZE + node->mn_ksize;
5674 if (F_ISSET(node->mn_flags, F_BIGDATA))
5675 sz += sizeof(pgno_t);
5677 sz += NODEDSZ(node);
5681 ptr = mp->mp_ptrs[indx];
5682 numkeys = NUMKEYS(mp);
5683 for (i = j = 0; i < numkeys; i++) {
5685 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5686 if (mp->mp_ptrs[i] < ptr)
5687 mp->mp_ptrs[j] += sz;
5692 base = (char *)mp + mp->mp_upper;
5693 memmove(base + sz, base, ptr - mp->mp_upper);
5695 mp->mp_lower -= sizeof(indx_t);
5699 /** Compact the main page after deleting a node on a subpage.
5700 * @param[in] mp The main page to operate on.
5701 * @param[in] indx The index of the subpage on the main page.
5704 mdb_node_shrink(MDB_page *mp, indx_t indx)
5711 indx_t i, numkeys, ptr;
5713 node = NODEPTR(mp, indx);
5714 sp = (MDB_page *)NODEDATA(node);
5715 osize = NODEDSZ(node);
5717 delta = sp->mp_upper - sp->mp_lower;
5718 SETDSZ(node, osize - delta);
5719 xp = (MDB_page *)((char *)sp + delta);
5721 /* shift subpage upward */
5723 nsize = NUMKEYS(sp) * sp->mp_pad;
5724 memmove(METADATA(xp), METADATA(sp), nsize);
5727 nsize = osize - sp->mp_upper;
5728 numkeys = NUMKEYS(sp);
5729 for (i=numkeys-1; i>=0; i--)
5730 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5732 xp->mp_upper = sp->mp_lower;
5733 xp->mp_lower = sp->mp_lower;
5734 xp->mp_flags = sp->mp_flags;
5735 xp->mp_pad = sp->mp_pad;
5736 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5738 /* shift lower nodes upward */
5739 ptr = mp->mp_ptrs[indx];
5740 numkeys = NUMKEYS(mp);
5741 for (i = 0; i < numkeys; i++) {
5742 if (mp->mp_ptrs[i] <= ptr)
5743 mp->mp_ptrs[i] += delta;
5746 base = (char *)mp + mp->mp_upper;
5747 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5748 mp->mp_upper += delta;
5751 /** Initial setup of a sorted-dups cursor.
5752 * Sorted duplicates are implemented as a sub-database for the given key.
5753 * The duplicate data items are actually keys of the sub-database.
5754 * Operations on the duplicate data items are performed using a sub-cursor
5755 * initialized when the sub-database is first accessed. This function does
5756 * the preliminary setup of the sub-cursor, filling in the fields that
5757 * depend only on the parent DB.
5758 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5761 mdb_xcursor_init0(MDB_cursor *mc)
5763 MDB_xcursor *mx = mc->mc_xcursor;
5765 mx->mx_cursor.mc_xcursor = NULL;
5766 mx->mx_cursor.mc_txn = mc->mc_txn;
5767 mx->mx_cursor.mc_db = &mx->mx_db;
5768 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5769 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5770 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5771 mx->mx_cursor.mc_snum = 0;
5772 mx->mx_cursor.mc_top = 0;
5773 mx->mx_cursor.mc_flags = C_SUB;
5774 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5775 mx->mx_dbx.md_dcmp = NULL;
5776 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5779 /** Final setup of a sorted-dups cursor.
5780 * Sets up the fields that depend on the data from the main cursor.
5781 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5782 * @param[in] node The data containing the #MDB_db record for the
5783 * sorted-dup database.
