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 MDB_PIDLOCK 0
144 #define pthread_t DWORD
145 #define pthread_mutex_t HANDLE
146 #define pthread_key_t DWORD
147 #define pthread_self() GetCurrentThreadId()
148 #define pthread_key_create(x,y) \
149 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
150 #define pthread_key_delete(x) TlsFree(x)
151 #define pthread_getspecific(x) TlsGetValue(x)
152 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
153 #define pthread_mutex_unlock(x) ReleaseMutex(x)
154 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
155 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
156 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
157 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
158 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
159 #define getpid() GetCurrentProcessId()
160 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
161 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
162 #define ErrCode() GetLastError()
163 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
164 #define close(fd) (CloseHandle(fd) ? 0 : -1)
165 #define munmap(ptr,len) UnmapViewOfFile(ptr)
166 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
167 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
169 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
174 #define Z "z" /**< printf format modifier for size_t */
176 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
177 #define MDB_PIDLOCK 1
179 #ifdef MDB_USE_POSIX_SEM
181 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
182 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
183 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
184 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
187 mdb_sem_wait(sem_t *sem)
190 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
195 /** Lock the reader mutex.
197 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
198 /** Unlock the reader mutex.
200 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
202 /** Lock the writer mutex.
203 * Only a single write transaction is allowed at a time. Other writers
204 * will block waiting for this mutex.
206 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
207 /** Unlock the writer mutex.
209 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
210 #endif /* MDB_USE_POSIX_SEM */
212 /** Get the error code for the last failed system function.
214 #define ErrCode() errno
216 /** An abstraction for a file handle.
217 * On POSIX systems file handles are small integers. On Windows
218 * they're opaque pointers.
222 /** A value for an invalid file handle.
223 * Mainly used to initialize file variables and signify that they are
226 #define INVALID_HANDLE_VALUE (-1)
228 /** Get the size of a memory page for the system.
229 * This is the basic size that the platform's memory manager uses, and is
230 * fundamental to the use of memory-mapped files.
232 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
235 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
238 #define MNAME_LEN (sizeof(pthread_mutex_t))
244 /** A flag for opening a file and requesting synchronous data writes.
245 * This is only used when writing a meta page. It's not strictly needed;
246 * we could just do a normal write and then immediately perform a flush.
247 * But if this flag is available it saves us an extra system call.
249 * @note If O_DSYNC is undefined but exists in /usr/include,
250 * preferably set some compiler flag to get the definition.
251 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
254 # define MDB_DSYNC O_DSYNC
258 /** Function for flushing the data of a file. Define this to fsync
259 * if fdatasync() is not supported.
261 #ifndef MDB_FDATASYNC
262 # define MDB_FDATASYNC fdatasync
266 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
277 /** A page number in the database.
278 * Note that 64 bit page numbers are overkill, since pages themselves
279 * already represent 12-13 bits of addressable memory, and the OS will
280 * always limit applications to a maximum of 63 bits of address space.
282 * @note In the #MDB_node structure, we only store 48 bits of this value,
283 * which thus limits us to only 60 bits of addressable data.
285 typedef MDB_ID pgno_t;
287 /** A transaction ID.
288 * See struct MDB_txn.mt_txnid for details.
290 typedef MDB_ID txnid_t;
292 /** @defgroup debug Debug Macros
296 /** Enable debug output. Needs variable argument macros (a C99 feature).
297 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
298 * read from and written to the database (used for free space management).
304 static int mdb_debug;
305 static txnid_t mdb_debug_start;
307 /** Print a debug message with printf formatting.
308 * Requires double parenthesis around 2 or more args.
310 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
311 # define DPRINTF0(fmt, ...) \
312 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
314 # define DPRINTF(args) ((void) 0)
316 /** Print a debug string.
317 * The string is printed literally, with no format processing.
319 #define DPUTS(arg) DPRINTF(("%s", arg))
322 /** A default memory page size.
323 * The actual size is platform-dependent, but we use this for
324 * boot-strapping. We probably should not be using this any more.
325 * The #GET_PAGESIZE() macro is used to get the actual size.
327 * Note that we don't currently support Huge pages. On Linux,
328 * regular data files cannot use Huge pages, and in general
329 * Huge pages aren't actually pageable. We rely on the OS
330 * demand-pager to read our data and page it out when memory
331 * pressure from other processes is high. So until OSs have
332 * actual paging support for Huge pages, they're not viable.
334 #define MDB_PAGESIZE 4096
336 /** The minimum number of keys required in a database page.
337 * Setting this to a larger value will place a smaller bound on the
338 * maximum size of a data item. Data items larger than this size will
339 * be pushed into overflow pages instead of being stored directly in
340 * the B-tree node. This value used to default to 4. With a page size
341 * of 4096 bytes that meant that any item larger than 1024 bytes would
342 * go into an overflow page. That also meant that on average 2-3KB of
343 * each overflow page was wasted space. The value cannot be lower than
344 * 2 because then there would no longer be a tree structure. With this
345 * value, items larger than 2KB will go into overflow pages, and on
346 * average only 1KB will be wasted.
348 #define MDB_MINKEYS 2
350 /** A stamp that identifies a file as an MDB file.
351 * There's nothing special about this value other than that it is easily
352 * recognizable, and it will reflect any byte order mismatches.
354 #define MDB_MAGIC 0xBEEFC0DE
356 /** The version number for a database's datafile format. */
357 #define MDB_DATA_VERSION 1
358 /** The version number for a database's lockfile format. */
359 #define MDB_LOCK_VERSION 1
361 /** @brief The maximum size of a key in the database.
363 * The library rejects bigger keys, and cannot deal with records
364 * with bigger keys stored by a library with bigger max keysize.
366 * We require that keys all fit onto a regular page. This limit
367 * could be raised a bit further if needed; to something just
368 * under #MDB_PAGESIZE / #MDB_MINKEYS.
370 * Note that data items in an #MDB_DUPSORT database are actually keys
371 * of a subDB, so they're also limited to this size.
373 #ifndef MDB_MAXKEYSIZE
374 #define MDB_MAXKEYSIZE 511
377 /** @brief The maximum size of a data item.
379 * We only store a 32 bit value for node sizes.
381 #define MAXDATASIZE 0xffffffffUL
386 * This is used for printing a hex dump of a key's contents.
388 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
389 /** Display a key in hex.
391 * Invoke a function to display a key in hex.
393 #define DKEY(x) mdb_dkey(x, kbuf)
399 /** An invalid page number.
400 * Mainly used to denote an empty tree.
402 #define P_INVALID (~(pgno_t)0)
404 /** Test if the flags \b f are set in a flag word \b w. */
405 #define F_ISSET(w, f) (((w) & (f)) == (f))
407 /** Used for offsets within a single page.
408 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
411 typedef uint16_t indx_t;
413 /** Default size of memory map.
414 * This is certainly too small for any actual applications. Apps should always set
415 * the size explicitly using #mdb_env_set_mapsize().
417 #define DEFAULT_MAPSIZE 1048576
419 /** @defgroup readers Reader Lock Table
420 * Readers don't acquire any locks for their data access. Instead, they
421 * simply record their transaction ID in the reader table. The reader
422 * mutex is needed just to find an empty slot in the reader table. The
423 * slot's address is saved in thread-specific data so that subsequent read
424 * transactions started by the same thread need no further locking to proceed.
426 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
428 * No reader table is used if the database is on a read-only filesystem.
430 * Since the database uses multi-version concurrency control, readers don't
431 * actually need any locking. This table is used to keep track of which
432 * readers are using data from which old transactions, so that we'll know
433 * when a particular old transaction is no longer in use. Old transactions
434 * that have discarded any data pages can then have those pages reclaimed
435 * for use by a later write transaction.
437 * The lock table is constructed such that reader slots are aligned with the
438 * processor's cache line size. Any slot is only ever used by one thread.
439 * This alignment guarantees that there will be no contention or cache
440 * thrashing as threads update their own slot info, and also eliminates
441 * any need for locking when accessing a slot.
443 * A writer thread will scan every slot in the table to determine the oldest
444 * outstanding reader transaction. Any freed pages older than this will be
445 * reclaimed by the writer. The writer doesn't use any locks when scanning
446 * this table. This means that there's no guarantee that the writer will
447 * see the most up-to-date reader info, but that's not required for correct
448 * operation - all we need is to know the upper bound on the oldest reader,
449 * we don't care at all about the newest reader. So the only consequence of
450 * reading stale information here is that old pages might hang around a
451 * while longer before being reclaimed. That's actually good anyway, because
452 * the longer we delay reclaiming old pages, the more likely it is that a
453 * string of contiguous pages can be found after coalescing old pages from
454 * many old transactions together.
457 /** Number of slots in the reader table.
458 * This value was chosen somewhat arbitrarily. 126 readers plus a
459 * couple mutexes fit exactly into 8KB on my development machine.
460 * Applications should set the table size using #mdb_env_set_maxreaders().
462 #define DEFAULT_READERS 126
464 /** The size of a CPU cache line in bytes. We want our lock structures
465 * aligned to this size to avoid false cache line sharing in the
467 * This value works for most CPUs. For Itanium this should be 128.
473 /** The information we store in a single slot of the reader table.
474 * In addition to a transaction ID, we also record the process and
475 * thread ID that owns a slot, so that we can detect stale information,
476 * e.g. threads or processes that went away without cleaning up.
477 * @note We currently don't check for stale records. We simply re-init
478 * the table when we know that we're the only process opening the
481 typedef struct MDB_rxbody {
482 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
483 * Multiple readers that start at the same time will probably have the
484 * same ID here. Again, it's not important to exclude them from
485 * anything; all we need to know is which version of the DB they
486 * started from so we can avoid overwriting any data used in that
487 * particular version.
490 /** The process ID of the process owning this reader txn. */
492 /** The thread ID of the thread owning this txn. */
496 /** The actual reader record, with cacheline padding. */
497 typedef struct MDB_reader {
500 /** shorthand for mrb_txnid */
501 #define mr_txnid mru.mrx.mrb_txnid
502 #define mr_pid mru.mrx.mrb_pid
503 #define mr_tid mru.mrx.mrb_tid
504 /** cache line alignment */
505 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
509 /** The header for the reader table.
510 * The table resides in a memory-mapped file. (This is a different file
511 * than is used for the main database.)
513 * For POSIX the actual mutexes reside in the shared memory of this
514 * mapped file. On Windows, mutexes are named objects allocated by the
515 * kernel; we store the mutex names in this mapped file so that other
516 * processes can grab them. This same approach is also used on
517 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
518 * process-shared POSIX mutexes. For these cases where a named object
519 * is used, the object name is derived from a 64 bit FNV hash of the
520 * environment pathname. As such, naming collisions are extremely
521 * unlikely. If a collision occurs, the results are unpredictable.
523 typedef struct MDB_txbody {
524 /** Stamp identifying this as an MDB file. It must be set
527 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
529 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
530 char mtb_rmname[MNAME_LEN];
532 /** Mutex protecting access to this table.
533 * This is the reader lock that #LOCK_MUTEX_R acquires.
535 pthread_mutex_t mtb_mutex;
537 /** The ID of the last transaction committed to the database.
538 * This is recorded here only for convenience; the value can always
539 * be determined by reading the main database meta pages.
542 /** The number of slots that have been used in the reader table.
543 * This always records the maximum count, it is not decremented
544 * when readers release their slots.
546 unsigned mtb_numreaders;
549 /** The actual reader table definition. */
550 typedef struct MDB_txninfo {
553 #define mti_magic mt1.mtb.mtb_magic
554 #define mti_format mt1.mtb.mtb_format
555 #define mti_mutex mt1.mtb.mtb_mutex
556 #define mti_rmname mt1.mtb.mtb_rmname
557 #define mti_txnid mt1.mtb.mtb_txnid
558 #define mti_numreaders mt1.mtb.mtb_numreaders
559 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
562 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
563 char mt2_wmname[MNAME_LEN];
564 #define mti_wmname mt2.mt2_wmname
566 pthread_mutex_t mt2_wmutex;
567 #define mti_wmutex mt2.mt2_wmutex
569 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
571 MDB_reader mti_readers[1];
574 /** Lockfile format signature: version, features and field layout */
575 #define MDB_LOCK_FORMAT \
577 ((MDB_LOCK_VERSION) \
578 /* Flags which describe functionality */ \
579 + (((MDB_PIDLOCK) != 0) << 16)))
582 /** Common header for all page types.
583 * Overflow records occupy a number of contiguous pages with no
584 * headers on any page after the first.
586 typedef struct MDB_page {
587 #define mp_pgno mp_p.p_pgno
588 #define mp_next mp_p.p_next
590 pgno_t p_pgno; /**< page number */
591 void * p_next; /**< for in-memory list of freed structs */
594 /** @defgroup mdb_page Page Flags
596 * Flags for the page headers.
599 #define P_BRANCH 0x01 /**< branch page */
600 #define P_LEAF 0x02 /**< leaf page */
601 #define P_OVERFLOW 0x04 /**< overflow page */
602 #define P_META 0x08 /**< meta page */
603 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
604 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
605 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
606 #define P_KEEP 0x8000 /**< leave this page alone during spill */
608 uint16_t mp_flags; /**< @ref mdb_page */
609 #define mp_lower mp_pb.pb.pb_lower
610 #define mp_upper mp_pb.pb.pb_upper
611 #define mp_pages mp_pb.pb_pages
614 indx_t pb_lower; /**< lower bound of free space */
615 indx_t pb_upper; /**< upper bound of free space */
617 uint32_t pb_pages; /**< number of overflow pages */
619 indx_t mp_ptrs[1]; /**< dynamic size */
622 /** Size of the page header, excluding dynamic data at the end */
623 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
625 /** Address of first usable data byte in a page, after the header */
626 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
628 /** Number of nodes on a page */
629 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
631 /** The amount of space remaining in the page */
632 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
634 /** The percentage of space used in the page, in tenths of a percent. */
635 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
636 ((env)->me_psize - PAGEHDRSZ))
637 /** The minimum page fill factor, in tenths of a percent.
638 * Pages emptier than this are candidates for merging.
640 #define FILL_THRESHOLD 250
642 /** Test if a page is a leaf page */
643 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
644 /** Test if a page is a LEAF2 page */
645 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
646 /** Test if a page is a branch page */
647 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
648 /** Test if a page is an overflow page */
649 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
650 /** Test if a page is a sub page */
651 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
653 /** The number of overflow pages needed to store the given size. */
654 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
656 /** Header for a single key/data pair within a page.
657 * We guarantee 2-byte alignment for nodes.
659 typedef struct MDB_node {
660 /** lo and hi are used for data size on leaf nodes and for
661 * child pgno on branch nodes. On 64 bit platforms, flags
662 * is also used for pgno. (Branch nodes have no flags).
663 * They are in host byte order in case that lets some
664 * accesses be optimized into a 32-bit word access.
666 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
667 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
668 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
669 /** @defgroup mdb_node Node Flags
671 * Flags for node headers.
674 #define F_BIGDATA 0x01 /**< data put on overflow page */
675 #define F_SUBDATA 0x02 /**< data is a sub-database */
676 #define F_DUPDATA 0x04 /**< data has duplicates */
678 /** valid flags for #mdb_node_add() */
679 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
682 unsigned short mn_flags; /**< @ref mdb_node */
683 unsigned short mn_ksize; /**< key size */
684 char mn_data[1]; /**< key and data are appended here */
687 /** Size of the node header, excluding dynamic data at the end */
688 #define NODESIZE offsetof(MDB_node, mn_data)
690 /** Bit position of top word in page number, for shifting mn_flags */
691 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
693 /** Size of a node in a branch page with a given key.
694 * This is just the node header plus the key, there is no data.
696 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
698 /** Size of a node in a leaf page with a given key and data.
699 * This is node header plus key plus data size.
701 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
703 /** Address of node \b i in page \b p */
704 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
706 /** Address of the key for the node */
707 #define NODEKEY(node) (void *)((node)->mn_data)
709 /** Address of the data for a node */
710 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
712 /** Get the page number pointed to by a branch node */
713 #define NODEPGNO(node) \
714 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
715 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
716 /** Set the page number in a branch node */
717 #define SETPGNO(node,pgno) do { \
718 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
719 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
721 /** Get the size of the data in a leaf node */
722 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
723 /** Set the size of the data for a leaf node */
724 #define SETDSZ(node,size) do { \
725 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
726 /** The size of a key in a node */
727 #define NODEKSZ(node) ((node)->mn_ksize)
729 /** Copy a page number from src to dst */
731 #define COPY_PGNO(dst,src) dst = src
733 #if SIZE_MAX > 4294967295UL
734 #define COPY_PGNO(dst,src) do { \
735 unsigned short *s, *d; \
736 s = (unsigned short *)&(src); \
737 d = (unsigned short *)&(dst); \
744 #define COPY_PGNO(dst,src) do { \
745 unsigned short *s, *d; \
746 s = (unsigned short *)&(src); \
747 d = (unsigned short *)&(dst); \
753 /** The address of a key in a LEAF2 page.
754 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
755 * There are no node headers, keys are stored contiguously.
757 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
759 /** Set the \b node's key into \b keyptr, if requested. */
760 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
761 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
763 /** Set the \b node's key into \b key. */
764 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
766 /** Information about a single database in the environment. */
767 typedef struct MDB_db {
768 uint32_t md_pad; /**< also ksize for LEAF2 pages */
769 uint16_t md_flags; /**< @ref mdb_dbi_open */
770 uint16_t md_depth; /**< depth of this tree */
771 pgno_t md_branch_pages; /**< number of internal pages */
772 pgno_t md_leaf_pages; /**< number of leaf pages */
773 pgno_t md_overflow_pages; /**< number of overflow pages */
774 size_t md_entries; /**< number of data items */
775 pgno_t md_root; /**< the root page of this tree */
778 /** mdb_dbi_open flags */
779 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
780 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
781 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
782 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
784 /** Handle for the DB used to track free pages. */
786 /** Handle for the default DB. */
789 /** Meta page content.
790 * A meta page is the start point for accessing a database snapshot.
791 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
793 typedef struct MDB_meta {
794 /** Stamp identifying this as an MDB file. It must be set
797 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
799 void *mm_address; /**< address for fixed mapping */
800 size_t mm_mapsize; /**< size of mmap region */
801 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
802 /** The size of pages used in this DB */
803 #define mm_psize mm_dbs[0].md_pad
804 /** Any persistent environment flags. @ref mdb_env */
805 #define mm_flags mm_dbs[0].md_flags
806 pgno_t mm_last_pg; /**< last used page in file */
807 txnid_t mm_txnid; /**< txnid that committed this page */
810 /** Buffer for a stack-allocated dirty page.
811 * The members define size and alignment, and silence type
812 * aliasing warnings. They are not used directly; that could
813 * mean incorrectly using several union members in parallel.
815 typedef union MDB_pagebuf {
816 char mb_raw[MDB_PAGESIZE];
819 char mm_pad[PAGEHDRSZ];
824 /** Auxiliary DB info.
825 * The information here is mostly static/read-only. There is
826 * only a single copy of this record in the environment.
828 typedef struct MDB_dbx {
829 MDB_val md_name; /**< name of the database */
830 MDB_cmp_func *md_cmp; /**< function for comparing keys */
831 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
832 MDB_rel_func *md_rel; /**< user relocate function */
833 void *md_relctx; /**< user-provided context for md_rel */
836 /** A database transaction.
837 * Every operation requires a transaction handle.
840 MDB_txn *mt_parent; /**< parent of a nested txn */
841 MDB_txn *mt_child; /**< nested txn under this txn */
842 pgno_t mt_next_pgno; /**< next unallocated page */
843 /** The ID of this transaction. IDs are integers incrementing from 1.
844 * Only committed write transactions increment the ID. If a transaction
845 * aborts, the ID may be re-used by the next writer.
848 MDB_env *mt_env; /**< the DB environment */
849 /** The list of pages that became unused during this transaction.
852 /** The sorted list of dirty pages we temporarily wrote to disk
853 * because the dirty list was full. page numbers in here are
854 * shifted left by 1, deleted slots have the LSB set.
856 MDB_IDL mt_spill_pgs;
858 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
860 /** For read txns: This thread/txn's reader table slot, or NULL. */
863 /** Array of records for each DB known in the environment. */
865 /** Array of MDB_db records for each known DB */
867 /** @defgroup mt_dbflag Transaction DB Flags
871 #define DB_DIRTY 0x01 /**< DB was written in this txn */
872 #define DB_STALE 0x02 /**< DB record is older than txnID */
873 #define DB_NEW 0x04 /**< DB handle opened in this txn */
874 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
876 /** In write txns, array of cursors for each DB */
877 MDB_cursor **mt_cursors;
878 /** Array of flags for each DB */
879 unsigned char *mt_dbflags;
880 /** Number of DB records in use. This number only ever increments;
881 * we don't decrement it when individual DB handles are closed.
885 /** @defgroup mdb_txn Transaction Flags
889 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
890 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
891 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
892 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
894 unsigned int mt_flags; /**< @ref mdb_txn */
895 /** dirty_list room: Array size - #dirty pages visible to this txn.
896 * Includes ancestor txns' dirty pages not hidden by other txns'
897 * dirty/spilled pages. Thus commit(nested txn) has room to merge
898 * dirty_list into mt_parent after freeing hidden mt_parent pages.
900 unsigned int mt_dirty_room;
901 /** Tracks which of the two meta pages was used at the start
902 * of this transaction.
904 unsigned int mt_toggle;
907 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
908 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
909 * raise this on a 64 bit machine.
911 #define CURSOR_STACK 32
915 /** Cursors are used for all DB operations.
916 * A cursor holds a path of (page pointer, key index) from the DB
917 * root to a position in the DB, plus other state. #MDB_DUPSORT
918 * cursors include an xcursor to the current data item. Write txns
919 * track their cursors and keep them up to date when data moves.
920 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
921 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
924 /** Next cursor on this DB in this txn */
926 /** Backup of the original cursor if this cursor is a shadow */
927 MDB_cursor *mc_backup;
928 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
929 struct MDB_xcursor *mc_xcursor;
930 /** The transaction that owns this cursor */
932 /** The database handle this cursor operates on */
934 /** The database record for this cursor */
936 /** The database auxiliary record for this cursor */
938 /** The @ref mt_dbflag for this database */
939 unsigned char *mc_dbflag;
940 unsigned short mc_snum; /**< number of pushed pages */
941 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
942 /** @defgroup mdb_cursor Cursor Flags
944 * Cursor state flags.
