2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define pthread_t DWORD
144 #define pthread_mutex_t HANDLE
145 #define pthread_key_t DWORD
146 #define pthread_self() GetCurrentThreadId()
147 #define pthread_key_create(x,y) \
148 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
149 #define pthread_key_delete(x) TlsFree(x)
150 #define pthread_getspecific(x) TlsGetValue(x)
151 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
152 #define pthread_mutex_unlock(x) ReleaseMutex(x)
153 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
154 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
158 #define getpid() GetCurrentProcessId()
159 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
160 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
161 #define ErrCode() GetLastError()
162 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
163 #define close(fd) (CloseHandle(fd) ? 0 : -1)
164 #define munmap(ptr,len) UnmapViewOfFile(ptr)
167 #ifdef MDB_USE_POSIX_SEM
169 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
170 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
171 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
172 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
175 mdb_sem_wait(sem_t *sem)
178 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
183 /** Lock the reader mutex.
185 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
186 /** Unlock the reader mutex.
188 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
190 /** Lock the writer mutex.
191 * Only a single write transaction is allowed at a time. Other writers
192 * will block waiting for this mutex.
194 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
195 /** Unlock the writer mutex.
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
198 #endif /* MDB_USE_POSIX_SEM */
200 /** Get the error code for the last failed system function.
202 #define ErrCode() errno
204 /** An abstraction for a file handle.
205 * On POSIX systems file handles are small integers. On Windows
206 * they're opaque pointers.
210 /** A value for an invalid file handle.
211 * Mainly used to initialize file variables and signify that they are
214 #define INVALID_HANDLE_VALUE (-1)
216 /** Get the size of a memory page for the system.
217 * This is the basic size that the platform's memory manager uses, and is
218 * fundamental to the use of memory-mapped files.
220 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
223 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
226 #define MNAME_LEN (sizeof(pthread_mutex_t))
232 /** A flag for opening a file and requesting synchronous data writes.
233 * This is only used when writing a meta page. It's not strictly needed;
234 * we could just do a normal write and then immediately perform a flush.
235 * But if this flag is available it saves us an extra system call.
237 * @note If O_DSYNC is undefined but exists in /usr/include,
238 * preferably set some compiler flag to get the definition.
239 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
242 # define MDB_DSYNC O_DSYNC
246 /** Function for flushing the data of a file. Define this to fsync
247 * if fdatasync() is not supported.
249 #ifndef MDB_FDATASYNC
250 # define MDB_FDATASYNC fdatasync
254 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
265 /** A page number in the database.
266 * Note that 64 bit page numbers are overkill, since pages themselves
267 * already represent 12-13 bits of addressable memory, and the OS will
268 * always limit applications to a maximum of 63 bits of address space.
270 * @note In the #MDB_node structure, we only store 48 bits of this value,
271 * which thus limits us to only 60 bits of addressable data.
273 typedef MDB_ID pgno_t;
275 /** A transaction ID.
276 * See struct MDB_txn.mt_txnid for details.
278 typedef MDB_ID txnid_t;
280 /** @defgroup debug Debug Macros
284 /** Enable debug output.
285 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
286 * read from and written to the database (used for free space management).
291 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
294 # define DPRINTF (void) /* Vararg macros may be unsupported */
296 static int mdb_debug;
297 static txnid_t mdb_debug_start;
299 /** Print a debug message with printf formatting. */
300 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
301 ((void) ((mdb_debug) && \
302 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
304 # define DPRINTF(fmt, ...) ((void) 0)
305 # define MDB_DEBUG_SKIP
307 /** Print a debug string.
308 * The string is printed literally, with no format processing.
310 #define DPUTS(arg) DPRINTF("%s", arg)
313 /** A default memory page size.
314 * The actual size is platform-dependent, but we use this for
315 * boot-strapping. We probably should not be using this any more.
316 * The #GET_PAGESIZE() macro is used to get the actual size.
318 * Note that we don't currently support Huge pages. On Linux,
319 * regular data files cannot use Huge pages, and in general
320 * Huge pages aren't actually pageable. We rely on the OS
321 * demand-pager to read our data and page it out when memory
322 * pressure from other processes is high. So until OSs have
323 * actual paging support for Huge pages, they're not viable.
325 #define MDB_PAGESIZE 4096
327 /** The minimum number of keys required in a database page.
328 * Setting this to a larger value will place a smaller bound on the
329 * maximum size of a data item. Data items larger than this size will
330 * be pushed into overflow pages instead of being stored directly in
331 * the B-tree node. This value used to default to 4. With a page size
332 * of 4096 bytes that meant that any item larger than 1024 bytes would
333 * go into an overflow page. That also meant that on average 2-3KB of
334 * each overflow page was wasted space. The value cannot be lower than
335 * 2 because then there would no longer be a tree structure. With this
336 * value, items larger than 2KB will go into overflow pages, and on
337 * average only 1KB will be wasted.
339 #define MDB_MINKEYS 2
341 /** A stamp that identifies a file as an MDB file.
342 * There's nothing special about this value other than that it is easily
343 * recognizable, and it will reflect any byte order mismatches.
345 #define MDB_MAGIC 0xBEEFC0DE
347 /** The version number for a database's file format. */
348 #define MDB_VERSION 1
350 /** @brief The maximum size of a key in the database.
352 * The library rejects bigger keys, and cannot deal with records
353 * with bigger keys stored by a library with bigger max keysize.
355 * We require that keys all fit onto a regular page. This limit
356 * could be raised a bit further if needed; to something just
357 * under #MDB_PAGESIZE / #MDB_MINKEYS.
359 * Note that data items in an #MDB_DUPSORT database are actually keys
360 * of a subDB, so they're also limited to this size.
362 #ifndef MDB_MAXKEYSIZE
363 #define MDB_MAXKEYSIZE 511
366 /** @brief The maximum size of a data item.
368 * We only store a 32 bit value for node sizes.
370 #define MAXDATASIZE 0xffffffffUL
375 * This is used for printing a hex dump of a key's contents.
377 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
378 /** Display a key in hex.
380 * Invoke a function to display a key in hex.
382 #define DKEY(x) mdb_dkey(x, kbuf)
384 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
388 /** An invalid page number.
389 * Mainly used to denote an empty tree.
391 #define P_INVALID (~(pgno_t)0)
393 /** Test if the flags \b f are set in a flag word \b w. */
394 #define F_ISSET(w, f) (((w) & (f)) == (f))
396 /** Used for offsets within a single page.
397 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
400 typedef uint16_t indx_t;
402 /** Default size of memory map.
403 * This is certainly too small for any actual applications. Apps should always set
404 * the size explicitly using #mdb_env_set_mapsize().
406 #define DEFAULT_MAPSIZE 1048576
408 /** @defgroup readers Reader Lock Table
409 * Readers don't acquire any locks for their data access. Instead, they
410 * simply record their transaction ID in the reader table. The reader
411 * mutex is needed just to find an empty slot in the reader table. The
412 * slot's address is saved in thread-specific data so that subsequent read
413 * transactions started by the same thread need no further locking to proceed.
415 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
417 * No reader table is used if the database is on a read-only filesystem.
419 * Since the database uses multi-version concurrency control, readers don't
420 * actually need any locking. This table is used to keep track of which
421 * readers are using data from which old transactions, so that we'll know
422 * when a particular old transaction is no longer in use. Old transactions
423 * that have discarded any data pages can then have those pages reclaimed
424 * for use by a later write transaction.
426 * The lock table is constructed such that reader slots are aligned with the
427 * processor's cache line size. Any slot is only ever used by one thread.
428 * This alignment guarantees that there will be no contention or cache
429 * thrashing as threads update their own slot info, and also eliminates
430 * any need for locking when accessing a slot.
432 * A writer thread will scan every slot in the table to determine the oldest
433 * outstanding reader transaction. Any freed pages older than this will be
434 * reclaimed by the writer. The writer doesn't use any locks when scanning
435 * this table. This means that there's no guarantee that the writer will
436 * see the most up-to-date reader info, but that's not required for correct
437 * operation - all we need is to know the upper bound on the oldest reader,
438 * we don't care at all about the newest reader. So the only consequence of
439 * reading stale information here is that old pages might hang around a
440 * while longer before being reclaimed. That's actually good anyway, because
441 * the longer we delay reclaiming old pages, the more likely it is that a
442 * string of contiguous pages can be found after coalescing old pages from
443 * many old transactions together.
446 /** Number of slots in the reader table.
447 * This value was chosen somewhat arbitrarily. 126 readers plus a
448 * couple mutexes fit exactly into 8KB on my development machine.
449 * Applications should set the table size using #mdb_env_set_maxreaders().
451 #define DEFAULT_READERS 126
453 /** The size of a CPU cache line in bytes. We want our lock structures
454 * aligned to this size to avoid false cache line sharing in the
456 * This value works for most CPUs. For Itanium this should be 128.
462 /** The information we store in a single slot of the reader table.
463 * In addition to a transaction ID, we also record the process and
464 * thread ID that owns a slot, so that we can detect stale information,
465 * e.g. threads or processes that went away without cleaning up.
466 * @note We currently don't check for stale records. We simply re-init
467 * the table when we know that we're the only process opening the
470 typedef struct MDB_rxbody {
471 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
472 * Multiple readers that start at the same time will probably have the
473 * same ID here. Again, it's not important to exclude them from
474 * anything; all we need to know is which version of the DB they
475 * started from so we can avoid overwriting any data used in that
476 * particular version.
479 /** The process ID of the process owning this reader txn. */
481 /** The thread ID of the thread owning this txn. */
485 /** The actual reader record, with cacheline padding. */
486 typedef struct MDB_reader {
489 /** shorthand for mrb_txnid */
490 #define mr_txnid mru.mrx.mrb_txnid
491 #define mr_pid mru.mrx.mrb_pid
492 #define mr_tid mru.mrx.mrb_tid
493 /** cache line alignment */
494 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
498 /** The header for the reader table.
499 * The table resides in a memory-mapped file. (This is a different file
500 * than is used for the main database.)
502 * For POSIX the actual mutexes reside in the shared memory of this
503 * mapped file. On Windows, mutexes are named objects allocated by the
504 * kernel; we store the mutex names in this mapped file so that other
505 * processes can grab them. This same approach is also used on
506 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
507 * process-shared POSIX mutexes. For these cases where a named object
508 * is used, the object name is derived from a 64 bit FNV hash of the
509 * environment pathname. As such, naming collisions are extremely
510 * unlikely. If a collision occurs, the results are unpredictable.
512 typedef struct MDB_txbody {
513 /** Stamp identifying this as an MDB file. It must be set
516 /** Version number of this lock file. Must be set to #MDB_VERSION. */
517 uint32_t mtb_version;
518 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
519 char mtb_rmname[MNAME_LEN];
521 /** Mutex protecting access to this table.
522 * This is the reader lock that #LOCK_MUTEX_R acquires.
524 pthread_mutex_t mtb_mutex;
526 /** The ID of the last transaction committed to the database.
527 * This is recorded here only for convenience; the value can always
528 * be determined by reading the main database meta pages.
531 /** The number of slots that have been used in the reader table.
532 * This always records the maximum count, it is not decremented
533 * when readers release their slots.
535 unsigned mtb_numreaders;
538 /** The actual reader table definition. */
539 typedef struct MDB_txninfo {
542 #define mti_magic mt1.mtb.mtb_magic
543 #define mti_version mt1.mtb.mtb_version
544 #define mti_mutex mt1.mtb.mtb_mutex
545 #define mti_rmname mt1.mtb.mtb_rmname
546 #define mti_txnid mt1.mtb.mtb_txnid
547 #define mti_numreaders mt1.mtb.mtb_numreaders
548 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
551 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
552 char mt2_wmname[MNAME_LEN];
553 #define mti_wmname mt2.mt2_wmname
555 pthread_mutex_t mt2_wmutex;
556 #define mti_wmutex mt2.mt2_wmutex
558 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
560 MDB_reader mti_readers[1];
564 /** Common header for all page types.
565 * Overflow records occupy a number of contiguous pages with no
566 * headers on any page after the first.
568 typedef struct MDB_page {
569 #define mp_pgno mp_p.p_pgno
570 #define mp_next mp_p.p_next
572 pgno_t p_pgno; /**< page number */
573 void * p_next; /**< for in-memory list of freed structs */
576 /** @defgroup mdb_page Page Flags
578 * Flags for the page headers.
581 #define P_BRANCH 0x01 /**< branch page */
582 #define P_LEAF 0x02 /**< leaf page */
583 #define P_OVERFLOW 0x04 /**< overflow page */
584 #define P_META 0x08 /**< meta page */
585 #define P_DIRTY 0x10 /**< dirty page */
586 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
587 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
589 uint16_t mp_flags; /**< @ref mdb_page */
590 #define mp_lower mp_pb.pb.pb_lower
591 #define mp_upper mp_pb.pb.pb_upper
592 #define mp_pages mp_pb.pb_pages
595 indx_t pb_lower; /**< lower bound of free space */
596 indx_t pb_upper; /**< upper bound of free space */
598 uint32_t pb_pages; /**< number of overflow pages */
600 indx_t mp_ptrs[1]; /**< dynamic size */
603 /** Size of the page header, excluding dynamic data at the end */
604 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
606 /** Address of first usable data byte in a page, after the header */
607 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
609 /** Number of nodes on a page */
610 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
612 /** The amount of space remaining in the page */
613 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
615 /** The percentage of space used in the page, in tenths of a percent. */
616 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
617 ((env)->me_psize - PAGEHDRSZ))
618 /** The minimum page fill factor, in tenths of a percent.
619 * Pages emptier than this are candidates for merging.
621 #define FILL_THRESHOLD 250
623 /** Test if a page is a leaf page */
624 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
625 /** Test if a page is a LEAF2 page */
626 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
627 /** Test if a page is a branch page */
628 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
629 /** Test if a page is an overflow page */
630 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
631 /** Test if a page is a sub page */
632 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
634 /** The number of overflow pages needed to store the given size. */
635 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
637 /** Header for a single key/data pair within a page.
638 * We guarantee 2-byte alignment for nodes.
640 typedef struct MDB_node {
641 /** lo and hi are used for data size on leaf nodes and for
642 * child pgno on branch nodes. On 64 bit platforms, flags
643 * is also used for pgno. (Branch nodes have no flags).
644 * They are in host byte order in case that lets some
645 * accesses be optimized into a 32-bit word access.
647 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
648 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
649 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
650 /** @defgroup mdb_node Node Flags
652 * Flags for node headers.
655 #define F_BIGDATA 0x01 /**< data put on overflow page */
656 #define F_SUBDATA 0x02 /**< data is a sub-database */
657 #define F_DUPDATA 0x04 /**< data has duplicates */
659 /** valid flags for #mdb_node_add() */
660 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
663 unsigned short mn_flags; /**< @ref mdb_node */
664 unsigned short mn_ksize; /**< key size */
665 char mn_data[1]; /**< key and data are appended here */
668 /** Size of the node header, excluding dynamic data at the end */
669 #define NODESIZE offsetof(MDB_node, mn_data)
671 /** Bit position of top word in page number, for shifting mn_flags */
672 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
674 /** Size of a node in a branch page with a given key.
675 * This is just the node header plus the key, there is no data.
677 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
679 /** Size of a node in a leaf page with a given key and data.
680 * This is node header plus key plus data size.
682 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
684 /** Address of node \b i in page \b p */
685 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
687 /** Address of the key for the node */
688 #define NODEKEY(node) (void *)((node)->mn_data)
690 /** Address of the data for a node */
691 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
693 /** Get the page number pointed to by a branch node */
694 #define NODEPGNO(node) \
695 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
696 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
697 /** Set the page number in a branch node */
698 #define SETPGNO(node,pgno) do { \
699 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
700 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
702 /** Get the size of the data in a leaf node */
703 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
704 /** Set the size of the data for a leaf node */
705 #define SETDSZ(node,size) do { \
706 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
707 /** The size of a key in a node */
708 #define NODEKSZ(node) ((node)->mn_ksize)
710 /** Copy a page number from src to dst */
712 #define COPY_PGNO(dst,src) dst = src
714 #if SIZE_MAX > 4294967295UL
715 #define COPY_PGNO(dst,src) do { \
716 unsigned short *s, *d; \
717 s = (unsigned short *)&(src); \
718 d = (unsigned short *)&(dst); \
725 #define COPY_PGNO(dst,src) do { \
726 unsigned short *s, *d; \
727 s = (unsigned short *)&(src); \
728 d = (unsigned short *)&(dst); \
734 /** The address of a key in a LEAF2 page.
735 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
736 * There are no node headers, keys are stored contiguously.
738 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
740 /** Set the \b node's key into \b key, if requested. */
741 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
742 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
744 /** Information about a single database in the environment. */
745 typedef struct MDB_db {
746 uint32_t md_pad; /**< also ksize for LEAF2 pages */
747 uint16_t md_flags; /**< @ref mdb_dbi_open */
748 uint16_t md_depth; /**< depth of this tree */
749 pgno_t md_branch_pages; /**< number of internal pages */
750 pgno_t md_leaf_pages; /**< number of leaf pages */
751 pgno_t md_overflow_pages; /**< number of overflow pages */
752 size_t md_entries; /**< number of data items */
753 pgno_t md_root; /**< the root page of this tree */
756 /** mdb_dbi_open flags */
757 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
758 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
759 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
760 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
762 /** Handle for the DB used to track free pages. */
764 /** Handle for the default DB. */
767 /** Meta page content. */
768 typedef struct MDB_meta {
769 /** Stamp identifying this as an MDB file. It must be set
772 /** Version number of this lock file. Must be set to #MDB_VERSION. */
774 void *mm_address; /**< address for fixed mapping */
775 size_t mm_mapsize; /**< size of mmap region */
776 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
777 /** The size of pages used in this DB */
778 #define mm_psize mm_dbs[0].md_pad
779 /** Any persistent environment flags. @ref mdb_env */
780 #define mm_flags mm_dbs[0].md_flags
781 pgno_t mm_last_pg; /**< last used page in file */
782 txnid_t mm_txnid; /**< txnid that committed this page */
785 /** Buffer for a stack-allocated dirty page.
786 * The members define size and alignment, and silence type
787 * aliasing warnings. They are not used directly; that could
788 * mean incorrectly using several union members in parallel.
790 typedef union MDB_pagebuf {
791 char mb_raw[MDB_PAGESIZE];
794 char mm_pad[PAGEHDRSZ];
799 /** Auxiliary DB info.
800 * The information here is mostly static/read-only. There is
801 * only a single copy of this record in the environment.
803 typedef struct MDB_dbx {
804 MDB_val md_name; /**< name of the database */
805 MDB_cmp_func *md_cmp; /**< function for comparing keys */
806 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
807 MDB_rel_func *md_rel; /**< user relocate function */
808 void *md_relctx; /**< user-provided context for md_rel */
811 /** A database transaction.
812 * Every operation requires a transaction handle.
815 MDB_txn *mt_parent; /**< parent of a nested txn */
816 MDB_txn *mt_child; /**< nested txn under this txn */
817 pgno_t mt_next_pgno; /**< next unallocated page */
818 /** The ID of this transaction. IDs are integers incrementing from 1.
819 * Only committed write transactions increment the ID. If a transaction
820 * aborts, the ID may be re-used by the next writer.
823 MDB_env *mt_env; /**< the DB environment */
824 /** The list of pages that became unused during this transaction.
828 MDB_ID2L dirty_list; /**< for write txns: modified pages */
829 MDB_reader *reader; /**< this thread's reader table slot or NULL */
831 /** Array of records for each DB known in the environment. */
833 /** Array of MDB_db records for each known DB */
835 /** @defgroup mt_dbflag Transaction DB Flags
839 #define DB_DIRTY 0x01 /**< DB was written in this txn */
840 #define DB_STALE 0x02 /**< DB record is older than txnID */
841 #define DB_NEW 0x04 /**< DB handle opened in this txn */
842 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
844 /** In write txns, array of cursors for each DB */
845 MDB_cursor **mt_cursors;
846 /** Array of flags for each DB */
847 unsigned char *mt_dbflags;
848 /** Number of DB records in use. This number only ever increments;
849 * we don't decrement it when individual DB handles are closed.
