2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2015 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.
42 /* We use native NT APIs to setup the memory map, so that we can
43 * let the DB file grow incrementally instead of always preallocating
44 * the full size. These APIs are defined in <wdm.h> and <ntifs.h>
45 * but those headers are meant for driver-level development and
46 * conflict with the regular user-level headers, so we explicitly
47 * declare them here. Using these APIs also means we must link to
48 * ntdll.dll, which is not linked by default in user code.
51 NtCreateSection(OUT PHANDLE sh, IN ACCESS_MASK acc,
52 IN void * oa OPTIONAL,
53 IN PLARGE_INTEGER ms OPTIONAL,
54 IN ULONG pp, IN ULONG aa, IN HANDLE fh OPTIONAL);
56 typedef enum _SECTION_INHERIT {
62 NtMapViewOfSection(IN PHANDLE sh, IN HANDLE ph,
63 IN OUT PVOID *addr, IN ULONG_PTR zbits,
64 IN SIZE_T cs, IN OUT PLARGE_INTEGER off OPTIONAL,
65 IN OUT PSIZE_T vs, IN SECTION_INHERIT ih,
66 IN ULONG at, IN ULONG pp);
71 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
72 * as int64 which is wrong. MSVC doesn't define it at all, so just
76 #define MDB_THR_T DWORD
77 #include <sys/types.h>
80 # include <sys/param.h>
82 # define LITTLE_ENDIAN 1234
83 # define BIG_ENDIAN 4321
84 # define BYTE_ORDER LITTLE_ENDIAN
86 # define SSIZE_MAX INT_MAX
90 #include <sys/types.h>
92 #define MDB_PID_T pid_t
93 #define MDB_THR_T pthread_t
94 #include <sys/param.h>
97 #ifdef HAVE_SYS_FILE_H
103 #if defined(__mips) && defined(__linux)
104 /* MIPS has cache coherency issues, requires explicit cache control */
105 #include <asm/cachectl.h>
106 extern int cacheflush(char *addr, int nbytes, int cache);
107 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
109 #define CACHEFLUSH(addr, bytes, cache)
112 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
113 /** fdatasync is broken on ext3/ext4fs on older kernels, see
114 * description in #mdb_env_open2 comments. You can safely
115 * define MDB_FDATASYNC_WORKS if this code will only be run
116 * on kernels 3.6 and newer.
118 #define BROKEN_FDATASYNC
124 #include <inttypes.h>
132 typedef SSIZE_T ssize_t;
137 #if defined(__sun) || defined(ANDROID)
138 /* Most platforms have posix_memalign, older may only have memalign */
139 #define HAVE_MEMALIGN 1
143 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
144 #include <netinet/in.h>
145 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
148 #if defined(__APPLE__) || defined (BSD)
149 # if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
150 # define MDB_USE_SYSV_SEM 1
152 # define MDB_FDATASYNC fsync
153 #elif defined(ANDROID)
154 # define MDB_FDATASYNC fsync
159 #ifdef MDB_USE_POSIX_SEM
160 # define MDB_USE_HASH 1
161 #include <semaphore.h>
162 #elif defined(MDB_USE_SYSV_SEM)
165 #ifdef _SEM_SEMUN_UNDEFINED
168 struct semid_ds *buf;
169 unsigned short *array;
171 #endif /* _SEM_SEMUN_UNDEFINED */
173 #define MDB_USE_POSIX_MUTEX 1
174 #endif /* MDB_USE_POSIX_SEM */
177 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
178 + defined(MDB_USE_POSIX_MUTEX) != 1
179 # error "Ambiguous shared-lock implementation"
183 #include <valgrind/memcheck.h>
184 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
185 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
186 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
187 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
188 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
190 #define VGMEMP_CREATE(h,r,z)
191 #define VGMEMP_ALLOC(h,a,s)
192 #define VGMEMP_FREE(h,a)
193 #define VGMEMP_DESTROY(h)
194 #define VGMEMP_DEFINED(a,s)
198 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
199 /* Solaris just defines one or the other */
200 # define LITTLE_ENDIAN 1234
201 # define BIG_ENDIAN 4321
202 # ifdef _LITTLE_ENDIAN
203 # define BYTE_ORDER LITTLE_ENDIAN
205 # define BYTE_ORDER BIG_ENDIAN
208 # define BYTE_ORDER __BYTE_ORDER
212 #ifndef LITTLE_ENDIAN
213 #define LITTLE_ENDIAN __LITTLE_ENDIAN
216 #define BIG_ENDIAN __BIG_ENDIAN
219 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
220 #define MISALIGNED_OK 1
226 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
227 # error "Unknown or unsupported endianness (BYTE_ORDER)"
228 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
229 # error "Two's complement, reasonably sized integer types, please"
233 /** Put infrequently used env functions in separate section */
235 # define ESECT __attribute__ ((section("__TEXT,text_env")))
237 # define ESECT __attribute__ ((section("text_env")))
244 #define CALL_CONV WINAPI
249 /** @defgroup internal LMDB Internals
252 /** @defgroup compat Compatibility Macros
253 * A bunch of macros to minimize the amount of platform-specific ifdefs
254 * needed throughout the rest of the code. When the features this library
255 * needs are similar enough to POSIX to be hidden in a one-or-two line
256 * replacement, this macro approach is used.
260 /** Features under development */
265 /** Wrapper around __func__, which is a C99 feature */
266 #if __STDC_VERSION__ >= 199901L
267 # define mdb_func_ __func__
268 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
269 # define mdb_func_ __FUNCTION__
271 /* If a debug message says <mdb_unknown>(), update the #if statements above */
272 # define mdb_func_ "<mdb_unknown>"
275 /* Internal error codes, not exposed outside liblmdb */
276 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
278 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
279 #elif defined MDB_USE_SYSV_SEM
280 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
281 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
282 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
286 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
288 /** Some platforms define the EOWNERDEAD error code
289 * even though they don't support Robust Mutexes.
290 * Compile with -DMDB_USE_ROBUST=0, or use some other
291 * mechanism like -DMDB_USE_SYSV_SEM instead of
292 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
293 * also Robust, but some systems don't support them
296 #ifndef MDB_USE_ROBUST
297 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
298 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
299 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
300 # define MDB_USE_ROBUST 0
302 # define MDB_USE_ROBUST 1
303 /* glibc < 2.12 only provided _np API */
304 # if defined(__GLIBC__) && GLIBC_VER < 0x02000c
305 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
306 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
307 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
310 #endif /* MDB_USE_ROBUST */
312 #if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
313 #define MDB_ROBUST_SUPPORTED 1
317 #define MDB_USE_HASH 1
318 #define MDB_PIDLOCK 0
319 #define THREAD_RET DWORD
320 #define pthread_t HANDLE
321 #define pthread_mutex_t HANDLE
322 #define pthread_cond_t HANDLE
323 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
324 #define pthread_key_t DWORD
325 #define pthread_self() GetCurrentThreadId()
326 #define pthread_key_create(x,y) \
327 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
328 #define pthread_key_delete(x) TlsFree(x)
329 #define pthread_getspecific(x) TlsGetValue(x)
330 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
331 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
332 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
333 #define pthread_cond_signal(x) SetEvent(*x)
334 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
335 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
336 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
337 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
338 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
339 #define mdb_mutex_consistent(mutex) 0
340 #define getpid() GetCurrentProcessId()
341 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
342 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
343 #define ErrCode() GetLastError()
344 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
345 #define close(fd) (CloseHandle(fd) ? 0 : -1)
346 #define munmap(ptr,len) UnmapViewOfFile(ptr)
347 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
348 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
350 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
354 #define THREAD_RET void *
355 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
356 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
357 #define Z "z" /**< printf format modifier for size_t */
359 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
360 #define MDB_PIDLOCK 1
362 #ifdef MDB_USE_POSIX_SEM
364 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
365 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
366 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
369 mdb_sem_wait(sem_t *sem)
372 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
376 #elif defined MDB_USE_SYSV_SEM
378 typedef struct mdb_mutex {
382 } mdb_mutex_t[1], *mdb_mutexref_t;
384 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
385 #define UNLOCK_MUTEX(mutex) do { \
386 struct sembuf sb = { 0, 1, SEM_UNDO }; \
387 sb.sem_num = (mutex)->semnum; \
388 *(mutex)->locked = 0; \
389 semop((mutex)->semid, &sb, 1); \
393 mdb_sem_wait(mdb_mutexref_t sem)
395 int rc, *locked = sem->locked;
396 struct sembuf sb = { 0, -1, SEM_UNDO };
397 sb.sem_num = sem->semnum;
399 if (!semop(sem->semid, &sb, 1)) {
400 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
404 } while ((rc = errno) == EINTR);
408 #define mdb_mutex_consistent(mutex) 0
410 #else /* MDB_USE_POSIX_MUTEX: */
411 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
412 * local variables keep it (mdb_mutexref_t).
414 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
415 * be the same, or an array[size 1] and a pointer.
418 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
420 /** Lock the reader or writer mutex.
421 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
423 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
424 /** Unlock the reader or writer mutex.
426 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
427 /** Mark mutex-protected data as repaired, after death of previous owner.
429 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
430 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
432 /** Get the error code for the last failed system function.
434 #define ErrCode() errno
436 /** An abstraction for a file handle.
437 * On POSIX systems file handles are small integers. On Windows
438 * they're opaque pointers.
442 /** A value for an invalid file handle.
443 * Mainly used to initialize file variables and signify that they are
446 #define INVALID_HANDLE_VALUE (-1)
448 /** Get the size of a memory page for the system.
449 * This is the basic size that the platform's memory manager uses, and is
450 * fundamental to the use of memory-mapped files.
452 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
455 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
457 #elif defined(MDB_USE_SYSV_SEM)
458 #define MNAME_LEN (sizeof(int))
460 #define MNAME_LEN (sizeof(pthread_mutex_t))
463 #ifdef MDB_USE_SYSV_SEM
464 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
466 #define SYSV_SEM_FLAG 0
471 #ifdef MDB_ROBUST_SUPPORTED
472 /** Lock mutex, handle any error, set rc = result.
473 * Return 0 on success, nonzero (not rc) on error.
475 #define LOCK_MUTEX(rc, env, mutex) \
476 (((rc) = LOCK_MUTEX0(mutex)) && \
477 ((rc) = mdb_mutex_failed(env, mutex, rc)))
478 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
480 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
481 #define mdb_mutex_failed(env, mutex, rc) (rc)
485 /** A flag for opening a file and requesting synchronous data writes.
486 * This is only used when writing a meta page. It's not strictly needed;
487 * we could just do a normal write and then immediately perform a flush.
488 * But if this flag is available it saves us an extra system call.
490 * @note If O_DSYNC is undefined but exists in /usr/include,
491 * preferably set some compiler flag to get the definition.
495 # define MDB_DSYNC O_DSYNC
497 # define MDB_DSYNC O_SYNC
502 /** Function for flushing the data of a file. Define this to fsync
503 * if fdatasync() is not supported.
505 #ifndef MDB_FDATASYNC
506 # define MDB_FDATASYNC fdatasync
510 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
521 /** A page number in the database.
522 * Note that 64 bit page numbers are overkill, since pages themselves
523 * already represent 12-13 bits of addressable memory, and the OS will
524 * always limit applications to a maximum of 63 bits of address space.
526 * @note In the #MDB_node structure, we only store 48 bits of this value,
527 * which thus limits us to only 60 bits of addressable data.
529 typedef MDB_ID pgno_t;
531 /** A transaction ID.
532 * See struct MDB_txn.mt_txnid for details.
534 typedef MDB_ID txnid_t;
536 /** @defgroup debug Debug Macros
540 /** Enable debug output. Needs variable argument macros (a C99 feature).
541 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
542 * read from and written to the database (used for free space management).
548 static int mdb_debug;
549 static txnid_t mdb_debug_start;
551 /** Print a debug message with printf formatting.
552 * Requires double parenthesis around 2 or more args.
554 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
555 # define DPRINTF0(fmt, ...) \
556 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
558 # define DPRINTF(args) ((void) 0)
560 /** Print a debug string.
561 * The string is printed literally, with no format processing.
563 #define DPUTS(arg) DPRINTF(("%s", arg))
564 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
566 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
569 /** @brief The maximum size of a database page.
571 * It is 32k or 64k, since value-PAGEBASE must fit in
572 * #MDB_page.%mp_upper.
574 * LMDB will use database pages < OS pages if needed.
575 * That causes more I/O in write transactions: The OS must
576 * know (read) the whole page before writing a partial page.
578 * Note that we don't currently support Huge pages. On Linux,
579 * regular data files cannot use Huge pages, and in general
580 * Huge pages aren't actually pageable. We rely on the OS
581 * demand-pager to read our data and page it out when memory
582 * pressure from other processes is high. So until OSs have
583 * actual paging support for Huge pages, they're not viable.
585 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
587 /** The minimum number of keys required in a database page.
588 * Setting this to a larger value will place a smaller bound on the
589 * maximum size of a data item. Data items larger than this size will
590 * be pushed into overflow pages instead of being stored directly in
591 * the B-tree node. This value used to default to 4. With a page size
592 * of 4096 bytes that meant that any item larger than 1024 bytes would
593 * go into an overflow page. That also meant that on average 2-3KB of
594 * each overflow page was wasted space. The value cannot be lower than
595 * 2 because then there would no longer be a tree structure. With this
596 * value, items larger than 2KB will go into overflow pages, and on
597 * average only 1KB will be wasted.
599 #define MDB_MINKEYS 2
601 /** A stamp that identifies a file as an LMDB file.
602 * There's nothing special about this value other than that it is easily
603 * recognizable, and it will reflect any byte order mismatches.
605 #define MDB_MAGIC 0xBEEFC0DE
607 /** The version number for a database's datafile format. */
608 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
609 /** The version number for a database's lockfile format. */
610 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
612 /** @brief The max size of a key we can write, or 0 for computed max.
614 * This macro should normally be left alone or set to 0.
615 * Note that a database with big keys or dupsort data cannot be
616 * reliably modified by a liblmdb which uses a smaller max.
617 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
619 * Other values are allowed, for backwards compat. However:
620 * A value bigger than the computed max can break if you do not
621 * know what you are doing, and liblmdb <= 0.9.10 can break when
622 * modifying a DB with keys/dupsort data bigger than its max.
624 * Data items in an #MDB_DUPSORT database are also limited to
625 * this size, since they're actually keys of a sub-DB. Keys and
626 * #MDB_DUPSORT data items must fit on a node in a regular page.
628 #ifndef MDB_MAXKEYSIZE
629 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
632 /** The maximum size of a key we can write to the environment. */
634 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
636 #define ENV_MAXKEY(env) ((env)->me_maxkey)
639 /** @brief The maximum size of a data item.
641 * We only store a 32 bit value for node sizes.
643 #define MAXDATASIZE 0xffffffffUL
646 /** Key size which fits in a #DKBUF.
649 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
652 * This is used for printing a hex dump of a key's contents.
654 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
655 /** Display a key in hex.
657 * Invoke a function to display a key in hex.
659 #define DKEY(x) mdb_dkey(x, kbuf)
665 /** An invalid page number.
666 * Mainly used to denote an empty tree.
668 #define P_INVALID (~(pgno_t)0)
670 /** Test if the flags \b f are set in a flag word \b w. */
671 #define F_ISSET(w, f) (((w) & (f)) == (f))
673 /** Round \b n up to an even number. */
674 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
676 /** Used for offsets within a single page.
677 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
680 typedef uint16_t indx_t;
682 /** Default size of memory map.
683 * This is certainly too small for any actual applications. Apps should always set
684 * the size explicitly using #mdb_env_set_mapsize().
686 #define DEFAULT_MAPSIZE 1048576
688 /** @defgroup readers Reader Lock Table
689 * Readers don't acquire any locks for their data access. Instead, they
690 * simply record their transaction ID in the reader table. The reader
691 * mutex is needed just to find an empty slot in the reader table. The
692 * slot's address is saved in thread-specific data so that subsequent read
693 * transactions started by the same thread need no further locking to proceed.
695 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
697 * No reader table is used if the database is on a read-only filesystem, or
698 * if #MDB_NOLOCK is set.
700 * Since the database uses multi-version concurrency control, readers don't
701 * actually need any locking. This table is used to keep track of which
702 * readers are using data from which old transactions, so that we'll know
703 * when a particular old transaction is no longer in use. Old transactions
704 * that have discarded any data pages can then have those pages reclaimed
705 * for use by a later write transaction.
707 * The lock table is constructed such that reader slots are aligned with the
708 * processor's cache line size. Any slot is only ever used by one thread.
709 * This alignment guarantees that there will be no contention or cache
710 * thrashing as threads update their own slot info, and also eliminates
711 * any need for locking when accessing a slot.
713 * A writer thread will scan every slot in the table to determine the oldest
714 * outstanding reader transaction. Any freed pages older than this will be
715 * reclaimed by the writer. The writer doesn't use any locks when scanning
716 * this table. This means that there's no guarantee that the writer will
717 * see the most up-to-date reader info, but that's not required for correct
718 * operation - all we need is to know the upper bound on the oldest reader,
719 * we don't care at all about the newest reader. So the only consequence of
720 * reading stale information here is that old pages might hang around a
721 * while longer before being reclaimed. That's actually good anyway, because
722 * the longer we delay reclaiming old pages, the more likely it is that a
723 * string of contiguous pages can be found after coalescing old pages from
724 * many old transactions together.
727 /** Number of slots in the reader table.
728 * This value was chosen somewhat arbitrarily. 126 readers plus a
729 * couple mutexes fit exactly into 8KB on my development machine.
730 * Applications should set the table size using #mdb_env_set_maxreaders().
732 #define DEFAULT_READERS 126
734 /** The size of a CPU cache line in bytes. We want our lock structures
735 * aligned to this size to avoid false cache line sharing in the
737 * This value works for most CPUs. For Itanium this should be 128.
743 /** The information we store in a single slot of the reader table.
744 * In addition to a transaction ID, we also record the process and
745 * thread ID that owns a slot, so that we can detect stale information,
746 * e.g. threads or processes that went away without cleaning up.
747 * @note We currently don't check for stale records. We simply re-init
748 * the table when we know that we're the only process opening the
751 typedef struct MDB_rxbody {
752 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
753 * Multiple readers that start at the same time will probably have the
754 * same ID here. Again, it's not important to exclude them from
755 * anything; all we need to know is which version of the DB they
756 * started from so we can avoid overwriting any data used in that
757 * particular version.
759 volatile txnid_t mrb_txnid;
760 /** The process ID of the process owning this reader txn. */
761 volatile MDB_PID_T mrb_pid;
762 /** The thread ID of the thread owning this txn. */
763 volatile MDB_THR_T mrb_tid;
766 /** The actual reader record, with cacheline padding. */
767 typedef struct MDB_reader {
770 /** shorthand for mrb_txnid */
771 #define mr_txnid mru.mrx.mrb_txnid
772 #define mr_pid mru.mrx.mrb_pid
773 #define mr_tid mru.mrx.mrb_tid
774 /** cache line alignment */
775 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
779 /** The header for the reader table.
780 * The table resides in a memory-mapped file. (This is a different file
781 * than is used for the main database.)
783 * For POSIX the actual mutexes reside in the shared memory of this
784 * mapped file. On Windows, mutexes are named objects allocated by the
785 * kernel; we store the mutex names in this mapped file so that other
786 * processes can grab them. This same approach is also used on
787 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
788 * process-shared POSIX mutexes. For these cases where a named object
789 * is used, the object name is derived from a 64 bit FNV hash of the
790 * environment pathname. As such, naming collisions are extremely
791 * unlikely. If a collision occurs, the results are unpredictable.
793 typedef struct MDB_txbody {
794 /** Stamp identifying this as an LMDB file. It must be set
797 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
799 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
800 char mtb_rmname[MNAME_LEN];
801 #elif defined(MDB_USE_SYSV_SEM)
805 /** Mutex protecting access to this table.
806 * This is the reader table lock used with LOCK_MUTEX().
808 mdb_mutex_t mtb_rmutex;
810 /** The ID of the last transaction committed to the database.
811 * This is recorded here only for convenience; the value can always
812 * be determined by reading the main database meta pages.
814 volatile txnid_t mtb_txnid;
815 /** The number of slots that have been used in the reader table.
816 * This always records the maximum count, it is not decremented
817 * when readers release their slots.
819 volatile unsigned mtb_numreaders;
822 /** The actual reader table definition. */
823 typedef struct MDB_txninfo {
826 #define mti_magic mt1.mtb.mtb_magic
827 #define mti_format mt1.mtb.mtb_format
828 #define mti_rmutex mt1.mtb.mtb_rmutex
829 #define mti_rmname mt1.mtb.mtb_rmname
830 #define mti_txnid mt1.mtb.mtb_txnid
831 #define mti_numreaders mt1.mtb.mtb_numreaders
832 #ifdef MDB_USE_SYSV_SEM
833 #define mti_semid mt1.mtb.mtb_semid
834 #define mti_rlocked mt1.mtb.mtb_rlocked
836 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
839 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
840 char mt2_wmname[MNAME_LEN];
841 #define mti_wmname mt2.mt2_wmname
842 #elif defined MDB_USE_SYSV_SEM
844 #define mti_wlocked mt2.mt2_wlocked
846 mdb_mutex_t mt2_wmutex;
847 #define mti_wmutex mt2.mt2_wmutex
849 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
851 MDB_reader mti_readers[1];
854 /** Lockfile format signature: version, features and field layout */
855 #define MDB_LOCK_FORMAT \
857 ((MDB_LOCK_VERSION) \
858 /* Flags which describe functionality */ \
859 + (SYSV_SEM_FLAG << 18) \
860 + (((MDB_PIDLOCK) != 0) << 16)))
863 /** Common header for all page types.
864 * Overflow records occupy a number of contiguous pages with no
865 * headers on any page after the first.
867 typedef struct MDB_page {
868 #define mp_pgno mp_p.p_pgno
869 #define mp_next mp_p.p_next
871 pgno_t p_pgno; /**< page number */
872 struct MDB_page *p_next; /**< for in-memory list of freed pages */
875 /** @defgroup mdb_page Page Flags
877 * Flags for the page headers.
880 #define P_BRANCH 0x01 /**< branch page */
881 #define P_LEAF 0x02 /**< leaf page */
882 #define P_OVERFLOW 0x04 /**< overflow page */
883 #define P_META 0x08 /**< meta page */
884 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
885 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
886 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
887 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
888 #define P_KEEP 0x8000 /**< leave this page alone during spill */
890 uint16_t mp_flags; /**< @ref mdb_page */
891 #define mp_lower mp_pb.pb.pb_lower
892 #define mp_upper mp_pb.pb.pb_upper
893 #define mp_pages mp_pb.pb_pages
896 indx_t pb_lower; /**< lower bound of free space */
897 indx_t pb_upper; /**< upper bound of free space */
899 uint32_t pb_pages; /**< number of overflow pages */
901 indx_t mp_ptrs[1]; /**< dynamic size */
904 /** Size of the page header, excluding dynamic data at the end */
905 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
907 /** Address of first usable data byte in a page, after the header */
908 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
910 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
911 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
913 /** Number of nodes on a page */
914 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
916 /** The amount of space remaining in the page */
917 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
919 /** The percentage of space used in the page, in tenths of a percent. */
920 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
921 ((env)->me_psize - PAGEHDRSZ))
922 /** The minimum page fill factor, in tenths of a percent.
