2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define pthread_t DWORD
144 #define pthread_mutex_t HANDLE
145 #define pthread_key_t DWORD
146 #define pthread_self() GetCurrentThreadId()
147 #define pthread_key_create(x,y) \
148 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
149 #define pthread_key_delete(x) TlsFree(x)
150 #define pthread_getspecific(x) TlsGetValue(x)
151 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
152 #define pthread_mutex_unlock(x) ReleaseMutex(x)
153 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
154 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
158 #define getpid() GetCurrentProcessId()
159 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
160 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
161 #define ErrCode() GetLastError()
162 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
163 #define close(fd) (CloseHandle(fd) ? 0 : -1)
164 #define munmap(ptr,len) UnmapViewOfFile(ptr)
167 #ifdef MDB_USE_POSIX_SEM
169 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
170 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
171 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
172 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
175 mdb_sem_wait(sem_t *sem)
178 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
183 /** Lock the reader mutex.
185 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
186 /** Unlock the reader mutex.
188 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
190 /** Lock the writer mutex.
191 * Only a single write transaction is allowed at a time. Other writers
192 * will block waiting for this mutex.
194 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
195 /** Unlock the writer mutex.
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
198 #endif /* MDB_USE_POSIX_SEM */
200 /** Get the error code for the last failed system function.
202 #define ErrCode() errno
204 /** An abstraction for a file handle.
205 * On POSIX systems file handles are small integers. On Windows
206 * they're opaque pointers.
210 /** A value for an invalid file handle.
211 * Mainly used to initialize file variables and signify that they are
214 #define INVALID_HANDLE_VALUE (-1)
216 /** Get the size of a memory page for the system.
217 * This is the basic size that the platform's memory manager uses, and is
218 * fundamental to the use of memory-mapped files.
220 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
223 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
226 #define MNAME_LEN (sizeof(pthread_mutex_t))
232 /** A flag for opening a file and requesting synchronous data writes.
233 * This is only used when writing a meta page. It's not strictly needed;
234 * we could just do a normal write and then immediately perform a flush.
235 * But if this flag is available it saves us an extra system call.
237 * @note If O_DSYNC is undefined but exists in /usr/include,
238 * preferably set some compiler flag to get the definition.
239 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
242 # define MDB_DSYNC O_DSYNC
246 /** Function for flushing the data of a file. Define this to fsync
247 * if fdatasync() is not supported.
249 #ifndef MDB_FDATASYNC
250 # define MDB_FDATASYNC fdatasync
254 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
265 /** A page number in the database.
266 * Note that 64 bit page numbers are overkill, since pages themselves
267 * already represent 12-13 bits of addressable memory, and the OS will
268 * always limit applications to a maximum of 63 bits of address space.
270 * @note In the #MDB_node structure, we only store 48 bits of this value,
271 * which thus limits us to only 60 bits of addressable data.
273 typedef MDB_ID pgno_t;
275 /** A transaction ID.
276 * See struct MDB_txn.mt_txnid for details.
278 typedef MDB_ID txnid_t;
280 /** @defgroup debug Debug Macros
284 /** Enable debug output.
285 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
286 * read from and written to the database (used for free space management).
291 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
294 # define DPRINTF (void) /* Vararg macros may be unsupported */
296 static int mdb_debug;
297 static txnid_t mdb_debug_start;
299 /** Print a debug message with printf formatting. */
300 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
301 ((void) ((mdb_debug) && \
302 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
304 # define DPRINTF(fmt, ...) ((void) 0)
305 # define MDB_DEBUG_SKIP
307 /** Print a debug string.
308 * The string is printed literally, with no format processing.
310 #define DPUTS(arg) DPRINTF("%s", arg)
313 /** A default memory page size.
314 * The actual size is platform-dependent, but we use this for
315 * boot-strapping. We probably should not be using this any more.
316 * The #GET_PAGESIZE() macro is used to get the actual size.
318 * Note that we don't currently support Huge pages. On Linux,
319 * regular data files cannot use Huge pages, and in general
320 * Huge pages aren't actually pageable. We rely on the OS
321 * demand-pager to read our data and page it out when memory
322 * pressure from other processes is high. So until OSs have
323 * actual paging support for Huge pages, they're not viable.
325 #define MDB_PAGESIZE 4096
327 /** The minimum number of keys required in a database page.
328 * Setting this to a larger value will place a smaller bound on the
329 * maximum size of a data item. Data items larger than this size will
330 * be pushed into overflow pages instead of being stored directly in
331 * the B-tree node. This value used to default to 4. With a page size
332 * of 4096 bytes that meant that any item larger than 1024 bytes would
333 * go into an overflow page. That also meant that on average 2-3KB of
334 * each overflow page was wasted space. The value cannot be lower than
335 * 2 because then there would no longer be a tree structure. With this
336 * value, items larger than 2KB will go into overflow pages, and on
337 * average only 1KB will be wasted.
339 #define MDB_MINKEYS 2
341 /** A stamp that identifies a file as an MDB file.
342 * There's nothing special about this value other than that it is easily
343 * recognizable, and it will reflect any byte order mismatches.
345 #define MDB_MAGIC 0xBEEFC0DE
347 /** The version number for a database's file format. */
348 #define MDB_VERSION 1
350 /** @brief The maximum size of a key in the database.
352 * The library rejects bigger keys, and cannot deal with records
353 * with bigger keys stored by a library with bigger max keysize.
355 * We require that keys all fit onto a regular page. This limit
356 * could be raised a bit further if needed; to something just
357 * under #MDB_PAGESIZE / #MDB_MINKEYS.
359 * Note that data items in an #MDB_DUPSORT database are actually keys
360 * of a subDB, so they're also limited to this size.
362 #ifndef MDB_MAXKEYSIZE
363 #define MDB_MAXKEYSIZE 511
366 /** @brief The maximum size of a data item.
368 * We only store a 32 bit value for node sizes.
370 #define MAXDATASIZE 0xffffffffUL
375 * This is used for printing a hex dump of a key's contents.
377 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
378 /** Display a key in hex.
380 * Invoke a function to display a key in hex.
382 #define DKEY(x) mdb_dkey(x, kbuf)
384 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
388 /** An invalid page number.
389 * Mainly used to denote an empty tree.
391 #define P_INVALID (~(pgno_t)0)
393 /** Test if the flags \b f are set in a flag word \b w. */
394 #define F_ISSET(w, f) (((w) & (f)) == (f))
396 /** Used for offsets within a single page.
397 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
400 typedef uint16_t indx_t;
402 /** Default size of memory map.
403 * This is certainly too small for any actual applications. Apps should always set
404 * the size explicitly using #mdb_env_set_mapsize().
406 #define DEFAULT_MAPSIZE 1048576
408 /** @defgroup readers Reader Lock Table
409 * Readers don't acquire any locks for their data access. Instead, they
410 * simply record their transaction ID in the reader table. The reader
411 * mutex is needed just to find an empty slot in the reader table. The
412 * slot's address is saved in thread-specific data so that subsequent read
413 * transactions started by the same thread need no further locking to proceed.
415 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
417 * No reader table is used if the database is on a read-only filesystem.
419 * Since the database uses multi-version concurrency control, readers don't
420 * actually need any locking. This table is used to keep track of which
421 * readers are using data from which old transactions, so that we'll know
422 * when a particular old transaction is no longer in use. Old transactions
423 * that have discarded any data pages can then have those pages reclaimed
424 * for use by a later write transaction.
426 * The lock table is constructed such that reader slots are aligned with the
427 * processor's cache line size. Any slot is only ever used by one thread.
428 * This alignment guarantees that there will be no contention or cache
429 * thrashing as threads update their own slot info, and also eliminates
430 * any need for locking when accessing a slot.
432 * A writer thread will scan every slot in the table to determine the oldest
433 * outstanding reader transaction. Any freed pages older than this will be
434 * reclaimed by the writer. The writer doesn't use any locks when scanning
435 * this table. This means that there's no guarantee that the writer will
436 * see the most up-to-date reader info, but that's not required for correct
437 * operation - all we need is to know the upper bound on the oldest reader,
438 * we don't care at all about the newest reader. So the only consequence of
439 * reading stale information here is that old pages might hang around a
440 * while longer before being reclaimed. That's actually good anyway, because
441 * the longer we delay reclaiming old pages, the more likely it is that a
442 * string of contiguous pages can be found after coalescing old pages from
443 * many old transactions together.
446 /** Number of slots in the reader table.
447 * This value was chosen somewhat arbitrarily. 126 readers plus a
448 * couple mutexes fit exactly into 8KB on my development machine.
449 * Applications should set the table size using #mdb_env_set_maxreaders().
451 #define DEFAULT_READERS 126
453 /** The size of a CPU cache line in bytes. We want our lock structures
454 * aligned to this size to avoid false cache line sharing in the
456 * This value works for most CPUs. For Itanium this should be 128.
462 /** The information we store in a single slot of the reader table.
463 * In addition to a transaction ID, we also record the process and
464 * thread ID that owns a slot, so that we can detect stale information,
465 * e.g. threads or processes that went away without cleaning up.
466 * @note We currently don't check for stale records. We simply re-init
467 * the table when we know that we're the only process opening the
470 typedef struct MDB_rxbody {
471 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
472 * Multiple readers that start at the same time will probably have the
473 * same ID here. Again, it's not important to exclude them from
474 * anything; all we need to know is which version of the DB they
475 * started from so we can avoid overwriting any data used in that
476 * particular version.
479 /** The process ID of the process owning this reader txn. */
481 /** The thread ID of the thread owning this txn. */
485 /** The actual reader record, with cacheline padding. */
486 typedef struct MDB_reader {
489 /** shorthand for mrb_txnid */
490 #define mr_txnid mru.mrx.mrb_txnid
491 #define mr_pid mru.mrx.mrb_pid
492 #define mr_tid mru.mrx.mrb_tid
493 /** cache line alignment */
494 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
498 /** The header for the reader table.
499 * The table resides in a memory-mapped file. (This is a different file
500 * than is used for the main database.)
502 * For POSIX the actual mutexes reside in the shared memory of this
503 * mapped file. On Windows, mutexes are named objects allocated by the
504 * kernel; we store the mutex names in this mapped file so that other
505 * processes can grab them. This same approach is also used on
506 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
507 * process-shared POSIX mutexes. For these cases where a named object
508 * is used, the object name is derived from a 64 bit FNV hash of the
509 * environment pathname. As such, naming collisions are extremely
510 * unlikely. If a collision occurs, the results are unpredictable.
512 typedef struct MDB_txbody {
513 /** Stamp identifying this as an MDB file. It must be set
516 /** Version number of this lock file. Must be set to #MDB_VERSION. */
517 uint32_t mtb_version;
518 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
519 char mtb_rmname[MNAME_LEN];
521 /** Mutex protecting access to this table.
522 * This is the reader lock that #LOCK_MUTEX_R acquires.
524 pthread_mutex_t mtb_mutex;
526 /** The ID of the last transaction committed to the database.
527 * This is recorded here only for convenience; the value can always
528 * be determined by reading the main database meta pages.
531 /** The number of slots that have been used in the reader table.
532 * This always records the maximum count, it is not decremented
533 * when readers release their slots.
535 unsigned mtb_numreaders;
538 /** The actual reader table definition. */
539 typedef struct MDB_txninfo {
542 #define mti_magic mt1.mtb.mtb_magic
543 #define mti_version mt1.mtb.mtb_version
544 #define mti_mutex mt1.mtb.mtb_mutex
545 #define mti_rmname mt1.mtb.mtb_rmname
546 #define mti_txnid mt1.mtb.mtb_txnid
547 #define mti_numreaders mt1.mtb.mtb_numreaders
548 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
551 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
552 char mt2_wmname[MNAME_LEN];
553 #define mti_wmname mt2.mt2_wmname
555 pthread_mutex_t mt2_wmutex;
556 #define mti_wmutex mt2.mt2_wmutex
558 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
560 MDB_reader mti_readers[1];
564 /** Common header for all page types.
565 * Overflow records occupy a number of contiguous pages with no
566 * headers on any page after the first.
568 typedef struct MDB_page {
569 #define mp_pgno mp_p.p_pgno
570 #define mp_next mp_p.p_next
572 pgno_t p_pgno; /**< page number */
573 void * p_next; /**< for in-memory list of freed structs */
576 /** @defgroup mdb_page Page Flags
578 * Flags for the page headers.
581 #define P_BRANCH 0x01 /**< branch page */
582 #define P_LEAF 0x02 /**< leaf page */
583 #define P_OVERFLOW 0x04 /**< overflow page */
584 #define P_META 0x08 /**< meta page */
585 #define P_DIRTY 0x10 /**< dirty page */
586 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
587 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
589 uint16_t mp_flags; /**< @ref mdb_page */
590 #define mp_lower mp_pb.pb.pb_lower
591 #define mp_upper mp_pb.pb.pb_upper
592 #define mp_pages mp_pb.pb_pages
595 indx_t pb_lower; /**< lower bound of free space */
596 indx_t pb_upper; /**< upper bound of free space */
598 uint32_t pb_pages; /**< number of overflow pages */
600 indx_t mp_ptrs[1]; /**< dynamic size */
603 /** Size of the page header, excluding dynamic data at the end */
604 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
606 /** Address of first usable data byte in a page, after the header */
607 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
609 /** Number of nodes on a page */
610 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
612 /** The amount of space remaining in the page */
613 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
615 /** The percentage of space used in the page, in tenths of a percent. */
616 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
617 ((env)->me_psize - PAGEHDRSZ))
618 /** The minimum page fill factor, in tenths of a percent.
619 * Pages emptier than this are candidates for merging.
621 #define FILL_THRESHOLD 250
623 /** Test if a page is a leaf page */
624 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
625 /** Test if a page is a LEAF2 page */
626 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
627 /** Test if a page is a branch page */
628 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
629 /** Test if a page is an overflow page */
630 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
631 /** Test if a page is a sub page */
632 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
634 /** The number of overflow pages needed to store the given size. */
635 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
637 /** Header for a single key/data pair within a page.
638 * We guarantee 2-byte alignment for nodes.
640 typedef struct MDB_node {
641 /** lo and hi are used for data size on leaf nodes and for
642 * child pgno on branch nodes. On 64 bit platforms, flags
643 * is also used for pgno. (Branch nodes have no flags).
644 * They are in host byte order in case that lets some
645 * accesses be optimized into a 32-bit word access.
647 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
648 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
649 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
650 /** @defgroup mdb_node Node Flags
652 * Flags for node headers.
655 #define F_BIGDATA 0x01 /**< data put on overflow page */
656 #define F_SUBDATA 0x02 /**< data is a sub-database */
657 #define F_DUPDATA 0x04 /**< data has duplicates */
659 /** valid flags for #mdb_node_add() */
660 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
663 unsigned short mn_flags; /**< @ref mdb_node */
664 unsigned short mn_ksize; /**< key size */
665 char mn_data[1]; /**< key and data are appended here */
668 /** Size of the node header, excluding dynamic data at the end */
669 #define NODESIZE offsetof(MDB_node, mn_data)
671 /** Bit position of top word in page number, for shifting mn_flags */
672 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
674 /** Size of a node in a branch page with a given key.
675 * This is just the node header plus the key, there is no data.
677 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
679 /** Size of a node in a leaf page with a given key and data.
680 * This is node header plus key plus data size.
682 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
684 /** Address of node \b i in page \b p */
685 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
687 /** Address of the key for the node */
688 #define NODEKEY(node) (void *)((node)->mn_data)
690 /** Address of the data for a node */
691 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
693 /** Get the page number pointed to by a branch node */
694 #define NODEPGNO(node) \
695 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
696 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
697 /** Set the page number in a branch node */
698 #define SETPGNO(node,pgno) do { \
699 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
700 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
702 /** Get the size of the data in a leaf node */
703 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
704 /** Set the size of the data for a leaf node */
705 #define SETDSZ(node,size) do { \
706 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
707 /** The size of a key in a node */
708 #define NODEKSZ(node) ((node)->mn_ksize)
710 /** Copy a page number from src to dst */
712 #define COPY_PGNO(dst,src) dst = src
714 #if SIZE_MAX > 4294967295UL
715 #define COPY_PGNO(dst,src) do { \
716 unsigned short *s, *d; \
717 s = (unsigned short *)&(src); \
718 d = (unsigned short *)&(dst); \
725 #define COPY_PGNO(dst,src) do { \
726 unsigned short *s, *d; \
727 s = (unsigned short *)&(src); \
728 d = (unsigned short *)&(dst); \
734 /** The address of a key in a LEAF2 page.
735 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
736 * There are no node headers, keys are stored contiguously.
738 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
740 /** Set the \b node's key into \b key, if requested. */
741 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
742 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
744 /** Information about a single database in the environment. */
745 typedef struct MDB_db {
746 uint32_t md_pad; /**< also ksize for LEAF2 pages */
747 uint16_t md_flags; /**< @ref mdb_dbi_open */
748 uint16_t md_depth; /**< depth of this tree */
749 pgno_t md_branch_pages; /**< number of internal pages */
750 pgno_t md_leaf_pages; /**< number of leaf pages */
751 pgno_t md_overflow_pages; /**< number of overflow pages */
752 size_t md_entries; /**< number of data items */
753 pgno_t md_root; /**< the root page of this tree */
756 /** mdb_dbi_open flags */
757 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
758 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
759 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
760 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
762 /** Handle for the DB used to track free pages. */
764 /** Handle for the default DB. */
767 /** Meta page content. */
768 typedef struct MDB_meta {
769 /** Stamp identifying this as an MDB file. It must be set
772 /** Version number of this lock file. Must be set to #MDB_VERSION. */
774 void *mm_address; /**< address for fixed mapping */
775 size_t mm_mapsize; /**< size of mmap region */
776 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
777 /** The size of pages used in this DB */
778 #define mm_psize mm_dbs[0].md_pad
779 /** Any persistent environment flags. @ref mdb_env */
780 #define mm_flags mm_dbs[0].md_flags
781 pgno_t mm_last_pg; /**< last used page in file */
782 txnid_t mm_txnid; /**< txnid that committed this page */
785 /** Buffer for a stack-allocated dirty page.