853 /** @defgroup mdb_txn Transaction Flags
857 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
858 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
859 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
861 unsigned int mt_flags; /**< @ref mdb_txn */
862 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
863 unsigned int mt_dirty_room;
864 /** Tracks which of the two meta pages was used at the start
865 * of this transaction.
867 unsigned int mt_toggle;
870 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
871 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
872 * raise this on a 64 bit machine.
874 #define CURSOR_STACK 32
878 /** Cursors are used for all DB operations */
880 /** Next cursor on this DB in this txn */
882 /** Backup of the original cursor if this cursor is a shadow */
883 MDB_cursor *mc_backup;
884 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
885 struct MDB_xcursor *mc_xcursor;
886 /** The transaction that owns this cursor */
888 /** The database handle this cursor operates on */
890 /** The database record for this cursor */
892 /** The database auxiliary record for this cursor */
894 /** The @ref mt_dbflag for this database */
895 unsigned char *mc_dbflag;
896 unsigned short mc_snum; /**< number of pushed pages */
897 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
898 /** @defgroup mdb_cursor Cursor Flags
900 * Cursor state flags.
903 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
904 #define C_EOF 0x02 /**< No more data */
905 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
906 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
907 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
909 unsigned int mc_flags; /**< @ref mdb_cursor */
910 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
911 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
914 /** Context for sorted-dup records.
915 * We could have gone to a fully recursive design, with arbitrarily
916 * deep nesting of sub-databases. But for now we only handle these
917 * levels - main DB, optional sub-DB, sorted-duplicate DB.
919 typedef struct MDB_xcursor {
920 /** A sub-cursor for traversing the Dup DB */
921 MDB_cursor mx_cursor;
922 /** The database record for this Dup DB */
924 /** The auxiliary DB record for this Dup DB */
926 /** The @ref mt_dbflag for this Dup DB */
927 unsigned char mx_dbflag;
930 /** State of FreeDB old pages, stored in the MDB_env */
931 typedef struct MDB_pgstate {
932 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
933 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
936 /** The database environment. */
938 HANDLE me_fd; /**< The main data file */
939 HANDLE me_lfd; /**< The lock file */
940 HANDLE me_mfd; /**< just for writing the meta pages */
941 /** Failed to update the meta page. Probably an I/O error. */
942 #define MDB_FATAL_ERROR 0x80000000U
943 /** Some fields are initialized. */
944 #define MDB_ENV_ACTIVE 0x20000000U
945 /** me_txkey is set */
946 #define MDB_ENV_TXKEY 0x10000000U
947 uint32_t me_flags; /**< @ref mdb_env */
948 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
949 unsigned int me_maxreaders; /**< size of the reader table */
950 unsigned int me_numreaders; /**< max numreaders set by this env */
951 MDB_dbi me_numdbs; /**< number of DBs opened */
952 MDB_dbi me_maxdbs; /**< size of the DB table */
953 pid_t me_pid; /**< process ID of this env */
954 char *me_path; /**< path to the DB files */
955 char *me_map; /**< the memory map of the data file */
956 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
957 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
958 MDB_txn *me_txn; /**< current write transaction */
959 size_t me_mapsize; /**< size of the data memory map */
960 off_t me_size; /**< current file size */
961 pgno_t me_maxpg; /**< me_mapsize / me_psize */
962 MDB_dbx *me_dbxs; /**< array of static DB info */
963 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
964 pthread_key_t me_txkey; /**< thread-key for readers */
965 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
966 # define me_pglast me_pgstate.mf_pglast
967 # define me_pghead me_pgstate.mf_pghead
968 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
969 /** IDL of pages that became unused in a write txn */
971 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
972 MDB_ID2L me_dirty_list;
973 /** Max number of freelist items that can fit in a single overflow page */
975 /** Max size of a node on a page */
976 unsigned int me_nodemax;
978 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
980 #elif defined(MDB_USE_POSIX_SEM)
981 sem_t *me_rmutex; /* Shared mutexes are not supported */
986 /** Nested transaction */
987 typedef struct MDB_ntxn {
988 MDB_txn mnt_txn; /* the transaction */
989 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
992 /** max number of pages to commit in one writev() call */
993 #define MDB_COMMIT_PAGES 64
994 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
995 #undef MDB_COMMIT_PAGES
996 #define MDB_COMMIT_PAGES IOV_MAX
999 /* max bytes to write in one call */
1000 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1002 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1003 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1004 static int mdb_page_touch(MDB_cursor *mc);
1006 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1007 static int mdb_page_search_root(MDB_cursor *mc,
1008 MDB_val *key, int modify);
1009 #define MDB_PS_MODIFY 1
1010 #define MDB_PS_ROOTONLY 2
1011 static int mdb_page_search(MDB_cursor *mc,
1012 MDB_val *key, int flags);
1013 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1015 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1016 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1017 pgno_t newpgno, unsigned int nflags);
1019 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1020 static int mdb_env_pick_meta(const MDB_env *env);
1021 static int mdb_env_write_meta(MDB_txn *txn);
1022 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1023 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1025 static void mdb_env_close0(MDB_env *env, int excl);
1027 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1028 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1029 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1030 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1031 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1032 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1033 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1034 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1035 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1037 static int mdb_rebalance(MDB_cursor *mc);
1038 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1040 static void mdb_cursor_pop(MDB_cursor *mc);
1041 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1043 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1044 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1045 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1046 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1047 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1049 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1050 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1052 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1053 static void mdb_xcursor_init0(MDB_cursor *mc);
1054 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1056 static int mdb_drop0(MDB_cursor *mc, int subs);
1057 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1060 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1064 static SECURITY_DESCRIPTOR mdb_null_sd;
1065 static SECURITY_ATTRIBUTES mdb_all_sa;
1066 static int mdb_sec_inited;
1069 /** Return the library version info. */
1071 mdb_version(int *major, int *minor, int *patch)
1073 if (major) *major = MDB_VERSION_MAJOR;
1074 if (minor) *minor = MDB_VERSION_MINOR;
1075 if (patch) *patch = MDB_VERSION_PATCH;
1076 return MDB_VERSION_STRING;
1079 /** Table of descriptions for MDB @ref errors */
1080 static char *const mdb_errstr[] = {
1081 "MDB_KEYEXIST: Key/data pair already exists",
1082 "MDB_NOTFOUND: No matching key/data pair found",
1083 "MDB_PAGE_NOTFOUND: Requested page not found",
1084 "MDB_CORRUPTED: Located page was wrong type",
1085 "MDB_PANIC: Update of meta page failed",
1086 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1087 "MDB_INVALID: File is not an MDB file",
1088 "MDB_MAP_FULL: Environment mapsize limit reached",
1089 "MDB_DBS_FULL: Environment maxdbs limit reached",
1090 "MDB_READERS_FULL: Environment maxreaders limit reached",
1091 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1092 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1093 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1094 "MDB_PAGE_FULL: Internal error - page has no more space",
1095 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1096 "MDB_INCOMPATIBLE: Database flags changed or would change",
1097 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1101 mdb_strerror(int err)
1105 return ("Successful return: 0");
1107 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1108 i = err - MDB_KEYEXIST;
1109 return mdb_errstr[i];
1112 return strerror(err);
1116 /** Display a key in hexadecimal and return the address of the result.
1117 * @param[in] key the key to display
1118 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1119 * @return The key in hexadecimal form.
1122 mdb_dkey(MDB_val *key, char *buf)
1125 unsigned char *c = key->mv_data;
1131 if (key->mv_size > MDB_MAXKEYSIZE)
1132 return "MDB_MAXKEYSIZE";
1133 /* may want to make this a dynamic check: if the key is mostly
1134 * printable characters, print it as-is instead of converting to hex.
1138 for (i=0; i<key->mv_size; i++)
1139 ptr += sprintf(ptr, "%02x", *c++);
1141 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1146 /** Display all the keys in the page. */
1148 mdb_page_list(MDB_page *mp)
1151 unsigned int i, nkeys, nsize;
1155 nkeys = NUMKEYS(mp);
1156 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1157 for (i=0; i<nkeys; i++) {
1158 node = NODEPTR(mp, i);
1159 key.mv_size = node->mn_ksize;
1160 key.mv_data = node->mn_data;
1161 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1162 if (IS_BRANCH(mp)) {
1163 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1166 if (F_ISSET(node->mn_flags, F_BIGDATA))
1167 nsize += sizeof(pgno_t);
1169 nsize += NODEDSZ(node);
1170 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1176 mdb_cursor_chk(MDB_cursor *mc)
1182 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1183 for (i=0; i<mc->mc_top; i++) {
1185 node = NODEPTR(mp, mc->mc_ki[i]);
1186 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1189 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1195 /** Count all the pages in each DB and in the freelist
1196 * and make sure it matches the actual number of pages
1199 static void mdb_audit(MDB_txn *txn)
1203 MDB_ID freecount, count;
1208 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1209 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1210 freecount += *(MDB_ID *)data.mv_data;
1213 for (i = 0; i<txn->mt_numdbs; i++) {
1215 mdb_cursor_init(&mc, txn, i, &mx);
1216 if (txn->mt_dbs[i].md_root == P_INVALID)
1218 count += txn->mt_dbs[i].md_branch_pages +
1219 txn->mt_dbs[i].md_leaf_pages +
1220 txn->mt_dbs[i].md_overflow_pages;
1221 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1222 mdb_page_search(&mc, NULL, 0);
1226 mp = mc.mc_pg[mc.mc_top];
1227 for (j=0; j<NUMKEYS(mp); j++) {
1228 MDB_node *leaf = NODEPTR(mp, j);
1229 if (leaf->mn_flags & F_SUBDATA) {
1231 memcpy(&db, NODEDATA(leaf), sizeof(db));
1232 count += db.md_branch_pages + db.md_leaf_pages +
1233 db.md_overflow_pages;
1237 while (mdb_cursor_sibling(&mc, 1) == 0);
1240 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1241 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1242 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1248 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1250 return txn->mt_dbxs[dbi].md_cmp(a, b);
1254 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1256 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1259 /** Allocate a page.
1260 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1263 mdb_page_malloc(MDB_txn *txn, unsigned num)
1265 MDB_env *env = txn->mt_env;
1266 MDB_page *ret = env->me_dpages;
1267 size_t sz = env->me_psize;
1270 VGMEMP_ALLOC(env, ret, sz);
1271 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1272 env->me_dpages = ret->mp_next;
1278 if ((ret = malloc(sz)) != NULL) {
1279 VGMEMP_ALLOC(env, ret, sz);
1284 /** Free a single page.
1285 * Saves single pages to a list, for future reuse.
1286 * (This is not used for multi-page overflow pages.)
1289 mdb_page_free(MDB_env *env, MDB_page *mp)
1291 mp->mp_next = env->me_dpages;
1292 VGMEMP_FREE(env, mp);
1293 env->me_dpages = mp;
1296 /* Free a dirty page */
1298 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1300 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1301 mdb_page_free(env, dp);
1303 /* large pages just get freed directly */
1304 VGMEMP_FREE(env, dp);
1309 /* Return all dirty pages to dpage list */
1311 mdb_dlist_free(MDB_txn *txn)
1313 MDB_env *env = txn->mt_env;
1314 MDB_ID2L dl = txn->mt_u.dirty_list;
1315 unsigned i, n = dl[0].mid;
1317 for (i = 1; i <= n; i++) {
1318 mdb_dpage_free(env, dl[i].mptr);
1323 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1325 mdb_find_oldest(MDB_txn *txn)
1328 txnid_t mr, oldest = txn->mt_txnid - 1;
1329 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1330 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1340 /** Allocate pages for writing.
1341 * If there are free pages available from older transactions, they
1342 * will be re-used first. Otherwise a new page will be allocated.
1343 * @param[in] mc cursor A cursor handle identifying the transaction and
1344 * database for which we are allocating.
1345 * @param[in] num the number of pages to allocate.
1346 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1347 * will always be satisfied by a single contiguous chunk of memory.
1348 * @return 0 on success, non-zero on failure.
1351 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1353 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1354 /* Get at most <Max_retries> more freeDB records once me_pghead
1355 * has enough pages. If not enough, use new pages from the map.
1356 * If <Paranoid> and mc is updating the freeDB, only get new
1357 * records if me_pghead is empty. Then the freelist cannot play
1358 * catch-up with itself by growing while trying to save it.
1360 enum { Paranoid = 1, Max_retries = 500 };
1362 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1364 int rc, n2 = num-1, retry = Max_retries;
1365 MDB_txn *txn = mc->mc_txn;
1366 MDB_env *env = txn->mt_env;
1367 pgno_t pgno, *mop = env->me_pghead;
1368 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1371 txnid_t oldest = 0, last;
1374 int (*insert)(MDB_ID2L, MDB_ID2 *);
1378 /* If our dirty list is already full, we can't do anything */
1379 if (txn->mt_dirty_room == 0)
1380 return MDB_TXN_FULL;
1382 for (op = MDB_FIRST;; op = MDB_NEXT) {
1385 pgno_t *idl, old_id, new_id;
1387 /* Seek a big enough contiguous page range. Prefer
1388 * pages at the tail, just truncating the list.
1390 if (mop_len >= (unsigned)num) {
1394 if (mop[i-n2] == pgno+n2)
1396 } while (--i >= (unsigned)num);
1397 if (Max_retries < INT_MAX && --retry < 0)
1401 if (op == MDB_FIRST) { /* 1st iteration */
1402 /* Prepare to fetch more and coalesce */
1403 oldest = mdb_find_oldest(txn);
1404 last = env->me_pglast;
1405 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1408 key.mv_data = &last; /* will loop up last+1 */
1409 key.mv_size = sizeof(last);
1411 if (Paranoid && mc->mc_dbi == FREE_DBI)
1414 if (Paranoid && retry < 0 && mop_len)
1418 /* Do not fetch more if the record will be too recent */
1421 rc = mdb_cursor_get(&m2, &key, NULL, op);
1423 if (rc == MDB_NOTFOUND)
1427 last = *(txnid_t*)key.mv_data;
1430 np = m2.mc_pg[m2.mc_top];
1431 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1432 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1435 idl = (MDB_ID *) data.mv_data;
1438 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1441 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1443 mop = env->me_pghead;
1445 env->me_pglast = last;
1447 DPRINTF("IDL read txn %zu root %zu num %u",
1448 last, txn->mt_dbs[FREE_DBI].md_root, i);
1450 DPRINTF("IDL %zu", idl[k]);
1452 /* Merge in descending sorted order */
1455 mop[0] = (pgno_t)-1;
1459 for (; old_id < new_id; old_id = mop[--j])
1466 /* Use new pages from the map when nothing suitable in the freeDB */
1468 pgno = txn->mt_next_pgno;
1469 if (pgno + num >= env->me_maxpg) {
1470 DPUTS("DB size maxed out");
1471 return MDB_MAP_FULL;
1475 if (env->me_flags & MDB_WRITEMAP) {
1476 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1477 insert = mdb_mid2l_append;
1479 if (!(np = mdb_page_malloc(txn, num)))
1481 insert = mdb_mid2l_insert;
1484 mop[0] = mop_len -= num;
1485 /* Move any stragglers down */
1486 for (j = i-num; j < mop_len; )
1487 mop[++j] = mop[++i];
1489 txn->mt_next_pgno = pgno + num;
1491 mid.mid = np->mp_pgno = pgno;
1493 insert(txn->mt_u.dirty_list, &mid);
1494 txn->mt_dirty_room--;
1500 /** Copy the used portions of a non-overflow page.
1501 * @param[in] dst page to copy into
1502 * @param[in] src page to copy from
1503 * @param[in] psize size of a page
1506 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1508 enum { Align = sizeof(pgno_t) };
1509 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1511 /* If page isn't full, just copy the used portion. Adjust
1512 * alignment so memcpy may copy words instead of bytes.
1514 if ((unused &= -Align) && !IS_LEAF2(src)) {
1516 memcpy(dst, src, (lower + (Align-1)) & -Align);
1517 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1520 memcpy(dst, src, psize - unused);
1524 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1525 * @param[in] mc cursor pointing to the page to be touched
1526 * @return 0 on success, non-zero on failure.
1529 mdb_page_touch(MDB_cursor *mc)
1531 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1532 MDB_txn *txn = mc->mc_txn;
1533 MDB_cursor *m2, *m3;
1538 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1539 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1540 (rc = mdb_page_alloc(mc, 1, &np)))
1543 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi,mp->mp_pgno,pgno);
1544 assert(mp->mp_pgno != pgno);
1545 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1546 /* Update the parent page, if any, to point to the new page */
1548 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1549 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1550 SETPGNO(node, pgno);
1552 mc->mc_db->md_root = pgno;
1554 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1555 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1557 /* If txn has a parent, make sure the page is in our
1561 unsigned x = mdb_mid2l_search(dl, pgno);
1562 if (x <= dl[0].mid && dl[x].mid == pgno) {
1565 mc->mc_pg[mc->mc_top] = np;
1569 assert(dl[0].mid < MDB_IDL_UM_MAX);
1571 np = mdb_page_malloc(txn, 1);
1576 mdb_mid2l_insert(dl, &mid);
1581 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1583 np->mp_flags |= P_DIRTY;
1585 /* Adjust cursors pointing to mp */
1586 mc->mc_pg[mc->mc_top] = np;
1588 if (mc->mc_flags & C_SUB) {
1590 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1591 m3 = &m2->mc_xcursor->mx_cursor;
1592 if (m3->mc_snum < mc->mc_snum) continue;
1593 if (m3->mc_pg[mc->mc_top] == mp)
1594 m3->mc_pg[mc->mc_top] = np;
1597 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1598 if (m2->mc_snum < mc->mc_snum) continue;
1599 if (m2->mc_pg[mc->mc_top] == mp) {
1600 m2->mc_pg[mc->mc_top] = np;
1601 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1602 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1604 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1605 if (!(leaf->mn_flags & F_SUBDATA))
1606 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1615 mdb_env_sync(MDB_env *env, int force)
1618 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1619 if (env->me_flags & MDB_WRITEMAP) {
1620 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1621 ? MS_ASYNC : MS_SYNC;
1622 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1625 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1629 if (MDB_FDATASYNC(env->me_fd))
1636 /** Back up parent txn's cursors, then grab the originals for tracking */
1638 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1640 MDB_cursor *mc, *bk;
1645 for (i = src->mt_numdbs; --i >= 0; ) {
1646 if ((mc = src->mt_cursors[i]) != NULL) {
1647 size = sizeof(MDB_cursor);
1649 size += sizeof(MDB_xcursor);
1650 for (; mc; mc = bk->mc_next) {
1656 mc->mc_db = &dst->mt_dbs[i];
1657 /* Kill pointers into src - and dst to reduce abuse: The
1658 * user may not use mc until dst ends. Otherwise we'd...
1660 mc->mc_txn = NULL; /* ...set this to dst */
1661 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
1662 if ((mx = mc->mc_xcursor) != NULL) {
1663 *(MDB_xcursor *)(bk+1) = *mx;
1664 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
1666 mc->mc_next = dst->mt_cursors[i];
1667 dst->mt_cursors[i] = mc;
1674 /** Close this write txn's cursors, give parent txn's cursors back to parent.
1675 * @param[in] txn the transaction handle.
1676 * @param[in] merge true to keep changes to parent cursors, false to revert.
1677 * @return 0 on success, non-zero on failure.