947 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
948 #define C_EOF 0x02 /**< No more data */
949 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
950 #define C_DEL 0x08 /**< last op was a cursor_del */
951 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
952 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
954 unsigned int mc_flags; /**< @ref mdb_cursor */
955 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
956 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
959 /** Context for sorted-dup records.
960 * We could have gone to a fully recursive design, with arbitrarily
961 * deep nesting of sub-databases. But for now we only handle these
962 * levels - main DB, optional sub-DB, sorted-duplicate DB.
964 typedef struct MDB_xcursor {
965 /** A sub-cursor for traversing the Dup DB */
966 MDB_cursor mx_cursor;
967 /** The database record for this Dup DB */
969 /** The auxiliary DB record for this Dup DB */
971 /** The @ref mt_dbflag for this Dup DB */
972 unsigned char mx_dbflag;
975 /** State of FreeDB old pages, stored in the MDB_env */
976 typedef struct MDB_pgstate {
977 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
978 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
981 /** The database environment. */
983 HANDLE me_fd; /**< The main data file */
984 HANDLE me_lfd; /**< The lock file */
985 HANDLE me_mfd; /**< just for writing the meta pages */
986 /** Failed to update the meta page. Probably an I/O error. */
987 #define MDB_FATAL_ERROR 0x80000000U
988 /** Some fields are initialized. */
989 #define MDB_ENV_ACTIVE 0x20000000U
990 /** me_txkey is set */
991 #define MDB_ENV_TXKEY 0x10000000U
992 /** Have liveness lock in reader table */
993 #define MDB_LIVE_READER 0x08000000U
994 uint32_t me_flags; /**< @ref mdb_env */
995 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
996 unsigned int me_maxreaders; /**< size of the reader table */
997 unsigned int me_numreaders; /**< max numreaders set by this env */
998 MDB_dbi me_numdbs; /**< number of DBs opened */
999 MDB_dbi me_maxdbs; /**< size of the DB table */
1000 pid_t me_pid; /**< process ID of this env */
1001 char *me_path; /**< path to the DB files */
1002 char *me_map; /**< the memory map of the data file */
1003 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1004 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1005 MDB_txn *me_txn; /**< current write transaction */
1006 size_t me_mapsize; /**< size of the data memory map */
1007 off_t me_size; /**< current file size */
1008 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1009 MDB_dbx *me_dbxs; /**< array of static DB info */
1010 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1011 pthread_key_t me_txkey; /**< thread-key for readers */
1012 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1013 # define me_pglast me_pgstate.mf_pglast
1014 # define me_pghead me_pgstate.mf_pghead
1015 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1016 /** IDL of pages that became unused in a write txn */
1017 MDB_IDL me_free_pgs;
1018 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1019 MDB_ID2L me_dirty_list;
1020 /** Max number of freelist items that can fit in a single overflow page */
1022 /** Max size of a node on a page */
1023 unsigned int me_nodemax;
1025 int me_pidquery; /**< Used in OpenProcess */
1026 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1028 #elif defined(MDB_USE_POSIX_SEM)
1029 sem_t *me_rmutex; /* Shared mutexes are not supported */
1034 /** Nested transaction */
1035 typedef struct MDB_ntxn {
1036 MDB_txn mnt_txn; /**< the transaction */
1037 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1040 /** max number of pages to commit in one writev() call */
1041 #define MDB_COMMIT_PAGES 64
1042 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1043 #undef MDB_COMMIT_PAGES
1044 #define MDB_COMMIT_PAGES IOV_MAX
1047 /* max bytes to write in one call */
1048 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1050 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1051 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1052 static int mdb_page_touch(MDB_cursor *mc);
1054 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1055 static int mdb_page_search_root(MDB_cursor *mc,
1056 MDB_val *key, int modify);
1057 #define MDB_PS_MODIFY 1
1058 #define MDB_PS_ROOTONLY 2
1059 static int mdb_page_search(MDB_cursor *mc,
1060 MDB_val *key, int flags);
1061 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1063 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1064 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1065 pgno_t newpgno, unsigned int nflags);
1067 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1068 static int mdb_env_pick_meta(const MDB_env *env);
1069 static int mdb_env_write_meta(MDB_txn *txn);
1070 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1071 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1073 static void mdb_env_close0(MDB_env *env, int excl);
1075 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1076 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1077 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1078 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1079 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1080 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1081 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1082 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1083 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1085 static int mdb_rebalance(MDB_cursor *mc);
1086 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1088 static void mdb_cursor_pop(MDB_cursor *mc);
1089 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1091 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1092 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1093 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1094 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1095 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1097 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1098 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1100 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1101 static void mdb_xcursor_init0(MDB_cursor *mc);
1102 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1104 static int mdb_drop0(MDB_cursor *mc, int subs);
1105 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1108 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1112 static SECURITY_DESCRIPTOR mdb_null_sd;
1113 static SECURITY_ATTRIBUTES mdb_all_sa;
1114 static int mdb_sec_inited;
1117 /** Return the library version info. */
1119 mdb_version(int *major, int *minor, int *patch)
1121 if (major) *major = MDB_VERSION_MAJOR;
1122 if (minor) *minor = MDB_VERSION_MINOR;
1123 if (patch) *patch = MDB_VERSION_PATCH;
1124 return MDB_VERSION_STRING;
1127 /** Table of descriptions for MDB @ref errors */
1128 static char *const mdb_errstr[] = {
1129 "MDB_KEYEXIST: Key/data pair already exists",
1130 "MDB_NOTFOUND: No matching key/data pair found",
1131 "MDB_PAGE_NOTFOUND: Requested page not found",
1132 "MDB_CORRUPTED: Located page was wrong type",
1133 "MDB_PANIC: Update of meta page failed",
1134 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1135 "MDB_INVALID: File is not an MDB file",
1136 "MDB_MAP_FULL: Environment mapsize limit reached",
1137 "MDB_DBS_FULL: Environment maxdbs limit reached",
1138 "MDB_READERS_FULL: Environment maxreaders limit reached",
1139 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1140 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1141 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1142 "MDB_PAGE_FULL: Internal error - page has no more space",
1143 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1144 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1145 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1146 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1147 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1151 mdb_strerror(int err)
1155 return ("Successful return: 0");
1157 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1158 i = err - MDB_KEYEXIST;
1159 return mdb_errstr[i];
1162 return strerror(err);
1166 /** Display a key in hexadecimal and return the address of the result.
1167 * @param[in] key the key to display
1168 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1169 * @return The key in hexadecimal form.
1172 mdb_dkey(MDB_val *key, char *buf)
1175 unsigned char *c = key->mv_data;
1181 if (key->mv_size > MDB_MAXKEYSIZE)
1182 return "MDB_MAXKEYSIZE";
1183 /* may want to make this a dynamic check: if the key is mostly
1184 * printable characters, print it as-is instead of converting to hex.
1188 for (i=0; i<key->mv_size; i++)
1189 ptr += sprintf(ptr, "%02x", *c++);
1191 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1196 /** Display all the keys in the page. */
1198 mdb_page_list(MDB_page *mp)
1201 unsigned int i, nkeys, nsize;
1205 nkeys = NUMKEYS(mp);
1206 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1207 for (i=0; i<nkeys; i++) {
1208 node = NODEPTR(mp, i);
1209 key.mv_size = node->mn_ksize;
1210 key.mv_data = node->mn_data;
1211 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1212 if (IS_BRANCH(mp)) {
1213 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1216 if (F_ISSET(node->mn_flags, F_BIGDATA))
1217 nsize += sizeof(pgno_t);
1219 nsize += NODEDSZ(node);
1220 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1226 mdb_cursor_chk(MDB_cursor *mc)
1232 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1233 for (i=0; i<mc->mc_top; i++) {
1235 node = NODEPTR(mp, mc->mc_ki[i]);
1236 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1239 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1245 /** Count all the pages in each DB and in the freelist
1246 * and make sure it matches the actual number of pages
1249 static void mdb_audit(MDB_txn *txn)
1253 MDB_ID freecount, count;
1258 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1259 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1260 freecount += *(MDB_ID *)data.mv_data;
1263 for (i = 0; i<txn->mt_numdbs; i++) {
1265 mdb_cursor_init(&mc, txn, i, &mx);
1266 if (txn->mt_dbs[i].md_root == P_INVALID)
1268 count += txn->mt_dbs[i].md_branch_pages +
1269 txn->mt_dbs[i].md_leaf_pages +
1270 txn->mt_dbs[i].md_overflow_pages;
1271 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1272 mdb_page_search(&mc, NULL, 0);
1276 mp = mc.mc_pg[mc.mc_top];
1277 for (j=0; j<NUMKEYS(mp); j++) {
1278 MDB_node *leaf = NODEPTR(mp, j);
1279 if (leaf->mn_flags & F_SUBDATA) {
1281 memcpy(&db, NODEDATA(leaf), sizeof(db));
1282 count += db.md_branch_pages + db.md_leaf_pages +
1283 db.md_overflow_pages;
1287 while (mdb_cursor_sibling(&mc, 1) == 0);
1290 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1291 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1292 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1298 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1300 return txn->mt_dbxs[dbi].md_cmp(a, b);
1304 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1306 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1309 /** Allocate memory for a page.
1310 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1313 mdb_page_malloc(MDB_txn *txn, unsigned num)
1315 MDB_env *env = txn->mt_env;
1316 MDB_page *ret = env->me_dpages;
1317 size_t sz = env->me_psize;
1320 VGMEMP_ALLOC(env, ret, sz);
1321 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1322 env->me_dpages = ret->mp_next;
1328 if ((ret = malloc(sz)) != NULL) {
1329 VGMEMP_ALLOC(env, ret, sz);
1334 /** Free a single page.
1335 * Saves single pages to a list, for future reuse.
1336 * (This is not used for multi-page overflow pages.)
1339 mdb_page_free(MDB_env *env, MDB_page *mp)
1341 mp->mp_next = env->me_dpages;
1342 VGMEMP_FREE(env, mp);
1343 env->me_dpages = mp;
1346 /** Free a dirty page */
1348 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1350 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1351 mdb_page_free(env, dp);
1353 /* large pages just get freed directly */
1354 VGMEMP_FREE(env, dp);
1359 /** Return all dirty pages to dpage list */
1361 mdb_dlist_free(MDB_txn *txn)
1363 MDB_env *env = txn->mt_env;
1364 MDB_ID2L dl = txn->mt_u.dirty_list;
1365 unsigned i, n = dl[0].mid;
1367 for (i = 1; i <= n; i++) {
1368 mdb_dpage_free(env, dl[i].mptr);
1373 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1374 * @param[in] mc A cursor handle for the current operation.
1375 * @param[in] pflags Flags of the pages to update:
1376 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1377 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1378 * @return 0 on success, non-zero on failure.
1381 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1383 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1384 MDB_txn *txn = mc->mc_txn;
1390 int rc = MDB_SUCCESS, level;
1392 /* Mark pages seen by cursors */
1393 if (mc->mc_flags & C_UNTRACK)
1394 mc = NULL; /* will find mc in mt_cursors */
1395 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1396 for (; mc; mc=mc->mc_next) {
1397 if (!(mc->mc_flags & C_INITIALIZED))
1399 for (m3 = mc;; m3 = &mx->mx_cursor) {
1401 for (j=0; j<m3->mc_snum; j++) {
1403 if ((mp->mp_flags & Mask) == pflags)
1404 mp->mp_flags ^= P_KEEP;
1406 mx = m3->mc_xcursor;
1407 /* Proceed to mx if it is at a sub-database */
1408 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1410 if (! (mp && (mp->mp_flags & P_LEAF)))
1412 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1413 if (!(leaf->mn_flags & F_SUBDATA))
1422 /* Mark dirty root pages */
1423 for (i=0; i<txn->mt_numdbs; i++) {
1424 if (txn->mt_dbflags[i] & DB_DIRTY) {
1425 pgno_t pgno = txn->mt_dbs[i].md_root;
1426 if (pgno == P_INVALID)
1428 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1430 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1431 dp->mp_flags ^= P_KEEP;
1439 static int mdb_page_flush(MDB_txn *txn, int keep);
1441 /** Spill pages from the dirty list back to disk.
1442 * This is intended to prevent running into #MDB_TXN_FULL situations,
1443 * but note that they may still occur in a few cases:
1444 * 1) our estimate of the txn size could be too small. Currently this
1445 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1446 * 2) child txns may run out of space if their parents dirtied a
1447 * lot of pages and never spilled them. TODO: we probably should do
1448 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1449 * the parent's dirty_room is below a given threshold.
1451 * Otherwise, if not using nested txns, it is expected that apps will
1452 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1453 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1454 * If the txn never references them again, they can be left alone.
1455 * If the txn only reads them, they can be used without any fuss.
1456 * If the txn writes them again, they can be dirtied immediately without
1457 * going thru all of the work of #mdb_page_touch(). Such references are
1458 * handled by #mdb_page_unspill().
1460 * Also note, we never spill DB root pages, nor pages of active cursors,
1461 * because we'll need these back again soon anyway. And in nested txns,
1462 * we can't spill a page in a child txn if it was already spilled in a
1463 * parent txn. That would alter the parent txns' data even though
1464 * the child hasn't committed yet, and we'd have no way to undo it if
1465 * the child aborted.
1467 * @param[in] m0 cursor A cursor handle identifying the transaction and
1468 * database for which we are checking space.
1469 * @param[in] key For a put operation, the key being stored.
1470 * @param[in] data For a put operation, the data being stored.
1471 * @return 0 on success, non-zero on failure.
1474 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1476 MDB_txn *txn = m0->mc_txn;
1478 MDB_ID2L dl = txn->mt_u.dirty_list;
1479 unsigned int i, j, need;
1482 if (m0->mc_flags & C_SUB)
1485 /* Estimate how much space this op will take */
1486 i = m0->mc_db->md_depth;
1487 /* Named DBs also dirty the main DB */
1488 if (m0->mc_dbi > MAIN_DBI)
1489 i += txn->mt_dbs[MAIN_DBI].md_depth;
1490 /* For puts, roughly factor in the key+data size */
1492 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1493 i += i; /* double it for good measure */
1496 if (txn->mt_dirty_room > i)
1499 if (!txn->mt_spill_pgs) {
1500 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1501 if (!txn->mt_spill_pgs)
1504 /* purge deleted slots */
1505 MDB_IDL sl = txn->mt_spill_pgs;
1506 unsigned int num = sl[0];
1508 for (i=1; i<=num; i++) {
1515 /* Preserve pages which may soon be dirtied again */
1516 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1519 /* Less aggressive spill - we originally spilled the entire dirty list,
1520 * with a few exceptions for cursor pages and DB root pages. But this
1521 * turns out to be a lot of wasted effort because in a large txn many
1522 * of those pages will need to be used again. So now we spill only 1/8th
1523 * of the dirty pages. Testing revealed this to be a good tradeoff,
1524 * better than 1/2, 1/4, or 1/10.
1526 if (need < MDB_IDL_UM_MAX / 8)
1527 need = MDB_IDL_UM_MAX / 8;
1529 /* Save the page IDs of all the pages we're flushing */
1530 /* flush from the tail forward, this saves a lot of shifting later on. */
1531 for (i=dl[0].mid; i && need; i--) {
1532 MDB_ID pn = dl[i].mid << 1;
1534 if (dp->mp_flags & P_KEEP)
1536 /* Can't spill twice, make sure it's not already in a parent's
1539 if (txn->mt_parent) {
1541 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1542 if (tx2->mt_spill_pgs) {
1543 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1544 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1545 dp->mp_flags |= P_KEEP;
1553 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1557 mdb_midl_sort(txn->mt_spill_pgs);
1559 /* Flush the spilled part of dirty list */
1560 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1563 /* Reset any dirty pages we kept that page_flush didn't see */
1564 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1567 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1571 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1573 mdb_find_oldest(MDB_txn *txn)
1576 txnid_t mr, oldest = txn->mt_txnid - 1;
1577 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1578 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1588 /** Add a page to the txn's dirty list */
1590 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1593 int (*insert)(MDB_ID2L, MDB_ID2 *);
1595 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1596 insert = mdb_mid2l_append;
1598 insert = mdb_mid2l_insert;
1600 mid.mid = mp->mp_pgno;
1602 insert(txn->mt_u.dirty_list, &mid);
1603 txn->mt_dirty_room--;
1606 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1607 * me_pghead and mt_next_pgno.
1609 * If there are free pages available from older transactions, they
1610 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1611 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1612 * and move me_pglast to say which records were consumed. Only this
1613 * function can create me_pghead and move me_pglast/mt_next_pgno.
1614 * @param[in] mc cursor A cursor handle identifying the transaction and
1615 * database for which we are allocating.
1616 * @param[in] num the number of pages to allocate.
1617 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1618 * will always be satisfied by a single contiguous chunk of memory.
1619 * @return 0 on success, non-zero on failure.
1622 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1624 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1625 /* Get at most <Max_retries> more freeDB records once me_pghead
1626 * has enough pages. If not enough, use new pages from the map.
1627 * If <Paranoid> and mc is updating the freeDB, only get new
1628 * records if me_pghead is empty. Then the freelist cannot play
1629 * catch-up with itself by growing while trying to save it.
1631 enum { Paranoid = 1, Max_retries = 500 };
1633 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1635 int rc, n2 = num-1, retry = Max_retries;
1636 MDB_txn *txn = mc->mc_txn;
1637 MDB_env *env = txn->mt_env;
1638 pgno_t pgno, *mop = env->me_pghead;
1639 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1641 txnid_t oldest = 0, last;
1647 /* If our dirty list is already full, we can't do anything */
1648 if (txn->mt_dirty_room == 0)
1649 return MDB_TXN_FULL;
1651 for (op = MDB_FIRST;; op = MDB_NEXT) {
1654 pgno_t *idl, old_id, new_id;
1656 /* Seek a big enough contiguous page range. Prefer
1657 * pages at the tail, just truncating the list.
1659 if (mop_len >= (unsigned)num) {
1663 if (mop[i-n2] == pgno+n2)
1665 } while (--i >= (unsigned)num);
1666 if (Max_retries < INT_MAX && --retry < 0)
1670 if (op == MDB_FIRST) { /* 1st iteration */
1671 /* Prepare to fetch more and coalesce */
1672 oldest = mdb_find_oldest(txn);
1673 last = env->me_pglast;
1674 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1677 key.mv_data = &last; /* will look up last+1 */
1678 key.mv_size = sizeof(last);
1680 if (Paranoid && mc->mc_dbi == FREE_DBI)
1683 if (Paranoid && retry < 0 && mop_len)
1687 /* Do not fetch more if the record will be too recent */
1690 rc = mdb_cursor_get(&m2, &key, NULL, op);
1692 if (rc == MDB_NOTFOUND)
1696 last = *(txnid_t*)key.mv_data;
1699 np = m2.mc_pg[m2.mc_top];
1700 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1701 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1704 idl = (MDB_ID *) data.mv_data;
1707 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1710 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1712 mop = env->me_pghead;
1714 env->me_pglast = last;
1716 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1717 last, txn->mt_dbs[FREE_DBI].md_root, i));
1719 DPRINTF(("IDL %"Z"u", idl[k]));
1721 /* Merge in descending sorted order */
1724 mop[0] = (pgno_t)-1;
1728 for (; old_id < new_id; old_id = mop[--j])
1735 /* Use new pages from the map when nothing suitable in the freeDB */
1737 pgno = txn->mt_next_pgno;
1738 if (pgno + num >= env->me_maxpg) {
1739 DPUTS("DB size maxed out");
1740 return MDB_MAP_FULL;
1744 if (env->me_flags & MDB_WRITEMAP) {
1745 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1747 if (!(np = mdb_page_malloc(txn, num)))
1751 mop[0] = mop_len -= num;
1752 /* Move any stragglers down */
1753 for (j = i-num; j < mop_len; )
1754 mop[++j] = mop[++i];
1756 txn->mt_next_pgno = pgno + num;
1759 mdb_page_dirty(txn, np);
1765 /** Copy the used portions of a non-overflow page.
1766 * @param[in] dst page to copy into
1767 * @param[in] src page to copy from
1768 * @param[in] psize size of a page
1771 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1773 enum { Align = sizeof(pgno_t) };
1774 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1776 /* If page isn't full, just copy the used portion. Adjust
1777 * alignment so memcpy may copy words instead of bytes.
1779 if ((unused &= -Align) && !IS_LEAF2(src)) {
1781 memcpy(dst, src, (lower + (Align-1)) & -Align);
1782 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1785 memcpy(dst, src, psize - unused);
1789 /** Pull a page off the txn's spill list, if present.
1790 * If a page being referenced was spilled to disk in this txn, bring
1791 * it back and make it dirty/writable again.
1792 * @param[in] txn the transaction handle.
1793 * @param[in] mp the page being referenced.
1794 * @param[out] ret the writable page, if any. ret is unchanged if
1795 * mp wasn't spilled.
1798 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1800 MDB_env *env = txn->mt_env;
1803 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1805 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1806 if (!tx2->mt_spill_pgs)
1808 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1809 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1812 if (txn->mt_dirty_room == 0)
1813 return MDB_TXN_FULL;
1814 if (IS_OVERFLOW(mp))
1818 if (env->me_flags & MDB_WRITEMAP) {
1821 np = mdb_page_malloc(txn, num);
1825 memcpy(np, mp, num * env->me_psize);
1827 mdb_page_copy(np, mp, env->me_psize);
1830 /* If in current txn, this page is no longer spilled.
1831 * If it happens to be the last page, truncate the spill list.
1832 * Otherwise mark it as deleted by setting the LSB.
1834 if (x == txn->mt_spill_pgs[0])
1835 txn->mt_spill_pgs[0]--;
1837 txn->mt_spill_pgs[x] |= 1;
1838 } /* otherwise, if belonging to a parent txn, the
1839 * page remains spilled until child commits
1842 mdb_page_dirty(txn, np);
1843 np->mp_flags |= P_DIRTY;
1851 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1852 * @param[in] mc cursor pointing to the page to be touched
1853 * @return 0 on success, non-zero on failure.