786 * The members define size and alignment, and silence type
787 * aliasing warnings. They are not used directly; that could
788 * mean incorrectly using several union members in parallel.
790 typedef union MDB_pagebuf {
791 char mb_raw[MDB_PAGESIZE];
794 char mm_pad[PAGEHDRSZ];
799 /** Auxiliary DB info.
800 * The information here is mostly static/read-only. There is
801 * only a single copy of this record in the environment.
803 typedef struct MDB_dbx {
804 MDB_val md_name; /**< name of the database */
805 MDB_cmp_func *md_cmp; /**< function for comparing keys */
806 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
807 MDB_rel_func *md_rel; /**< user relocate function */
808 void *md_relctx; /**< user-provided context for md_rel */
811 /** A database transaction.
812 * Every operation requires a transaction handle.
815 MDB_txn *mt_parent; /**< parent of a nested txn */
816 MDB_txn *mt_child; /**< nested txn under this txn */
817 pgno_t mt_next_pgno; /**< next unallocated page */
818 /** The ID of this transaction. IDs are integers incrementing from 1.
819 * Only committed write transactions increment the ID. If a transaction
820 * aborts, the ID may be re-used by the next writer.
823 MDB_env *mt_env; /**< the DB environment */
824 /** The list of pages that became unused during this transaction.
828 MDB_ID2L dirty_list; /**< for write txns: modified pages */
829 MDB_reader *reader; /**< this thread's reader table slot or NULL */
831 /** Array of records for each DB known in the environment. */
833 /** Array of MDB_db records for each known DB */
835 /** @defgroup mt_dbflag Transaction DB Flags
839 #define DB_DIRTY 0x01 /**< DB was written in this txn */
840 #define DB_STALE 0x02 /**< DB record is older than txnID */
841 #define DB_NEW 0x04 /**< DB handle opened in this txn */
842 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
844 /** In write txns, array of cursors for each DB */
845 MDB_cursor **mt_cursors;
846 /** Array of flags for each DB */
847 unsigned char *mt_dbflags;
848 /** Number of DB records in use. This number only ever increments;
849 * we don't decrement it when individual DB handles are closed.
853 /** @defgroup mdb_txn Transaction Flags
857 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
858 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
859 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
861 unsigned int mt_flags; /**< @ref mdb_txn */
862 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
863 unsigned int mt_dirty_room;
864 /** Tracks which of the two meta pages was used at the start
865 * of this transaction.
867 unsigned int mt_toggle;
870 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
871 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
872 * raise this on a 64 bit machine.
874 #define CURSOR_STACK 32
878 /** Cursors are used for all DB operations */
880 /** Next cursor on this DB in this txn */
882 /** Original cursor if this is a shadow */
884 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
885 struct MDB_xcursor *mc_xcursor;
886 /** The transaction that owns this cursor */
888 /** The database handle this cursor operates on */
890 /** The database record for this cursor */
892 /** The database auxiliary record for this cursor */
894 /** The @ref mt_dbflag for this database */
895 unsigned char *mc_dbflag;
896 unsigned short mc_snum; /**< number of pushed pages */
897 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
898 /** @defgroup mdb_cursor Cursor Flags
900 * Cursor state flags.
903 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
904 #define C_EOF 0x02 /**< No more data */
905 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
906 #define C_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 /** Make shadow copies of all of parent txn's cursors */
1638 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1640 MDB_cursor *mc, *m2;
1641 unsigned int i, j, size;
1643 for (i=0;i<src->mt_numdbs; i++) {
1644 if (src->mt_cursors[i]) {
1645 size = sizeof(MDB_cursor);
1646 if (src->mt_cursors[i]->mc_xcursor)
1647 size += sizeof(MDB_xcursor);
1648 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1655 mc->mc_db = &dst->mt_dbs[i];
1656 mc->mc_dbx = m2->mc_dbx;
1657 mc->mc_dbflag = &dst->mt_dbflags[i];
1658 mc->mc_snum = m2->mc_snum;
1659 mc->mc_top = m2->mc_top;
1660 mc->mc_flags = m2->mc_flags;
1661 for (j=0; j<mc->mc_snum; j++) {
1662 mc->mc_pg[j] = m2->mc_pg[j];
1663 mc->mc_ki[j] = m2->mc_ki[j];
1665 if (m2->mc_xcursor) {
1666 MDB_xcursor *mx, *mx2;
1667 mx = (MDB_xcursor *)(mc+1);
1668 mc->mc_xcursor = mx;
1669 mx2 = m2->mc_xcursor;
1670 mx->mx_db = mx2->mx_db;
1671 mx->mx_dbx = mx2->mx_dbx;
1672 mx->mx_dbflag = mx2->mx_dbflag;
1673 mx->mx_cursor.mc_txn = dst;
1674 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1675 mx->mx_cursor.mc_db = &mx->mx_db;
1676 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1677 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1678 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1679 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1680 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags;
1681 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1682 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1683 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1686 mc->mc_xcursor = NULL;
1688 mc->mc_next = dst->mt_cursors[i];
1689 dst->mt_cursors[i] = mc;
1696 /** Close this write txn's cursors, after optionally merging its shadow
1697 * cursors back into parent's.
1698 * @param[in] txn the transaction handle.
1699 * @param[in] merge zero to not merge cursors, non-zero to merge.
1700 * @return 0 on success, non-zero on failure.
1703 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1705 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1708 for (i = txn->mt_numdbs; --i >= 0; ) {
1709 for (mc = cursors[i]; mc; mc = next) {
1711 if (merge && mc->mc_orig) {
1712 MDB_cursor *m2 = mc->mc_orig;
1713 m2->mc_snum = mc->mc_snum;
1714 m2->mc_top = mc->mc_top;
1715 for (j = mc->mc_snum; --j >= 0; ) {
1716 m2->mc_pg[j] = mc->mc_pg[j];
1717 m2->mc_ki[j] = mc->mc_ki[j];
1720 /* Only malloced cursors are permanently tracked. */
1727 #ifdef MDB_DEBUG_SKIP
1728 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
1731 mdb_txn_reset0(MDB_txn *txn, const char *act);
1733 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1734 * @param[in] txn the transaction handle to initialize
1735 * @return 0 on success, non-zero on failure.
1738 mdb_txn_renew0(MDB_txn *txn)
1740 MDB_env *env = txn->mt_env;
1743 int rc, new_notls = 0;
1746 txn->mt_numdbs = env->me_numdbs;
1747 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1749 if (txn->mt_flags & MDB_TXN_RDONLY) {
1750 if (!env->me_txns) {
1751 i = mdb_env_pick_meta(env);
1752 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1753 txn->mt_u.reader = NULL;
1755 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1756 pthread_getspecific(env->me_txkey);
1758 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1759 return MDB_BAD_RSLOT;
1761 pid_t pid = env->me_pid;
1762 pthread_t tid = pthread_self();
1765 for (i=0; i<env->me_txns->mti_numreaders; i++)
1766 if (env->me_txns->mti_readers[i].mr_pid == 0)
1768 if (i == env->me_maxreaders) {
1769 UNLOCK_MUTEX_R(env);
1770 return MDB_READERS_FULL;
1772 env->me_txns->mti_readers[i].mr_pid = pid;
1773 env->me_txns->mti_readers[i].mr_tid = tid;
1774 if (i >= env->me_txns->mti_numreaders)
1775 env->me_txns->mti_numreaders = i+1;
1776 /* Save numreaders for un-mutexed mdb_env_close() */
1777 env->me_numreaders = env->me_txns->mti_numreaders;
1778 UNLOCK_MUTEX_R(env);
1779 r = &env->me_txns->mti_readers[i];
1780 new_notls = (env->me_flags & MDB_NOTLS);
1781 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1786 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1787 txn->mt_u.reader = r;
1789 txn->mt_toggle = txn->mt_txnid & 1;
1793 txn->mt_txnid = env->me_txns->mti_txnid;
1794 txn->mt_toggle = txn->mt_txnid & 1;
1797 if (txn->mt_txnid == mdb_debug_start)
1800 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1801 txn->mt_u.dirty_list = env->me_dirty_list;
1802 txn->mt_u.dirty_list[0].mid = 0;
1803 txn->mt_free_pgs = env->me_free_pgs;
1804 txn->mt_free_pgs[0] = 0;
1808 /* Copy the DB info and flags */
1809 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1811 /* Moved to here to avoid a data race in read TXNs */
1812 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1814 for (i=2; i<txn->mt_numdbs; i++) {
1815 x = env->me_dbflags[i];
1816 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1817 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1819 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1821 if (env->me_maxpg < txn->mt_next_pgno) {
1822 mdb_txn_reset0(txn, "renew0-mapfail");
1824 txn->mt_u.reader->mr_pid = 0;
1825 txn->mt_u.reader = NULL;
1827 return MDB_MAP_RESIZED;
1834 mdb_txn_renew(MDB_txn *txn)
1838 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1841 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1842 DPUTS("environment had fatal error, must shutdown!");
1846 rc = mdb_txn_renew0(txn);
1847 if (rc == MDB_SUCCESS) {
1848 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1849 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1850 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1856 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1860 int rc, size, tsize = sizeof(MDB_txn);
1862 if (env->me_flags & MDB_FATAL_ERROR) {
1863 DPUTS("environment had fatal error, must shutdown!");
1866 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1869 /* Nested transactions: Max 1 child, write txns only, no writemap */
1870 if (parent->mt_child ||
1871 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1872 (env->me_flags & MDB_WRITEMAP))
1876 tsize = sizeof(MDB_ntxn);
1878 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1879 if (!(flags & MDB_RDONLY))
1880 size += env->me_maxdbs * sizeof(MDB_cursor *);
1882 if ((txn = calloc(1, size)) == NULL) {
1883 DPRINTF("calloc: %s", strerror(ErrCode()));
1886 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1887 if (flags & MDB_RDONLY) {
1888 txn->mt_flags |= MDB_TXN_RDONLY;
1889 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1891 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1892 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1898 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1899 if (!txn->mt_u.dirty_list ||
1900 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1902 free(txn->mt_u.dirty_list);
1906 txn->mt_txnid = parent->mt_txnid;
1907 txn->mt_toggle = parent->mt_toggle;
1908 txn->mt_dirty_room = parent->mt_dirty_room;
1909 txn->mt_u.dirty_list[0].mid = 0;
1910 txn->mt_next_pgno = parent->mt_next_pgno;
1911 parent->mt_child = txn;
1912 txn->mt_parent = parent;
1913 txn->mt_numdbs = parent->mt_numdbs;
1914 txn->mt_flags = parent->mt_flags;
1915 txn->mt_dbxs = parent->mt_dbxs;
1916 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1917 /* Copy parent's mt_dbflags, but clear DB_NEW */
1918 for (i=0; i<txn->mt_numdbs; i++)
1919 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1921 ntxn = (MDB_ntxn *)txn;
1922 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1923 if (env->me_pghead) {
1924 size = MDB_IDL_SIZEOF(env->me_pghead);
1925 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1927 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1932 rc = mdb_cursor_shadow(parent, txn);
1934 mdb_txn_reset0(txn, "beginchild-fail");
1936 rc = mdb_txn_renew0(txn);
1942 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1943 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1944 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1950 /** Export or close DBI handles opened in this txn. */
1952 mdb_dbis_update(MDB_txn *txn, int keep)
1955 MDB_dbi n = txn->mt_numdbs;
1956 MDB_env *env = txn->mt_env;
1957 unsigned char *tdbflags = txn->mt_dbflags;
1959 for (i = n; --i >= 2;) {
1960 if (tdbflags[i] & DB_NEW) {
1962 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
1964 char *ptr = env->me_dbxs[i].md_name.mv_data;
1965 env->me_dbxs[i].md_name.mv_data = NULL;
1966 env->me_dbxs[i].md_name.mv_size = 0;
1967 env->me_dbflags[i] = 0;
1972 if (keep && env->me_numdbs < n)
1976 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1977 * May be called twice for readonly txns: First reset it, then abort.
1978 * @param[in] txn the transaction handle to reset
1981 mdb_txn_reset0(MDB_txn *txn, const char *act)
1983 MDB_env *env = txn->mt_env;
1985 /* Close any DBI handles opened in this txn */
1986 mdb_dbis_update(txn, 0);
1988 DPRINTF("%s txn %zu%c %p on mdbenv %p, root page %zu",
1989 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1990 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1992 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1993 if (txn->mt_u.reader) {
1994 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1995 if (!(env->me_flags & MDB_NOTLS))
1996 txn->mt_u.reader = NULL; /* txn does not own reader */
1998 txn->mt_numdbs = 0; /* close nothing if called again */
1999 txn->mt_dbxs = NULL; /* mark txn as reset */
2001 mdb_cursors_close(txn, 0);
2003 if (!(env->me_flags & MDB_WRITEMAP)) {
2004 mdb_dlist_free(txn);
2006 mdb_midl_free(env->me_pghead);
2008 if (txn->mt_parent) {
2009 txn->mt_parent->mt_child = NULL;
2010 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2011 mdb_midl_free(txn->mt_free_pgs);
2012 free(txn->mt_u.dirty_list);
2016 if (mdb_midl_shrink(&txn->mt_free_pgs))
2017 env->me_free_pgs = txn->mt_free_pgs;
2018 env->me_pghead = NULL;
2022 /* The writer mutex was locked in mdb_txn_begin. */
2023 UNLOCK_MUTEX_W(env);
2028 mdb_txn_reset(MDB_txn *txn)
2033 /* This call is only valid for read-only txns */
2034 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2037 mdb_txn_reset0(txn, "reset");
2041 mdb_txn_abort(MDB_txn *txn)
2047 mdb_txn_abort(txn->mt_child);
2049 mdb_txn_reset0(txn, "abort");
2050 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2051 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2052 txn->mt_u.reader->mr_pid = 0;
2057 /** Save the freelist as of this transaction to the freeDB.
2058 * This changes the freelist. Keep trying until it stabilizes.
2061 mdb_freelist_save(MDB_txn *txn)
2063 /* env->me_pghead[] can grow and shrink during this call.
2064 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2065 * Page numbers cannot disappear from txn->mt_free_pgs[].
2068 MDB_env *env = txn->mt_env;
2069 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2070 txnid_t pglast = 0, head_id = 0;
2071 pgno_t freecnt = 0, *free_pgs, *mop;
2072 ssize_t head_room = 0, total_room = 0, mop_len;
2074 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2076 if (env->me_pghead) {
2077 /* Make sure first page of freeDB is touched and on freelist */
2078 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2079 if (rc && rc != MDB_NOTFOUND)
2084 /* Come back here after each Put() in case freelist changed */
2087 /* If using records from freeDB which we have not yet
2088 * deleted, delete them and any we reserved for me_pghead.
2090 while (pglast < env->me_pglast) {
2091 rc = mdb_cursor_first(&mc, &key, NULL);
2094 pglast = head_id = *(txnid_t *)key.mv_data;
2095 total_room = head_room = 0;
2096 assert(pglast <= env->me_pglast);
2097 rc = mdb_cursor_del(&mc, 0);
2102 /* Save the IDL of pages freed by this txn, to a single record */
2103 if (freecnt < txn->mt_free_pgs[0]) {
2105 /* Make sure last page of freeDB is touched and on freelist */
2106 key.mv_size = MDB_MAXKEYSIZE+1;
2108 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2109 if (rc && rc != MDB_NOTFOUND)
2112 free_pgs = txn->mt_free_pgs;
2113 /* Write to last page of freeDB */
2114 key.mv_size = sizeof(txn->mt_txnid);
2115 key.mv_data = &txn->mt_txnid;
2117 freecnt = free_pgs[0];
2118 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2119 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2122 /* Retry if mt_free_pgs[] grew during the Put() */
2123 free_pgs = txn->mt_free_pgs;
2124 } while (freecnt < free_pgs[0]);
2125 mdb_midl_sort(free_pgs);
2126 memcpy(data.mv_data, free_pgs, data.mv_size);
2129 unsigned int i = free_pgs[0];
2130 DPRINTF("IDL write txn %zu root %zu num %u",
2131 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2133 DPRINTF("IDL %zu", free_pgs[i]);
2139 mop = env->me_pghead;
2140 mop_len = mop ? mop[0] : 0;
2142 /* Reserve records for me_pghead[]. Split it if multi-page,
2143 * to avoid searching freeDB for a page range. Use keys in
2144 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2146 if (total_room >= mop_len) {
2147 if (total_room == mop_len || --more < 0)
2149 } else if (head_room >= maxfree_1pg && head_id > 1) {
2150 /* Keep current record (overflow page), add a new one */
2154 /* (Re)write {key = head_id, IDL length = head_room} */
2155 total_room -= head_room;
2156 head_room = mop_len - total_room;
2157 if (head_room > maxfree_1pg && head_id > 1) {
2158 /* Overflow multi-page for part of me_pghead */
2159 head_room /= head_id; /* amortize page sizes */
2160 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2161 } else if (head_room < 0) {
2162 /* Rare case, not bothering to delete this record */
2165 key.mv_size = sizeof(head_id);
2166 key.mv_data = &head_id;
2167 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2168 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2171 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2172 total_room += head_room;
2175 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2176 * so they cannot rearrange anything, just read the destinations.
2183 rc = mdb_cursor_first(&mc, &key, &data);
2184 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2185 MDB_IDL dest = data.mv_data;
2186 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2188 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2192 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2193 if (! (mop_len -= len))
2200 /** Flush dirty pages to the map, after clearing their dirty flag.