923 * Pages emptier than this are candidates for merging.
925 #define FILL_THRESHOLD 250
927 /** Test if a page is a leaf page */
928 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
929 /** Test if a page is a LEAF2 page */
930 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
931 /** Test if a page is a branch page */
932 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
933 /** Test if a page is an overflow page */
934 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
935 /** Test if a page is a sub page */
936 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
938 /** The number of overflow pages needed to store the given size. */
939 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
941 /** Link in #MDB_txn.%mt_loose_pgs list */
942 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
944 /** Header for a single key/data pair within a page.
945 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
946 * We guarantee 2-byte alignment for 'MDB_node's.
948 typedef struct MDB_node {
949 /** lo and hi are used for data size on leaf nodes and for
950 * child pgno on branch nodes. On 64 bit platforms, flags
951 * is also used for pgno. (Branch nodes have no flags).
952 * They are in host byte order in case that lets some
953 * accesses be optimized into a 32-bit word access.
955 #if BYTE_ORDER == LITTLE_ENDIAN
956 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
958 unsigned short mn_hi, mn_lo;
960 /** @defgroup mdb_node Node Flags
962 * Flags for node headers.
965 #define F_BIGDATA 0x01 /**< data put on overflow page */
966 #define F_SUBDATA 0x02 /**< data is a sub-database */
967 #define F_DUPDATA 0x04 /**< data has duplicates */
969 /** valid flags for #mdb_node_add() */
970 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
973 unsigned short mn_flags; /**< @ref mdb_node */
974 unsigned short mn_ksize; /**< key size */
975 char mn_data[1]; /**< key and data are appended here */
978 /** Size of the node header, excluding dynamic data at the end */
979 #define NODESIZE offsetof(MDB_node, mn_data)
981 /** Bit position of top word in page number, for shifting mn_flags */
982 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
984 /** Size of a node in a branch page with a given key.
985 * This is just the node header plus the key, there is no data.
987 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
989 /** Size of a node in a leaf page with a given key and data.
990 * This is node header plus key plus data size.
992 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
994 /** Address of node \b i in page \b p */
995 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
997 /** Address of the key for the node */
998 #define NODEKEY(node) (void *)((node)->mn_data)
1000 /** Address of the data for a node */
1001 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
1003 /** Get the page number pointed to by a branch node */
1004 #define NODEPGNO(node) \
1005 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
1006 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
1007 /** Set the page number in a branch node */
1008 #define SETPGNO(node,pgno) do { \
1009 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
1010 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
1012 /** Get the size of the data in a leaf node */
1013 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
1014 /** Set the size of the data for a leaf node */
1015 #define SETDSZ(node,size) do { \
1016 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
1017 /** The size of a key in a node */
1018 #define NODEKSZ(node) ((node)->mn_ksize)
1020 /** Copy a page number from src to dst */
1021 #ifdef MISALIGNED_OK
1022 #define COPY_PGNO(dst,src) dst = src
1024 #if SIZE_MAX > 4294967295UL
1025 #define COPY_PGNO(dst,src) do { \
1026 unsigned short *s, *d; \
1027 s = (unsigned short *)&(src); \
1028 d = (unsigned short *)&(dst); \
1035 #define COPY_PGNO(dst,src) do { \
1036 unsigned short *s, *d; \
1037 s = (unsigned short *)&(src); \
1038 d = (unsigned short *)&(dst); \
1044 /** The address of a key in a LEAF2 page.
1045 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
1046 * There are no node headers, keys are stored contiguously.
1048 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
1050 /** Set the \b node's key into \b keyptr, if requested. */
1051 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
1052 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
1054 /** Set the \b node's key into \b key. */
1055 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
1057 /** Information about a single database in the environment. */
1058 typedef struct MDB_db {
1059 uint32_t md_pad; /**< also ksize for LEAF2 pages */
1060 uint16_t md_flags; /**< @ref mdb_dbi_open */
1061 uint16_t md_depth; /**< depth of this tree */
1062 pgno_t md_branch_pages; /**< number of internal pages */
1063 pgno_t md_leaf_pages; /**< number of leaf pages */
1064 pgno_t md_overflow_pages; /**< number of overflow pages */
1065 size_t md_entries; /**< number of data items */
1066 pgno_t md_root; /**< the root page of this tree */
1069 /** mdb_dbi_open flags */
1070 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1071 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1072 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1073 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1075 /** Handle for the DB used to track free pages. */
1077 /** Handle for the default DB. */
1079 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1082 /** Number of meta pages - also hardcoded elsewhere */
1085 /** Meta page content.
1086 * A meta page is the start point for accessing a database snapshot.
1087 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1089 typedef struct MDB_meta {
1090 /** Stamp identifying this as an LMDB file. It must be set
1093 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1094 uint32_t mm_version;
1095 void *mm_address; /**< address for fixed mapping */
1096 size_t mm_mapsize; /**< size of mmap region */
1097 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1098 /** The size of pages used in this DB */
1099 #define mm_psize mm_dbs[FREE_DBI].md_pad
1100 /** Any persistent environment flags. @ref mdb_env */
1101 #define mm_flags mm_dbs[FREE_DBI].md_flags
1102 pgno_t mm_last_pg; /**< last used page in file */
1103 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1106 /** Buffer for a stack-allocated meta page.
1107 * The members define size and alignment, and silence type
1108 * aliasing warnings. They are not used directly; that could
1109 * mean incorrectly using several union members in parallel.
1111 typedef union MDB_metabuf {
1114 char mm_pad[PAGEHDRSZ];
1119 /** Auxiliary DB info.
1120 * The information here is mostly static/read-only. There is
1121 * only a single copy of this record in the environment.
1123 typedef struct MDB_dbx {
1124 MDB_val md_name; /**< name of the database */
1125 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1126 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1127 MDB_rel_func *md_rel; /**< user relocate function */
1128 void *md_relctx; /**< user-provided context for md_rel */
1131 /** A database transaction.
1132 * Every operation requires a transaction handle.
1135 MDB_txn *mt_parent; /**< parent of a nested txn */
1136 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1138 pgno_t mt_next_pgno; /**< next unallocated page */
1139 /** The ID of this transaction. IDs are integers incrementing from 1.
1140 * Only committed write transactions increment the ID. If a transaction
1141 * aborts, the ID may be re-used by the next writer.
1144 MDB_env *mt_env; /**< the DB environment */
1145 /** The list of pages that became unused during this transaction.
1147 MDB_IDL mt_free_pgs;
1148 /** The list of loose pages that became unused and may be reused
1149 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1151 MDB_page *mt_loose_pgs;
1152 /* #Number of loose pages (#mt_loose_pgs) */
1154 /** The sorted list of dirty pages we temporarily wrote to disk
1155 * because the dirty list was full. page numbers in here are
1156 * shifted left by 1, deleted slots have the LSB set.
1158 MDB_IDL mt_spill_pgs;
1160 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1161 MDB_ID2L dirty_list;
1162 /** For read txns: This thread/txn's reader table slot, or NULL. */
1165 /** Array of records for each DB known in the environment. */
1167 /** Array of MDB_db records for each known DB */
1169 /** Array of sequence numbers for each DB handle */
1170 unsigned int *mt_dbiseqs;
1171 /** @defgroup mt_dbflag Transaction DB Flags
1175 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1176 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1177 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1178 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1179 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1181 /** In write txns, array of cursors for each DB */
1182 MDB_cursor **mt_cursors;
1183 /** Array of flags for each DB */
1184 unsigned char *mt_dbflags;
1185 /** Number of DB records in use, or 0 when the txn is finished.
1186 * This number only ever increments until the txn finishes; we
1187 * don't decrement it when individual DB handles are closed.
1191 /** @defgroup mdb_txn Transaction Flags
1195 /** #mdb_txn_begin() flags */
1196 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1197 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1198 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1199 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1200 /* internal txn flags */
1201 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1202 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1203 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1204 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1205 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1206 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1207 /** most operations on the txn are currently illegal */
1208 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1210 unsigned int mt_flags; /**< @ref mdb_txn */
1211 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1212 * Includes ancestor txns' dirty pages not hidden by other txns'
1213 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1214 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1216 unsigned int mt_dirty_room;
1219 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1220 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1221 * raise this on a 64 bit machine.
1223 #define CURSOR_STACK 32
1227 /** Cursors are used for all DB operations.
1228 * A cursor holds a path of (page pointer, key index) from the DB
1229 * root to a position in the DB, plus other state. #MDB_DUPSORT
1230 * cursors include an xcursor to the current data item. Write txns
1231 * track their cursors and keep them up to date when data moves.
1232 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1233 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1236 /** Next cursor on this DB in this txn */
1237 MDB_cursor *mc_next;
1238 /** Backup of the original cursor if this cursor is a shadow */
1239 MDB_cursor *mc_backup;
1240 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1241 struct MDB_xcursor *mc_xcursor;
1242 /** The transaction that owns this cursor */
1244 /** The database handle this cursor operates on */
1246 /** The database record for this cursor */
1248 /** The database auxiliary record for this cursor */
1250 /** The @ref mt_dbflag for this database */
1251 unsigned char *mc_dbflag;
1252 unsigned short mc_snum; /**< number of pushed pages */
1253 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1254 /** @defgroup mdb_cursor Cursor Flags
1256 * Cursor state flags.
1259 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1260 #define C_EOF 0x02 /**< No more data */
1261 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1262 #define C_DEL 0x08 /**< last op was a cursor_del */
1263 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1265 unsigned int mc_flags; /**< @ref mdb_cursor */
1266 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1267 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1270 /** Context for sorted-dup records.
1271 * We could have gone to a fully recursive design, with arbitrarily
1272 * deep nesting of sub-databases. But for now we only handle these
1273 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1275 typedef struct MDB_xcursor {
1276 /** A sub-cursor for traversing the Dup DB */
1277 MDB_cursor mx_cursor;
1278 /** The database record for this Dup DB */
1280 /** The auxiliary DB record for this Dup DB */
1282 /** The @ref mt_dbflag for this Dup DB */
1283 unsigned char mx_dbflag;
1286 /** State of FreeDB old pages, stored in the MDB_env */
1287 typedef struct MDB_pgstate {
1288 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1289 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1292 /** The database environment. */
1294 HANDLE me_fd; /**< The main data file */
1295 HANDLE me_lfd; /**< The lock file */
1296 HANDLE me_mfd; /**< just for writing the meta pages */
1297 /** Failed to update the meta page. Probably an I/O error. */
1298 #define MDB_FATAL_ERROR 0x80000000U
1299 /** Some fields are initialized. */
1300 #define MDB_ENV_ACTIVE 0x20000000U
1301 /** me_txkey is set */
1302 #define MDB_ENV_TXKEY 0x10000000U
1303 /** fdatasync is unreliable */
1304 #define MDB_FSYNCONLY 0x08000000U
1305 uint32_t me_flags; /**< @ref mdb_env */
1306 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1307 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1308 unsigned int me_maxreaders; /**< size of the reader table */
1309 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1310 volatile int me_close_readers;
1311 MDB_dbi me_numdbs; /**< number of DBs opened */
1312 MDB_dbi me_maxdbs; /**< size of the DB table */
1313 MDB_PID_T me_pid; /**< process ID of this env */
1314 char *me_path; /**< path to the DB files */
1315 char *me_map; /**< the memory map of the data file */
1316 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1317 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1318 void *me_pbuf; /**< scratch area for DUPSORT put() */
1319 MDB_txn *me_txn; /**< current write transaction */
1320 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1321 size_t me_mapsize; /**< size of the data memory map */
1322 off_t me_size; /**< current file size */
1323 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1324 MDB_dbx *me_dbxs; /**< array of static DB info */
1325 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1326 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1327 pthread_key_t me_txkey; /**< thread-key for readers */
1328 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1329 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1330 # define me_pglast me_pgstate.mf_pglast
1331 # define me_pghead me_pgstate.mf_pghead
1332 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1333 /** IDL of pages that became unused in a write txn */
1334 MDB_IDL me_free_pgs;
1335 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1336 MDB_ID2L me_dirty_list;
1337 /** Max number of freelist items that can fit in a single overflow page */
1339 /** Max size of a node on a page */
1340 unsigned int me_nodemax;
1341 #if !(MDB_MAXKEYSIZE)
1342 unsigned int me_maxkey; /**< max size of a key */
1344 int me_live_reader; /**< have liveness lock in reader table */
1346 int me_pidquery; /**< Used in OpenProcess */
1348 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1349 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1350 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1352 mdb_mutex_t me_rmutex;
1353 mdb_mutex_t me_wmutex;
1355 void *me_userctx; /**< User-settable context */
1356 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1359 /** Nested transaction */
1360 typedef struct MDB_ntxn {
1361 MDB_txn mnt_txn; /**< the transaction */
1362 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1365 /** max number of pages to commit in one writev() call */
1366 #define MDB_COMMIT_PAGES 64
1367 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1368 #undef MDB_COMMIT_PAGES
1369 #define MDB_COMMIT_PAGES IOV_MAX
1372 /** max bytes to write in one call */
1373 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1375 /** Check \b txn and \b dbi arguments to a function */
1376 #define TXN_DBI_EXIST(txn, dbi, validity) \
1377 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1379 /** Check for misused \b dbi handles */
1380 #define TXN_DBI_CHANGED(txn, dbi) \
1381 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1383 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1384 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1385 static int mdb_page_touch(MDB_cursor *mc);
1387 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1388 "reset-tmp", "fail-begin", "fail-beginchild"}
1390 /* mdb_txn_end operation number, for logging */
1391 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1392 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1394 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1395 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1396 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1397 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1398 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1400 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1401 static int mdb_page_search_root(MDB_cursor *mc,
1402 MDB_val *key, int modify);
1403 #define MDB_PS_MODIFY 1
1404 #define MDB_PS_ROOTONLY 2
1405 #define MDB_PS_FIRST 4
1406 #define MDB_PS_LAST 8
1407 static int mdb_page_search(MDB_cursor *mc,
1408 MDB_val *key, int flags);
1409 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1411 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1412 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1413 pgno_t newpgno, unsigned int nflags);
1415 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1416 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1417 static int mdb_env_write_meta(MDB_txn *txn);
1418 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1419 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1421 static void mdb_env_close0(MDB_env *env, int excl);
1423 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1424 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1425 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1426 static void mdb_node_del(MDB_cursor *mc, int ksize);
1427 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1428 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1429 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1430 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1431 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1433 static int mdb_rebalance(MDB_cursor *mc);
1434 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1436 static void mdb_cursor_pop(MDB_cursor *mc);
1437 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1439 static int mdb_cursor_del0(MDB_cursor *mc);
1440 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1441 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1442 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1443 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1444 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1446 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1447 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1449 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1450 static void mdb_xcursor_init0(MDB_cursor *mc);
1451 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1452 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1454 static int mdb_drop0(MDB_cursor *mc, int subs);
1455 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1456 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1459 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1462 /** Compare two items pointing at size_t's of unknown alignment. */
1463 #ifdef MISALIGNED_OK
1464 # define mdb_cmp_clong mdb_cmp_long
1466 # define mdb_cmp_clong mdb_cmp_cint
1470 static SECURITY_DESCRIPTOR mdb_null_sd;
1471 static SECURITY_ATTRIBUTES mdb_all_sa;
1472 static int mdb_sec_inited;
1474 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize);
1477 /** Return the library version info. */
1479 mdb_version(int *major, int *minor, int *patch)
1481 if (major) *major = MDB_VERSION_MAJOR;
1482 if (minor) *minor = MDB_VERSION_MINOR;
1483 if (patch) *patch = MDB_VERSION_PATCH;
1484 return MDB_VERSION_STRING;
1487 /** Table of descriptions for LMDB @ref errors */
1488 static char *const mdb_errstr[] = {
1489 "MDB_KEYEXIST: Key/data pair already exists",
1490 "MDB_NOTFOUND: No matching key/data pair found",
1491 "MDB_PAGE_NOTFOUND: Requested page not found",
1492 "MDB_CORRUPTED: Located page was wrong type",
1493 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1494 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1495 "MDB_INVALID: File is not an LMDB file",
1496 "MDB_MAP_FULL: Environment mapsize limit reached",
1497 "MDB_DBS_FULL: Environment maxdbs limit reached",
1498 "MDB_READERS_FULL: Environment maxreaders limit reached",
1499 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1500 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1501 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1502 "MDB_PAGE_FULL: Internal error - page has no more space",
1503 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1504 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1505 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1506 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1507 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1508 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1512 mdb_strerror(int err)
1515 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1516 * This works as long as no function between the call to mdb_strerror
1517 * and the actual use of the message uses more than 4K of stack.
1520 char buf[1024], *ptr = buf;
1524 return ("Successful return: 0");
1526 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1527 i = err - MDB_KEYEXIST;
1528 return mdb_errstr[i];
1532 /* These are the C-runtime error codes we use. The comment indicates
1533 * their numeric value, and the Win32 error they would correspond to
1534 * if the error actually came from a Win32 API. A major mess, we should
1535 * have used LMDB-specific error codes for everything.
1538 case ENOENT: /* 2, FILE_NOT_FOUND */
1539 case EIO: /* 5, ACCESS_DENIED */
1540 case ENOMEM: /* 12, INVALID_ACCESS */
1541 case EACCES: /* 13, INVALID_DATA */
1542 case EBUSY: /* 16, CURRENT_DIRECTORY */
1543 case EINVAL: /* 22, BAD_COMMAND */
1544 case ENOSPC: /* 28, OUT_OF_PAPER */
1545 return strerror(err);
1550 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1551 FORMAT_MESSAGE_IGNORE_INSERTS,
1552 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1555 return strerror(err);
1559 /** assert(3) variant in cursor context */
1560 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1561 /** assert(3) variant in transaction context */
1562 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1563 /** assert(3) variant in environment context */
1564 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1567 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1568 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1571 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1572 const char *func, const char *file, int line)
1575 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1576 file, line, expr_txt, func);
1577 if (env->me_assert_func)
1578 env->me_assert_func(env, buf);
1579 fprintf(stderr, "%s\n", buf);
1583 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1587 /** Return the page number of \b mp which may be sub-page, for debug output */
1589 mdb_dbg_pgno(MDB_page *mp)
1592 COPY_PGNO(ret, mp->mp_pgno);
1596 /** Display a key in hexadecimal and return the address of the result.
1597 * @param[in] key the key to display
1598 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1599 * @return The key in hexadecimal form.
1602 mdb_dkey(MDB_val *key, char *buf)
1605 unsigned char *c = key->mv_data;
1611 if (key->mv_size > DKBUF_MAXKEYSIZE)
1612 return "MDB_MAXKEYSIZE";
1613 /* may want to make this a dynamic check: if the key is mostly
1614 * printable characters, print it as-is instead of converting to hex.
1618 for (i=0; i<key->mv_size; i++)
1619 ptr += sprintf(ptr, "%02x", *c++);
1621 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1627 mdb_leafnode_type(MDB_node *n)
1629 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1630 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1631 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1634 /** Display all the keys in the page. */
1636 mdb_page_list(MDB_page *mp)
1638 pgno_t pgno = mdb_dbg_pgno(mp);
1639 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1641 unsigned int i, nkeys, nsize, total = 0;
1645 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1646 case P_BRANCH: type = "Branch page"; break;
1647 case P_LEAF: type = "Leaf page"; break;
1648 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1649 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1650 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1652 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1653 pgno, mp->mp_pages, state);
1656 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1657 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1660 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1664 nkeys = NUMKEYS(mp);
1665 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1667 for (i=0; i<nkeys; i++) {
1668 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1669 key.mv_size = nsize = mp->mp_pad;
1670 key.mv_data = LEAF2KEY(mp, i, nsize);
1672 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1675 node = NODEPTR(mp, i);
1676 key.mv_size = node->mn_ksize;
1677 key.mv_data = node->mn_data;
1678 nsize = NODESIZE + key.mv_size;
1679 if (IS_BRANCH(mp)) {
1680 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1684 if (F_ISSET(node->mn_flags, F_BIGDATA))
1685 nsize += sizeof(pgno_t);
1687 nsize += NODEDSZ(node);
1689 nsize += sizeof(indx_t);
1690 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1691 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1693 total = EVEN(total);
1695 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1696 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1700 mdb_cursor_chk(MDB_cursor *mc)
1706 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1707 for (i=0; i<mc->mc_top; i++) {
1709 node = NODEPTR(mp, mc->mc_ki[i]);
1710 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1713 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1715 if (mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
1716 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1717 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1718 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1726 /** Count all the pages in each DB and in the freelist
1727 * and make sure it matches the actual number of pages
1729 * All named DBs must be open for a correct count.
1731 static void mdb_audit(MDB_txn *txn)
1735 MDB_ID freecount, count;
1740 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1741 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1742 freecount += *(MDB_ID *)data.mv_data;
1743 mdb_tassert(txn, rc == MDB_NOTFOUND);
1746 for (i = 0; i<txn->mt_numdbs; i++) {
1748 if (!(txn->mt_dbflags[i] & DB_VALID))
1750 mdb_cursor_init(&mc, txn, i, &mx);
1751 if (txn->mt_dbs[i].md_root == P_INVALID)
1753 count += txn->mt_dbs[i].md_branch_pages +
1754 txn->mt_dbs[i].md_leaf_pages +
1755 txn->mt_dbs[i].md_overflow_pages;
1756 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1757 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1758 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1761 mp = mc.mc_pg[mc.mc_top];
1762 for (j=0; j<NUMKEYS(mp); j++) {
1763 MDB_node *leaf = NODEPTR(mp, j);
1764 if (leaf->mn_flags & F_SUBDATA) {
1766 memcpy(&db, NODEDATA(leaf), sizeof(db));
1767 count += db.md_branch_pages + db.md_leaf_pages +
1768 db.md_overflow_pages;
1772 mdb_tassert(txn, rc == MDB_NOTFOUND);
1775 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1776 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1777 txn->mt_txnid, freecount, count+NUM_METAS,
1778 freecount+count+NUM_METAS, txn->mt_next_pgno);
1784 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1786 return txn->mt_dbxs[dbi].md_cmp(a, b);
1790 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1792 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1793 #if UINT_MAX < SIZE_MAX
1794 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1795 dcmp = mdb_cmp_clong;
1800 /** Allocate memory for a page.
1801 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1804 mdb_page_malloc(MDB_txn *txn, unsigned num)
1806 MDB_env *env = txn->mt_env;
1807 MDB_page *ret = env->me_dpages;
1808 size_t psize = env->me_psize, sz = psize, off;
1809 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1810 * For a single page alloc, we init everything after the page header.
1811 * For multi-page, we init the final page; if the caller needed that
1812 * many pages they will be filling in at least up to the last page.
1816 VGMEMP_ALLOC(env, ret, sz);
1817 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1818 env->me_dpages = ret->mp_next;
1821 psize -= off = PAGEHDRSZ;
1826 if ((ret = malloc(sz)) != NULL) {
1827 VGMEMP_ALLOC(env, ret, sz);
1828 if (!(env->me_flags & MDB_NOMEMINIT)) {
1829 memset((char *)ret + off, 0, psize);
1833 txn->mt_flags |= MDB_TXN_ERROR;
1837 /** Free a single page.