1680 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1682 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
1686 for (i = txn->mt_numdbs; --i >= 0; ) {
1687 for (mc = cursors[i]; mc; mc = next) {
1689 if ((bk = mc->mc_backup) != NULL) {
1691 /* Commit changes to parent txn */
1692 mc->mc_next = bk->mc_next;
1693 mc->mc_backup = bk->mc_backup;
1694 mc->mc_txn = bk->mc_txn;
1695 mc->mc_db = bk->mc_db;
1696 mc->mc_dbflag = bk->mc_dbflag;
1697 if ((mx = mc->mc_xcursor) != NULL)
1698 mx->mx_cursor.mc_txn = bk->mc_txn;
1700 /* Abort nested txn */
1702 if ((mx = mc->mc_xcursor) != NULL)
1703 *mx = *(MDB_xcursor *)(bk+1);
1713 #ifdef MDB_DEBUG_SKIP
1714 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
1717 mdb_txn_reset0(MDB_txn *txn, const char *act);
1719 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1720 * @param[in] txn the transaction handle to initialize
1721 * @return 0 on success, non-zero on failure.
1724 mdb_txn_renew0(MDB_txn *txn)
1726 MDB_env *env = txn->mt_env;
1729 int rc, new_notls = 0;
1732 txn->mt_numdbs = env->me_numdbs;
1733 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1735 if (txn->mt_flags & MDB_TXN_RDONLY) {
1736 if (!env->me_txns) {
1737 i = mdb_env_pick_meta(env);
1738 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1739 txn->mt_u.reader = NULL;
1741 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1742 pthread_getspecific(env->me_txkey);
1744 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1745 return MDB_BAD_RSLOT;
1747 pid_t pid = env->me_pid;
1748 pthread_t tid = pthread_self();
1751 for (i=0; i<env->me_txns->mti_numreaders; i++)
1752 if (env->me_txns->mti_readers[i].mr_pid == 0)
1754 if (i == env->me_maxreaders) {
1755 UNLOCK_MUTEX_R(env);
1756 return MDB_READERS_FULL;
1758 env->me_txns->mti_readers[i].mr_pid = pid;
1759 env->me_txns->mti_readers[i].mr_tid = tid;
1760 if (i >= env->me_txns->mti_numreaders)
1761 env->me_txns->mti_numreaders = i+1;
1762 /* Save numreaders for un-mutexed mdb_env_close() */
1763 env->me_numreaders = env->me_txns->mti_numreaders;
1764 UNLOCK_MUTEX_R(env);
1765 r = &env->me_txns->mti_readers[i];
1766 new_notls = (env->me_flags & MDB_NOTLS);
1767 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1772 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1773 txn->mt_u.reader = r;
1775 txn->mt_toggle = txn->mt_txnid & 1;
1779 txn->mt_txnid = env->me_txns->mti_txnid;
1780 txn->mt_toggle = txn->mt_txnid & 1;
1783 if (txn->mt_txnid == mdb_debug_start)
1786 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1787 txn->mt_u.dirty_list = env->me_dirty_list;
1788 txn->mt_u.dirty_list[0].mid = 0;
1789 txn->mt_free_pgs = env->me_free_pgs;
1790 txn->mt_free_pgs[0] = 0;
1794 /* Copy the DB info and flags */
1795 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1797 /* Moved to here to avoid a data race in read TXNs */
1798 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1800 for (i=2; i<txn->mt_numdbs; i++) {
1801 x = env->me_dbflags[i];
1802 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1803 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1805 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1807 if (env->me_maxpg < txn->mt_next_pgno) {
1808 mdb_txn_reset0(txn, "renew0-mapfail");
1810 txn->mt_u.reader->mr_pid = 0;
1811 txn->mt_u.reader = NULL;
1813 return MDB_MAP_RESIZED;
1820 mdb_txn_renew(MDB_txn *txn)
1824 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1827 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1828 DPUTS("environment had fatal error, must shutdown!");
1832 rc = mdb_txn_renew0(txn);
1833 if (rc == MDB_SUCCESS) {
1834 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1835 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1836 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1842 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1846 int rc, size, tsize = sizeof(MDB_txn);
1848 if (env->me_flags & MDB_FATAL_ERROR) {
1849 DPUTS("environment had fatal error, must shutdown!");
1852 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1855 /* Nested transactions: Max 1 child, write txns only, no writemap */
1856 if (parent->mt_child ||
1857 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1858 (env->me_flags & MDB_WRITEMAP))
1862 tsize = sizeof(MDB_ntxn);
1864 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1865 if (!(flags & MDB_RDONLY))
1866 size += env->me_maxdbs * sizeof(MDB_cursor *);
1868 if ((txn = calloc(1, size)) == NULL) {
1869 DPRINTF("calloc: %s", strerror(ErrCode()));
1872 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1873 if (flags & MDB_RDONLY) {
1874 txn->mt_flags |= MDB_TXN_RDONLY;
1875 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1877 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1878 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1884 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1885 if (!txn->mt_u.dirty_list ||
1886 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1888 free(txn->mt_u.dirty_list);
1892 txn->mt_txnid = parent->mt_txnid;
1893 txn->mt_toggle = parent->mt_toggle;
1894 txn->mt_dirty_room = parent->mt_dirty_room;
1895 txn->mt_u.dirty_list[0].mid = 0;
1896 txn->mt_next_pgno = parent->mt_next_pgno;
1897 parent->mt_child = txn;
1898 txn->mt_parent = parent;
1899 txn->mt_numdbs = parent->mt_numdbs;
1900 txn->mt_flags = parent->mt_flags;
1901 txn->mt_dbxs = parent->mt_dbxs;
1902 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1903 /* Copy parent's mt_dbflags, but clear DB_NEW */
1904 for (i=0; i<txn->mt_numdbs; i++)
1905 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1907 ntxn = (MDB_ntxn *)txn;
1908 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1909 if (env->me_pghead) {
1910 size = MDB_IDL_SIZEOF(env->me_pghead);
1911 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1913 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1918 rc = mdb_cursor_shadow(parent, txn);
1920 mdb_txn_reset0(txn, "beginchild-fail");
1922 rc = mdb_txn_renew0(txn);
1928 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1929 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1930 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1936 /** Export or close DBI handles opened in this txn. */
1938 mdb_dbis_update(MDB_txn *txn, int keep)
1941 MDB_dbi n = txn->mt_numdbs;
1942 MDB_env *env = txn->mt_env;
1943 unsigned char *tdbflags = txn->mt_dbflags;
1945 for (i = n; --i >= 2;) {
1946 if (tdbflags[i] & DB_NEW) {
1948 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
1950 char *ptr = env->me_dbxs[i].md_name.mv_data;
1951 env->me_dbxs[i].md_name.mv_data = NULL;
1952 env->me_dbxs[i].md_name.mv_size = 0;
1953 env->me_dbflags[i] = 0;
1958 if (keep && env->me_numdbs < n)
1962 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1963 * May be called twice for readonly txns: First reset it, then abort.
1964 * @param[in] txn the transaction handle to reset
1967 mdb_txn_reset0(MDB_txn *txn, const char *act)
1969 MDB_env *env = txn->mt_env;
1971 /* Close any DBI handles opened in this txn */
1972 mdb_dbis_update(txn, 0);
1974 DPRINTF("%s txn %zu%c %p on mdbenv %p, root page %zu",
1975 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1976 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1978 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1979 if (txn->mt_u.reader) {
1980 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1981 if (!(env->me_flags & MDB_NOTLS))
1982 txn->mt_u.reader = NULL; /* txn does not own reader */
1984 txn->mt_numdbs = 0; /* close nothing if called again */
1985 txn->mt_dbxs = NULL; /* mark txn as reset */
1987 mdb_cursors_close(txn, 0);
1989 if (!(env->me_flags & MDB_WRITEMAP)) {
1990 mdb_dlist_free(txn);
1992 mdb_midl_free(env->me_pghead);
1994 if (txn->mt_parent) {
1995 txn->mt_parent->mt_child = NULL;
1996 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
1997 mdb_midl_free(txn->mt_free_pgs);
1998 free(txn->mt_u.dirty_list);
2002 if (mdb_midl_shrink(&txn->mt_free_pgs))
2003 env->me_free_pgs = txn->mt_free_pgs;
2004 env->me_pghead = NULL;
2008 /* The writer mutex was locked in mdb_txn_begin. */
2009 UNLOCK_MUTEX_W(env);
2014 mdb_txn_reset(MDB_txn *txn)
2019 /* This call is only valid for read-only txns */
2020 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2023 mdb_txn_reset0(txn, "reset");
2027 mdb_txn_abort(MDB_txn *txn)
2033 mdb_txn_abort(txn->mt_child);
2035 mdb_txn_reset0(txn, "abort");
2036 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2037 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2038 txn->mt_u.reader->mr_pid = 0;
2043 /** Save the freelist as of this transaction to the freeDB.
2044 * This changes the freelist. Keep trying until it stabilizes.
2047 mdb_freelist_save(MDB_txn *txn)
2049 /* env->me_pghead[] can grow and shrink during this call.
2050 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2051 * Page numbers cannot disappear from txn->mt_free_pgs[].
2054 MDB_env *env = txn->mt_env;
2055 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2056 txnid_t pglast = 0, head_id = 0;
2057 pgno_t freecnt = 0, *free_pgs, *mop;
2058 ssize_t head_room = 0, total_room = 0, mop_len;
2060 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2062 if (env->me_pghead) {
2063 /* Make sure first page of freeDB is touched and on freelist */
2064 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2065 if (rc && rc != MDB_NOTFOUND)
2070 /* Come back here after each Put() in case freelist changed */
2073 /* If using records from freeDB which we have not yet
2074 * deleted, delete them and any we reserved for me_pghead.
2076 while (pglast < env->me_pglast) {
2077 rc = mdb_cursor_first(&mc, &key, NULL);
2080 pglast = head_id = *(txnid_t *)key.mv_data;
2081 total_room = head_room = 0;
2082 assert(pglast <= env->me_pglast);
2083 rc = mdb_cursor_del(&mc, 0);
2088 /* Save the IDL of pages freed by this txn, to a single record */
2089 if (freecnt < txn->mt_free_pgs[0]) {
2091 /* Make sure last page of freeDB is touched and on freelist */
2092 key.mv_size = MDB_MAXKEYSIZE+1;
2094 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2095 if (rc && rc != MDB_NOTFOUND)
2098 free_pgs = txn->mt_free_pgs;
2099 /* Write to last page of freeDB */
2100 key.mv_size = sizeof(txn->mt_txnid);
2101 key.mv_data = &txn->mt_txnid;
2103 freecnt = free_pgs[0];
2104 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2105 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2108 /* Retry if mt_free_pgs[] grew during the Put() */
2109 free_pgs = txn->mt_free_pgs;
2110 } while (freecnt < free_pgs[0]);
2111 mdb_midl_sort(free_pgs);
2112 memcpy(data.mv_data, free_pgs, data.mv_size);
2115 unsigned int i = free_pgs[0];
2116 DPRINTF("IDL write txn %zu root %zu num %u",
2117 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2119 DPRINTF("IDL %zu", free_pgs[i]);
2125 mop = env->me_pghead;
2126 mop_len = mop ? mop[0] : 0;
2128 /* Reserve records for me_pghead[]. Split it if multi-page,
2129 * to avoid searching freeDB for a page range. Use keys in
2130 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2132 if (total_room >= mop_len) {
2133 if (total_room == mop_len || --more < 0)
2135 } else if (head_room >= maxfree_1pg && head_id > 1) {
2136 /* Keep current record (overflow page), add a new one */
2140 /* (Re)write {key = head_id, IDL length = head_room} */
2141 total_room -= head_room;
2142 head_room = mop_len - total_room;
2143 if (head_room > maxfree_1pg && head_id > 1) {
2144 /* Overflow multi-page for part of me_pghead */
2145 head_room /= head_id; /* amortize page sizes */
2146 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2147 } else if (head_room < 0) {
2148 /* Rare case, not bothering to delete this record */
2151 key.mv_size = sizeof(head_id);
2152 key.mv_data = &head_id;
2153 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2154 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2157 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2158 total_room += head_room;
2161 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2162 * so they cannot rearrange anything, just read the destinations.
2169 rc = mdb_cursor_first(&mc, &key, &data);
2170 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2171 MDB_IDL dest = data.mv_data;
2172 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2174 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2178 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2179 if (! (mop_len -= len))
2186 /** Flush dirty pages to the map, after clearing their dirty flag.
2189 mdb_page_flush(MDB_txn *txn)
2191 MDB_env *env = txn->mt_env;
2192 MDB_ID2L dl = txn->mt_u.dirty_list;
2193 unsigned psize = env->me_psize;
2194 int i, pagecount = dl[0].mid, rc;
2195 size_t size = 0, pos = 0;
2197 MDB_page *dp = NULL;
2201 struct iovec iov[MDB_COMMIT_PAGES];
2202 ssize_t wpos = 0, wsize = 0, wres;
2203 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2207 if (env->me_flags & MDB_WRITEMAP) {
2208 /* Clear dirty flags */
2209 for (i = pagecount; i; i--) {
2211 dp->mp_flags &= ~P_DIRTY;
2217 /* Write the pages */
2219 if (i <= pagecount) {
2222 /* clear dirty flag */
2223 dp->mp_flags &= ~P_DIRTY;
2226 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2231 /* Windows actually supports scatter/gather I/O, but only on
2232 * unbuffered file handles. Since we're relying on the OS page
2233 * cache for all our data, that's self-defeating. So we just
2234 * write pages one at a time. We use the ov structure to set
2235 * the write offset, to at least save the overhead of a Seek
2238 DPRINTF("committing page %zu", pgno);
2239 memset(&ov, 0, sizeof(ov));
2240 ov.Offset = pos & 0xffffffff;
2241 ov.OffsetHigh = pos >> 16 >> 16;
2242 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2244 DPRINTF("WriteFile: %d", rc);
2248 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2249 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2251 /* Write previous page(s) */
2252 #ifdef MDB_USE_PWRITEV
2253 wres = pwritev(env->me_fd, iov, n, wpos);
2256 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2258 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2260 DPRINTF("lseek: %s", strerror(rc));
2263 wres = writev(env->me_fd, iov, n);
2266 if (wres != wsize) {
2269 DPRINTF("Write error: %s", strerror(rc));
2271 rc = EIO; /* TODO: Use which error code? */
2272 DPUTS("short write, filesystem full?");
2283 DPRINTF("committing page %zu", pgno);
2284 next_pos = pos + size;
2285 iov[n].iov_len = size;
2286 iov[n].iov_base = (char *)dp;
2292 mdb_dlist_free(txn);
2298 mdb_txn_commit(MDB_txn *txn)
2304 assert(txn != NULL);
2305 assert(txn->mt_env != NULL);
2307 if (txn->mt_child) {
2308 rc = mdb_txn_commit(txn->mt_child);
2309 txn->mt_child = NULL;
2316 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2317 mdb_dbis_update(txn, 1);
2318 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2323 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2324 DPUTS("error flag is set, can't commit");
2326 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2331 if (txn->mt_parent) {
2332 MDB_txn *parent = txn->mt_parent;
2336 /* Append our free list to parent's */
2337 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2340 mdb_midl_free(txn->mt_free_pgs);
2342 parent->mt_next_pgno = txn->mt_next_pgno;
2343 parent->mt_flags = txn->mt_flags;
2345 /* Merge our cursors into parent's and close them */
2346 mdb_cursors_close(txn, 1);
2348 /* Update parent's DB table. */
2349 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2350 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2351 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2352 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2353 for (i=2; i<txn->mt_numdbs; i++) {
2354 /* preserve parent's DB_NEW status */
2355 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2356 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2359 dst = txn->mt_parent->mt_u.dirty_list;
2360 src = txn->mt_u.dirty_list;
2361 /* Find len = length of merging our dirty list with parent's */
2363 dst[0].mid = 0; /* simplify loops */
2364 if (parent->mt_parent) {
2365 len = x + src[0].mid;
2366 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2367 for (i = x; y && i; y--) {
2368 pgno_t yp = src[y].mid;
2369 while (yp < dst[i].mid)
2371 if (yp == dst[i].mid) {
2376 } else { /* Simplify the above for single-ancestor case */
2377 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2379 /* Merge our dirty list with parent's */
2381 for (i = len; y; dst[i--] = src[y--]) {
2382 pgno_t yp = src[y].mid;
2383 while (yp < dst[x].mid)
2384 dst[i--] = dst[x--];
2385 if (yp == dst[x].mid)
2386 free(dst[x--].mptr);
2390 free(txn->mt_u.dirty_list);
2391 parent->mt_dirty_room = txn->mt_dirty_room;
2393 txn->mt_parent->mt_child = NULL;
2394 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2399 if (txn != env->me_txn) {
2400 DPUTS("attempt to commit unknown transaction");
2405 mdb_cursors_close(txn, 0);
2407 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2410 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2411 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2413 /* Update DB root pointers */
2414 if (txn->mt_numdbs > 2) {
2418 data.mv_size = sizeof(MDB_db);
2420 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2421 for (i = 2; i < txn->mt_numdbs; i++) {
2422 if (txn->mt_dbflags[i] & DB_DIRTY) {
2423 data.mv_data = &txn->mt_dbs[i];
2424 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2431 rc = mdb_freelist_save(txn);
2435 mdb_midl_free(env->me_pghead);
2436 env->me_pghead = NULL;
2437 if (mdb_midl_shrink(&txn->mt_free_pgs))
2438 env->me_free_pgs = txn->mt_free_pgs;
2444 if ((rc = mdb_page_flush(txn)) ||
2445 (rc = mdb_env_sync(env, 0)) ||
2446 (rc = mdb_env_write_meta(txn)))
2452 mdb_dbis_update(txn, 1);
2454 UNLOCK_MUTEX_W(env);
2464 /** Read the environment parameters of a DB environment before
2465 * mapping it into memory.
2466 * @param[in] env the environment handle
2467 * @param[out] meta address of where to store the meta information
2468 * @return 0 on success, non-zero on failure.
2471 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2478 /* We don't know the page size yet, so use a minimum value.
2479 * Read both meta pages so we can use the latest one.
2482 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2486 memset(&ov, 0, sizeof(ov));
2488 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2490 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2492 if (rc != MDB_PAGESIZE) {
2493 if (rc == 0 && off == 0)
2495 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2496 DPRINTF("read: %s", mdb_strerror(rc));
2500 p = (MDB_page *)&pbuf;
2502 if (!F_ISSET(p->mp_flags, P_META)) {
2503 DPRINTF("page %zu not a meta page", p->mp_pgno);
2508 if (m->mm_magic != MDB_MAGIC) {
2509 DPUTS("meta has invalid magic");
2513 if (m->mm_version != MDB_VERSION) {
2514 DPRINTF("database is version %u, expected version %u",
2515 m->mm_version, MDB_VERSION);
2516 return MDB_VERSION_MISMATCH;
2519 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2525 /** Write the environment parameters of a freshly created DB environment.
2526 * @param[in] env the environment handle
2527 * @param[out] meta address of where to store the meta information
2528 * @return 0 on success, non-zero on failure.
2531 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2537 DPUTS("writing new meta page");
2539 GET_PAGESIZE(psize);
2541 meta->mm_magic = MDB_MAGIC;
2542 meta->mm_version = MDB_VERSION;
2543 meta->mm_mapsize = env->me_mapsize;
2544 meta->mm_psize = psize;
2545 meta->mm_last_pg = 1;
2546 meta->mm_flags = env->me_flags & 0xffff;
2547 meta->mm_flags |= MDB_INTEGERKEY;
2548 meta->mm_dbs[0].md_root = P_INVALID;
2549 meta->mm_dbs[1].md_root = P_INVALID;
2551 p = calloc(2, psize);
2553 p->mp_flags = P_META;
2554 *(MDB_meta *)METADATA(p) = *meta;
2556 q = (MDB_page *)((char *)p + psize);
2558 q->mp_flags = P_META;
2559 *(MDB_meta *)METADATA(q) = *meta;
2565 memset(&ov, 0, sizeof(ov));
2566 rc = WriteFile(env->me_fd, p, psize * 2, &len, &ov);
2567 rc = rc ? (len == psize * 2 ? MDB_SUCCESS : EIO) : ErrCode();
2570 rc = pwrite(env->me_fd, p, psize * 2, 0);
2571 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
2577 /** Update the environment info to commit a transaction.