2203 mdb_page_flush(MDB_txn *txn)
2205 MDB_env *env = txn->mt_env;
2206 MDB_ID2L dl = txn->mt_u.dirty_list;
2207 unsigned psize = env->me_psize;
2208 int i, pagecount = dl[0].mid, rc;
2209 size_t size = 0, pos = 0;
2211 MDB_page *dp = NULL;
2215 struct iovec iov[MDB_COMMIT_PAGES];
2216 ssize_t wpos = 0, wsize = 0, wres;
2217 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2221 if (env->me_flags & MDB_WRITEMAP) {
2222 /* Clear dirty flags */
2223 for (i = pagecount; i; i--) {
2225 dp->mp_flags &= ~P_DIRTY;
2231 /* Write the pages */
2233 if (i <= pagecount) {
2236 /* clear dirty flag */
2237 dp->mp_flags &= ~P_DIRTY;
2240 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2245 /* Windows actually supports scatter/gather I/O, but only on
2246 * unbuffered file handles. Since we're relying on the OS page
2247 * cache for all our data, that's self-defeating. So we just
2248 * write pages one at a time. We use the ov structure to set
2249 * the write offset, to at least save the overhead of a Seek
2252 DPRINTF("committing page %zu", pgno);
2253 memset(&ov, 0, sizeof(ov));
2254 ov.Offset = pos & 0xffffffff;
2255 ov.OffsetHigh = pos >> 16 >> 16;
2256 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2258 DPRINTF("WriteFile: %d", rc);
2262 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2263 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2265 /* Write previous page(s) */
2266 #ifdef MDB_USE_PWRITEV
2267 wres = pwritev(env->me_fd, iov, n, wpos);
2270 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2272 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2274 DPRINTF("lseek: %s", strerror(rc));
2277 wres = writev(env->me_fd, iov, n);
2280 if (wres != wsize) {
2283 DPRINTF("Write error: %s", strerror(rc));
2285 rc = EIO; /* TODO: Use which error code? */
2286 DPUTS("short write, filesystem full?");
2297 DPRINTF("committing page %zu", pgno);
2298 next_pos = pos + size;
2299 iov[n].iov_len = size;
2300 iov[n].iov_base = (char *)dp;
2306 mdb_dlist_free(txn);
2312 mdb_txn_commit(MDB_txn *txn)
2318 assert(txn != NULL);
2319 assert(txn->mt_env != NULL);
2321 if (txn->mt_child) {
2322 rc = mdb_txn_commit(txn->mt_child);
2323 txn->mt_child = NULL;
2330 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2331 mdb_dbis_update(txn, 1);
2332 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2337 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2338 DPUTS("error flag is set, can't commit");
2340 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2345 if (txn->mt_parent) {
2346 MDB_txn *parent = txn->mt_parent;
2350 /* Append our free list to parent's */
2351 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2354 mdb_midl_free(txn->mt_free_pgs);
2356 parent->mt_next_pgno = txn->mt_next_pgno;
2357 parent->mt_flags = txn->mt_flags;
2359 /* Merge our cursors into parent's and close them */
2360 mdb_cursors_close(txn, 1);
2362 /* Update parent's DB table. */
2363 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2364 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2365 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2366 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2367 for (i=2; i<txn->mt_numdbs; i++) {
2368 /* preserve parent's DB_NEW status */
2369 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2370 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2373 dst = txn->mt_parent->mt_u.dirty_list;
2374 src = txn->mt_u.dirty_list;
2375 /* Find len = length of merging our dirty list with parent's */
2377 dst[0].mid = 0; /* simplify loops */
2378 if (parent->mt_parent) {
2379 len = x + src[0].mid;
2380 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2381 for (i = x; y && i; y--) {
2382 pgno_t yp = src[y].mid;
2383 while (yp < dst[i].mid)
2385 if (yp == dst[i].mid) {
2390 } else { /* Simplify the above for single-ancestor case */
2391 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2393 /* Merge our dirty list with parent's */
2395 for (i = len; y; dst[i--] = src[y--]) {
2396 pgno_t yp = src[y].mid;
2397 while (yp < dst[x].mid)
2398 dst[i--] = dst[x--];
2399 if (yp == dst[x].mid)
2400 free(dst[x--].mptr);
2404 free(txn->mt_u.dirty_list);
2405 parent->mt_dirty_room = txn->mt_dirty_room;
2407 txn->mt_parent->mt_child = NULL;
2408 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2413 if (txn != env->me_txn) {
2414 DPUTS("attempt to commit unknown transaction");
2419 mdb_cursors_close(txn, 0);
2421 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2424 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2425 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2427 /* Update DB root pointers */
2428 if (txn->mt_numdbs > 2) {
2432 data.mv_size = sizeof(MDB_db);
2434 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2435 for (i = 2; i < txn->mt_numdbs; i++) {
2436 if (txn->mt_dbflags[i] & DB_DIRTY) {
2437 data.mv_data = &txn->mt_dbs[i];
2438 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2445 rc = mdb_freelist_save(txn);
2449 mdb_midl_free(env->me_pghead);
2450 env->me_pghead = NULL;
2451 if (mdb_midl_shrink(&txn->mt_free_pgs))
2452 env->me_free_pgs = txn->mt_free_pgs;
2458 if ((rc = mdb_page_flush(txn)) ||
2459 (rc = mdb_env_sync(env, 0)) ||
2460 (rc = mdb_env_write_meta(txn)))
2466 mdb_dbis_update(txn, 1);
2468 UNLOCK_MUTEX_W(env);
2478 /** Read the environment parameters of a DB environment before
2479 * mapping it into memory.
2480 * @param[in] env the environment handle
2481 * @param[out] meta address of where to store the meta information
2482 * @return 0 on success, non-zero on failure.
2485 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2492 /* We don't know the page size yet, so use a minimum value.
2493 * Read both meta pages so we can use the latest one.
2496 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2500 memset(&ov, 0, sizeof(ov));
2502 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2504 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2506 if (rc != MDB_PAGESIZE) {
2507 if (rc == 0 && off == 0)
2509 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2510 DPRINTF("read: %s", mdb_strerror(rc));
2514 p = (MDB_page *)&pbuf;
2516 if (!F_ISSET(p->mp_flags, P_META)) {
2517 DPRINTF("page %zu not a meta page", p->mp_pgno);
2522 if (m->mm_magic != MDB_MAGIC) {
2523 DPUTS("meta has invalid magic");
2527 if (m->mm_version != MDB_VERSION) {
2528 DPRINTF("database is version %u, expected version %u",
2529 m->mm_version, MDB_VERSION);
2530 return MDB_VERSION_MISMATCH;
2533 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2539 /** Write the environment parameters of a freshly created DB environment.
2540 * @param[in] env the environment handle
2541 * @param[out] meta address of where to store the meta information
2542 * @return 0 on success, non-zero on failure.
2545 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2551 DPUTS("writing new meta page");
2553 GET_PAGESIZE(psize);
2555 meta->mm_magic = MDB_MAGIC;
2556 meta->mm_version = MDB_VERSION;
2557 meta->mm_mapsize = env->me_mapsize;
2558 meta->mm_psize = psize;
2559 meta->mm_last_pg = 1;
2560 meta->mm_flags = env->me_flags & 0xffff;
2561 meta->mm_flags |= MDB_INTEGERKEY;
2562 meta->mm_dbs[0].md_root = P_INVALID;
2563 meta->mm_dbs[1].md_root = P_INVALID;
2565 p = calloc(2, psize);
2567 p->mp_flags = P_META;
2568 *(MDB_meta *)METADATA(p) = *meta;
2570 q = (MDB_page *)((char *)p + psize);
2572 q->mp_flags = P_META;
2573 *(MDB_meta *)METADATA(q) = *meta;
2579 memset(&ov, 0, sizeof(ov));
2580 rc = WriteFile(env->me_fd, p, psize * 2, &len, &ov);
2581 rc = rc ? (len == psize * 2 ? MDB_SUCCESS : EIO) : ErrCode();
2584 rc = pwrite(env->me_fd, p, psize * 2, 0);
2585 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
2591 /** Update the environment info to commit a transaction.
2592 * @param[in] txn the transaction that's being committed
2593 * @return 0 on success, non-zero on failure.
2596 mdb_env_write_meta(MDB_txn *txn)
2599 MDB_meta meta, metab, *mp;
2601 int rc, len, toggle;
2610 assert(txn != NULL);
2611 assert(txn->mt_env != NULL);
2613 toggle = !txn->mt_toggle;
2614 DPRINTF("writing meta page %d for root page %zu",
2615 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2618 mp = env->me_metas[toggle];
2620 if (env->me_flags & MDB_WRITEMAP) {
2621 /* Persist any increases of mapsize config */
2622 if (env->me_mapsize > mp->mm_mapsize)
2623 mp->mm_mapsize = env->me_mapsize;
2624 mp->mm_dbs[0] = txn->mt_dbs[0];
2625 mp->mm_dbs[1] = txn->mt_dbs[1];
2626 mp->mm_last_pg = txn->mt_next_pgno - 1;
2627 mp->mm_txnid = txn->mt_txnid;
2628 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2629 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2632 ptr += env->me_psize;
2633 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2640 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2641 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2643 ptr = (char *)&meta;
2644 if (env->me_mapsize > mp->mm_mapsize) {
2645 /* Persist any increases of mapsize config */
2646 meta.mm_mapsize = env->me_mapsize;
2647 off = offsetof(MDB_meta, mm_mapsize);
2649 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2651 len = sizeof(MDB_meta) - off;
2654 meta.mm_dbs[0] = txn->mt_dbs[0];
2655 meta.mm_dbs[1] = txn->mt_dbs[1];
2656 meta.mm_last_pg = txn->mt_next_pgno - 1;
2657 meta.mm_txnid = txn->mt_txnid;
2660 off += env->me_psize;
2663 /* Write to the SYNC fd */
2664 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2665 env->me_fd : env->me_mfd;
2668 memset(&ov, 0, sizeof(ov));
2670 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
2674 rc = pwrite(mfd, ptr, len, off);
2677 rc = rc < 0 ? ErrCode() : EIO;
2678 DPUTS("write failed, disk error?");
2679 /* On a failure, the pagecache still contains the new data.
2680 * Write some old data back, to prevent it from being used.
2681 * Use the non-SYNC fd; we know it will fail anyway.
2683 meta.mm_last_pg = metab.mm_last_pg;
2684 meta.mm_txnid = metab.mm_txnid;
2686 memset(&ov, 0, sizeof(ov));
2688 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2690 r2 = pwrite(env->me_fd, ptr, len, off);
2693 env->me_flags |= MDB_FATAL_ERROR;
2697 /* Memory ordering issues are irrelevant; since the entire writer
2698 * is wrapped by wmutex, all of these changes will become visible
2699 * after the wmutex is unlocked. Since the DB is multi-version,
2700 * readers will get consistent data regardless of how fresh or
2701 * how stale their view of these values is.
2703 env->me_txns->mti_txnid = txn->mt_txnid;
2708 /** Check both meta pages to see which one is newer.
2709 * @param[in] env the environment handle
2710 * @return meta toggle (0 or 1).
2713 mdb_env_pick_meta(const MDB_env *env)
2715 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2719 mdb_env_create(MDB_env **env)
2723 e = calloc(1, sizeof(MDB_env));
2727 e->me_maxreaders = DEFAULT_READERS;
2728 e->me_maxdbs = e->me_numdbs = 2;
2729 e->me_fd = INVALID_HANDLE_VALUE;
2730 e->me_lfd = INVALID_HANDLE_VALUE;
2731 e->me_mfd = INVALID_HANDLE_VALUE;
2732 #ifdef MDB_USE_POSIX_SEM
2733 e->me_rmutex = SEM_FAILED;
2734 e->me_wmutex = SEM_FAILED;
2736 e->me_pid = getpid();
2737 VGMEMP_CREATE(e,0,0);
2743 mdb_env_set_mapsize(MDB_env *env, size_t size)
2747 env->me_mapsize = size;
2749 env->me_maxpg = env->me_mapsize / env->me_psize;
2754 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2758 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2763 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2765 if (env->me_map || readers < 1)
2767 env->me_maxreaders = readers;
2772 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2774 if (!env || !readers)
2776 *readers = env->me_maxreaders;
2780 /** Further setup required for opening an MDB environment
2783 mdb_env_open2(MDB_env *env)
2785 unsigned int flags = env->me_flags;
2793 memset(&meta, 0, sizeof(meta));
2795 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2798 DPUTS("new mdbenv");
2802 /* Was a mapsize configured? */
2803 if (!env->me_mapsize) {
2804 /* If this is a new environment, take the default,
2805 * else use the size recorded in the existing env.
2807 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2808 } else if (env->me_mapsize < meta.mm_mapsize) {
2809 /* If the configured size is smaller, make sure it's
2810 * still big enough. Silently round up to minimum if not.
2812 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2813 if (env->me_mapsize < minsize)
2814 env->me_mapsize = minsize;
2821 LONG sizelo, sizehi;
2822 sizelo = env->me_mapsize & 0xffffffff;
2823 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
2824 /* Windows won't create mappings for zero length files.
2825 * Just allocate the maxsize right now.
2828 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
2829 || !SetEndOfFile(env->me_fd)
2830 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
2833 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2834 PAGE_READWRITE : PAGE_READONLY,
2835 sizehi, sizelo, NULL);
2838 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2839 FILE_MAP_WRITE : FILE_MAP_READ,
2840 0, 0, env->me_mapsize, meta.mm_address);
2841 rc = env->me_map ? 0 : ErrCode();
2849 if (flags & MDB_WRITEMAP) {
2851 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2854 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2856 if (env->me_map == MAP_FAILED) {
2860 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2862 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2864 #ifdef POSIX_MADV_RANDOM
2865 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2866 #endif /* POSIX_MADV_RANDOM */
2867 #endif /* MADV_RANDOM */
2871 if (flags & MDB_FIXEDMAP)
2872 meta.mm_address = env->me_map;
2873 i = mdb_env_init_meta(env, &meta);
2874 if (i != MDB_SUCCESS) {
2877 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2878 /* Can happen because the address argument to mmap() is just a
2879 * hint. mmap() can pick another, e.g. if the range is in use.
2880 * The MAP_FIXED flag would prevent that, but then mmap could
2881 * instead unmap existing pages to make room for the new map.
2883 return EBUSY; /* TODO: Make a new MDB_* error code? */
2885 env->me_psize = meta.mm_psize;
2886 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2887 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2889 env->me_maxpg = env->me_mapsize / env->me_psize;
2891 p = (MDB_page *)env->me_map;
2892 env->me_metas[0] = METADATA(p);
2893 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2897 int toggle = mdb_env_pick_meta(env);
2898 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2900 DPRINTF("opened database version %u, pagesize %u",
2901 env->me_metas[0]->mm_version, env->me_psize);
2902 DPRINTF("using meta page %d", toggle);
2903 DPRINTF("depth: %u", db->md_depth);
2904 DPRINTF("entries: %zu", db->md_entries);
2905 DPRINTF("branch pages: %zu", db->md_branch_pages);
2906 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2907 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2908 DPRINTF("root: %zu", db->md_root);
2916 /** Release a reader thread's slot in the reader lock table.
2917 * This function is called automatically when a thread exits.
2918 * @param[in] ptr This points to the slot in the reader lock table.
2921 mdb_env_reader_dest(void *ptr)
2923 MDB_reader *reader = ptr;
2929 /** Junk for arranging thread-specific callbacks on Windows. This is
2930 * necessarily platform and compiler-specific. Windows supports up
2931 * to 1088 keys. Let's assume nobody opens more than 64 environments
2932 * in a single process, for now. They can override this if needed.
2934 #ifndef MAX_TLS_KEYS
2935 #define MAX_TLS_KEYS 64
2937 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2938 static int mdb_tls_nkeys;
2940 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2944 case DLL_PROCESS_ATTACH: break;
2945 case DLL_THREAD_ATTACH: break;
2946 case DLL_THREAD_DETACH:
2947 for (i=0; i<mdb_tls_nkeys; i++) {
2948 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2949 mdb_env_reader_dest(r);
2952 case DLL_PROCESS_DETACH: break;
2957 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2959 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2963 /* Force some symbol references.
2964 * _tls_used forces the linker to create the TLS directory if not already done
2965 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2967 #pragma comment(linker, "/INCLUDE:_tls_used")
2968 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2969 #pragma const_seg(".CRT$XLB")
2970 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2971 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2974 #pragma comment(linker, "/INCLUDE:__tls_used")
2975 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2976 #pragma data_seg(".CRT$XLB")
2977 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2979 #endif /* WIN 32/64 */
2980 #endif /* !__GNUC__ */
2983 /** Downgrade the exclusive lock on the region back to shared */
2985 mdb_env_share_locks(MDB_env *env, int *excl)
2987 int rc = 0, toggle = mdb_env_pick_meta(env);
2989 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2994 /* First acquire a shared lock. The Unlock will
2995 * then release the existing exclusive lock.
2997 memset(&ov, 0, sizeof(ov));
2998 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3001 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3007 struct flock lock_info;
3008 /* The shared lock replaces the existing lock */
3009 memset((void *)&lock_info, 0, sizeof(lock_info));
3010 lock_info.l_type = F_RDLCK;
3011 lock_info.l_whence = SEEK_SET;
3012 lock_info.l_start = 0;
3013 lock_info.l_len = 1;
3014 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3015 (rc = ErrCode()) == EINTR) ;
3016 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3023 /** Try to get exlusive lock, otherwise shared.
3024 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3027 mdb_env_excl_lock(MDB_env *env, int *excl)
3031 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3035 memset(&ov, 0, sizeof(ov));
3036 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3043 struct flock lock_info;
3044 memset((void *)&lock_info, 0, sizeof(lock_info));
3045 lock_info.l_type = F_WRLCK;
3046 lock_info.l_whence = SEEK_SET;
3047 lock_info.l_start = 0;
3048 lock_info.l_len = 1;
3049 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3050 (rc = ErrCode()) == EINTR) ;
3054 # ifdef MDB_USE_POSIX_SEM
3055 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3058 lock_info.l_type = F_RDLCK;
3059 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3060 (rc = ErrCode()) == EINTR) ;
3068 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3070 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3072 * @(#) $Revision: 5.1 $
3073 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3074 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3076 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3080 * Please do not copyright this code. This code is in the public domain.