1838 * Saves single pages to a list, for future reuse.
1839 * (This is not used for multi-page overflow pages.)
1842 mdb_page_free(MDB_env *env, MDB_page *mp)
1844 mp->mp_next = env->me_dpages;
1845 VGMEMP_FREE(env, mp);
1846 env->me_dpages = mp;
1849 /** Free a dirty page */
1851 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1853 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1854 mdb_page_free(env, dp);
1856 /* large pages just get freed directly */
1857 VGMEMP_FREE(env, dp);
1862 /** Return all dirty pages to dpage list */
1864 mdb_dlist_free(MDB_txn *txn)
1866 MDB_env *env = txn->mt_env;
1867 MDB_ID2L dl = txn->mt_u.dirty_list;
1868 unsigned i, n = dl[0].mid;
1870 for (i = 1; i <= n; i++) {
1871 mdb_dpage_free(env, dl[i].mptr);
1876 /** Loosen or free a single page.
1877 * Saves single pages to a list for future reuse
1878 * in this same txn. It has been pulled from the freeDB
1879 * and already resides on the dirty list, but has been
1880 * deleted. Use these pages first before pulling again
1883 * If the page wasn't dirtied in this txn, just add it
1884 * to this txn's free list.
1887 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1890 pgno_t pgno = mp->mp_pgno;
1891 MDB_txn *txn = mc->mc_txn;
1893 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1894 if (txn->mt_parent) {
1895 MDB_ID2 *dl = txn->mt_u.dirty_list;
1896 /* If txn has a parent, make sure the page is in our
1900 unsigned x = mdb_mid2l_search(dl, pgno);
1901 if (x <= dl[0].mid && dl[x].mid == pgno) {
1902 if (mp != dl[x].mptr) { /* bad cursor? */
1903 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1904 txn->mt_flags |= MDB_TXN_ERROR;
1905 return MDB_CORRUPTED;
1912 /* no parent txn, so it's just ours */
1917 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1919 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1920 txn->mt_loose_pgs = mp;
1921 txn->mt_loose_count++;
1922 mp->mp_flags |= P_LOOSE;
1924 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1932 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1933 * @param[in] mc A cursor handle for the current operation.
1934 * @param[in] pflags Flags of the pages to update:
1935 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1936 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1937 * @return 0 on success, non-zero on failure.
1940 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1942 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1943 MDB_txn *txn = mc->mc_txn;
1949 int rc = MDB_SUCCESS, level;
1951 /* Mark pages seen by cursors */
1952 if (mc->mc_flags & C_UNTRACK)
1953 mc = NULL; /* will find mc in mt_cursors */
1954 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1955 for (; mc; mc=mc->mc_next) {
1956 if (!(mc->mc_flags & C_INITIALIZED))
1958 for (m3 = mc;; m3 = &mx->mx_cursor) {
1960 for (j=0; j<m3->mc_snum; j++) {
1962 if ((mp->mp_flags & Mask) == pflags)
1963 mp->mp_flags ^= P_KEEP;
1965 mx = m3->mc_xcursor;
1966 /* Proceed to mx if it is at a sub-database */
1967 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1969 if (! (mp && (mp->mp_flags & P_LEAF)))
1971 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1972 if (!(leaf->mn_flags & F_SUBDATA))
1981 /* Mark dirty root pages */
1982 for (i=0; i<txn->mt_numdbs; i++) {
1983 if (txn->mt_dbflags[i] & DB_DIRTY) {
1984 pgno_t pgno = txn->mt_dbs[i].md_root;
1985 if (pgno == P_INVALID)
1987 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1989 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1990 dp->mp_flags ^= P_KEEP;
1998 static int mdb_page_flush(MDB_txn *txn, int keep);
2000 /** Spill pages from the dirty list back to disk.
2001 * This is intended to prevent running into #MDB_TXN_FULL situations,
2002 * but note that they may still occur in a few cases:
2003 * 1) our estimate of the txn size could be too small. Currently this
2004 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
2005 * 2) child txns may run out of space if their parents dirtied a
2006 * lot of pages and never spilled them. TODO: we probably should do
2007 * a preemptive spill during #mdb_txn_begin() of a child txn, if
2008 * the parent's dirty_room is below a given threshold.
2010 * Otherwise, if not using nested txns, it is expected that apps will
2011 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
2012 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
2013 * If the txn never references them again, they can be left alone.
2014 * If the txn only reads them, they can be used without any fuss.
2015 * If the txn writes them again, they can be dirtied immediately without
2016 * going thru all of the work of #mdb_page_touch(). Such references are
2017 * handled by #mdb_page_unspill().
2019 * Also note, we never spill DB root pages, nor pages of active cursors,
2020 * because we'll need these back again soon anyway. And in nested txns,
2021 * we can't spill a page in a child txn if it was already spilled in a
2022 * parent txn. That would alter the parent txns' data even though
2023 * the child hasn't committed yet, and we'd have no way to undo it if
2024 * the child aborted.
2026 * @param[in] m0 cursor A cursor handle identifying the transaction and
2027 * database for which we are checking space.
2028 * @param[in] key For a put operation, the key being stored.
2029 * @param[in] data For a put operation, the data being stored.
2030 * @return 0 on success, non-zero on failure.
2033 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
2035 MDB_txn *txn = m0->mc_txn;
2037 MDB_ID2L dl = txn->mt_u.dirty_list;
2038 unsigned int i, j, need;
2041 if (m0->mc_flags & C_SUB)
2044 /* Estimate how much space this op will take */
2045 i = m0->mc_db->md_depth;
2046 /* Named DBs also dirty the main DB */
2047 if (m0->mc_dbi >= CORE_DBS)
2048 i += txn->mt_dbs[MAIN_DBI].md_depth;
2049 /* For puts, roughly factor in the key+data size */
2051 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2052 i += i; /* double it for good measure */
2055 if (txn->mt_dirty_room > i)
2058 if (!txn->mt_spill_pgs) {
2059 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2060 if (!txn->mt_spill_pgs)
2063 /* purge deleted slots */
2064 MDB_IDL sl = txn->mt_spill_pgs;
2065 unsigned int num = sl[0];
2067 for (i=1; i<=num; i++) {
2074 /* Preserve pages which may soon be dirtied again */
2075 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2078 /* Less aggressive spill - we originally spilled the entire dirty list,
2079 * with a few exceptions for cursor pages and DB root pages. But this
2080 * turns out to be a lot of wasted effort because in a large txn many
2081 * of those pages will need to be used again. So now we spill only 1/8th
2082 * of the dirty pages. Testing revealed this to be a good tradeoff,
2083 * better than 1/2, 1/4, or 1/10.
2085 if (need < MDB_IDL_UM_MAX / 8)
2086 need = MDB_IDL_UM_MAX / 8;
2088 /* Save the page IDs of all the pages we're flushing */
2089 /* flush from the tail forward, this saves a lot of shifting later on. */
2090 for (i=dl[0].mid; i && need; i--) {
2091 MDB_ID pn = dl[i].mid << 1;
2093 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2095 /* Can't spill twice, make sure it's not already in a parent's
2098 if (txn->mt_parent) {
2100 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2101 if (tx2->mt_spill_pgs) {
2102 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2103 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2104 dp->mp_flags |= P_KEEP;
2112 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2116 mdb_midl_sort(txn->mt_spill_pgs);
2118 /* Flush the spilled part of dirty list */
2119 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2122 /* Reset any dirty pages we kept that page_flush didn't see */
2123 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2126 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2130 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2132 mdb_find_oldest(MDB_txn *txn)
2135 txnid_t mr, oldest = txn->mt_txnid - 1;
2136 if (txn->mt_env->me_txns) {
2137 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2138 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2149 /** Add a page to the txn's dirty list */
2151 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2154 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2156 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2157 insert = mdb_mid2l_append;
2159 insert = mdb_mid2l_insert;
2161 mid.mid = mp->mp_pgno;
2163 rc = insert(txn->mt_u.dirty_list, &mid);
2164 mdb_tassert(txn, rc == 0);
2165 txn->mt_dirty_room--;
2168 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2169 * me_pghead and mt_next_pgno.
2171 * If there are free pages available from older transactions, they
2172 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2173 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2174 * and move me_pglast to say which records were consumed. Only this
2175 * function can create me_pghead and move me_pglast/mt_next_pgno.
2176 * @param[in] mc cursor A cursor handle identifying the transaction and
2177 * database for which we are allocating.
2178 * @param[in] num the number of pages to allocate.
2179 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2180 * will always be satisfied by a single contiguous chunk of memory.
2181 * @return 0 on success, non-zero on failure.
2184 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2186 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2187 /* Get at most <Max_retries> more freeDB records once me_pghead
2188 * has enough pages. If not enough, use new pages from the map.
2189 * If <Paranoid> and mc is updating the freeDB, only get new
2190 * records if me_pghead is empty. Then the freelist cannot play
2191 * catch-up with itself by growing while trying to save it.
2193 enum { Paranoid = 1, Max_retries = 500 };
2195 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2197 int rc, retry = num * 60;
2198 MDB_txn *txn = mc->mc_txn;
2199 MDB_env *env = txn->mt_env;
2200 pgno_t pgno, *mop = env->me_pghead;
2201 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2203 txnid_t oldest = 0, last;
2208 /* If there are any loose pages, just use them */
2209 if (num == 1 && txn->mt_loose_pgs) {
2210 np = txn->mt_loose_pgs;
2211 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2212 txn->mt_loose_count--;
2213 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2221 /* If our dirty list is already full, we can't do anything */
2222 if (txn->mt_dirty_room == 0) {
2227 for (op = MDB_FIRST;; op = MDB_NEXT) {
2232 /* Seek a big enough contiguous page range. Prefer
2233 * pages at the tail, just truncating the list.
2239 if (mop[i-n2] == pgno+n2)
2246 if (op == MDB_FIRST) { /* 1st iteration */
2247 /* Prepare to fetch more and coalesce */
2248 last = env->me_pglast;
2249 oldest = env->me_pgoldest;
2250 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2253 key.mv_data = &last; /* will look up last+1 */
2254 key.mv_size = sizeof(last);
2256 if (Paranoid && mc->mc_dbi == FREE_DBI)
2259 if (Paranoid && retry < 0 && mop_len)
2263 /* Do not fetch more if the record will be too recent */
2264 if (oldest <= last) {
2266 oldest = mdb_find_oldest(txn);
2267 env->me_pgoldest = oldest;
2273 rc = mdb_cursor_get(&m2, &key, NULL, op);
2275 if (rc == MDB_NOTFOUND)
2279 last = *(txnid_t*)key.mv_data;
2280 if (oldest <= last) {
2282 oldest = mdb_find_oldest(txn);
2283 env->me_pgoldest = oldest;
2289 np = m2.mc_pg[m2.mc_top];
2290 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2291 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2294 idl = (MDB_ID *) data.mv_data;
2297 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2302 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2304 mop = env->me_pghead;
2306 env->me_pglast = last;
2308 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2309 last, txn->mt_dbs[FREE_DBI].md_root, i));
2311 DPRINTF(("IDL %"Z"u", idl[j]));
2313 /* Merge in descending sorted order */
2314 mdb_midl_xmerge(mop, idl);
2318 /* Use new pages from the map when nothing suitable in the freeDB */
2320 pgno = txn->mt_next_pgno;
2321 if (pgno + num >= env->me_maxpg) {
2322 DPUTS("DB size maxed out");
2327 if (env->me_flags & MDB_WRITEMAP) {
2329 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
2330 p = VirtualAlloc(p, env->me_psize * num, MEM_COMMIT,
2331 (env->me_flags & MDB_WRITEMAP) ? PAGE_READWRITE:
2334 DPUTS("VirtualAlloc failed");
2342 if (env->me_flags & MDB_WRITEMAP) {
2343 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2345 if (!(np = mdb_page_malloc(txn, num))) {
2351 mop[0] = mop_len -= num;
2352 /* Move any stragglers down */
2353 for (j = i-num; j < mop_len; )
2354 mop[++j] = mop[++i];
2356 txn->mt_next_pgno = pgno + num;
2359 mdb_page_dirty(txn, np);
2365 txn->mt_flags |= MDB_TXN_ERROR;
2369 /** Copy the used portions of a non-overflow page.
2370 * @param[in] dst page to copy into
2371 * @param[in] src page to copy from
2372 * @param[in] psize size of a page
2375 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2377 enum { Align = sizeof(pgno_t) };
2378 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2380 /* If page isn't full, just copy the used portion. Adjust
2381 * alignment so memcpy may copy words instead of bytes.
2383 if ((unused &= -Align) && !IS_LEAF2(src)) {
2384 upper = (upper + PAGEBASE) & -Align;
2385 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2386 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2389 memcpy(dst, src, psize - unused);
2393 /** Pull a page off the txn's spill list, if present.
2394 * If a page being referenced was spilled to disk in this txn, bring
2395 * it back and make it dirty/writable again.
2396 * @param[in] txn the transaction handle.
2397 * @param[in] mp the page being referenced. It must not be dirty.
2398 * @param[out] ret the writable page, if any. ret is unchanged if
2399 * mp wasn't spilled.
2402 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2404 MDB_env *env = txn->mt_env;
2407 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2409 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2410 if (!tx2->mt_spill_pgs)
2412 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2413 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2416 if (txn->mt_dirty_room == 0)
2417 return MDB_TXN_FULL;
2418 if (IS_OVERFLOW(mp))
2422 if (env->me_flags & MDB_WRITEMAP) {
2425 np = mdb_page_malloc(txn, num);
2429 memcpy(np, mp, num * env->me_psize);
2431 mdb_page_copy(np, mp, env->me_psize);
2434 /* If in current txn, this page is no longer spilled.
2435 * If it happens to be the last page, truncate the spill list.
2436 * Otherwise mark it as deleted by setting the LSB.
2438 if (x == txn->mt_spill_pgs[0])
2439 txn->mt_spill_pgs[0]--;
2441 txn->mt_spill_pgs[x] |= 1;
2442 } /* otherwise, if belonging to a parent txn, the
2443 * page remains spilled until child commits
2446 mdb_page_dirty(txn, np);
2447 np->mp_flags |= P_DIRTY;
2455 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2456 * @param[in] mc cursor pointing to the page to be touched
2457 * @return 0 on success, non-zero on failure.
2460 mdb_page_touch(MDB_cursor *mc)
2462 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2463 MDB_txn *txn = mc->mc_txn;
2464 MDB_cursor *m2, *m3;
2468 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2469 if (txn->mt_flags & MDB_TXN_SPILLS) {
2471 rc = mdb_page_unspill(txn, mp, &np);
2477 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2478 (rc = mdb_page_alloc(mc, 1, &np)))
2481 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2482 mp->mp_pgno, pgno));
2483 mdb_cassert(mc, mp->mp_pgno != pgno);
2484 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2485 /* Update the parent page, if any, to point to the new page */
2487 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2488 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2489 SETPGNO(node, pgno);
2491 mc->mc_db->md_root = pgno;
2493 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2494 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2496 /* If txn has a parent, make sure the page is in our
2500 unsigned x = mdb_mid2l_search(dl, pgno);
2501 if (x <= dl[0].mid && dl[x].mid == pgno) {
2502 if (mp != dl[x].mptr) { /* bad cursor? */
2503 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2504 txn->mt_flags |= MDB_TXN_ERROR;
2505 return MDB_CORRUPTED;
2510 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2512 np = mdb_page_malloc(txn, 1);
2517 rc = mdb_mid2l_insert(dl, &mid);
2518 mdb_cassert(mc, rc == 0);
2523 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2525 np->mp_flags |= P_DIRTY;
2528 /* Adjust cursors pointing to mp */
2529 mc->mc_pg[mc->mc_top] = np;
2530 m2 = txn->mt_cursors[mc->mc_dbi];
2531 if (mc->mc_flags & C_SUB) {
2532 for (; m2; m2=m2->mc_next) {
2533 m3 = &m2->mc_xcursor->mx_cursor;
2534 if (m3->mc_snum < mc->mc_snum) continue;
2535 if (m3->mc_pg[mc->mc_top] == mp)
2536 m3->mc_pg[mc->mc_top] = np;
2539 for (; m2; m2=m2->mc_next) {
2540 if (m2->mc_snum < mc->mc_snum) continue;
2541 if (m2 == mc) continue;
2542 if (m2->mc_pg[mc->mc_top] == mp) {
2543 m2->mc_pg[mc->mc_top] = np;
2544 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2546 (m2->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
2548 MDB_node *leaf = NODEPTR(np, m2->mc_ki[mc->mc_top]);
2549 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
2550 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2558 txn->mt_flags |= MDB_TXN_ERROR;
2563 mdb_env_sync(MDB_env *env, int force)
2566 if (env->me_flags & MDB_RDONLY)
2568 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2569 if (env->me_flags & MDB_WRITEMAP) {
2570 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2571 ? MS_ASYNC : MS_SYNC;
2572 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2575 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2579 #ifdef BROKEN_FDATASYNC
2580 if (env->me_flags & MDB_FSYNCONLY) {
2581 if (fsync(env->me_fd))
2585 if (MDB_FDATASYNC(env->me_fd))
2592 /** Back up parent txn's cursors, then grab the originals for tracking */
2594 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2596 MDB_cursor *mc, *bk;
2601 for (i = src->mt_numdbs; --i >= 0; ) {
2602 if ((mc = src->mt_cursors[i]) != NULL) {
2603 size = sizeof(MDB_cursor);
2605 size += sizeof(MDB_xcursor);
2606 for (; mc; mc = bk->mc_next) {
2612 mc->mc_db = &dst->mt_dbs[i];
2613 /* Kill pointers into src to reduce abuse: The
2614 * user may not use mc until dst ends. But we need a valid
2615 * txn pointer here for cursor fixups to keep working.
2618 mc->mc_dbflag = &dst->mt_dbflags[i];
2619 if ((mx = mc->mc_xcursor) != NULL) {
2620 *(MDB_xcursor *)(bk+1) = *mx;
2621 mx->mx_cursor.mc_txn = dst;
2623 mc->mc_next = dst->mt_cursors[i];
2624 dst->mt_cursors[i] = mc;
2631 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2632 * @param[in] txn the transaction handle.
2633 * @param[in] merge true to keep changes to parent cursors, false to revert.
2634 * @return 0 on success, non-zero on failure.
2637 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2639 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2643 for (i = txn->mt_numdbs; --i >= 0; ) {
2644 for (mc = cursors[i]; mc; mc = next) {
2646 if ((bk = mc->mc_backup) != NULL) {
2648 /* Commit changes to parent txn */
2649 mc->mc_next = bk->mc_next;
2650 mc->mc_backup = bk->mc_backup;
2651 mc->mc_txn = bk->mc_txn;
2652 mc->mc_db = bk->mc_db;
2653 mc->mc_dbflag = bk->mc_dbflag;
2654 if ((mx = mc->mc_xcursor) != NULL)
2655 mx->mx_cursor.mc_txn = bk->mc_txn;
2657 /* Abort nested txn */
2659 if ((mx = mc->mc_xcursor) != NULL)
2660 *mx = *(MDB_xcursor *)(bk+1);
2664 /* Only malloced cursors are permanently tracked. */
2671 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2677 Pidset = F_SETLK, Pidcheck = F_GETLK
2681 /** Set or check a pid lock. Set returns 0 on success.
2682 * Check returns 0 if the process is certainly dead, nonzero if it may
2683 * be alive (the lock exists or an error happened so we do not know).
2685 * On Windows Pidset is a no-op, we merely check for the existence
2686 * of the process with the given pid. On POSIX we use a single byte
2687 * lock on the lockfile, set at an offset equal to the pid.
2690 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2692 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2695 if (op == Pidcheck) {
2696 h = OpenProcess(env->me_pidquery, FALSE, pid);
2697 /* No documented "no such process" code, but other program use this: */
2699 return ErrCode() != ERROR_INVALID_PARAMETER;
2700 /* A process exists until all handles to it close. Has it exited? */
2701 ret = WaitForSingleObject(h, 0) != 0;
2708 struct flock lock_info;
2709 memset(&lock_info, 0, sizeof(lock_info));
2710 lock_info.l_type = F_WRLCK;
2711 lock_info.l_whence = SEEK_SET;
2712 lock_info.l_start = pid;
2713 lock_info.l_len = 1;
2714 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2715 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2717 } else if ((rc = ErrCode()) == EINTR) {
2725 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2726 * @param[in] txn the transaction handle to initialize
2727 * @return 0 on success, non-zero on failure.
2730 mdb_txn_renew0(MDB_txn *txn)
2732 MDB_env *env = txn->mt_env;
2733 MDB_txninfo *ti = env->me_txns;
2735 unsigned int i, nr, flags = txn->mt_flags;
2737 int rc, new_notls = 0;
2739 if ((flags &= MDB_TXN_RDONLY) != 0) {
2741 meta = mdb_env_pick_meta(env);
2742 txn->mt_txnid = meta->mm_txnid;
2743 txn->mt_u.reader = NULL;
2745 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2746 pthread_getspecific(env->me_txkey);
2748 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2749 return MDB_BAD_RSLOT;
2751 MDB_PID_T pid = env->me_pid;
2752 MDB_THR_T tid = pthread_self();
2753 mdb_mutexref_t rmutex = env->me_rmutex;
2755 if (!env->me_live_reader) {
2756 rc = mdb_reader_pid(env, Pidset, pid);
2759 env->me_live_reader = 1;
2762 if (LOCK_MUTEX(rc, env, rmutex))
2764 nr = ti->mti_numreaders;
2765 for (i=0; i<nr; i++)
2766 if (ti->mti_readers[i].mr_pid == 0)
2768 if (i == env->me_maxreaders) {
2769 UNLOCK_MUTEX(rmutex);
2770 return MDB_READERS_FULL;
2772 r = &ti->mti_readers[i];
2773 /* Claim the reader slot, carefully since other code
2774 * uses the reader table un-mutexed: First reset the
2775 * slot, next publish it in mti_numreaders. After
2776 * that, it is safe for mdb_env_close() to touch it.
2777 * When it will be closed, we can finally claim it.