1856 mdb_page_touch(MDB_cursor *mc)
1858 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1859 MDB_txn *txn = mc->mc_txn;
1860 MDB_cursor *m2, *m3;
1865 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1866 if (txn->mt_flags & MDB_TXN_SPILLS) {
1868 rc = mdb_page_unspill(txn, mp, &np);
1874 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1875 (rc = mdb_page_alloc(mc, 1, &np)))
1878 DPRINTF(("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno));
1879 assert(mp->mp_pgno != pgno);
1880 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1881 /* Update the parent page, if any, to point to the new page */
1883 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1884 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1885 SETPGNO(node, pgno);
1887 mc->mc_db->md_root = pgno;
1889 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1890 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1892 /* If txn has a parent, make sure the page is in our
1896 unsigned x = mdb_mid2l_search(dl, pgno);
1897 if (x <= dl[0].mid && dl[x].mid == pgno) {
1898 if (mp != dl[x].mptr) { /* bad cursor? */
1899 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1900 return MDB_CORRUPTED;
1905 assert(dl[0].mid < MDB_IDL_UM_MAX);
1907 np = mdb_page_malloc(txn, 1);
1912 mdb_mid2l_insert(dl, &mid);
1917 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1919 np->mp_flags |= P_DIRTY;
1922 /* Adjust cursors pointing to mp */
1923 mc->mc_pg[mc->mc_top] = np;
1925 if (mc->mc_flags & C_SUB) {
1927 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1928 m3 = &m2->mc_xcursor->mx_cursor;
1929 if (m3->mc_snum < mc->mc_snum) continue;
1930 if (m3->mc_pg[mc->mc_top] == mp)
1931 m3->mc_pg[mc->mc_top] = np;
1934 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1935 if (m2->mc_snum < mc->mc_snum) continue;
1936 if (m2->mc_pg[mc->mc_top] == mp) {
1937 m2->mc_pg[mc->mc_top] = np;
1938 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1939 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1941 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1942 if (!(leaf->mn_flags & F_SUBDATA))
1943 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1952 mdb_env_sync(MDB_env *env, int force)
1955 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1956 if (env->me_flags & MDB_WRITEMAP) {
1957 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1958 ? MS_ASYNC : MS_SYNC;
1959 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1962 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1966 if (MDB_FDATASYNC(env->me_fd))
1973 /** Back up parent txn's cursors, then grab the originals for tracking */
1975 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1977 MDB_cursor *mc, *bk;
1982 for (i = src->mt_numdbs; --i >= 0; ) {
1983 if ((mc = src->mt_cursors[i]) != NULL) {
1984 size = sizeof(MDB_cursor);
1986 size += sizeof(MDB_xcursor);
1987 for (; mc; mc = bk->mc_next) {
1993 mc->mc_db = &dst->mt_dbs[i];
1994 /* Kill pointers into src - and dst to reduce abuse: The
1995 * user may not use mc until dst ends. Otherwise we'd...
1997 mc->mc_txn = NULL; /* ...set this to dst */
1998 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
1999 if ((mx = mc->mc_xcursor) != NULL) {
2000 *(MDB_xcursor *)(bk+1) = *mx;
2001 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2003 mc->mc_next = dst->mt_cursors[i];
2004 dst->mt_cursors[i] = mc;
2011 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2012 * @param[in] txn the transaction handle.
2013 * @param[in] merge true to keep changes to parent cursors, false to revert.
2014 * @return 0 on success, non-zero on failure.
2017 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2019 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2023 for (i = txn->mt_numdbs; --i >= 0; ) {
2024 for (mc = cursors[i]; mc; mc = next) {
2026 if ((bk = mc->mc_backup) != NULL) {
2028 /* Commit changes to parent txn */
2029 mc->mc_next = bk->mc_next;
2030 mc->mc_backup = bk->mc_backup;
2031 mc->mc_txn = bk->mc_txn;
2032 mc->mc_db = bk->mc_db;
2033 mc->mc_dbflag = bk->mc_dbflag;
2034 if ((mx = mc->mc_xcursor) != NULL)
2035 mx->mx_cursor.mc_txn = bk->mc_txn;
2037 /* Abort nested txn */
2039 if ((mx = mc->mc_xcursor) != NULL)
2040 *mx = *(MDB_xcursor *)(bk+1);
2044 /* Only malloced cursors are permanently tracked. */
2052 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2055 mdb_txn_reset0(MDB_txn *txn, const char *act);
2057 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2063 Pidset = F_SETLK, Pidcheck = F_GETLK
2067 /** Set or check a pid lock. Set returns 0 on success.
2068 * Check returns 0 if the process is certainly dead, nonzero if it may
2069 * be alive (the lock exists or an error happened so we do not know).
2071 * On Windows Pidset is a no-op, we merely check for the existence
2072 * of the process with the given pid. On POSIX we use a single byte
2073 * lock on the lockfile, set at an offset equal to the pid.
2076 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2078 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2081 if (op == Pidcheck) {
2082 h = OpenProcess(env->me_pidquery, FALSE, pid);
2083 /* No documented "no such process" code, but other program use this: */
2085 return ErrCode() != ERROR_INVALID_PARAMETER;
2086 /* A process exists until all handles to it close. Has it exited? */
2087 ret = WaitForSingleObject(h, 0) != 0;
2094 struct flock lock_info;
2095 memset(&lock_info, 0, sizeof(lock_info));
2096 lock_info.l_type = F_WRLCK;
2097 lock_info.l_whence = SEEK_SET;
2098 lock_info.l_start = pid;
2099 lock_info.l_len = 1;
2100 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2101 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2103 } else if ((rc = ErrCode()) == EINTR) {
2111 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2112 * @param[in] txn the transaction handle to initialize
2113 * @return 0 on success, non-zero on failure.
2116 mdb_txn_renew0(MDB_txn *txn)
2118 MDB_env *env = txn->mt_env;
2121 int rc, new_notls = 0;
2124 txn->mt_numdbs = env->me_numdbs;
2125 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2127 if (txn->mt_flags & MDB_TXN_RDONLY) {
2128 if (!env->me_txns) {
2129 i = mdb_env_pick_meta(env);
2130 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2131 txn->mt_u.reader = NULL;
2133 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2134 pthread_getspecific(env->me_txkey);
2136 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2137 return MDB_BAD_RSLOT;
2139 pid_t pid = env->me_pid;
2140 pthread_t tid = pthread_self();
2142 if (!(env->me_flags & MDB_LIVE_READER)) {
2143 rc = mdb_reader_pid(env, Pidset, pid);
2145 UNLOCK_MUTEX_R(env);
2148 env->me_flags |= MDB_LIVE_READER;
2152 for (i=0; i<env->me_txns->mti_numreaders; i++)
2153 if (env->me_txns->mti_readers[i].mr_pid == 0)
2155 if (i == env->me_maxreaders) {
2156 UNLOCK_MUTEX_R(env);
2157 return MDB_READERS_FULL;
2159 env->me_txns->mti_readers[i].mr_pid = pid;
2160 env->me_txns->mti_readers[i].mr_tid = tid;
2161 if (i >= env->me_txns->mti_numreaders)
2162 env->me_txns->mti_numreaders = i+1;
2163 /* Save numreaders for un-mutexed mdb_env_close() */
2164 env->me_numreaders = env->me_txns->mti_numreaders;
2165 UNLOCK_MUTEX_R(env);
2166 r = &env->me_txns->mti_readers[i];
2167 new_notls = (env->me_flags & MDB_NOTLS);
2168 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2173 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2174 txn->mt_u.reader = r;
2176 txn->mt_toggle = txn->mt_txnid & 1;
2180 txn->mt_txnid = env->me_txns->mti_txnid;
2181 txn->mt_toggle = txn->mt_txnid & 1;
2184 if (txn->mt_txnid == mdb_debug_start)
2187 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2188 txn->mt_u.dirty_list = env->me_dirty_list;
2189 txn->mt_u.dirty_list[0].mid = 0;
2190 txn->mt_free_pgs = env->me_free_pgs;
2191 txn->mt_free_pgs[0] = 0;
2192 txn->mt_spill_pgs = NULL;
2196 /* Copy the DB info and flags */
2197 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2199 /* Moved to here to avoid a data race in read TXNs */
2200 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2202 for (i=2; i<txn->mt_numdbs; i++) {
2203 x = env->me_dbflags[i];
2204 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2205 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2207 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2209 if (env->me_maxpg < txn->mt_next_pgno) {
2210 mdb_txn_reset0(txn, "renew0-mapfail");
2212 txn->mt_u.reader->mr_pid = 0;
2213 txn->mt_u.reader = NULL;
2215 return MDB_MAP_RESIZED;
2222 mdb_txn_renew(MDB_txn *txn)
2226 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2229 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2230 DPUTS("environment had fatal error, must shutdown!");
2234 rc = mdb_txn_renew0(txn);
2235 if (rc == MDB_SUCCESS) {
2236 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2237 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2238 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2244 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2248 int rc, size, tsize = sizeof(MDB_txn);
2250 if (env->me_flags & MDB_FATAL_ERROR) {
2251 DPUTS("environment had fatal error, must shutdown!");
2254 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2257 /* Nested transactions: Max 1 child, write txns only, no writemap */
2258 if (parent->mt_child ||
2259 (flags & MDB_RDONLY) ||
2260 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2261 (env->me_flags & MDB_WRITEMAP))
2263 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2265 tsize = sizeof(MDB_ntxn);
2267 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2268 if (!(flags & MDB_RDONLY))
2269 size += env->me_maxdbs * sizeof(MDB_cursor *);
2271 if ((txn = calloc(1, size)) == NULL) {
2272 DPRINTF(("calloc: %s", strerror(ErrCode())));
2275 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2276 if (flags & MDB_RDONLY) {
2277 txn->mt_flags |= MDB_TXN_RDONLY;
2278 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2280 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2281 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2287 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2288 if (!txn->mt_u.dirty_list ||
2289 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2291 free(txn->mt_u.dirty_list);
2295 txn->mt_txnid = parent->mt_txnid;
2296 txn->mt_toggle = parent->mt_toggle;
2297 txn->mt_dirty_room = parent->mt_dirty_room;
2298 txn->mt_u.dirty_list[0].mid = 0;
2299 txn->mt_spill_pgs = NULL;
2300 txn->mt_next_pgno = parent->mt_next_pgno;
2301 parent->mt_child = txn;
2302 txn->mt_parent = parent;
2303 txn->mt_numdbs = parent->mt_numdbs;
2304 txn->mt_flags = parent->mt_flags;
2305 txn->mt_dbxs = parent->mt_dbxs;
2306 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2307 /* Copy parent's mt_dbflags, but clear DB_NEW */
2308 for (i=0; i<txn->mt_numdbs; i++)
2309 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2311 ntxn = (MDB_ntxn *)txn;
2312 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2313 if (env->me_pghead) {
2314 size = MDB_IDL_SIZEOF(env->me_pghead);
2315 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2317 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2322 rc = mdb_cursor_shadow(parent, txn);
2324 mdb_txn_reset0(txn, "beginchild-fail");
2326 rc = mdb_txn_renew0(txn);
2332 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2333 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2334 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2341 mdb_txn_env(MDB_txn *txn)
2343 if(!txn) return NULL;
2347 /** Export or close DBI handles opened in this txn. */
2349 mdb_dbis_update(MDB_txn *txn, int keep)
2352 MDB_dbi n = txn->mt_numdbs;
2353 MDB_env *env = txn->mt_env;
2354 unsigned char *tdbflags = txn->mt_dbflags;
2356 for (i = n; --i >= 2;) {
2357 if (tdbflags[i] & DB_NEW) {
2359 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2361 char *ptr = env->me_dbxs[i].md_name.mv_data;
2362 env->me_dbxs[i].md_name.mv_data = NULL;
2363 env->me_dbxs[i].md_name.mv_size = 0;
2364 env->me_dbflags[i] = 0;
2369 if (keep && env->me_numdbs < n)
2373 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2374 * May be called twice for readonly txns: First reset it, then abort.
2375 * @param[in] txn the transaction handle to reset
2376 * @param[in] act why the transaction is being reset
2379 mdb_txn_reset0(MDB_txn *txn, const char *act)
2381 MDB_env *env = txn->mt_env;
2383 /* Close any DBI handles opened in this txn */
2384 mdb_dbis_update(txn, 0);
2386 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2387 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2388 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2390 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2391 if (txn->mt_u.reader) {
2392 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2393 if (!(env->me_flags & MDB_NOTLS))
2394 txn->mt_u.reader = NULL; /* txn does not own reader */
2396 txn->mt_numdbs = 0; /* close nothing if called again */
2397 txn->mt_dbxs = NULL; /* mark txn as reset */
2399 mdb_cursors_close(txn, 0);
2401 if (!(env->me_flags & MDB_WRITEMAP)) {
2402 mdb_dlist_free(txn);
2404 mdb_midl_free(env->me_pghead);
2406 if (txn->mt_parent) {
2407 txn->mt_parent->mt_child = NULL;
2408 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2409 mdb_midl_free(txn->mt_free_pgs);
2410 mdb_midl_free(txn->mt_spill_pgs);
2411 free(txn->mt_u.dirty_list);
2415 if (mdb_midl_shrink(&txn->mt_free_pgs))
2416 env->me_free_pgs = txn->mt_free_pgs;
2417 env->me_pghead = NULL;
2421 /* The writer mutex was locked in mdb_txn_begin. */
2422 UNLOCK_MUTEX_W(env);
2427 mdb_txn_reset(MDB_txn *txn)
2432 /* This call is only valid for read-only txns */
2433 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2436 mdb_txn_reset0(txn, "reset");
2440 mdb_txn_abort(MDB_txn *txn)
2446 mdb_txn_abort(txn->mt_child);
2448 mdb_txn_reset0(txn, "abort");
2449 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2450 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2451 txn->mt_u.reader->mr_pid = 0;
2456 /** Save the freelist as of this transaction to the freeDB.
2457 * This changes the freelist. Keep trying until it stabilizes.
2460 mdb_freelist_save(MDB_txn *txn)
2462 /* env->me_pghead[] can grow and shrink during this call.
2463 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2464 * Page numbers cannot disappear from txn->mt_free_pgs[].
2467 MDB_env *env = txn->mt_env;
2468 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2469 txnid_t pglast = 0, head_id = 0;
2470 pgno_t freecnt = 0, *free_pgs, *mop;
2471 ssize_t head_room = 0, total_room = 0, mop_len;
2473 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2475 if (env->me_pghead) {
2476 /* Make sure first page of freeDB is touched and on freelist */
2477 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2478 if (rc && rc != MDB_NOTFOUND)
2483 /* Come back here after each Put() in case freelist changed */
2486 /* If using records from freeDB which we have not yet
2487 * deleted, delete them and any we reserved for me_pghead.
2489 while (pglast < env->me_pglast) {
2490 rc = mdb_cursor_first(&mc, &key, NULL);
2493 pglast = head_id = *(txnid_t *)key.mv_data;
2494 total_room = head_room = 0;
2495 assert(pglast <= env->me_pglast);
2496 rc = mdb_cursor_del(&mc, 0);
2501 /* Save the IDL of pages freed by this txn, to a single record */
2502 if (freecnt < txn->mt_free_pgs[0]) {
2504 /* Make sure last page of freeDB is touched and on freelist */
2505 key.mv_size = MDB_MAXKEYSIZE+1;
2507 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2508 if (rc && rc != MDB_NOTFOUND)
2511 free_pgs = txn->mt_free_pgs;
2512 /* Write to last page of freeDB */
2513 key.mv_size = sizeof(txn->mt_txnid);
2514 key.mv_data = &txn->mt_txnid;
2516 freecnt = free_pgs[0];
2517 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2518 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2521 /* Retry if mt_free_pgs[] grew during the Put() */
2522 free_pgs = txn->mt_free_pgs;
2523 } while (freecnt < free_pgs[0]);
2524 mdb_midl_sort(free_pgs);
2525 memcpy(data.mv_data, free_pgs, data.mv_size);
2528 unsigned int i = free_pgs[0];
2529 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2530 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2532 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2538 mop = env->me_pghead;
2539 mop_len = mop ? mop[0] : 0;
2541 /* Reserve records for me_pghead[]. Split it if multi-page,
2542 * to avoid searching freeDB for a page range. Use keys in
2543 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2545 if (total_room >= mop_len) {
2546 if (total_room == mop_len || --more < 0)
2548 } else if (head_room >= maxfree_1pg && head_id > 1) {
2549 /* Keep current record (overflow page), add a new one */
2553 /* (Re)write {key = head_id, IDL length = head_room} */
2554 total_room -= head_room;
2555 head_room = mop_len - total_room;
2556 if (head_room > maxfree_1pg && head_id > 1) {
2557 /* Overflow multi-page for part of me_pghead */
2558 head_room /= head_id; /* amortize page sizes */
2559 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2560 } else if (head_room < 0) {
2561 /* Rare case, not bothering to delete this record */
2564 key.mv_size = sizeof(head_id);
2565 key.mv_data = &head_id;
2566 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2567 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2570 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2571 total_room += head_room;
2574 /* Fill in the reserved me_pghead records */
2580 rc = mdb_cursor_first(&mc, &key, &data);
2581 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2582 unsigned flags = MDB_CURRENT;
2583 txnid_t id = *(txnid_t *)key.mv_data;
2584 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2587 assert(len >= 0 && id <= env->me_pglast);
2589 if (len > mop_len) {
2591 data.mv_size = (len + 1) * sizeof(MDB_ID);
2594 data.mv_data = mop -= len;
2597 rc = mdb_cursor_put(&mc, &key, &data, flags);
2599 if (rc || !(mop_len -= len))
2606 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2607 * @param[in] txn the transaction that's being committed
2608 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2609 * @return 0 on success, non-zero on failure.
2612 mdb_page_flush(MDB_txn *txn, int keep)
2614 MDB_env *env = txn->mt_env;
2615 MDB_ID2L dl = txn->mt_u.dirty_list;
2616 unsigned psize = env->me_psize, j;
2617 int i, pagecount = dl[0].mid, rc;
2618 size_t size = 0, pos = 0;
2620 MDB_page *dp = NULL;
2624 struct iovec iov[MDB_COMMIT_PAGES];
2625 ssize_t wpos = 0, wsize = 0, wres;
2626 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2632 if (env->me_flags & MDB_WRITEMAP) {
2633 /* Clear dirty flags */
2634 while (++i <= pagecount) {
2636 /* Don't flush this page yet */
2637 if (dp->mp_flags & P_KEEP) {
2638 dp->mp_flags ^= P_KEEP;
2642 dp->mp_flags &= ~P_DIRTY;
2647 /* Write the pages */
2649 if (++i <= pagecount) {
2651 /* Don't flush this page yet */
2652 if (dp->mp_flags & P_KEEP) {
2653 dp->mp_flags ^= P_KEEP;
2658 /* clear dirty flag */
2659 dp->mp_flags &= ~P_DIRTY;
2662 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2667 /* Windows actually supports scatter/gather I/O, but only on
2668 * unbuffered file handles. Since we're relying on the OS page
2669 * cache for all our data, that's self-defeating. So we just
2670 * write pages one at a time. We use the ov structure to set
2671 * the write offset, to at least save the overhead of a Seek
2674 DPRINTF(("committing page %"Z"u", pgno));
2675 memset(&ov, 0, sizeof(ov));
2676 ov.Offset = pos & 0xffffffff;
2677 ov.OffsetHigh = pos >> 16 >> 16;
2678 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2680 DPRINTF(("WriteFile: %d", rc));
2684 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2685 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2687 /* Write previous page(s) */
2688 #ifdef MDB_USE_PWRITEV
2689 wres = pwritev(env->me_fd, iov, n, wpos);
2692 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2694 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2696 DPRINTF(("lseek: %s", strerror(rc)));
2699 wres = writev(env->me_fd, iov, n);
2702 if (wres != wsize) {
2705 DPRINTF(("Write error: %s", strerror(rc)));
2707 rc = EIO; /* TODO: Use which error code? */
2708 DPUTS("short write, filesystem full?");
2719 DPRINTF(("committing page %"Z"u", pgno));
2720 next_pos = pos + size;
2721 iov[n].iov_len = size;
2722 iov[n].iov_base = (char *)dp;
2728 for (i = keep; ++i <= pagecount; ) {
2730 /* This is a page we skipped above */
2733 dl[j].mid = dp->mp_pgno;
2736 mdb_dpage_free(env, dp);
2741 txn->mt_dirty_room += i - j;
2747 mdb_txn_commit(MDB_txn *txn)
2753 assert(txn != NULL);
2754 assert(txn->mt_env != NULL);
2756 if (txn->mt_child) {
2757 rc = mdb_txn_commit(txn->mt_child);
2758 txn->mt_child = NULL;
2765 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2766 mdb_dbis_update(txn, 1);
2767 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2772 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2773 DPUTS("error flag is set, can't commit");
2775 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2780 if (txn->mt_parent) {
2781 MDB_txn *parent = txn->mt_parent;
2784 unsigned x, y, len, ps_len;
2786 /* Append our free list to parent's */
2787 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2790 mdb_midl_free(txn->mt_free_pgs);
2791 /* Failures after this must either undo the changes
2792 * to the parent or set MDB_TXN_ERROR in the parent.