3082 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3083 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3084 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3085 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3086 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3087 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3088 * PERFORMANCE OF THIS SOFTWARE.
3091 * chongo <Landon Curt Noll> /\oo/\
3092 * http://www.isthe.com/chongo/
3094 * Share and Enjoy! :-)
3097 typedef unsigned long long mdb_hash_t;
3098 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3100 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3101 * @param[in] str string to hash
3102 * @param[in] hval initial value for hash
3103 * @return 64 bit hash
3105 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3106 * hval arg on the first call.
3109 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3111 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3112 unsigned char *end = s + val->mv_size;
3114 * FNV-1a hash each octet of the string
3117 /* xor the bottom with the current octet */
3118 hval ^= (mdb_hash_t)*s++;
3120 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3121 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3122 (hval << 7) + (hval << 8) + (hval << 40);
3124 /* return our new hash value */
3128 /** Hash the string and output the hash in hex.
3129 * @param[in] str string to hash
3130 * @param[out] hexbuf an array of 17 chars to hold the hash
3133 mdb_hash_hex(MDB_val *val, char *hexbuf)
3136 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3137 for (i=0; i<8; i++) {
3138 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3144 /** Open and/or initialize the lock region for the environment.
3145 * @param[in] env The MDB environment.
3146 * @param[in] lpath The pathname of the file used for the lock region.
3147 * @param[in] mode The Unix permissions for the file, if we create it.
3148 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3149 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3150 * @return 0 on success, non-zero on failure.
3153 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3156 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3158 # define MDB_ERRCODE_ROFS EROFS
3159 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3160 # define MDB_CLOEXEC O_CLOEXEC
3163 # define MDB_CLOEXEC 0
3170 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3171 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3172 FILE_ATTRIBUTE_NORMAL, NULL);
3174 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3176 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3178 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3183 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3184 /* Lose record locks when exec*() */
3185 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3186 fcntl(env->me_lfd, F_SETFD, fdflags);
3189 if (!(env->me_flags & MDB_NOTLS)) {
3190 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3193 env->me_flags |= MDB_ENV_TXKEY;
3195 /* Windows TLS callbacks need help finding their TLS info. */
3196 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3200 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3204 /* Try to get exclusive lock. If we succeed, then
3205 * nobody is using the lock region and we should initialize it.
3207 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3210 size = GetFileSize(env->me_lfd, NULL);
3212 size = lseek(env->me_lfd, 0, SEEK_END);
3213 if (size == -1) goto fail_errno;
3215 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3216 if (size < rsize && *excl > 0) {
3218 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3219 || !SetEndOfFile(env->me_lfd))
3222 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3226 size = rsize - sizeof(MDB_txninfo);
3227 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3232 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3234 if (!mh) goto fail_errno;
3235 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3237 if (!env->me_txns) goto fail_errno;
3239 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3241 if (m == MAP_FAILED) goto fail_errno;
3247 BY_HANDLE_FILE_INFORMATION stbuf;
3256 if (!mdb_sec_inited) {
3257 InitializeSecurityDescriptor(&mdb_null_sd,
3258 SECURITY_DESCRIPTOR_REVISION);
3259 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3260 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3261 mdb_all_sa.bInheritHandle = FALSE;
3262 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3265 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3266 idbuf.volume = stbuf.dwVolumeSerialNumber;
3267 idbuf.nhigh = stbuf.nFileIndexHigh;
3268 idbuf.nlow = stbuf.nFileIndexLow;
3269 val.mv_data = &idbuf;
3270 val.mv_size = sizeof(idbuf);
3271 mdb_hash_hex(&val, hexbuf);
3272 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3273 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3274 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3275 if (!env->me_rmutex) goto fail_errno;
3276 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3277 if (!env->me_wmutex) goto fail_errno;
3278 #elif defined(MDB_USE_POSIX_SEM)
3287 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3288 idbuf.dev = stbuf.st_dev;
3289 idbuf.ino = stbuf.st_ino;
3290 val.mv_data = &idbuf;
3291 val.mv_size = sizeof(idbuf);
3292 mdb_hash_hex(&val, hexbuf);
3293 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3294 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3295 /* Clean up after a previous run, if needed: Try to
3296 * remove both semaphores before doing anything else.
3298 sem_unlink(env->me_txns->mti_rmname);
3299 sem_unlink(env->me_txns->mti_wmname);
3300 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3301 O_CREAT|O_EXCL, mode, 1);
3302 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3303 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3304 O_CREAT|O_EXCL, mode, 1);
3305 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3306 #else /* MDB_USE_POSIX_SEM */
3307 pthread_mutexattr_t mattr;
3309 if ((rc = pthread_mutexattr_init(&mattr))
3310 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3311 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3312 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3314 pthread_mutexattr_destroy(&mattr);
3315 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3317 env->me_txns->mti_version = MDB_VERSION;
3318 env->me_txns->mti_magic = MDB_MAGIC;
3319 env->me_txns->mti_txnid = 0;
3320 env->me_txns->mti_numreaders = 0;
3323 if (env->me_txns->mti_magic != MDB_MAGIC) {
3324 DPUTS("lock region has invalid magic");
3328 if (env->me_txns->mti_version != MDB_VERSION) {
3329 DPRINTF("lock region is version %u, expected version %u",
3330 env->me_txns->mti_version, MDB_VERSION);
3331 rc = MDB_VERSION_MISMATCH;
3335 if (rc && rc != EACCES && rc != EAGAIN) {
3339 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3340 if (!env->me_rmutex) goto fail_errno;
3341 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3342 if (!env->me_wmutex) goto fail_errno;
3343 #elif defined(MDB_USE_POSIX_SEM)
3344 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3345 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3346 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3347 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3358 /** The name of the lock file in the DB environment */
3359 #define LOCKNAME "/lock.mdb"
3360 /** The name of the data file in the DB environment */
3361 #define DATANAME "/data.mdb"
3362 /** The suffix of the lock file when no subdir is used */
3363 #define LOCKSUFF "-lock"
3364 /** Only a subset of the @ref mdb_env flags can be changed
3365 * at runtime. Changing other flags requires closing the
3366 * environment and re-opening it with the new flags.
3368 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3369 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3372 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3374 int oflags, rc, len, excl = -1;
3375 char *lpath, *dpath;
3377 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3381 if (flags & MDB_NOSUBDIR) {
3382 rc = len + sizeof(LOCKSUFF) + len + 1;
3384 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3389 if (flags & MDB_NOSUBDIR) {
3390 dpath = lpath + len + sizeof(LOCKSUFF);
3391 sprintf(lpath, "%s" LOCKSUFF, path);
3392 strcpy(dpath, path);
3394 dpath = lpath + len + sizeof(LOCKNAME);
3395 sprintf(lpath, "%s" LOCKNAME, path);
3396 sprintf(dpath, "%s" DATANAME, path);
3400 flags |= env->me_flags;
3401 if (flags & MDB_RDONLY) {
3402 /* silently ignore WRITEMAP when we're only getting read access */
3403 flags &= ~MDB_WRITEMAP;
3405 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3406 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3409 env->me_flags = flags |= MDB_ENV_ACTIVE;
3413 env->me_path = strdup(path);
3414 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3415 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3416 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3421 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3426 if (F_ISSET(flags, MDB_RDONLY)) {
3427 oflags = GENERIC_READ;
3428 len = OPEN_EXISTING;
3430 oflags = GENERIC_READ|GENERIC_WRITE;
3433 mode = FILE_ATTRIBUTE_NORMAL;
3434 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3435 NULL, len, mode, NULL);
3437 if (F_ISSET(flags, MDB_RDONLY))
3440 oflags = O_RDWR | O_CREAT;
3442 env->me_fd = open(dpath, oflags, mode);
3444 if (env->me_fd == INVALID_HANDLE_VALUE) {
3449 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3450 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3451 env->me_mfd = env->me_fd;
3453 /* Synchronous fd for meta writes. Needed even with
3454 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3457 env->me_mfd = CreateFile(dpath, oflags,
3458 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3459 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3461 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3463 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3468 DPRINTF("opened dbenv %p", (void *) env);
3470 rc = mdb_env_share_locks(env, &excl);
3476 mdb_env_close0(env, excl);
3482 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3484 mdb_env_close0(MDB_env *env, int excl)
3488 if (!(env->me_flags & MDB_ENV_ACTIVE))
3491 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3492 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3493 free(env->me_dbxs[i].md_name.mv_data);
3495 free(env->me_dbflags);
3498 free(env->me_dirty_list);
3499 mdb_midl_free(env->me_free_pgs);
3501 if (env->me_flags & MDB_ENV_TXKEY) {
3502 pthread_key_delete(env->me_txkey);
3504 /* Delete our key from the global list */
3505 for (i=0; i<mdb_tls_nkeys; i++)
3506 if (mdb_tls_keys[i] == env->me_txkey) {
3507 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3515 munmap(env->me_map, env->me_mapsize);
3517 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3518 (void) close(env->me_mfd);
3519 if (env->me_fd != INVALID_HANDLE_VALUE)
3520 (void) close(env->me_fd);
3522 pid_t pid = env->me_pid;
3523 /* Clearing readers is done in this function because
3524 * me_txkey with its destructor must be disabled first.
3526 for (i = env->me_numreaders; --i >= 0; )
3527 if (env->me_txns->mti_readers[i].mr_pid == pid)
3528 env->me_txns->mti_readers[i].mr_pid = 0;
3530 if (env->me_rmutex) {
3531 CloseHandle(env->me_rmutex);
3532 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3534 /* Windows automatically destroys the mutexes when
3535 * the last handle closes.
3537 #elif defined(MDB_USE_POSIX_SEM)
3538 if (env->me_rmutex != SEM_FAILED) {
3539 sem_close(env->me_rmutex);
3540 if (env->me_wmutex != SEM_FAILED)
3541 sem_close(env->me_wmutex);
3542 /* If we have the filelock: If we are the
3543 * only remaining user, clean up semaphores.
3546 mdb_env_excl_lock(env, &excl);
3548 sem_unlink(env->me_txns->mti_rmname);
3549 sem_unlink(env->me_txns->mti_wmname);
3553 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3555 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3558 /* Unlock the lockfile. Windows would have unlocked it
3559 * after closing anyway, but not necessarily at once.
3561 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3564 (void) close(env->me_lfd);
3567 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3571 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3573 MDB_txn *txn = NULL;
3578 /* Do the lock/unlock of the reader mutex before starting the
3579 * write txn. Otherwise other read txns could block writers.
3581 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3586 /* We must start the actual read txn after blocking writers */
3587 mdb_txn_reset0(txn, "reset-stage1");
3589 /* Temporarily block writers until we snapshot the meta pages */
3592 rc = mdb_txn_renew0(txn);
3594 UNLOCK_MUTEX_W(env);
3599 wsize = env->me_psize * 2;
3603 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3604 rc = rc ? (len == wsize ? MDB_SUCCESS : EIO) : ErrCode();
3607 rc = write(fd, env->me_map, wsize);
3608 rc = rc == (int)wsize ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
3611 UNLOCK_MUTEX_W(env);
3616 ptr = env->me_map + wsize;
3617 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3621 if (wsize > MAX_WRITE)
3625 rc = WriteFile(fd, ptr, w2, &len, NULL);
3626 rc = rc ? (len == w2 ? MDB_SUCCESS : EIO) : ErrCode();
3635 if (wsize > MAX_WRITE)
3639 wres = write(fd, ptr, w2);
3640 rc = wres == (ssize_t)w2 ? MDB_SUCCESS : wres < 0 ? ErrCode() : EIO;
3653 mdb_env_copy(MDB_env *env, const char *path)
3657 HANDLE newfd = INVALID_HANDLE_VALUE;
3659 if (env->me_flags & MDB_NOSUBDIR) {
3660 lpath = (char *)path;
3663 len += sizeof(DATANAME);
3664 lpath = malloc(len);
3667 sprintf(lpath, "%s" DATANAME, path);
3670 /* The destination path must exist, but the destination file must not.
3671 * We don't want the OS to cache the writes, since the source data is
3672 * already in the OS cache.
3675 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3676 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3678 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3684 if (newfd == INVALID_HANDLE_VALUE) {
3689 #ifdef F_NOCACHE /* __APPLE__ */
3690 rc = fcntl(newfd, F_NOCACHE, 1);
3697 rc = mdb_env_copyfd(env, newfd);
3700 if (!(env->me_flags & MDB_NOSUBDIR))
3702 if (newfd != INVALID_HANDLE_VALUE)
3703 if (close(newfd) < 0 && rc == MDB_SUCCESS)
3710 mdb_env_close(MDB_env *env)
3717 VGMEMP_DESTROY(env);
3718 while ((dp = env->me_dpages) != NULL) {
3719 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3720 env->me_dpages = dp->mp_next;
3724 mdb_env_close0(env, 0);
3728 /** Compare two items pointing at aligned size_t's */
3730 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3732 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3733 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3736 /** Compare two items pointing at aligned int's */
3738 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3740 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3741 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3744 /** Compare two items pointing at ints of unknown alignment.
3745 * Nodes and keys are guaranteed to be 2-byte aligned.
3748 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3750 #if BYTE_ORDER == LITTLE_ENDIAN
3751 unsigned short *u, *c;
3754 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3755 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3758 } while(!x && u > (unsigned short *)a->mv_data);
3761 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3765 /** Compare two items lexically */
3767 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3774 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3780 diff = memcmp(a->mv_data, b->mv_data, len);
3781 return diff ? diff : len_diff<0 ? -1 : len_diff;
3784 /** Compare two items in reverse byte order */
3786 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3788 const unsigned char *p1, *p2, *p1_lim;
3792 p1_lim = (const unsigned char *)a->mv_data;
3793 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3794 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3796 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3802 while (p1 > p1_lim) {
3803 diff = *--p1 - *--p2;
3807 return len_diff<0 ? -1 : len_diff;
3810 /** Search for key within a page, using binary search.
3811 * Returns the smallest entry larger or equal to the key.
3812 * If exactp is non-null, stores whether the found entry was an exact match
3813 * in *exactp (1 or 0).
3814 * Updates the cursor index with the index of the found entry.
3815 * If no entry larger or equal to the key is found, returns NULL.
3818 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3820 unsigned int i = 0, nkeys;
3823 MDB_page *mp = mc->mc_pg[mc->mc_top];
3824 MDB_node *node = NULL;
3829 nkeys = NUMKEYS(mp);
3834 COPY_PGNO(pgno, mp->mp_pgno);
3835 DPRINTF("searching %u keys in %s %spage %zu",
3836 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3843 low = IS_LEAF(mp) ? 0 : 1;
3845 cmp = mc->mc_dbx->md_cmp;
3847 /* Branch pages have no data, so if using integer keys,
3848 * alignment is guaranteed. Use faster mdb_cmp_int.
3850 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3851 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3858 nodekey.mv_size = mc->mc_db->md_pad;
3859 node = NODEPTR(mp, 0); /* fake */
3860 while (low <= high) {
3861 i = (low + high) >> 1;
3862 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3863 rc = cmp(key, &nodekey);
3864 DPRINTF("found leaf index %u [%s], rc = %i",
3865 i, DKEY(&nodekey), rc);
3874 while (low <= high) {
3875 i = (low + high) >> 1;
3877 node = NODEPTR(mp, i);
3878 nodekey.mv_size = NODEKSZ(node);
3879 nodekey.mv_data = NODEKEY(node);
3881 rc = cmp(key, &nodekey);
3884 DPRINTF("found leaf index %u [%s], rc = %i",
3885 i, DKEY(&nodekey), rc);
3887 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3888 i, DKEY(&nodekey), NODEPGNO(node), rc);
3899 if (rc > 0) { /* Found entry is less than the key. */
3900 i++; /* Skip to get the smallest entry larger than key. */
3902 node = NODEPTR(mp, i);
3905 *exactp = (rc == 0);
3906 /* store the key index */
3907 mc->mc_ki[mc->mc_top] = i;
3909 /* There is no entry larger or equal to the key. */
3912 /* nodeptr is fake for LEAF2 */
3918 mdb_cursor_adjust(MDB_cursor *mc, func)
3922 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3923 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3930 /** Pop a page off the top of the cursor's stack. */
3932 mdb_cursor_pop(MDB_cursor *mc)
3935 #ifndef MDB_DEBUG_SKIP
3936 MDB_page *top = mc->mc_pg[mc->mc_top];
3942 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3943 mc->mc_dbi, (void *) mc);
3947 /** Push a page onto the top of the cursor's stack. */
3949 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3951 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3952 mc->mc_dbi, (void *) mc);
3954 if (mc->mc_snum >= CURSOR_STACK) {
3955 assert(mc->mc_snum < CURSOR_STACK);
3956 return MDB_CURSOR_FULL;
3959 mc->mc_top = mc->mc_snum++;
3960 mc->mc_pg[mc->mc_top] = mp;
3961 mc->mc_ki[mc->mc_top] = 0;
3966 /** Find the address of the page corresponding to a given page number.
3967 * @param[in] txn the transaction for this access.
3968 * @param[in] pgno the page number for the page to retrieve.
3969 * @param[out] ret address of a pointer where the page's address will be stored.
3970 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
3971 * @return 0 on success, non-zero on failure.
3974 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
3979 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3980 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3985 MDB_ID2L dl = tx2->mt_u.dirty_list;
3987 unsigned x = mdb_mid2l_search(dl, pgno);
3988 if (x <= dl[0].mid && dl[x].mid == pgno) {
3994 } while ((tx2 = tx2->mt_parent) != NULL);
3997 if (pgno < txn->mt_next_pgno) {
3999 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4001 DPRINTF("page %zu not found", pgno);
4003 return MDB_PAGE_NOTFOUND;
4013 /** Search for the page a given key should be in.