5786 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5788 MDB_xcursor *mx = mc->mc_xcursor;
5790 if (node->mn_flags & F_SUBDATA) {
5791 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5792 mx->mx_cursor.mc_pg[0] = 0;
5793 mx->mx_cursor.mc_snum = 0;
5794 mx->mx_cursor.mc_flags = C_SUB;
5796 MDB_page *fp = NODEDATA(node);
5797 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5798 mx->mx_db.md_flags = 0;
5799 mx->mx_db.md_depth = 1;
5800 mx->mx_db.md_branch_pages = 0;
5801 mx->mx_db.md_leaf_pages = 1;
5802 mx->mx_db.md_overflow_pages = 0;
5803 mx->mx_db.md_entries = NUMKEYS(fp);
5804 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5805 mx->mx_cursor.mc_snum = 1;
5806 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5807 mx->mx_cursor.mc_top = 0;
5808 mx->mx_cursor.mc_pg[0] = fp;
5809 mx->mx_cursor.mc_ki[0] = 0;
5810 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5811 mx->mx_db.md_flags = MDB_DUPFIXED;
5812 mx->mx_db.md_pad = fp->mp_pad;
5813 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5814 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5817 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5819 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5821 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5822 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5823 #if UINT_MAX < SIZE_MAX
5824 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5825 #ifdef MISALIGNED_OK
5826 mx->mx_dbx.md_cmp = mdb_cmp_long;
5828 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5833 /** Initialize a cursor for a given transaction and database. */
5835 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5840 mc->mc_db = &txn->mt_dbs[dbi];
5841 mc->mc_dbx = &txn->mt_dbxs[dbi];
5842 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5847 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5849 mc->mc_xcursor = mx;
5850 mdb_xcursor_init0(mc);
5852 mc->mc_xcursor = NULL;
5854 if (*mc->mc_dbflag & DB_STALE) {
5855 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5860 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5863 size_t size = sizeof(MDB_cursor);
5865 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5868 /* Allow read access to the freelist */
5869 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5872 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5873 size += sizeof(MDB_xcursor);
5875 if ((mc = malloc(size)) != NULL) {
5876 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5877 if (txn->mt_cursors) {
5878 mc->mc_next = txn->mt_cursors[dbi];
5879 txn->mt_cursors[dbi] = mc;
5880 mc->mc_flags |= C_UNTRACK;
5882 mc->mc_flags |= C_ALLOCD;
5893 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5897 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5900 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5903 flags = mc->mc_flags;
5905 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5907 mc->mc_flags |= (flags & C_ALLOCD);
5911 /* Return the count of duplicate data items for the current key */
5913 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5917 if (mc == NULL || countp == NULL)
5920 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5923 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5924 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5927 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5930 *countp = mc->mc_xcursor->mx_db.md_entries;
5936 mdb_cursor_close(MDB_cursor *mc)
5939 /* remove from txn, if tracked */
5940 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
5941 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5942 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5944 *prev = mc->mc_next;
5946 if (mc->mc_flags & C_ALLOCD)
5952 mdb_cursor_txn(MDB_cursor *mc)
5954 if (!mc) return NULL;
5959 mdb_cursor_dbi(MDB_cursor *mc)
5965 /** Replace the key for a node with a new key.
5966 * @param[in] mc Cursor pointing to the node to operate on.
5967 * @param[in] key The new key to use.
5968 * @return 0 on success, non-zero on failure.
5971 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5978 indx_t ptr, i, numkeys, indx;
5981 indx = mc->mc_ki[mc->mc_top];
5982 mp = mc->mc_pg[mc->mc_top];
5983 node = NODEPTR(mp, indx);
5984 ptr = mp->mp_ptrs[indx];
5988 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5989 k2.mv_data = NODEKEY(node);
5990 k2.mv_size = node->mn_ksize;
5991 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5993 mdb_dkey(&k2, kbuf2),
5999 delta0 = delta = key->mv_size - node->mn_ksize;
6001 /* Must be 2-byte aligned. If new key is
6002 * shorter by 1, the shift will be skipped.
6004 delta += (delta & 1);
6006 if (delta > 0 && SIZELEFT(mp) < delta) {
6008 /* not enough space left, do a delete and split */
6009 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6010 pgno = NODEPGNO(node);
6011 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6012 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6015 numkeys = NUMKEYS(mp);
6016 for (i = 0; i < numkeys; i++) {
6017 if (mp->mp_ptrs[i] <= ptr)
6018 mp->mp_ptrs[i] -= delta;
6021 base = (char *)mp + mp->mp_upper;
6022 len = ptr - mp->mp_upper + NODESIZE;
6023 memmove(base - delta, base, len);
6024 mp->mp_upper -= delta;
6026 node = NODEPTR(mp, indx);
6029 /* But even if no shift was needed, update ksize */
6031 node->mn_ksize = key->mv_size;
6034 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6040 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6042 /** Move a node from csrc to cdst.
6045 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6052 unsigned short flags;
6056 /* Mark src and dst as dirty. */
6057 if ((rc = mdb_page_touch(csrc)) ||
6058 (rc = mdb_page_touch(cdst)))
6061 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6062 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6063 key.mv_size = csrc->mc_db->md_pad;
6064 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6066 data.mv_data = NULL;
6070 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6071 assert(!((long)srcnode&1));
6072 srcpg = NODEPGNO(srcnode);
6073 flags = srcnode->mn_flags;
6074 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6075 unsigned int snum = csrc->mc_snum;
6077 /* must find the lowest key below src */
6078 mdb_page_search_lowest(csrc);
6079 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6080 key.mv_size = csrc->mc_db->md_pad;
6081 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6083 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6084 key.mv_size = NODEKSZ(s2);
6085 key.mv_data = NODEKEY(s2);
6087 csrc->mc_snum = snum--;
6088 csrc->mc_top = snum;
6090 key.mv_size = NODEKSZ(srcnode);
6091 key.mv_data = NODEKEY(srcnode);
6093 data.mv_size = NODEDSZ(srcnode);
6094 data.mv_data = NODEDATA(srcnode);
6096 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6097 unsigned int snum = cdst->mc_snum;
6100 /* must find the lowest key below dst */
6101 mdb_page_search_lowest(cdst);
6102 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6103 bkey.mv_size = cdst->mc_db->md_pad;
6104 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6106 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6107 bkey.mv_size = NODEKSZ(s2);
6108 bkey.mv_data = NODEKEY(s2);
6110 cdst->mc_snum = snum--;
6111 cdst->mc_top = snum;
6112 mdb_cursor_copy(cdst, &mn);
6114 rc = mdb_update_key(&mn, &bkey);
6119 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6120 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6121 csrc->mc_ki[csrc->mc_top],
6123 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6124 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6126 /* Add the node to the destination page.