2795 parent->mt_next_pgno = txn->mt_next_pgno;
2796 parent->mt_flags = txn->mt_flags;
2798 /* Merge our cursors into parent's and close them */
2799 mdb_cursors_close(txn, 1);
2801 /* Update parent's DB table. */
2802 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2803 parent->mt_numdbs = txn->mt_numdbs;
2804 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2805 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2806 for (i=2; i<txn->mt_numdbs; i++) {
2807 /* preserve parent's DB_NEW status */
2808 x = parent->mt_dbflags[i] & DB_NEW;
2809 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2812 dst = parent->mt_u.dirty_list;
2813 src = txn->mt_u.dirty_list;
2814 /* Remove anything in our dirty list from parent's spill list */
2815 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2817 pspill[0] = (pgno_t)-1;
2818 /* Mark our dirty pages as deleted in parent spill list */
2819 for (i=0, len=src[0].mid; ++i <= len; ) {
2820 MDB_ID pn = src[i].mid << 1;
2821 while (pn > pspill[x])
2823 if (pn == pspill[x]) {
2828 /* Squash deleted pagenums if we deleted any */
2829 for (x=y; ++x <= ps_len; )
2830 if (!(pspill[x] & 1))
2831 pspill[++y] = pspill[x];
2835 /* Find len = length of merging our dirty list with parent's */
2837 dst[0].mid = 0; /* simplify loops */
2838 if (parent->mt_parent) {
2839 len = x + src[0].mid;
2840 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2841 for (i = x; y && i; y--) {
2842 pgno_t yp = src[y].mid;
2843 while (yp < dst[i].mid)
2845 if (yp == dst[i].mid) {
2850 } else { /* Simplify the above for single-ancestor case */
2851 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2853 /* Merge our dirty list with parent's */
2855 for (i = len; y; dst[i--] = src[y--]) {
2856 pgno_t yp = src[y].mid;
2857 while (yp < dst[x].mid)
2858 dst[i--] = dst[x--];
2859 if (yp == dst[x].mid)
2860 free(dst[x--].mptr);
2864 free(txn->mt_u.dirty_list);
2865 parent->mt_dirty_room = txn->mt_dirty_room;
2866 if (txn->mt_spill_pgs) {
2867 if (parent->mt_spill_pgs) {
2868 /* TODO: Prevent failure here, so parent does not fail */
2869 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2871 parent->mt_flags |= MDB_TXN_ERROR;
2872 mdb_midl_free(txn->mt_spill_pgs);
2873 mdb_midl_sort(parent->mt_spill_pgs);
2875 parent->mt_spill_pgs = txn->mt_spill_pgs;
2879 parent->mt_child = NULL;
2880 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2885 if (txn != env->me_txn) {
2886 DPUTS("attempt to commit unknown transaction");
2891 mdb_cursors_close(txn, 0);
2893 if (!txn->mt_u.dirty_list[0].mid &&
2894 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2897 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2898 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2900 /* Update DB root pointers */
2901 if (txn->mt_numdbs > 2) {
2905 data.mv_size = sizeof(MDB_db);
2907 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2908 for (i = 2; i < txn->mt_numdbs; i++) {
2909 if (txn->mt_dbflags[i] & DB_DIRTY) {
2910 data.mv_data = &txn->mt_dbs[i];
2911 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2918 rc = mdb_freelist_save(txn);
2922 mdb_midl_free(env->me_pghead);
2923 env->me_pghead = NULL;
2924 if (mdb_midl_shrink(&txn->mt_free_pgs))
2925 env->me_free_pgs = txn->mt_free_pgs;
2931 if ((rc = mdb_page_flush(txn, 0)) ||
2932 (rc = mdb_env_sync(env, 0)) ||
2933 (rc = mdb_env_write_meta(txn)))
2939 mdb_dbis_update(txn, 1);
2941 UNLOCK_MUTEX_W(env);
2951 /** Read the environment parameters of a DB environment before
2952 * mapping it into memory.
2953 * @param[in] env the environment handle
2954 * @param[out] meta address of where to store the meta information
2955 * @return 0 on success, non-zero on failure.
2958 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2965 /* We don't know the page size yet, so use a minimum value.
2966 * Read both meta pages so we can use the latest one.
2969 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2973 memset(&ov, 0, sizeof(ov));
2975 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2976 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2979 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2981 if (rc != MDB_PAGESIZE) {
2982 if (rc == 0 && off == 0)
2984 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2985 DPRINTF(("read: %s", mdb_strerror(rc)));
2989 p = (MDB_page *)&pbuf;
2991 if (!F_ISSET(p->mp_flags, P_META)) {
2992 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
2997 if (m->mm_magic != MDB_MAGIC) {
2998 DPUTS("meta has invalid magic");
3002 if (m->mm_version != MDB_DATA_VERSION) {
3003 DPRINTF(("database is version %u, expected version %u",
3004 m->mm_version, MDB_DATA_VERSION));
3005 return MDB_VERSION_MISMATCH;
3008 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3014 /** Write the environment parameters of a freshly created DB environment.
3015 * @param[in] env the environment handle
3016 * @param[out] meta address of where to store the meta information
3017 * @return 0 on success, non-zero on failure.
3020 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3028 memset(&ov, 0, sizeof(ov));
3029 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3031 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3034 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3035 len = pwrite(fd, ptr, size, pos); \
3036 rc = (len >= 0); } while(0)
3039 DPUTS("writing new meta page");
3041 psize = env->me_psize;
3043 meta->mm_magic = MDB_MAGIC;
3044 meta->mm_version = MDB_DATA_VERSION;
3045 meta->mm_mapsize = env->me_mapsize;
3046 meta->mm_psize = psize;
3047 meta->mm_last_pg = 1;
3048 meta->mm_flags = env->me_flags & 0xffff;
3049 meta->mm_flags |= MDB_INTEGERKEY;
3050 meta->mm_dbs[0].md_root = P_INVALID;
3051 meta->mm_dbs[1].md_root = P_INVALID;
3053 p = calloc(2, psize);
3055 p->mp_flags = P_META;
3056 *(MDB_meta *)METADATA(p) = *meta;
3058 q = (MDB_page *)((char *)p + psize);
3060 q->mp_flags = P_META;
3061 *(MDB_meta *)METADATA(q) = *meta;
3063 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3066 else if ((unsigned) len == psize * 2)
3074 /** Update the environment info to commit a transaction.
3075 * @param[in] txn the transaction that's being committed
3076 * @return 0 on success, non-zero on failure.
3079 mdb_env_write_meta(MDB_txn *txn)
3082 MDB_meta meta, metab, *mp;
3084 int rc, len, toggle;
3093 assert(txn != NULL);
3094 assert(txn->mt_env != NULL);
3096 toggle = !txn->mt_toggle;
3097 DPRINTF(("writing meta page %d for root page %"Z"u",
3098 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3101 mp = env->me_metas[toggle];
3103 if (env->me_flags & MDB_WRITEMAP) {
3104 /* Persist any increases of mapsize config */
3105 if (env->me_mapsize > mp->mm_mapsize)
3106 mp->mm_mapsize = env->me_mapsize;
3107 mp->mm_dbs[0] = txn->mt_dbs[0];
3108 mp->mm_dbs[1] = txn->mt_dbs[1];
3109 mp->mm_last_pg = txn->mt_next_pgno - 1;
3110 mp->mm_txnid = txn->mt_txnid;
3111 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3112 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3115 ptr += env->me_psize;
3116 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3123 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3124 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3126 ptr = (char *)&meta;
3127 if (env->me_mapsize > mp->mm_mapsize) {
3128 /* Persist any increases of mapsize config */
3129 meta.mm_mapsize = env->me_mapsize;
3130 off = offsetof(MDB_meta, mm_mapsize);
3132 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3134 len = sizeof(MDB_meta) - off;
3137 meta.mm_dbs[0] = txn->mt_dbs[0];
3138 meta.mm_dbs[1] = txn->mt_dbs[1];
3139 meta.mm_last_pg = txn->mt_next_pgno - 1;
3140 meta.mm_txnid = txn->mt_txnid;
3143 off += env->me_psize;
3146 /* Write to the SYNC fd */
3147 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3148 env->me_fd : env->me_mfd;
3151 memset(&ov, 0, sizeof(ov));
3153 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3157 rc = pwrite(mfd, ptr, len, off);
3160 rc = rc < 0 ? ErrCode() : EIO;
3161 DPUTS("write failed, disk error?");
3162 /* On a failure, the pagecache still contains the new data.
3163 * Write some old data back, to prevent it from being used.
3164 * Use the non-SYNC fd; we know it will fail anyway.
3166 meta.mm_last_pg = metab.mm_last_pg;
3167 meta.mm_txnid = metab.mm_txnid;
3169 memset(&ov, 0, sizeof(ov));
3171 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3173 r2 = pwrite(env->me_fd, ptr, len, off);
3174 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3177 env->me_flags |= MDB_FATAL_ERROR;
3181 /* Memory ordering issues are irrelevant; since the entire writer
3182 * is wrapped by wmutex, all of these changes will become visible
3183 * after the wmutex is unlocked. Since the DB is multi-version,
3184 * readers will get consistent data regardless of how fresh or
3185 * how stale their view of these values is.
3187 env->me_txns->mti_txnid = txn->mt_txnid;
3192 /** Check both meta pages to see which one is newer.
3193 * @param[in] env the environment handle
3194 * @return meta toggle (0 or 1).
3197 mdb_env_pick_meta(const MDB_env *env)
3199 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3203 mdb_env_create(MDB_env **env)
3207 e = calloc(1, sizeof(MDB_env));
3211 e->me_maxreaders = DEFAULT_READERS;
3212 e->me_maxdbs = e->me_numdbs = 2;
3213 e->me_fd = INVALID_HANDLE_VALUE;
3214 e->me_lfd = INVALID_HANDLE_VALUE;
3215 e->me_mfd = INVALID_HANDLE_VALUE;
3216 #ifdef MDB_USE_POSIX_SEM
3217 e->me_rmutex = SEM_FAILED;
3218 e->me_wmutex = SEM_FAILED;
3220 e->me_pid = getpid();
3221 VGMEMP_CREATE(e,0,0);
3227 mdb_env_map(MDB_env *env, void *addr, int newsize)
3230 unsigned int flags = env->me_flags;
3234 LONG sizelo, sizehi;
3235 sizelo = env->me_mapsize & 0xffffffff;
3236 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3238 /* Windows won't create mappings for zero length files.
3239 * Just allocate the maxsize right now.
3242 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3243 || !SetEndOfFile(env->me_fd)
3244 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3247 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3248 PAGE_READWRITE : PAGE_READONLY,
3249 sizehi, sizelo, NULL);
3252 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3253 FILE_MAP_WRITE : FILE_MAP_READ,
3254 0, 0, env->me_mapsize, addr);
3255 rc = env->me_map ? 0 : ErrCode();
3260 int prot = PROT_READ;
3261 if (flags & MDB_WRITEMAP) {
3263 if (newsize && ftruncate(env->me_fd, env->me_mapsize) < 0)
3266 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3268 if (env->me_map == MAP_FAILED) {
3272 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3274 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3276 #ifdef POSIX_MADV_RANDOM
3277 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3278 #endif /* POSIX_MADV_RANDOM */
3279 #endif /* MADV_RANDOM */
3282 /* Can happen because the address argument to mmap() is just a
3283 * hint. mmap() can pick another, e.g. if the range is in use.
3284 * The MAP_FIXED flag would prevent that, but then mmap could
3285 * instead unmap existing pages to make room for the new map.
3287 if (addr && env->me_map != addr)
3288 return EBUSY; /* TODO: Make a new MDB_* error code? */
3290 p = (MDB_page *)env->me_map;
3291 env->me_metas[0] = METADATA(p);
3292 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3298 mdb_env_set_mapsize(MDB_env *env, size_t size)
3300 /* If env is already open, caller is responsible for making
3301 * sure there are no active txns.
3309 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3310 munmap(env->me_map, env->me_mapsize);
3311 env->me_mapsize = size;
3312 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3313 rc = mdb_env_map(env, old, 1);
3317 env->me_mapsize = size;
3319 env->me_maxpg = env->me_mapsize / env->me_psize;
3324 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3328 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3333 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3335 if (env->me_map || readers < 1)
3337 env->me_maxreaders = readers;
3342 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3344 if (!env || !readers)
3346 *readers = env->me_maxreaders;
3350 /** Further setup required for opening an MDB environment
3353 mdb_env_open2(MDB_env *env)
3355 unsigned int flags = env->me_flags;
3356 int i, newenv = 0, rc;
3360 /* See if we should use QueryLimited */
3362 if ((rc & 0xff) > 5)
3363 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3365 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3368 memset(&meta, 0, sizeof(meta));
3370 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3373 DPUTS("new mdbenv");
3375 GET_PAGESIZE(env->me_psize);
3377 env->me_psize = meta.mm_psize;
3380 /* Was a mapsize configured? */
3381 if (!env->me_mapsize) {
3382 /* If this is a new environment, take the default,
3383 * else use the size recorded in the existing env.
3385 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3386 } else if (env->me_mapsize < meta.mm_mapsize) {
3387 /* If the configured size is smaller, make sure it's
3388 * still big enough. Silently round up to minimum if not.
3390 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3391 if (env->me_mapsize < minsize)
3392 env->me_mapsize = minsize;
3395 rc = mdb_env_map(env, meta.mm_address, newenv);
3400 if (flags & MDB_FIXEDMAP)
3401 meta.mm_address = env->me_map;
3402 i = mdb_env_init_meta(env, &meta);
3403 if (i != MDB_SUCCESS) {
3407 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3408 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3410 env->me_maxpg = env->me_mapsize / env->me_psize;
3413 int toggle = mdb_env_pick_meta(env);
3414 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3416 DPRINTF(("opened database version %u, pagesize %u",
3417 env->me_metas[0]->mm_version, env->me_psize));
3418 DPRINTF(("using meta page %d", toggle));
3419 DPRINTF(("depth: %u", db->md_depth));
3420 DPRINTF(("entries: %"Z"u", db->md_entries));
3421 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3422 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3423 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3424 DPRINTF(("root: %"Z"u", db->md_root));
3432 /** Release a reader thread's slot in the reader lock table.
3433 * This function is called automatically when a thread exits.
3434 * @param[in] ptr This points to the slot in the reader lock table.
3437 mdb_env_reader_dest(void *ptr)
3439 MDB_reader *reader = ptr;
3445 /** Junk for arranging thread-specific callbacks on Windows. This is
3446 * necessarily platform and compiler-specific. Windows supports up
3447 * to 1088 keys. Let's assume nobody opens more than 64 environments
3448 * in a single process, for now. They can override this if needed.
3450 #ifndef MAX_TLS_KEYS
3451 #define MAX_TLS_KEYS 64
3453 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3454 static int mdb_tls_nkeys;
3456 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3460 case DLL_PROCESS_ATTACH: break;
3461 case DLL_THREAD_ATTACH: break;
3462 case DLL_THREAD_DETACH:
3463 for (i=0; i<mdb_tls_nkeys; i++) {
3464 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3465 mdb_env_reader_dest(r);
3468 case DLL_PROCESS_DETACH: break;
3473 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3475 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3479 /* Force some symbol references.
3480 * _tls_used forces the linker to create the TLS directory if not already done
3481 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3483 #pragma comment(linker, "/INCLUDE:_tls_used")
3484 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3485 #pragma const_seg(".CRT$XLB")
3486 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3487 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3490 #pragma comment(linker, "/INCLUDE:__tls_used")
3491 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3492 #pragma data_seg(".CRT$XLB")
3493 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3495 #endif /* WIN 32/64 */
3496 #endif /* !__GNUC__ */
3499 /** Downgrade the exclusive lock on the region back to shared */
3501 mdb_env_share_locks(MDB_env *env, int *excl)
3503 int rc = 0, toggle = mdb_env_pick_meta(env);
3505 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3510 /* First acquire a shared lock. The Unlock will
3511 * then release the existing exclusive lock.
3513 memset(&ov, 0, sizeof(ov));
3514 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3517 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3523 struct flock lock_info;
3524 /* The shared lock replaces the existing lock */
3525 memset((void *)&lock_info, 0, sizeof(lock_info));
3526 lock_info.l_type = F_RDLCK;
3527 lock_info.l_whence = SEEK_SET;
3528 lock_info.l_start = 0;
3529 lock_info.l_len = 1;
3530 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3531 (rc = ErrCode()) == EINTR) ;
3532 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3539 /** Try to get exlusive lock, otherwise shared.
3540 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3543 mdb_env_excl_lock(MDB_env *env, int *excl)
3547 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3551 memset(&ov, 0, sizeof(ov));
3552 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3559 struct flock lock_info;
3560 memset((void *)&lock_info, 0, sizeof(lock_info));
3561 lock_info.l_type = F_WRLCK;
3562 lock_info.l_whence = SEEK_SET;
3563 lock_info.l_start = 0;
3564 lock_info.l_len = 1;
3565 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3566 (rc = ErrCode()) == EINTR) ;
3570 # ifdef MDB_USE_POSIX_SEM
3571 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3574 lock_info.l_type = F_RDLCK;
3575 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3576 (rc = ErrCode()) == EINTR) ;
3584 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3586 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3588 * @(#) $Revision: 5.1 $
3589 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3590 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3592 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3596 * Please do not copyright this code. This code is in the public domain.
3598 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3599 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3600 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3601 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3602 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3603 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3604 * PERFORMANCE OF THIS SOFTWARE.
3607 * chongo <Landon Curt Noll> /\oo/\
3608 * http://www.isthe.com/chongo/
3610 * Share and Enjoy! :-)
3613 typedef unsigned long long mdb_hash_t;
3614 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3616 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3617 * @param[in] val value to hash
3618 * @param[in] hval initial value for hash
3619 * @return 64 bit hash
3621 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3622 * hval arg on the first call.
3625 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3627 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3628 unsigned char *end = s + val->mv_size;
3630 * FNV-1a hash each octet of the string
3633 /* xor the bottom with the current octet */
3634 hval ^= (mdb_hash_t)*s++;
3636 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3637 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3638 (hval << 7) + (hval << 8) + (hval << 40);
3640 /* return our new hash value */
3644 /** Hash the string and output the encoded hash.
3645 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3646 * very short name limits. We don't care about the encoding being reversible,
3647 * we just want to preserve as many bits of the input as possible in a
3648 * small printable string.
3649 * @param[in] str string to hash
3650 * @param[out] encbuf an array of 11 chars to hold the hash
3652 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3655 mdb_pack85(unsigned long l, char *out)
3659 for (i=0; i<5; i++) {
3660 *out++ = mdb_a85[l % 85];
3666 mdb_hash_enc(MDB_val *val, char *encbuf)
3668 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3670 mdb_pack85(h, encbuf);
3671 mdb_pack85(h>>32, encbuf+5);
3676 /** Open and/or initialize the lock region for the environment.
3677 * @param[in] env The MDB environment.
3678 * @param[in] lpath The pathname of the file used for the lock region.
3679 * @param[in] mode The Unix permissions for the file, if we create it.
3680 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3681 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3682 * @return 0 on success, non-zero on failure.
3685 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3688 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3690 # define MDB_ERRCODE_ROFS EROFS
3691 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3692 # define MDB_CLOEXEC O_CLOEXEC
3695 # define MDB_CLOEXEC 0
3702 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3703 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3704 FILE_ATTRIBUTE_NORMAL, NULL);
3706 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3708 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3710 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3715 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3716 /* Lose record locks when exec*() */
3717 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3718 fcntl(env->me_lfd, F_SETFD, fdflags);
3721 if (!(env->me_flags & MDB_NOTLS)) {
3722 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3725 env->me_flags |= MDB_ENV_TXKEY;
3727 /* Windows TLS callbacks need help finding their TLS info. */
3728 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3732 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3736 /* Try to get exclusive lock. If we succeed, then
3737 * nobody is using the lock region and we should initialize it.
3739 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3742 size = GetFileSize(env->me_lfd, NULL);
3744 size = lseek(env->me_lfd, 0, SEEK_END);
3745 if (size == -1) goto fail_errno;
3747 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3748 if (size < rsize && *excl > 0) {
3750 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3751 || !SetEndOfFile(env->me_lfd))
3754 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3758 size = rsize - sizeof(MDB_txninfo);
3759 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3764 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3766 if (!mh) goto fail_errno;
3767 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3769 if (!env->me_txns) goto fail_errno;
3771 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3773 if (m == MAP_FAILED) goto fail_errno;
3779 BY_HANDLE_FILE_INFORMATION stbuf;
3788 if (!mdb_sec_inited) {
3789 InitializeSecurityDescriptor(&mdb_null_sd,
3790 SECURITY_DESCRIPTOR_REVISION);
3791 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3792 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3793 mdb_all_sa.bInheritHandle = FALSE;
3794 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3797 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3798 idbuf.volume = stbuf.dwVolumeSerialNumber;
3799 idbuf.nhigh = stbuf.nFileIndexHigh;
3800 idbuf.nlow = stbuf.nFileIndexLow;
3801 val.mv_data = &idbuf;
3802 val.mv_size = sizeof(idbuf);
3803 mdb_hash_enc(&val, encbuf);
3804 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3805 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3806 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3807 if (!env->me_rmutex) goto fail_errno;
3808 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3809 if (!env->me_wmutex) goto fail_errno;
3810 #elif defined(MDB_USE_POSIX_SEM)
3819 #if defined(__NetBSD__)
3820 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3822 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3823 idbuf.dev = stbuf.st_dev;
3824 idbuf.ino = stbuf.st_ino;
3825 val.mv_data = &idbuf;
3826 val.mv_size = sizeof(idbuf);
3827 mdb_hash_enc(&val, encbuf);
3828 #ifdef MDB_SHORT_SEMNAMES
3829 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3831 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3832 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3833 /* Clean up after a previous run, if needed: Try to
3834 * remove both semaphores before doing anything else.
3836 sem_unlink(env->me_txns->mti_rmname);
3837 sem_unlink(env->me_txns->mti_wmname);
3838 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3839 O_CREAT|O_EXCL, mode, 1);
3840 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3841 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3842 O_CREAT|O_EXCL, mode, 1);
3843 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3844 #else /* MDB_USE_POSIX_SEM */
3845 pthread_mutexattr_t mattr;
3847 if ((rc = pthread_mutexattr_init(&mattr))
3848 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3849 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3850 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3852 pthread_mutexattr_destroy(&mattr);
3853 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3855 env->me_txns->mti_magic = MDB_MAGIC;
3856 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3857 env->me_txns->mti_txnid = 0;
3858 env->me_txns->mti_numreaders = 0;
3861 if (env->me_txns->mti_magic != MDB_MAGIC) {
3862 DPUTS("lock region has invalid magic");
3866 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3867 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3868 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3869 rc = MDB_VERSION_MISMATCH;
3873 if (rc && rc != EACCES && rc != EAGAIN) {
3877 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3878 if (!env->me_rmutex) goto fail_errno;
3879 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3880 if (!env->me_wmutex) goto fail_errno;
3881 #elif defined(MDB_USE_POSIX_SEM)
3882 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3883 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3884 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3885 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3896 /** The name of the lock file in the DB environment */
3897 #define LOCKNAME "/lock.mdb"
3898 /** The name of the data file in the DB environment */
3899 #define DATANAME "/data.mdb"
3900 /** The suffix of the lock file when no subdir is used */
3901 #define LOCKSUFF "-lock"
3902 /** Only a subset of the @ref mdb_env flags can be changed
3903 * at runtime. Changing other flags requires closing the
3904 * environment and re-opening it with the new flags.