4014 * Pushes parent pages on the cursor stack. This function continues a
4015 * search on a cursor that has already been initialized. (Usually by
4016 * #mdb_page_search() but also by #mdb_node_move().)
4017 * @param[in,out] mc the cursor for this operation.
4018 * @param[in] key the key to search for. If NULL, search for the lowest
4019 * page. (This is used by #mdb_cursor_first().)
4020 * @param[in] modify If true, visited pages are updated with new page numbers.
4021 * @return 0 on success, non-zero on failure.
4024 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4026 MDB_page *mp = mc->mc_pg[mc->mc_top];
4031 while (IS_BRANCH(mp)) {
4035 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4036 assert(NUMKEYS(mp) > 1);
4037 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4039 if (key == NULL) /* Initialize cursor to first page. */
4041 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4042 /* cursor to last page */
4046 node = mdb_node_search(mc, key, &exact);
4048 i = NUMKEYS(mp) - 1;
4050 i = mc->mc_ki[mc->mc_top];
4059 DPRINTF("following index %u for key [%s]",
4061 assert(i < NUMKEYS(mp));
4062 node = NODEPTR(mp, i);
4064 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4067 mc->mc_ki[mc->mc_top] = i;
4068 if ((rc = mdb_cursor_push(mc, mp)))
4072 if ((rc = mdb_page_touch(mc)) != 0)
4074 mp = mc->mc_pg[mc->mc_top];
4079 DPRINTF("internal error, index points to a %02X page!?",
4081 return MDB_CORRUPTED;
4084 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4085 key ? DKEY(key) : NULL);
4090 /** Search for the lowest key under the current branch page.
4091 * This just bypasses a NUMKEYS check in the current page
4092 * before calling mdb_page_search_root(), because the callers
4093 * are all in situations where the current page is known to
4097 mdb_page_search_lowest(MDB_cursor *mc)
4099 MDB_page *mp = mc->mc_pg[mc->mc_top];
4100 MDB_node *node = NODEPTR(mp, 0);
4103 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4106 mc->mc_ki[mc->mc_top] = 0;
4107 if ((rc = mdb_cursor_push(mc, mp)))
4109 return mdb_page_search_root(mc, NULL, 0);
4112 /** Search for the page a given key should be in.
4113 * Pushes parent pages on the cursor stack. This function just sets up
4114 * the search; it finds the root page for \b mc's database and sets this
4115 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4116 * called to complete the search.
4117 * @param[in,out] mc the cursor for this operation.
4118 * @param[in] key the key to search for. If NULL, search for the lowest
4119 * page. (This is used by #mdb_cursor_first().)
4120 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4121 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4122 * @return 0 on success, non-zero on failure.
4125 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4130 /* Make sure the txn is still viable, then find the root from
4131 * the txn's db table.
4133 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4134 DPUTS("transaction has failed, must abort");
4137 /* Make sure we're using an up-to-date root */
4138 if (mc->mc_dbi > MAIN_DBI) {
4139 if ((*mc->mc_dbflag & DB_STALE) ||
4140 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4142 unsigned char dbflag = 0;
4143 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4144 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4147 if (*mc->mc_dbflag & DB_STALE) {
4151 MDB_node *leaf = mdb_node_search(&mc2,
4152 &mc->mc_dbx->md_name, &exact);
4154 return MDB_NOTFOUND;
4155 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4158 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4160 /* The txn may not know this DBI, or another process may
4161 * have dropped and recreated the DB with other flags.
4163 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4164 return MDB_INCOMPATIBLE;
4165 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4167 if (flags & MDB_PS_MODIFY)
4169 *mc->mc_dbflag &= ~DB_STALE;
4170 *mc->mc_dbflag |= dbflag;
4173 root = mc->mc_db->md_root;
4175 if (root == P_INVALID) { /* Tree is empty. */
4176 DPUTS("tree is empty");
4177 return MDB_NOTFOUND;
4182 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4183 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4189 DPRINTF("db %u root page %zu has flags 0x%X",
4190 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4192 if (flags & MDB_PS_MODIFY) {
4193 if ((rc = mdb_page_touch(mc)))
4197 if (flags & MDB_PS_ROOTONLY)
4200 return mdb_page_search_root(mc, key, flags);
4204 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4206 MDB_txn *txn = mc->mc_txn;
4207 pgno_t pg = mp->mp_pgno;
4208 unsigned i, ovpages = mp->mp_pages;
4209 MDB_env *env = txn->mt_env;
4212 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4213 /* If the page is dirty we just acquired it, so we should
4214 * give it back to our current free list, if any.
4215 * Not currently supported in nested txns.
4216 * Otherwise put it onto the list of pages we freed in this txn.
4218 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && env->me_pghead) {
4221 MDB_ID2 *dl, ix, iy;
4222 rc = mdb_midl_need(&env->me_pghead, ovpages);
4225 /* Remove from dirty list */
4226 dl = txn->mt_u.dirty_list;
4228 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4236 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4237 txn->mt_flags |= MDB_TXN_ERROR;
4238 return MDB_CORRUPTED;
4241 if (!(env->me_flags & MDB_WRITEMAP))
4242 mdb_dpage_free(env, mp);
4243 /* Insert in me_pghead */
4244 mop = env->me_pghead;
4245 j = mop[0] + ovpages;
4246 for (i = mop[0]; i && mop[i] < pg; i--)
4252 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4256 mc->mc_db->md_overflow_pages -= ovpages;
4260 /** Return the data associated with a given node.
4261 * @param[in] txn The transaction for this operation.
4262 * @param[in] leaf The node being read.
4263 * @param[out] data Updated to point to the node's data.
4264 * @return 0 on success, non-zero on failure.
4267 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4269 MDB_page *omp; /* overflow page */
4273 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4274 data->mv_size = NODEDSZ(leaf);
4275 data->mv_data = NODEDATA(leaf);
4279 /* Read overflow data.
4281 data->mv_size = NODEDSZ(leaf);
4282 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4283 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4284 DPRINTF("read overflow page %zu failed", pgno);
4287 data->mv_data = METADATA(omp);
4293 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4294 MDB_val *key, MDB_val *data)
4303 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4305 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4308 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4312 mdb_cursor_init(&mc, txn, dbi, &mx);
4313 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4316 /** Find a sibling for a page.
4317 * Replaces the page at the top of the cursor's stack with the
4318 * specified sibling, if one exists.
4319 * @param[in] mc The cursor for this operation.
4320 * @param[in] move_right Non-zero if the right sibling is requested,
4321 * otherwise the left sibling.
4322 * @return 0 on success, non-zero on failure.
4325 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4331 if (mc->mc_snum < 2) {
4332 return MDB_NOTFOUND; /* root has no siblings */
4336 DPRINTF("parent page is page %zu, index %u",
4337 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4339 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4340 : (mc->mc_ki[mc->mc_top] == 0)) {
4341 DPRINTF("no more keys left, moving to %s sibling",
4342 move_right ? "right" : "left");
4343 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4344 /* undo cursor_pop before returning */
4351 mc->mc_ki[mc->mc_top]++;
4353 mc->mc_ki[mc->mc_top]--;
4354 DPRINTF("just moving to %s index key %u",
4355 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4357 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4359 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4360 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4363 mdb_cursor_push(mc, mp);
4365 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4370 /** Move the cursor to the next data item. */
4372 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4378 if (mc->mc_flags & C_EOF) {
4379 return MDB_NOTFOUND;
4382 assert(mc->mc_flags & C_INITIALIZED);
4384 mp = mc->mc_pg[mc->mc_top];
4386 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4387 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4388 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4389 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4390 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4391 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4395 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4396 if (op == MDB_NEXT_DUP)
4397 return MDB_NOTFOUND;
4401 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4403 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4404 DPUTS("=====> move to next sibling page");
4405 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4406 mc->mc_flags |= C_EOF;
4407 mc->mc_flags &= ~C_INITIALIZED;
4410 mp = mc->mc_pg[mc->mc_top];
4411 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4413 mc->mc_ki[mc->mc_top]++;
4415 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4416 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4419 key->mv_size = mc->mc_db->md_pad;
4420 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4424 assert(IS_LEAF(mp));
4425 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4427 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4428 mdb_xcursor_init1(mc, leaf);
4431 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4434 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4435 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4436 if (rc != MDB_SUCCESS)
4441 MDB_GET_KEY(leaf, key);
4445 /** Move the cursor to the previous data item. */
4447 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4453 assert(mc->mc_flags & C_INITIALIZED);
4455 mp = mc->mc_pg[mc->mc_top];
4457 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4458 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4459 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4460 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4461 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4462 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4465 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4466 if (op == MDB_PREV_DUP)
4467 return MDB_NOTFOUND;
4472 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4474 if (mc->mc_ki[mc->mc_top] == 0) {
4475 DPUTS("=====> move to prev sibling page");
4476 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
4477 mc->mc_flags &= ~C_INITIALIZED;
4480 mp = mc->mc_pg[mc->mc_top];
4481 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4482 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4484 mc->mc_ki[mc->mc_top]--;
4486 mc->mc_flags &= ~C_EOF;
4488 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4489 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4492 key->mv_size = mc->mc_db->md_pad;
4493 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4497 assert(IS_LEAF(mp));
4498 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4500 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4501 mdb_xcursor_init1(mc, leaf);
4504 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4507 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4508 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4509 if (rc != MDB_SUCCESS)
4514 MDB_GET_KEY(leaf, key);
4518 /** Set the cursor on a specific data item. */
4520 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4521 MDB_cursor_op op, int *exactp)
4525 MDB_node *leaf = NULL;
4530 assert(key->mv_size > 0);
4532 /* See if we're already on the right page */
4533 if (mc->mc_flags & C_INITIALIZED) {
4536 mp = mc->mc_pg[mc->mc_top];
4538 mc->mc_ki[mc->mc_top] = 0;
4539 return MDB_NOTFOUND;
4541 if (mp->mp_flags & P_LEAF2) {
4542 nodekey.mv_size = mc->mc_db->md_pad;
4543 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4545 leaf = NODEPTR(mp, 0);
4546 MDB_GET_KEY(leaf, &nodekey);
4548 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4550 /* Probably happens rarely, but first node on the page
4551 * was the one we wanted.
4553 mc->mc_ki[mc->mc_top] = 0;
4560 unsigned int nkeys = NUMKEYS(mp);
4562 if (mp->mp_flags & P_LEAF2) {
4563 nodekey.mv_data = LEAF2KEY(mp,
4564 nkeys-1, nodekey.mv_size);
4566 leaf = NODEPTR(mp, nkeys-1);
4567 MDB_GET_KEY(leaf, &nodekey);
4569 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4571 /* last node was the one we wanted */
4572 mc->mc_ki[mc->mc_top] = nkeys-1;
4578 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4579 /* This is definitely the right page, skip search_page */
4580 if (mp->mp_flags & P_LEAF2) {
4581 nodekey.mv_data = LEAF2KEY(mp,
4582 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4584 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4585 MDB_GET_KEY(leaf, &nodekey);
4587 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4589 /* current node was the one we wanted */
4599 /* If any parents have right-sibs, search.
4600 * Otherwise, there's nothing further.
4602 for (i=0; i<mc->mc_top; i++)
4604 NUMKEYS(mc->mc_pg[i])-1)
4606 if (i == mc->mc_top) {
4607 /* There are no other pages */
4608 mc->mc_ki[mc->mc_top] = nkeys;
4609 return MDB_NOTFOUND;
4613 /* There are no other pages */
4614 mc->mc_ki[mc->mc_top] = 0;
4615 return MDB_NOTFOUND;
4619 rc = mdb_page_search(mc, key, 0);
4620 if (rc != MDB_SUCCESS)
4623 mp = mc->mc_pg[mc->mc_top];
4624 assert(IS_LEAF(mp));
4627 leaf = mdb_node_search(mc, key, exactp);
4628 if (exactp != NULL && !*exactp) {
4629 /* MDB_SET specified and not an exact match. */
4630 return MDB_NOTFOUND;
4634 DPUTS("===> inexact leaf not found, goto sibling");
4635 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4636 return rc; /* no entries matched */
4637 mp = mc->mc_pg[mc->mc_top];
4638 assert(IS_LEAF(mp));
4639 leaf = NODEPTR(mp, 0);
4643 mc->mc_flags |= C_INITIALIZED;
4644 mc->mc_flags &= ~C_EOF;
4647 key->mv_size = mc->mc_db->md_pad;
4648 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4652 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4653 mdb_xcursor_init1(mc, leaf);
4656 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4657 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4658 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4661 if (op == MDB_GET_BOTH) {
4667 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4668 if (rc != MDB_SUCCESS)
4671 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4673 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4675 rc = mc->mc_dbx->md_dcmp(data, &d2);
4677 if (op == MDB_GET_BOTH || rc > 0)
4678 return MDB_NOTFOUND;
4683 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4684 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4689 /* The key already matches in all other cases */
4690 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4691 MDB_GET_KEY(leaf, key);
4692 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4697 /** Move the cursor to the first item in the database. */
4699 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4704 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4705 rc = mdb_page_search(mc, NULL, 0);
4706 if (rc != MDB_SUCCESS)
4709 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4711 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4712 mc->mc_flags |= C_INITIALIZED;
4713 mc->mc_flags &= ~C_EOF;
4715 mc->mc_ki[mc->mc_top] = 0;
4717 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4718 key->mv_size = mc->mc_db->md_pad;
4719 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4724 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4725 mdb_xcursor_init1(mc, leaf);
4726 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4731 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4732 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4736 MDB_GET_KEY(leaf, key);
4740 /** Move the cursor to the last item in the database. */
4742 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4747 if (!(mc->mc_flags & C_EOF)) {
4749 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4752 lkey.mv_size = MDB_MAXKEYSIZE+1;
4753 lkey.mv_data = NULL;
4754 rc = mdb_page_search(mc, &lkey, 0);
4755 if (rc != MDB_SUCCESS)
4758 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4761 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4762 mc->mc_flags |= C_INITIALIZED|C_EOF;
4763 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4765 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4766 key->mv_size = mc->mc_db->md_pad;
4767 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4772 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4773 mdb_xcursor_init1(mc, leaf);
4774 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4779 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4780 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4785 MDB_GET_KEY(leaf, key);
4790 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4799 case MDB_GET_CURRENT:
4800 if (!(mc->mc_flags & C_INITIALIZED)) {
4803 MDB_page *mp = mc->mc_pg[mc->mc_top];
4805 mc->mc_ki[mc->mc_top] = 0;
4811 key->mv_size = mc->mc_db->md_pad;
4812 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4814 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4815 MDB_GET_KEY(leaf, key);
4817 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4818 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4820 rc = mdb_node_read(mc->mc_txn, leaf, data);
4827 case MDB_GET_BOTH_RANGE:
4828 if (data == NULL || mc->mc_xcursor == NULL) {
4836 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4838 } else if (op == MDB_SET_RANGE)
4839 rc = mdb_cursor_set(mc, key, data, op, NULL);
4841 rc = mdb_cursor_set(mc, key, data, op, &exact);
4843 case MDB_GET_MULTIPLE:
4845 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4846 !(mc->mc_flags & C_INITIALIZED)) {
4851 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4852 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4855 case MDB_NEXT_MULTIPLE:
4857 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4861 if (!(mc->mc_flags & C_INITIALIZED))
4862 rc = mdb_cursor_first(mc, key, data);
4864 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4865 if (rc == MDB_SUCCESS) {
4866 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4869 mx = &mc->mc_xcursor->mx_cursor;
4870 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4872 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4873 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4881 case MDB_NEXT_NODUP:
4882 if (!(mc->mc_flags & C_INITIALIZED))
4883 rc = mdb_cursor_first(mc, key, data);
4885 rc = mdb_cursor_next(mc, key, data, op);
4889 case MDB_PREV_NODUP:
4890 if (!(mc->mc_flags & C_INITIALIZED)) {
4891 rc = mdb_cursor_last(mc, key, data);
4894 mc->mc_flags |= C_INITIALIZED;
4895 mc->mc_ki[mc->mc_top]++;
4897 rc = mdb_cursor_prev(mc, key, data, op);
4900 rc = mdb_cursor_first(mc, key, data);
4904 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4905 !(mc->mc_flags & C_INITIALIZED) ||
4906 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4910 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4913 rc = mdb_cursor_last(mc, key, data);
4917 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4918 !(mc->mc_flags & C_INITIALIZED) ||
4919 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4923 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4926 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4934 /** Touch all the pages in the cursor stack.
4935 * Makes sure all the pages are writable, before attempting a write operation.
4936 * @param[in] mc The cursor to operate on.