2578 * @param[in] txn the transaction that's being committed
2579 * @return 0 on success, non-zero on failure.
2582 mdb_env_write_meta(MDB_txn *txn)
2585 MDB_meta meta, metab, *mp;
2587 int rc, len, toggle;
2596 assert(txn != NULL);
2597 assert(txn->mt_env != NULL);
2599 toggle = !txn->mt_toggle;
2600 DPRINTF("writing meta page %d for root page %zu",
2601 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2604 mp = env->me_metas[toggle];
2606 if (env->me_flags & MDB_WRITEMAP) {
2607 /* Persist any increases of mapsize config */
2608 if (env->me_mapsize > mp->mm_mapsize)
2609 mp->mm_mapsize = env->me_mapsize;
2610 mp->mm_dbs[0] = txn->mt_dbs[0];
2611 mp->mm_dbs[1] = txn->mt_dbs[1];
2612 mp->mm_last_pg = txn->mt_next_pgno - 1;
2613 mp->mm_txnid = txn->mt_txnid;
2614 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2615 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2618 ptr += env->me_psize;
2619 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2626 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2627 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2629 ptr = (char *)&meta;
2630 if (env->me_mapsize > mp->mm_mapsize) {
2631 /* Persist any increases of mapsize config */
2632 meta.mm_mapsize = env->me_mapsize;
2633 off = offsetof(MDB_meta, mm_mapsize);
2635 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2637 len = sizeof(MDB_meta) - off;
2640 meta.mm_dbs[0] = txn->mt_dbs[0];
2641 meta.mm_dbs[1] = txn->mt_dbs[1];
2642 meta.mm_last_pg = txn->mt_next_pgno - 1;
2643 meta.mm_txnid = txn->mt_txnid;
2646 off += env->me_psize;
2649 /* Write to the SYNC fd */
2650 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2651 env->me_fd : env->me_mfd;
2654 memset(&ov, 0, sizeof(ov));
2656 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
2660 rc = pwrite(mfd, ptr, len, off);
2663 rc = rc < 0 ? ErrCode() : EIO;
2664 DPUTS("write failed, disk error?");
2665 /* On a failure, the pagecache still contains the new data.
2666 * Write some old data back, to prevent it from being used.
2667 * Use the non-SYNC fd; we know it will fail anyway.
2669 meta.mm_last_pg = metab.mm_last_pg;
2670 meta.mm_txnid = metab.mm_txnid;
2672 memset(&ov, 0, sizeof(ov));
2674 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2676 r2 = pwrite(env->me_fd, ptr, len, off);
2679 env->me_flags |= MDB_FATAL_ERROR;
2683 /* Memory ordering issues are irrelevant; since the entire writer
2684 * is wrapped by wmutex, all of these changes will become visible
2685 * after the wmutex is unlocked. Since the DB is multi-version,
2686 * readers will get consistent data regardless of how fresh or
2687 * how stale their view of these values is.
2689 env->me_txns->mti_txnid = txn->mt_txnid;
2694 /** Check both meta pages to see which one is newer.
2695 * @param[in] env the environment handle
2696 * @return meta toggle (0 or 1).
2699 mdb_env_pick_meta(const MDB_env *env)
2701 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2705 mdb_env_create(MDB_env **env)
2709 e = calloc(1, sizeof(MDB_env));
2713 e->me_maxreaders = DEFAULT_READERS;
2714 e->me_maxdbs = e->me_numdbs = 2;
2715 e->me_fd = INVALID_HANDLE_VALUE;
2716 e->me_lfd = INVALID_HANDLE_VALUE;
2717 e->me_mfd = INVALID_HANDLE_VALUE;
2718 #ifdef MDB_USE_POSIX_SEM
2719 e->me_rmutex = SEM_FAILED;
2720 e->me_wmutex = SEM_FAILED;
2722 e->me_pid = getpid();
2723 VGMEMP_CREATE(e,0,0);
2729 mdb_env_set_mapsize(MDB_env *env, size_t size)
2733 env->me_mapsize = size;
2735 env->me_maxpg = env->me_mapsize / env->me_psize;
2740 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2744 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2749 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2751 if (env->me_map || readers < 1)
2753 env->me_maxreaders = readers;
2758 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2760 if (!env || !readers)
2762 *readers = env->me_maxreaders;
2766 /** Further setup required for opening an MDB environment
2769 mdb_env_open2(MDB_env *env)
2771 unsigned int flags = env->me_flags;
2779 memset(&meta, 0, sizeof(meta));
2781 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2784 DPUTS("new mdbenv");
2788 /* Was a mapsize configured? */
2789 if (!env->me_mapsize) {
2790 /* If this is a new environment, take the default,
2791 * else use the size recorded in the existing env.
2793 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2794 } else if (env->me_mapsize < meta.mm_mapsize) {
2795 /* If the configured size is smaller, make sure it's
2796 * still big enough. Silently round up to minimum if not.
2798 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2799 if (env->me_mapsize < minsize)
2800 env->me_mapsize = minsize;
2807 LONG sizelo, sizehi;
2808 sizelo = env->me_mapsize & 0xffffffff;
2809 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
2810 /* Windows won't create mappings for zero length files.
2811 * Just allocate the maxsize right now.
2814 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
2815 || !SetEndOfFile(env->me_fd)
2816 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
2819 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2820 PAGE_READWRITE : PAGE_READONLY,
2821 sizehi, sizelo, NULL);
2824 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2825 FILE_MAP_WRITE : FILE_MAP_READ,
2826 0, 0, env->me_mapsize, meta.mm_address);
2827 rc = env->me_map ? 0 : ErrCode();
2835 if (flags & MDB_WRITEMAP) {
2837 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2840 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2842 if (env->me_map == MAP_FAILED) {
2846 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2848 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2850 #ifdef POSIX_MADV_RANDOM
2851 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2852 #endif /* POSIX_MADV_RANDOM */
2853 #endif /* MADV_RANDOM */
2857 if (flags & MDB_FIXEDMAP)
2858 meta.mm_address = env->me_map;
2859 i = mdb_env_init_meta(env, &meta);
2860 if (i != MDB_SUCCESS) {
2863 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2864 /* Can happen because the address argument to mmap() is just a
2865 * hint. mmap() can pick another, e.g. if the range is in use.
2866 * The MAP_FIXED flag would prevent that, but then mmap could
2867 * instead unmap existing pages to make room for the new map.
2869 return EBUSY; /* TODO: Make a new MDB_* error code? */
2871 env->me_psize = meta.mm_psize;
2872 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2873 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2875 env->me_maxpg = env->me_mapsize / env->me_psize;
2877 p = (MDB_page *)env->me_map;
2878 env->me_metas[0] = METADATA(p);
2879 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2883 int toggle = mdb_env_pick_meta(env);
2884 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2886 DPRINTF("opened database version %u, pagesize %u",
2887 env->me_metas[0]->mm_version, env->me_psize);
2888 DPRINTF("using meta page %d", toggle);
2889 DPRINTF("depth: %u", db->md_depth);
2890 DPRINTF("entries: %zu", db->md_entries);
2891 DPRINTF("branch pages: %zu", db->md_branch_pages);
2892 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2893 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2894 DPRINTF("root: %zu", db->md_root);
2902 /** Release a reader thread's slot in the reader lock table.
2903 * This function is called automatically when a thread exits.
2904 * @param[in] ptr This points to the slot in the reader lock table.
2907 mdb_env_reader_dest(void *ptr)
2909 MDB_reader *reader = ptr;
2915 /** Junk for arranging thread-specific callbacks on Windows. This is
2916 * necessarily platform and compiler-specific. Windows supports up
2917 * to 1088 keys. Let's assume nobody opens more than 64 environments
2918 * in a single process, for now. They can override this if needed.
2920 #ifndef MAX_TLS_KEYS
2921 #define MAX_TLS_KEYS 64
2923 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2924 static int mdb_tls_nkeys;
2926 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2930 case DLL_PROCESS_ATTACH: break;
2931 case DLL_THREAD_ATTACH: break;
2932 case DLL_THREAD_DETACH:
2933 for (i=0; i<mdb_tls_nkeys; i++) {
2934 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2935 mdb_env_reader_dest(r);
2938 case DLL_PROCESS_DETACH: break;
2943 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2945 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2949 /* Force some symbol references.
2950 * _tls_used forces the linker to create the TLS directory if not already done
2951 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2953 #pragma comment(linker, "/INCLUDE:_tls_used")
2954 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2955 #pragma const_seg(".CRT$XLB")
2956 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2957 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2960 #pragma comment(linker, "/INCLUDE:__tls_used")
2961 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2962 #pragma data_seg(".CRT$XLB")
2963 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2965 #endif /* WIN 32/64 */
2966 #endif /* !__GNUC__ */
2969 /** Downgrade the exclusive lock on the region back to shared */
2971 mdb_env_share_locks(MDB_env *env, int *excl)
2973 int rc = 0, toggle = mdb_env_pick_meta(env);
2975 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2980 /* First acquire a shared lock. The Unlock will
2981 * then release the existing exclusive lock.
2983 memset(&ov, 0, sizeof(ov));
2984 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2987 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2993 struct flock lock_info;
2994 /* The shared lock replaces the existing lock */
2995 memset((void *)&lock_info, 0, sizeof(lock_info));
2996 lock_info.l_type = F_RDLCK;
2997 lock_info.l_whence = SEEK_SET;
2998 lock_info.l_start = 0;
2999 lock_info.l_len = 1;
3000 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3001 (rc = ErrCode()) == EINTR) ;
3002 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3009 /** Try to get exlusive lock, otherwise shared.
3010 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3013 mdb_env_excl_lock(MDB_env *env, int *excl)
3017 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3021 memset(&ov, 0, sizeof(ov));
3022 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3029 struct flock lock_info;
3030 memset((void *)&lock_info, 0, sizeof(lock_info));
3031 lock_info.l_type = F_WRLCK;
3032 lock_info.l_whence = SEEK_SET;
3033 lock_info.l_start = 0;
3034 lock_info.l_len = 1;
3035 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3036 (rc = ErrCode()) == EINTR) ;
3040 # ifdef MDB_USE_POSIX_SEM
3041 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3044 lock_info.l_type = F_RDLCK;
3045 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3046 (rc = ErrCode()) == EINTR) ;
3054 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3056 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3058 * @(#) $Revision: 5.1 $
3059 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3060 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3062 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3066 * Please do not copyright this code. This code is in the public domain.
3068 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3069 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3070 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3071 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3072 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3073 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3074 * PERFORMANCE OF THIS SOFTWARE.
3077 * chongo <Landon Curt Noll> /\oo/\
3078 * http://www.isthe.com/chongo/
3080 * Share and Enjoy! :-)
3083 typedef unsigned long long mdb_hash_t;
3084 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3086 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3087 * @param[in] str string to hash
3088 * @param[in] hval initial value for hash
3089 * @return 64 bit hash
3091 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3092 * hval arg on the first call.
3095 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3097 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3098 unsigned char *end = s + val->mv_size;
3100 * FNV-1a hash each octet of the string
3103 /* xor the bottom with the current octet */
3104 hval ^= (mdb_hash_t)*s++;
3106 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3107 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3108 (hval << 7) + (hval << 8) + (hval << 40);
3110 /* return our new hash value */
3114 /** Hash the string and output the hash in hex.
3115 * @param[in] str string to hash
3116 * @param[out] hexbuf an array of 17 chars to hold the hash
3119 mdb_hash_hex(MDB_val *val, char *hexbuf)
3122 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3123 for (i=0; i<8; i++) {
3124 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3130 /** Open and/or initialize the lock region for the environment.
3131 * @param[in] env The MDB environment.
3132 * @param[in] lpath The pathname of the file used for the lock region.
3133 * @param[in] mode The Unix permissions for the file, if we create it.
3134 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3135 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3136 * @return 0 on success, non-zero on failure.
3139 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3142 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3144 # define MDB_ERRCODE_ROFS EROFS
3145 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3146 # define MDB_CLOEXEC O_CLOEXEC
3149 # define MDB_CLOEXEC 0
3156 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3157 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3158 FILE_ATTRIBUTE_NORMAL, NULL);
3160 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3162 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3164 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3169 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3170 /* Lose record locks when exec*() */
3171 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3172 fcntl(env->me_lfd, F_SETFD, fdflags);
3175 if (!(env->me_flags & MDB_NOTLS)) {
3176 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3179 env->me_flags |= MDB_ENV_TXKEY;
3181 /* Windows TLS callbacks need help finding their TLS info. */
3182 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3186 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3190 /* Try to get exclusive lock. If we succeed, then
3191 * nobody is using the lock region and we should initialize it.
3193 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3196 size = GetFileSize(env->me_lfd, NULL);
3198 size = lseek(env->me_lfd, 0, SEEK_END);
3199 if (size == -1) goto fail_errno;
3201 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3202 if (size < rsize && *excl > 0) {
3204 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3205 || !SetEndOfFile(env->me_lfd))
3208 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3212 size = rsize - sizeof(MDB_txninfo);
3213 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3218 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3220 if (!mh) goto fail_errno;
3221 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3223 if (!env->me_txns) goto fail_errno;
3225 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3227 if (m == MAP_FAILED) goto fail_errno;
3233 BY_HANDLE_FILE_INFORMATION stbuf;
3242 if (!mdb_sec_inited) {
3243 InitializeSecurityDescriptor(&mdb_null_sd,
3244 SECURITY_DESCRIPTOR_REVISION);
3245 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3246 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3247 mdb_all_sa.bInheritHandle = FALSE;
3248 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3251 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3252 idbuf.volume = stbuf.dwVolumeSerialNumber;
3253 idbuf.nhigh = stbuf.nFileIndexHigh;
3254 idbuf.nlow = stbuf.nFileIndexLow;
3255 val.mv_data = &idbuf;
3256 val.mv_size = sizeof(idbuf);
3257 mdb_hash_hex(&val, hexbuf);
3258 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3259 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3260 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3261 if (!env->me_rmutex) goto fail_errno;
3262 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3263 if (!env->me_wmutex) goto fail_errno;
3264 #elif defined(MDB_USE_POSIX_SEM)
3273 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3274 idbuf.dev = stbuf.st_dev;
3275 idbuf.ino = stbuf.st_ino;
3276 val.mv_data = &idbuf;
3277 val.mv_size = sizeof(idbuf);
3278 mdb_hash_hex(&val, hexbuf);
3279 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3280 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3281 /* Clean up after a previous run, if needed: Try to
3282 * remove both semaphores before doing anything else.
3284 sem_unlink(env->me_txns->mti_rmname);
3285 sem_unlink(env->me_txns->mti_wmname);
3286 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3287 O_CREAT|O_EXCL, mode, 1);
3288 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3289 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3290 O_CREAT|O_EXCL, mode, 1);
3291 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3292 #else /* MDB_USE_POSIX_SEM */
3293 pthread_mutexattr_t mattr;
3295 if ((rc = pthread_mutexattr_init(&mattr))
3296 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3297 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3298 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3300 pthread_mutexattr_destroy(&mattr);
3301 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3303 env->me_txns->mti_version = MDB_VERSION;
3304 env->me_txns->mti_magic = MDB_MAGIC;
3305 env->me_txns->mti_txnid = 0;
3306 env->me_txns->mti_numreaders = 0;
3309 if (env->me_txns->mti_magic != MDB_MAGIC) {
3310 DPUTS("lock region has invalid magic");
3314 if (env->me_txns->mti_version != MDB_VERSION) {
3315 DPRINTF("lock region is version %u, expected version %u",
3316 env->me_txns->mti_version, MDB_VERSION);
3317 rc = MDB_VERSION_MISMATCH;
3321 if (rc && rc != EACCES && rc != EAGAIN) {
3325 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3326 if (!env->me_rmutex) goto fail_errno;
3327 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3328 if (!env->me_wmutex) goto fail_errno;
3329 #elif defined(MDB_USE_POSIX_SEM)
3330 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3331 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3332 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3333 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3344 /** The name of the lock file in the DB environment */
3345 #define LOCKNAME "/lock.mdb"
3346 /** The name of the data file in the DB environment */
3347 #define DATANAME "/data.mdb"
3348 /** The suffix of the lock file when no subdir is used */
3349 #define LOCKSUFF "-lock"
3350 /** Only a subset of the @ref mdb_env flags can be changed
3351 * at runtime. Changing other flags requires closing the
3352 * environment and re-opening it with the new flags.
3354 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3355 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3358 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3360 int oflags, rc, len, excl = -1;
3361 char *lpath, *dpath;
3363 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3367 if (flags & MDB_NOSUBDIR) {
3368 rc = len + sizeof(LOCKSUFF) + len + 1;
3370 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3375 if (flags & MDB_NOSUBDIR) {
3376 dpath = lpath + len + sizeof(LOCKSUFF);
3377 sprintf(lpath, "%s" LOCKSUFF, path);
3378 strcpy(dpath, path);
3380 dpath = lpath + len + sizeof(LOCKNAME);
3381 sprintf(lpath, "%s" LOCKNAME, path);
3382 sprintf(dpath, "%s" DATANAME, path);
3386 flags |= env->me_flags;
3387 if (flags & MDB_RDONLY) {
3388 /* silently ignore WRITEMAP when we're only getting read access */
3389 flags &= ~MDB_WRITEMAP;
3391 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3392 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3395 env->me_flags = flags |= MDB_ENV_ACTIVE;
3399 env->me_path = strdup(path);
3400 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3401 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3402 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3407 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3412 if (F_ISSET(flags, MDB_RDONLY)) {
3413 oflags = GENERIC_READ;
3414 len = OPEN_EXISTING;
3416 oflags = GENERIC_READ|GENERIC_WRITE;
3419 mode = FILE_ATTRIBUTE_NORMAL;
3420 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3421 NULL, len, mode, NULL);
3423 if (F_ISSET(flags, MDB_RDONLY))
3426 oflags = O_RDWR | O_CREAT;
3428 env->me_fd = open(dpath, oflags, mode);
3430 if (env->me_fd == INVALID_HANDLE_VALUE) {
3435 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3436 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3437 env->me_mfd = env->me_fd;
3439 /* Synchronous fd for meta writes. Needed even with
3440 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3443 env->me_mfd = CreateFile(dpath, oflags,
3444 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3445 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3447 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3449 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3454 DPRINTF("opened dbenv %p", (void *) env);
3456 rc = mdb_env_share_locks(env, &excl);
3462 mdb_env_close0(env, excl);
3468 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3470 mdb_env_close0(MDB_env *env, int excl)
3474 if (!(env->me_flags & MDB_ENV_ACTIVE))
3477 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3478 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3479 free(env->me_dbxs[i].md_name.mv_data);
3481 free(env->me_dbflags);
3484 free(env->me_dirty_list);
3485 mdb_midl_free(env->me_free_pgs);
3487 if (env->me_flags & MDB_ENV_TXKEY) {
3488 pthread_key_delete(env->me_txkey);
3490 /* Delete our key from the global list */
3491 for (i=0; i<mdb_tls_nkeys; i++)
3492 if (mdb_tls_keys[i] == env->me_txkey) {
3493 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3501 munmap(env->me_map, env->me_mapsize);
3503 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3504 (void) close(env->me_mfd);
3505 if (env->me_fd != INVALID_HANDLE_VALUE)
3506 (void) close(env->me_fd);
3508 pid_t pid = env->me_pid;
3509 /* Clearing readers is done in this function because
3510 * me_txkey with its destructor must be disabled first.
3512 for (i = env->me_numreaders; --i >= 0; )
3513 if (env->me_txns->mti_readers[i].mr_pid == pid)
3514 env->me_txns->mti_readers[i].mr_pid = 0;
3516 if (env->me_rmutex) {
3517 CloseHandle(env->me_rmutex);
3518 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3520 /* Windows automatically destroys the mutexes when
3521 * the last handle closes.