2780 r->mr_txnid = (txnid_t)-1;
2783 ti->mti_numreaders = ++nr;
2784 env->me_close_readers = nr;
2786 UNLOCK_MUTEX(rmutex);
2788 new_notls = (env->me_flags & MDB_NOTLS);
2789 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2794 do /* LY: Retry on a race, ITS#7970. */
2795 r->mr_txnid = ti->mti_txnid;
2796 while(r->mr_txnid != ti->mti_txnid);
2797 txn->mt_txnid = r->mr_txnid;
2798 txn->mt_u.reader = r;
2799 meta = env->me_metas[txn->mt_txnid & 1];
2803 /* Not yet touching txn == env->me_txn0, it may be active */
2805 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2807 txn->mt_txnid = ti->mti_txnid;
2808 meta = env->me_metas[txn->mt_txnid & 1];
2810 meta = mdb_env_pick_meta(env);
2811 txn->mt_txnid = meta->mm_txnid;
2815 if (txn->mt_txnid == mdb_debug_start)
2818 txn->mt_child = NULL;
2819 txn->mt_loose_pgs = NULL;
2820 txn->mt_loose_count = 0;
2821 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2822 txn->mt_u.dirty_list = env->me_dirty_list;
2823 txn->mt_u.dirty_list[0].mid = 0;
2824 txn->mt_free_pgs = env->me_free_pgs;
2825 txn->mt_free_pgs[0] = 0;
2826 txn->mt_spill_pgs = NULL;
2828 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2831 /* Copy the DB info and flags */
2832 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2834 /* Moved to here to avoid a data race in read TXNs */
2835 txn->mt_next_pgno = meta->mm_last_pg+1;
2837 txn->mt_flags = flags;
2840 txn->mt_numdbs = env->me_numdbs;
2841 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2842 x = env->me_dbflags[i];
2843 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2844 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2846 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2847 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2849 if (env->me_flags & MDB_FATAL_ERROR) {
2850 DPUTS("environment had fatal error, must shutdown!");
2852 } else if (env->me_maxpg < txn->mt_next_pgno) {
2853 rc = MDB_MAP_RESIZED;
2857 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2862 mdb_txn_renew(MDB_txn *txn)
2866 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2869 rc = mdb_txn_renew0(txn);
2870 if (rc == MDB_SUCCESS) {
2871 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2872 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2873 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2879 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2883 int rc, size, tsize;
2885 flags &= MDB_TXN_BEGIN_FLAGS;
2886 flags |= env->me_flags & MDB_WRITEMAP;
2888 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2892 /* Nested transactions: Max 1 child, write txns only, no writemap */
2893 flags |= parent->mt_flags;
2894 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2895 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2897 /* Child txns save MDB_pgstate and use own copy of cursors */
2898 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2899 size += tsize = sizeof(MDB_ntxn);
2900 } else if (flags & MDB_RDONLY) {
2901 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2902 size += tsize = sizeof(MDB_txn);
2904 /* Reuse preallocated write txn. However, do not touch it until
2905 * mdb_txn_renew0() succeeds, since it currently may be active.
2910 if ((txn = calloc(1, size)) == NULL) {
2911 DPRINTF(("calloc: %s", strerror(errno)));
2914 txn->mt_dbxs = env->me_dbxs; /* static */
2915 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2916 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2917 txn->mt_flags = flags;
2922 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2923 txn->mt_dbiseqs = parent->mt_dbiseqs;
2924 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2925 if (!txn->mt_u.dirty_list ||
2926 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2928 free(txn->mt_u.dirty_list);
2932 txn->mt_txnid = parent->mt_txnid;
2933 txn->mt_dirty_room = parent->mt_dirty_room;
2934 txn->mt_u.dirty_list[0].mid = 0;
2935 txn->mt_spill_pgs = NULL;
2936 txn->mt_next_pgno = parent->mt_next_pgno;
2937 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2938 parent->mt_child = txn;
2939 txn->mt_parent = parent;
2940 txn->mt_numdbs = parent->mt_numdbs;
2941 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2942 /* Copy parent's mt_dbflags, but clear DB_NEW */
2943 for (i=0; i<txn->mt_numdbs; i++)
2944 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2946 ntxn = (MDB_ntxn *)txn;
2947 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2948 if (env->me_pghead) {
2949 size = MDB_IDL_SIZEOF(env->me_pghead);
2950 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2952 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2957 rc = mdb_cursor_shadow(parent, txn);
2959 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2960 } else { /* MDB_RDONLY */
2961 txn->mt_dbiseqs = env->me_dbiseqs;
2963 rc = mdb_txn_renew0(txn);
2966 if (txn != env->me_txn0)
2969 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2971 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2972 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2973 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2980 mdb_txn_env(MDB_txn *txn)
2982 if(!txn) return NULL;
2987 mdb_txn_id(MDB_txn *txn)
2990 return txn->mt_txnid;
2993 /** Export or close DBI handles opened in this txn. */
2995 mdb_dbis_update(MDB_txn *txn, int keep)
2998 MDB_dbi n = txn->mt_numdbs;
2999 MDB_env *env = txn->mt_env;
3000 unsigned char *tdbflags = txn->mt_dbflags;
3002 for (i = n; --i >= CORE_DBS;) {
3003 if (tdbflags[i] & DB_NEW) {
3005 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
3007 char *ptr = env->me_dbxs[i].md_name.mv_data;
3009 env->me_dbxs[i].md_name.mv_data = NULL;
3010 env->me_dbxs[i].md_name.mv_size = 0;
3011 env->me_dbflags[i] = 0;
3012 env->me_dbiseqs[i]++;
3018 if (keep && env->me_numdbs < n)
3022 /** End a transaction, except successful commit of a nested transaction.
3023 * May be called twice for readonly txns: First reset it, then abort.
3024 * @param[in] txn the transaction handle to end
3025 * @param[in] mode why and how to end the transaction
3028 mdb_txn_end(MDB_txn *txn, unsigned mode)
3030 MDB_env *env = txn->mt_env;
3032 static const char *const names[] = MDB_END_NAMES;
3035 /* Export or close DBI handles opened in this txn */
3036 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
3038 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
3039 names[mode & MDB_END_OPMASK],
3040 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3041 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
3043 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3044 if (txn->mt_u.reader) {
3045 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
3046 if (!(env->me_flags & MDB_NOTLS)) {
3047 txn->mt_u.reader = NULL; /* txn does not own reader */
3048 } else if (mode & MDB_END_SLOT) {
3049 txn->mt_u.reader->mr_pid = 0;
3050 txn->mt_u.reader = NULL;
3051 } /* else txn owns the slot until it does MDB_END_SLOT */
3053 txn->mt_numdbs = 0; /* prevent further DBI activity */
3054 txn->mt_flags |= MDB_TXN_FINISHED;
3056 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
3057 pgno_t *pghead = env->me_pghead;
3059 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
3060 mdb_cursors_close(txn, 0);
3061 if (!(env->me_flags & MDB_WRITEMAP)) {
3062 mdb_dlist_free(txn);
3066 txn->mt_flags = MDB_TXN_FINISHED;
3068 if (!txn->mt_parent) {
3069 mdb_midl_shrink(&txn->mt_free_pgs);
3070 env->me_free_pgs = txn->mt_free_pgs;
3072 env->me_pghead = NULL;
3076 mode = 0; /* txn == env->me_txn0, do not free() it */
3078 /* The writer mutex was locked in mdb_txn_begin. */
3080 UNLOCK_MUTEX(env->me_wmutex);
3082 txn->mt_parent->mt_child = NULL;
3083 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3084 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3085 mdb_midl_free(txn->mt_free_pgs);
3086 mdb_midl_free(txn->mt_spill_pgs);
3087 free(txn->mt_u.dirty_list);
3090 mdb_midl_free(pghead);
3093 if (mode & MDB_END_FREE)
3098 mdb_txn_reset(MDB_txn *txn)
3103 /* This call is only valid for read-only txns */
3104 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3107 mdb_txn_end(txn, MDB_END_RESET);
3111 mdb_txn_abort(MDB_txn *txn)
3117 mdb_txn_abort(txn->mt_child);
3119 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3122 /** Save the freelist as of this transaction to the freeDB.
3123 * This changes the freelist. Keep trying until it stabilizes.
3126 mdb_freelist_save(MDB_txn *txn)
3128 /* env->me_pghead[] can grow and shrink during this call.
3129 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3130 * Page numbers cannot disappear from txn->mt_free_pgs[].
3133 MDB_env *env = txn->mt_env;
3134 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3135 txnid_t pglast = 0, head_id = 0;
3136 pgno_t freecnt = 0, *free_pgs, *mop;
3137 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3139 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3141 if (env->me_pghead) {
3142 /* Make sure first page of freeDB is touched and on freelist */
3143 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3144 if (rc && rc != MDB_NOTFOUND)
3148 if (!env->me_pghead && txn->mt_loose_pgs) {
3149 /* Put loose page numbers in mt_free_pgs, since
3150 * we may be unable to return them to me_pghead.
3152 MDB_page *mp = txn->mt_loose_pgs;
3153 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3155 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3156 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3157 txn->mt_loose_pgs = NULL;
3158 txn->mt_loose_count = 0;
3161 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3162 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3163 ? SSIZE_MAX : maxfree_1pg;
3166 /* Come back here after each Put() in case freelist changed */
3171 /* If using records from freeDB which we have not yet
3172 * deleted, delete them and any we reserved for me_pghead.
3174 while (pglast < env->me_pglast) {
3175 rc = mdb_cursor_first(&mc, &key, NULL);
3178 pglast = head_id = *(txnid_t *)key.mv_data;
3179 total_room = head_room = 0;
3180 mdb_tassert(txn, pglast <= env->me_pglast);
3181 rc = mdb_cursor_del(&mc, 0);
3186 /* Save the IDL of pages freed by this txn, to a single record */
3187 if (freecnt < txn->mt_free_pgs[0]) {
3189 /* Make sure last page of freeDB is touched and on freelist */
3190 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3191 if (rc && rc != MDB_NOTFOUND)
3194 free_pgs = txn->mt_free_pgs;
3195 /* Write to last page of freeDB */
3196 key.mv_size = sizeof(txn->mt_txnid);
3197 key.mv_data = &txn->mt_txnid;
3199 freecnt = free_pgs[0];
3200 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3201 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3204 /* Retry if mt_free_pgs[] grew during the Put() */
3205 free_pgs = txn->mt_free_pgs;
3206 } while (freecnt < free_pgs[0]);
3207 mdb_midl_sort(free_pgs);
3208 memcpy(data.mv_data, free_pgs, data.mv_size);
3211 unsigned int i = free_pgs[0];
3212 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3213 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3215 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3221 mop = env->me_pghead;
3222 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3224 /* Reserve records for me_pghead[]. Split it if multi-page,
3225 * to avoid searching freeDB for a page range. Use keys in
3226 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3228 if (total_room >= mop_len) {
3229 if (total_room == mop_len || --more < 0)
3231 } else if (head_room >= maxfree_1pg && head_id > 1) {
3232 /* Keep current record (overflow page), add a new one */
3236 /* (Re)write {key = head_id, IDL length = head_room} */
3237 total_room -= head_room;
3238 head_room = mop_len - total_room;
3239 if (head_room > maxfree_1pg && head_id > 1) {
3240 /* Overflow multi-page for part of me_pghead */
3241 head_room /= head_id; /* amortize page sizes */
3242 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3243 } else if (head_room < 0) {
3244 /* Rare case, not bothering to delete this record */
3247 key.mv_size = sizeof(head_id);
3248 key.mv_data = &head_id;
3249 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3250 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3253 /* IDL is initially empty, zero out at least the length */
3254 pgs = (pgno_t *)data.mv_data;
3255 j = head_room > clean_limit ? head_room : 0;
3259 total_room += head_room;
3262 /* Return loose page numbers to me_pghead, though usually none are
3263 * left at this point. The pages themselves remain in dirty_list.
3265 if (txn->mt_loose_pgs) {
3266 MDB_page *mp = txn->mt_loose_pgs;
3267 unsigned count = txn->mt_loose_count;
3269 /* Room for loose pages + temp IDL with same */
3270 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3272 mop = env->me_pghead;
3273 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3274 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3275 loose[ ++count ] = mp->mp_pgno;
3277 mdb_midl_sort(loose);
3278 mdb_midl_xmerge(mop, loose);
3279 txn->mt_loose_pgs = NULL;
3280 txn->mt_loose_count = 0;
3284 /* Fill in the reserved me_pghead records */
3290 rc = mdb_cursor_first(&mc, &key, &data);
3291 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3292 txnid_t id = *(txnid_t *)key.mv_data;
3293 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3296 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3298 if (len > mop_len) {
3300 data.mv_size = (len + 1) * sizeof(MDB_ID);
3302 data.mv_data = mop -= len;
3305 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3307 if (rc || !(mop_len -= len))
3314 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3315 * @param[in] txn the transaction that's being committed
3316 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3317 * @return 0 on success, non-zero on failure.
3320 mdb_page_flush(MDB_txn *txn, int keep)
3322 MDB_env *env = txn->mt_env;
3323 MDB_ID2L dl = txn->mt_u.dirty_list;
3324 unsigned psize = env->me_psize, j;
3325 int i, pagecount = dl[0].mid, rc;
3326 size_t size = 0, pos = 0;
3328 MDB_page *dp = NULL;
3332 struct iovec iov[MDB_COMMIT_PAGES];
3333 ssize_t wpos = 0, wsize = 0, wres;
3334 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3340 if (env->me_flags & MDB_WRITEMAP) {
3341 /* Clear dirty flags */
3342 while (++i <= pagecount) {
3344 /* Don't flush this page yet */
3345 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3346 dp->mp_flags &= ~P_KEEP;
3350 dp->mp_flags &= ~P_DIRTY;
3355 /* Write the pages */
3357 if (++i <= pagecount) {
3359 /* Don't flush this page yet */
3360 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3361 dp->mp_flags &= ~P_KEEP;
3366 /* clear dirty flag */
3367 dp->mp_flags &= ~P_DIRTY;
3370 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3375 /* Windows actually supports scatter/gather I/O, but only on
3376 * unbuffered file handles. Since we're relying on the OS page
3377 * cache for all our data, that's self-defeating. So we just
3378 * write pages one at a time. We use the ov structure to set
3379 * the write offset, to at least save the overhead of a Seek
3382 DPRINTF(("committing page %"Z"u", pgno));
3383 memset(&ov, 0, sizeof(ov));
3384 ov.Offset = pos & 0xffffffff;
3385 ov.OffsetHigh = pos >> 16 >> 16;
3386 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3388 DPRINTF(("WriteFile: %d", rc));
3392 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3393 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3396 /* Write previous page(s) */
3397 #ifdef MDB_USE_PWRITEV
3398 wres = pwritev(env->me_fd, iov, n, wpos);
3401 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3404 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3408 DPRINTF(("lseek: %s", strerror(rc)));
3411 wres = writev(env->me_fd, iov, n);
3414 if (wres != wsize) {
3419 DPRINTF(("Write error: %s", strerror(rc)));
3421 rc = EIO; /* TODO: Use which error code? */
3422 DPUTS("short write, filesystem full?");
3433 DPRINTF(("committing page %"Z"u", pgno));
3434 next_pos = pos + size;
3435 iov[n].iov_len = size;
3436 iov[n].iov_base = (char *)dp;
3442 /* MIPS has cache coherency issues, this is a no-op everywhere else
3443 * Note: for any size >= on-chip cache size, entire on-chip cache is
3446 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3448 for (i = keep; ++i <= pagecount; ) {
3450 /* This is a page we skipped above */
3453 dl[j].mid = dp->mp_pgno;
3456 mdb_dpage_free(env, dp);
3461 txn->mt_dirty_room += i - j;
3467 mdb_txn_commit(MDB_txn *txn)
3470 unsigned int i, end_mode;
3476 /* mdb_txn_end() mode for a commit which writes nothing */
3477 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3479 if (txn->mt_child) {
3480 rc = mdb_txn_commit(txn->mt_child);
3487 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3491 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3492 DPUTS("txn has failed/finished, can't commit");
3494 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3499 if (txn->mt_parent) {
3500 MDB_txn *parent = txn->mt_parent;
3504 unsigned x, y, len, ps_len;
3506 /* Append our free list to parent's */
3507 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3510 mdb_midl_free(txn->mt_free_pgs);
3511 /* Failures after this must either undo the changes
3512 * to the parent or set MDB_TXN_ERROR in the parent.
3515 parent->mt_next_pgno = txn->mt_next_pgno;
3516 parent->mt_flags = txn->mt_flags;
3518 /* Merge our cursors into parent's and close them */
3519 mdb_cursors_close(txn, 1);
3521 /* Update parent's DB table. */
3522 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3523 parent->mt_numdbs = txn->mt_numdbs;
3524 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3525 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3526 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3527 /* preserve parent's DB_NEW status */
3528 x = parent->mt_dbflags[i] & DB_NEW;
3529 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3532 dst = parent->mt_u.dirty_list;
3533 src = txn->mt_u.dirty_list;
3534 /* Remove anything in our dirty list from parent's spill list */
3535 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3537 pspill[0] = (pgno_t)-1;
3538 /* Mark our dirty pages as deleted in parent spill list */
3539 for (i=0, len=src[0].mid; ++i <= len; ) {
3540 MDB_ID pn = src[i].mid << 1;
3541 while (pn > pspill[x])
3543 if (pn == pspill[x]) {
3548 /* Squash deleted pagenums if we deleted any */
3549 for (x=y; ++x <= ps_len; )
3550 if (!(pspill[x] & 1))
3551 pspill[++y] = pspill[x];
3555 /* Remove anything in our spill list from parent's dirty list */
3556 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3557 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3558 MDB_ID pn = txn->mt_spill_pgs[i];
3560 continue; /* deleted spillpg */
3562 y = mdb_mid2l_search(dst, pn);
3563 if (y <= dst[0].mid && dst[y].mid == pn) {
3565 while (y < dst[0].mid) {
3574 /* Find len = length of merging our dirty list with parent's */
3576 dst[0].mid = 0; /* simplify loops */
3577 if (parent->mt_parent) {
3578 len = x + src[0].mid;
3579 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3580 for (i = x; y && i; y--) {
3581 pgno_t yp = src[y].mid;
3582 while (yp < dst[i].mid)
3584 if (yp == dst[i].mid) {
3589 } else { /* Simplify the above for single-ancestor case */
3590 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3592 /* Merge our dirty list with parent's */
3594 for (i = len; y; dst[i--] = src[y--]) {
3595 pgno_t yp = src[y].mid;
3596 while (yp < dst[x].mid)
3597 dst[i--] = dst[x--];
3598 if (yp == dst[x].mid)
3599 free(dst[x--].mptr);
3601 mdb_tassert(txn, i == x);
3603 free(txn->mt_u.dirty_list);
3604 parent->mt_dirty_room = txn->mt_dirty_room;
3605 if (txn->mt_spill_pgs) {
3606 if (parent->mt_spill_pgs) {
3607 /* TODO: Prevent failure here, so parent does not fail */
3608 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3610 parent->mt_flags |= MDB_TXN_ERROR;
3611 mdb_midl_free(txn->mt_spill_pgs);
3612 mdb_midl_sort(parent->mt_spill_pgs);
3614 parent->mt_spill_pgs = txn->mt_spill_pgs;
3618 /* Append our loose page list to parent's */
3619 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3621 *lp = txn->mt_loose_pgs;
3622 parent->mt_loose_count += txn->mt_loose_count;
3624 parent->mt_child = NULL;
3625 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3630 if (txn != env->me_txn) {
3631 DPUTS("attempt to commit unknown transaction");
3636 mdb_cursors_close(txn, 0);
3638 if (!txn->mt_u.dirty_list[0].mid &&
3639 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3642 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3643 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3645 /* Update DB root pointers */
3646 if (txn->mt_numdbs > CORE_DBS) {
3650 data.mv_size = sizeof(MDB_db);
3652 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3653 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3654 if (txn->mt_dbflags[i] & DB_DIRTY) {
3655 if (TXN_DBI_CHANGED(txn, i)) {
3659 data.mv_data = &txn->mt_dbs[i];
3660 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3668 rc = mdb_freelist_save(txn);
3672 mdb_midl_free(env->me_pghead);
3673 env->me_pghead = NULL;
3674 mdb_midl_shrink(&txn->mt_free_pgs);
3680 if ((rc = mdb_page_flush(txn, 0)))
3682 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3683 (rc = mdb_env_sync(env, 0)))
3685 if ((rc = mdb_env_write_meta(txn)))
3687 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3690 mdb_txn_end(txn, end_mode);
3698 /** Read the environment parameters of a DB environment before
3699 * mapping it into memory.
3700 * @param[in] env the environment handle
3701 * @param[out] meta address of where to store the meta information
3702 * @return 0 on success, non-zero on failure.
3705 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3711 enum { Size = sizeof(pbuf) };
3713 /* We don't know the page size yet, so use a minimum value.
3714 * Read both meta pages so we can use the latest one.
3717 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3721 memset(&ov, 0, sizeof(ov));
3723 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3724 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3727 rc = pread(env->me_fd, &pbuf, Size, off);
3730 if (rc == 0 && off == 0)
3732 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3733 DPRINTF(("read: %s", mdb_strerror(rc)));
3737 p = (MDB_page *)&pbuf;
3739 if (!F_ISSET(p->mp_flags, P_META)) {
3740 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3745 if (m->mm_magic != MDB_MAGIC) {
3746 DPUTS("meta has invalid magic");
3750 if (m->mm_version != MDB_DATA_VERSION) {
3751 DPRINTF(("database is version %u, expected version %u",
3752 m->mm_version, MDB_DATA_VERSION));
3753 return MDB_VERSION_MISMATCH;
3756 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3762 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3764 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3766 meta->mm_magic = MDB_MAGIC;
3767 meta->mm_version = MDB_DATA_VERSION;
3768 meta->mm_mapsize = env->me_mapsize;
3769 meta->mm_psize = env->me_psize;
3770 meta->mm_last_pg = NUM_METAS-1;
3771 meta->mm_flags = env->me_flags & 0xffff;
3772 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3773 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3774 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3777 /** Write the environment parameters of a freshly created DB environment.
3778 * @param[in] env the environment handle
3779 * @param[in] meta the #MDB_meta to write
3780 * @return 0 on success, non-zero on failure.
3783 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3791 memset(&ov, 0, sizeof(ov));
3792 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3794 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3797 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3798 len = pwrite(fd, ptr, size, pos); \
3799 if (len == -1 && ErrCode() == EINTR) continue; \
3800 rc = (len >= 0); break; } while(1)
3803 DPUTS("writing new meta page");
3805 psize = env->me_psize;
3807 p = calloc(NUM_METAS, psize);
3811 p->mp_flags = P_META;
3812 *(MDB_meta *)METADATA(p) = *meta;
3814 q = (MDB_page *)((char *)p + psize);
3816 q->mp_flags = P_META;
3817 *(MDB_meta *)METADATA(q) = *meta;
3819 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3822 else if ((unsigned) len == psize * NUM_METAS)
3830 /** Update the environment info to commit a transaction.
3831 * @param[in] txn the transaction that's being committed
3832 * @return 0 on success, non-zero on failure.