3906 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3907 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3910 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3912 int oflags, rc, len, excl = -1;
3913 char *lpath, *dpath;
3915 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3919 if (flags & MDB_NOSUBDIR) {
3920 rc = len + sizeof(LOCKSUFF) + len + 1;
3922 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3927 if (flags & MDB_NOSUBDIR) {
3928 dpath = lpath + len + sizeof(LOCKSUFF);
3929 sprintf(lpath, "%s" LOCKSUFF, path);
3930 strcpy(dpath, path);
3932 dpath = lpath + len + sizeof(LOCKNAME);
3933 sprintf(lpath, "%s" LOCKNAME, path);
3934 sprintf(dpath, "%s" DATANAME, path);
3938 flags |= env->me_flags;
3939 if (flags & MDB_RDONLY) {
3940 /* silently ignore WRITEMAP when we're only getting read access */
3941 flags &= ~MDB_WRITEMAP;
3943 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3944 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3947 env->me_flags = flags |= MDB_ENV_ACTIVE;
3951 env->me_path = strdup(path);
3952 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3953 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3954 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3959 /* For RDONLY, get lockfile after we know datafile exists */
3960 if (!F_ISSET(flags, MDB_RDONLY)) {
3961 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3967 if (F_ISSET(flags, MDB_RDONLY)) {
3968 oflags = GENERIC_READ;
3969 len = OPEN_EXISTING;
3971 oflags = GENERIC_READ|GENERIC_WRITE;
3974 mode = FILE_ATTRIBUTE_NORMAL;
3975 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3976 NULL, len, mode, NULL);
3978 if (F_ISSET(flags, MDB_RDONLY))
3981 oflags = O_RDWR | O_CREAT;
3983 env->me_fd = open(dpath, oflags, mode);
3985 if (env->me_fd == INVALID_HANDLE_VALUE) {
3990 if (F_ISSET(flags, MDB_RDONLY)) {
3991 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3996 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3997 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3998 env->me_mfd = env->me_fd;
4000 /* Synchronous fd for meta writes. Needed even with
4001 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4004 len = OPEN_EXISTING;
4005 env->me_mfd = CreateFile(dpath, oflags,
4006 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4007 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4010 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4012 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4017 DPRINTF(("opened dbenv %p", (void *) env));
4019 rc = mdb_env_share_locks(env, &excl);
4025 mdb_env_close0(env, excl);
4031 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4033 mdb_env_close0(MDB_env *env, int excl)
4037 if (!(env->me_flags & MDB_ENV_ACTIVE))
4040 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4041 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4042 free(env->me_dbxs[i].md_name.mv_data);
4044 free(env->me_dbflags);
4047 free(env->me_dirty_list);
4048 mdb_midl_free(env->me_free_pgs);
4050 if (env->me_flags & MDB_ENV_TXKEY) {
4051 pthread_key_delete(env->me_txkey);
4053 /* Delete our key from the global list */
4054 for (i=0; i<mdb_tls_nkeys; i++)
4055 if (mdb_tls_keys[i] == env->me_txkey) {
4056 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4064 munmap(env->me_map, env->me_mapsize);
4066 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4067 (void) close(env->me_mfd);
4068 if (env->me_fd != INVALID_HANDLE_VALUE)
4069 (void) close(env->me_fd);
4071 pid_t pid = env->me_pid;
4072 /* Clearing readers is done in this function because
4073 * me_txkey with its destructor must be disabled first.
4075 for (i = env->me_numreaders; --i >= 0; )
4076 if (env->me_txns->mti_readers[i].mr_pid == pid)
4077 env->me_txns->mti_readers[i].mr_pid = 0;
4079 if (env->me_rmutex) {
4080 CloseHandle(env->me_rmutex);
4081 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4083 /* Windows automatically destroys the mutexes when
4084 * the last handle closes.
4086 #elif defined(MDB_USE_POSIX_SEM)
4087 if (env->me_rmutex != SEM_FAILED) {
4088 sem_close(env->me_rmutex);
4089 if (env->me_wmutex != SEM_FAILED)
4090 sem_close(env->me_wmutex);
4091 /* If we have the filelock: If we are the
4092 * only remaining user, clean up semaphores.
4095 mdb_env_excl_lock(env, &excl);
4097 sem_unlink(env->me_txns->mti_rmname);
4098 sem_unlink(env->me_txns->mti_wmname);
4102 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4104 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4107 /* Unlock the lockfile. Windows would have unlocked it
4108 * after closing anyway, but not necessarily at once.
4110 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4113 (void) close(env->me_lfd);
4116 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4120 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4122 MDB_txn *txn = NULL;
4128 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4132 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4135 /* Do the lock/unlock of the reader mutex before starting the
4136 * write txn. Otherwise other read txns could block writers.
4138 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4143 /* We must start the actual read txn after blocking writers */
4144 mdb_txn_reset0(txn, "reset-stage1");
4146 /* Temporarily block writers until we snapshot the meta pages */
4149 rc = mdb_txn_renew0(txn);
4151 UNLOCK_MUTEX_W(env);
4156 wsize = env->me_psize * 2;
4160 DO_WRITE(rc, fd, ptr, w2, len);
4164 } else if (len > 0) {
4170 /* Non-blocking or async handles are not supported */
4176 UNLOCK_MUTEX_W(env);
4181 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4183 if (wsize > MAX_WRITE)
4187 DO_WRITE(rc, fd, ptr, w2, len);
4191 } else if (len > 0) {
4208 mdb_env_copy(MDB_env *env, const char *path)
4212 HANDLE newfd = INVALID_HANDLE_VALUE;
4214 if (env->me_flags & MDB_NOSUBDIR) {
4215 lpath = (char *)path;
4218 len += sizeof(DATANAME);
4219 lpath = malloc(len);
4222 sprintf(lpath, "%s" DATANAME, path);
4225 /* The destination path must exist, but the destination file must not.
4226 * We don't want the OS to cache the writes, since the source data is
4227 * already in the OS cache.
4230 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4231 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4233 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4235 if (newfd == INVALID_HANDLE_VALUE) {
4241 /* Set O_DIRECT if the file system supports it */
4242 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4243 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4245 #ifdef F_NOCACHE /* __APPLE__ */
4246 rc = fcntl(newfd, F_NOCACHE, 1);
4253 rc = mdb_env_copyfd(env, newfd);
4256 if (!(env->me_flags & MDB_NOSUBDIR))
4258 if (newfd != INVALID_HANDLE_VALUE)
4259 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4266 mdb_env_close(MDB_env *env)
4273 VGMEMP_DESTROY(env);
4274 while ((dp = env->me_dpages) != NULL) {
4275 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4276 env->me_dpages = dp->mp_next;
4280 mdb_env_close0(env, 0);
4284 /** Compare two items pointing at aligned size_t's */
4286 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4288 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4289 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4292 /** Compare two items pointing at aligned unsigned int's */
4294 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4296 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4297 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4300 /** Compare two items pointing at unsigned ints of unknown alignment.
4301 * Nodes and keys are guaranteed to be 2-byte aligned.
4304 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4306 #if BYTE_ORDER == LITTLE_ENDIAN
4307 unsigned short *u, *c;
4310 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4311 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4314 } while(!x && u > (unsigned short *)a->mv_data);
4317 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4321 /** Compare two items lexically */
4323 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4330 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4336 diff = memcmp(a->mv_data, b->mv_data, len);
4337 return diff ? diff : len_diff<0 ? -1 : len_diff;
4340 /** Compare two items in reverse byte order */
4342 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4344 const unsigned char *p1, *p2, *p1_lim;
4348 p1_lim = (const unsigned char *)a->mv_data;
4349 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4350 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4352 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4358 while (p1 > p1_lim) {
4359 diff = *--p1 - *--p2;
4363 return len_diff<0 ? -1 : len_diff;
4366 /** Search for key within a page, using binary search.
4367 * Returns the smallest entry larger or equal to the key.
4368 * If exactp is non-null, stores whether the found entry was an exact match
4369 * in *exactp (1 or 0).
4370 * Updates the cursor index with the index of the found entry.
4371 * If no entry larger or equal to the key is found, returns NULL.
4374 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4376 unsigned int i = 0, nkeys;
4379 MDB_page *mp = mc->mc_pg[mc->mc_top];
4380 MDB_node *node = NULL;
4385 nkeys = NUMKEYS(mp);
4390 COPY_PGNO(pgno, mp->mp_pgno);
4391 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4392 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4399 low = IS_LEAF(mp) ? 0 : 1;
4401 cmp = mc->mc_dbx->md_cmp;
4403 /* Branch pages have no data, so if using integer keys,
4404 * alignment is guaranteed. Use faster mdb_cmp_int.
4406 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4407 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4414 nodekey.mv_size = mc->mc_db->md_pad;
4415 node = NODEPTR(mp, 0); /* fake */
4416 while (low <= high) {
4417 i = (low + high) >> 1;
4418 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4419 rc = cmp(key, &nodekey);
4420 DPRINTF(("found leaf index %u [%s], rc = %i",
4421 i, DKEY(&nodekey), rc));
4430 while (low <= high) {
4431 i = (low + high) >> 1;
4433 node = NODEPTR(mp, i);
4434 nodekey.mv_size = NODEKSZ(node);
4435 nodekey.mv_data = NODEKEY(node);
4437 rc = cmp(key, &nodekey);
4440 DPRINTF(("found leaf index %u [%s], rc = %i",
4441 i, DKEY(&nodekey), rc));
4443 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4444 i, DKEY(&nodekey), NODEPGNO(node), rc));
4455 if (rc > 0) { /* Found entry is less than the key. */
4456 i++; /* Skip to get the smallest entry larger than key. */
4458 node = NODEPTR(mp, i);
4461 *exactp = (rc == 0);
4462 /* store the key index */
4463 mc->mc_ki[mc->mc_top] = i;
4465 /* There is no entry larger or equal to the key. */
4468 /* nodeptr is fake for LEAF2 */
4474 mdb_cursor_adjust(MDB_cursor *mc, func)
4478 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4479 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4486 /** Pop a page off the top of the cursor's stack. */
4488 mdb_cursor_pop(MDB_cursor *mc)
4492 MDB_page *top = mc->mc_pg[mc->mc_top];
4498 DPRINTF(("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4499 mc->mc_dbi, (void *) mc));
4503 /** Push a page onto the top of the cursor's stack. */
4505 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4507 DPRINTF(("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4508 mc->mc_dbi, (void *) mc));
4510 if (mc->mc_snum >= CURSOR_STACK) {
4511 assert(mc->mc_snum < CURSOR_STACK);
4512 return MDB_CURSOR_FULL;
4515 mc->mc_top = mc->mc_snum++;
4516 mc->mc_pg[mc->mc_top] = mp;
4517 mc->mc_ki[mc->mc_top] = 0;
4522 /** Find the address of the page corresponding to a given page number.
4523 * @param[in] txn the transaction for this access.
4524 * @param[in] pgno the page number for the page to retrieve.
4525 * @param[out] ret address of a pointer where the page's address will be stored.
4526 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4527 * @return 0 on success, non-zero on failure.
4530 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4532 MDB_env *env = txn->mt_env;
4536 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4540 MDB_ID2L dl = tx2->mt_u.dirty_list;
4542 /* Spilled pages were dirtied in this txn and flushed
4543 * because the dirty list got full. Bring this page
4544 * back in from the map (but don't unspill it here,
4545 * leave that unless page_touch happens again).
4547 if (tx2->mt_spill_pgs) {
4548 MDB_ID pn = pgno << 1;
4549 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4550 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4551 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4556 unsigned x = mdb_mid2l_search(dl, pgno);
4557 if (x <= dl[0].mid && dl[x].mid == pgno) {
4563 } while ((tx2 = tx2->mt_parent) != NULL);
4566 if (pgno < txn->mt_next_pgno) {
4568 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4570 DPRINTF(("page %"Z"u not found", pgno));
4572 return MDB_PAGE_NOTFOUND;
4582 /** Search for the page a given key should be in.
4583 * Pushes parent pages on the cursor stack. This function continues a
4584 * search on a cursor that has already been initialized. (Usually by
4585 * #mdb_page_search() but also by #mdb_node_move().)
4586 * @param[in,out] mc the cursor for this operation.
4587 * @param[in] key the key to search for. If NULL, search for the lowest
4588 * page. (This is used by #mdb_cursor_first().)
4589 * @param[in] modify If true, visited pages are updated with new page numbers.
4590 * @return 0 on success, non-zero on failure.
4593 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4595 MDB_page *mp = mc->mc_pg[mc->mc_top];
4599 while (IS_BRANCH(mp)) {
4603 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4604 assert(NUMKEYS(mp) > 1);
4605 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4607 if (key == NULL) /* Initialize cursor to first page. */
4609 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4610 /* cursor to last page */
4614 node = mdb_node_search(mc, key, &exact);
4616 i = NUMKEYS(mp) - 1;
4618 i = mc->mc_ki[mc->mc_top];
4627 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4628 assert(i < NUMKEYS(mp));
4629 node = NODEPTR(mp, i);
4631 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4634 mc->mc_ki[mc->mc_top] = i;
4635 if ((rc = mdb_cursor_push(mc, mp)))
4639 if ((rc = mdb_page_touch(mc)) != 0)
4641 mp = mc->mc_pg[mc->mc_top];
4646 DPRINTF(("internal error, index points to a %02X page!?",
4648 return MDB_CORRUPTED;
4651 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4652 key ? DKEY(key) : "null"));
4653 mc->mc_flags |= C_INITIALIZED;
4654 mc->mc_flags &= ~C_EOF;
4659 /** Search for the lowest key under the current branch page.
4660 * This just bypasses a NUMKEYS check in the current page
4661 * before calling mdb_page_search_root(), because the callers
4662 * are all in situations where the current page is known to
4666 mdb_page_search_lowest(MDB_cursor *mc)
4668 MDB_page *mp = mc->mc_pg[mc->mc_top];
4669 MDB_node *node = NODEPTR(mp, 0);
4672 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4675 mc->mc_ki[mc->mc_top] = 0;
4676 if ((rc = mdb_cursor_push(mc, mp)))
4678 return mdb_page_search_root(mc, NULL, 0);
4681 /** Search for the page a given key should be in.
4682 * Pushes parent pages on the cursor stack. This function just sets up
4683 * the search; it finds the root page for \b mc's database and sets this
4684 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4685 * called to complete the search.
4686 * @param[in,out] mc the cursor for this operation.
4687 * @param[in] key the key to search for. If NULL, search for the lowest
4688 * page. (This is used by #mdb_cursor_first().)
4689 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4690 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4691 * @return 0 on success, non-zero on failure.
4694 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4699 /* Make sure the txn is still viable, then find the root from
4700 * the txn's db table.
4702 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4703 DPUTS("transaction has failed, must abort");
4706 /* Make sure we're using an up-to-date root */
4707 if (mc->mc_dbi > MAIN_DBI) {
4708 if ((*mc->mc_dbflag & DB_STALE) ||
4709 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4711 unsigned char dbflag = 0;
4712 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4713 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4716 if (*mc->mc_dbflag & DB_STALE) {
4720 MDB_node *leaf = mdb_node_search(&mc2,
4721 &mc->mc_dbx->md_name, &exact);
4723 return MDB_NOTFOUND;
4724 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4727 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4729 /* The txn may not know this DBI, or another process may
4730 * have dropped and recreated the DB with other flags.
4732 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4733 return MDB_INCOMPATIBLE;
4734 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4736 if (flags & MDB_PS_MODIFY)
4738 *mc->mc_dbflag &= ~DB_STALE;
4739 *mc->mc_dbflag |= dbflag;
4742 root = mc->mc_db->md_root;
4744 if (root == P_INVALID) { /* Tree is empty. */
4745 DPUTS("tree is empty");
4746 return MDB_NOTFOUND;
4751 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4752 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4758 DPRINTF(("db %u root page %"Z"u has flags 0x%X",
4759 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags));
4761 if (flags & MDB_PS_MODIFY) {
4762 if ((rc = mdb_page_touch(mc)))
4766 if (flags & MDB_PS_ROOTONLY)
4769 return mdb_page_search_root(mc, key, flags);
4773 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4775 MDB_txn *txn = mc->mc_txn;
4776 pgno_t pg = mp->mp_pgno;
4777 unsigned x = 0, ovpages = mp->mp_pages;
4778 MDB_env *env = txn->mt_env;
4779 MDB_IDL sl = txn->mt_spill_pgs;
4780 MDB_ID pn = pg << 1;
4783 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4784 /* If the page is dirty or on the spill list we just acquired it,
4785 * so we should give it back to our current free list, if any.
4786 * Otherwise put it onto the list of pages we freed in this txn.
4788 * Won't create me_pghead: me_pglast must be inited along with it.
4789 * Unsupported in nested txns: They would need to hide the page
4790 * range in ancestor txns' dirty and spilled lists.
4792 if (env->me_pghead &&
4794 ((mp->mp_flags & P_DIRTY) ||
4795 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4799 MDB_ID2 *dl, ix, iy;
4800 rc = mdb_midl_need(&env->me_pghead, ovpages);
4803 if (!(mp->mp_flags & P_DIRTY)) {
4804 /* This page is no longer spilled */
4811 /* Remove from dirty list */
4812 dl = txn->mt_u.dirty_list;
4814 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4822 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4823 txn->mt_flags |= MDB_TXN_ERROR;
4824 return MDB_CORRUPTED;
4827 if (!(env->me_flags & MDB_WRITEMAP))
4828 mdb_dpage_free(env, mp);
4830 /* Insert in me_pghead */
4831 mop = env->me_pghead;
4832 j = mop[0] + ovpages;
4833 for (i = mop[0]; i && mop[i] < pg; i--)
4839 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4843 mc->mc_db->md_overflow_pages -= ovpages;
4847 /** Return the data associated with a given node.
4848 * @param[in] txn The transaction for this operation.
4849 * @param[in] leaf The node being read.
4850 * @param[out] data Updated to point to the node's data.
4851 * @return 0 on success, non-zero on failure.
4854 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4856 MDB_page *omp; /* overflow page */
4860 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4861 data->mv_size = NODEDSZ(leaf);
4862 data->mv_data = NODEDATA(leaf);
4866 /* Read overflow data.
4868 data->mv_size = NODEDSZ(leaf);
4869 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4870 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4871 DPRINTF(("read overflow page %"Z"u failed", pgno));
4874 data->mv_data = METADATA(omp);
4880 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4881 MDB_val *key, MDB_val *data)
4890 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4892 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4895 if (txn->mt_flags & MDB_TXN_ERROR)
4898 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4899 return MDB_BAD_VALSIZE;
4902 mdb_cursor_init(&mc, txn, dbi, &mx);
4903 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4906 /** Find a sibling for a page.
4907 * Replaces the page at the top of the cursor's stack with the
4908 * specified sibling, if one exists.
4909 * @param[in] mc The cursor for this operation.
4910 * @param[in] move_right Non-zero if the right sibling is requested,
4911 * otherwise the left sibling.
4912 * @return 0 on success, non-zero on failure.
4915 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4921 if (mc->mc_snum < 2) {
4922 return MDB_NOTFOUND; /* root has no siblings */
4926 DPRINTF(("parent page is page %"Z"u, index %u",
4927 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4929 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4930 : (mc->mc_ki[mc->mc_top] == 0)) {
4931 DPRINTF(("no more keys left, moving to %s sibling",
4932 move_right ? "right" : "left"));
4933 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4934 /* undo cursor_pop before returning */
4941 mc->mc_ki[mc->mc_top]++;
4943 mc->mc_ki[mc->mc_top]--;
4944 DPRINTF(("just moving to %s index key %u",
4945 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4947 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4949 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4950 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4953 mdb_cursor_push(mc, mp);
4955 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4960 /** Move the cursor to the next data item. */
4962 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4968 if (mc->mc_flags & C_EOF) {
4969 return MDB_NOTFOUND;
4972 assert(mc->mc_flags & C_INITIALIZED);
4974 mp = mc->mc_pg[mc->mc_top];
4976 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4977 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4978 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4979 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4980 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4981 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
4982 if (rc == MDB_SUCCESS)
4983 MDB_GET_KEY(leaf, key);
4988 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4989 if (op == MDB_NEXT_DUP)
4990 return MDB_NOTFOUND;
4994 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
4995 if (mc->mc_flags & C_DEL)
4998 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4999 DPUTS("=====> move to next sibling page");
5000 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5001 mc->mc_flags |= C_EOF;
5004 mp = mc->mc_pg[mc->mc_top];
5005 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5007 mc->mc_ki[mc->mc_top]++;
5010 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5011 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5014 key->mv_size = mc->mc_db->md_pad;
5015 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5019 assert(IS_LEAF(mp));
5020 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5022 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5023 mdb_xcursor_init1(mc, leaf);
5026 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5029 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5030 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5031 if (rc != MDB_SUCCESS)
5036 MDB_GET_KEY(leaf, key);
5040 /** Move the cursor to the previous data item. */
5042 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5048 assert(mc->mc_flags & C_INITIALIZED);
5050 mp = mc->mc_pg[mc->mc_top];
5052 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5053 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5054 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5055 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5056 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5057 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5058 if (rc == MDB_SUCCESS)
5059 MDB_GET_KEY(leaf, key);
5063 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5064 if (op == MDB_PREV_DUP)
5065 return MDB_NOTFOUND;
5070 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5072 if (mc->mc_ki[mc->mc_top] == 0) {
5073 DPUTS("=====> move to prev sibling page");
5074 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5077 mp = mc->mc_pg[mc->mc_top];
5078 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5079 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5081 mc->mc_ki[mc->mc_top]--;
5083 mc->mc_flags &= ~C_EOF;
5085 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5086 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5089 key->mv_size = mc->mc_db->md_pad;
5090 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5094 assert(IS_LEAF(mp));
5095 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5097 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5098 mdb_xcursor_init1(mc, leaf);
5101 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5104 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5105 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5106 if (rc != MDB_SUCCESS)
5111 MDB_GET_KEY(leaf, key);
5115 /** Set the cursor on a specific data item. */
5117 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5118 MDB_cursor_op op, int *exactp)
5122 MDB_node *leaf = NULL;
5127 assert(key->mv_size > 0);
5130 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5132 /* See if we're already on the right page */
5133 if (mc->mc_flags & C_INITIALIZED) {
5136 mp = mc->mc_pg[mc->mc_top];
5138 mc->mc_ki[mc->mc_top] = 0;
5139 return MDB_NOTFOUND;
5141 if (mp->mp_flags & P_LEAF2) {
5142 nodekey.mv_size = mc->mc_db->md_pad;
5143 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5145 leaf = NODEPTR(mp, 0);
5146 MDB_GET_KEY2(leaf, nodekey);
5148 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5150 /* Probably happens rarely, but first node on the page
5151 * was the one we wanted.