4939 mdb_cursor_touch(MDB_cursor *mc)
4943 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4946 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
4947 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4950 *mc->mc_dbflag |= DB_DIRTY;
4952 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4953 rc = mdb_page_touch(mc);
4957 mc->mc_top = mc->mc_snum-1;
4962 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4965 MDB_node *leaf = NULL;
4966 MDB_val xdata, *rdata, dkey;
4969 int do_sub = 0, insert = 0;
4970 unsigned int mcount = 0, dcount = 0;
4974 char dbuf[MDB_MAXKEYSIZE+1];
4975 unsigned int nflags;
4978 /* Check this first so counter will always be zero on any
4981 if (flags & MDB_MULTIPLE) {
4982 dcount = data[1].mv_size;
4983 data[1].mv_size = 0;
4984 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
4988 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4991 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4994 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4997 #if SIZE_MAX > MAXDATASIZE
4998 if (data->mv_size > MAXDATASIZE)
5002 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5003 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5007 if (flags == MDB_CURRENT) {
5008 if (!(mc->mc_flags & C_INITIALIZED))
5011 } else if (mc->mc_db->md_root == P_INVALID) {
5013 /* new database, write a root leaf page */
5014 DPUTS("allocating new root leaf page");
5015 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5019 mdb_cursor_push(mc, np);
5020 mc->mc_db->md_root = np->mp_pgno;
5021 mc->mc_db->md_depth++;
5022 *mc->mc_dbflag |= DB_DIRTY;
5023 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5025 np->mp_flags |= P_LEAF2;
5026 mc->mc_flags |= C_INITIALIZED;
5032 if (flags & MDB_APPEND) {
5034 rc = mdb_cursor_last(mc, &k2, &d2);
5036 rc = mc->mc_dbx->md_cmp(key, &k2);
5039 mc->mc_ki[mc->mc_top]++;
5041 /* new key is <= last key */
5046 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5048 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5049 DPRINTF("duplicate key [%s]", DKEY(key));
5051 return MDB_KEYEXIST;
5053 if (rc && rc != MDB_NOTFOUND)
5057 /* Cursor is positioned, now make sure all pages are writable */
5058 rc2 = mdb_cursor_touch(mc);
5063 /* The key already exists */
5064 if (rc == MDB_SUCCESS) {
5065 /* there's only a key anyway, so this is a no-op */
5066 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5067 unsigned int ksize = mc->mc_db->md_pad;
5068 if (key->mv_size != ksize)
5070 if (flags == MDB_CURRENT) {
5071 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5072 memcpy(ptr, key->mv_data, ksize);
5077 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5080 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5081 /* Was a single item before, must convert now */
5083 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5084 /* Just overwrite the current item */
5085 if (flags == MDB_CURRENT)
5088 dkey.mv_size = NODEDSZ(leaf);
5089 dkey.mv_data = NODEDATA(leaf);
5090 #if UINT_MAX < SIZE_MAX
5091 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5092 #ifdef MISALIGNED_OK
5093 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5095 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5098 /* if data matches, ignore it */
5099 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5100 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5102 /* create a fake page for the dup items */
5103 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5104 dkey.mv_data = dbuf;
5105 fp = (MDB_page *)&pbuf;
5106 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5107 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5108 fp->mp_lower = PAGEHDRSZ;
5109 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5110 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5111 fp->mp_flags |= P_LEAF2;
5112 fp->mp_pad = data->mv_size;
5113 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5115 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5116 (dkey.mv_size & 1) + (data->mv_size & 1);
5118 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5121 xdata.mv_size = fp->mp_upper;
5126 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5127 /* See if we need to convert from fake page to subDB */
5129 unsigned int offset;
5133 fp = NODEDATA(leaf);
5134 if (flags == MDB_CURRENT) {
5136 fp->mp_flags |= P_DIRTY;
5137 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5138 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5142 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5143 offset = fp->mp_pad;
5144 if (SIZELEFT(fp) >= offset)
5146 offset *= 4; /* space for 4 more */
5148 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5150 offset += offset & 1;
5151 fp_flags = fp->mp_flags;
5152 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5153 offset >= mc->mc_txn->mt_env->me_nodemax) {
5154 /* yes, convert it */
5156 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5157 dummy.md_pad = fp->mp_pad;
5158 dummy.md_flags = MDB_DUPFIXED;
5159 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5160 dummy.md_flags |= MDB_INTEGERKEY;
5163 dummy.md_branch_pages = 0;
5164 dummy.md_leaf_pages = 1;
5165 dummy.md_overflow_pages = 0;
5166 dummy.md_entries = NUMKEYS(fp);
5168 xdata.mv_size = sizeof(MDB_db);
5169 xdata.mv_data = &dummy;
5170 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5172 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5173 flags |= F_DUPDATA|F_SUBDATA;
5174 dummy.md_root = mp->mp_pgno;
5175 fp_flags &= ~P_SUBP;
5177 /* no, just grow it */
5179 xdata.mv_size = NODEDSZ(leaf) + offset;
5180 xdata.mv_data = &pbuf;
5181 mp = (MDB_page *)&pbuf;
5182 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5185 mp->mp_flags = fp_flags | P_DIRTY;
5186 mp->mp_pad = fp->mp_pad;
5187 mp->mp_lower = fp->mp_lower;
5188 mp->mp_upper = fp->mp_upper + offset;
5190 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5192 nsize = NODEDSZ(leaf) - fp->mp_upper;
5193 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5194 for (i=0; i<NUMKEYS(fp); i++)
5195 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5197 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5201 /* data is on sub-DB, just store it */
5202 flags |= F_DUPDATA|F_SUBDATA;
5206 /* overflow page overwrites need special handling */
5207 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5210 unsigned psize = mc->mc_txn->mt_env->me_psize;
5211 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5213 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5214 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5216 ovpages = omp->mp_pages;
5218 /* Is the ov page writable and large enough? */
5219 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5220 /* yes, overwrite it. Note in this case we don't
5221 * bother to try shrinking the page if the new data
5222 * is smaller than the overflow threshold.
5225 /* It is writable only in a parent txn */
5226 size_t sz = (size_t) psize * ovpages, off;
5227 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5233 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5234 if (!(flags & MDB_RESERVE)) {
5235 /* Copy end of page, adjusting alignment so
5236 * compiler may copy words instead of bytes.
5238 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5239 memcpy((size_t *)((char *)np + off),
5240 (size_t *)((char *)omp + off), sz - off);
5243 memcpy(np, omp, sz); /* Copy beginning of page */
5246 SETDSZ(leaf, data->mv_size);
5247 if (F_ISSET(flags, MDB_RESERVE))
5248 data->mv_data = METADATA(omp);
5250 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5253 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5256 } else if (NODEDSZ(leaf) == data->mv_size) {
5257 /* same size, just replace it. Note that we could
5258 * also reuse this node if the new data is smaller,
5259 * but instead we opt to shrink the node in that case.
5261 if (F_ISSET(flags, MDB_RESERVE))
5262 data->mv_data = NODEDATA(leaf);
5263 else if (data->mv_size)
5264 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5266 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5269 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5270 mc->mc_db->md_entries--;
5272 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5279 nflags = flags & NODE_ADD_FLAGS;
5280 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5281 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5282 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5283 nflags &= ~MDB_APPEND;
5285 nflags |= MDB_SPLIT_REPLACE;
5286 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5288 /* There is room already in this leaf page. */
5289 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5290 if (rc == 0 && !do_sub && insert) {
5291 /* Adjust other cursors pointing to mp */
5292 MDB_cursor *m2, *m3;
5293 MDB_dbi dbi = mc->mc_dbi;
5294 unsigned i = mc->mc_top;
5295 MDB_page *mp = mc->mc_pg[i];
5297 if (mc->mc_flags & C_SUB)
5300 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5301 if (mc->mc_flags & C_SUB)
5302 m3 = &m2->mc_xcursor->mx_cursor;
5305 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5306 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5313 if (rc != MDB_SUCCESS)
5314 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5316 /* Now store the actual data in the child DB. Note that we're
5317 * storing the user data in the keys field, so there are strict
5318 * size limits on dupdata. The actual data fields of the child
5319 * DB are all zero size.
5326 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5327 if (flags & MDB_CURRENT) {
5328 xflags = MDB_CURRENT;
5330 mdb_xcursor_init1(mc, leaf);
5331 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5333 /* converted, write the original data first */
5335 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5339 /* Adjust other cursors pointing to mp */
5341 unsigned i = mc->mc_top;
5342 MDB_page *mp = mc->mc_pg[i];
5344 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5345 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5346 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5347 mdb_xcursor_init1(m2, leaf);
5351 /* we've done our job */
5354 if (flags & MDB_APPENDDUP)
5355 xflags |= MDB_APPEND;
5356 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5357 if (flags & F_SUBDATA) {
5358 void *db = NODEDATA(leaf);
5359 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5362 /* sub-writes might have failed so check rc again.
5363 * Don't increment count if we just replaced an existing item.
5365 if (!rc && !(flags & MDB_CURRENT))
5366 mc->mc_db->md_entries++;
5367 if (flags & MDB_MULTIPLE) {
5370 if (mcount < dcount) {
5371 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5372 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5376 /* let caller know how many succeeded, if any */
5377 data[1].mv_size = mcount;
5381 /* If we succeeded and the key didn't exist before, make sure
5382 * the cursor is marked valid.
5385 mc->mc_flags |= C_INITIALIZED;
5390 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5395 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5398 if (!(mc->mc_flags & C_INITIALIZED))
5401 rc = mdb_cursor_touch(mc);
5405 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5407 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5408 if (flags != MDB_NODUPDATA) {
5409 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5410 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5412 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5413 /* If sub-DB still has entries, we're done */
5414 if (mc->mc_xcursor->mx_db.md_entries) {
5415 if (leaf->mn_flags & F_SUBDATA) {
5416 /* update subDB info */
5417 void *db = NODEDATA(leaf);
5418 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5421 /* shrink fake page */
5422 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5423 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5424 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5425 /* fix other sub-DB cursors pointed at this fake page */
5426 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5427 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5428 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5429 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5430 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5433 mc->mc_db->md_entries--;
5436 /* otherwise fall thru and delete the sub-DB */
5439 if (leaf->mn_flags & F_SUBDATA) {
5440 /* add all the child DB's pages to the free list */
5441 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5442 if (rc == MDB_SUCCESS) {
5443 mc->mc_db->md_entries -=
5444 mc->mc_xcursor->mx_db.md_entries;
5449 return mdb_cursor_del0(mc, leaf);
5452 /** Allocate and initialize new pages for a database.
5453 * @param[in] mc a cursor on the database being added to.
5454 * @param[in] flags flags defining what type of page is being allocated.
5455 * @param[in] num the number of pages to allocate. This is usually 1,
5456 * unless allocating overflow pages for a large record.
5457 * @param[out] mp Address of a page, or NULL on failure.
5458 * @return 0 on success, non-zero on failure.
5461 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5466 if ((rc = mdb_page_alloc(mc, num, &np)))
5468 DPRINTF("allocated new mpage %zu, page size %u",
5469 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5470 np->mp_flags = flags | P_DIRTY;
5471 np->mp_lower = PAGEHDRSZ;
5472 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5475 mc->mc_db->md_branch_pages++;
5476 else if (IS_LEAF(np))
5477 mc->mc_db->md_leaf_pages++;
5478 else if (IS_OVERFLOW(np)) {
5479 mc->mc_db->md_overflow_pages += num;
5487 /** Calculate the size of a leaf node.
5488 * The size depends on the environment's page size; if a data item
5489 * is too large it will be put onto an overflow page and the node
5490 * size will only include the key and not the data. Sizes are always
5491 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5492 * of the #MDB_node headers.
5493 * @param[in] env The environment handle.
5494 * @param[in] key The key for the node.
5495 * @param[in] data The data for the node.
5496 * @return The number of bytes needed to store the node.
5499 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5503 sz = LEAFSIZE(key, data);
5504 if (sz >= env->me_nodemax) {
5505 /* put on overflow page */
5506 sz -= data->mv_size - sizeof(pgno_t);
5510 return sz + sizeof(indx_t);
5513 /** Calculate the size of a branch node.
5514 * The size should depend on the environment's page size but since
5515 * we currently don't support spilling large keys onto overflow
5516 * pages, it's simply the size of the #MDB_node header plus the
5517 * size of the key. Sizes are always rounded up to an even number
5518 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5519 * @param[in] env The environment handle.
5520 * @param[in] key The key for the node.
5521 * @return The number of bytes needed to store the node.
5524 mdb_branch_size(MDB_env *env, MDB_val *key)
5529 if (sz >= env->me_nodemax) {
5530 /* put on overflow page */
5531 /* not implemented */
5532 /* sz -= key->size - sizeof(pgno_t); */
5535 return sz + sizeof(indx_t);
5538 /** Add a node to the page pointed to by the cursor.
5539 * @param[in] mc The cursor for this operation.
5540 * @param[in] indx The index on the page where the new node should be added.
5541 * @param[in] key The key for the new node.
5542 * @param[in] data The data for the new node, if any.
5543 * @param[in] pgno The page number, if adding a branch node.
5544 * @param[in] flags Flags for the node.
5545 * @return 0 on success, non-zero on failure. Possible errors are:
5547 * <li>ENOMEM - failed to allocate overflow pages for the node.
5548 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5549 * should never happen since all callers already calculate the
5550 * page's free space before calling this function.
5554 mdb_node_add(MDB_cursor *mc, indx_t indx,
5555 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5558 size_t node_size = NODESIZE;
5561 MDB_page *mp = mc->mc_pg[mc->mc_top];
5562 MDB_page *ofp = NULL; /* overflow page */
5565 assert(mp->mp_upper >= mp->mp_lower);
5567 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5568 IS_LEAF(mp) ? "leaf" : "branch",
5569 IS_SUBP(mp) ? "sub-" : "",
5570 mp->mp_pgno, indx, data ? data->mv_size : 0,
5571 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5574 /* Move higher keys up one slot. */
5575 int ksize = mc->mc_db->md_pad, dif;
5576 char *ptr = LEAF2KEY(mp, indx, ksize);
5577 dif = NUMKEYS(mp) - indx;
5579 memmove(ptr+ksize, ptr, dif*ksize);
5580 /* insert new key */
5581 memcpy(ptr, key->mv_data, ksize);
5583 /* Just using these for counting */
5584 mp->mp_lower += sizeof(indx_t);
5585 mp->mp_upper -= ksize - sizeof(indx_t);
5590 node_size += key->mv_size;
5594 if (F_ISSET(flags, F_BIGDATA)) {
5595 /* Data already on overflow page. */
5596 node_size += sizeof(pgno_t);
5597 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5598 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5600 /* Put data on overflow page. */
5601 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5602 data->mv_size, node_size+data->mv_size);
5603 node_size += sizeof(pgno_t);
5604 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5606 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5609 node_size += data->mv_size;
5612 node_size += node_size & 1;
5614 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5615 DPRINTF("not enough room in page %zu, got %u ptrs",
5616 mp->mp_pgno, NUMKEYS(mp));
5617 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5618 mp->mp_upper - mp->mp_lower);
5619 DPRINTF("node size = %zu", node_size);
5620 return MDB_PAGE_FULL;
5623 /* Move higher pointers up one slot. */
5624 for (i = NUMKEYS(mp); i > indx; i--)
5625 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5627 /* Adjust free space offsets. */
5628 ofs = mp->mp_upper - node_size;
5629 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5630 mp->mp_ptrs[indx] = ofs;
5632 mp->mp_lower += sizeof(indx_t);
5634 /* Write the node data. */
5635 node = NODEPTR(mp, indx);
5636 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5637 node->mn_flags = flags;
5639 SETDSZ(node,data->mv_size);
5644 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5649 if (F_ISSET(flags, F_BIGDATA))
5650 memcpy(node->mn_data + key->mv_size, data->mv_data,
5652 else if (F_ISSET(flags, MDB_RESERVE))
5653 data->mv_data = node->mn_data + key->mv_size;
5655 memcpy(node->mn_data + key->mv_size, data->mv_data,
5658 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5660 if (F_ISSET(flags, MDB_RESERVE))
5661 data->mv_data = METADATA(ofp);
5663 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5670 /** Delete the specified node from a page.
5671 * @param[in] mp The page to operate on.
5672 * @param[in] indx The index of the node to delete.
5673 * @param[in] ksize The size of a node. Only used if the page is
5674 * part of a #MDB_DUPFIXED database.
5677 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5680 indx_t i, j, numkeys, ptr;
5687 COPY_PGNO(pgno, mp->mp_pgno);
5688 DPRINTF("delete node %u on %s page %zu", indx,
5689 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5692 assert(indx < NUMKEYS(mp));
5695 int x = NUMKEYS(mp) - 1 - indx;
5696 base = LEAF2KEY(mp, indx, ksize);
5698 memmove(base, base + ksize, x * ksize);
5699 mp->mp_lower -= sizeof(indx_t);
5700 mp->mp_upper += ksize - sizeof(indx_t);
5704 node = NODEPTR(mp, indx);
5705 sz = NODESIZE + node->mn_ksize;
5707 if (F_ISSET(node->mn_flags, F_BIGDATA))
5708 sz += sizeof(pgno_t);
5710 sz += NODEDSZ(node);
5714 ptr = mp->mp_ptrs[indx];
5715 numkeys = NUMKEYS(mp);
5716 for (i = j = 0; i < numkeys; i++) {
5718 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5719 if (mp->mp_ptrs[i] < ptr)
5720 mp->mp_ptrs[j] += sz;
5725 base = (char *)mp + mp->mp_upper;
5726 memmove(base + sz, base, ptr - mp->mp_upper);
5728 mp->mp_lower -= sizeof(indx_t);
5732 /** Compact the main page after deleting a node on a subpage.
5733 * @param[in] mp The main page to operate on.
5734 * @param[in] indx The index of the subpage on the main page.
5737 mdb_node_shrink(MDB_page *mp, indx_t indx)
5744 indx_t i, numkeys, ptr;
5746 node = NODEPTR(mp, indx);
5747 sp = (MDB_page *)NODEDATA(node);
5748 osize = NODEDSZ(node);
5750 delta = sp->mp_upper - sp->mp_lower;
5751 SETDSZ(node, osize - delta);
5752 xp = (MDB_page *)((char *)sp + delta);
5754 /* shift subpage upward */
5756 nsize = NUMKEYS(sp) * sp->mp_pad;
5757 memmove(METADATA(xp), METADATA(sp), nsize);
5760 nsize = osize - sp->mp_upper;
5761 numkeys = NUMKEYS(sp);
5762 for (i=numkeys-1; i>=0; i--)
5763 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5765 xp->mp_upper = sp->mp_lower;
5766 xp->mp_lower = sp->mp_lower;
5767 xp->mp_flags = sp->mp_flags;
5768 xp->mp_pad = sp->mp_pad;
5769 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5771 /* shift lower nodes upward */
5772 ptr = mp->mp_ptrs[indx];
5773 numkeys = NUMKEYS(mp);
5774 for (i = 0; i < numkeys; i++) {
5775 if (mp->mp_ptrs[i] <= ptr)
5776 mp->mp_ptrs[i] += delta;
5779 base = (char *)mp + mp->mp_upper;
5780 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5781 mp->mp_upper += delta;
5784 /** Initial setup of a sorted-dups cursor.