3835 mdb_env_write_meta(MDB_txn *txn)
3838 MDB_meta meta, metab, *mp;
3842 int rc, len, toggle;
3851 toggle = txn->mt_txnid & 1;
3852 DPRINTF(("writing meta page %d for root page %"Z"u",
3853 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3856 flags = txn->mt_flags | env->me_flags;
3857 mp = env->me_metas[toggle];
3858 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3859 /* Persist any increases of mapsize config */
3860 if (mapsize < env->me_mapsize)
3861 mapsize = env->me_mapsize;
3863 if (flags & MDB_WRITEMAP) {
3864 mp->mm_mapsize = mapsize;
3865 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3866 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3867 mp->mm_last_pg = txn->mt_next_pgno - 1;
3868 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3869 !(defined(__i386__) || defined(__x86_64__))
3870 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3871 __sync_synchronize();
3873 mp->mm_txnid = txn->mt_txnid;
3874 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3875 unsigned meta_size = env->me_psize;
3876 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3877 ptr = (char *)mp - PAGEHDRSZ;
3878 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3879 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3883 if (MDB_MSYNC(ptr, meta_size, rc)) {
3890 metab.mm_txnid = mp->mm_txnid;
3891 metab.mm_last_pg = mp->mm_last_pg;
3893 meta.mm_mapsize = mapsize;
3894 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3895 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3896 meta.mm_last_pg = txn->mt_next_pgno - 1;
3897 meta.mm_txnid = txn->mt_txnid;
3899 off = offsetof(MDB_meta, mm_mapsize);
3900 ptr = (char *)&meta + off;
3901 len = sizeof(MDB_meta) - off;
3902 off += (char *)mp - env->me_map;
3904 /* Write to the SYNC fd */
3905 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3908 memset(&ov, 0, sizeof(ov));
3910 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3915 rc = pwrite(mfd, ptr, len, off);
3918 rc = rc < 0 ? ErrCode() : EIO;
3923 DPUTS("write failed, disk error?");
3924 /* On a failure, the pagecache still contains the new data.
3925 * Write some old data back, to prevent it from being used.
3926 * Use the non-SYNC fd; we know it will fail anyway.
3928 meta.mm_last_pg = metab.mm_last_pg;
3929 meta.mm_txnid = metab.mm_txnid;
3931 memset(&ov, 0, sizeof(ov));
3933 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3935 r2 = pwrite(env->me_fd, ptr, len, off);
3936 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3939 env->me_flags |= MDB_FATAL_ERROR;
3942 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3943 CACHEFLUSH(env->me_map + off, len, DCACHE);
3945 /* Memory ordering issues are irrelevant; since the entire writer
3946 * is wrapped by wmutex, all of these changes will become visible
3947 * after the wmutex is unlocked. Since the DB is multi-version,
3948 * readers will get consistent data regardless of how fresh or
3949 * how stale their view of these values is.
3952 env->me_txns->mti_txnid = txn->mt_txnid;
3957 /** Check both meta pages to see which one is newer.
3958 * @param[in] env the environment handle
3959 * @return newest #MDB_meta.
3962 mdb_env_pick_meta(const MDB_env *env)
3964 MDB_meta *const *metas = env->me_metas;
3965 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3969 mdb_env_create(MDB_env **env)
3973 e = calloc(1, sizeof(MDB_env));
3977 e->me_maxreaders = DEFAULT_READERS;
3978 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3979 e->me_fd = INVALID_HANDLE_VALUE;
3980 e->me_lfd = INVALID_HANDLE_VALUE;
3981 e->me_mfd = INVALID_HANDLE_VALUE;
3982 #ifdef MDB_USE_POSIX_SEM
3983 e->me_rmutex = SEM_FAILED;
3984 e->me_wmutex = SEM_FAILED;
3985 #elif defined MDB_USE_SYSV_SEM
3986 e->me_rmutex->semid = -1;
3987 e->me_wmutex->semid = -1;
3989 e->me_pid = getpid();
3990 GET_PAGESIZE(e->me_os_psize);
3991 VGMEMP_CREATE(e,0,0);
3997 mdb_env_map(MDB_env *env, void *addr)
4000 unsigned int flags = env->me_flags;
4003 int access = SECTION_MAP_READ;
4007 ULONG pageprot = PAGE_READONLY, secprot, alloctype;
4009 if (flags & MDB_WRITEMAP) {
4010 access |= SECTION_MAP_WRITE;
4011 pageprot = PAGE_READWRITE;
4013 if (flags & MDB_RDONLY) {
4014 secprot = PAGE_READONLY;
4018 secprot = PAGE_READWRITE;
4019 msize = env->me_mapsize;
4020 alloctype = MEM_RESERVE;
4023 rc = NtCreateSection(&mh, access, NULL, NULL, secprot, SEC_RESERVE, env->me_fd);
4027 rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
4033 int prot = PROT_READ;
4034 if (flags & MDB_WRITEMAP) {
4036 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
4039 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
4041 if (env->me_map == MAP_FAILED) {
4046 if (flags & MDB_NORDAHEAD) {
4047 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
4049 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
4051 #ifdef POSIX_MADV_RANDOM
4052 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
4053 #endif /* POSIX_MADV_RANDOM */
4054 #endif /* MADV_RANDOM */
4058 /* Can happen because the address argument to mmap() is just a
4059 * hint. mmap() can pick another, e.g. if the range is in use.
4060 * The MAP_FIXED flag would prevent that, but then mmap could
4061 * instead unmap existing pages to make room for the new map.
4063 if (addr && env->me_map != addr)
4064 return EBUSY; /* TODO: Make a new MDB_* error code? */
4066 p = (MDB_page *)env->me_map;
4067 env->me_metas[0] = METADATA(p);
4068 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4074 mdb_env_set_mapsize(MDB_env *env, size_t size)
4076 /* If env is already open, caller is responsible for making
4077 * sure there are no active txns.
4085 meta = mdb_env_pick_meta(env);
4087 size = meta->mm_mapsize;
4089 /* Silently round up to minimum if the size is too small */
4090 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4094 munmap(env->me_map, env->me_mapsize);
4095 env->me_mapsize = size;
4096 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4097 rc = mdb_env_map(env, old);
4101 env->me_mapsize = size;
4103 env->me_maxpg = env->me_mapsize / env->me_psize;
4108 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4112 env->me_maxdbs = dbs + CORE_DBS;
4117 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4119 if (env->me_map || readers < 1)
4121 env->me_maxreaders = readers;
4126 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4128 if (!env || !readers)
4130 *readers = env->me_maxreaders;
4135 mdb_fsize(HANDLE fd, size_t *size)
4138 LARGE_INTEGER fsize;
4140 if (!GetFileSizeEx(fd, &fsize))
4143 *size = fsize.QuadPart;
4155 #ifdef BROKEN_FDATASYNC
4156 #include <sys/utsname.h>
4157 #include <sys/vfs.h>
4160 /** Further setup required for opening an LMDB environment
4163 mdb_env_open2(MDB_env *env)
4165 unsigned int flags = env->me_flags;
4166 int i, newenv = 0, rc;
4170 /* See if we should use QueryLimited */
4172 if ((rc & 0xff) > 5)
4173 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4175 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4178 #ifdef BROKEN_FDATASYNC
4179 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4180 * https://lkml.org/lkml/2012/9/3/83
4181 * Kernels after 3.6-rc6 are known good.
4182 * https://lkml.org/lkml/2012/9/10/556
4183 * See if the DB is on ext3/ext4, then check for new enough kernel
4184 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4189 fstatfs(env->me_fd, &st);
4190 while (st.f_type == 0xEF53) {
4194 if (uts.release[0] < '3') {
4195 if (!strncmp(uts.release, "2.6.32.", 7)) {
4196 i = atoi(uts.release+7);
4198 break; /* 2.6.32.60 and newer is OK */
4199 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4200 i = atoi(uts.release+7);
4202 break; /* 2.6.34.15 and newer is OK */
4204 } else if (uts.release[0] == '3') {
4205 i = atoi(uts.release+2);
4207 break; /* 3.6 and newer is OK */
4209 i = atoi(uts.release+4);
4211 break; /* 3.5.4 and newer is OK */
4212 } else if (i == 2) {
4213 i = atoi(uts.release+4);
4215 break; /* 3.2.30 and newer is OK */
4217 } else { /* 4.x and newer is OK */
4220 env->me_flags |= MDB_FSYNCONLY;
4226 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4229 DPUTS("new mdbenv");
4231 env->me_psize = env->me_os_psize;
4232 if (env->me_psize > MAX_PAGESIZE)
4233 env->me_psize = MAX_PAGESIZE;
4234 memset(&meta, 0, sizeof(meta));
4235 mdb_env_init_meta0(env, &meta);
4236 meta.mm_mapsize = DEFAULT_MAPSIZE;
4238 env->me_psize = meta.mm_psize;
4241 /* Was a mapsize configured? */
4242 if (!env->me_mapsize) {
4243 env->me_mapsize = meta.mm_mapsize;
4246 /* Make sure mapsize >= committed data size. Even when using
4247 * mm_mapsize, which could be broken in old files (ITS#7789).
4249 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4250 if (env->me_mapsize < minsize)
4251 env->me_mapsize = minsize;
4253 meta.mm_mapsize = env->me_mapsize;
4255 if (newenv && !(flags & MDB_FIXEDMAP)) {
4256 /* mdb_env_map() may grow the datafile. Write the metapages
4257 * first, so the file will be valid if initialization fails.
4258 * Except with FIXEDMAP, since we do not yet know mm_address.
4259 * We could fill in mm_address later, but then a different
4260 * program might end up doing that - one with a memory layout
4261 * and map address which does not suit the main program.
4263 rc = mdb_env_init_meta(env, &meta);
4269 /* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
4273 rc = WriteFile(env->me_fd, &dummy, 1, &len, NULL);
4281 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4286 if (flags & MDB_FIXEDMAP)
4287 meta.mm_address = env->me_map;
4288 i = mdb_env_init_meta(env, &meta);
4289 if (i != MDB_SUCCESS) {
4294 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4295 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4297 #if !(MDB_MAXKEYSIZE)
4298 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4300 env->me_maxpg = env->me_mapsize / env->me_psize;
4304 MDB_meta *meta = mdb_env_pick_meta(env);
4305 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4307 DPRINTF(("opened database version %u, pagesize %u",
4308 meta->mm_version, env->me_psize));
4309 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4310 DPRINTF(("depth: %u", db->md_depth));
4311 DPRINTF(("entries: %"Z"u", db->md_entries));
4312 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4313 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4314 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4315 DPRINTF(("root: %"Z"u", db->md_root));
4323 /** Release a reader thread's slot in the reader lock table.
4324 * This function is called automatically when a thread exits.
4325 * @param[in] ptr This points to the slot in the reader lock table.
4328 mdb_env_reader_dest(void *ptr)
4330 MDB_reader *reader = ptr;
4336 /** Junk for arranging thread-specific callbacks on Windows. This is
4337 * necessarily platform and compiler-specific. Windows supports up
4338 * to 1088 keys. Let's assume nobody opens more than 64 environments
4339 * in a single process, for now. They can override this if needed.
4341 #ifndef MAX_TLS_KEYS
4342 #define MAX_TLS_KEYS 64
4344 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4345 static int mdb_tls_nkeys;
4347 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4351 case DLL_PROCESS_ATTACH: break;
4352 case DLL_THREAD_ATTACH: break;
4353 case DLL_THREAD_DETACH:
4354 for (i=0; i<mdb_tls_nkeys; i++) {
4355 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4357 mdb_env_reader_dest(r);
4361 case DLL_PROCESS_DETACH: break;
4366 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4368 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4372 /* Force some symbol references.
4373 * _tls_used forces the linker to create the TLS directory if not already done
4374 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4376 #pragma comment(linker, "/INCLUDE:_tls_used")
4377 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4378 #pragma const_seg(".CRT$XLB")
4379 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4380 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4383 #pragma comment(linker, "/INCLUDE:__tls_used")
4384 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4385 #pragma data_seg(".CRT$XLB")
4386 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4388 #endif /* WIN 32/64 */
4389 #endif /* !__GNUC__ */
4392 /** Downgrade the exclusive lock on the region back to shared */
4394 mdb_env_share_locks(MDB_env *env, int *excl)
4397 MDB_meta *meta = mdb_env_pick_meta(env);
4399 env->me_txns->mti_txnid = meta->mm_txnid;
4404 /* First acquire a shared lock. The Unlock will
4405 * then release the existing exclusive lock.
4407 memset(&ov, 0, sizeof(ov));
4408 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4411 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4417 struct flock lock_info;
4418 /* The shared lock replaces the existing lock */
4419 memset((void *)&lock_info, 0, sizeof(lock_info));
4420 lock_info.l_type = F_RDLCK;
4421 lock_info.l_whence = SEEK_SET;
4422 lock_info.l_start = 0;
4423 lock_info.l_len = 1;
4424 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4425 (rc = ErrCode()) == EINTR) ;
4426 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4433 /** Try to get exclusive lock, otherwise shared.
4434 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4437 mdb_env_excl_lock(MDB_env *env, int *excl)
4441 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4445 memset(&ov, 0, sizeof(ov));
4446 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4453 struct flock lock_info;
4454 memset((void *)&lock_info, 0, sizeof(lock_info));
4455 lock_info.l_type = F_WRLCK;
4456 lock_info.l_whence = SEEK_SET;
4457 lock_info.l_start = 0;
4458 lock_info.l_len = 1;
4459 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4460 (rc = ErrCode()) == EINTR) ;
4464 # ifndef MDB_USE_POSIX_MUTEX
4465 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4468 lock_info.l_type = F_RDLCK;
4469 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4470 (rc = ErrCode()) == EINTR) ;
4480 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4482 * @(#) $Revision: 5.1 $
4483 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4484 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4486 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4490 * Please do not copyright this code. This code is in the public domain.
4492 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4493 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4494 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4495 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4496 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4497 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4498 * PERFORMANCE OF THIS SOFTWARE.
4501 * chongo <Landon Curt Noll> /\oo/\
4502 * http://www.isthe.com/chongo/
4504 * Share and Enjoy! :-)
4507 typedef unsigned long long mdb_hash_t;
4508 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4510 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4511 * @param[in] val value to hash
4512 * @param[in] hval initial value for hash
4513 * @return 64 bit hash
4515 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4516 * hval arg on the first call.
4519 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4521 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4522 unsigned char *end = s + val->mv_size;
4524 * FNV-1a hash each octet of the string
4527 /* xor the bottom with the current octet */
4528 hval ^= (mdb_hash_t)*s++;
4530 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4531 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4532 (hval << 7) + (hval << 8) + (hval << 40);
4534 /* return our new hash value */
4538 /** Hash the string and output the encoded hash.
4539 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4540 * very short name limits. We don't care about the encoding being reversible,
4541 * we just want to preserve as many bits of the input as possible in a
4542 * small printable string.
4543 * @param[in] str string to hash
4544 * @param[out] encbuf an array of 11 chars to hold the hash
4546 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4549 mdb_pack85(unsigned long l, char *out)
4553 for (i=0; i<5; i++) {
4554 *out++ = mdb_a85[l % 85];
4560 mdb_hash_enc(MDB_val *val, char *encbuf)
4562 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4564 mdb_pack85(h, encbuf);
4565 mdb_pack85(h>>32, encbuf+5);
4570 /** Open and/or initialize the lock region for the environment.
4571 * @param[in] env The LMDB environment.
4572 * @param[in] lpath The pathname of the file used for the lock region.
4573 * @param[in] mode The Unix permissions for the file, if we create it.
4574 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4575 * @return 0 on success, non-zero on failure.
4578 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4581 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4583 # define MDB_ERRCODE_ROFS EROFS
4584 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4585 # define MDB_CLOEXEC O_CLOEXEC
4588 # define MDB_CLOEXEC 0
4591 #ifdef MDB_USE_SYSV_SEM
4600 utf8_to_utf16(lpath, -1, &wlpath, NULL);
4601 env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
4602 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4603 FILE_ATTRIBUTE_NORMAL, NULL);
4606 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4608 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4610 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4615 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4616 /* Lose record locks when exec*() */
4617 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4618 fcntl(env->me_lfd, F_SETFD, fdflags);
4621 if (!(env->me_flags & MDB_NOTLS)) {
4622 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4625 env->me_flags |= MDB_ENV_TXKEY;
4627 /* Windows TLS callbacks need help finding their TLS info. */
4628 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4632 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4636 /* Try to get exclusive lock. If we succeed, then
4637 * nobody is using the lock region and we should initialize it.
4639 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4642 size = GetFileSize(env->me_lfd, NULL);
4644 size = lseek(env->me_lfd, 0, SEEK_END);
4645 if (size == -1) goto fail_errno;
4647 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4648 if (size < rsize && *excl > 0) {
4650 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4651 || !SetEndOfFile(env->me_lfd))
4654 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4658 size = rsize - sizeof(MDB_txninfo);
4659 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4664 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4666 if (!mh) goto fail_errno;
4667 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4669 if (!env->me_txns) goto fail_errno;
4671 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4673 if (m == MAP_FAILED) goto fail_errno;
4679 BY_HANDLE_FILE_INFORMATION stbuf;
4688 if (!mdb_sec_inited) {
4689 InitializeSecurityDescriptor(&mdb_null_sd,
4690 SECURITY_DESCRIPTOR_REVISION);
4691 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4692 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4693 mdb_all_sa.bInheritHandle = FALSE;
4694 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4697 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4698 idbuf.volume = stbuf.dwVolumeSerialNumber;
4699 idbuf.nhigh = stbuf.nFileIndexHigh;
4700 idbuf.nlow = stbuf.nFileIndexLow;
4701 val.mv_data = &idbuf;
4702 val.mv_size = sizeof(idbuf);
4703 mdb_hash_enc(&val, encbuf);
4704 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4705 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4706 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4707 if (!env->me_rmutex) goto fail_errno;
4708 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4709 if (!env->me_wmutex) goto fail_errno;
4710 #elif defined(MDB_USE_POSIX_SEM)
4719 #if defined(__NetBSD__)
4720 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4722 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4723 idbuf.dev = stbuf.st_dev;
4724 idbuf.ino = stbuf.st_ino;
4725 val.mv_data = &idbuf;
4726 val.mv_size = sizeof(idbuf);
4727 mdb_hash_enc(&val, encbuf);
4728 #ifdef MDB_SHORT_SEMNAMES
4729 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4731 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4732 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4733 /* Clean up after a previous run, if needed: Try to
4734 * remove both semaphores before doing anything else.
4736 sem_unlink(env->me_txns->mti_rmname);
4737 sem_unlink(env->me_txns->mti_wmname);
4738 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4739 O_CREAT|O_EXCL, mode, 1);
4740 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4741 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4742 O_CREAT|O_EXCL, mode, 1);
4743 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4744 #elif defined(MDB_USE_SYSV_SEM)
4745 unsigned short vals[2] = {1, 1};
4746 key_t key = ftok(lpath, 'M');
4749 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4753 if (semctl(semid, 0, SETALL, semu) < 0)
4755 env->me_txns->mti_semid = semid;
4756 #else /* MDB_USE_POSIX_MUTEX: */
4757 pthread_mutexattr_t mattr;
4759 if ((rc = pthread_mutexattr_init(&mattr))
4760 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4761 #ifdef MDB_ROBUST_SUPPORTED
4762 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4764 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4765 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4767 pthread_mutexattr_destroy(&mattr);
4768 #endif /* _WIN32 || ... */
4770 env->me_txns->mti_magic = MDB_MAGIC;
4771 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4772 env->me_txns->mti_txnid = 0;
4773 env->me_txns->mti_numreaders = 0;
4776 #ifdef MDB_USE_SYSV_SEM
4777 struct semid_ds buf;
4779 if (env->me_txns->mti_magic != MDB_MAGIC) {
4780 DPUTS("lock region has invalid magic");
4784 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4785 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4786 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4787 rc = MDB_VERSION_MISMATCH;
4791 if (rc && rc != EACCES && rc != EAGAIN) {
4795 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4796 if (!env->me_rmutex) goto fail_errno;
4797 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4798 if (!env->me_wmutex) goto fail_errno;
4799 #elif defined(MDB_USE_POSIX_SEM)
4800 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4801 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4802 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4803 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4804 #elif defined(MDB_USE_SYSV_SEM)
4805 semid = env->me_txns->mti_semid;
4807 /* check for read access */
4808 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4810 /* check for write access */
4811 if (semctl(semid, 0, IPC_SET, semu) < 0)
4815 #ifdef MDB_USE_SYSV_SEM
4816 env->me_rmutex->semid = semid;
4817 env->me_wmutex->semid = semid;
4818 env->me_rmutex->semnum = 0;
4819 env->me_wmutex->semnum = 1;
4820 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4821 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4832 /** The name of the lock file in the DB environment */
4833 #define LOCKNAME "/lock.mdb"
4834 /** The name of the data file in the DB environment */
4835 #define DATANAME "/data.mdb"
4836 /** The suffix of the lock file when no subdir is used */
4837 #define LOCKSUFF "-lock"
4838 /** Only a subset of the @ref mdb_env flags can be changed
4839 * at runtime. Changing other flags requires closing the
4840 * environment and re-opening it with the new flags.
4842 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4843 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4844 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4846 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4847 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4851 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4853 int oflags, rc, len, excl = -1;
4854 char *lpath, *dpath;
4859 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4863 if (flags & MDB_NOSUBDIR) {
4864 rc = len + sizeof(LOCKSUFF) + len + 1;
4866 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4871 if (flags & MDB_NOSUBDIR) {
4872 dpath = lpath + len + sizeof(LOCKSUFF);
4873 sprintf(lpath, "%s" LOCKSUFF, path);
4874 strcpy(dpath, path);
4876 dpath = lpath + len + sizeof(LOCKNAME);
4877 sprintf(lpath, "%s" LOCKNAME, path);
4878 sprintf(dpath, "%s" DATANAME, path);
4882 flags |= env->me_flags;
4883 if (flags & MDB_RDONLY) {
4884 /* silently ignore WRITEMAP when we're only getting read access */
4885 flags &= ~MDB_WRITEMAP;
4887 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4888 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4891 env->me_flags = flags |= MDB_ENV_ACTIVE;
4895 env->me_path = strdup(path);
4896 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4897 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4898 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4899 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4903 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4905 /* For RDONLY, get lockfile after we know datafile exists */
4906 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4907 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4913 if (F_ISSET(flags, MDB_RDONLY)) {
4914 oflags = GENERIC_READ;
4915 len = OPEN_EXISTING;
4917 oflags = GENERIC_READ|GENERIC_WRITE;
4920 mode = FILE_ATTRIBUTE_NORMAL;
4921 utf8_to_utf16(dpath, -1, &wpath, NULL);
4922 env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4923 NULL, len, mode, NULL);
4926 if (F_ISSET(flags, MDB_RDONLY))
4929 oflags = O_RDWR | O_CREAT;
4931 env->me_fd = open(dpath, oflags, mode);
4933 if (env->me_fd == INVALID_HANDLE_VALUE) {
4938 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4939 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4944 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4945 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4946 env->me_mfd = env->me_fd;
4948 /* Synchronous fd for meta writes. Needed even with
4949 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4952 len = OPEN_EXISTING;
4953 utf8_to_utf16(dpath, -1, &wpath, NULL);
4954 env->me_mfd = CreateFileW(wpath, oflags,
4955 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4956 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4960 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4962 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4967 DPRINTF(("opened dbenv %p", (void *) env));
4969 rc = mdb_env_share_locks(env, &excl);
4973 if (!(flags & MDB_RDONLY)) {
4975 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4976 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4977 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4978 (txn = calloc(1, size)))
4980 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4981 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4982 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4983 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4985 txn->mt_dbxs = env->me_dbxs;
4986 txn->mt_flags = MDB_TXN_FINISHED;
4996 mdb_env_close0(env, excl);
5002 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5004 mdb_env_close0(MDB_env *env, int excl)
5008 if (!(env->me_flags & MDB_ENV_ACTIVE))
5011 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5013 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5014 free(env->me_dbxs[i].md_name.mv_data);
5019 free(env->me_dbiseqs);
5020 free(env->me_dbflags);
5022 free(env->me_dirty_list);
5024 mdb_midl_free(env->me_free_pgs);
5026 if (env->me_flags & MDB_ENV_TXKEY) {
5027 pthread_key_delete(env->me_txkey);
5029 /* Delete our key from the global list */
5030 for (i=0; i<mdb_tls_nkeys; i++)
5031 if (mdb_tls_keys[i] == env->me_txkey) {
5032 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5040 munmap(env->me_map, env->me_mapsize);
5042 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
5043 (void) close(env->me_mfd);
5044 if (env->me_fd != INVALID_HANDLE_VALUE)
5045 (void) close(env->me_fd);
5047 MDB_PID_T pid = env->me_pid;
5048 /* Clearing readers is done in this function because
5049 * me_txkey with its destructor must be disabled first.