5153 mc->mc_ki[mc->mc_top] = 0;
5160 unsigned int nkeys = NUMKEYS(mp);
5162 if (mp->mp_flags & P_LEAF2) {
5163 nodekey.mv_data = LEAF2KEY(mp,
5164 nkeys-1, nodekey.mv_size);
5166 leaf = NODEPTR(mp, nkeys-1);
5167 MDB_GET_KEY2(leaf, nodekey);
5169 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5171 /* last node was the one we wanted */
5172 mc->mc_ki[mc->mc_top] = nkeys-1;
5178 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5179 /* This is definitely the right page, skip search_page */
5180 if (mp->mp_flags & P_LEAF2) {
5181 nodekey.mv_data = LEAF2KEY(mp,
5182 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5184 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5185 MDB_GET_KEY2(leaf, nodekey);
5187 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5189 /* current node was the one we wanted */
5199 /* If any parents have right-sibs, search.
5200 * Otherwise, there's nothing further.
5202 for (i=0; i<mc->mc_top; i++)
5204 NUMKEYS(mc->mc_pg[i])-1)
5206 if (i == mc->mc_top) {
5207 /* There are no other pages */
5208 mc->mc_ki[mc->mc_top] = nkeys;
5209 return MDB_NOTFOUND;
5213 /* There are no other pages */
5214 mc->mc_ki[mc->mc_top] = 0;
5215 if (op == MDB_SET_RANGE) {
5219 return MDB_NOTFOUND;
5223 rc = mdb_page_search(mc, key, 0);
5224 if (rc != MDB_SUCCESS)
5227 mp = mc->mc_pg[mc->mc_top];
5228 assert(IS_LEAF(mp));
5231 leaf = mdb_node_search(mc, key, exactp);
5232 if (exactp != NULL && !*exactp) {
5233 /* MDB_SET specified and not an exact match. */
5234 return MDB_NOTFOUND;
5238 DPUTS("===> inexact leaf not found, goto sibling");
5239 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5240 return rc; /* no entries matched */
5241 mp = mc->mc_pg[mc->mc_top];
5242 assert(IS_LEAF(mp));
5243 leaf = NODEPTR(mp, 0);
5247 mc->mc_flags |= C_INITIALIZED;
5248 mc->mc_flags &= ~C_EOF;
5251 key->mv_size = mc->mc_db->md_pad;
5252 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5256 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5257 mdb_xcursor_init1(mc, leaf);
5260 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5261 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5262 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5265 if (op == MDB_GET_BOTH) {
5271 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5272 if (rc != MDB_SUCCESS)
5275 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5277 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5279 rc = mc->mc_dbx->md_dcmp(data, &d2);
5281 if (op == MDB_GET_BOTH || rc > 0)
5282 return MDB_NOTFOUND;
5288 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5289 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5294 /* The key already matches in all other cases */
5295 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5296 MDB_GET_KEY(leaf, key);
5297 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5302 /** Move the cursor to the first item in the database. */
5304 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5310 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5312 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5313 rc = mdb_page_search(mc, NULL, 0);
5314 if (rc != MDB_SUCCESS)
5317 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5319 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5320 mc->mc_flags |= C_INITIALIZED;
5321 mc->mc_flags &= ~C_EOF;
5323 mc->mc_ki[mc->mc_top] = 0;
5325 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5326 key->mv_size = mc->mc_db->md_pad;
5327 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5332 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5333 mdb_xcursor_init1(mc, leaf);
5334 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5338 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5342 MDB_GET_KEY(leaf, key);
5346 /** Move the cursor to the last item in the database. */
5348 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5354 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5356 if (!(mc->mc_flags & C_EOF)) {
5358 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5361 lkey.mv_size = MDB_MAXKEYSIZE+1;
5362 lkey.mv_data = NULL;
5363 rc = mdb_page_search(mc, &lkey, 0);
5364 if (rc != MDB_SUCCESS)
5367 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5370 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5371 mc->mc_flags |= C_INITIALIZED|C_EOF;
5372 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5374 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5375 key->mv_size = mc->mc_db->md_pad;
5376 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5381 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5382 mdb_xcursor_init1(mc, leaf);
5383 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5387 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5392 MDB_GET_KEY(leaf, key);
5397 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5402 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5406 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5410 case MDB_GET_CURRENT:
5411 if (!(mc->mc_flags & C_INITIALIZED)) {
5414 MDB_page *mp = mc->mc_pg[mc->mc_top];
5416 mc->mc_ki[mc->mc_top] = 0;
5422 key->mv_size = mc->mc_db->md_pad;
5423 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5425 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5426 MDB_GET_KEY(leaf, key);
5428 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5429 if (mc->mc_flags & C_DEL)
5430 mdb_xcursor_init1(mc, leaf);
5431 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5433 rc = mdb_node_read(mc->mc_txn, leaf, data);
5440 case MDB_GET_BOTH_RANGE:
5445 if (mc->mc_xcursor == NULL) {
5446 rc = MDB_INCOMPATIBLE;
5455 } else if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5456 rc = MDB_BAD_VALSIZE;
5457 } else if (op == MDB_SET_RANGE)
5458 rc = mdb_cursor_set(mc, key, data, op, NULL);
5460 rc = mdb_cursor_set(mc, key, data, op, &exact);
5462 case MDB_GET_MULTIPLE:
5463 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5467 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5468 rc = MDB_INCOMPATIBLE;
5472 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5473 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5476 case MDB_NEXT_MULTIPLE:
5481 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5482 rc = MDB_INCOMPATIBLE;
5485 if (!(mc->mc_flags & C_INITIALIZED))
5486 rc = mdb_cursor_first(mc, key, data);
5488 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5489 if (rc == MDB_SUCCESS) {
5490 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5493 mx = &mc->mc_xcursor->mx_cursor;
5494 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5496 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5497 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5505 case MDB_NEXT_NODUP:
5506 if (!(mc->mc_flags & C_INITIALIZED))
5507 rc = mdb_cursor_first(mc, key, data);
5509 rc = mdb_cursor_next(mc, key, data, op);
5513 case MDB_PREV_NODUP:
5514 if (!(mc->mc_flags & C_INITIALIZED)) {
5515 rc = mdb_cursor_last(mc, key, data);
5518 mc->mc_flags |= C_INITIALIZED;
5519 mc->mc_ki[mc->mc_top]++;
5521 rc = mdb_cursor_prev(mc, key, data, op);
5524 rc = mdb_cursor_first(mc, key, data);
5527 mfunc = mdb_cursor_first;
5529 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5533 if (mc->mc_xcursor == NULL) {
5534 rc = MDB_INCOMPATIBLE;
5537 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5541 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5544 rc = mdb_cursor_last(mc, key, data);
5547 mfunc = mdb_cursor_last;
5550 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5555 if (mc->mc_flags & C_DEL)
5556 mc->mc_flags ^= C_DEL;
5561 /** Touch all the pages in the cursor stack.
5562 * Makes sure all the pages are writable, before attempting a write operation.
5563 * @param[in] mc The cursor to operate on.
5566 mdb_cursor_touch(MDB_cursor *mc)
5570 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5573 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5574 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5577 *mc->mc_dbflag |= DB_DIRTY;
5579 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5580 rc = mdb_page_touch(mc);
5584 mc->mc_top = mc->mc_snum-1;
5588 /** Do not spill pages to disk if txn is getting full, may fail instead */
5589 #define MDB_NOSPILL 0x8000
5592 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5595 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5596 MDB_node *leaf = NULL;
5597 MDB_val xdata, *rdata, dkey;
5600 int do_sub = 0, insert = 0;
5601 unsigned int mcount = 0, dcount = 0, nospill;
5605 char dbuf[MDB_MAXKEYSIZE+1];
5606 unsigned int nflags;
5609 /* Check this first so counter will always be zero on any
5612 if (flags & MDB_MULTIPLE) {
5613 dcount = data[1].mv_size;
5614 data[1].mv_size = 0;
5615 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5616 return MDB_INCOMPATIBLE;
5619 nospill = flags & MDB_NOSPILL;
5620 flags &= ~MDB_NOSPILL;
5622 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5623 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5625 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5626 return MDB_BAD_VALSIZE;
5628 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5629 return MDB_BAD_VALSIZE;
5631 #if SIZE_MAX > MAXDATASIZE
5632 if (data->mv_size > MAXDATASIZE)
5633 return MDB_BAD_VALSIZE;
5636 DPRINTF(("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5637 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size));
5641 if (flags == MDB_CURRENT) {
5642 if (!(mc->mc_flags & C_INITIALIZED))
5645 } else if (mc->mc_db->md_root == P_INVALID) {
5646 /* new database, cursor has nothing to point to */
5648 mc->mc_flags &= ~C_INITIALIZED;
5653 if (flags & MDB_APPEND) {
5655 rc = mdb_cursor_last(mc, &k2, &d2);
5657 rc = mc->mc_dbx->md_cmp(key, &k2);
5660 mc->mc_ki[mc->mc_top]++;
5662 /* new key is <= last key */
5667 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5669 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5670 DPRINTF(("duplicate key [%s]", DKEY(key)));
5672 return MDB_KEYEXIST;
5674 if (rc && rc != MDB_NOTFOUND)
5678 if (mc->mc_flags & C_DEL)
5679 mc->mc_flags ^= C_DEL;
5681 /* Cursor is positioned, check for room in the dirty list */
5683 if (flags & MDB_MULTIPLE) {
5685 xdata.mv_size = data->mv_size * dcount;
5689 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5693 if (rc == MDB_NO_ROOT) {
5695 /* new database, write a root leaf page */
5696 DPUTS("allocating new root leaf page");
5697 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5700 mdb_cursor_push(mc, np);
5701 mc->mc_db->md_root = np->mp_pgno;
5702 mc->mc_db->md_depth++;
5703 *mc->mc_dbflag |= DB_DIRTY;
5704 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5706 np->mp_flags |= P_LEAF2;
5707 mc->mc_flags |= C_INITIALIZED;
5709 /* make sure all cursor pages are writable */
5710 rc2 = mdb_cursor_touch(mc);
5715 /* The key already exists */
5716 if (rc == MDB_SUCCESS) {
5717 /* there's only a key anyway, so this is a no-op */
5718 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5719 unsigned int ksize = mc->mc_db->md_pad;
5720 if (key->mv_size != ksize)
5721 return MDB_BAD_VALSIZE;
5722 if (flags == MDB_CURRENT) {
5723 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5724 memcpy(ptr, key->mv_data, ksize);
5729 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5732 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5733 /* Was a single item before, must convert now */
5735 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5736 /* Just overwrite the current item */
5737 if (flags == MDB_CURRENT)
5740 dkey.mv_size = NODEDSZ(leaf);
5741 dkey.mv_data = NODEDATA(leaf);
5742 #if UINT_MAX < SIZE_MAX
5743 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5744 #ifdef MISALIGNED_OK
5745 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5747 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5750 /* if data matches, skip it */
5751 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5752 if (flags & MDB_NODUPDATA)
5754 else if (flags & MDB_MULTIPLE)
5761 /* create a fake page for the dup items */
5762 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5763 dkey.mv_data = dbuf;
5764 fp = (MDB_page *)&pbuf;
5765 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5766 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5767 fp->mp_lower = PAGEHDRSZ;
5768 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5769 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5770 fp->mp_flags |= P_LEAF2;
5771 fp->mp_pad = data->mv_size;
5772 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5774 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5775 (dkey.mv_size & 1) + (data->mv_size & 1);
5777 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5780 xdata.mv_size = fp->mp_upper;
5785 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5786 /* See if we need to convert from fake page to subDB */
5788 unsigned int offset;
5792 fp = NODEDATA(leaf);
5793 if (flags == MDB_CURRENT) {
5795 fp->mp_flags |= P_DIRTY;
5796 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5797 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5801 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5802 offset = fp->mp_pad;
5803 if (SIZELEFT(fp) >= offset)
5805 offset *= 4; /* space for 4 more */
5807 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5809 offset += offset & 1;
5810 fp_flags = fp->mp_flags;
5811 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5812 offset >= mc->mc_txn->mt_env->me_nodemax) {
5813 /* yes, convert it */
5815 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5816 dummy.md_pad = fp->mp_pad;
5817 dummy.md_flags = MDB_DUPFIXED;
5818 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5819 dummy.md_flags |= MDB_INTEGERKEY;
5822 dummy.md_branch_pages = 0;
5823 dummy.md_leaf_pages = 1;
5824 dummy.md_overflow_pages = 0;
5825 dummy.md_entries = NUMKEYS(fp);
5827 xdata.mv_size = sizeof(MDB_db);
5828 xdata.mv_data = &dummy;
5829 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5831 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5832 flags |= F_DUPDATA|F_SUBDATA;
5833 dummy.md_root = mp->mp_pgno;
5834 fp_flags &= ~P_SUBP;
5836 /* no, just grow it */
5838 xdata.mv_size = NODEDSZ(leaf) + offset;
5839 xdata.mv_data = &pbuf;
5840 mp = (MDB_page *)&pbuf;
5841 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5844 mp->mp_flags = fp_flags | P_DIRTY;
5845 mp->mp_pad = fp->mp_pad;
5846 mp->mp_lower = fp->mp_lower;
5847 mp->mp_upper = fp->mp_upper + offset;
5849 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5851 nsize = NODEDSZ(leaf) - fp->mp_upper;
5852 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5853 for (i=0; i<NUMKEYS(fp); i++)
5854 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5856 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5860 /* data is on sub-DB, just store it */
5861 flags |= F_DUPDATA|F_SUBDATA;
5865 /* overflow page overwrites need special handling */
5866 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5869 unsigned psize = mc->mc_txn->mt_env->me_psize;
5870 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5872 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5873 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5875 ovpages = omp->mp_pages;
5877 /* Is the ov page large enough? */
5878 if (ovpages >= dpages) {
5879 if (!(omp->mp_flags & P_DIRTY) &&
5880 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5882 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5885 level = 0; /* dirty in this txn or clean */
5888 if (omp->mp_flags & P_DIRTY) {
5889 /* yes, overwrite it. Note in this case we don't
5890 * bother to try shrinking the page if the new data
5891 * is smaller than the overflow threshold.
5894 /* It is writable only in a parent txn */
5895 size_t sz = (size_t) psize * ovpages, off;
5896 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5902 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5903 if (!(flags & MDB_RESERVE)) {
5904 /* Copy end of page, adjusting alignment so
5905 * compiler may copy words instead of bytes.
5907 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5908 memcpy((size_t *)((char *)np + off),
5909 (size_t *)((char *)omp + off), sz - off);
5912 memcpy(np, omp, sz); /* Copy beginning of page */
5915 SETDSZ(leaf, data->mv_size);
5916 if (F_ISSET(flags, MDB_RESERVE))
5917 data->mv_data = METADATA(omp);
5919 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5923 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5925 } else if (NODEDSZ(leaf) == data->mv_size) {
5926 /* same size, just replace it. Note that we could
5927 * also reuse this node if the new data is smaller,
5928 * but instead we opt to shrink the node in that case.
5930 if (F_ISSET(flags, MDB_RESERVE))
5931 data->mv_data = NODEDATA(leaf);
5932 else if (data->mv_size)
5933 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5935 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5938 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5939 mc->mc_db->md_entries--;
5941 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5948 nflags = flags & NODE_ADD_FLAGS;
5949 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5950 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5951 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5952 nflags &= ~MDB_APPEND;
5954 nflags |= MDB_SPLIT_REPLACE;
5955 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5957 /* There is room already in this leaf page. */
5958 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5959 if (rc == 0 && !do_sub && insert) {
5960 /* Adjust other cursors pointing to mp */
5961 MDB_cursor *m2, *m3;
5962 MDB_dbi dbi = mc->mc_dbi;
5963 unsigned i = mc->mc_top;
5964 MDB_page *mp = mc->mc_pg[i];
5966 if (mc->mc_flags & C_SUB)
5969 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5970 if (mc->mc_flags & C_SUB)
5971 m3 = &m2->mc_xcursor->mx_cursor;
5974 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5975 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5982 if (rc != MDB_SUCCESS)
5983 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5985 /* Now store the actual data in the child DB. Note that we're
5986 * storing the user data in the keys field, so there are strict
5987 * size limits on dupdata. The actual data fields of the child
5988 * DB are all zero size.
5995 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5996 if (flags & MDB_CURRENT) {
5997 xflags = MDB_CURRENT|MDB_NOSPILL;
5999 mdb_xcursor_init1(mc, leaf);
6000 xflags = (flags & MDB_NODUPDATA) ?
6001 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6003 /* converted, write the original data first */
6005 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6009 /* Adjust other cursors pointing to mp */
6011 unsigned i = mc->mc_top;
6012 MDB_page *mp = mc->mc_pg[i];
6014 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6015 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6016 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6017 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6018 mdb_xcursor_init1(m2, leaf);
6022 /* we've done our job */
6025 if (flags & MDB_APPENDDUP)
6026 xflags |= MDB_APPEND;
6027 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6028 if (flags & F_SUBDATA) {
6029 void *db = NODEDATA(leaf);
6030 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6033 /* sub-writes might have failed so check rc again.
6034 * Don't increment count if we just replaced an existing item.
6036 if (!rc && !(flags & MDB_CURRENT))
6037 mc->mc_db->md_entries++;
6038 if (flags & MDB_MULTIPLE) {
6042 /* let caller know how many succeeded, if any */
6043 data[1].mv_size = mcount;
6044 if (mcount < dcount) {
6045 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6046 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6053 /* If we succeeded and the key didn't exist before, make sure
6054 * the cursor is marked valid.
6057 mc->mc_flags |= C_INITIALIZED;
6062 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6067 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6068 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6070 if (!(mc->mc_flags & C_INITIALIZED))
6073 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6075 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
6077 rc = mdb_cursor_touch(mc);
6081 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6083 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6084 if (!(flags & MDB_NODUPDATA)) {
6085 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6086 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6088 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6089 /* If sub-DB still has entries, we're done */
6090 if (mc->mc_xcursor->mx_db.md_entries) {
6091 if (leaf->mn_flags & F_SUBDATA) {
6092 /* update subDB info */
6093 void *db = NODEDATA(leaf);
6094 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6097 /* shrink fake page */
6098 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6099 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6100 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6101 /* fix other sub-DB cursors pointed at this fake page */
6102 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6103 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6104 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6105 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6106 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6109 mc->mc_db->md_entries--;
6110 mc->mc_flags |= C_DEL;
6113 /* otherwise fall thru and delete the sub-DB */
6116 if (leaf->mn_flags & F_SUBDATA) {
6117 /* add all the child DB's pages to the free list */
6118 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6119 if (rc == MDB_SUCCESS) {
6120 mc->mc_db->md_entries -=
6121 mc->mc_xcursor->mx_db.md_entries;
6126 return mdb_cursor_del0(mc, leaf);
6129 /** Allocate and initialize new pages for a database.
6130 * @param[in] mc a cursor on the database being added to.
6131 * @param[in] flags flags defining what type of page is being allocated.
6132 * @param[in] num the number of pages to allocate. This is usually 1,
6133 * unless allocating overflow pages for a large record.
6134 * @param[out] mp Address of a page, or NULL on failure.
6135 * @return 0 on success, non-zero on failure.
6138 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6143 if ((rc = mdb_page_alloc(mc, num, &np)))
6145 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6146 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6147 np->mp_flags = flags | P_DIRTY;
6148 np->mp_lower = PAGEHDRSZ;
6149 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6152 mc->mc_db->md_branch_pages++;
6153 else if (IS_LEAF(np))
6154 mc->mc_db->md_leaf_pages++;
6155 else if (IS_OVERFLOW(np)) {
6156 mc->mc_db->md_overflow_pages += num;
6164 /** Calculate the size of a leaf node.
6165 * The size depends on the environment's page size; if a data item
6166 * is too large it will be put onto an overflow page and the node
6167 * size will only include the key and not the data. Sizes are always
6168 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6169 * of the #MDB_node headers.
6170 * @param[in] env The environment handle.
6171 * @param[in] key The key for the node.
6172 * @param[in] data The data for the node.
6173 * @return The number of bytes needed to store the node.
6176 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6180 sz = LEAFSIZE(key, data);
6181 if (sz >= env->me_nodemax) {
6182 /* put on overflow page */
6183 sz -= data->mv_size - sizeof(pgno_t);
6187 return sz + sizeof(indx_t);
6190 /** Calculate the size of a branch node.
6191 * The size should depend on the environment's page size but since
6192 * we currently don't support spilling large keys onto overflow
6193 * pages, it's simply the size of the #MDB_node header plus the
6194 * size of the key. Sizes are always rounded up to an even number
6195 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6196 * @param[in] env The environment handle.
6197 * @param[in] key The key for the node.
6198 * @return The number of bytes needed to store the node.
6201 mdb_branch_size(MDB_env *env, MDB_val *key)
6206 if (sz >= env->me_nodemax) {
6207 /* put on overflow page */
6208 /* not implemented */
6209 /* sz -= key->size - sizeof(pgno_t); */
6212 return sz + sizeof(indx_t);
6215 /** Add a node to the page pointed to by the cursor.
6216 * @param[in] mc The cursor for this operation.
6217 * @param[in] indx The index on the page where the new node should be added.
6218 * @param[in] key The key for the new node.
6219 * @param[in] data The data for the new node, if any.
6220 * @param[in] pgno The page number, if adding a branch node.
6221 * @param[in] flags Flags for the node.
6222 * @return 0 on success, non-zero on failure. Possible errors are:
6224 * <li>ENOMEM - failed to allocate overflow pages for the node.
6225 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6226 * should never happen since all callers already calculate the
6227 * page's free space before calling this function.