5785 * Sorted duplicates are implemented as a sub-database for the given key.
5786 * The duplicate data items are actually keys of the sub-database.
5787 * Operations on the duplicate data items are performed using a sub-cursor
5788 * initialized when the sub-database is first accessed. This function does
5789 * the preliminary setup of the sub-cursor, filling in the fields that
5790 * depend only on the parent DB.
5791 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5794 mdb_xcursor_init0(MDB_cursor *mc)
5796 MDB_xcursor *mx = mc->mc_xcursor;
5798 mx->mx_cursor.mc_xcursor = NULL;
5799 mx->mx_cursor.mc_txn = mc->mc_txn;
5800 mx->mx_cursor.mc_db = &mx->mx_db;
5801 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5802 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5803 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5804 mx->mx_cursor.mc_snum = 0;
5805 mx->mx_cursor.mc_top = 0;
5806 mx->mx_cursor.mc_flags = C_SUB;
5807 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5808 mx->mx_dbx.md_dcmp = NULL;
5809 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5812 /** Final setup of a sorted-dups cursor.
5813 * Sets up the fields that depend on the data from the main cursor.
5814 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5815 * @param[in] node The data containing the #MDB_db record for the
5816 * sorted-dup database.
5819 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5821 MDB_xcursor *mx = mc->mc_xcursor;
5823 if (node->mn_flags & F_SUBDATA) {
5824 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5825 mx->mx_cursor.mc_pg[0] = 0;
5826 mx->mx_cursor.mc_snum = 0;
5827 mx->mx_cursor.mc_flags = C_SUB;
5829 MDB_page *fp = NODEDATA(node);
5830 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5831 mx->mx_db.md_flags = 0;
5832 mx->mx_db.md_depth = 1;
5833 mx->mx_db.md_branch_pages = 0;
5834 mx->mx_db.md_leaf_pages = 1;
5835 mx->mx_db.md_overflow_pages = 0;
5836 mx->mx_db.md_entries = NUMKEYS(fp);
5837 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5838 mx->mx_cursor.mc_snum = 1;
5839 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5840 mx->mx_cursor.mc_top = 0;
5841 mx->mx_cursor.mc_pg[0] = fp;
5842 mx->mx_cursor.mc_ki[0] = 0;
5843 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5844 mx->mx_db.md_flags = MDB_DUPFIXED;
5845 mx->mx_db.md_pad = fp->mp_pad;
5846 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5847 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5850 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5852 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5854 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5855 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5856 #if UINT_MAX < SIZE_MAX
5857 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5858 #ifdef MISALIGNED_OK
5859 mx->mx_dbx.md_cmp = mdb_cmp_long;
5861 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5866 /** Initialize a cursor for a given transaction and database. */
5868 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5873 mc->mc_db = &txn->mt_dbs[dbi];
5874 mc->mc_dbx = &txn->mt_dbxs[dbi];
5875 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5880 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5882 mc->mc_xcursor = mx;
5883 mdb_xcursor_init0(mc);
5885 mc->mc_xcursor = NULL;
5887 if (*mc->mc_dbflag & DB_STALE) {
5888 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5893 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5896 size_t size = sizeof(MDB_cursor);
5898 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5901 /* Allow read access to the freelist */
5902 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5905 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5906 size += sizeof(MDB_xcursor);
5908 if ((mc = malloc(size)) != NULL) {
5909 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5910 if (txn->mt_cursors) {
5911 mc->mc_next = txn->mt_cursors[dbi];
5912 txn->mt_cursors[dbi] = mc;
5913 mc->mc_flags |= C_UNTRACK;
5925 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5927 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5930 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5933 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5937 /* Return the count of duplicate data items for the current key */
5939 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5943 if (mc == NULL || countp == NULL)
5946 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5949 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5950 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5953 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5956 *countp = mc->mc_xcursor->mx_db.md_entries;
5962 mdb_cursor_close(MDB_cursor *mc)
5965 /* remove from txn, if tracked */
5966 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
5967 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5968 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5970 *prev = mc->mc_next;
5977 mdb_cursor_txn(MDB_cursor *mc)
5979 if (!mc) return NULL;
5984 mdb_cursor_dbi(MDB_cursor *mc)
5990 /** Replace the key for a node with a new key.
5991 * @param[in] mc Cursor pointing to the node to operate on.
5992 * @param[in] key The new key to use.
5993 * @return 0 on success, non-zero on failure.
5996 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6003 indx_t ptr, i, numkeys, indx;
6006 indx = mc->mc_ki[mc->mc_top];
6007 mp = mc->mc_pg[mc->mc_top];
6008 node = NODEPTR(mp, indx);
6009 ptr = mp->mp_ptrs[indx];
6013 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6014 k2.mv_data = NODEKEY(node);
6015 k2.mv_size = node->mn_ksize;
6016 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6018 mdb_dkey(&k2, kbuf2),
6024 delta0 = delta = key->mv_size - node->mn_ksize;
6026 /* Must be 2-byte aligned. If new key is
6027 * shorter by 1, the shift will be skipped.
6029 delta += (delta & 1);
6031 if (delta > 0 && SIZELEFT(mp) < delta) {
6033 /* not enough space left, do a delete and split */
6034 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6035 pgno = NODEPGNO(node);
6036 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6037 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6040 numkeys = NUMKEYS(mp);
6041 for (i = 0; i < numkeys; i++) {
6042 if (mp->mp_ptrs[i] <= ptr)
6043 mp->mp_ptrs[i] -= delta;
6046 base = (char *)mp + mp->mp_upper;
6047 len = ptr - mp->mp_upper + NODESIZE;
6048 memmove(base - delta, base, len);
6049 mp->mp_upper -= delta;
6051 node = NODEPTR(mp, indx);
6054 /* But even if no shift was needed, update ksize */
6056 node->mn_ksize = key->mv_size;
6059 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6065 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6067 /** Move a node from csrc to cdst.
6070 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6077 unsigned short flags;
6081 /* Mark src and dst as dirty. */
6082 if ((rc = mdb_page_touch(csrc)) ||
6083 (rc = mdb_page_touch(cdst)))
6086 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6087 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6088 key.mv_size = csrc->mc_db->md_pad;
6089 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6091 data.mv_data = NULL;
6095 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6096 assert(!((long)srcnode&1));
6097 srcpg = NODEPGNO(srcnode);
6098 flags = srcnode->mn_flags;
6099 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6100 unsigned int snum = csrc->mc_snum;
6102 /* must find the lowest key below src */
6103 mdb_page_search_lowest(csrc);
6104 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6105 key.mv_size = csrc->mc_db->md_pad;
6106 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6108 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6109 key.mv_size = NODEKSZ(s2);
6110 key.mv_data = NODEKEY(s2);
6112 csrc->mc_snum = snum--;
6113 csrc->mc_top = snum;
6115 key.mv_size = NODEKSZ(srcnode);
6116 key.mv_data = NODEKEY(srcnode);
6118 data.mv_size = NODEDSZ(srcnode);
6119 data.mv_data = NODEDATA(srcnode);
6121 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6122 unsigned int snum = cdst->mc_snum;
6125 /* must find the lowest key below dst */
6126 mdb_page_search_lowest(cdst);
6127 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6128 bkey.mv_size = cdst->mc_db->md_pad;
6129 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6131 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6132 bkey.mv_size = NODEKSZ(s2);
6133 bkey.mv_data = NODEKEY(s2);
6135 cdst->mc_snum = snum--;
6136 cdst->mc_top = snum;
6137 mdb_cursor_copy(cdst, &mn);
6139 rc = mdb_update_key(&mn, &bkey);
6144 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6145 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6146 csrc->mc_ki[csrc->mc_top],
6148 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6149 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6151 /* Add the node to the destination page.
6153 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6154 if (rc != MDB_SUCCESS)
6157 /* Delete the node from the source page.
6159 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6162 /* Adjust other cursors pointing to mp */
6163 MDB_cursor *m2, *m3;
6164 MDB_dbi dbi = csrc->mc_dbi;
6165 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6167 if (csrc->mc_flags & C_SUB)
6170 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6171 if (csrc->mc_flags & C_SUB)
6172 m3 = &m2->mc_xcursor->mx_cursor;
6175 if (m3 == csrc) continue;
6176 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6177 csrc->mc_ki[csrc->mc_top]) {
6178 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6179 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6184 /* Update the parent separators.
6186 if (csrc->mc_ki[csrc->mc_top] == 0) {
6187 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6188 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6189 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6191 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6192 key.mv_size = NODEKSZ(srcnode);
6193 key.mv_data = NODEKEY(srcnode);
6195 DPRINTF("update separator for source page %zu to [%s]",
6196 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6197 mdb_cursor_copy(csrc, &mn);
6200 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6203 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6205 indx_t ix = csrc->mc_ki[csrc->mc_top];
6206 nullkey.mv_size = 0;
6207 csrc->mc_ki[csrc->mc_top] = 0;
6208 rc = mdb_update_key(csrc, &nullkey);
6209 csrc->mc_ki[csrc->mc_top] = ix;
6210 assert(rc == MDB_SUCCESS);
6214 if (cdst->mc_ki[cdst->mc_top] == 0) {
6215 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6216 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6217 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6219 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6220 key.mv_size = NODEKSZ(srcnode);
6221 key.mv_data = NODEKEY(srcnode);
6223 DPRINTF("update separator for destination page %zu to [%s]",
6224 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6225 mdb_cursor_copy(cdst, &mn);
6228 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6231 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6233 indx_t ix = cdst->mc_ki[cdst->mc_top];
6234 nullkey.mv_size = 0;
6235 cdst->mc_ki[cdst->mc_top] = 0;
6236 rc = mdb_update_key(cdst, &nullkey);
6237 cdst->mc_ki[cdst->mc_top] = ix;
6238 assert(rc == MDB_SUCCESS);
6245 /** Merge one page into another.
6246 * The nodes from the page pointed to by \b csrc will
6247 * be copied to the page pointed to by \b cdst and then
6248 * the \b csrc page will be freed.
6249 * @param[in] csrc Cursor pointing to the source page.
6250 * @param[in] cdst Cursor pointing to the destination page.
6253 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6261 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6262 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6264 assert(csrc->mc_snum > 1); /* can't merge root page */
6265 assert(cdst->mc_snum > 1);
6267 /* Mark dst as dirty. */
6268 if ((rc = mdb_page_touch(cdst)))
6271 /* Move all nodes from src to dst.
6273 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6274 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6275 key.mv_size = csrc->mc_db->md_pad;
6276 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6277 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6278 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6279 if (rc != MDB_SUCCESS)
6281 key.mv_data = (char *)key.mv_data + key.mv_size;
6284 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6285 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6286 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6287 unsigned int snum = csrc->mc_snum;
6289 /* must find the lowest key below src */
6290 mdb_page_search_lowest(csrc);
6291 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6292 key.mv_size = csrc->mc_db->md_pad;
6293 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6295 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6296 key.mv_size = NODEKSZ(s2);
6297 key.mv_data = NODEKEY(s2);
6299 csrc->mc_snum = snum--;
6300 csrc->mc_top = snum;
6302 key.mv_size = srcnode->mn_ksize;
6303 key.mv_data = NODEKEY(srcnode);
6306 data.mv_size = NODEDSZ(srcnode);
6307 data.mv_data = NODEDATA(srcnode);
6308 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6309 if (rc != MDB_SUCCESS)
6314 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6315 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);
6317 /* Unlink the src page from parent and add to free list.
6319 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6320 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6323 rc = mdb_update_key(csrc, &key);
6329 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6330 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6333 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6334 csrc->mc_db->md_leaf_pages--;
6336 csrc->mc_db->md_branch_pages--;
6338 /* Adjust other cursors pointing to mp */
6339 MDB_cursor *m2, *m3;
6340 MDB_dbi dbi = csrc->mc_dbi;
6341 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6343 if (csrc->mc_flags & C_SUB)
6346 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6347 if (csrc->mc_flags & C_SUB)
6348 m3 = &m2->mc_xcursor->mx_cursor;
6351 if (m3 == csrc) continue;
6352 if (m3->mc_snum < csrc->mc_snum) continue;
6353 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6354 m3->mc_pg[csrc->mc_top] = mp;
6355 m3->mc_ki[csrc->mc_top] += nkeys;
6359 mdb_cursor_pop(csrc);
6361 return mdb_rebalance(csrc);
6364 /** Copy the contents of a cursor.
6365 * @param[in] csrc The cursor to copy from.
6366 * @param[out] cdst The cursor to copy to.
6369 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6373 cdst->mc_txn = csrc->mc_txn;
6374 cdst->mc_dbi = csrc->mc_dbi;
6375 cdst->mc_db = csrc->mc_db;
6376 cdst->mc_dbx = csrc->mc_dbx;
6377 cdst->mc_snum = csrc->mc_snum;
6378 cdst->mc_top = csrc->mc_top;
6379 cdst->mc_flags = csrc->mc_flags;
6381 for (i=0; i<csrc->mc_snum; i++) {
6382 cdst->mc_pg[i] = csrc->mc_pg[i];
6383 cdst->mc_ki[i] = csrc->mc_ki[i];
6387 /** Rebalance the tree after a delete operation.
6388 * @param[in] mc Cursor pointing to the page where rebalancing
6390 * @return 0 on success, non-zero on failure.
6393 mdb_rebalance(MDB_cursor *mc)
6397 unsigned int ptop, minkeys;
6400 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6404 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6405 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6406 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6407 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6411 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6412 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6415 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6416 DPRINTF("no need to rebalance page %zu, above fill threshold",
6422 if (mc->mc_snum < 2) {
6423 MDB_page *mp = mc->mc_pg[0];
6425 DPUTS("Can't rebalance a subpage, ignoring");
6428 if (NUMKEYS(mp) == 0) {
6429 DPUTS("tree is completely empty");
6430 mc->mc_db->md_root = P_INVALID;
6431 mc->mc_db->md_depth = 0;
6432 mc->mc_db->md_leaf_pages = 0;
6433 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6436 /* Adjust cursors pointing to mp */
6440 MDB_cursor *m2, *m3;
6441 MDB_dbi dbi = mc->mc_dbi;
6443 if (mc->mc_flags & C_SUB)
6446 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6447 if (mc->mc_flags & C_SUB)
6448 m3 = &m2->mc_xcursor->mx_cursor;
6451 if (m3->mc_snum < mc->mc_snum) continue;
6452 if (m3->mc_pg[0] == mp) {
6458 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6459 DPUTS("collapsing root page!");
6460 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6463 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6464 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6467 mc->mc_db->md_depth--;
6468 mc->mc_db->md_branch_pages--;
6470 /* Adjust other cursors pointing to mp */
6471 MDB_cursor *m2, *m3;
6472 MDB_dbi dbi = mc->mc_dbi;
6474 if (mc->mc_flags & C_SUB)
6477 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6478 if (mc->mc_flags & C_SUB)
6479 m3 = &m2->mc_xcursor->mx_cursor;
6482 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6483 if (m3->mc_pg[0] == mp) {
6484 m3->mc_pg[0] = mc->mc_pg[0];
6491 DPUTS("root page doesn't need rebalancing");
6495 /* The parent (branch page) must have at least 2 pointers,
6496 * otherwise the tree is invalid.
6498 ptop = mc->mc_top-1;
6499 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6501 /* Leaf page fill factor is below the threshold.
6502 * Try to move keys from left or right neighbor, or
6503 * merge with a neighbor page.
6508 mdb_cursor_copy(mc, &mn);
6509 mn.mc_xcursor = NULL;
6511 if (mc->mc_ki[ptop] == 0) {
6512 /* We're the leftmost leaf in our parent.
6514 DPUTS("reading right neighbor");
6516 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6517 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6520 mn.mc_ki[mn.mc_top] = 0;
6521 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6523 /* There is at least one neighbor to the left.
6525 DPUTS("reading left neighbor");
6527 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6528 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6531 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6532 mc->mc_ki[mc->mc_top] = 0;
6535 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6536 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);
6538 /* If the neighbor page is above threshold and has enough keys,
6539 * move one key from it. Otherwise we should try to merge them.
6540 * (A branch page must never have less than 2 keys.)
6542 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6543 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6544 return mdb_node_move(&mn, mc);
6546 if (mc->mc_ki[ptop] == 0)
6547 rc = mdb_page_merge(&mn, mc);
6549 rc = mdb_page_merge(mc, &mn);
6550 mc->mc_flags &= ~C_INITIALIZED;
6555 /** Complete a delete operation started by #mdb_cursor_del(). */
6557 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6561 /* add overflow pages to free list */
6562 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6566 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6567 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6568 (rc = mdb_ovpage_free(mc, omp)))
6571 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6572 mc->mc_db->md_entries--;
6573 rc = mdb_rebalance(mc);
6574 if (rc != MDB_SUCCESS)
6575 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6576 /* if mc points past last node in page, invalidate */
6577 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6578 mc->mc_flags &= ~C_INITIALIZED;
6584 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6585 MDB_val *key, MDB_val *data)
6590 MDB_val rdata, *xdata;
6594 assert(key != NULL);
6596 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6598 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6601 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6605 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6609 mdb_cursor_init(&mc, txn, dbi, &mx);
6620 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6622 /* let mdb_page_split know about this cursor if needed:
6623 * delete will trigger a rebalance; if it needs to move
6624 * a node from one page to another, it will have to
6625 * update the parent's separator key(s). If the new sepkey
6626 * is larger than the current one, the parent page may
6627 * run out of space, triggering a split. We need this
6628 * cursor to be consistent until the end of the rebalance.
6630 mc.mc_next = txn->mt_cursors[dbi];
6631 txn->mt_cursors[dbi] = &mc;
6632 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6633 txn->mt_cursors[dbi] = mc.mc_next;
6638 /** Split a page and insert a new node.