5051 * We skip the the reader mutex, so we touch only
5052 * data owned by this process (me_close_readers and
5053 * our readers), and clear each reader atomically.
5055 for (i = env->me_close_readers; --i >= 0; )
5056 if (env->me_txns->mti_readers[i].mr_pid == pid)
5057 env->me_txns->mti_readers[i].mr_pid = 0;
5059 if (env->me_rmutex) {
5060 CloseHandle(env->me_rmutex);
5061 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5063 /* Windows automatically destroys the mutexes when
5064 * the last handle closes.
5066 #elif defined(MDB_USE_POSIX_SEM)
5067 if (env->me_rmutex != SEM_FAILED) {
5068 sem_close(env->me_rmutex);
5069 if (env->me_wmutex != SEM_FAILED)
5070 sem_close(env->me_wmutex);
5071 /* If we have the filelock: If we are the
5072 * only remaining user, clean up semaphores.
5075 mdb_env_excl_lock(env, &excl);
5077 sem_unlink(env->me_txns->mti_rmname);
5078 sem_unlink(env->me_txns->mti_wmname);
5081 #elif defined(MDB_USE_SYSV_SEM)
5082 if (env->me_rmutex->semid != -1) {
5083 /* If we have the filelock: If we are the
5084 * only remaining user, clean up semaphores.
5087 mdb_env_excl_lock(env, &excl);
5089 semctl(env->me_rmutex->semid, 0, IPC_RMID);
5092 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5094 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5097 /* Unlock the lockfile. Windows would have unlocked it
5098 * after closing anyway, but not necessarily at once.
5100 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5103 (void) close(env->me_lfd);
5106 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5110 mdb_env_close(MDB_env *env)
5117 VGMEMP_DESTROY(env);
5118 while ((dp = env->me_dpages) != NULL) {
5119 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5120 env->me_dpages = dp->mp_next;
5124 mdb_env_close0(env, 0);
5128 /** Compare two items pointing at aligned size_t's */
5130 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5132 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5133 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5136 /** Compare two items pointing at aligned unsigned int's.
5138 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5139 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5142 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5144 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5145 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5148 /** Compare two items pointing at unsigned ints of unknown alignment.
5149 * Nodes and keys are guaranteed to be 2-byte aligned.
5152 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5154 #if BYTE_ORDER == LITTLE_ENDIAN
5155 unsigned short *u, *c;
5158 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5159 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5162 } while(!x && u > (unsigned short *)a->mv_data);
5165 unsigned short *u, *c, *end;
5168 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5169 u = (unsigned short *)a->mv_data;
5170 c = (unsigned short *)b->mv_data;
5173 } while(!x && u < end);
5178 /** Compare two items lexically */
5180 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5187 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5193 diff = memcmp(a->mv_data, b->mv_data, len);
5194 return diff ? diff : len_diff<0 ? -1 : len_diff;
5197 /** Compare two items in reverse byte order */
5199 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5201 const unsigned char *p1, *p2, *p1_lim;
5205 p1_lim = (const unsigned char *)a->mv_data;
5206 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5207 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5209 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5215 while (p1 > p1_lim) {
5216 diff = *--p1 - *--p2;
5220 return len_diff<0 ? -1 : len_diff;
5223 /** Search for key within a page, using binary search.
5224 * Returns the smallest entry larger or equal to the key.
5225 * If exactp is non-null, stores whether the found entry was an exact match
5226 * in *exactp (1 or 0).
5227 * Updates the cursor index with the index of the found entry.
5228 * If no entry larger or equal to the key is found, returns NULL.
5231 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5233 unsigned int i = 0, nkeys;
5236 MDB_page *mp = mc->mc_pg[mc->mc_top];
5237 MDB_node *node = NULL;
5242 nkeys = NUMKEYS(mp);
5244 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5245 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5248 low = IS_LEAF(mp) ? 0 : 1;
5250 cmp = mc->mc_dbx->md_cmp;
5252 /* Branch pages have no data, so if using integer keys,
5253 * alignment is guaranteed. Use faster mdb_cmp_int.
5255 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5256 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5263 nodekey.mv_size = mc->mc_db->md_pad;
5264 node = NODEPTR(mp, 0); /* fake */
5265 while (low <= high) {
5266 i = (low + high) >> 1;
5267 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5268 rc = cmp(key, &nodekey);
5269 DPRINTF(("found leaf index %u [%s], rc = %i",
5270 i, DKEY(&nodekey), rc));
5279 while (low <= high) {
5280 i = (low + high) >> 1;
5282 node = NODEPTR(mp, i);
5283 nodekey.mv_size = NODEKSZ(node);
5284 nodekey.mv_data = NODEKEY(node);
5286 rc = cmp(key, &nodekey);
5289 DPRINTF(("found leaf index %u [%s], rc = %i",
5290 i, DKEY(&nodekey), rc));
5292 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5293 i, DKEY(&nodekey), NODEPGNO(node), rc));
5304 if (rc > 0) { /* Found entry is less than the key. */
5305 i++; /* Skip to get the smallest entry larger than key. */
5307 node = NODEPTR(mp, i);
5310 *exactp = (rc == 0 && nkeys > 0);
5311 /* store the key index */
5312 mc->mc_ki[mc->mc_top] = i;
5314 /* There is no entry larger or equal to the key. */
5317 /* nodeptr is fake for LEAF2 */
5323 mdb_cursor_adjust(MDB_cursor *mc, func)
5327 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5328 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5335 /** Pop a page off the top of the cursor's stack. */
5337 mdb_cursor_pop(MDB_cursor *mc)
5340 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5341 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5347 mc->mc_flags &= ~C_INITIALIZED;
5352 /** Push a page onto the top of the cursor's stack. */
5354 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5356 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5357 DDBI(mc), (void *) mc));
5359 if (mc->mc_snum >= CURSOR_STACK) {
5360 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5361 return MDB_CURSOR_FULL;
5364 mc->mc_top = mc->mc_snum++;
5365 mc->mc_pg[mc->mc_top] = mp;
5366 mc->mc_ki[mc->mc_top] = 0;
5371 /** Find the address of the page corresponding to a given page number.
5372 * @param[in] txn the transaction for this access.
5373 * @param[in] pgno the page number for the page to retrieve.
5374 * @param[out] ret address of a pointer where the page's address will be stored.
5375 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5376 * @return 0 on success, non-zero on failure.
5379 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5381 MDB_env *env = txn->mt_env;
5385 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5389 MDB_ID2L dl = tx2->mt_u.dirty_list;
5391 /* Spilled pages were dirtied in this txn and flushed
5392 * because the dirty list got full. Bring this page
5393 * back in from the map (but don't unspill it here,
5394 * leave that unless page_touch happens again).
5396 if (tx2->mt_spill_pgs) {
5397 MDB_ID pn = pgno << 1;
5398 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5399 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5400 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5405 unsigned x = mdb_mid2l_search(dl, pgno);
5406 if (x <= dl[0].mid && dl[x].mid == pgno) {
5412 } while ((tx2 = tx2->mt_parent) != NULL);
5415 if (pgno < txn->mt_next_pgno) {
5417 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5419 DPRINTF(("page %"Z"u not found", pgno));
5420 txn->mt_flags |= MDB_TXN_ERROR;
5421 return MDB_PAGE_NOTFOUND;
5431 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5432 * The cursor is at the root page, set up the rest of it.
5435 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5437 MDB_page *mp = mc->mc_pg[mc->mc_top];
5441 while (IS_BRANCH(mp)) {
5445 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5446 /* Don't assert on branch pages in the FreeDB. We can get here
5447 * while in the process of rebalancing a FreeDB branch page; we must
5448 * let that proceed. ITS#8336
5450 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
5451 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5453 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5455 if (flags & MDB_PS_LAST)
5456 i = NUMKEYS(mp) - 1;
5459 node = mdb_node_search(mc, key, &exact);
5461 i = NUMKEYS(mp) - 1;
5463 i = mc->mc_ki[mc->mc_top];
5465 mdb_cassert(mc, i > 0);
5469 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5472 mdb_cassert(mc, i < NUMKEYS(mp));
5473 node = NODEPTR(mp, i);
5475 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5478 mc->mc_ki[mc->mc_top] = i;
5479 if ((rc = mdb_cursor_push(mc, mp)))
5482 if (flags & MDB_PS_MODIFY) {
5483 if ((rc = mdb_page_touch(mc)) != 0)
5485 mp = mc->mc_pg[mc->mc_top];
5490 DPRINTF(("internal error, index points to a %02X page!?",
5492 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5493 return MDB_CORRUPTED;
5496 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5497 key ? DKEY(key) : "null"));
5498 mc->mc_flags |= C_INITIALIZED;
5499 mc->mc_flags &= ~C_EOF;
5504 /** Search for the lowest key under the current branch page.
5505 * This just bypasses a NUMKEYS check in the current page
5506 * before calling mdb_page_search_root(), because the callers
5507 * are all in situations where the current page is known to
5511 mdb_page_search_lowest(MDB_cursor *mc)
5513 MDB_page *mp = mc->mc_pg[mc->mc_top];
5514 MDB_node *node = NODEPTR(mp, 0);
5517 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5520 mc->mc_ki[mc->mc_top] = 0;
5521 if ((rc = mdb_cursor_push(mc, mp)))
5523 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5526 /** Search for the page a given key should be in.
5527 * Push it and its parent pages on the cursor stack.
5528 * @param[in,out] mc the cursor for this operation.
5529 * @param[in] key the key to search for, or NULL for first/last page.
5530 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5531 * are touched (updated with new page numbers).
5532 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5533 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5534 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5535 * @return 0 on success, non-zero on failure.
5538 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5543 /* Make sure the txn is still viable, then find the root from
5544 * the txn's db table and set it as the root of the cursor's stack.
5546 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5547 DPUTS("transaction may not be used now");
5550 /* Make sure we're using an up-to-date root */
5551 if (*mc->mc_dbflag & DB_STALE) {
5553 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5555 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5556 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5563 MDB_node *leaf = mdb_node_search(&mc2,
5564 &mc->mc_dbx->md_name, &exact);
5566 return MDB_NOTFOUND;
5567 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5568 return MDB_INCOMPATIBLE; /* not a named DB */
5569 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5572 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5574 /* The txn may not know this DBI, or another process may
5575 * have dropped and recreated the DB with other flags.
5577 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5578 return MDB_INCOMPATIBLE;
5579 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5581 *mc->mc_dbflag &= ~DB_STALE;
5583 root = mc->mc_db->md_root;
5585 if (root == P_INVALID) { /* Tree is empty. */
5586 DPUTS("tree is empty");
5587 return MDB_NOTFOUND;
5591 mdb_cassert(mc, root > 1);
5592 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5593 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5599 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5600 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5602 if (flags & MDB_PS_MODIFY) {
5603 if ((rc = mdb_page_touch(mc)))
5607 if (flags & MDB_PS_ROOTONLY)
5610 return mdb_page_search_root(mc, key, flags);
5614 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5616 MDB_txn *txn = mc->mc_txn;
5617 pgno_t pg = mp->mp_pgno;
5618 unsigned x = 0, ovpages = mp->mp_pages;
5619 MDB_env *env = txn->mt_env;
5620 MDB_IDL sl = txn->mt_spill_pgs;
5621 MDB_ID pn = pg << 1;
5624 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5625 /* If the page is dirty or on the spill list we just acquired it,
5626 * so we should give it back to our current free list, if any.
5627 * Otherwise put it onto the list of pages we freed in this txn.
5629 * Won't create me_pghead: me_pglast must be inited along with it.
5630 * Unsupported in nested txns: They would need to hide the page
5631 * range in ancestor txns' dirty and spilled lists.
5633 if (env->me_pghead &&
5635 ((mp->mp_flags & P_DIRTY) ||
5636 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5640 MDB_ID2 *dl, ix, iy;
5641 rc = mdb_midl_need(&env->me_pghead, ovpages);
5644 if (!(mp->mp_flags & P_DIRTY)) {
5645 /* This page is no longer spilled */
5652 /* Remove from dirty list */
5653 dl = txn->mt_u.dirty_list;
5655 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5661 mdb_cassert(mc, x > 1);
5663 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5664 txn->mt_flags |= MDB_TXN_ERROR;
5665 return MDB_CORRUPTED;
5668 txn->mt_dirty_room++;
5669 if (!(env->me_flags & MDB_WRITEMAP))
5670 mdb_dpage_free(env, mp);
5672 /* Insert in me_pghead */
5673 mop = env->me_pghead;
5674 j = mop[0] + ovpages;
5675 for (i = mop[0]; i && mop[i] < pg; i--)
5681 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5685 mc->mc_db->md_overflow_pages -= ovpages;
5689 /** Return the data associated with a given node.
5690 * @param[in] txn The transaction for this operation.
5691 * @param[in] leaf The node being read.
5692 * @param[out] data Updated to point to the node's data.
5693 * @return 0 on success, non-zero on failure.
5696 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5698 MDB_page *omp; /* overflow page */
5702 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5703 data->mv_size = NODEDSZ(leaf);
5704 data->mv_data = NODEDATA(leaf);
5708 /* Read overflow data.
5710 data->mv_size = NODEDSZ(leaf);
5711 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5712 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5713 DPRINTF(("read overflow page %"Z"u failed", pgno));
5716 data->mv_data = METADATA(omp);
5722 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5723 MDB_val *key, MDB_val *data)
5730 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5732 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5735 if (txn->mt_flags & MDB_TXN_BLOCKED)
5738 mdb_cursor_init(&mc, txn, dbi, &mx);
5739 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5742 /** Find a sibling for a page.
5743 * Replaces the page at the top of the cursor's stack with the
5744 * specified sibling, if one exists.
5745 * @param[in] mc The cursor for this operation.
5746 * @param[in] move_right Non-zero if the right sibling is requested,
5747 * otherwise the left sibling.
5748 * @return 0 on success, non-zero on failure.
5751 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5757 if (mc->mc_snum < 2) {
5758 return MDB_NOTFOUND; /* root has no siblings */
5762 DPRINTF(("parent page is page %"Z"u, index %u",
5763 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5765 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5766 : (mc->mc_ki[mc->mc_top] == 0)) {
5767 DPRINTF(("no more keys left, moving to %s sibling",
5768 move_right ? "right" : "left"));
5769 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5770 /* undo cursor_pop before returning */
5777 mc->mc_ki[mc->mc_top]++;
5779 mc->mc_ki[mc->mc_top]--;
5780 DPRINTF(("just moving to %s index key %u",
5781 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5783 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5785 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5786 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5787 /* mc will be inconsistent if caller does mc_snum++ as above */
5788 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5792 mdb_cursor_push(mc, mp);
5794 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5799 /** Move the cursor to the next data item. */
5801 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5807 if (mc->mc_flags & C_EOF) {
5808 return MDB_NOTFOUND;
5811 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5813 mp = mc->mc_pg[mc->mc_top];
5815 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5816 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5817 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5818 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5819 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5820 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5821 if (rc == MDB_SUCCESS)
5822 MDB_GET_KEY(leaf, key);
5827 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5828 if (op == MDB_NEXT_DUP)
5829 return MDB_NOTFOUND;
5833 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5834 mdb_dbg_pgno(mp), (void *) mc));
5835 if (mc->mc_flags & C_DEL)
5838 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5839 DPUTS("=====> move to next sibling page");
5840 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5841 mc->mc_flags |= C_EOF;
5844 mp = mc->mc_pg[mc->mc_top];
5845 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5847 mc->mc_ki[mc->mc_top]++;
5850 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5851 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5854 key->mv_size = mc->mc_db->md_pad;
5855 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5859 mdb_cassert(mc, IS_LEAF(mp));
5860 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5862 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5863 mdb_xcursor_init1(mc, leaf);
5866 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5869 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5870 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5871 if (rc != MDB_SUCCESS)
5876 MDB_GET_KEY(leaf, key);
5880 /** Move the cursor to the previous data item. */
5882 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5888 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5890 mp = mc->mc_pg[mc->mc_top];
5892 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5893 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5894 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5895 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5896 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5897 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5898 if (rc == MDB_SUCCESS) {
5899 MDB_GET_KEY(leaf, key);
5900 mc->mc_flags &= ~C_EOF;
5906 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5907 if (op == MDB_PREV_DUP)
5908 return MDB_NOTFOUND;
5912 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5913 mdb_dbg_pgno(mp), (void *) mc));
5915 if (mc->mc_ki[mc->mc_top] == 0) {
5916 DPUTS("=====> move to prev sibling page");
5917 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5920 mp = mc->mc_pg[mc->mc_top];
5921 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5922 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5924 mc->mc_ki[mc->mc_top]--;
5926 mc->mc_flags &= ~C_EOF;
5928 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5929 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5932 key->mv_size = mc->mc_db->md_pad;
5933 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5937 mdb_cassert(mc, IS_LEAF(mp));
5938 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5940 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5941 mdb_xcursor_init1(mc, leaf);
5944 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5947 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5948 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5949 if (rc != MDB_SUCCESS)
5954 MDB_GET_KEY(leaf, key);
5958 /** Set the cursor on a specific data item. */
5960 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5961 MDB_cursor_op op, int *exactp)
5965 MDB_node *leaf = NULL;
5968 if (key->mv_size == 0)
5969 return MDB_BAD_VALSIZE;
5972 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5974 /* See if we're already on the right page */
5975 if (mc->mc_flags & C_INITIALIZED) {
5978 mp = mc->mc_pg[mc->mc_top];
5980 mc->mc_ki[mc->mc_top] = 0;
5981 return MDB_NOTFOUND;
5983 if (mp->mp_flags & P_LEAF2) {
5984 nodekey.mv_size = mc->mc_db->md_pad;
5985 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5987 leaf = NODEPTR(mp, 0);
5988 MDB_GET_KEY2(leaf, nodekey);
5990 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5992 /* Probably happens rarely, but first node on the page
5993 * was the one we wanted.
5995 mc->mc_ki[mc->mc_top] = 0;
6002 unsigned int nkeys = NUMKEYS(mp);
6004 if (mp->mp_flags & P_LEAF2) {
6005 nodekey.mv_data = LEAF2KEY(mp,
6006 nkeys-1, nodekey.mv_size);
6008 leaf = NODEPTR(mp, nkeys-1);
6009 MDB_GET_KEY2(leaf, nodekey);
6011 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6013 /* last node was the one we wanted */
6014 mc->mc_ki[mc->mc_top] = nkeys-1;
6020 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6021 /* This is definitely the right page, skip search_page */
6022 if (mp->mp_flags & P_LEAF2) {
6023 nodekey.mv_data = LEAF2KEY(mp,
6024 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6026 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6027 MDB_GET_KEY2(leaf, nodekey);
6029 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6031 /* current node was the one we wanted */
6041 /* If any parents have right-sibs, search.
6042 * Otherwise, there's nothing further.
6044 for (i=0; i<mc->mc_top; i++)
6046 NUMKEYS(mc->mc_pg[i])-1)
6048 if (i == mc->mc_top) {
6049 /* There are no other pages */
6050 mc->mc_ki[mc->mc_top] = nkeys;
6051 return MDB_NOTFOUND;
6055 /* There are no other pages */
6056 mc->mc_ki[mc->mc_top] = 0;
6057 if (op == MDB_SET_RANGE && !exactp) {
6061 return MDB_NOTFOUND;
6067 rc = mdb_page_search(mc, key, 0);
6068 if (rc != MDB_SUCCESS)
6071 mp = mc->mc_pg[mc->mc_top];
6072 mdb_cassert(mc, IS_LEAF(mp));
6075 leaf = mdb_node_search(mc, key, exactp);
6076 if (exactp != NULL && !*exactp) {
6077 /* MDB_SET specified and not an exact match. */
6078 return MDB_NOTFOUND;
6082 DPUTS("===> inexact leaf not found, goto sibling");
6083 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6084 mc->mc_flags |= C_EOF;
6085 return rc; /* no entries matched */
6087 mp = mc->mc_pg[mc->mc_top];
6088 mdb_cassert(mc, IS_LEAF(mp));
6089 leaf = NODEPTR(mp, 0);
6093 mc->mc_flags |= C_INITIALIZED;
6094 mc->mc_flags &= ~C_EOF;
6097 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6098 key->mv_size = mc->mc_db->md_pad;
6099 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6104 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6105 mdb_xcursor_init1(mc, leaf);
6108 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6109 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6110 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6113 if (op == MDB_GET_BOTH) {
6119 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6120 if (rc != MDB_SUCCESS)
6123 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6126 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
6128 dcmp = mc->mc_dbx->md_dcmp;
6129 #if UINT_MAX < SIZE_MAX
6130 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6131 dcmp = mdb_cmp_clong;
6133 rc = dcmp(data, &olddata);
6135 if (op == MDB_GET_BOTH || rc > 0)
6136 return MDB_NOTFOUND;
6143 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6144 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6149 /* The key already matches in all other cases */
6150 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6151 MDB_GET_KEY(leaf, key);
6152 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6157 /** Move the cursor to the first item in the database. */
6159 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6165 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6167 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6168 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6169 if (rc != MDB_SUCCESS)
6172 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6174 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6175 mc->mc_flags |= C_INITIALIZED;
6176 mc->mc_flags &= ~C_EOF;
6178 mc->mc_ki[mc->mc_top] = 0;
6180 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6181 key->mv_size = mc->mc_db->md_pad;
6182 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6187 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6188 mdb_xcursor_init1(mc, leaf);
6189 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6193 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6197 MDB_GET_KEY(leaf, key);
6201 /** Move the cursor to the last item in the database. */
6203 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6209 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6211 if (!(mc->mc_flags & C_EOF)) {
6213 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6214 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6215 if (rc != MDB_SUCCESS)
6218 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6221 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6222 mc->mc_flags |= C_INITIALIZED|C_EOF;
6223 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6225 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6226 key->mv_size = mc->mc_db->md_pad;
6227 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6232 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6233 mdb_xcursor_init1(mc, leaf);
6234 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6238 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6243 MDB_GET_KEY(leaf, key);
6248 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6253 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6258 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6262 case MDB_GET_CURRENT:
6263 if (!(mc->mc_flags & C_INITIALIZED)) {
6266 MDB_page *mp = mc->mc_pg[mc->mc_top];
6267 int nkeys = NUMKEYS(mp);
6268 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6269 mc->mc_ki[mc->mc_top] = nkeys;
6275 key->mv_size = mc->mc_db->md_pad;
6276 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6278 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6279 MDB_GET_KEY(leaf, key);
6281 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6282 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6284 rc = mdb_node_read(mc->mc_txn, leaf, data);
6291 case MDB_GET_BOTH_RANGE:
6296 if (mc->mc_xcursor == NULL) {
6297 rc = MDB_INCOMPATIBLE;
6307 rc = mdb_cursor_set(mc, key, data, op,
6308 op == MDB_SET_RANGE ? NULL : &exact);
6311 case MDB_GET_MULTIPLE:
6312 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6316 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6317 rc = MDB_INCOMPATIBLE;
6321 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6322 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6325 case MDB_NEXT_MULTIPLE:
6330 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6331 rc = MDB_INCOMPATIBLE;
6334 if (!(mc->mc_flags & C_INITIALIZED))
6335 rc = mdb_cursor_first(mc, key, data);
6337 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6338 if (rc == MDB_SUCCESS) {
6339 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6342 mx = &mc->mc_xcursor->mx_cursor;
6343 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6345 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6346 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6354 case MDB_NEXT_NODUP:
6355 if (!(mc->mc_flags & C_INITIALIZED))
6356 rc = mdb_cursor_first(mc, key, data);
6358 rc = mdb_cursor_next(mc, key, data, op);
6362 case MDB_PREV_NODUP:
6363 if (!(mc->mc_flags & C_INITIALIZED)) {
6364 rc = mdb_cursor_last(mc, key, data);
6367 mc->mc_flags |= C_INITIALIZED;
6368 mc->mc_ki[mc->mc_top]++;
6370 rc = mdb_cursor_prev(mc, key, data, op);
6373 rc = mdb_cursor_first(mc, key, data);
6376 mfunc = mdb_cursor_first;
6378 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6382 if (mc->mc_xcursor == NULL) {
6383 rc = MDB_INCOMPATIBLE;
6387 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6388 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6389 MDB_GET_KEY(leaf, key);
6390 rc = mdb_node_read(mc->mc_txn, leaf, data);
6394 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6398 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6401 rc = mdb_cursor_last(mc, key, data);
6404 mfunc = mdb_cursor_last;
6407 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6412 if (mc->mc_flags & C_DEL)
6413 mc->mc_flags ^= C_DEL;
6418 /** Touch all the pages in the cursor stack. Set mc_top.