6128 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6129 if (rc != MDB_SUCCESS)
6132 /* Delete the node from the source page.
6134 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6137 /* Adjust other cursors pointing to mp */
6138 MDB_cursor *m2, *m3;
6139 MDB_dbi dbi = csrc->mc_dbi;
6140 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6142 if (csrc->mc_flags & C_SUB)
6145 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6146 if (m2 == csrc) continue;
6147 if (csrc->mc_flags & C_SUB)
6148 m3 = &m2->mc_xcursor->mx_cursor;
6151 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6152 csrc->mc_ki[csrc->mc_top]) {
6153 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6154 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6159 /* Update the parent separators.
6161 if (csrc->mc_ki[csrc->mc_top] == 0) {
6162 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6163 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6164 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6166 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6167 key.mv_size = NODEKSZ(srcnode);
6168 key.mv_data = NODEKEY(srcnode);
6170 DPRINTF("update separator for source page %zu to [%s]",
6171 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6172 mdb_cursor_copy(csrc, &mn);
6175 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6178 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6180 indx_t ix = csrc->mc_ki[csrc->mc_top];
6181 nullkey.mv_size = 0;
6182 csrc->mc_ki[csrc->mc_top] = 0;
6183 rc = mdb_update_key(csrc, &nullkey);
6184 csrc->mc_ki[csrc->mc_top] = ix;
6185 assert(rc == MDB_SUCCESS);
6189 if (cdst->mc_ki[cdst->mc_top] == 0) {
6190 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6191 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6192 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6194 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6195 key.mv_size = NODEKSZ(srcnode);
6196 key.mv_data = NODEKEY(srcnode);
6198 DPRINTF("update separator for destination page %zu to [%s]",
6199 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6200 mdb_cursor_copy(cdst, &mn);
6203 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6206 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6208 indx_t ix = cdst->mc_ki[cdst->mc_top];
6209 nullkey.mv_size = 0;
6210 cdst->mc_ki[cdst->mc_top] = 0;
6211 rc = mdb_update_key(cdst, &nullkey);
6212 cdst->mc_ki[cdst->mc_top] = ix;
6213 assert(rc == MDB_SUCCESS);
6220 /** Merge one page into another.
6221 * The nodes from the page pointed to by \b csrc will
6222 * be copied to the page pointed to by \b cdst and then
6223 * the \b csrc page will be freed.
6224 * @param[in] csrc Cursor pointing to the source page.
6225 * @param[in] cdst Cursor pointing to the destination page.
6228 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6236 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6237 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6239 assert(csrc->mc_snum > 1); /* can't merge root page */
6240 assert(cdst->mc_snum > 1);
6242 /* Mark dst as dirty. */
6243 if ((rc = mdb_page_touch(cdst)))
6246 /* Move all nodes from src to dst.
6248 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6249 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6250 key.mv_size = csrc->mc_db->md_pad;
6251 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6252 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6253 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6254 if (rc != MDB_SUCCESS)
6256 key.mv_data = (char *)key.mv_data + key.mv_size;
6259 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6260 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6261 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6262 unsigned int snum = csrc->mc_snum;
6264 /* must find the lowest key below src */
6265 mdb_page_search_lowest(csrc);
6266 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6267 key.mv_size = csrc->mc_db->md_pad;
6268 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6270 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6271 key.mv_size = NODEKSZ(s2);
6272 key.mv_data = NODEKEY(s2);
6274 csrc->mc_snum = snum--;
6275 csrc->mc_top = snum;
6277 key.mv_size = srcnode->mn_ksize;
6278 key.mv_data = NODEKEY(srcnode);
6281 data.mv_size = NODEDSZ(srcnode);
6282 data.mv_data = NODEDATA(srcnode);
6283 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6284 if (rc != MDB_SUCCESS)
6289 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6290 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);
6292 /* Unlink the src page from parent and add to free list.