6639 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6640 * The cursor will be updated to point to the actual page and index where
6641 * the node got inserted after the split.
6642 * @param[in] newkey The key for the newly inserted node.
6643 * @param[in] newdata The data for the newly inserted node.
6644 * @param[in] newpgno The page number, if the new node is a branch node.
6645 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6646 * @return 0 on success, non-zero on failure.
6649 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6650 unsigned int nflags)
6653 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6656 unsigned int i, j, split_indx, nkeys, pmax;
6658 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6660 MDB_page *mp, *rp, *pp;
6665 mp = mc->mc_pg[mc->mc_top];
6666 newindx = mc->mc_ki[mc->mc_top];
6668 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6669 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6670 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6672 /* Create a right sibling. */
6673 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6675 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6677 if (mc->mc_snum < 2) {
6678 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6680 /* shift current top to make room for new parent */
6681 mc->mc_pg[1] = mc->mc_pg[0];
6682 mc->mc_ki[1] = mc->mc_ki[0];
6685 mc->mc_db->md_root = pp->mp_pgno;
6686 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6687 mc->mc_db->md_depth++;
6690 /* Add left (implicit) pointer. */
6691 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6692 /* undo the pre-push */
6693 mc->mc_pg[0] = mc->mc_pg[1];
6694 mc->mc_ki[0] = mc->mc_ki[1];
6695 mc->mc_db->md_root = mp->mp_pgno;
6696 mc->mc_db->md_depth--;
6703 ptop = mc->mc_top-1;
6704 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6707 mc->mc_flags |= C_SPLITTING;
6708 mdb_cursor_copy(mc, &mn);
6709 mn.mc_pg[mn.mc_top] = rp;
6710 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6712 if (nflags & MDB_APPEND) {
6713 mn.mc_ki[mn.mc_top] = 0;
6715 split_indx = newindx;
6720 nkeys = NUMKEYS(mp);
6721 split_indx = nkeys / 2;
6722 if (newindx < split_indx)
6728 unsigned int lsize, rsize, ksize;
6729 /* Move half of the keys to the right sibling */
6731 x = mc->mc_ki[mc->mc_top] - split_indx;
6732 ksize = mc->mc_db->md_pad;
6733 split = LEAF2KEY(mp, split_indx, ksize);
6734 rsize = (nkeys - split_indx) * ksize;
6735 lsize = (nkeys - split_indx) * sizeof(indx_t);
6736 mp->mp_lower -= lsize;
6737 rp->mp_lower += lsize;
6738 mp->mp_upper += rsize - lsize;
6739 rp->mp_upper -= rsize - lsize;
6740 sepkey.mv_size = ksize;
6741 if (newindx == split_indx) {
6742 sepkey.mv_data = newkey->mv_data;
6744 sepkey.mv_data = split;
6747 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6748 memcpy(rp->mp_ptrs, split, rsize);
6749 sepkey.mv_data = rp->mp_ptrs;
6750 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6751 memcpy(ins, newkey->mv_data, ksize);
6752 mp->mp_lower += sizeof(indx_t);
6753 mp->mp_upper -= ksize - sizeof(indx_t);
6756 memcpy(rp->mp_ptrs, split, x * ksize);
6757 ins = LEAF2KEY(rp, x, ksize);
6758 memcpy(ins, newkey->mv_data, ksize);
6759 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6760 rp->mp_lower += sizeof(indx_t);
6761 rp->mp_upper -= ksize - sizeof(indx_t);
6762 mc->mc_ki[mc->mc_top] = x;
6763 mc->mc_pg[mc->mc_top] = rp;
6768 /* For leaf pages, check the split point based on what
6769 * fits where, since otherwise mdb_node_add can fail.
6771 * This check is only needed when the data items are
6772 * relatively large, such that being off by one will
6773 * make the difference between success or failure.
6775 * It's also relevant if a page happens to be laid out
6776 * such that one half of its nodes are all "small" and
6777 * the other half of its nodes are "large." If the new
6778 * item is also "large" and falls on the half with
6779 * "large" nodes, it also may not fit.
6782 unsigned int psize, nsize;
6783 /* Maximum free space in an empty page */
6784 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6785 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6786 if ((nkeys < 20) || (nsize > pmax/16)) {
6787 if (newindx <= split_indx) {
6790 for (i=0; i<split_indx; i++) {
6791 node = NODEPTR(mp, i);
6792 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6793 if (F_ISSET(node->mn_flags, F_BIGDATA))
6794 psize += sizeof(pgno_t);
6796 psize += NODEDSZ(node);
6800 split_indx = newindx;
6811 for (i=nkeys-1; i>=split_indx; i--) {
6812 node = NODEPTR(mp, i);
6813 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6814 if (F_ISSET(node->mn_flags, F_BIGDATA))
6815 psize += sizeof(pgno_t);
6817 psize += NODEDSZ(node);
6821 split_indx = newindx;
6832 /* First find the separating key between the split pages.
6833 * The case where newindx == split_indx is ambiguous; the
6834 * new item could go to the new page or stay on the original
6835 * page. If newpos == 1 it goes to the new page.
6837 if (newindx == split_indx && newpos) {
6838 sepkey.mv_size = newkey->mv_size;
6839 sepkey.mv_data = newkey->mv_data;
6841 node = NODEPTR(mp, split_indx);
6842 sepkey.mv_size = node->mn_ksize;
6843 sepkey.mv_data = NODEKEY(node);
6847 DPRINTF("separator is [%s]", DKEY(&sepkey));
6849 /* Copy separator key to the parent.
6851 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6855 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6858 if (mn.mc_snum == mc->mc_snum) {
6859 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6860 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6861 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6862 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6867 /* Right page might now have changed parent.
6868 * Check if left page also changed parent.
6870 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6871 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6872 for (i=0; i<ptop; i++) {
6873 mc->mc_pg[i] = mn.mc_pg[i];
6874 mc->mc_ki[i] = mn.mc_ki[i];
6876 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6877 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6881 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6884 mc->mc_flags ^= C_SPLITTING;
6885 if (rc != MDB_SUCCESS) {
6888 if (nflags & MDB_APPEND) {
6889 mc->mc_pg[mc->mc_top] = rp;
6890 mc->mc_ki[mc->mc_top] = 0;
6891 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6894 for (i=0; i<mc->mc_top; i++)
6895 mc->mc_ki[i] = mn.mc_ki[i];
6902 /* Move half of the keys to the right sibling. */
6904 /* grab a page to hold a temporary copy */
6905 copy = mdb_page_malloc(mc->mc_txn, 1);
6909 copy->mp_pgno = mp->mp_pgno;
6910 copy->mp_flags = mp->mp_flags;
6911 copy->mp_lower = PAGEHDRSZ;
6912 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6913 mc->mc_pg[mc->mc_top] = copy;
6914 for (i = j = 0; i <= nkeys; j++) {
6915 if (i == split_indx) {
6916 /* Insert in right sibling. */
6917 /* Reset insert index for right sibling. */
6918 if (i != newindx || (newpos ^ ins_new)) {
6920 mc->mc_pg[mc->mc_top] = rp;
6924 if (i == newindx && !ins_new) {
6925 /* Insert the original entry that caused the split. */
6926 rkey.mv_data = newkey->mv_data;
6927 rkey.mv_size = newkey->mv_size;
6936 /* Update index for the new key. */
6937 mc->mc_ki[mc->mc_top] = j;
6938 } else if (i == nkeys) {
6941 node = NODEPTR(mp, i);
6942 rkey.mv_data = NODEKEY(node);
6943 rkey.mv_size = node->mn_ksize;
6945 xdata.mv_data = NODEDATA(node);
6946 xdata.mv_size = NODEDSZ(node);
6949 pgno = NODEPGNO(node);
6950 flags = node->mn_flags;
6955 if (!IS_LEAF(mp) && j == 0) {
6956 /* First branch index doesn't need key data. */
6960 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6964 nkeys = NUMKEYS(copy);
6965 for (i=0; i<nkeys; i++)
6966 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6967 mp->mp_lower = copy->mp_lower;
6968 mp->mp_upper = copy->mp_upper;
6969 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6970 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6972 /* reset back to original page */
6973 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6974 mc->mc_pg[mc->mc_top] = mp;
6975 if (nflags & MDB_RESERVE) {
6976 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6977 if (!(node->mn_flags & F_BIGDATA))
6978 newdata->mv_data = NODEDATA(node);
6982 /* Make sure mc_ki is still valid.
6984 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6985 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6986 for (i=0; i<ptop; i++) {
6987 mc->mc_pg[i] = mn.mc_pg[i];
6988 mc->mc_ki[i] = mn.mc_ki[i];
6990 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6991 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6995 /* return tmp page to freelist */
6996 mdb_page_free(mc->mc_txn->mt_env, copy);
6999 /* Adjust other cursors pointing to mp */
7000 MDB_cursor *m2, *m3;
7001 MDB_dbi dbi = mc->mc_dbi;
7002 int fixup = NUMKEYS(mp);
7004 if (mc->mc_flags & C_SUB)
7007 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7008 if (mc->mc_flags & C_SUB)
7009 m3 = &m2->mc_xcursor->mx_cursor;
7014 if (!(m3->mc_flags & C_INITIALIZED))
7016 if (m3->mc_flags & C_SPLITTING)
7021 for (k=m3->mc_top; k>=0; k--) {
7022 m3->mc_ki[k+1] = m3->mc_ki[k];
7023 m3->mc_pg[k+1] = m3->mc_pg[k];
7025 if (m3->mc_ki[0] >= split_indx) {
7030 m3->mc_pg[0] = mc->mc_pg[0];
7034 if (m3->mc_pg[mc->mc_top] == mp) {
7035 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7036 m3->mc_ki[mc->mc_top]++;
7037 if (m3->mc_ki[mc->mc_top] >= fixup) {
7038 m3->mc_pg[mc->mc_top] = rp;
7039 m3->mc_ki[mc->mc_top] -= fixup;
7040 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7042 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7043 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7052 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7053 MDB_val *key, MDB_val *data, unsigned int flags)
7058 assert(key != NULL);
7059 assert(data != NULL);
7061 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7064 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7068 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7072 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7075 mdb_cursor_init(&mc, txn, dbi, &mx);
7076 return mdb_cursor_put(&mc, key, data, flags);
7080 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7082 if ((flag & CHANGEABLE) != flag)
7085 env->me_flags |= flag;
7087 env->me_flags &= ~flag;
7092 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7097 *arg = env->me_flags;
7102 mdb_env_get_path(MDB_env *env, const char **arg)
7107 *arg = env->me_path;
7111 /** Common code for #mdb_stat() and #mdb_env_stat().
7112 * @param[in] env the environment to operate in.
7113 * @param[in] db the #MDB_db record containing the stats to return.
7114 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7115 * @return 0, this function always succeeds.
7118 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7120 arg->ms_psize = env->me_psize;
7121 arg->ms_depth = db->md_depth;
7122 arg->ms_branch_pages = db->md_branch_pages;
7123 arg->ms_leaf_pages = db->md_leaf_pages;
7124 arg->ms_overflow_pages = db->md_overflow_pages;
7125 arg->ms_entries = db->md_entries;
7130 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7134 if (env == NULL || arg == NULL)
7137 toggle = mdb_env_pick_meta(env);
7139 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7143 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7147 if (env == NULL || arg == NULL)
7150 toggle = mdb_env_pick_meta(env);
7151 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7152 arg->me_mapsize = env->me_mapsize;
7153 arg->me_maxreaders = env->me_maxreaders;
7154 arg->me_numreaders = env->me_numreaders;
7155 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7156 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7160 /** Set the default comparison functions for a database.
7161 * Called immediately after a database is opened to set the defaults.
7162 * The user can then override them with #mdb_set_compare() or
7163 * #mdb_set_dupsort().
7164 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7165 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7168 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7170 uint16_t f = txn->mt_dbs[dbi].md_flags;
7172 txn->mt_dbxs[dbi].md_cmp =
7173 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7174 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7176 txn->mt_dbxs[dbi].md_dcmp =
7177 !(f & MDB_DUPSORT) ? 0 :
7178 ((f & MDB_INTEGERDUP)
7179 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7180 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7183 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7188 int rc, dbflag, exact;
7189 unsigned int unused = 0;
7192 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7193 mdb_default_cmp(txn, FREE_DBI);
7196 if ((flags & VALID_FLAGS) != flags)
7202 if (flags & PERSISTENT_FLAGS) {
7203 uint16_t f2 = flags & PERSISTENT_FLAGS;
7204 /* make sure flag changes get committed */
7205 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7206 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7207 txn->mt_flags |= MDB_TXN_DIRTY;
7210 mdb_default_cmp(txn, MAIN_DBI);
7214 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7215 mdb_default_cmp(txn, MAIN_DBI);
7218 /* Is the DB already open? */
7220 for (i=2; i<txn->mt_numdbs; i++) {
7221 if (!txn->mt_dbxs[i].md_name.mv_size) {
7222 /* Remember this free slot */
7223 if (!unused) unused = i;
7226 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7227 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7233 /* If no free slot and max hit, fail */
7234 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7235 return MDB_DBS_FULL;
7237 /* Cannot mix named databases with some mainDB flags */
7238 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7239 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7241 /* Find the DB info */
7242 dbflag = DB_NEW|DB_VALID;
7245 key.mv_data = (void *)name;
7246 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7247 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7248 if (rc == MDB_SUCCESS) {
7249 /* make sure this is actually a DB */
7250 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7251 if (!(node->mn_flags & F_SUBDATA))
7253 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7254 /* Create if requested */
7256 data.mv_size = sizeof(MDB_db);
7257 data.mv_data = &dummy;
7258 memset(&dummy, 0, sizeof(dummy));
7259 dummy.md_root = P_INVALID;
7260 dummy.md_flags = flags & PERSISTENT_FLAGS;
7261 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7265 /* OK, got info, add to table */
7266 if (rc == MDB_SUCCESS) {
7267 unsigned int slot = unused ? unused : txn->mt_numdbs;
7268 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7269 txn->mt_dbxs[slot].md_name.mv_size = len;
7270 txn->mt_dbxs[slot].md_rel = NULL;
7271 txn->mt_dbflags[slot] = dbflag;
7272 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7274 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7275 mdb_default_cmp(txn, slot);
7284 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7286 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7289 if (txn->mt_dbflags[dbi] & DB_STALE) {
7292 /* Stale, must read the DB's root. cursor_init does it for us. */
7293 mdb_cursor_init(&mc, txn, dbi, &mx);
7295 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7298 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7301 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7303 ptr = env->me_dbxs[dbi].md_name.mv_data;
7304 env->me_dbxs[dbi].md_name.mv_data = NULL;
7305 env->me_dbxs[dbi].md_name.mv_size = 0;
7306 env->me_dbflags[dbi] = 0;
7310 /** Add all the DB's pages to the free list.
7311 * @param[in] mc Cursor on the DB to free.
7312 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7313 * @return 0 on success, non-zero on failure.
7316 mdb_drop0(MDB_cursor *mc, int subs)
7320 rc = mdb_page_search(mc, NULL, 0);
7321 if (rc == MDB_SUCCESS) {
7322 MDB_txn *txn = mc->mc_txn;
7327 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7328 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7331 mdb_cursor_copy(mc, &mx);
7332 while (mc->mc_snum > 0) {
7333 MDB_page *mp = mc->mc_pg[mc->mc_top];
7334 unsigned n = NUMKEYS(mp);
7336 for (i=0; i<n; i++) {
7337 ni = NODEPTR(mp, i);
7338 if (ni->mn_flags & F_BIGDATA) {
7341 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7342 rc = mdb_page_get(txn, pg, &omp, NULL);
7345 assert(IS_OVERFLOW(omp));
7346 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7350 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7351 mdb_xcursor_init1(mc, ni);
7352 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7358 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7360 for (i=0; i<n; i++) {
7362 ni = NODEPTR(mp, i);
7365 mdb_midl_xappend(txn->mt_free_pgs, pg);
7370 mc->mc_ki[mc->mc_top] = i;
7371 rc = mdb_cursor_sibling(mc, 1);
7373 /* no more siblings, go back to beginning
7374 * of previous level.
7378 for (i=1; i<mc->mc_snum; i++) {
7380 mc->mc_pg[i] = mx.mc_pg[i];
7385 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
7386 } else if (rc == MDB_NOTFOUND) {
7392 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7394 MDB_cursor *mc, *m2;
7397 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7400 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7403 rc = mdb_cursor_open(txn, dbi, &mc);
7407 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7408 /* Invalidate the dropped DB's cursors */
7409 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7410 m2->mc_flags &= ~C_INITIALIZED;
7414 /* Can't delete the main DB */
7415 if (del && dbi > MAIN_DBI) {
7416 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7418 txn->mt_dbflags[dbi] = DB_STALE;
7419 mdb_dbi_close(txn->mt_env, dbi);
7422 /* reset the DB record, mark it dirty */
7423 txn->mt_dbflags[dbi] |= DB_DIRTY;
7424 txn->mt_dbs[dbi].md_depth = 0;
7425 txn->mt_dbs[dbi].md_branch_pages = 0;
7426 txn->mt_dbs[dbi].md_leaf_pages = 0;
7427 txn->mt_dbs[dbi].md_overflow_pages = 0;
7428 txn->mt_dbs[dbi].md_entries = 0;
7429 txn->mt_dbs[dbi].md_root = P_INVALID;
7431 txn->mt_flags |= MDB_TXN_DIRTY;
7434 mdb_cursor_close(mc);
7438 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7440 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7443 txn->mt_dbxs[dbi].md_cmp = cmp;
7447 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7449 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7452 txn->mt_dbxs[dbi].md_dcmp = cmp;
7456 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7458 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7461 txn->mt_dbxs[dbi].md_rel = rel;
7465 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7467 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7470 txn->mt_dbxs[dbi].md_relctx = ctx;