3523 #elif defined(MDB_USE_POSIX_SEM)
3524 if (env->me_rmutex != SEM_FAILED) {
3525 sem_close(env->me_rmutex);
3526 if (env->me_wmutex != SEM_FAILED)
3527 sem_close(env->me_wmutex);
3528 /* If we have the filelock: If we are the
3529 * only remaining user, clean up semaphores.
3532 mdb_env_excl_lock(env, &excl);
3534 sem_unlink(env->me_txns->mti_rmname);
3535 sem_unlink(env->me_txns->mti_wmname);
3539 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3541 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3544 /* Unlock the lockfile. Windows would have unlocked it
3545 * after closing anyway, but not necessarily at once.
3547 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3550 (void) close(env->me_lfd);
3553 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3557 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3559 MDB_txn *txn = NULL;
3564 /* Do the lock/unlock of the reader mutex before starting the
3565 * write txn. Otherwise other read txns could block writers.
3567 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3572 /* We must start the actual read txn after blocking writers */
3573 mdb_txn_reset0(txn, "reset-stage1");
3575 /* Temporarily block writers until we snapshot the meta pages */
3578 rc = mdb_txn_renew0(txn);
3580 UNLOCK_MUTEX_W(env);
3585 wsize = env->me_psize * 2;
3589 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3590 rc = rc ? (len == wsize ? MDB_SUCCESS : EIO) : ErrCode();
3593 rc = write(fd, env->me_map, wsize);
3594 rc = rc == (int)wsize ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
3597 UNLOCK_MUTEX_W(env);
3602 ptr = env->me_map + wsize;
3603 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3607 if (wsize > MAX_WRITE)
3611 rc = WriteFile(fd, ptr, w2, &len, NULL);
3612 rc = rc ? (len == w2 ? MDB_SUCCESS : EIO) : ErrCode();
3621 if (wsize > MAX_WRITE)
3625 wres = write(fd, ptr, w2);
3626 rc = wres == (ssize_t)w2 ? MDB_SUCCESS : wres < 0 ? ErrCode() : EIO;
3639 mdb_env_copy(MDB_env *env, const char *path)
3643 HANDLE newfd = INVALID_HANDLE_VALUE;
3645 if (env->me_flags & MDB_NOSUBDIR) {
3646 lpath = (char *)path;
3649 len += sizeof(DATANAME);
3650 lpath = malloc(len);
3653 sprintf(lpath, "%s" DATANAME, path);
3656 /* The destination path must exist, but the destination file must not.
3657 * We don't want the OS to cache the writes, since the source data is
3658 * already in the OS cache.
3661 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3662 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3664 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3670 if (newfd == INVALID_HANDLE_VALUE) {
3675 #ifdef F_NOCACHE /* __APPLE__ */
3676 rc = fcntl(newfd, F_NOCACHE, 1);
3683 rc = mdb_env_copyfd(env, newfd);
3686 if (!(env->me_flags & MDB_NOSUBDIR))
3688 if (newfd != INVALID_HANDLE_VALUE)
3689 if (close(newfd) < 0 && rc == MDB_SUCCESS)
3696 mdb_env_close(MDB_env *env)
3703 VGMEMP_DESTROY(env);
3704 while ((dp = env->me_dpages) != NULL) {
3705 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3706 env->me_dpages = dp->mp_next;
3710 mdb_env_close0(env, 0);
3714 /** Compare two items pointing at aligned size_t's */
3716 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3718 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3719 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3722 /** Compare two items pointing at aligned int's */
3724 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3726 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3727 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3730 /** Compare two items pointing at ints of unknown alignment.
3731 * Nodes and keys are guaranteed to be 2-byte aligned.
3734 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3736 #if BYTE_ORDER == LITTLE_ENDIAN
3737 unsigned short *u, *c;
3740 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3741 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3744 } while(!x && u > (unsigned short *)a->mv_data);
3747 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3751 /** Compare two items lexically */
3753 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3760 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3766 diff = memcmp(a->mv_data, b->mv_data, len);
3767 return diff ? diff : len_diff<0 ? -1 : len_diff;
3770 /** Compare two items in reverse byte order */
3772 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3774 const unsigned char *p1, *p2, *p1_lim;
3778 p1_lim = (const unsigned char *)a->mv_data;
3779 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3780 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3782 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3788 while (p1 > p1_lim) {
3789 diff = *--p1 - *--p2;
3793 return len_diff<0 ? -1 : len_diff;
3796 /** Search for key within a page, using binary search.
3797 * Returns the smallest entry larger or equal to the key.
3798 * If exactp is non-null, stores whether the found entry was an exact match
3799 * in *exactp (1 or 0).
3800 * Updates the cursor index with the index of the found entry.
3801 * If no entry larger or equal to the key is found, returns NULL.
3804 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3806 unsigned int i = 0, nkeys;
3809 MDB_page *mp = mc->mc_pg[mc->mc_top];
3810 MDB_node *node = NULL;
3815 nkeys = NUMKEYS(mp);
3820 COPY_PGNO(pgno, mp->mp_pgno);
3821 DPRINTF("searching %u keys in %s %spage %zu",
3822 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3829 low = IS_LEAF(mp) ? 0 : 1;
3831 cmp = mc->mc_dbx->md_cmp;
3833 /* Branch pages have no data, so if using integer keys,
3834 * alignment is guaranteed. Use faster mdb_cmp_int.
3836 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3837 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3844 nodekey.mv_size = mc->mc_db->md_pad;
3845 node = NODEPTR(mp, 0); /* fake */
3846 while (low <= high) {
3847 i = (low + high) >> 1;
3848 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3849 rc = cmp(key, &nodekey);
3850 DPRINTF("found leaf index %u [%s], rc = %i",
3851 i, DKEY(&nodekey), rc);
3860 while (low <= high) {
3861 i = (low + high) >> 1;
3863 node = NODEPTR(mp, i);
3864 nodekey.mv_size = NODEKSZ(node);
3865 nodekey.mv_data = NODEKEY(node);
3867 rc = cmp(key, &nodekey);
3870 DPRINTF("found leaf index %u [%s], rc = %i",
3871 i, DKEY(&nodekey), rc);
3873 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3874 i, DKEY(&nodekey), NODEPGNO(node), rc);
3885 if (rc > 0) { /* Found entry is less than the key. */
3886 i++; /* Skip to get the smallest entry larger than key. */
3888 node = NODEPTR(mp, i);
3891 *exactp = (rc == 0);
3892 /* store the key index */
3893 mc->mc_ki[mc->mc_top] = i;
3895 /* There is no entry larger or equal to the key. */
3898 /* nodeptr is fake for LEAF2 */
3904 mdb_cursor_adjust(MDB_cursor *mc, func)
3908 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3909 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3916 /** Pop a page off the top of the cursor's stack. */
3918 mdb_cursor_pop(MDB_cursor *mc)
3921 #ifndef MDB_DEBUG_SKIP
3922 MDB_page *top = mc->mc_pg[mc->mc_top];
3928 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3929 mc->mc_dbi, (void *) mc);
3933 /** Push a page onto the top of the cursor's stack. */
3935 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3937 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3938 mc->mc_dbi, (void *) mc);
3940 if (mc->mc_snum >= CURSOR_STACK) {
3941 assert(mc->mc_snum < CURSOR_STACK);
3942 return MDB_CURSOR_FULL;
3945 mc->mc_top = mc->mc_snum++;
3946 mc->mc_pg[mc->mc_top] = mp;
3947 mc->mc_ki[mc->mc_top] = 0;
3952 /** Find the address of the page corresponding to a given page number.
3953 * @param[in] txn the transaction for this access.
3954 * @param[in] pgno the page number for the page to retrieve.
3955 * @param[out] ret address of a pointer where the page's address will be stored.
3956 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
3957 * @return 0 on success, non-zero on failure.
3960 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
3965 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3966 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3971 MDB_ID2L dl = tx2->mt_u.dirty_list;
3973 unsigned x = mdb_mid2l_search(dl, pgno);
3974 if (x <= dl[0].mid && dl[x].mid == pgno) {
3980 } while ((tx2 = tx2->mt_parent) != NULL);
3983 if (pgno < txn->mt_next_pgno) {
3985 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3987 DPRINTF("page %zu not found", pgno);
3989 return MDB_PAGE_NOTFOUND;
3999 /** Search for the page a given key should be in.
4000 * Pushes parent pages on the cursor stack. This function continues a
4001 * search on a cursor that has already been initialized. (Usually by
4002 * #mdb_page_search() but also by #mdb_node_move().)
4003 * @param[in,out] mc the cursor for this operation.
4004 * @param[in] key the key to search for. If NULL, search for the lowest
4005 * page. (This is used by #mdb_cursor_first().)
4006 * @param[in] modify If true, visited pages are updated with new page numbers.
4007 * @return 0 on success, non-zero on failure.
4010 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4012 MDB_page *mp = mc->mc_pg[mc->mc_top];
4017 while (IS_BRANCH(mp)) {
4021 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4022 assert(NUMKEYS(mp) > 1);
4023 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4025 if (key == NULL) /* Initialize cursor to first page. */
4027 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4028 /* cursor to last page */
4032 node = mdb_node_search(mc, key, &exact);
4034 i = NUMKEYS(mp) - 1;
4036 i = mc->mc_ki[mc->mc_top];
4045 DPRINTF("following index %u for key [%s]",
4047 assert(i < NUMKEYS(mp));
4048 node = NODEPTR(mp, i);
4050 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4053 mc->mc_ki[mc->mc_top] = i;
4054 if ((rc = mdb_cursor_push(mc, mp)))
4058 if ((rc = mdb_page_touch(mc)) != 0)
4060 mp = mc->mc_pg[mc->mc_top];
4065 DPRINTF("internal error, index points to a %02X page!?",
4067 return MDB_CORRUPTED;
4070 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4071 key ? DKEY(key) : NULL);
4076 /** Search for the lowest key under the current branch page.
4077 * This just bypasses a NUMKEYS check in the current page
4078 * before calling mdb_page_search_root(), because the callers
4079 * are all in situations where the current page is known to
4083 mdb_page_search_lowest(MDB_cursor *mc)
4085 MDB_page *mp = mc->mc_pg[mc->mc_top];
4086 MDB_node *node = NODEPTR(mp, 0);
4089 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4092 mc->mc_ki[mc->mc_top] = 0;
4093 if ((rc = mdb_cursor_push(mc, mp)))
4095 return mdb_page_search_root(mc, NULL, 0);
4098 /** Search for the page a given key should be in.
4099 * Pushes parent pages on the cursor stack. This function just sets up
4100 * the search; it finds the root page for \b mc's database and sets this
4101 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4102 * called to complete the search.
4103 * @param[in,out] mc the cursor for this operation.
4104 * @param[in] key the key to search for. If NULL, search for the lowest
4105 * page. (This is used by #mdb_cursor_first().)
4106 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4107 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4108 * @return 0 on success, non-zero on failure.
4111 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4116 /* Make sure the txn is still viable, then find the root from
4117 * the txn's db table.
4119 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4120 DPUTS("transaction has failed, must abort");
4123 /* Make sure we're using an up-to-date root */
4124 if (mc->mc_dbi > MAIN_DBI) {
4125 if ((*mc->mc_dbflag & DB_STALE) ||
4126 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4128 unsigned char dbflag = 0;
4129 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4130 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4133 if (*mc->mc_dbflag & DB_STALE) {
4137 MDB_node *leaf = mdb_node_search(&mc2,
4138 &mc->mc_dbx->md_name, &exact);
4140 return MDB_NOTFOUND;
4141 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4144 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4146 /* The txn may not know this DBI, or another process may
4147 * have dropped and recreated the DB with other flags.
4149 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4150 return MDB_INCOMPATIBLE;
4151 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4153 if (flags & MDB_PS_MODIFY)
4155 *mc->mc_dbflag &= ~DB_STALE;
4156 *mc->mc_dbflag |= dbflag;
4159 root = mc->mc_db->md_root;
4161 if (root == P_INVALID) { /* Tree is empty. */
4162 DPUTS("tree is empty");
4163 return MDB_NOTFOUND;
4168 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4169 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4175 DPRINTF("db %u root page %zu has flags 0x%X",
4176 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4178 if (flags & MDB_PS_MODIFY) {
4179 if ((rc = mdb_page_touch(mc)))
4183 if (flags & MDB_PS_ROOTONLY)
4186 return mdb_page_search_root(mc, key, flags);
4190 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4192 MDB_txn *txn = mc->mc_txn;
4193 pgno_t pg = mp->mp_pgno;
4194 unsigned i, ovpages = mp->mp_pages;
4195 MDB_env *env = txn->mt_env;
4198 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4199 /* If the page is dirty we just acquired it, so we should
4200 * give it back to our current free list, if any.
4201 * Not currently supported in nested txns.
4202 * Otherwise put it onto the list of pages we freed in this txn.
4204 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && env->me_pghead) {
4207 MDB_ID2 *dl, ix, iy;
4208 rc = mdb_midl_need(&env->me_pghead, ovpages);
4211 /* Remove from dirty list */
4212 dl = txn->mt_u.dirty_list;
4214 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4222 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4223 txn->mt_flags |= MDB_TXN_ERROR;
4224 return MDB_CORRUPTED;
4227 if (!(env->me_flags & MDB_WRITEMAP))
4228 mdb_dpage_free(env, mp);
4229 /* Insert in me_pghead */
4230 mop = env->me_pghead;
4231 j = mop[0] + ovpages;
4232 for (i = mop[0]; i && mop[i] < pg; i--)
4238 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4242 mc->mc_db->md_overflow_pages -= ovpages;
4246 /** Return the data associated with a given node.
4247 * @param[in] txn The transaction for this operation.
4248 * @param[in] leaf The node being read.
4249 * @param[out] data Updated to point to the node's data.
4250 * @return 0 on success, non-zero on failure.
4253 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4255 MDB_page *omp; /* overflow page */
4259 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4260 data->mv_size = NODEDSZ(leaf);
4261 data->mv_data = NODEDATA(leaf);
4265 /* Read overflow data.
4267 data->mv_size = NODEDSZ(leaf);
4268 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4269 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4270 DPRINTF("read overflow page %zu failed", pgno);
4273 data->mv_data = METADATA(omp);
4279 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4280 MDB_val *key, MDB_val *data)
4289 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4291 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4294 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4298 mdb_cursor_init(&mc, txn, dbi, &mx);
4299 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4302 /** Find a sibling for a page.
4303 * Replaces the page at the top of the cursor's stack with the
4304 * specified sibling, if one exists.
4305 * @param[in] mc The cursor for this operation.
4306 * @param[in] move_right Non-zero if the right sibling is requested,
4307 * otherwise the left sibling.
4308 * @return 0 on success, non-zero on failure.
4311 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4317 if (mc->mc_snum < 2) {
4318 return MDB_NOTFOUND; /* root has no siblings */
4322 DPRINTF("parent page is page %zu, index %u",
4323 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4325 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4326 : (mc->mc_ki[mc->mc_top] == 0)) {
4327 DPRINTF("no more keys left, moving to %s sibling",
4328 move_right ? "right" : "left");
4329 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4330 /* undo cursor_pop before returning */
4337 mc->mc_ki[mc->mc_top]++;
4339 mc->mc_ki[mc->mc_top]--;
4340 DPRINTF("just moving to %s index key %u",
4341 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4343 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4345 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4346 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4349 mdb_cursor_push(mc, mp);
4351 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4356 /** Move the cursor to the next data item. */
4358 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4364 if (mc->mc_flags & C_EOF) {
4365 return MDB_NOTFOUND;
4368 assert(mc->mc_flags & C_INITIALIZED);
4370 mp = mc->mc_pg[mc->mc_top];
4372 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4373 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4374 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4375 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4376 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4377 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4381 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4382 if (op == MDB_NEXT_DUP)
4383 return MDB_NOTFOUND;
4387 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4389 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4390 DPUTS("=====> move to next sibling page");
4391 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4392 mc->mc_flags |= C_EOF;
4393 mc->mc_flags &= ~C_INITIALIZED;
4396 mp = mc->mc_pg[mc->mc_top];
4397 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4399 mc->mc_ki[mc->mc_top]++;
4401 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4402 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4405 key->mv_size = mc->mc_db->md_pad;
4406 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4410 assert(IS_LEAF(mp));
4411 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4413 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4414 mdb_xcursor_init1(mc, leaf);
4417 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4420 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4421 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4422 if (rc != MDB_SUCCESS)
4427 MDB_GET_KEY(leaf, key);
4431 /** Move the cursor to the previous data item. */
4433 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4439 assert(mc->mc_flags & C_INITIALIZED);
4441 mp = mc->mc_pg[mc->mc_top];
4443 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4444 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4445 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4446 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4447 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4448 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4451 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4452 if (op == MDB_PREV_DUP)
4453 return MDB_NOTFOUND;
4458 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4460 if (mc->mc_ki[mc->mc_top] == 0) {
4461 DPUTS("=====> move to prev sibling page");
4462 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
4463 mc->mc_flags &= ~C_INITIALIZED;
4466 mp = mc->mc_pg[mc->mc_top];
4467 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4468 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4470 mc->mc_ki[mc->mc_top]--;
4472 mc->mc_flags &= ~C_EOF;
4474 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4475 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4478 key->mv_size = mc->mc_db->md_pad;
4479 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4483 assert(IS_LEAF(mp));
4484 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4486 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4487 mdb_xcursor_init1(mc, leaf);
4490 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4493 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4494 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4495 if (rc != MDB_SUCCESS)
4500 MDB_GET_KEY(leaf, key);
4504 /** Set the cursor on a specific data item. */
4506 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4507 MDB_cursor_op op, int *exactp)
4511 MDB_node *leaf = NULL;
4516 assert(key->mv_size > 0);
4518 /* See if we're already on the right page */
4519 if (mc->mc_flags & C_INITIALIZED) {
4522 mp = mc->mc_pg[mc->mc_top];
4524 mc->mc_ki[mc->mc_top] = 0;
4525 return MDB_NOTFOUND;
4527 if (mp->mp_flags & P_LEAF2) {
4528 nodekey.mv_size = mc->mc_db->md_pad;
4529 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4531 leaf = NODEPTR(mp, 0);
4532 MDB_GET_KEY(leaf, &nodekey);
4534 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4536 /* Probably happens rarely, but first node on the page
4537 * was the one we wanted.
4539 mc->mc_ki[mc->mc_top] = 0;
4546 unsigned int nkeys = NUMKEYS(mp);
4548 if (mp->mp_flags & P_LEAF2) {
4549 nodekey.mv_data = LEAF2KEY(mp,
4550 nkeys-1, nodekey.mv_size);
4552 leaf = NODEPTR(mp, nkeys-1);
4553 MDB_GET_KEY(leaf, &nodekey);
4555 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4557 /* last node was the one we wanted */
4558 mc->mc_ki[mc->mc_top] = nkeys-1;
4564 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4565 /* This is definitely the right page, skip search_page */
4566 if (mp->mp_flags & P_LEAF2) {
4567 nodekey.mv_data = LEAF2KEY(mp,
4568 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4570 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4571 MDB_GET_KEY(leaf, &nodekey);
4573 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4575 /* current node was the one we wanted */
4585 /* If any parents have right-sibs, search.
4586 * Otherwise, there's nothing further.