6294 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6295 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6298 rc = mdb_update_key(csrc, &key);
6304 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6305 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6306 csrc->mc_db->md_leaf_pages--;
6308 csrc->mc_db->md_branch_pages--;
6310 /* Adjust other cursors pointing to mp */
6311 MDB_cursor *m2, *m3;
6312 MDB_dbi dbi = csrc->mc_dbi;
6313 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6315 if (csrc->mc_flags & C_SUB)
6318 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6319 if (csrc->mc_flags & C_SUB)
6320 m3 = &m2->mc_xcursor->mx_cursor;
6323 if (m3 == csrc) continue;
6324 if (m3->mc_snum < csrc->mc_snum) continue;
6325 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6326 m3->mc_pg[csrc->mc_top] = mp;
6327 m3->mc_ki[csrc->mc_top] += nkeys;
6331 mdb_cursor_pop(csrc);
6333 return mdb_rebalance(csrc);
6336 /** Copy the contents of a cursor.
6337 * @param[in] csrc The cursor to copy from.
6338 * @param[out] cdst The cursor to copy to.
6341 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6345 cdst->mc_txn = csrc->mc_txn;
6346 cdst->mc_dbi = csrc->mc_dbi;
6347 cdst->mc_db = csrc->mc_db;
6348 cdst->mc_dbx = csrc->mc_dbx;
6349 cdst->mc_snum = csrc->mc_snum;
6350 cdst->mc_top = csrc->mc_top;
6351 cdst->mc_flags = csrc->mc_flags;
6353 for (i=0; i<csrc->mc_snum; i++) {
6354 cdst->mc_pg[i] = csrc->mc_pg[i];
6355 cdst->mc_ki[i] = csrc->mc_ki[i];
6359 /** Rebalance the tree after a delete operation.
6360 * @param[in] mc Cursor pointing to the page where rebalancing
6362 * @return 0 on success, non-zero on failure.
6365 mdb_rebalance(MDB_cursor *mc)
6369 unsigned int ptop, minkeys;
6372 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6376 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6377 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6378 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6379 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6383 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6384 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6387 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6388 DPRINTF("no need to rebalance page %zu, above fill threshold",
6394 if (mc->mc_snum < 2) {
6395 MDB_page *mp = mc->mc_pg[0];
6397 DPUTS("Can't rebalance a subpage, ignoring");
6400 if (NUMKEYS(mp) == 0) {
6401 DPUTS("tree is completely empty");
6402 mc->mc_db->md_root = P_INVALID;
6403 mc->mc_db->md_depth = 0;
6404 mc->mc_db->md_leaf_pages = 0;
6405 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6409 /* Adjust other cursors pointing to mp */
6410 MDB_cursor *m2, *m3;
6411 MDB_dbi dbi = mc->mc_dbi;
6413 if (mc->mc_flags & C_SUB)
6416 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6417 if (m2 == mc) continue;
6418 if (mc->mc_flags & C_SUB)
6419 m3 = &m2->mc_xcursor->mx_cursor;
6422 if (m3->mc_snum < mc->mc_snum) continue;
6423 if (m3->mc_pg[0] == mp) {
6429 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6430 DPUTS("collapsing root page!");
6431 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6432 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6433 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6436 mc->mc_db->md_depth--;
6437 mc->mc_db->md_branch_pages--;
6439 /* Adjust other cursors pointing to mp */
6440 MDB_cursor *m2, *m3;
6441 MDB_dbi dbi = mc->mc_dbi;
6443 if (mc->mc_flags & C_SUB)
6446 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6447 if (m2 == mc) continue;
6448 if (mc->mc_flags & C_SUB)
6449 m3 = &m2->mc_xcursor->mx_cursor;
6452 if (m3->mc_snum < mc->mc_snum) continue;
6453 if (m3->mc_pg[0] == mp) {
6454 m3->mc_pg[0] = mc->mc_pg[0];
6461 DPUTS("root page doesn't need rebalancing");
6465 /* The parent (branch page) must have at least 2 pointers,
6466 * otherwise the tree is invalid.
6468 ptop = mc->mc_top-1;
6469 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6471 /* Leaf page fill factor is below the threshold.
6472 * Try to move keys from left or right neighbor, or
6473 * merge with a neighbor page.
6478 mdb_cursor_copy(mc, &mn);
6479 mn.mc_xcursor = NULL;
6481 if (mc->mc_ki[ptop] == 0) {
6482 /* We're the leftmost leaf in our parent.
6484 DPUTS("reading right neighbor");
6486 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6487 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6490 mn.mc_ki[mn.mc_top] = 0;
6491 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6493 /* There is at least one neighbor to the left.
6495 DPUTS("reading left neighbor");
6497 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6498 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6501 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6502 mc->mc_ki[mc->mc_top] = 0;
6505 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6506 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);
6508 /* If the neighbor page is above threshold and has enough keys,
6509 * move one key from it. Otherwise we should try to merge them.
6510 * (A branch page must never have less than 2 keys.)