6419 * Makes sure all the pages are writable, before attempting a write operation.
6420 * @param[in] mc The cursor to operate on.
6423 mdb_cursor_touch(MDB_cursor *mc)
6425 int rc = MDB_SUCCESS;
6427 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6430 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6432 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6433 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6436 *mc->mc_dbflag |= DB_DIRTY;
6441 rc = mdb_page_touch(mc);
6442 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6443 mc->mc_top = mc->mc_snum-1;
6448 /** Do not spill pages to disk if txn is getting full, may fail instead */
6449 #define MDB_NOSPILL 0x8000
6452 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6456 MDB_node *leaf = NULL;
6457 MDB_page *fp, *mp, *sub_root = NULL;
6459 MDB_val xdata, *rdata, dkey, olddata;
6461 int do_sub = 0, insert_key, insert_data;
6462 unsigned int mcount = 0, dcount = 0, nospill;
6465 unsigned int nflags;
6468 if (mc == NULL || key == NULL)
6471 env = mc->mc_txn->mt_env;
6473 /* Check this first so counter will always be zero on any
6476 if (flags & MDB_MULTIPLE) {
6477 dcount = data[1].mv_size;
6478 data[1].mv_size = 0;
6479 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6480 return MDB_INCOMPATIBLE;
6483 nospill = flags & MDB_NOSPILL;
6484 flags &= ~MDB_NOSPILL;
6486 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6487 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6489 if (key->mv_size-1 >= ENV_MAXKEY(env))
6490 return MDB_BAD_VALSIZE;
6492 #if SIZE_MAX > MAXDATASIZE
6493 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6494 return MDB_BAD_VALSIZE;
6496 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6497 return MDB_BAD_VALSIZE;
6500 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6501 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6505 if (flags == MDB_CURRENT) {
6506 if (!(mc->mc_flags & C_INITIALIZED))
6509 } else if (mc->mc_db->md_root == P_INVALID) {
6510 /* new database, cursor has nothing to point to */
6513 mc->mc_flags &= ~C_INITIALIZED;
6518 if (flags & MDB_APPEND) {
6520 rc = mdb_cursor_last(mc, &k2, &d2);
6522 rc = mc->mc_dbx->md_cmp(key, &k2);
6525 mc->mc_ki[mc->mc_top]++;
6527 /* new key is <= last key */
6532 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6534 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6535 DPRINTF(("duplicate key [%s]", DKEY(key)));
6537 return MDB_KEYEXIST;
6539 if (rc && rc != MDB_NOTFOUND)
6543 if (mc->mc_flags & C_DEL)
6544 mc->mc_flags ^= C_DEL;
6546 /* Cursor is positioned, check for room in the dirty list */
6548 if (flags & MDB_MULTIPLE) {
6550 xdata.mv_size = data->mv_size * dcount;
6554 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6558 if (rc == MDB_NO_ROOT) {
6560 /* new database, write a root leaf page */
6561 DPUTS("allocating new root leaf page");
6562 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6565 mdb_cursor_push(mc, np);
6566 mc->mc_db->md_root = np->mp_pgno;
6567 mc->mc_db->md_depth++;
6568 *mc->mc_dbflag |= DB_DIRTY;
6569 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6571 np->mp_flags |= P_LEAF2;
6572 mc->mc_flags |= C_INITIALIZED;
6574 /* make sure all cursor pages are writable */
6575 rc2 = mdb_cursor_touch(mc);
6580 insert_key = insert_data = rc;
6582 /* The key does not exist */
6583 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6584 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6585 LEAFSIZE(key, data) > env->me_nodemax)
6587 /* Too big for a node, insert in sub-DB. Set up an empty
6588 * "old sub-page" for prep_subDB to expand to a full page.
6590 fp_flags = P_LEAF|P_DIRTY;
6592 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6593 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6594 olddata.mv_size = PAGEHDRSZ;
6598 /* there's only a key anyway, so this is a no-op */
6599 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6601 unsigned int ksize = mc->mc_db->md_pad;
6602 if (key->mv_size != ksize)
6603 return MDB_BAD_VALSIZE;
6604 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6605 memcpy(ptr, key->mv_data, ksize);
6607 /* if overwriting slot 0 of leaf, need to
6608 * update branch key if there is a parent page
6610 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6611 unsigned short dtop = 1;
6613 /* slot 0 is always an empty key, find real slot */
6614 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6618 if (mc->mc_ki[mc->mc_top])
6619 rc2 = mdb_update_key(mc, key);
6630 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6631 olddata.mv_size = NODEDSZ(leaf);
6632 olddata.mv_data = NODEDATA(leaf);
6635 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6636 /* Prepare (sub-)page/sub-DB to accept the new item,
6637 * if needed. fp: old sub-page or a header faking
6638 * it. mp: new (sub-)page. offset: growth in page
6639 * size. xdata: node data with new page or DB.
6641 unsigned i, offset = 0;
6642 mp = fp = xdata.mv_data = env->me_pbuf;
6643 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6645 /* Was a single item before, must convert now */
6646 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6648 /* Just overwrite the current item */
6649 if (flags == MDB_CURRENT)
6651 dcmp = mc->mc_dbx->md_dcmp;
6652 #if UINT_MAX < SIZE_MAX
6653 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6654 dcmp = mdb_cmp_clong;
6656 /* does data match? */
6657 if (!dcmp(data, &olddata)) {
6658 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
6659 return MDB_KEYEXIST;
6664 /* Back up original data item */
6665 dkey.mv_size = olddata.mv_size;
6666 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6668 /* Make sub-page header for the dup items, with dummy body */
6669 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6670 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6671 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6672 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6673 fp->mp_flags |= P_LEAF2;
6674 fp->mp_pad = data->mv_size;
6675 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6677 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6678 (dkey.mv_size & 1) + (data->mv_size & 1);
6680 fp->mp_upper = xdata.mv_size - PAGEBASE;
6681 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6682 } else if (leaf->mn_flags & F_SUBDATA) {
6683 /* Data is on sub-DB, just store it */
6684 flags |= F_DUPDATA|F_SUBDATA;
6687 /* Data is on sub-page */
6688 fp = olddata.mv_data;
6691 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6692 offset = EVEN(NODESIZE + sizeof(indx_t) +
6696 offset = fp->mp_pad;
6697 if (SIZELEFT(fp) < offset) {
6698 offset *= 4; /* space for 4 more */
6701 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6703 fp->mp_flags |= P_DIRTY;
6704 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6705 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6709 xdata.mv_size = olddata.mv_size + offset;
6712 fp_flags = fp->mp_flags;
6713 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6714 /* Too big for a sub-page, convert to sub-DB */
6715 fp_flags &= ~P_SUBP;
6717 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6718 fp_flags |= P_LEAF2;
6719 dummy.md_pad = fp->mp_pad;
6720 dummy.md_flags = MDB_DUPFIXED;
6721 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6722 dummy.md_flags |= MDB_INTEGERKEY;
6728 dummy.md_branch_pages = 0;
6729 dummy.md_leaf_pages = 1;
6730 dummy.md_overflow_pages = 0;
6731 dummy.md_entries = NUMKEYS(fp);
6732 xdata.mv_size = sizeof(MDB_db);
6733 xdata.mv_data = &dummy;
6734 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6736 offset = env->me_psize - olddata.mv_size;
6737 flags |= F_DUPDATA|F_SUBDATA;
6738 dummy.md_root = mp->mp_pgno;
6742 mp->mp_flags = fp_flags | P_DIRTY;
6743 mp->mp_pad = fp->mp_pad;
6744 mp->mp_lower = fp->mp_lower;
6745 mp->mp_upper = fp->mp_upper + offset;
6746 if (fp_flags & P_LEAF2) {
6747 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6749 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6750 olddata.mv_size - fp->mp_upper - PAGEBASE);
6751 for (i=0; i<NUMKEYS(fp); i++)
6752 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6760 mdb_node_del(mc, 0);
6764 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6765 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6766 return MDB_INCOMPATIBLE;
6767 /* overflow page overwrites need special handling */
6768 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6771 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6773 memcpy(&pg, olddata.mv_data, sizeof(pg));
6774 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6776 ovpages = omp->mp_pages;
6778 /* Is the ov page large enough? */
6779 if (ovpages >= dpages) {
6780 if (!(omp->mp_flags & P_DIRTY) &&
6781 (level || (env->me_flags & MDB_WRITEMAP)))
6783 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6786 level = 0; /* dirty in this txn or clean */
6789 if (omp->mp_flags & P_DIRTY) {
6790 /* yes, overwrite it. Note in this case we don't
6791 * bother to try shrinking the page if the new data
6792 * is smaller than the overflow threshold.
6795 /* It is writable only in a parent txn */
6796 size_t sz = (size_t) env->me_psize * ovpages, off;
6797 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6803 /* Note - this page is already counted in parent's dirty_room */
6804 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6805 mdb_cassert(mc, rc2 == 0);
6806 if (!(flags & MDB_RESERVE)) {
6807 /* Copy end of page, adjusting alignment so
6808 * compiler may copy words instead of bytes.
6810 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6811 memcpy((size_t *)((char *)np + off),
6812 (size_t *)((char *)omp + off), sz - off);
6815 memcpy(np, omp, sz); /* Copy beginning of page */
6818 SETDSZ(leaf, data->mv_size);
6819 if (F_ISSET(flags, MDB_RESERVE))
6820 data->mv_data = METADATA(omp);
6822 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6826 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6828 } else if (data->mv_size == olddata.mv_size) {
6829 /* same size, just replace it. Note that we could
6830 * also reuse this node if the new data is smaller,
6831 * but instead we opt to shrink the node in that case.
6833 if (F_ISSET(flags, MDB_RESERVE))
6834 data->mv_data = olddata.mv_data;
6835 else if (!(mc->mc_flags & C_SUB))
6836 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6838 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6843 mdb_node_del(mc, 0);
6849 nflags = flags & NODE_ADD_FLAGS;
6850 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6851 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6852 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6853 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6855 nflags |= MDB_SPLIT_REPLACE;
6856 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6858 /* There is room already in this leaf page. */
6859 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6861 /* Adjust other cursors pointing to mp */
6862 MDB_cursor *m2, *m3;
6863 MDB_dbi dbi = mc->mc_dbi;
6864 unsigned i = mc->mc_top;
6865 MDB_page *mp = mc->mc_pg[i];
6867 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6868 if (mc->mc_flags & C_SUB)
6869 m3 = &m2->mc_xcursor->mx_cursor;
6872 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
6873 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
6876 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6877 MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
6878 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
6879 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6885 if (rc == MDB_SUCCESS) {
6886 /* Now store the actual data in the child DB. Note that we're
6887 * storing the user data in the keys field, so there are strict
6888 * size limits on dupdata. The actual data fields of the child
6889 * DB are all zero size.
6892 int xflags, new_dupdata;
6897 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6898 if (flags & MDB_CURRENT) {
6899 xflags = MDB_CURRENT|MDB_NOSPILL;
6901 mdb_xcursor_init1(mc, leaf);
6902 xflags = (flags & MDB_NODUPDATA) ?
6903 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6906 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6907 new_dupdata = (int)dkey.mv_size;
6908 /* converted, write the original data first */
6910 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6913 /* we've done our job */
6916 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6917 /* Adjust other cursors pointing to mp */
6919 MDB_xcursor *mx = mc->mc_xcursor;
6920 unsigned i = mc->mc_top;
6921 MDB_page *mp = mc->mc_pg[i];
6922 int nkeys = NUMKEYS(mp);
6924 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6925 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6926 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6927 if (m2->mc_pg[i] == mp) {
6928 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6929 mdb_xcursor_init2(m2, mx, new_dupdata);
6930 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
6931 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6932 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
6933 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6938 ecount = mc->mc_xcursor->mx_db.md_entries;
6939 if (flags & MDB_APPENDDUP)
6940 xflags |= MDB_APPEND;
6941 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6942 if (flags & F_SUBDATA) {
6943 void *db = NODEDATA(leaf);
6944 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6946 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6948 /* Increment count unless we just replaced an existing item. */
6950 mc->mc_db->md_entries++;
6952 /* Invalidate txn if we created an empty sub-DB */
6955 /* If we succeeded and the key didn't exist before,
6956 * make sure the cursor is marked valid.
6958 mc->mc_flags |= C_INITIALIZED;
6960 if (flags & MDB_MULTIPLE) {
6963 /* let caller know how many succeeded, if any */
6964 data[1].mv_size = mcount;
6965 if (mcount < dcount) {
6966 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6967 insert_key = insert_data = 0;
6974 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6977 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6982 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6988 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6989 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6991 if (!(mc->mc_flags & C_INITIALIZED))
6994 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6995 return MDB_NOTFOUND;
6997 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7000 rc = mdb_cursor_touch(mc);
7004 mp = mc->mc_pg[mc->mc_top];
7007 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7009 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7010 if (flags & MDB_NODUPDATA) {
7011 /* mdb_cursor_del0() will subtract the final entry */
7012 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7013 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7015 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7016 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7018 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7021 /* If sub-DB still has entries, we're done */
7022 if (mc->mc_xcursor->mx_db.md_entries) {
7023 if (leaf->mn_flags & F_SUBDATA) {
7024 /* update subDB info */
7025 void *db = NODEDATA(leaf);
7026 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7029 /* shrink fake page */
7030 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7031 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7032 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7033 /* fix other sub-DB cursors pointed at fake pages on this page */
7034 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7035 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7036 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7037 if (m2->mc_pg[mc->mc_top] == mp) {
7038 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
7039 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7041 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7042 if (!(n2->mn_flags & F_SUBDATA))
7043 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7048 mc->mc_db->md_entries--;
7051 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7053 /* otherwise fall thru and delete the sub-DB */
7056 if (leaf->mn_flags & F_SUBDATA) {
7057 /* add all the child DB's pages to the free list */
7058 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7063 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7064 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7065 rc = MDB_INCOMPATIBLE;
7069 /* add overflow pages to free list */
7070 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7074 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7075 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7076 (rc = mdb_ovpage_free(mc, omp)))
7081 return mdb_cursor_del0(mc);
7084 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7088 /** Allocate and initialize new pages for a database.
7089 * @param[in] mc a cursor on the database being added to.
7090 * @param[in] flags flags defining what type of page is being allocated.
7091 * @param[in] num the number of pages to allocate. This is usually 1,
7092 * unless allocating overflow pages for a large record.
7093 * @param[out] mp Address of a page, or NULL on failure.
7094 * @return 0 on success, non-zero on failure.
7097 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7102 if ((rc = mdb_page_alloc(mc, num, &np)))
7104 DPRINTF(("allocated new mpage %"Z"u, page size %u",
7105 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7106 np->mp_flags = flags | P_DIRTY;
7107 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7108 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7111 mc->mc_db->md_branch_pages++;
7112 else if (IS_LEAF(np))
7113 mc->mc_db->md_leaf_pages++;
7114 else if (IS_OVERFLOW(np)) {
7115 mc->mc_db->md_overflow_pages += num;
7123 /** Calculate the size of a leaf node.
7124 * The size depends on the environment's page size; if a data item
7125 * is too large it will be put onto an overflow page and the node
7126 * size will only include the key and not the data. Sizes are always
7127 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7128 * of the #MDB_node headers.
7129 * @param[in] env The environment handle.
7130 * @param[in] key The key for the node.
7131 * @param[in] data The data for the node.
7132 * @return The number of bytes needed to store the node.
7135 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7139 sz = LEAFSIZE(key, data);
7140 if (sz > env->me_nodemax) {
7141 /* put on overflow page */
7142 sz -= data->mv_size - sizeof(pgno_t);
7145 return EVEN(sz + sizeof(indx_t));
7148 /** Calculate the size of a branch node.
7149 * The size should depend on the environment's page size but since
7150 * we currently don't support spilling large keys onto overflow
7151 * pages, it's simply the size of the #MDB_node header plus the
7152 * size of the key. Sizes are always rounded up to an even number
7153 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7154 * @param[in] env The environment handle.
7155 * @param[in] key The key for the node.
7156 * @return The number of bytes needed to store the node.
7159 mdb_branch_size(MDB_env *env, MDB_val *key)
7164 if (sz > env->me_nodemax) {
7165 /* put on overflow page */
7166 /* not implemented */
7167 /* sz -= key->size - sizeof(pgno_t); */
7170 return sz + sizeof(indx_t);
7173 /** Add a node to the page pointed to by the cursor.
7174 * @param[in] mc The cursor for this operation.
7175 * @param[in] indx The index on the page where the new node should be added.
7176 * @param[in] key The key for the new node.
7177 * @param[in] data The data for the new node, if any.
7178 * @param[in] pgno The page number, if adding a branch node.
7179 * @param[in] flags Flags for the node.
7180 * @return 0 on success, non-zero on failure. Possible errors are:
7182 * <li>ENOMEM - failed to allocate overflow pages for the node.
7183 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7184 * should never happen since all callers already calculate the
7185 * page's free space before calling this function.
7189 mdb_node_add(MDB_cursor *mc, indx_t indx,
7190 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7193 size_t node_size = NODESIZE;
7197 MDB_page *mp = mc->mc_pg[mc->mc_top];
7198 MDB_page *ofp = NULL; /* overflow page */
7202 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7204 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7205 IS_LEAF(mp) ? "leaf" : "branch",
7206 IS_SUBP(mp) ? "sub-" : "",
7207 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7208 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7211 /* Move higher keys up one slot. */
7212 int ksize = mc->mc_db->md_pad, dif;
7213 char *ptr = LEAF2KEY(mp, indx, ksize);
7214 dif = NUMKEYS(mp) - indx;
7216 memmove(ptr+ksize, ptr, dif*ksize);
7217 /* insert new key */
7218 memcpy(ptr, key->mv_data, ksize);
7220 /* Just using these for counting */
7221 mp->mp_lower += sizeof(indx_t);
7222 mp->mp_upper -= ksize - sizeof(indx_t);
7226 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7228 node_size += key->mv_size;
7230 mdb_cassert(mc, key && data);
7231 if (F_ISSET(flags, F_BIGDATA)) {
7232 /* Data already on overflow page. */
7233 node_size += sizeof(pgno_t);
7234 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7235 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7237 /* Put data on overflow page. */
7238 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7239 data->mv_size, node_size+data->mv_size));
7240 node_size = EVEN(node_size + sizeof(pgno_t));
7241 if ((ssize_t)node_size > room)
7243 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7245 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7249 node_size += data->mv_size;
7252 node_size = EVEN(node_size);
7253 if ((ssize_t)node_size > room)
7257 /* Move higher pointers up one slot. */
7258 for (i = NUMKEYS(mp); i > indx; i--)
7259 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7261 /* Adjust free space offsets. */
7262 ofs = mp->mp_upper - node_size;
7263 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7264 mp->mp_ptrs[indx] = ofs;
7266 mp->mp_lower += sizeof(indx_t);
7268 /* Write the node data. */
7269 node = NODEPTR(mp, indx);
7270 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7271 node->mn_flags = flags;
7273 SETDSZ(node,data->mv_size);
7278 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7281 ndata = NODEDATA(node);
7283 if (F_ISSET(flags, F_BIGDATA))
7284 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7285 else if (F_ISSET(flags, MDB_RESERVE))
7286 data->mv_data = ndata;
7288 memcpy(ndata, data->mv_data, data->mv_size);
7290 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7291 ndata = METADATA(ofp);
7292 if (F_ISSET(flags, MDB_RESERVE))
7293 data->mv_data = ndata;
7295 memcpy(ndata, data->mv_data, data->mv_size);
7302 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7303 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7304 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7305 DPRINTF(("node size = %"Z"u", node_size));
7306 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7307 return MDB_PAGE_FULL;
7310 /** Delete the specified node from a page.
7311 * @param[in] mc Cursor pointing to the node to delete.
7312 * @param[in] ksize The size of a node. Only used if the page is
7313 * part of a #MDB_DUPFIXED database.
7316 mdb_node_del(MDB_cursor *mc, int ksize)
7318 MDB_page *mp = mc->mc_pg[mc->mc_top];
7319 indx_t indx = mc->mc_ki[mc->mc_top];
7321 indx_t i, j, numkeys, ptr;
7325 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7326 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7327 numkeys = NUMKEYS(mp);
7328 mdb_cassert(mc, indx < numkeys);
7331 int x = numkeys - 1 - indx;
7332 base = LEAF2KEY(mp, indx, ksize);
7334 memmove(base, base + ksize, x * ksize);
7335 mp->mp_lower -= sizeof(indx_t);
7336 mp->mp_upper += ksize - sizeof(indx_t);
7340 node = NODEPTR(mp, indx);
7341 sz = NODESIZE + node->mn_ksize;
7343 if (F_ISSET(node->mn_flags, F_BIGDATA))
7344 sz += sizeof(pgno_t);
7346 sz += NODEDSZ(node);
7350 ptr = mp->mp_ptrs[indx];
7351 for (i = j = 0; i < numkeys; i++) {
7353 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7354 if (mp->mp_ptrs[i] < ptr)
7355 mp->mp_ptrs[j] += sz;
7360 base = (char *)mp + mp->mp_upper + PAGEBASE;
7361 memmove(base + sz, base, ptr - mp->mp_upper);
7363 mp->mp_lower -= sizeof(indx_t);
7367 /** Compact the main page after deleting a node on a subpage.