4588 for (i=0; i<mc->mc_top; i++)
4590 NUMKEYS(mc->mc_pg[i])-1)
4592 if (i == mc->mc_top) {
4593 /* There are no other pages */
4594 mc->mc_ki[mc->mc_top] = nkeys;
4595 return MDB_NOTFOUND;
4599 /* There are no other pages */
4600 mc->mc_ki[mc->mc_top] = 0;
4601 return MDB_NOTFOUND;
4605 rc = mdb_page_search(mc, key, 0);
4606 if (rc != MDB_SUCCESS)
4609 mp = mc->mc_pg[mc->mc_top];
4610 assert(IS_LEAF(mp));
4613 leaf = mdb_node_search(mc, key, exactp);
4614 if (exactp != NULL && !*exactp) {
4615 /* MDB_SET specified and not an exact match. */
4616 return MDB_NOTFOUND;
4620 DPUTS("===> inexact leaf not found, goto sibling");
4621 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4622 return rc; /* no entries matched */
4623 mp = mc->mc_pg[mc->mc_top];
4624 assert(IS_LEAF(mp));
4625 leaf = NODEPTR(mp, 0);
4629 mc->mc_flags |= C_INITIALIZED;
4630 mc->mc_flags &= ~C_EOF;
4633 key->mv_size = mc->mc_db->md_pad;
4634 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4638 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4639 mdb_xcursor_init1(mc, leaf);
4642 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4643 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4644 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4647 if (op == MDB_GET_BOTH) {
4653 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4654 if (rc != MDB_SUCCESS)
4657 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4659 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4661 rc = mc->mc_dbx->md_dcmp(data, &d2);
4663 if (op == MDB_GET_BOTH || rc > 0)
4664 return MDB_NOTFOUND;
4669 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4670 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4675 /* The key already matches in all other cases */
4676 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4677 MDB_GET_KEY(leaf, key);
4678 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4683 /** Move the cursor to the first item in the database. */
4685 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4690 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4691 rc = mdb_page_search(mc, NULL, 0);
4692 if (rc != MDB_SUCCESS)
4695 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4697 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4698 mc->mc_flags |= C_INITIALIZED;
4699 mc->mc_flags &= ~C_EOF;
4701 mc->mc_ki[mc->mc_top] = 0;
4703 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4704 key->mv_size = mc->mc_db->md_pad;
4705 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4710 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4711 mdb_xcursor_init1(mc, leaf);
4712 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4717 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4718 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4722 MDB_GET_KEY(leaf, key);
4726 /** Move the cursor to the last item in the database. */
4728 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4733 if (!(mc->mc_flags & C_EOF)) {
4735 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4738 lkey.mv_size = MDB_MAXKEYSIZE+1;
4739 lkey.mv_data = NULL;
4740 rc = mdb_page_search(mc, &lkey, 0);
4741 if (rc != MDB_SUCCESS)
4744 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4747 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4748 mc->mc_flags |= C_INITIALIZED|C_EOF;
4749 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4751 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4752 key->mv_size = mc->mc_db->md_pad;
4753 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4758 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4759 mdb_xcursor_init1(mc, leaf);
4760 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4765 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4766 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4771 MDB_GET_KEY(leaf, key);
4776 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4785 case MDB_GET_CURRENT:
4786 if (!(mc->mc_flags & C_INITIALIZED)) {
4789 MDB_page *mp = mc->mc_pg[mc->mc_top];
4791 mc->mc_ki[mc->mc_top] = 0;
4797 key->mv_size = mc->mc_db->md_pad;
4798 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4800 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4801 MDB_GET_KEY(leaf, key);
4803 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4804 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4806 rc = mdb_node_read(mc->mc_txn, leaf, data);
4813 case MDB_GET_BOTH_RANGE:
4814 if (data == NULL || mc->mc_xcursor == NULL) {
4822 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4824 } else if (op == MDB_SET_RANGE)
4825 rc = mdb_cursor_set(mc, key, data, op, NULL);
4827 rc = mdb_cursor_set(mc, key, data, op, &exact);
4829 case MDB_GET_MULTIPLE:
4831 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4832 !(mc->mc_flags & C_INITIALIZED)) {
4837 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4838 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4841 case MDB_NEXT_MULTIPLE:
4843 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4847 if (!(mc->mc_flags & C_INITIALIZED))
4848 rc = mdb_cursor_first(mc, key, data);
4850 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4851 if (rc == MDB_SUCCESS) {
4852 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4855 mx = &mc->mc_xcursor->mx_cursor;
4856 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4858 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4859 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4867 case MDB_NEXT_NODUP:
4868 if (!(mc->mc_flags & C_INITIALIZED))
4869 rc = mdb_cursor_first(mc, key, data);
4871 rc = mdb_cursor_next(mc, key, data, op);
4875 case MDB_PREV_NODUP:
4876 if (!(mc->mc_flags & C_INITIALIZED)) {
4877 rc = mdb_cursor_last(mc, key, data);
4880 mc->mc_flags |= C_INITIALIZED;
4881 mc->mc_ki[mc->mc_top]++;
4883 rc = mdb_cursor_prev(mc, key, data, op);
4886 rc = mdb_cursor_first(mc, key, data);
4890 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4891 !(mc->mc_flags & C_INITIALIZED) ||
4892 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4896 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4899 rc = mdb_cursor_last(mc, key, data);
4903 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4904 !(mc->mc_flags & C_INITIALIZED) ||
4905 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4909 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4912 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4920 /** Touch all the pages in the cursor stack.
4921 * Makes sure all the pages are writable, before attempting a write operation.
4922 * @param[in] mc The cursor to operate on.
4925 mdb_cursor_touch(MDB_cursor *mc)
4929 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4932 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
4933 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4936 *mc->mc_dbflag |= DB_DIRTY;
4938 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4939 rc = mdb_page_touch(mc);
4943 mc->mc_top = mc->mc_snum-1;
4948 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4951 MDB_node *leaf = NULL;
4952 MDB_val xdata, *rdata, dkey;
4955 int do_sub = 0, insert = 0;
4956 unsigned int mcount = 0, dcount = 0;
4960 char dbuf[MDB_MAXKEYSIZE+1];
4961 unsigned int nflags;
4964 /* Check this first so counter will always be zero on any
4967 if (flags & MDB_MULTIPLE) {
4968 dcount = data[1].mv_size;
4969 data[1].mv_size = 0;
4970 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
4974 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4977 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4980 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4983 #if SIZE_MAX > MAXDATASIZE
4984 if (data->mv_size > MAXDATASIZE)
4988 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4989 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4993 if (flags == MDB_CURRENT) {
4994 if (!(mc->mc_flags & C_INITIALIZED))
4997 } else if (mc->mc_db->md_root == P_INVALID) {
4999 /* new database, write a root leaf page */
5000 DPUTS("allocating new root leaf page");
5001 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5005 mdb_cursor_push(mc, np);
5006 mc->mc_db->md_root = np->mp_pgno;
5007 mc->mc_db->md_depth++;
5008 *mc->mc_dbflag |= DB_DIRTY;
5009 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5011 np->mp_flags |= P_LEAF2;
5012 mc->mc_flags |= C_INITIALIZED;
5018 if (flags & MDB_APPEND) {
5020 rc = mdb_cursor_last(mc, &k2, &d2);
5022 rc = mc->mc_dbx->md_cmp(key, &k2);
5025 mc->mc_ki[mc->mc_top]++;
5027 /* new key is <= last key */
5032 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5034 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5035 DPRINTF("duplicate key [%s]", DKEY(key));
5037 return MDB_KEYEXIST;
5039 if (rc && rc != MDB_NOTFOUND)
5043 /* Cursor is positioned, now make sure all pages are writable */
5044 rc2 = mdb_cursor_touch(mc);
5049 /* The key already exists */
5050 if (rc == MDB_SUCCESS) {
5051 /* there's only a key anyway, so this is a no-op */
5052 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5053 unsigned int ksize = mc->mc_db->md_pad;
5054 if (key->mv_size != ksize)
5056 if (flags == MDB_CURRENT) {
5057 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5058 memcpy(ptr, key->mv_data, ksize);
5063 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5066 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5067 /* Was a single item before, must convert now */
5069 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5070 /* Just overwrite the current item */
5071 if (flags == MDB_CURRENT)
5074 dkey.mv_size = NODEDSZ(leaf);
5075 dkey.mv_data = NODEDATA(leaf);
5076 #if UINT_MAX < SIZE_MAX
5077 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5078 #ifdef MISALIGNED_OK
5079 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5081 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5084 /* if data matches, ignore it */
5085 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5086 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5088 /* create a fake page for the dup items */
5089 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5090 dkey.mv_data = dbuf;
5091 fp = (MDB_page *)&pbuf;
5092 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5093 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5094 fp->mp_lower = PAGEHDRSZ;
5095 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5096 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5097 fp->mp_flags |= P_LEAF2;
5098 fp->mp_pad = data->mv_size;
5099 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5101 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5102 (dkey.mv_size & 1) + (data->mv_size & 1);
5104 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5107 xdata.mv_size = fp->mp_upper;
5112 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5113 /* See if we need to convert from fake page to subDB */
5115 unsigned int offset;
5119 fp = NODEDATA(leaf);
5120 if (flags == MDB_CURRENT) {
5122 fp->mp_flags |= P_DIRTY;
5123 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5124 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5128 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5129 offset = fp->mp_pad;
5130 if (SIZELEFT(fp) >= offset)
5132 offset *= 4; /* space for 4 more */
5134 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5136 offset += offset & 1;
5137 fp_flags = fp->mp_flags;
5138 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5139 offset >= mc->mc_txn->mt_env->me_nodemax) {
5140 /* yes, convert it */
5142 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5143 dummy.md_pad = fp->mp_pad;
5144 dummy.md_flags = MDB_DUPFIXED;
5145 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5146 dummy.md_flags |= MDB_INTEGERKEY;
5149 dummy.md_branch_pages = 0;
5150 dummy.md_leaf_pages = 1;
5151 dummy.md_overflow_pages = 0;
5152 dummy.md_entries = NUMKEYS(fp);
5154 xdata.mv_size = sizeof(MDB_db);
5155 xdata.mv_data = &dummy;
5156 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5158 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5159 flags |= F_DUPDATA|F_SUBDATA;
5160 dummy.md_root = mp->mp_pgno;
5161 fp_flags &= ~P_SUBP;
5163 /* no, just grow it */
5165 xdata.mv_size = NODEDSZ(leaf) + offset;
5166 xdata.mv_data = &pbuf;
5167 mp = (MDB_page *)&pbuf;
5168 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5171 mp->mp_flags = fp_flags | P_DIRTY;
5172 mp->mp_pad = fp->mp_pad;
5173 mp->mp_lower = fp->mp_lower;
5174 mp->mp_upper = fp->mp_upper + offset;
5176 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5178 nsize = NODEDSZ(leaf) - fp->mp_upper;
5179 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5180 for (i=0; i<NUMKEYS(fp); i++)
5181 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5183 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5187 /* data is on sub-DB, just store it */
5188 flags |= F_DUPDATA|F_SUBDATA;
5192 /* overflow page overwrites need special handling */
5193 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5196 unsigned psize = mc->mc_txn->mt_env->me_psize;
5197 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5199 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5200 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5202 ovpages = omp->mp_pages;
5204 /* Is the ov page writable and large enough? */
5205 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5206 /* yes, overwrite it. Note in this case we don't
5207 * bother to try shrinking the page if the new data
5208 * is smaller than the overflow threshold.
5211 /* It is writable only in a parent txn */
5212 size_t sz = (size_t) psize * ovpages, off;
5213 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5219 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5220 if (!(flags & MDB_RESERVE)) {
5221 /* Copy end of page, adjusting alignment so
5222 * compiler may copy words instead of bytes.
5224 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5225 memcpy((size_t *)((char *)np + off),
5226 (size_t *)((char *)omp + off), sz - off);
5229 memcpy(np, omp, sz); /* Copy beginning of page */
5232 SETDSZ(leaf, data->mv_size);
5233 if (F_ISSET(flags, MDB_RESERVE))
5234 data->mv_data = METADATA(omp);
5236 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5239 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5242 } else if (NODEDSZ(leaf) == data->mv_size) {
5243 /* same size, just replace it. Note that we could
5244 * also reuse this node if the new data is smaller,
5245 * but instead we opt to shrink the node in that case.
5247 if (F_ISSET(flags, MDB_RESERVE))
5248 data->mv_data = NODEDATA(leaf);
5249 else if (data->mv_size)
5250 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5252 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5255 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5256 mc->mc_db->md_entries--;
5258 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5265 nflags = flags & NODE_ADD_FLAGS;
5266 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5267 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5268 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5269 nflags &= ~MDB_APPEND;
5271 nflags |= MDB_SPLIT_REPLACE;
5272 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5274 /* There is room already in this leaf page. */
5275 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5276 if (rc == 0 && !do_sub && insert) {
5277 /* Adjust other cursors pointing to mp */
5278 MDB_cursor *m2, *m3;
5279 MDB_dbi dbi = mc->mc_dbi;
5280 unsigned i = mc->mc_top;
5281 MDB_page *mp = mc->mc_pg[i];
5283 if (mc->mc_flags & C_SUB)
5286 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5287 if (mc->mc_flags & C_SUB)
5288 m3 = &m2->mc_xcursor->mx_cursor;
5291 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5292 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5299 if (rc != MDB_SUCCESS)
5300 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5302 /* Now store the actual data in the child DB. Note that we're
5303 * storing the user data in the keys field, so there are strict
5304 * size limits on dupdata. The actual data fields of the child
5305 * DB are all zero size.
5312 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5313 if (flags & MDB_CURRENT) {
5314 xflags = MDB_CURRENT;
5316 mdb_xcursor_init1(mc, leaf);
5317 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5319 /* converted, write the original data first */
5321 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5325 /* Adjust other cursors pointing to mp */
5327 unsigned i = mc->mc_top;
5328 MDB_page *mp = mc->mc_pg[i];
5330 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5331 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5332 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5333 mdb_xcursor_init1(m2, leaf);
5337 /* we've done our job */
5340 if (flags & MDB_APPENDDUP)
5341 xflags |= MDB_APPEND;
5342 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5343 if (flags & F_SUBDATA) {
5344 void *db = NODEDATA(leaf);
5345 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5348 /* sub-writes might have failed so check rc again.
5349 * Don't increment count if we just replaced an existing item.
5351 if (!rc && !(flags & MDB_CURRENT))
5352 mc->mc_db->md_entries++;
5353 if (flags & MDB_MULTIPLE) {
5356 if (mcount < dcount) {
5357 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5358 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5362 /* let caller know how many succeeded, if any */
5363 data[1].mv_size = mcount;
5367 /* If we succeeded and the key didn't exist before, make sure
5368 * the cursor is marked valid.
5371 mc->mc_flags |= C_INITIALIZED;
5376 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5381 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5384 if (!(mc->mc_flags & C_INITIALIZED))
5387 rc = mdb_cursor_touch(mc);
5391 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5393 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5394 if (flags != MDB_NODUPDATA) {
5395 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5396 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5398 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5399 /* If sub-DB still has entries, we're done */
5400 if (mc->mc_xcursor->mx_db.md_entries) {
5401 if (leaf->mn_flags & F_SUBDATA) {
5402 /* update subDB info */
5403 void *db = NODEDATA(leaf);
5404 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5407 /* shrink fake page */
5408 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5409 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5410 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5411 /* fix other sub-DB cursors pointed at this fake page */
5412 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5413 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5414 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5415 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5416 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5419 mc->mc_db->md_entries--;
5422 /* otherwise fall thru and delete the sub-DB */
5425 if (leaf->mn_flags & F_SUBDATA) {
5426 /* add all the child DB's pages to the free list */
5427 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5428 if (rc == MDB_SUCCESS) {
5429 mc->mc_db->md_entries -=
5430 mc->mc_xcursor->mx_db.md_entries;
5435 return mdb_cursor_del0(mc, leaf);
5438 /** Allocate and initialize new pages for a database.
5439 * @param[in] mc a cursor on the database being added to.
5440 * @param[in] flags flags defining what type of page is being allocated.
5441 * @param[in] num the number of pages to allocate. This is usually 1,
5442 * unless allocating overflow pages for a large record.
5443 * @param[out] mp Address of a page, or NULL on failure.
5444 * @return 0 on success, non-zero on failure.
5447 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5452 if ((rc = mdb_page_alloc(mc, num, &np)))
5454 DPRINTF("allocated new mpage %zu, page size %u",
5455 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5456 np->mp_flags = flags | P_DIRTY;
5457 np->mp_lower = PAGEHDRSZ;
5458 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5461 mc->mc_db->md_branch_pages++;
5462 else if (IS_LEAF(np))
5463 mc->mc_db->md_leaf_pages++;
5464 else if (IS_OVERFLOW(np)) {
5465 mc->mc_db->md_overflow_pages += num;
5473 /** Calculate the size of a leaf node.
5474 * The size depends on the environment's page size; if a data item
5475 * is too large it will be put onto an overflow page and the node
5476 * size will only include the key and not the data. Sizes are always
5477 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5478 * of the #MDB_node headers.
5479 * @param[in] env The environment handle.
5480 * @param[in] key The key for the node.
5481 * @param[in] data The data for the node.
5482 * @return The number of bytes needed to store the node.
5485 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5489 sz = LEAFSIZE(key, data);
5490 if (sz >= env->me_nodemax) {
5491 /* put on overflow page */
5492 sz -= data->mv_size - sizeof(pgno_t);
5496 return sz + sizeof(indx_t);
5499 /** Calculate the size of a branch node.
5500 * The size should depend on the environment's page size but since
5501 * we currently don't support spilling large keys onto overflow
5502 * pages, it's simply the size of the #MDB_node header plus the
5503 * size of the key. Sizes are always rounded up to an even number
5504 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5505 * @param[in] env The environment handle.
5506 * @param[in] key The key for the node.
5507 * @return The number of bytes needed to store the node.
5510 mdb_branch_size(MDB_env *env, MDB_val *key)
5515 if (sz >= env->me_nodemax) {
5516 /* put on overflow page */
5517 /* not implemented */
5518 /* sz -= key->size - sizeof(pgno_t); */
5521 return sz + sizeof(indx_t);
5524 /** Add a node to the page pointed to by the cursor.
5525 * @param[in] mc The cursor for this operation.
5526 * @param[in] indx The index on the page where the new node should be added.
5527 * @param[in] key The key for the new node.
5528 * @param[in] data The data for the new node, if any.
5529 * @param[in] pgno The page number, if adding a branch node.
5530 * @param[in] flags Flags for the node.
5531 * @return 0 on success, non-zero on failure. Possible errors are:
5533 * <li>ENOMEM - failed to allocate overflow pages for the node.
5534 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5535 * should never happen since all callers already calculate the
5536 * page's free space before calling this function.
5540 mdb_node_add(MDB_cursor *mc, indx_t indx,
5541 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5544 size_t node_size = NODESIZE;
5547 MDB_page *mp = mc->mc_pg[mc->mc_top];
5548 MDB_page *ofp = NULL; /* overflow page */
5551 assert(mp->mp_upper >= mp->mp_lower);
5553 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5554 IS_LEAF(mp) ? "leaf" : "branch",
5555 IS_SUBP(mp) ? "sub-" : "",
5556 mp->mp_pgno, indx, data ? data->mv_size : 0,
5557 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5560 /* Move higher keys up one slot. */
5561 int ksize = mc->mc_db->md_pad, dif;
5562 char *ptr = LEAF2KEY(mp, indx, ksize);
5563 dif = NUMKEYS(mp) - indx;
5565 memmove(ptr+ksize, ptr, dif*ksize);
5566 /* insert new key */
5567 memcpy(ptr, key->mv_data, ksize);
5569 /* Just using these for counting */
5570 mp->mp_lower += sizeof(indx_t);
5571 mp->mp_upper -= ksize - sizeof(indx_t);
5576 node_size += key->mv_size;
5580 if (F_ISSET(flags, F_BIGDATA)) {
5581 /* Data already on overflow page. */
5582 node_size += sizeof(pgno_t);
5583 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5584 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5586 /* Put data on overflow page. */
5587 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5588 data->mv_size, node_size+data->mv_size);
5589 node_size += sizeof(pgno_t);
5590 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5592 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5595 node_size += data->mv_size;
5598 node_size += node_size & 1;
5600 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5601 DPRINTF("not enough room in page %zu, got %u ptrs",
5602 mp->mp_pgno, NUMKEYS(mp));
5603 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5604 mp->mp_upper - mp->mp_lower);
5605 DPRINTF("node size = %zu", node_size);
5606 return MDB_PAGE_FULL;
5609 /* Move higher pointers up one slot. */
5610 for (i = NUMKEYS(mp); i > indx; i--)
5611 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5613 /* Adjust free space offsets. */
5614 ofs = mp->mp_upper - node_size;
5615 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5616 mp->mp_ptrs[indx] = ofs;
5618 mp->mp_lower += sizeof(indx_t);
5620 /* Write the node data. */
5621 node = NODEPTR(mp, indx);
5622 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5623 node->mn_flags = flags;
5625 SETDSZ(node,data->mv_size);
5630 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5635 if (F_ISSET(flags, F_BIGDATA))
5636 memcpy(node->mn_data + key->mv_size, data->mv_data,
5638 else if (F_ISSET(flags, MDB_RESERVE))
5639 data->mv_data = node->mn_data + key->mv_size;
5641 memcpy(node->mn_data + key->mv_size, data->mv_data,
5644 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5646 if (F_ISSET(flags, MDB_RESERVE))
5647 data->mv_data = METADATA(ofp);
5649 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5656 /** Delete the specified node from a page.