6512 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6513 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6514 return mdb_node_move(&mn, mc);
6516 if (mc->mc_ki[ptop] == 0)
6517 rc = mdb_page_merge(&mn, mc);
6519 rc = mdb_page_merge(mc, &mn);
6520 mc->mc_flags &= ~C_INITIALIZED;
6525 /** Complete a delete operation started by #mdb_cursor_del(). */
6527 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6531 /* add overflow pages to free list */
6532 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6537 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6538 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6540 assert(IS_OVERFLOW(omp));
6541 ovpages = omp->mp_pages;
6542 mc->mc_db->md_overflow_pages -= ovpages;
6543 for (i=0; i<ovpages; i++) {
6544 DPRINTF("freed ov page %zu", pg);
6545 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6549 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6550 mc->mc_db->md_entries--;
6551 rc = mdb_rebalance(mc);
6552 if (rc != MDB_SUCCESS)
6553 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6554 /* if mc points past last node in page, invalidate */
6555 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6556 mc->mc_flags &= ~C_INITIALIZED;
6562 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6563 MDB_val *key, MDB_val *data)
6568 MDB_val rdata, *xdata;
6572 assert(key != NULL);
6574 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6576 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6579 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6583 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6587 mdb_cursor_init(&mc, txn, dbi, &mx);
6598 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6600 /* let mdb_page_split know about this cursor if needed:
6601 * delete will trigger a rebalance; if it needs to move
6602 * a node from one page to another, it will have to
6603 * update the parent's separator key(s). If the new sepkey
6604 * is larger than the current one, the parent page may
6605 * run out of space, triggering a split. We need this
6606 * cursor to be consistent until the end of the rebalance.
6608 mc.mc_next = txn->mt_cursors[dbi];
6609 txn->mt_cursors[dbi] = &mc;
6610 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6611 txn->mt_cursors[dbi] = mc.mc_next;
6616 /** Split a page and insert a new node.
6617 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6618 * The cursor will be updated to point to the actual page and index where
6619 * the node got inserted after the split.
6620 * @param[in] newkey The key for the newly inserted node.
6621 * @param[in] newdata The data for the newly inserted node.
6622 * @param[in] newpgno The page number, if the new node is a branch node.
6623 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6624 * @return 0 on success, non-zero on failure.
6627 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6628 unsigned int nflags)
6631 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6634 unsigned int i, j, split_indx, nkeys, pmax;
6636 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6638 MDB_page *mp, *rp, *pp;
6643 mp = mc->mc_pg[mc->mc_top];
6644 newindx = mc->mc_ki[mc->mc_top];
6646 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6647 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6648 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6650 /* Create a right sibling. */
6651 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6653 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6655 if (mc->mc_snum < 2) {
6656 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6658 /* shift current top to make room for new parent */
6659 mc->mc_pg[1] = mc->mc_pg[0];
6660 mc->mc_ki[1] = mc->mc_ki[0];
6663 mc->mc_db->md_root = pp->mp_pgno;
6664 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6665 mc->mc_db->md_depth++;
6668 /* Add left (implicit) pointer. */
6669 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6670 /* undo the pre-push */
6671 mc->mc_pg[0] = mc->mc_pg[1];
6672 mc->mc_ki[0] = mc->mc_ki[1];
6673 mc->mc_db->md_root = mp->mp_pgno;
6674 mc->mc_db->md_depth--;
6681 ptop = mc->mc_top-1;
6682 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6685 mc->mc_flags |= C_SPLITTING;
6686 mdb_cursor_copy(mc, &mn);
6687 mn.mc_pg[mn.mc_top] = rp;
6688 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6690 if (nflags & MDB_APPEND) {
6691 mn.mc_ki[mn.mc_top] = 0;
6693 split_indx = newindx;
6698 nkeys = NUMKEYS(mp);
6699 split_indx = nkeys / 2;
6700 if (newindx < split_indx)
6706 unsigned int lsize, rsize, ksize;
6707 /* Move half of the keys to the right sibling */
6709 x = mc->mc_ki[mc->mc_top] - split_indx;
6710 ksize = mc->mc_db->md_pad;
6711 split = LEAF2KEY(mp, split_indx, ksize);
6712 rsize = (nkeys - split_indx) * ksize;
6713 lsize = (nkeys - split_indx) * sizeof(indx_t);
6714 mp->mp_lower -= lsize;
6715 rp->mp_lower += lsize;
6716 mp->mp_upper += rsize - lsize;
6717 rp->mp_upper -= rsize - lsize;
6718 sepkey.mv_size = ksize;
6719 if (newindx == split_indx) {
6720 sepkey.mv_data = newkey->mv_data;
6722 sepkey.mv_data = split;
6725 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6726 memcpy(rp->mp_ptrs, split, rsize);
6727 sepkey.mv_data = rp->mp_ptrs;
6728 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6729 memcpy(ins, newkey->mv_data, ksize);
6730 mp->mp_lower += sizeof(indx_t);
6731 mp->mp_upper -= ksize - sizeof(indx_t);
6734 memcpy(rp->mp_ptrs, split, x * ksize);
6735 ins = LEAF2KEY(rp, x, ksize);
6736 memcpy(ins, newkey->mv_data, ksize);
6737 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6738 rp->mp_lower += sizeof(indx_t);
6739 rp->mp_upper -= ksize - sizeof(indx_t);
6740 mc->mc_ki[mc->mc_top] = x;
6741 mc->mc_pg[mc->mc_top] = rp;
6746 /* For leaf pages, check the split point based on what
6747 * fits where, since otherwise mdb_node_add can fail.