6231 mdb_node_add(MDB_cursor *mc, indx_t indx,
6232 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6235 size_t node_size = NODESIZE;
6238 MDB_page *mp = mc->mc_pg[mc->mc_top];
6239 MDB_page *ofp = NULL; /* overflow page */
6242 assert(mp->mp_upper >= mp->mp_lower);
6244 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6245 IS_LEAF(mp) ? "leaf" : "branch",
6246 IS_SUBP(mp) ? "sub-" : "",
6247 mp->mp_pgno, indx, data ? data->mv_size : 0,
6248 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6251 /* Move higher keys up one slot. */
6252 int ksize = mc->mc_db->md_pad, dif;
6253 char *ptr = LEAF2KEY(mp, indx, ksize);
6254 dif = NUMKEYS(mp) - indx;
6256 memmove(ptr+ksize, ptr, dif*ksize);
6257 /* insert new key */
6258 memcpy(ptr, key->mv_data, ksize);
6260 /* Just using these for counting */
6261 mp->mp_lower += sizeof(indx_t);
6262 mp->mp_upper -= ksize - sizeof(indx_t);
6267 node_size += key->mv_size;
6271 if (F_ISSET(flags, F_BIGDATA)) {
6272 /* Data already on overflow page. */
6273 node_size += sizeof(pgno_t);
6274 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6275 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6277 /* Put data on overflow page. */
6278 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6279 data->mv_size, node_size+data->mv_size));
6280 node_size += sizeof(pgno_t);
6281 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6283 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6286 node_size += data->mv_size;
6289 node_size += node_size & 1;
6291 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6292 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6293 mp->mp_pgno, NUMKEYS(mp)));
6294 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6295 mp->mp_upper - mp->mp_lower));
6296 DPRINTF(("node size = %"Z"u", node_size));
6297 return MDB_PAGE_FULL;
6300 /* Move higher pointers up one slot. */
6301 for (i = NUMKEYS(mp); i > indx; i--)
6302 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6304 /* Adjust free space offsets. */
6305 ofs = mp->mp_upper - node_size;
6306 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6307 mp->mp_ptrs[indx] = ofs;
6309 mp->mp_lower += sizeof(indx_t);
6311 /* Write the node data. */
6312 node = NODEPTR(mp, indx);
6313 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6314 node->mn_flags = flags;
6316 SETDSZ(node,data->mv_size);
6321 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6326 if (F_ISSET(flags, F_BIGDATA))
6327 memcpy(node->mn_data + key->mv_size, data->mv_data,
6329 else if (F_ISSET(flags, MDB_RESERVE))
6330 data->mv_data = node->mn_data + key->mv_size;
6332 memcpy(node->mn_data + key->mv_size, data->mv_data,
6335 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6337 if (F_ISSET(flags, MDB_RESERVE))
6338 data->mv_data = METADATA(ofp);
6340 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6347 /** Delete the specified node from a page.
6348 * @param[in] mp The page to operate on.
6349 * @param[in] indx The index of the node to delete.
6350 * @param[in] ksize The size of a node. Only used if the page is
6351 * part of a #MDB_DUPFIXED database.
6354 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6357 indx_t i, j, numkeys, ptr;
6364 COPY_PGNO(pgno, mp->mp_pgno);
6365 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6366 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6369 assert(indx < NUMKEYS(mp));
6372 int x = NUMKEYS(mp) - 1 - indx;
6373 base = LEAF2KEY(mp, indx, ksize);
6375 memmove(base, base + ksize, x * ksize);
6376 mp->mp_lower -= sizeof(indx_t);
6377 mp->mp_upper += ksize - sizeof(indx_t);
6381 node = NODEPTR(mp, indx);
6382 sz = NODESIZE + node->mn_ksize;
6384 if (F_ISSET(node->mn_flags, F_BIGDATA))
6385 sz += sizeof(pgno_t);
6387 sz += NODEDSZ(node);
6391 ptr = mp->mp_ptrs[indx];
6392 numkeys = NUMKEYS(mp);
6393 for (i = j = 0; i < numkeys; i++) {
6395 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6396 if (mp->mp_ptrs[i] < ptr)
6397 mp->mp_ptrs[j] += sz;
6402 base = (char *)mp + mp->mp_upper;
6403 memmove(base + sz, base, ptr - mp->mp_upper);
6405 mp->mp_lower -= sizeof(indx_t);
6409 /** Compact the main page after deleting a node on a subpage.
6410 * @param[in] mp The main page to operate on.
6411 * @param[in] indx The index of the subpage on the main page.
6414 mdb_node_shrink(MDB_page *mp, indx_t indx)
6421 indx_t i, numkeys, ptr;
6423 node = NODEPTR(mp, indx);
6424 sp = (MDB_page *)NODEDATA(node);
6425 osize = NODEDSZ(node);
6427 delta = sp->mp_upper - sp->mp_lower;
6428 SETDSZ(node, osize - delta);
6429 xp = (MDB_page *)((char *)sp + delta);
6431 /* shift subpage upward */
6433 nsize = NUMKEYS(sp) * sp->mp_pad;
6434 memmove(METADATA(xp), METADATA(sp), nsize);
6437 nsize = osize - sp->mp_upper;
6438 numkeys = NUMKEYS(sp);
6439 for (i=numkeys-1; i>=0; i--)
6440 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6442 xp->mp_upper = sp->mp_lower;
6443 xp->mp_lower = sp->mp_lower;
6444 xp->mp_flags = sp->mp_flags;
6445 xp->mp_pad = sp->mp_pad;
6446 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6448 /* shift lower nodes upward */
6449 ptr = mp->mp_ptrs[indx];
6450 numkeys = NUMKEYS(mp);
6451 for (i = 0; i < numkeys; i++) {
6452 if (mp->mp_ptrs[i] <= ptr)
6453 mp->mp_ptrs[i] += delta;
6456 base = (char *)mp + mp->mp_upper;
6457 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6458 mp->mp_upper += delta;
6461 /** Initial setup of a sorted-dups cursor.
6462 * Sorted duplicates are implemented as a sub-database for the given key.
6463 * The duplicate data items are actually keys of the sub-database.
6464 * Operations on the duplicate data items are performed using a sub-cursor
6465 * initialized when the sub-database is first accessed. This function does
6466 * the preliminary setup of the sub-cursor, filling in the fields that
6467 * depend only on the parent DB.
6468 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6471 mdb_xcursor_init0(MDB_cursor *mc)
6473 MDB_xcursor *mx = mc->mc_xcursor;
6475 mx->mx_cursor.mc_xcursor = NULL;
6476 mx->mx_cursor.mc_txn = mc->mc_txn;
6477 mx->mx_cursor.mc_db = &mx->mx_db;
6478 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6479 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6480 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6481 mx->mx_cursor.mc_snum = 0;
6482 mx->mx_cursor.mc_top = 0;
6483 mx->mx_cursor.mc_flags = C_SUB;
6484 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6485 mx->mx_dbx.md_dcmp = NULL;
6486 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6489 /** Final setup of a sorted-dups cursor.
6490 * Sets up the fields that depend on the data from the main cursor.
6491 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6492 * @param[in] node The data containing the #MDB_db record for the
6493 * sorted-dup database.
6496 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6498 MDB_xcursor *mx = mc->mc_xcursor;
6500 if (node->mn_flags & F_SUBDATA) {
6501 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6502 mx->mx_cursor.mc_pg[0] = 0;
6503 mx->mx_cursor.mc_snum = 0;
6504 mx->mx_cursor.mc_flags = C_SUB;
6506 MDB_page *fp = NODEDATA(node);
6507 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6508 mx->mx_db.md_flags = 0;
6509 mx->mx_db.md_depth = 1;
6510 mx->mx_db.md_branch_pages = 0;
6511 mx->mx_db.md_leaf_pages = 1;
6512 mx->mx_db.md_overflow_pages = 0;
6513 mx->mx_db.md_entries = NUMKEYS(fp);
6514 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6515 mx->mx_cursor.mc_snum = 1;
6516 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6517 mx->mx_cursor.mc_top = 0;
6518 mx->mx_cursor.mc_pg[0] = fp;
6519 mx->mx_cursor.mc_ki[0] = 0;
6520 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6521 mx->mx_db.md_flags = MDB_DUPFIXED;
6522 mx->mx_db.md_pad = fp->mp_pad;
6523 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6524 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6527 DPRINTF(("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6528 mx->mx_db.md_root));
6529 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6531 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6532 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6533 #if UINT_MAX < SIZE_MAX
6534 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6535 #ifdef MISALIGNED_OK
6536 mx->mx_dbx.md_cmp = mdb_cmp_long;
6538 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6543 /** Initialize a cursor for a given transaction and database. */
6545 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6548 mc->mc_backup = NULL;
6551 mc->mc_db = &txn->mt_dbs[dbi];
6552 mc->mc_dbx = &txn->mt_dbxs[dbi];
6553 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6558 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6560 mc->mc_xcursor = mx;
6561 mdb_xcursor_init0(mc);
6563 mc->mc_xcursor = NULL;
6565 if (*mc->mc_dbflag & DB_STALE) {
6566 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6571 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6574 size_t size = sizeof(MDB_cursor);
6576 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6579 if (txn->mt_flags & MDB_TXN_ERROR)
6582 /* Allow read access to the freelist */
6583 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6586 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6587 size += sizeof(MDB_xcursor);
6589 if ((mc = malloc(size)) != NULL) {
6590 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6591 if (txn->mt_cursors) {
6592 mc->mc_next = txn->mt_cursors[dbi];
6593 txn->mt_cursors[dbi] = mc;
6594 mc->mc_flags |= C_UNTRACK;
6606 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6608 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6611 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6614 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6618 /* Return the count of duplicate data items for the current key */
6620 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6624 if (mc == NULL || countp == NULL)
6627 if (mc->mc_xcursor == NULL)
6628 return MDB_INCOMPATIBLE;
6630 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6631 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6634 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6637 *countp = mc->mc_xcursor->mx_db.md_entries;
6643 mdb_cursor_close(MDB_cursor *mc)
6645 if (mc && !mc->mc_backup) {
6646 /* remove from txn, if tracked */
6647 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6648 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6649 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6651 *prev = mc->mc_next;
6658 mdb_cursor_txn(MDB_cursor *mc)
6660 if (!mc) return NULL;
6665 mdb_cursor_dbi(MDB_cursor *mc)
6671 /** Replace the key for a node with a new key.
6672 * @param[in] mc Cursor pointing to the node to operate on.
6673 * @param[in] key The new key to use.
6674 * @return 0 on success, non-zero on failure.
6677 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6684 indx_t ptr, i, numkeys, indx;
6687 indx = mc->mc_ki[mc->mc_top];
6688 mp = mc->mc_pg[mc->mc_top];
6689 node = NODEPTR(mp, indx);
6690 ptr = mp->mp_ptrs[indx];
6694 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6695 k2.mv_data = NODEKEY(node);
6696 k2.mv_size = node->mn_ksize;
6697 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6699 mdb_dkey(&k2, kbuf2),
6705 delta0 = delta = key->mv_size - node->mn_ksize;
6707 /* Must be 2-byte aligned. If new key is
6708 * shorter by 1, the shift will be skipped.
6710 delta += (delta & 1);
6712 if (delta > 0 && SIZELEFT(mp) < delta) {
6714 /* not enough space left, do a delete and split */
6715 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6716 pgno = NODEPGNO(node);
6717 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6718 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6721 numkeys = NUMKEYS(mp);
6722 for (i = 0; i < numkeys; i++) {
6723 if (mp->mp_ptrs[i] <= ptr)
6724 mp->mp_ptrs[i] -= delta;
6727 base = (char *)mp + mp->mp_upper;
6728 len = ptr - mp->mp_upper + NODESIZE;
6729 memmove(base - delta, base, len);
6730 mp->mp_upper -= delta;
6732 node = NODEPTR(mp, indx);
6735 /* But even if no shift was needed, update ksize */
6737 node->mn_ksize = key->mv_size;
6740 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6746 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6748 /** Move a node from csrc to cdst.
6751 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6758 unsigned short flags;
6762 /* Mark src and dst as dirty. */
6763 if ((rc = mdb_page_touch(csrc)) ||
6764 (rc = mdb_page_touch(cdst)))
6767 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6768 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6769 key.mv_size = csrc->mc_db->md_pad;
6770 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6772 data.mv_data = NULL;
6776 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6777 assert(!((long)srcnode&1));
6778 srcpg = NODEPGNO(srcnode);
6779 flags = srcnode->mn_flags;
6780 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6781 unsigned int snum = csrc->mc_snum;
6783 /* must find the lowest key below src */
6784 mdb_page_search_lowest(csrc);
6785 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6786 key.mv_size = csrc->mc_db->md_pad;
6787 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6789 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6790 key.mv_size = NODEKSZ(s2);
6791 key.mv_data = NODEKEY(s2);
6793 csrc->mc_snum = snum--;
6794 csrc->mc_top = snum;
6796 key.mv_size = NODEKSZ(srcnode);
6797 key.mv_data = NODEKEY(srcnode);
6799 data.mv_size = NODEDSZ(srcnode);
6800 data.mv_data = NODEDATA(srcnode);
6802 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6803 unsigned int snum = cdst->mc_snum;
6806 /* must find the lowest key below dst */
6807 mdb_page_search_lowest(cdst);
6808 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6809 bkey.mv_size = cdst->mc_db->md_pad;
6810 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6812 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6813 bkey.mv_size = NODEKSZ(s2);
6814 bkey.mv_data = NODEKEY(s2);
6816 cdst->mc_snum = snum--;
6817 cdst->mc_top = snum;
6818 mdb_cursor_copy(cdst, &mn);
6820 rc = mdb_update_key(&mn, &bkey);
6825 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6826 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6827 csrc->mc_ki[csrc->mc_top],
6829 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6830 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6832 /* Add the node to the destination page.
6834 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6835 if (rc != MDB_SUCCESS)
6838 /* Delete the node from the source page.
6840 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6843 /* Adjust other cursors pointing to mp */
6844 MDB_cursor *m2, *m3;
6845 MDB_dbi dbi = csrc->mc_dbi;
6846 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6848 if (csrc->mc_flags & C_SUB)
6851 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6852 if (csrc->mc_flags & C_SUB)
6853 m3 = &m2->mc_xcursor->mx_cursor;
6856 if (m3 == csrc) continue;
6857 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6858 csrc->mc_ki[csrc->mc_top]) {
6859 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6860 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6865 /* Update the parent separators.
6867 if (csrc->mc_ki[csrc->mc_top] == 0) {
6868 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6869 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6870 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6872 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6873 key.mv_size = NODEKSZ(srcnode);
6874 key.mv_data = NODEKEY(srcnode);
6876 DPRINTF(("update separator for source page %"Z"u to [%s]",
6877 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6878 mdb_cursor_copy(csrc, &mn);
6881 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6884 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6886 indx_t ix = csrc->mc_ki[csrc->mc_top];
6887 nullkey.mv_size = 0;
6888 csrc->mc_ki[csrc->mc_top] = 0;
6889 rc = mdb_update_key(csrc, &nullkey);
6890 csrc->mc_ki[csrc->mc_top] = ix;
6891 assert(rc == MDB_SUCCESS);
6895 if (cdst->mc_ki[cdst->mc_top] == 0) {
6896 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6897 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6898 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6900 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6901 key.mv_size = NODEKSZ(srcnode);
6902 key.mv_data = NODEKEY(srcnode);
6904 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6905 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6906 mdb_cursor_copy(cdst, &mn);
6909 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6912 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6914 indx_t ix = cdst->mc_ki[cdst->mc_top];
6915 nullkey.mv_size = 0;
6916 cdst->mc_ki[cdst->mc_top] = 0;
6917 rc = mdb_update_key(cdst, &nullkey);
6918 cdst->mc_ki[cdst->mc_top] = ix;
6919 assert(rc == MDB_SUCCESS);
6926 /** Merge one page into another.
6927 * The nodes from the page pointed to by \b csrc will
6928 * be copied to the page pointed to by \b cdst and then
6929 * the \b csrc page will be freed.
6930 * @param[in] csrc Cursor pointing to the source page.
6931 * @param[in] cdst Cursor pointing to the destination page.
6934 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6942 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6943 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6945 assert(csrc->mc_snum > 1); /* can't merge root page */
6946 assert(cdst->mc_snum > 1);
6948 /* Mark dst as dirty. */
6949 if ((rc = mdb_page_touch(cdst)))
6952 /* Move all nodes from src to dst.
6954 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6955 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6956 key.mv_size = csrc->mc_db->md_pad;
6957 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6958 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6959 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6960 if (rc != MDB_SUCCESS)
6962 key.mv_data = (char *)key.mv_data + key.mv_size;
6965 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6966 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6967 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6968 unsigned int snum = csrc->mc_snum;
6970 /* must find the lowest key below src */
6971 mdb_page_search_lowest(csrc);
6972 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6973 key.mv_size = csrc->mc_db->md_pad;
6974 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6976 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6977 key.mv_size = NODEKSZ(s2);
6978 key.mv_data = NODEKEY(s2);
6980 csrc->mc_snum = snum--;
6981 csrc->mc_top = snum;
6983 key.mv_size = srcnode->mn_ksize;
6984 key.mv_data = NODEKEY(srcnode);
6987 data.mv_size = NODEDSZ(srcnode);
6988 data.mv_data = NODEDATA(srcnode);
6989 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6990 if (rc != MDB_SUCCESS)
6995 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
6996 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
6997 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
6999 /* Unlink the src page from parent and add to free list.
7001 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
7002 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7005 rc = mdb_update_key(csrc, &key);
7011 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7012 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7015 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7016 csrc->mc_db->md_leaf_pages--;
7018 csrc->mc_db->md_branch_pages--;
7020 /* Adjust other cursors pointing to mp */
7021 MDB_cursor *m2, *m3;
7022 MDB_dbi dbi = csrc->mc_dbi;
7023 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7025 if (csrc->mc_flags & C_SUB)
7028 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7029 if (csrc->mc_flags & C_SUB)
7030 m3 = &m2->mc_xcursor->mx_cursor;
7033 if (m3 == csrc) continue;
7034 if (m3->mc_snum < csrc->mc_snum) continue;
7035 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7036 m3->mc_pg[csrc->mc_top] = mp;
7037 m3->mc_ki[csrc->mc_top] += nkeys;
7041 mdb_cursor_pop(csrc);
7043 return mdb_rebalance(csrc);
7046 /** Copy the contents of a cursor.
7047 * @param[in] csrc The cursor to copy from.
7048 * @param[out] cdst The cursor to copy to.
7051 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7055 cdst->mc_txn = csrc->mc_txn;
7056 cdst->mc_dbi = csrc->mc_dbi;
7057 cdst->mc_db = csrc->mc_db;
7058 cdst->mc_dbx = csrc->mc_dbx;
7059 cdst->mc_snum = csrc->mc_snum;
7060 cdst->mc_top = csrc->mc_top;
7061 cdst->mc_flags = csrc->mc_flags;
7063 for (i=0; i<csrc->mc_snum; i++) {
7064 cdst->mc_pg[i] = csrc->mc_pg[i];
7065 cdst->mc_ki[i] = csrc->mc_ki[i];
7069 /** Rebalance the tree after a delete operation.
7070 * @param[in] mc Cursor pointing to the page where rebalancing
7072 * @return 0 on success, non-zero on failure.
7075 mdb_rebalance(MDB_cursor *mc)
7079 unsigned int ptop, minkeys;
7082 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7086 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7087 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7088 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7089 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7090 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7094 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7095 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7098 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7099 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7105 if (mc->mc_snum < 2) {
7106 MDB_page *mp = mc->mc_pg[0];
7108 DPUTS("Can't rebalance a subpage, ignoring");
7111 if (NUMKEYS(mp) == 0) {
7112 DPUTS("tree is completely empty");
7113 mc->mc_db->md_root = P_INVALID;
7114 mc->mc_db->md_depth = 0;
7115 mc->mc_db->md_leaf_pages = 0;
7116 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7119 /* Adjust cursors pointing to mp */
7123 MDB_cursor *m2, *m3;
7124 MDB_dbi dbi = mc->mc_dbi;
7126 if (mc->mc_flags & C_SUB)
7129 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7130 if (mc->mc_flags & C_SUB)
7131 m3 = &m2->mc_xcursor->mx_cursor;
7134 if (m3->mc_snum < mc->mc_snum) continue;
7135 if (m3->mc_pg[0] == mp) {
7141 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7142 DPUTS("collapsing root page!");
7143 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7146 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7147 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7150 mc->mc_db->md_depth--;
7151 mc->mc_db->md_branch_pages--;
7152 mc->mc_ki[0] = mc->mc_ki[1];
7154 /* Adjust other cursors pointing to mp */
7155 MDB_cursor *m2, *m3;
7156 MDB_dbi dbi = mc->mc_dbi;
7158 if (mc->mc_flags & C_SUB)
7161 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7162 if (mc->mc_flags & C_SUB)
7163 m3 = &m2->mc_xcursor->mx_cursor;
7166 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7167 if (m3->mc_pg[0] == mp) {
7168 m3->mc_pg[0] = mc->mc_pg[0];
7171 m3->mc_ki[0] = m3->mc_ki[1];
7176 DPUTS("root page doesn't need rebalancing");
7180 /* The parent (branch page) must have at least 2 pointers,
7181 * otherwise the tree is invalid.
7183 ptop = mc->mc_top-1;
7184 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7186 /* Leaf page fill factor is below the threshold.
7187 * Try to move keys from left or right neighbor, or
7188 * merge with a neighbor page.
7193 mdb_cursor_copy(mc, &mn);
7194 mn.mc_xcursor = NULL;
7196 if (mc->mc_ki[ptop] == 0) {
7197 /* We're the leftmost leaf in our parent.
7199 DPUTS("reading right neighbor");
7201 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7202 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7205 mn.mc_ki[mn.mc_top] = 0;
7206 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7208 /* There is at least one neighbor to the left.
7210 DPUTS("reading left neighbor");
7212 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7213 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7216 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7217 mc->mc_ki[mc->mc_top] = 0;
7220 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7221 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7222 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7224 /* If the neighbor page is above threshold and has enough keys,
7225 * move one key from it. Otherwise we should try to merge them.
7226 * (A branch page must never have less than 2 keys.)