5657 * @param[in] mp The page to operate on.
5658 * @param[in] indx The index of the node to delete.
5659 * @param[in] ksize The size of a node. Only used if the page is
5660 * part of a #MDB_DUPFIXED database.
5663 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5666 indx_t i, j, numkeys, ptr;
5673 COPY_PGNO(pgno, mp->mp_pgno);
5674 DPRINTF("delete node %u on %s page %zu", indx,
5675 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5678 assert(indx < NUMKEYS(mp));
5681 int x = NUMKEYS(mp) - 1 - indx;
5682 base = LEAF2KEY(mp, indx, ksize);
5684 memmove(base, base + ksize, x * ksize);
5685 mp->mp_lower -= sizeof(indx_t);
5686 mp->mp_upper += ksize - sizeof(indx_t);
5690 node = NODEPTR(mp, indx);
5691 sz = NODESIZE + node->mn_ksize;
5693 if (F_ISSET(node->mn_flags, F_BIGDATA))
5694 sz += sizeof(pgno_t);
5696 sz += NODEDSZ(node);
5700 ptr = mp->mp_ptrs[indx];
5701 numkeys = NUMKEYS(mp);
5702 for (i = j = 0; i < numkeys; i++) {
5704 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5705 if (mp->mp_ptrs[i] < ptr)
5706 mp->mp_ptrs[j] += sz;
5711 base = (char *)mp + mp->mp_upper;
5712 memmove(base + sz, base, ptr - mp->mp_upper);
5714 mp->mp_lower -= sizeof(indx_t);
5718 /** Compact the main page after deleting a node on a subpage.
5719 * @param[in] mp The main page to operate on.
5720 * @param[in] indx The index of the subpage on the main page.
5723 mdb_node_shrink(MDB_page *mp, indx_t indx)
5730 indx_t i, numkeys, ptr;
5732 node = NODEPTR(mp, indx);
5733 sp = (MDB_page *)NODEDATA(node);
5734 osize = NODEDSZ(node);
5736 delta = sp->mp_upper - sp->mp_lower;
5737 SETDSZ(node, osize - delta);
5738 xp = (MDB_page *)((char *)sp + delta);
5740 /* shift subpage upward */
5742 nsize = NUMKEYS(sp) * sp->mp_pad;
5743 memmove(METADATA(xp), METADATA(sp), nsize);
5746 nsize = osize - sp->mp_upper;
5747 numkeys = NUMKEYS(sp);
5748 for (i=numkeys-1; i>=0; i--)
5749 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5751 xp->mp_upper = sp->mp_lower;
5752 xp->mp_lower = sp->mp_lower;
5753 xp->mp_flags = sp->mp_flags;
5754 xp->mp_pad = sp->mp_pad;
5755 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5757 /* shift lower nodes upward */
5758 ptr = mp->mp_ptrs[indx];
5759 numkeys = NUMKEYS(mp);
5760 for (i = 0; i < numkeys; i++) {
5761 if (mp->mp_ptrs[i] <= ptr)
5762 mp->mp_ptrs[i] += delta;
5765 base = (char *)mp + mp->mp_upper;
5766 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5767 mp->mp_upper += delta;
5770 /** Initial setup of a sorted-dups cursor.
5771 * Sorted duplicates are implemented as a sub-database for the given key.
5772 * The duplicate data items are actually keys of the sub-database.
5773 * Operations on the duplicate data items are performed using a sub-cursor
5774 * initialized when the sub-database is first accessed. This function does
5775 * the preliminary setup of the sub-cursor, filling in the fields that
5776 * depend only on the parent DB.
5777 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5780 mdb_xcursor_init0(MDB_cursor *mc)
5782 MDB_xcursor *mx = mc->mc_xcursor;
5784 mx->mx_cursor.mc_xcursor = NULL;
5785 mx->mx_cursor.mc_txn = mc->mc_txn;
5786 mx->mx_cursor.mc_db = &mx->mx_db;
5787 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5788 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5789 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5790 mx->mx_cursor.mc_snum = 0;
5791 mx->mx_cursor.mc_top = 0;
5792 mx->mx_cursor.mc_flags = C_SUB;
5793 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5794 mx->mx_dbx.md_dcmp = NULL;
5795 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5798 /** Final setup of a sorted-dups cursor.
5799 * Sets up the fields that depend on the data from the main cursor.
5800 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5801 * @param[in] node The data containing the #MDB_db record for the
5802 * sorted-dup database.
5805 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5807 MDB_xcursor *mx = mc->mc_xcursor;
5809 if (node->mn_flags & F_SUBDATA) {
5810 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5811 mx->mx_cursor.mc_pg[0] = 0;
5812 mx->mx_cursor.mc_snum = 0;
5813 mx->mx_cursor.mc_flags = C_SUB;
5815 MDB_page *fp = NODEDATA(node);
5816 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5817 mx->mx_db.md_flags = 0;
5818 mx->mx_db.md_depth = 1;
5819 mx->mx_db.md_branch_pages = 0;
5820 mx->mx_db.md_leaf_pages = 1;
5821 mx->mx_db.md_overflow_pages = 0;
5822 mx->mx_db.md_entries = NUMKEYS(fp);
5823 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5824 mx->mx_cursor.mc_snum = 1;
5825 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5826 mx->mx_cursor.mc_top = 0;
5827 mx->mx_cursor.mc_pg[0] = fp;
5828 mx->mx_cursor.mc_ki[0] = 0;
5829 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5830 mx->mx_db.md_flags = MDB_DUPFIXED;
5831 mx->mx_db.md_pad = fp->mp_pad;
5832 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5833 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5836 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5838 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5840 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5841 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5842 #if UINT_MAX < SIZE_MAX
5843 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5844 #ifdef MISALIGNED_OK
5845 mx->mx_dbx.md_cmp = mdb_cmp_long;
5847 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5852 /** Initialize a cursor for a given transaction and database. */
5854 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5856 mc->mc_backup = NULL;
5859 mc->mc_db = &txn->mt_dbs[dbi];
5860 mc->mc_dbx = &txn->mt_dbxs[dbi];
5861 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5866 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5868 mc->mc_xcursor = mx;
5869 mdb_xcursor_init0(mc);
5871 mc->mc_xcursor = NULL;
5873 if (*mc->mc_dbflag & DB_STALE) {
5874 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5879 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5882 size_t size = sizeof(MDB_cursor);
5884 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5887 /* Allow read access to the freelist */
5888 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5891 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5892 size += sizeof(MDB_xcursor);
5894 if ((mc = malloc(size)) != NULL) {
5895 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5896 if (txn->mt_cursors) {
5897 mc->mc_next = txn->mt_cursors[dbi];
5898 txn->mt_cursors[dbi] = mc;
5899 mc->mc_flags |= C_UNTRACK;
5911 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5913 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5916 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5919 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5923 /* Return the count of duplicate data items for the current key */
5925 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5929 if (mc == NULL || countp == NULL)
5932 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5935 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5936 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5939 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5942 *countp = mc->mc_xcursor->mx_db.md_entries;
5948 mdb_cursor_close(MDB_cursor *mc)
5950 if (mc && !mc->mc_backup) {
5951 /* remove from txn, if tracked */
5952 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
5953 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5954 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5956 *prev = mc->mc_next;
5963 mdb_cursor_txn(MDB_cursor *mc)
5965 if (!mc) return NULL;
5970 mdb_cursor_dbi(MDB_cursor *mc)
5976 /** Replace the key for a node with a new key.
5977 * @param[in] mc Cursor pointing to the node to operate on.
5978 * @param[in] key The new key to use.
5979 * @return 0 on success, non-zero on failure.
5982 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5989 indx_t ptr, i, numkeys, indx;
5992 indx = mc->mc_ki[mc->mc_top];
5993 mp = mc->mc_pg[mc->mc_top];
5994 node = NODEPTR(mp, indx);
5995 ptr = mp->mp_ptrs[indx];
5999 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6000 k2.mv_data = NODEKEY(node);
6001 k2.mv_size = node->mn_ksize;
6002 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6004 mdb_dkey(&k2, kbuf2),
6010 delta0 = delta = key->mv_size - node->mn_ksize;
6012 /* Must be 2-byte aligned. If new key is
6013 * shorter by 1, the shift will be skipped.
6015 delta += (delta & 1);
6017 if (delta > 0 && SIZELEFT(mp) < delta) {
6019 /* not enough space left, do a delete and split */
6020 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6021 pgno = NODEPGNO(node);
6022 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6023 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6026 numkeys = NUMKEYS(mp);
6027 for (i = 0; i < numkeys; i++) {
6028 if (mp->mp_ptrs[i] <= ptr)
6029 mp->mp_ptrs[i] -= delta;
6032 base = (char *)mp + mp->mp_upper;
6033 len = ptr - mp->mp_upper + NODESIZE;
6034 memmove(base - delta, base, len);
6035 mp->mp_upper -= delta;
6037 node = NODEPTR(mp, indx);
6040 /* But even if no shift was needed, update ksize */
6042 node->mn_ksize = key->mv_size;
6045 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6051 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6053 /** Move a node from csrc to cdst.
6056 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6063 unsigned short flags;
6067 /* Mark src and dst as dirty. */
6068 if ((rc = mdb_page_touch(csrc)) ||
6069 (rc = mdb_page_touch(cdst)))
6072 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6073 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6074 key.mv_size = csrc->mc_db->md_pad;
6075 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6077 data.mv_data = NULL;
6081 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6082 assert(!((long)srcnode&1));
6083 srcpg = NODEPGNO(srcnode);
6084 flags = srcnode->mn_flags;
6085 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6086 unsigned int snum = csrc->mc_snum;
6088 /* must find the lowest key below src */
6089 mdb_page_search_lowest(csrc);
6090 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6091 key.mv_size = csrc->mc_db->md_pad;
6092 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6094 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6095 key.mv_size = NODEKSZ(s2);
6096 key.mv_data = NODEKEY(s2);
6098 csrc->mc_snum = snum--;
6099 csrc->mc_top = snum;
6101 key.mv_size = NODEKSZ(srcnode);
6102 key.mv_data = NODEKEY(srcnode);
6104 data.mv_size = NODEDSZ(srcnode);
6105 data.mv_data = NODEDATA(srcnode);
6107 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6108 unsigned int snum = cdst->mc_snum;
6111 /* must find the lowest key below dst */
6112 mdb_page_search_lowest(cdst);
6113 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6114 bkey.mv_size = cdst->mc_db->md_pad;
6115 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6117 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6118 bkey.mv_size = NODEKSZ(s2);
6119 bkey.mv_data = NODEKEY(s2);
6121 cdst->mc_snum = snum--;
6122 cdst->mc_top = snum;
6123 mdb_cursor_copy(cdst, &mn);
6125 rc = mdb_update_key(&mn, &bkey);
6130 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6131 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6132 csrc->mc_ki[csrc->mc_top],
6134 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6135 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6137 /* Add the node to the destination page.
6139 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6140 if (rc != MDB_SUCCESS)
6143 /* Delete the node from the source page.
6145 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6148 /* Adjust other cursors pointing to mp */
6149 MDB_cursor *m2, *m3;
6150 MDB_dbi dbi = csrc->mc_dbi;
6151 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6153 if (csrc->mc_flags & C_SUB)
6156 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6157 if (csrc->mc_flags & C_SUB)
6158 m3 = &m2->mc_xcursor->mx_cursor;
6161 if (m3 == csrc) continue;
6162 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6163 csrc->mc_ki[csrc->mc_top]) {
6164 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6165 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6170 /* Update the parent separators.
6172 if (csrc->mc_ki[csrc->mc_top] == 0) {
6173 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6174 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6175 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6177 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6178 key.mv_size = NODEKSZ(srcnode);
6179 key.mv_data = NODEKEY(srcnode);
6181 DPRINTF("update separator for source page %zu to [%s]",
6182 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6183 mdb_cursor_copy(csrc, &mn);
6186 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6189 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6191 indx_t ix = csrc->mc_ki[csrc->mc_top];
6192 nullkey.mv_size = 0;
6193 csrc->mc_ki[csrc->mc_top] = 0;
6194 rc = mdb_update_key(csrc, &nullkey);
6195 csrc->mc_ki[csrc->mc_top] = ix;
6196 assert(rc == MDB_SUCCESS);
6200 if (cdst->mc_ki[cdst->mc_top] == 0) {
6201 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6202 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6203 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6205 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6206 key.mv_size = NODEKSZ(srcnode);
6207 key.mv_data = NODEKEY(srcnode);
6209 DPRINTF("update separator for destination page %zu to [%s]",
6210 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6211 mdb_cursor_copy(cdst, &mn);
6214 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6217 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6219 indx_t ix = cdst->mc_ki[cdst->mc_top];
6220 nullkey.mv_size = 0;
6221 cdst->mc_ki[cdst->mc_top] = 0;
6222 rc = mdb_update_key(cdst, &nullkey);
6223 cdst->mc_ki[cdst->mc_top] = ix;
6224 assert(rc == MDB_SUCCESS);
6231 /** Merge one page into another.
6232 * The nodes from the page pointed to by \b csrc will
6233 * be copied to the page pointed to by \b cdst and then
6234 * the \b csrc page will be freed.
6235 * @param[in] csrc Cursor pointing to the source page.
6236 * @param[in] cdst Cursor pointing to the destination page.
6239 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6247 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6248 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6250 assert(csrc->mc_snum > 1); /* can't merge root page */
6251 assert(cdst->mc_snum > 1);
6253 /* Mark dst as dirty. */
6254 if ((rc = mdb_page_touch(cdst)))
6257 /* Move all nodes from src to dst.
6259 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6260 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6261 key.mv_size = csrc->mc_db->md_pad;
6262 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6263 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6264 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6265 if (rc != MDB_SUCCESS)
6267 key.mv_data = (char *)key.mv_data + key.mv_size;
6270 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6271 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6272 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6273 unsigned int snum = csrc->mc_snum;
6275 /* must find the lowest key below src */
6276 mdb_page_search_lowest(csrc);
6277 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6278 key.mv_size = csrc->mc_db->md_pad;
6279 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6281 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6282 key.mv_size = NODEKSZ(s2);
6283 key.mv_data = NODEKEY(s2);
6285 csrc->mc_snum = snum--;
6286 csrc->mc_top = snum;
6288 key.mv_size = srcnode->mn_ksize;
6289 key.mv_data = NODEKEY(srcnode);
6292 data.mv_size = NODEDSZ(srcnode);
6293 data.mv_data = NODEDATA(srcnode);
6294 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6295 if (rc != MDB_SUCCESS)
6300 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6301 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);
6303 /* Unlink the src page from parent and add to free list.
6305 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6306 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6309 rc = mdb_update_key(csrc, &key);
6315 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6316 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6319 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6320 csrc->mc_db->md_leaf_pages--;
6322 csrc->mc_db->md_branch_pages--;
6324 /* Adjust other cursors pointing to mp */
6325 MDB_cursor *m2, *m3;
6326 MDB_dbi dbi = csrc->mc_dbi;
6327 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6329 if (csrc->mc_flags & C_SUB)
6332 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6333 if (csrc->mc_flags & C_SUB)
6334 m3 = &m2->mc_xcursor->mx_cursor;
6337 if (m3 == csrc) continue;
6338 if (m3->mc_snum < csrc->mc_snum) continue;
6339 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6340 m3->mc_pg[csrc->mc_top] = mp;
6341 m3->mc_ki[csrc->mc_top] += nkeys;
6345 mdb_cursor_pop(csrc);
6347 return mdb_rebalance(csrc);
6350 /** Copy the contents of a cursor.
6351 * @param[in] csrc The cursor to copy from.
6352 * @param[out] cdst The cursor to copy to.
6355 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6359 cdst->mc_txn = csrc->mc_txn;
6360 cdst->mc_dbi = csrc->mc_dbi;
6361 cdst->mc_db = csrc->mc_db;
6362 cdst->mc_dbx = csrc->mc_dbx;
6363 cdst->mc_snum = csrc->mc_snum;
6364 cdst->mc_top = csrc->mc_top;
6365 cdst->mc_flags = csrc->mc_flags;
6367 for (i=0; i<csrc->mc_snum; i++) {
6368 cdst->mc_pg[i] = csrc->mc_pg[i];
6369 cdst->mc_ki[i] = csrc->mc_ki[i];
6373 /** Rebalance the tree after a delete operation.
6374 * @param[in] mc Cursor pointing to the page where rebalancing
6376 * @return 0 on success, non-zero on failure.
6379 mdb_rebalance(MDB_cursor *mc)
6383 unsigned int ptop, minkeys;
6386 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6390 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6391 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6392 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6393 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6397 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6398 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6401 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6402 DPRINTF("no need to rebalance page %zu, above fill threshold",
6408 if (mc->mc_snum < 2) {
6409 MDB_page *mp = mc->mc_pg[0];
6411 DPUTS("Can't rebalance a subpage, ignoring");
6414 if (NUMKEYS(mp) == 0) {
6415 DPUTS("tree is completely empty");
6416 mc->mc_db->md_root = P_INVALID;
6417 mc->mc_db->md_depth = 0;
6418 mc->mc_db->md_leaf_pages = 0;
6419 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6422 /* Adjust cursors pointing to mp */
6426 MDB_cursor *m2, *m3;
6427 MDB_dbi dbi = mc->mc_dbi;
6429 if (mc->mc_flags & C_SUB)
6432 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6433 if (mc->mc_flags & C_SUB)
6434 m3 = &m2->mc_xcursor->mx_cursor;
6437 if (m3->mc_snum < mc->mc_snum) continue;
6438 if (m3->mc_pg[0] == mp) {
6444 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6445 DPUTS("collapsing root page!");
6446 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6449 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6450 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6453 mc->mc_db->md_depth--;
6454 mc->mc_db->md_branch_pages--;
6456 /* Adjust other cursors pointing to mp */
6457 MDB_cursor *m2, *m3;
6458 MDB_dbi dbi = mc->mc_dbi;
6460 if (mc->mc_flags & C_SUB)
6463 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6464 if (mc->mc_flags & C_SUB)
6465 m3 = &m2->mc_xcursor->mx_cursor;
6468 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6469 if (m3->mc_pg[0] == mp) {
6470 m3->mc_pg[0] = mc->mc_pg[0];
6477 DPUTS("root page doesn't need rebalancing");
6481 /* The parent (branch page) must have at least 2 pointers,
6482 * otherwise the tree is invalid.
6484 ptop = mc->mc_top-1;
6485 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6487 /* Leaf page fill factor is below the threshold.
6488 * Try to move keys from left or right neighbor, or
6489 * merge with a neighbor page.
6494 mdb_cursor_copy(mc, &mn);
6495 mn.mc_xcursor = NULL;
6497 if (mc->mc_ki[ptop] == 0) {
6498 /* We're the leftmost leaf in our parent.