6749 * This check is only needed when the data items are
6750 * relatively large, such that being off by one will
6751 * make the difference between success or failure.
6753 * It's also relevant if a page happens to be laid out
6754 * such that one half of its nodes are all "small" and
6755 * the other half of its nodes are "large." If the new
6756 * item is also "large" and falls on the half with
6757 * "large" nodes, it also may not fit.
6760 unsigned int psize, nsize;
6761 /* Maximum free space in an empty page */
6762 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6763 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6764 if ((nkeys < 20) || (nsize > pmax/16)) {
6765 if (newindx <= split_indx) {
6768 for (i=0; i<split_indx; i++) {
6769 node = NODEPTR(mp, i);
6770 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6771 if (F_ISSET(node->mn_flags, F_BIGDATA))
6772 psize += sizeof(pgno_t);
6774 psize += NODEDSZ(node);
6778 split_indx = newindx;
6789 for (i=nkeys-1; i>=split_indx; i--) {
6790 node = NODEPTR(mp, i);
6791 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6792 if (F_ISSET(node->mn_flags, F_BIGDATA))
6793 psize += sizeof(pgno_t);
6795 psize += NODEDSZ(node);
6799 split_indx = newindx;
6810 /* First find the separating key between the split pages.
6811 * The case where newindx == split_indx is ambiguous; the
6812 * new item could go to the new page or stay on the original
6813 * page. If newpos == 1 it goes to the new page.
6815 if (newindx == split_indx && newpos) {
6816 sepkey.mv_size = newkey->mv_size;
6817 sepkey.mv_data = newkey->mv_data;
6819 node = NODEPTR(mp, split_indx);
6820 sepkey.mv_size = node->mn_ksize;
6821 sepkey.mv_data = NODEKEY(node);
6825 DPRINTF("separator is [%s]", DKEY(&sepkey));
6827 /* Copy separator key to the parent.
6829 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6833 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6836 if (mn.mc_snum == mc->mc_snum) {
6837 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6838 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6839 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6840 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6845 /* Right page might now have changed parent.
6846 * Check if left page also changed parent.
6848 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6849 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6850 for (i=0; i<ptop; i++) {
6851 mc->mc_pg[i] = mn.mc_pg[i];
6852 mc->mc_ki[i] = mn.mc_ki[i];
6854 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6855 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6859 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6862 mc->mc_flags ^= C_SPLITTING;
6863 if (rc != MDB_SUCCESS) {
6866 if (nflags & MDB_APPEND) {
6867 mc->mc_pg[mc->mc_top] = rp;
6868 mc->mc_ki[mc->mc_top] = 0;
6869 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6872 for (i=0; i<mc->mc_top; i++)
6873 mc->mc_ki[i] = mn.mc_ki[i];
6880 /* Move half of the keys to the right sibling. */
6882 /* grab a page to hold a temporary copy */
6883 copy = mdb_page_malloc(mc, 1);
6887 copy->mp_pgno = mp->mp_pgno;
6888 copy->mp_flags = mp->mp_flags;
6889 copy->mp_lower = PAGEHDRSZ;
6890 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6891 mc->mc_pg[mc->mc_top] = copy;
6892 for (i = j = 0; i <= nkeys; j++) {
6893 if (i == split_indx) {
6894 /* Insert in right sibling. */
6895 /* Reset insert index for right sibling. */
6896 if (i != newindx || (newpos ^ ins_new)) {
6898 mc->mc_pg[mc->mc_top] = rp;
6902 if (i == newindx && !ins_new) {
6903 /* Insert the original entry that caused the split. */
6904 rkey.mv_data = newkey->mv_data;
6905 rkey.mv_size = newkey->mv_size;
6914 /* Update index for the new key. */
6915 mc->mc_ki[mc->mc_top] = j;
6916 } else if (i == nkeys) {
6919 node = NODEPTR(mp, i);
6920 rkey.mv_data = NODEKEY(node);
6921 rkey.mv_size = node->mn_ksize;
6923 xdata.mv_data = NODEDATA(node);
6924 xdata.mv_size = NODEDSZ(node);
6927 pgno = NODEPGNO(node);
6928 flags = node->mn_flags;
6933 if (!IS_LEAF(mp) && j == 0) {
6934 /* First branch index doesn't need key data. */
6938 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6942 nkeys = NUMKEYS(copy);
6943 for (i=0; i<nkeys; i++)
6944 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6945 mp->mp_lower = copy->mp_lower;
6946 mp->mp_upper = copy->mp_upper;
6947 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6948 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6950 /* reset back to original page */
6951 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6952 mc->mc_pg[mc->mc_top] = mp;
6953 if (nflags & MDB_RESERVE) {
6954 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6955 if (!(node->mn_flags & F_BIGDATA))
6956 newdata->mv_data = NODEDATA(node);
6960 /* Make sure mc_ki is still valid.