7228 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7229 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7230 return mdb_node_move(&mn, mc);
7232 if (mc->mc_ki[ptop] == 0)
7233 rc = mdb_page_merge(&mn, mc);
7235 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7236 rc = mdb_page_merge(mc, &mn);
7237 mdb_cursor_copy(&mn, mc);
7239 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7244 /** Complete a delete operation started by #mdb_cursor_del(). */
7246 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7253 mp = mc->mc_pg[mc->mc_top];
7254 ki = mc->mc_ki[mc->mc_top];
7256 /* add overflow pages to free list */
7257 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7261 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7262 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7263 (rc = mdb_ovpage_free(mc, omp)))
7266 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7267 mc->mc_db->md_entries--;
7268 rc = mdb_rebalance(mc);
7269 if (rc != MDB_SUCCESS)
7270 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7273 MDB_dbi dbi = mc->mc_dbi;
7275 mp = mc->mc_pg[mc->mc_top];
7276 nkeys = NUMKEYS(mp);
7278 /* if mc points past last node in page, find next sibling */
7279 if (mc->mc_ki[mc->mc_top] >= nkeys)
7280 mdb_cursor_sibling(mc, 1);
7282 /* Adjust other cursors pointing to mp */
7283 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7286 if (!(m2->mc_flags & C_INITIALIZED))
7288 if (m2->mc_pg[mc->mc_top] == mp) {
7289 if (m2->mc_ki[mc->mc_top] >= ki) {
7290 m2->mc_flags |= C_DEL;
7291 if (m2->mc_ki[mc->mc_top] > ki)
7292 m2->mc_ki[mc->mc_top]--;
7294 if (m2->mc_ki[mc->mc_top] >= nkeys)
7295 mdb_cursor_sibling(m2, 1);
7298 mc->mc_flags |= C_DEL;
7305 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7306 MDB_val *key, MDB_val *data)
7311 MDB_val rdata, *xdata;
7315 assert(key != NULL);
7317 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7319 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7322 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7323 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7325 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7326 return MDB_BAD_VALSIZE;
7329 mdb_cursor_init(&mc, txn, dbi, &mx);
7332 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7333 /* must ignore any data */
7344 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7346 /* let mdb_page_split know about this cursor if needed:
7347 * delete will trigger a rebalance; if it needs to move
7348 * a node from one page to another, it will have to
7349 * update the parent's separator key(s). If the new sepkey
7350 * is larger than the current one, the parent page may
7351 * run out of space, triggering a split. We need this
7352 * cursor to be consistent until the end of the rebalance.
7354 mc.mc_flags |= C_UNTRACK;
7355 mc.mc_next = txn->mt_cursors[dbi];
7356 txn->mt_cursors[dbi] = &mc;
7357 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7358 txn->mt_cursors[dbi] = mc.mc_next;
7363 /** Split a page and insert a new node.
7364 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7365 * The cursor will be updated to point to the actual page and index where
7366 * the node got inserted after the split.
7367 * @param[in] newkey The key for the newly inserted node.
7368 * @param[in] newdata The data for the newly inserted node.
7369 * @param[in] newpgno The page number, if the new node is a branch node.
7370 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7371 * @return 0 on success, non-zero on failure.
7374 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7375 unsigned int nflags)
7378 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7381 unsigned int i, j, split_indx, nkeys, pmax;
7383 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7385 MDB_page *mp, *rp, *pp;
7390 mp = mc->mc_pg[mc->mc_top];
7391 newindx = mc->mc_ki[mc->mc_top];
7393 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7394 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7395 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7397 /* Create a right sibling. */
7398 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7400 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7402 if (mc->mc_snum < 2) {
7403 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7405 /* shift current top to make room for new parent */
7406 mc->mc_pg[1] = mc->mc_pg[0];
7407 mc->mc_ki[1] = mc->mc_ki[0];
7410 mc->mc_db->md_root = pp->mp_pgno;
7411 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7412 mc->mc_db->md_depth++;
7415 /* Add left (implicit) pointer. */
7416 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7417 /* undo the pre-push */
7418 mc->mc_pg[0] = mc->mc_pg[1];
7419 mc->mc_ki[0] = mc->mc_ki[1];
7420 mc->mc_db->md_root = mp->mp_pgno;
7421 mc->mc_db->md_depth--;
7428 ptop = mc->mc_top-1;
7429 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7432 mc->mc_flags |= C_SPLITTING;
7433 mdb_cursor_copy(mc, &mn);
7434 mn.mc_pg[mn.mc_top] = rp;
7435 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7437 if (nflags & MDB_APPEND) {
7438 mn.mc_ki[mn.mc_top] = 0;
7440 split_indx = newindx;
7445 nkeys = NUMKEYS(mp);
7446 split_indx = nkeys / 2;
7447 if (newindx < split_indx)
7453 unsigned int lsize, rsize, ksize;
7454 /* Move half of the keys to the right sibling */
7456 x = mc->mc_ki[mc->mc_top] - split_indx;
7457 ksize = mc->mc_db->md_pad;
7458 split = LEAF2KEY(mp, split_indx, ksize);
7459 rsize = (nkeys - split_indx) * ksize;
7460 lsize = (nkeys - split_indx) * sizeof(indx_t);
7461 mp->mp_lower -= lsize;
7462 rp->mp_lower += lsize;
7463 mp->mp_upper += rsize - lsize;
7464 rp->mp_upper -= rsize - lsize;
7465 sepkey.mv_size = ksize;
7466 if (newindx == split_indx) {
7467 sepkey.mv_data = newkey->mv_data;
7469 sepkey.mv_data = split;
7472 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7473 memcpy(rp->mp_ptrs, split, rsize);
7474 sepkey.mv_data = rp->mp_ptrs;
7475 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7476 memcpy(ins, newkey->mv_data, ksize);
7477 mp->mp_lower += sizeof(indx_t);
7478 mp->mp_upper -= ksize - sizeof(indx_t);
7481 memcpy(rp->mp_ptrs, split, x * ksize);
7482 ins = LEAF2KEY(rp, x, ksize);
7483 memcpy(ins, newkey->mv_data, ksize);
7484 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7485 rp->mp_lower += sizeof(indx_t);
7486 rp->mp_upper -= ksize - sizeof(indx_t);
7487 mc->mc_ki[mc->mc_top] = x;
7488 mc->mc_pg[mc->mc_top] = rp;
7493 /* For leaf pages, check the split point based on what
7494 * fits where, since otherwise mdb_node_add can fail.
7496 * This check is only needed when the data items are
7497 * relatively large, such that being off by one will
7498 * make the difference between success or failure.
7500 * It's also relevant if a page happens to be laid out
7501 * such that one half of its nodes are all "small" and
7502 * the other half of its nodes are "large." If the new
7503 * item is also "large" and falls on the half with
7504 * "large" nodes, it also may not fit.
7507 unsigned int psize, nsize;
7508 /* Maximum free space in an empty page */
7509 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7510 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7511 if ((nkeys < 20) || (nsize > pmax/16)) {
7512 if (newindx <= split_indx) {
7515 for (i=0; i<split_indx; i++) {
7516 node = NODEPTR(mp, i);
7517 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7518 if (F_ISSET(node->mn_flags, F_BIGDATA))
7519 psize += sizeof(pgno_t);
7521 psize += NODEDSZ(node);
7525 split_indx = newindx;
7536 for (i=nkeys-1; i>=split_indx; i--) {
7537 node = NODEPTR(mp, i);
7538 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7539 if (F_ISSET(node->mn_flags, F_BIGDATA))
7540 psize += sizeof(pgno_t);
7542 psize += NODEDSZ(node);
7546 split_indx = newindx;
7557 /* First find the separating key between the split pages.
7558 * The case where newindx == split_indx is ambiguous; the
7559 * new item could go to the new page or stay on the original
7560 * page. If newpos == 1 it goes to the new page.
7562 if (newindx == split_indx && newpos) {
7563 sepkey.mv_size = newkey->mv_size;
7564 sepkey.mv_data = newkey->mv_data;
7566 node = NODEPTR(mp, split_indx);
7567 sepkey.mv_size = node->mn_ksize;
7568 sepkey.mv_data = NODEKEY(node);
7572 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7574 /* Copy separator key to the parent.
7576 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7580 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7583 if (mn.mc_snum == mc->mc_snum) {
7584 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7585 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7586 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7587 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7592 /* Right page might now have changed parent.
7593 * Check if left page also changed parent.
7595 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7596 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7597 for (i=0; i<ptop; i++) {
7598 mc->mc_pg[i] = mn.mc_pg[i];
7599 mc->mc_ki[i] = mn.mc_ki[i];
7601 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7602 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7606 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7609 mc->mc_flags ^= C_SPLITTING;
7610 if (rc != MDB_SUCCESS) {
7613 if (nflags & MDB_APPEND) {
7614 mc->mc_pg[mc->mc_top] = rp;
7615 mc->mc_ki[mc->mc_top] = 0;
7616 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7619 for (i=0; i<mc->mc_top; i++)
7620 mc->mc_ki[i] = mn.mc_ki[i];
7627 /* Move half of the keys to the right sibling. */
7629 /* grab a page to hold a temporary copy */
7630 copy = mdb_page_malloc(mc->mc_txn, 1);
7634 copy->mp_pgno = mp->mp_pgno;
7635 copy->mp_flags = mp->mp_flags;
7636 copy->mp_lower = PAGEHDRSZ;
7637 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7638 mc->mc_pg[mc->mc_top] = copy;
7639 for (i = j = 0; i <= nkeys; j++) {
7640 if (i == split_indx) {
7641 /* Insert in right sibling. */
7642 /* Reset insert index for right sibling. */
7643 if (i != newindx || (newpos ^ ins_new)) {
7645 mc->mc_pg[mc->mc_top] = rp;
7649 if (i == newindx && !ins_new) {
7650 /* Insert the original entry that caused the split. */
7651 rkey.mv_data = newkey->mv_data;
7652 rkey.mv_size = newkey->mv_size;
7661 /* Update index for the new key. */
7662 mc->mc_ki[mc->mc_top] = j;
7663 } else if (i == nkeys) {
7666 node = NODEPTR(mp, i);
7667 rkey.mv_data = NODEKEY(node);
7668 rkey.mv_size = node->mn_ksize;
7670 xdata.mv_data = NODEDATA(node);
7671 xdata.mv_size = NODEDSZ(node);
7674 pgno = NODEPGNO(node);
7675 flags = node->mn_flags;
7680 if (!IS_LEAF(mp) && j == 0) {
7681 /* First branch index doesn't need key data. */
7685 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7689 nkeys = NUMKEYS(copy);
7690 for (i=0; i<nkeys; i++)
7691 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7692 mp->mp_lower = copy->mp_lower;
7693 mp->mp_upper = copy->mp_upper;
7694 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7695 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7697 /* reset back to original page */
7698 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7699 mc->mc_pg[mc->mc_top] = mp;
7700 if (nflags & MDB_RESERVE) {
7701 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7702 if (!(node->mn_flags & F_BIGDATA))
7703 newdata->mv_data = NODEDATA(node);
7707 /* Make sure mc_ki is still valid.
7709 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7710 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7711 for (i=0; i<ptop; i++) {
7712 mc->mc_pg[i] = mn.mc_pg[i];
7713 mc->mc_ki[i] = mn.mc_ki[i];
7715 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7716 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7720 /* return tmp page to freelist */
7721 mdb_page_free(mc->mc_txn->mt_env, copy);
7724 /* Adjust other cursors pointing to mp */
7725 MDB_cursor *m2, *m3;
7726 MDB_dbi dbi = mc->mc_dbi;
7727 int fixup = NUMKEYS(mp);
7729 if (mc->mc_flags & C_SUB)
7732 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7733 if (mc->mc_flags & C_SUB)
7734 m3 = &m2->mc_xcursor->mx_cursor;
7739 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7741 if (m3->mc_flags & C_SPLITTING)
7746 for (k=m3->mc_top; k>=0; k--) {
7747 m3->mc_ki[k+1] = m3->mc_ki[k];
7748 m3->mc_pg[k+1] = m3->mc_pg[k];
7750 if (m3->mc_ki[0] >= split_indx) {
7755 m3->mc_pg[0] = mc->mc_pg[0];
7759 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7760 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7761 m3->mc_ki[mc->mc_top]++;
7762 if (m3->mc_ki[mc->mc_top] >= fixup) {
7763 m3->mc_pg[mc->mc_top] = rp;
7764 m3->mc_ki[mc->mc_top] -= fixup;
7765 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7767 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7768 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7777 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7778 MDB_val *key, MDB_val *data, unsigned int flags)
7783 assert(key != NULL);
7784 assert(data != NULL);
7786 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7789 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7790 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7792 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7793 return MDB_BAD_VALSIZE;
7796 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7799 mdb_cursor_init(&mc, txn, dbi, &mx);
7800 return mdb_cursor_put(&mc, key, data, flags);
7804 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7806 if ((flag & CHANGEABLE) != flag)
7809 env->me_flags |= flag;
7811 env->me_flags &= ~flag;
7816 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7821 *arg = env->me_flags;
7826 mdb_env_get_path(MDB_env *env, const char **arg)
7831 *arg = env->me_path;
7835 /** Common code for #mdb_stat() and #mdb_env_stat().
7836 * @param[in] env the environment to operate in.
7837 * @param[in] db the #MDB_db record containing the stats to return.
7838 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7839 * @return 0, this function always succeeds.
7842 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7844 arg->ms_psize = env->me_psize;
7845 arg->ms_depth = db->md_depth;
7846 arg->ms_branch_pages = db->md_branch_pages;
7847 arg->ms_leaf_pages = db->md_leaf_pages;
7848 arg->ms_overflow_pages = db->md_overflow_pages;
7849 arg->ms_entries = db->md_entries;
7854 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7858 if (env == NULL || arg == NULL)
7861 toggle = mdb_env_pick_meta(env);
7863 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7867 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7871 if (env == NULL || arg == NULL)
7874 toggle = mdb_env_pick_meta(env);
7875 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7876 arg->me_mapsize = env->me_mapsize;
7877 arg->me_maxreaders = env->me_maxreaders;
7879 /* me_numreaders may be zero if this process never used any readers. Use
7880 * the shared numreader count if it exists.
7882 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7884 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7885 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7889 /** Set the default comparison functions for a database.
7890 * Called immediately after a database is opened to set the defaults.
7891 * The user can then override them with #mdb_set_compare() or
7892 * #mdb_set_dupsort().
7893 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7894 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7897 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7899 uint16_t f = txn->mt_dbs[dbi].md_flags;
7901 txn->mt_dbxs[dbi].md_cmp =
7902 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7903 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7905 txn->mt_dbxs[dbi].md_dcmp =
7906 !(f & MDB_DUPSORT) ? 0 :
7907 ((f & MDB_INTEGERDUP)
7908 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7909 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7912 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7917 int rc, dbflag, exact;
7918 unsigned int unused = 0;
7921 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7922 mdb_default_cmp(txn, FREE_DBI);
7925 if ((flags & VALID_FLAGS) != flags)
7927 if (txn->mt_flags & MDB_TXN_ERROR)
7933 if (flags & PERSISTENT_FLAGS) {
7934 uint16_t f2 = flags & PERSISTENT_FLAGS;
7935 /* make sure flag changes get committed */
7936 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7937 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7938 txn->mt_flags |= MDB_TXN_DIRTY;
7941 mdb_default_cmp(txn, MAIN_DBI);
7945 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7946 mdb_default_cmp(txn, MAIN_DBI);
7949 /* Is the DB already open? */
7951 for (i=2; i<txn->mt_numdbs; i++) {
7952 if (!txn->mt_dbxs[i].md_name.mv_size) {
7953 /* Remember this free slot */
7954 if (!unused) unused = i;
7957 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7958 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7964 /* If no free slot and max hit, fail */
7965 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7966 return MDB_DBS_FULL;
7968 /* Cannot mix named databases with some mainDB flags */
7969 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7970 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7972 /* Find the DB info */
7973 dbflag = DB_NEW|DB_VALID;
7976 key.mv_data = (void *)name;
7977 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7978 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7979 if (rc == MDB_SUCCESS) {
7980 /* make sure this is actually a DB */
7981 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7982 if (!(node->mn_flags & F_SUBDATA))
7983 return MDB_INCOMPATIBLE;
7984 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7985 /* Create if requested */
7987 data.mv_size = sizeof(MDB_db);
7988 data.mv_data = &dummy;
7989 memset(&dummy, 0, sizeof(dummy));
7990 dummy.md_root = P_INVALID;
7991 dummy.md_flags = flags & PERSISTENT_FLAGS;
7992 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7996 /* OK, got info, add to table */
7997 if (rc == MDB_SUCCESS) {
7998 unsigned int slot = unused ? unused : txn->mt_numdbs;
7999 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8000 txn->mt_dbxs[slot].md_name.mv_size = len;
8001 txn->mt_dbxs[slot].md_rel = NULL;
8002 txn->mt_dbflags[slot] = dbflag;
8003 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8005 mdb_default_cmp(txn, slot);
8014 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8016 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8019 if (txn->mt_dbflags[dbi] & DB_STALE) {
8022 /* Stale, must read the DB's root. cursor_init does it for us. */
8023 mdb_cursor_init(&mc, txn, dbi, &mx);
8025 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8028 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8031 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8033 ptr = env->me_dbxs[dbi].md_name.mv_data;
8034 env->me_dbxs[dbi].md_name.mv_data = NULL;
8035 env->me_dbxs[dbi].md_name.mv_size = 0;
8036 env->me_dbflags[dbi] = 0;
8040 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8042 /* We could return the flags for the FREE_DBI too but what's the point? */
8043 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8045 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8049 /** Add all the DB's pages to the free list.
8050 * @param[in] mc Cursor on the DB to free.
8051 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8052 * @return 0 on success, non-zero on failure.
8055 mdb_drop0(MDB_cursor *mc, int subs)
8059 rc = mdb_page_search(mc, NULL, 0);
8060 if (rc == MDB_SUCCESS) {
8061 MDB_txn *txn = mc->mc_txn;
8066 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8067 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8070 mdb_cursor_copy(mc, &mx);
8071 while (mc->mc_snum > 0) {
8072 MDB_page *mp = mc->mc_pg[mc->mc_top];
8073 unsigned n = NUMKEYS(mp);
8075 for (i=0; i<n; i++) {
8076 ni = NODEPTR(mp, i);
8077 if (ni->mn_flags & F_BIGDATA) {
8080 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8081 rc = mdb_page_get(txn, pg, &omp, NULL);
8084 assert(IS_OVERFLOW(omp));
8085 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8089 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8090 mdb_xcursor_init1(mc, ni);
8091 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8097 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8099 for (i=0; i<n; i++) {
8101 ni = NODEPTR(mp, i);
8104 mdb_midl_xappend(txn->mt_free_pgs, pg);
8109 mc->mc_ki[mc->mc_top] = i;
8110 rc = mdb_cursor_sibling(mc, 1);
8112 /* no more siblings, go back to beginning
8113 * of previous level.
8117 for (i=1; i<mc->mc_snum; i++) {
8119 mc->mc_pg[i] = mx.mc_pg[i];
8124 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8125 } else if (rc == MDB_NOTFOUND) {
8131 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8133 MDB_cursor *mc, *m2;
8136 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8139 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8142 rc = mdb_cursor_open(txn, dbi, &mc);
8146 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8147 /* Invalidate the dropped DB's cursors */
8148 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8149 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8153 /* Can't delete the main DB */
8154 if (del && dbi > MAIN_DBI) {
8155 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8157 txn->mt_dbflags[dbi] = DB_STALE;
8158 mdb_dbi_close(txn->mt_env, dbi);
8161 /* reset the DB record, mark it dirty */
8162 txn->mt_dbflags[dbi] |= DB_DIRTY;
8163 txn->mt_dbs[dbi].md_depth = 0;
8164 txn->mt_dbs[dbi].md_branch_pages = 0;
8165 txn->mt_dbs[dbi].md_leaf_pages = 0;
8166 txn->mt_dbs[dbi].md_overflow_pages = 0;
8167 txn->mt_dbs[dbi].md_entries = 0;
8168 txn->mt_dbs[dbi].md_root = P_INVALID;
8170 txn->mt_flags |= MDB_TXN_DIRTY;
8173 mdb_cursor_close(mc);
8177 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8179 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8182 txn->mt_dbxs[dbi].md_cmp = cmp;
8186 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8188 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8191 txn->mt_dbxs[dbi].md_dcmp = cmp;
8195 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8197 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8200 txn->mt_dbxs[dbi].md_rel = rel;
8204 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8206 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8209 txn->mt_dbxs[dbi].md_relctx = ctx;
8213 int mdb_env_get_maxkeysize(MDB_env *env)
8215 return MDB_MAXKEYSIZE;
8218 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8220 unsigned int i, rdrs;
8227 if (!env->me_txns) {
8228 return func("(no reader locks)\n", ctx);
8230 rdrs = env->me_txns->mti_numreaders;
8231 mr = env->me_txns->mti_readers;
8232 for (i=0; i<rdrs; i++) {
8237 if (mr[i].mr_txnid == (txnid_t)-1) {
8238 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8240 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8244 func(" pid thread txnid\n", ctx);
8246 rc = func(buf, ctx);
8252 func("(no active readers)\n", ctx);
8257 /** Insert pid into list if not already present.
8258 * return -1 if already present.
8260 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8262 /* binary search of pid in list */
8264 unsigned cursor = 1;
8266 unsigned n = ids[0];
8269 unsigned pivot = n >> 1;
8270 cursor = base + pivot + 1;
8271 val = pid - ids[cursor];
8276 } else if ( val > 0 ) {
8281 /* found, so it's a duplicate */
8290 for (n = ids[0]; n > cursor; n--)
8296 int mdb_reader_check(MDB_env *env, int *dead)
8298 unsigned int i, j, rdrs;
8309 rdrs = env->me_txns->mti_numreaders;
8310 pids = malloc((rdrs+1) * sizeof(pid_t));
8314 mr = env->me_txns->mti_readers;
8316 for (i=0; i<rdrs; i++) {
8317 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8319 if (mdb_pid_insert(pids, pid) == 0) {
8320 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8322 /* Recheck, a new process may have reused pid */
8323 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8324 for (j=i; j<rdrs; j++)
8325 if (mr[j].mr_pid == pid) {
8326 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8327 (unsigned) pid, mr[j].mr_txnid));
8332 UNLOCK_MUTEX_R(env);