6500 DPUTS("reading right neighbor");
6502 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6503 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6506 mn.mc_ki[mn.mc_top] = 0;
6507 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6509 /* There is at least one neighbor to the left.
6511 DPUTS("reading left neighbor");
6513 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6514 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6517 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6518 mc->mc_ki[mc->mc_top] = 0;
6521 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6522 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);
6524 /* If the neighbor page is above threshold and has enough keys,
6525 * move one key from it. Otherwise we should try to merge them.
6526 * (A branch page must never have less than 2 keys.)
6528 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6529 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6530 return mdb_node_move(&mn, mc);
6532 if (mc->mc_ki[ptop] == 0)
6533 rc = mdb_page_merge(&mn, mc);
6535 rc = mdb_page_merge(mc, &mn);
6536 mc->mc_flags &= ~C_INITIALIZED;
6541 /** Complete a delete operation started by #mdb_cursor_del(). */
6543 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6547 /* add overflow pages to free list */
6548 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6552 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6553 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6554 (rc = mdb_ovpage_free(mc, omp)))
6557 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6558 mc->mc_db->md_entries--;
6559 rc = mdb_rebalance(mc);
6560 if (rc != MDB_SUCCESS)
6561 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6562 /* if mc points past last node in page, invalidate */
6563 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6564 mc->mc_flags &= ~C_INITIALIZED;
6570 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6571 MDB_val *key, MDB_val *data)
6576 MDB_val rdata, *xdata;
6580 assert(key != NULL);
6582 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6584 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6587 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6591 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6595 mdb_cursor_init(&mc, txn, dbi, &mx);
6606 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6608 /* let mdb_page_split know about this cursor if needed:
6609 * delete will trigger a rebalance; if it needs to move
6610 * a node from one page to another, it will have to
6611 * update the parent's separator key(s). If the new sepkey
6612 * is larger than the current one, the parent page may
6613 * run out of space, triggering a split. We need this
6614 * cursor to be consistent until the end of the rebalance.
6616 mc.mc_next = txn->mt_cursors[dbi];
6617 txn->mt_cursors[dbi] = &mc;
6618 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6619 txn->mt_cursors[dbi] = mc.mc_next;
6624 /** Split a page and insert a new node.
6625 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6626 * The cursor will be updated to point to the actual page and index where
6627 * the node got inserted after the split.
6628 * @param[in] newkey The key for the newly inserted node.
6629 * @param[in] newdata The data for the newly inserted node.
6630 * @param[in] newpgno The page number, if the new node is a branch node.
6631 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6632 * @return 0 on success, non-zero on failure.
6635 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6636 unsigned int nflags)
6639 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6642 unsigned int i, j, split_indx, nkeys, pmax;
6644 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6646 MDB_page *mp, *rp, *pp;
6651 mp = mc->mc_pg[mc->mc_top];
6652 newindx = mc->mc_ki[mc->mc_top];
6654 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6655 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6656 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6658 /* Create a right sibling. */
6659 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6661 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6663 if (mc->mc_snum < 2) {
6664 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6666 /* shift current top to make room for new parent */
6667 mc->mc_pg[1] = mc->mc_pg[0];
6668 mc->mc_ki[1] = mc->mc_ki[0];
6671 mc->mc_db->md_root = pp->mp_pgno;
6672 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6673 mc->mc_db->md_depth++;
6676 /* Add left (implicit) pointer. */
6677 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6678 /* undo the pre-push */
6679 mc->mc_pg[0] = mc->mc_pg[1];
6680 mc->mc_ki[0] = mc->mc_ki[1];
6681 mc->mc_db->md_root = mp->mp_pgno;
6682 mc->mc_db->md_depth--;
6689 ptop = mc->mc_top-1;
6690 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6693 mc->mc_flags |= C_SPLITTING;
6694 mdb_cursor_copy(mc, &mn);
6695 mn.mc_pg[mn.mc_top] = rp;
6696 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6698 if (nflags & MDB_APPEND) {
6699 mn.mc_ki[mn.mc_top] = 0;
6701 split_indx = newindx;
6706 nkeys = NUMKEYS(mp);
6707 split_indx = nkeys / 2;
6708 if (newindx < split_indx)
6714 unsigned int lsize, rsize, ksize;
6715 /* Move half of the keys to the right sibling */
6717 x = mc->mc_ki[mc->mc_top] - split_indx;
6718 ksize = mc->mc_db->md_pad;
6719 split = LEAF2KEY(mp, split_indx, ksize);
6720 rsize = (nkeys - split_indx) * ksize;
6721 lsize = (nkeys - split_indx) * sizeof(indx_t);
6722 mp->mp_lower -= lsize;
6723 rp->mp_lower += lsize;
6724 mp->mp_upper += rsize - lsize;
6725 rp->mp_upper -= rsize - lsize;
6726 sepkey.mv_size = ksize;
6727 if (newindx == split_indx) {
6728 sepkey.mv_data = newkey->mv_data;
6730 sepkey.mv_data = split;
6733 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6734 memcpy(rp->mp_ptrs, split, rsize);
6735 sepkey.mv_data = rp->mp_ptrs;
6736 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6737 memcpy(ins, newkey->mv_data, ksize);
6738 mp->mp_lower += sizeof(indx_t);
6739 mp->mp_upper -= ksize - sizeof(indx_t);
6742 memcpy(rp->mp_ptrs, split, x * ksize);
6743 ins = LEAF2KEY(rp, x, ksize);
6744 memcpy(ins, newkey->mv_data, ksize);
6745 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6746 rp->mp_lower += sizeof(indx_t);
6747 rp->mp_upper -= ksize - sizeof(indx_t);
6748 mc->mc_ki[mc->mc_top] = x;
6749 mc->mc_pg[mc->mc_top] = rp;
6754 /* For leaf pages, check the split point based on what
6755 * fits where, since otherwise mdb_node_add can fail.
6757 * This check is only needed when the data items are
6758 * relatively large, such that being off by one will
6759 * make the difference between success or failure.
6761 * It's also relevant if a page happens to be laid out
6762 * such that one half of its nodes are all "small" and
6763 * the other half of its nodes are "large." If the new
6764 * item is also "large" and falls on the half with
6765 * "large" nodes, it also may not fit.
6768 unsigned int psize, nsize;
6769 /* Maximum free space in an empty page */
6770 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6771 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6772 if ((nkeys < 20) || (nsize > pmax/16)) {
6773 if (newindx <= split_indx) {
6776 for (i=0; i<split_indx; i++) {
6777 node = NODEPTR(mp, i);
6778 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6779 if (F_ISSET(node->mn_flags, F_BIGDATA))
6780 psize += sizeof(pgno_t);
6782 psize += NODEDSZ(node);
6786 split_indx = newindx;
6797 for (i=nkeys-1; i>=split_indx; i--) {
6798 node = NODEPTR(mp, i);
6799 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6800 if (F_ISSET(node->mn_flags, F_BIGDATA))
6801 psize += sizeof(pgno_t);
6803 psize += NODEDSZ(node);
6807 split_indx = newindx;
6818 /* First find the separating key between the split pages.
6819 * The case where newindx == split_indx is ambiguous; the
6820 * new item could go to the new page or stay on the original
6821 * page. If newpos == 1 it goes to the new page.
6823 if (newindx == split_indx && newpos) {
6824 sepkey.mv_size = newkey->mv_size;
6825 sepkey.mv_data = newkey->mv_data;
6827 node = NODEPTR(mp, split_indx);
6828 sepkey.mv_size = node->mn_ksize;
6829 sepkey.mv_data = NODEKEY(node);
6833 DPRINTF("separator is [%s]", DKEY(&sepkey));
6835 /* Copy separator key to the parent.
6837 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6841 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6844 if (mn.mc_snum == mc->mc_snum) {
6845 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6846 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6847 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6848 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6853 /* Right page might now have changed parent.
6854 * Check if left page also changed parent.
6856 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6857 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6858 for (i=0; i<ptop; i++) {
6859 mc->mc_pg[i] = mn.mc_pg[i];
6860 mc->mc_ki[i] = mn.mc_ki[i];
6862 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6863 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6867 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6870 mc->mc_flags ^= C_SPLITTING;
6871 if (rc != MDB_SUCCESS) {
6874 if (nflags & MDB_APPEND) {
6875 mc->mc_pg[mc->mc_top] = rp;
6876 mc->mc_ki[mc->mc_top] = 0;
6877 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6880 for (i=0; i<mc->mc_top; i++)
6881 mc->mc_ki[i] = mn.mc_ki[i];
6888 /* Move half of the keys to the right sibling. */
6890 /* grab a page to hold a temporary copy */
6891 copy = mdb_page_malloc(mc->mc_txn, 1);
6895 copy->mp_pgno = mp->mp_pgno;
6896 copy->mp_flags = mp->mp_flags;
6897 copy->mp_lower = PAGEHDRSZ;
6898 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6899 mc->mc_pg[mc->mc_top] = copy;
6900 for (i = j = 0; i <= nkeys; j++) {
6901 if (i == split_indx) {
6902 /* Insert in right sibling. */
6903 /* Reset insert index for right sibling. */
6904 if (i != newindx || (newpos ^ ins_new)) {
6906 mc->mc_pg[mc->mc_top] = rp;
6910 if (i == newindx && !ins_new) {
6911 /* Insert the original entry that caused the split. */
6912 rkey.mv_data = newkey->mv_data;
6913 rkey.mv_size = newkey->mv_size;
6922 /* Update index for the new key. */
6923 mc->mc_ki[mc->mc_top] = j;
6924 } else if (i == nkeys) {
6927 node = NODEPTR(mp, i);
6928 rkey.mv_data = NODEKEY(node);
6929 rkey.mv_size = node->mn_ksize;
6931 xdata.mv_data = NODEDATA(node);
6932 xdata.mv_size = NODEDSZ(node);
6935 pgno = NODEPGNO(node);
6936 flags = node->mn_flags;
6941 if (!IS_LEAF(mp) && j == 0) {
6942 /* First branch index doesn't need key data. */
6946 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6950 nkeys = NUMKEYS(copy);
6951 for (i=0; i<nkeys; i++)
6952 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6953 mp->mp_lower = copy->mp_lower;
6954 mp->mp_upper = copy->mp_upper;
6955 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6956 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6958 /* reset back to original page */
6959 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6960 mc->mc_pg[mc->mc_top] = mp;
6961 if (nflags & MDB_RESERVE) {
6962 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6963 if (!(node->mn_flags & F_BIGDATA))
6964 newdata->mv_data = NODEDATA(node);
6968 /* Make sure mc_ki is still valid.
6970 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6971 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6972 for (i=0; i<ptop; i++) {
6973 mc->mc_pg[i] = mn.mc_pg[i];
6974 mc->mc_ki[i] = mn.mc_ki[i];
6976 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6977 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6981 /* return tmp page to freelist */
6982 mdb_page_free(mc->mc_txn->mt_env, copy);
6985 /* Adjust other cursors pointing to mp */
6986 MDB_cursor *m2, *m3;
6987 MDB_dbi dbi = mc->mc_dbi;
6988 int fixup = NUMKEYS(mp);
6990 if (mc->mc_flags & C_SUB)
6993 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6994 if (mc->mc_flags & C_SUB)
6995 m3 = &m2->mc_xcursor->mx_cursor;
7000 if (!(m3->mc_flags & C_INITIALIZED))
7002 if (m3->mc_flags & C_SPLITTING)
7007 for (k=m3->mc_top; k>=0; k--) {
7008 m3->mc_ki[k+1] = m3->mc_ki[k];
7009 m3->mc_pg[k+1] = m3->mc_pg[k];
7011 if (m3->mc_ki[0] >= split_indx) {
7016 m3->mc_pg[0] = mc->mc_pg[0];
7020 if (m3->mc_pg[mc->mc_top] == mp) {
7021 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7022 m3->mc_ki[mc->mc_top]++;
7023 if (m3->mc_ki[mc->mc_top] >= fixup) {
7024 m3->mc_pg[mc->mc_top] = rp;
7025 m3->mc_ki[mc->mc_top] -= fixup;
7026 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7028 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7029 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7038 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7039 MDB_val *key, MDB_val *data, unsigned int flags)
7044 assert(key != NULL);
7045 assert(data != NULL);
7047 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7050 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7054 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7058 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7061 mdb_cursor_init(&mc, txn, dbi, &mx);
7062 return mdb_cursor_put(&mc, key, data, flags);
7066 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7068 if ((flag & CHANGEABLE) != flag)
7071 env->me_flags |= flag;
7073 env->me_flags &= ~flag;
7078 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7083 *arg = env->me_flags;
7088 mdb_env_get_path(MDB_env *env, const char **arg)
7093 *arg = env->me_path;
7097 /** Common code for #mdb_stat() and #mdb_env_stat().
7098 * @param[in] env the environment to operate in.
7099 * @param[in] db the #MDB_db record containing the stats to return.
7100 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7101 * @return 0, this function always succeeds.
7104 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7106 arg->ms_psize = env->me_psize;
7107 arg->ms_depth = db->md_depth;
7108 arg->ms_branch_pages = db->md_branch_pages;
7109 arg->ms_leaf_pages = db->md_leaf_pages;
7110 arg->ms_overflow_pages = db->md_overflow_pages;
7111 arg->ms_entries = db->md_entries;
7116 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7120 if (env == NULL || arg == NULL)
7123 toggle = mdb_env_pick_meta(env);
7125 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7129 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7133 if (env == NULL || arg == NULL)
7136 toggle = mdb_env_pick_meta(env);
7137 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7138 arg->me_mapsize = env->me_mapsize;
7139 arg->me_maxreaders = env->me_maxreaders;
7140 arg->me_numreaders = env->me_numreaders;
7141 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7142 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7146 /** Set the default comparison functions for a database.
7147 * Called immediately after a database is opened to set the defaults.
7148 * The user can then override them with #mdb_set_compare() or
7149 * #mdb_set_dupsort().
7150 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7151 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7154 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7156 uint16_t f = txn->mt_dbs[dbi].md_flags;
7158 txn->mt_dbxs[dbi].md_cmp =
7159 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7160 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7162 txn->mt_dbxs[dbi].md_dcmp =
7163 !(f & MDB_DUPSORT) ? 0 :
7164 ((f & MDB_INTEGERDUP)
7165 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7166 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7169 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7174 int rc, dbflag, exact;
7175 unsigned int unused = 0;
7178 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7179 mdb_default_cmp(txn, FREE_DBI);
7182 if ((flags & VALID_FLAGS) != flags)
7188 if (flags & PERSISTENT_FLAGS) {
7189 uint16_t f2 = flags & PERSISTENT_FLAGS;
7190 /* make sure flag changes get committed */
7191 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7192 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7193 txn->mt_flags |= MDB_TXN_DIRTY;
7196 mdb_default_cmp(txn, MAIN_DBI);
7200 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7201 mdb_default_cmp(txn, MAIN_DBI);
7204 /* Is the DB already open? */
7206 for (i=2; i<txn->mt_numdbs; i++) {
7207 if (!txn->mt_dbxs[i].md_name.mv_size) {
7208 /* Remember this free slot */
7209 if (!unused) unused = i;
7212 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7213 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7219 /* If no free slot and max hit, fail */
7220 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7221 return MDB_DBS_FULL;
7223 /* Cannot mix named databases with some mainDB flags */
7224 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7225 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7227 /* Find the DB info */
7228 dbflag = DB_NEW|DB_VALID;
7231 key.mv_data = (void *)name;
7232 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7233 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7234 if (rc == MDB_SUCCESS) {
7235 /* make sure this is actually a DB */
7236 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7237 if (!(node->mn_flags & F_SUBDATA))
7239 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7240 /* Create if requested */
7242 data.mv_size = sizeof(MDB_db);
7243 data.mv_data = &dummy;
7244 memset(&dummy, 0, sizeof(dummy));
7245 dummy.md_root = P_INVALID;
7246 dummy.md_flags = flags & PERSISTENT_FLAGS;
7247 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7251 /* OK, got info, add to table */
7252 if (rc == MDB_SUCCESS) {
7253 unsigned int slot = unused ? unused : txn->mt_numdbs;
7254 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7255 txn->mt_dbxs[slot].md_name.mv_size = len;
7256 txn->mt_dbxs[slot].md_rel = NULL;
7257 txn->mt_dbflags[slot] = dbflag;
7258 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7260 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7261 mdb_default_cmp(txn, slot);
7270 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7272 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7275 if (txn->mt_dbflags[dbi] & DB_STALE) {
7278 /* Stale, must read the DB's root. cursor_init does it for us. */
7279 mdb_cursor_init(&mc, txn, dbi, &mx);
7281 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7284 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7287 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7289 ptr = env->me_dbxs[dbi].md_name.mv_data;
7290 env->me_dbxs[dbi].md_name.mv_data = NULL;
7291 env->me_dbxs[dbi].md_name.mv_size = 0;
7292 env->me_dbflags[dbi] = 0;
7296 /** Add all the DB's pages to the free list.
7297 * @param[in] mc Cursor on the DB to free.
7298 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7299 * @return 0 on success, non-zero on failure.
7302 mdb_drop0(MDB_cursor *mc, int subs)
7306 rc = mdb_page_search(mc, NULL, 0);
7307 if (rc == MDB_SUCCESS) {
7308 MDB_txn *txn = mc->mc_txn;
7313 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7314 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7317 mdb_cursor_copy(mc, &mx);
7318 while (mc->mc_snum > 0) {
7319 MDB_page *mp = mc->mc_pg[mc->mc_top];
7320 unsigned n = NUMKEYS(mp);
7322 for (i=0; i<n; i++) {
7323 ni = NODEPTR(mp, i);
7324 if (ni->mn_flags & F_BIGDATA) {
7327 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7328 rc = mdb_page_get(txn, pg, &omp, NULL);
7331 assert(IS_OVERFLOW(omp));
7332 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7336 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7337 mdb_xcursor_init1(mc, ni);
7338 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7344 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7346 for (i=0; i<n; i++) {
7348 ni = NODEPTR(mp, i);
7351 mdb_midl_xappend(txn->mt_free_pgs, pg);
7356 mc->mc_ki[mc->mc_top] = i;
7357 rc = mdb_cursor_sibling(mc, 1);
7359 /* no more siblings, go back to beginning
7360 * of previous level.
7364 for (i=1; i<mc->mc_snum; i++) {
7366 mc->mc_pg[i] = mx.mc_pg[i];
7371 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
7372 } else if (rc == MDB_NOTFOUND) {
7378 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7380 MDB_cursor *mc, *m2;
7383 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7386 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7389 rc = mdb_cursor_open(txn, dbi, &mc);
7393 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7394 /* Invalidate the dropped DB's cursors */
7395 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7396 m2->mc_flags &= ~C_INITIALIZED;
7400 /* Can't delete the main DB */
7401 if (del && dbi > MAIN_DBI) {
7402 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7404 txn->mt_dbflags[dbi] = DB_STALE;
7405 mdb_dbi_close(txn->mt_env, dbi);
7408 /* reset the DB record, mark it dirty */
7409 txn->mt_dbflags[dbi] |= DB_DIRTY;
7410 txn->mt_dbs[dbi].md_depth = 0;
7411 txn->mt_dbs[dbi].md_branch_pages = 0;
7412 txn->mt_dbs[dbi].md_leaf_pages = 0;
7413 txn->mt_dbs[dbi].md_overflow_pages = 0;
7414 txn->mt_dbs[dbi].md_entries = 0;
7415 txn->mt_dbs[dbi].md_root = P_INVALID;
7417 txn->mt_flags |= MDB_TXN_DIRTY;
7420 mdb_cursor_close(mc);
7424 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7426 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7429 txn->mt_dbxs[dbi].md_cmp = cmp;
7433 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7435 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7438 txn->mt_dbxs[dbi].md_dcmp = cmp;
7442 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7444 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7447 txn->mt_dbxs[dbi].md_rel = rel;
7451 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7453 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7456 txn->mt_dbxs[dbi].md_relctx = ctx;