6962 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6963 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6964 for (i=0; i<ptop; i++) {
6965 mc->mc_pg[i] = mn.mc_pg[i];
6966 mc->mc_ki[i] = mn.mc_ki[i];
6968 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6969 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6973 /* return tmp page to freelist */
6974 mdb_page_free(mc->mc_txn->mt_env, copy);
6977 /* Adjust other cursors pointing to mp */
6978 MDB_cursor *m2, *m3;
6979 MDB_dbi dbi = mc->mc_dbi;
6980 int fixup = NUMKEYS(mp);
6982 if (mc->mc_flags & C_SUB)
6985 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6986 if (m2 == mc) continue;
6987 if (mc->mc_flags & C_SUB)
6988 m3 = &m2->mc_xcursor->mx_cursor;
6991 if (!(m3->mc_flags & C_INITIALIZED))
6993 if (m3->mc_flags & C_SPLITTING)
6998 for (k=m3->mc_top; k>=0; k--) {
6999 m3->mc_ki[k+1] = m3->mc_ki[k];
7000 m3->mc_pg[k+1] = m3->mc_pg[k];
7002 if (m3->mc_ki[0] >= split_indx) {
7007 m3->mc_pg[0] = mc->mc_pg[0];
7011 if (m3->mc_pg[mc->mc_top] == mp) {
7012 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7013 m3->mc_ki[mc->mc_top]++;
7014 if (m3->mc_ki[mc->mc_top] >= fixup) {
7015 m3->mc_pg[mc->mc_top] = rp;
7016 m3->mc_ki[mc->mc_top] -= fixup;
7017 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7019 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7020 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7029 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7030 MDB_val *key, MDB_val *data, unsigned int flags)
7035 assert(key != NULL);
7036 assert(data != NULL);
7038 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7041 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7045 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7049 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7052 mdb_cursor_init(&mc, txn, dbi, &mx);
7053 return mdb_cursor_put(&mc, key, data, flags);
7057 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7059 if ((flag & CHANGEABLE) != flag)
7062 env->me_flags |= flag;
7064 env->me_flags &= ~flag;
7069 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7074 *arg = env->me_flags;
7079 mdb_env_get_path(MDB_env *env, const char **arg)
7084 *arg = env->me_path;
7088 /** Common code for #mdb_stat() and #mdb_env_stat().
7089 * @param[in] env the environment to operate in.
7090 * @param[in] db the #MDB_db record containing the stats to return.
7091 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7092 * @return 0, this function always succeeds.
7095 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7097 arg->ms_psize = env->me_psize;
7098 arg->ms_depth = db->md_depth;
7099 arg->ms_branch_pages = db->md_branch_pages;
7100 arg->ms_leaf_pages = db->md_leaf_pages;
7101 arg->ms_overflow_pages = db->md_overflow_pages;
7102 arg->ms_entries = db->md_entries;
7107 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7111 if (env == NULL || arg == NULL)
7114 toggle = mdb_env_pick_meta(env);
7116 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7120 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7124 if (env == NULL || arg == NULL)
7127 toggle = mdb_env_pick_meta(env);
7128 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7129 arg->me_mapsize = env->me_mapsize;
7130 arg->me_maxreaders = env->me_maxreaders;
7131 arg->me_numreaders = env->me_numreaders;
7132 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7133 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7137 /** Set the default comparison functions for a database.
7138 * Called immediately after a database is opened to set the defaults.
7139 * The user can then override them with #mdb_set_compare() or
7140 * #mdb_set_dupsort().
7141 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7142 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7145 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7147 uint16_t f = txn->mt_dbs[dbi].md_flags;
7149 txn->mt_dbxs[dbi].md_cmp =
7150 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7151 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7153 txn->mt_dbxs[dbi].md_dcmp =
7154 !(f & MDB_DUPSORT) ? 0 :
7155 ((f & MDB_INTEGERDUP)
7156 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7157 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7160 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7165 int rc, dbflag, exact;
7166 unsigned int unused = 0;
7169 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7170 mdb_default_cmp(txn, FREE_DBI);
7173 if ((flags & VALID_FLAGS) != flags)
7179 if (flags & PERSISTENT_FLAGS) {
7180 uint16_t f2 = flags & PERSISTENT_FLAGS;
7181 /* make sure flag changes get committed */
7182 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7183 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7184 txn->mt_flags |= MDB_TXN_DIRTY;
7187 mdb_default_cmp(txn, MAIN_DBI);
7191 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7192 mdb_default_cmp(txn, MAIN_DBI);
7195 /* Is the DB already open? */
7197 for (i=2; i<txn->mt_numdbs; i++) {
7198 if (!txn->mt_dbxs[i].md_name.mv_size) {
7199 /* Remember this free slot */
7200 if (!unused) unused = i;
7203 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7204 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7210 /* If no free slot and max hit, fail */
7211 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7212 return MDB_DBS_FULL;
7214 /* Cannot mix named databases with some mainDB flags */
7215 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7216 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7218 /* Find the DB info */
7219 dbflag = DB_NEW|DB_VALID;
7222 key.mv_data = (void *)name;
7223 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7224 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7225 if (rc == MDB_SUCCESS) {
7226 /* make sure this is actually a DB */
7227 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7228 if (!(node->mn_flags & F_SUBDATA))
7230 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7231 /* Create if requested */
7233 data.mv_size = sizeof(MDB_db);
7234 data.mv_data = &dummy;
7235 memset(&dummy, 0, sizeof(dummy));
7236 dummy.md_root = P_INVALID;
7237 dummy.md_flags = flags & PERSISTENT_FLAGS;
7238 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7242 /* OK, got info, add to table */
7243 if (rc == MDB_SUCCESS) {
7244 unsigned int slot = unused ? unused : txn->mt_numdbs;
7245 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7246 txn->mt_dbxs[slot].md_name.mv_size = len;
7247 txn->mt_dbxs[slot].md_rel = NULL;
7248 txn->mt_dbflags[slot] = dbflag;
7249 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7251 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7252 mdb_default_cmp(txn, slot);
7261 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7263 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7266 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7269 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7272 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7274 ptr = env->me_dbxs[dbi].md_name.mv_data;
7275 env->me_dbxs[dbi].md_name.mv_data = NULL;
7276 env->me_dbxs[dbi].md_name.mv_size = 0;
7277 env->me_dbflags[dbi] = 0;
7281 /** Add all the DB's pages to the free list.
7282 * @param[in] mc Cursor on the DB to free.
7283 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7284 * @return 0 on success, non-zero on failure.
7287 mdb_drop0(MDB_cursor *mc, int subs)
7291 rc = mdb_page_search(mc, NULL, 0);
7292 if (rc == MDB_SUCCESS) {
7297 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7298 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7301 mdb_cursor_copy(mc, &mx);
7302 while (mc->mc_snum > 0) {
7303 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7304 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7305 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7306 if (ni->mn_flags & F_BIGDATA) {
7310 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7311 rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL);
7314 assert(IS_OVERFLOW(omp));
7315 ovpages = omp->mp_pages;
7316 for (j=0; j<ovpages; j++) {
7317 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7320 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7321 mdb_xcursor_init1(mc, ni);
7322 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7328 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7330 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7333 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7338 mc->mc_ki[mc->mc_top] = i;
7339 rc = mdb_cursor_sibling(mc, 1);
7341 /* no more siblings, go back to beginning
7342 * of previous level.
7346 for (i=1; i<mc->mc_snum; i++) {
7348 mc->mc_pg[i] = mx.mc_pg[i];
7353 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7354 mc->mc_db->md_root);
7359 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7364 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7367 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7370 rc = mdb_cursor_open(txn, dbi, &mc);
7374 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7378 /* Can't delete the main DB */
7379 if (del && dbi > MAIN_DBI) {
7380 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7382 txn->mt_dbflags[dbi] = DB_STALE;
7383 mdb_dbi_close(txn->mt_env, dbi);
7386 /* reset the DB record, mark it dirty */
7387 txn->mt_dbflags[dbi] |= DB_DIRTY;
7388 txn->mt_dbs[dbi].md_depth = 0;
7389 txn->mt_dbs[dbi].md_branch_pages = 0;
7390 txn->mt_dbs[dbi].md_leaf_pages = 0;
7391 txn->mt_dbs[dbi].md_overflow_pages = 0;
7392 txn->mt_dbs[dbi].md_entries = 0;
7393 txn->mt_dbs[dbi].md_root = P_INVALID;
7395 txn->mt_flags |= MDB_TXN_DIRTY;
7398 mdb_cursor_close(mc);
7402 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7404 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7407 txn->mt_dbxs[dbi].md_cmp = cmp;
7411 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7413 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7416 txn->mt_dbxs[dbi].md_dcmp = cmp;
7420 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7422 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7425 txn->mt_dbxs[dbi].md_rel = rel;
7429 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7431 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7434 txn->mt_dbxs[dbi].md_relctx = ctx;