7368 * @param[in] mp The main page to operate on.
7369 * @param[in] indx The index of the subpage on the main page.
7372 mdb_node_shrink(MDB_page *mp, indx_t indx)
7377 indx_t delta, nsize, len, ptr;
7380 node = NODEPTR(mp, indx);
7381 sp = (MDB_page *)NODEDATA(node);
7382 delta = SIZELEFT(sp);
7383 nsize = NODEDSZ(node) - delta;
7385 /* Prepare to shift upward, set len = length(subpage part to shift) */
7389 return; /* do not make the node uneven-sized */
7391 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7392 for (i = NUMKEYS(sp); --i >= 0; )
7393 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7396 sp->mp_upper = sp->mp_lower;
7397 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7398 SETDSZ(node, nsize);
7400 /* Shift <lower nodes...initial part of subpage> upward */
7401 base = (char *)mp + mp->mp_upper + PAGEBASE;
7402 memmove(base + delta, base, (char *)sp + len - base);
7404 ptr = mp->mp_ptrs[indx];
7405 for (i = NUMKEYS(mp); --i >= 0; ) {
7406 if (mp->mp_ptrs[i] <= ptr)
7407 mp->mp_ptrs[i] += delta;
7409 mp->mp_upper += delta;
7412 /** Initial setup of a sorted-dups cursor.
7413 * Sorted duplicates are implemented as a sub-database for the given key.
7414 * The duplicate data items are actually keys of the sub-database.
7415 * Operations on the duplicate data items are performed using a sub-cursor
7416 * initialized when the sub-database is first accessed. This function does
7417 * the preliminary setup of the sub-cursor, filling in the fields that
7418 * depend only on the parent DB.
7419 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7422 mdb_xcursor_init0(MDB_cursor *mc)
7424 MDB_xcursor *mx = mc->mc_xcursor;
7426 mx->mx_cursor.mc_xcursor = NULL;
7427 mx->mx_cursor.mc_txn = mc->mc_txn;
7428 mx->mx_cursor.mc_db = &mx->mx_db;
7429 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7430 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7431 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7432 mx->mx_cursor.mc_snum = 0;
7433 mx->mx_cursor.mc_top = 0;
7434 mx->mx_cursor.mc_flags = C_SUB;
7435 mx->mx_dbx.md_name.mv_size = 0;
7436 mx->mx_dbx.md_name.mv_data = NULL;
7437 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7438 mx->mx_dbx.md_dcmp = NULL;
7439 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7442 /** Final setup of a sorted-dups cursor.
7443 * Sets up the fields that depend on the data from the main cursor.
7444 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7445 * @param[in] node The data containing the #MDB_db record for the
7446 * sorted-dup database.
7449 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7451 MDB_xcursor *mx = mc->mc_xcursor;
7453 if (node->mn_flags & F_SUBDATA) {
7454 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7455 mx->mx_cursor.mc_pg[0] = 0;
7456 mx->mx_cursor.mc_snum = 0;
7457 mx->mx_cursor.mc_top = 0;
7458 mx->mx_cursor.mc_flags = C_SUB;
7460 MDB_page *fp = NODEDATA(node);
7461 mx->mx_db.md_pad = 0;
7462 mx->mx_db.md_flags = 0;
7463 mx->mx_db.md_depth = 1;
7464 mx->mx_db.md_branch_pages = 0;
7465 mx->mx_db.md_leaf_pages = 1;
7466 mx->mx_db.md_overflow_pages = 0;
7467 mx->mx_db.md_entries = NUMKEYS(fp);
7468 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7469 mx->mx_cursor.mc_snum = 1;
7470 mx->mx_cursor.mc_top = 0;
7471 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7472 mx->mx_cursor.mc_pg[0] = fp;
7473 mx->mx_cursor.mc_ki[0] = 0;
7474 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7475 mx->mx_db.md_flags = MDB_DUPFIXED;
7476 mx->mx_db.md_pad = fp->mp_pad;
7477 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7478 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7481 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7482 mx->mx_db.md_root));
7483 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7484 #if UINT_MAX < SIZE_MAX
7485 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7486 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7491 /** Fixup a sorted-dups cursor due to underlying update.
7492 * Sets up some fields that depend on the data from the main cursor.
7493 * Almost the same as init1, but skips initialization steps if the
7494 * xcursor had already been used.
7495 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7496 * @param[in] src_mx The xcursor of an up-to-date cursor.
7497 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7500 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7502 MDB_xcursor *mx = mc->mc_xcursor;
7505 mx->mx_cursor.mc_snum = 1;
7506 mx->mx_cursor.mc_top = 0;
7507 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7508 mx->mx_cursor.mc_ki[0] = 0;
7509 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7510 #if UINT_MAX < SIZE_MAX
7511 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7513 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7516 mx->mx_db = src_mx->mx_db;
7517 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7518 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7519 mx->mx_db.md_root));
7522 /** Initialize a cursor for a given transaction and database. */
7524 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7527 mc->mc_backup = NULL;
7530 mc->mc_db = &txn->mt_dbs[dbi];
7531 mc->mc_dbx = &txn->mt_dbxs[dbi];
7532 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7538 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7539 mdb_tassert(txn, mx != NULL);
7540 mc->mc_xcursor = mx;
7541 mdb_xcursor_init0(mc);
7543 mc->mc_xcursor = NULL;
7545 if (*mc->mc_dbflag & DB_STALE) {
7546 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7551 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7554 size_t size = sizeof(MDB_cursor);
7556 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7559 if (txn->mt_flags & MDB_TXN_BLOCKED)
7562 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7565 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7566 size += sizeof(MDB_xcursor);
7568 if ((mc = malloc(size)) != NULL) {
7569 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7570 if (txn->mt_cursors) {
7571 mc->mc_next = txn->mt_cursors[dbi];
7572 txn->mt_cursors[dbi] = mc;
7573 mc->mc_flags |= C_UNTRACK;
7585 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7587 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7590 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7593 if (txn->mt_flags & MDB_TXN_BLOCKED)
7596 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7600 /* Return the count of duplicate data items for the current key */
7602 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7606 if (mc == NULL || countp == NULL)
7609 if (mc->mc_xcursor == NULL)
7610 return MDB_INCOMPATIBLE;
7612 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7615 if (!(mc->mc_flags & C_INITIALIZED))
7618 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7619 return MDB_NOTFOUND;
7621 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7622 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7625 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7628 *countp = mc->mc_xcursor->mx_db.md_entries;
7634 mdb_cursor_close(MDB_cursor *mc)
7636 if (mc && !mc->mc_backup) {
7637 /* remove from txn, if tracked */
7638 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7639 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7640 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7642 *prev = mc->mc_next;
7649 mdb_cursor_txn(MDB_cursor *mc)
7651 if (!mc) return NULL;
7656 mdb_cursor_dbi(MDB_cursor *mc)
7661 /** Replace the key for a branch node with a new key.
7662 * @param[in] mc Cursor pointing to the node to operate on.
7663 * @param[in] key The new key to use.
7664 * @return 0 on success, non-zero on failure.
7667 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7673 int delta, ksize, oksize;
7674 indx_t ptr, i, numkeys, indx;
7677 indx = mc->mc_ki[mc->mc_top];
7678 mp = mc->mc_pg[mc->mc_top];
7679 node = NODEPTR(mp, indx);
7680 ptr = mp->mp_ptrs[indx];
7684 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7685 k2.mv_data = NODEKEY(node);
7686 k2.mv_size = node->mn_ksize;
7687 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7689 mdb_dkey(&k2, kbuf2),
7695 /* Sizes must be 2-byte aligned. */
7696 ksize = EVEN(key->mv_size);
7697 oksize = EVEN(node->mn_ksize);
7698 delta = ksize - oksize;
7700 /* Shift node contents if EVEN(key length) changed. */
7702 if (delta > 0 && SIZELEFT(mp) < delta) {
7704 /* not enough space left, do a delete and split */
7705 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7706 pgno = NODEPGNO(node);
7707 mdb_node_del(mc, 0);
7708 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7711 numkeys = NUMKEYS(mp);
7712 for (i = 0; i < numkeys; i++) {
7713 if (mp->mp_ptrs[i] <= ptr)
7714 mp->mp_ptrs[i] -= delta;
7717 base = (char *)mp + mp->mp_upper + PAGEBASE;
7718 len = ptr - mp->mp_upper + NODESIZE;
7719 memmove(base - delta, base, len);
7720 mp->mp_upper -= delta;
7722 node = NODEPTR(mp, indx);
7725 /* But even if no shift was needed, update ksize */
7726 if (node->mn_ksize != key->mv_size)
7727 node->mn_ksize = key->mv_size;
7730 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7736 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7738 /** Perform \b act while tracking temporary cursor \b mn */
7739 #define WITH_CURSOR_TRACKING(mn, act) do { \
7740 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
7741 if ((mn).mc_flags & C_SUB) { \
7742 dummy.mc_flags = C_INITIALIZED; \
7743 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
7748 tracked->mc_next = *tp; \
7751 *tp = tracked->mc_next; \
7754 /** Move a node from csrc to cdst.
7757 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
7764 unsigned short flags;
7768 /* Mark src and dst as dirty. */
7769 if ((rc = mdb_page_touch(csrc)) ||
7770 (rc = mdb_page_touch(cdst)))
7773 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7774 key.mv_size = csrc->mc_db->md_pad;
7775 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7777 data.mv_data = NULL;
7781 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7782 mdb_cassert(csrc, !((size_t)srcnode & 1));
7783 srcpg = NODEPGNO(srcnode);
7784 flags = srcnode->mn_flags;
7785 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7786 unsigned int snum = csrc->mc_snum;
7788 /* must find the lowest key below src */
7789 rc = mdb_page_search_lowest(csrc);
7792 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7793 key.mv_size = csrc->mc_db->md_pad;
7794 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7796 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7797 key.mv_size = NODEKSZ(s2);
7798 key.mv_data = NODEKEY(s2);
7800 csrc->mc_snum = snum--;
7801 csrc->mc_top = snum;
7803 key.mv_size = NODEKSZ(srcnode);
7804 key.mv_data = NODEKEY(srcnode);
7806 data.mv_size = NODEDSZ(srcnode);
7807 data.mv_data = NODEDATA(srcnode);
7809 mn.mc_xcursor = NULL;
7810 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7811 unsigned int snum = cdst->mc_snum;
7814 /* must find the lowest key below dst */
7815 mdb_cursor_copy(cdst, &mn);
7816 rc = mdb_page_search_lowest(&mn);
7819 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7820 bkey.mv_size = mn.mc_db->md_pad;
7821 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7823 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7824 bkey.mv_size = NODEKSZ(s2);
7825 bkey.mv_data = NODEKEY(s2);
7827 mn.mc_snum = snum--;
7830 rc = mdb_update_key(&mn, &bkey);
7835 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7836 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7837 csrc->mc_ki[csrc->mc_top],
7839 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7840 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7842 /* Add the node to the destination page.
7844 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7845 if (rc != MDB_SUCCESS)
7848 /* Delete the node from the source page.
7850 mdb_node_del(csrc, key.mv_size);
7853 /* Adjust other cursors pointing to mp */
7854 MDB_cursor *m2, *m3;
7855 MDB_dbi dbi = csrc->mc_dbi;
7856 MDB_page *mpd, *mps;
7858 mps = csrc->mc_pg[csrc->mc_top];
7859 /* If we're adding on the left, bump others up */
7861 mpd = cdst->mc_pg[csrc->mc_top];
7862 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7863 if (csrc->mc_flags & C_SUB)
7864 m3 = &m2->mc_xcursor->mx_cursor;
7867 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7870 m3->mc_pg[csrc->mc_top] == mpd &&
7871 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7872 m3->mc_ki[csrc->mc_top]++;
7875 m3->mc_pg[csrc->mc_top] == mps &&
7876 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7877 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7878 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7879 m3->mc_ki[csrc->mc_top-1]++;
7881 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
7883 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7884 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
7885 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
7889 /* Adding on the right, bump others down */
7891 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7892 if (csrc->mc_flags & C_SUB)
7893 m3 = &m2->mc_xcursor->mx_cursor;
7896 if (m3 == csrc) continue;
7897 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7899 if (m3->mc_pg[csrc->mc_top] == mps) {
7900 if (!m3->mc_ki[csrc->mc_top]) {
7901 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7902 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7903 m3->mc_ki[csrc->mc_top-1]--;
7905 m3->mc_ki[csrc->mc_top]--;
7907 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
7909 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7910 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
7911 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
7918 /* Update the parent separators.
7920 if (csrc->mc_ki[csrc->mc_top] == 0) {
7921 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7922 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7923 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7925 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7926 key.mv_size = NODEKSZ(srcnode);
7927 key.mv_data = NODEKEY(srcnode);
7929 DPRINTF(("update separator for source page %"Z"u to [%s]",
7930 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7931 mdb_cursor_copy(csrc, &mn);
7934 /* We want mdb_rebalance to find mn when doing fixups */
7935 WITH_CURSOR_TRACKING(mn,
7936 rc = mdb_update_key(&mn, &key));
7940 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7942 indx_t ix = csrc->mc_ki[csrc->mc_top];
7943 nullkey.mv_size = 0;
7944 csrc->mc_ki[csrc->mc_top] = 0;
7945 rc = mdb_update_key(csrc, &nullkey);
7946 csrc->mc_ki[csrc->mc_top] = ix;
7947 mdb_cassert(csrc, rc == MDB_SUCCESS);
7951 if (cdst->mc_ki[cdst->mc_top] == 0) {
7952 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7953 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7954 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7956 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7957 key.mv_size = NODEKSZ(srcnode);
7958 key.mv_data = NODEKEY(srcnode);
7960 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7961 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7962 mdb_cursor_copy(cdst, &mn);
7965 /* We want mdb_rebalance to find mn when doing fixups */
7966 WITH_CURSOR_TRACKING(mn,
7967 rc = mdb_update_key(&mn, &key));
7971 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7973 indx_t ix = cdst->mc_ki[cdst->mc_top];
7974 nullkey.mv_size = 0;
7975 cdst->mc_ki[cdst->mc_top] = 0;
7976 rc = mdb_update_key(cdst, &nullkey);
7977 cdst->mc_ki[cdst->mc_top] = ix;
7978 mdb_cassert(cdst, rc == MDB_SUCCESS);
7985 /** Merge one page into another.
7986 * The nodes from the page pointed to by \b csrc will
7987 * be copied to the page pointed to by \b cdst and then
7988 * the \b csrc page will be freed.
7989 * @param[in] csrc Cursor pointing to the source page.
7990 * @param[in] cdst Cursor pointing to the destination page.
7991 * @return 0 on success, non-zero on failure.
7994 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7996 MDB_page *psrc, *pdst;
8003 psrc = csrc->mc_pg[csrc->mc_top];
8004 pdst = cdst->mc_pg[cdst->mc_top];
8006 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
8008 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8009 mdb_cassert(csrc, cdst->mc_snum > 1);
8011 /* Mark dst as dirty. */
8012 if ((rc = mdb_page_touch(cdst)))
8015 /* get dst page again now that we've touched it. */
8016 pdst = cdst->mc_pg[cdst->mc_top];
8018 /* Move all nodes from src to dst.
8020 j = nkeys = NUMKEYS(pdst);
8021 if (IS_LEAF2(psrc)) {
8022 key.mv_size = csrc->mc_db->md_pad;
8023 key.mv_data = METADATA(psrc);
8024 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8025 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8026 if (rc != MDB_SUCCESS)
8028 key.mv_data = (char *)key.mv_data + key.mv_size;
8031 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8032 srcnode = NODEPTR(psrc, i);
8033 if (i == 0 && IS_BRANCH(psrc)) {
8036 mdb_cursor_copy(csrc, &mn);
8037 mn.mc_xcursor = NULL;
8038 /* must find the lowest key below src */
8039 rc = mdb_page_search_lowest(&mn);
8042 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8043 key.mv_size = mn.mc_db->md_pad;
8044 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8046 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8047 key.mv_size = NODEKSZ(s2);
8048 key.mv_data = NODEKEY(s2);
8051 key.mv_size = srcnode->mn_ksize;
8052 key.mv_data = NODEKEY(srcnode);
8055 data.mv_size = NODEDSZ(srcnode);
8056 data.mv_data = NODEDATA(srcnode);
8057 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8058 if (rc != MDB_SUCCESS)
8063 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
8064 pdst->mp_pgno, NUMKEYS(pdst),
8065 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8067 /* Unlink the src page from parent and add to free list.
8070 mdb_node_del(csrc, 0);
8071 if (csrc->mc_ki[csrc->mc_top] == 0) {
8073 rc = mdb_update_key(csrc, &key);
8081 psrc = csrc->mc_pg[csrc->mc_top];
8082 /* If not operating on FreeDB, allow this page to be reused
8083 * in this txn. Otherwise just add to free list.
8085 rc = mdb_page_loose(csrc, psrc);
8089 csrc->mc_db->md_leaf_pages--;
8091 csrc->mc_db->md_branch_pages--;
8093 /* Adjust other cursors pointing to mp */
8094 MDB_cursor *m2, *m3;
8095 MDB_dbi dbi = csrc->mc_dbi;
8096 unsigned int top = csrc->mc_top;
8098 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8099 if (csrc->mc_flags & C_SUB)
8100 m3 = &m2->mc_xcursor->mx_cursor;
8103 if (m3 == csrc) continue;
8104 if (m3->mc_snum < csrc->mc_snum) continue;
8105 if (m3->mc_pg[top] == psrc) {
8106 m3->mc_pg[top] = pdst;
8107 m3->mc_ki[top] += nkeys;
8108 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8109 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8110 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8113 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8115 MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
8116 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8117 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8122 unsigned int snum = cdst->mc_snum;
8123 uint16_t depth = cdst->mc_db->md_depth;
8124 mdb_cursor_pop(cdst);
8125 rc = mdb_rebalance(cdst);
8126 /* Did the tree height change? */
8127 if (depth != cdst->mc_db->md_depth)
8128 snum += cdst->mc_db->md_depth - depth;
8129 cdst->mc_snum = snum;
8130 cdst->mc_top = snum-1;
8135 /** Copy the contents of a cursor.
8136 * @param[in] csrc The cursor to copy from.
8137 * @param[out] cdst The cursor to copy to.
8140 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8144 cdst->mc_txn = csrc->mc_txn;
8145 cdst->mc_dbi = csrc->mc_dbi;
8146 cdst->mc_db = csrc->mc_db;
8147 cdst->mc_dbx = csrc->mc_dbx;
8148 cdst->mc_snum = csrc->mc_snum;
8149 cdst->mc_top = csrc->mc_top;
8150 cdst->mc_flags = csrc->mc_flags;
8152 for (i=0; i<csrc->mc_snum; i++) {
8153 cdst->mc_pg[i] = csrc->mc_pg[i];
8154 cdst->mc_ki[i] = csrc->mc_ki[i];
8158 /** Rebalance the tree after a delete operation.
8159 * @param[in] mc Cursor pointing to the page where rebalancing
8161 * @return 0 on success, non-zero on failure.
8164 mdb_rebalance(MDB_cursor *mc)
8168 unsigned int ptop, minkeys, thresh;
8172 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8177 thresh = FILL_THRESHOLD;
8179 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
8180 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8181 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8182 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8184 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8185 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8186 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
8187 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8191 if (mc->mc_snum < 2) {
8192 MDB_page *mp = mc->mc_pg[0];
8194 DPUTS("Can't rebalance a subpage, ignoring");
8197 if (NUMKEYS(mp) == 0) {
8198 DPUTS("tree is completely empty");
8199 mc->mc_db->md_root = P_INVALID;
8200 mc->mc_db->md_depth = 0;
8201 mc->mc_db->md_leaf_pages = 0;
8202 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8205 /* Adjust cursors pointing to mp */
8208 mc->mc_flags &= ~C_INITIALIZED;
8210 MDB_cursor *m2, *m3;
8211 MDB_dbi dbi = mc->mc_dbi;
8213 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8214 if (mc->mc_flags & C_SUB)
8215 m3 = &m2->mc_xcursor->mx_cursor;
8218 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8220 if (m3->mc_pg[0] == mp) {
8223 m3->mc_flags &= ~C_INITIALIZED;
8227 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8229 DPUTS("collapsing root page!");
8230 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8233 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8234 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
8237 mc->mc_db->md_depth--;
8238 mc->mc_db->md_branch_pages--;
8239 mc->mc_ki[0] = mc->mc_ki[1];
8240 for (i = 1; i<mc->mc_db->md_depth; i++) {
8241 mc->mc_pg[i] = mc->mc_pg[i+1];
8242 mc->mc_ki[i] = mc->mc_ki[i+1];
8245 /* Adjust other cursors pointing to mp */
8246 MDB_cursor *m2, *m3;
8247 MDB_dbi dbi = mc->mc_dbi;
8249 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8250 if (mc->mc_flags & C_SUB)
8251 m3 = &m2->mc_xcursor->mx_cursor;
8254 if (m3 == mc) continue;
8255 if (!(m3->mc_flags & C_INITIALIZED))
8257 if (m3->mc_pg[0] == mp) {
8258 for (i=0; i<mc->mc_db->md_depth; i++) {
8259 m3->mc_pg[i] = m3->mc_pg[i+1];
8260 m3->mc_ki[i] = m3->mc_ki[i+1];
8268 DPUTS("root page doesn't need rebalancing");
8272 /* The parent (branch page) must have at least 2 pointers,
8273 * otherwise the tree is invalid.
8275 ptop = mc->mc_top-1;
8276 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8278 /* Leaf page fill factor is below the threshold.
8279 * Try to move keys from left or right neighbor, or
8280 * merge with a neighbor page.
8285 mdb_cursor_copy(mc, &mn);
8286 mn.mc_xcursor = NULL;
8288 oldki = mc->mc_ki[mc->mc_top];
8289 if (mc->mc_ki[ptop] == 0) {
8290 /* We're the leftmost leaf in our parent.
8292 DPUTS("reading right neighbor");
8294 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8295 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8298 mn.mc_ki[mn.mc_top] = 0;
8299 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8302 /* There is at least one neighbor to the left.
8304 DPUTS("reading left neighbor");
8306 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8307 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8310 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8311 mc->mc_ki[mc->mc_top] = 0;
8315 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8316 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8317 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8319 /* If the neighbor page is above threshold and has enough keys,
8320 * move one key from it. Otherwise we should try to merge them.
8321 * (A branch page must never have less than 2 keys.)
8323 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8324 rc = mdb_node_move(&mn, mc, fromleft);
8326 /* if we inserted on left, bump position up */
8331 rc = mdb_page_merge(&mn, mc);
8333 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8334 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8335 /* We want mdb_rebalance to find mn when doing fixups */
8336 WITH_CURSOR_TRACKING(mn,
8337 rc = mdb_page_merge(mc, &mn));
8338 mdb_cursor_copy(&mn, mc);
8340 mc->mc_flags &= ~C_EOF;
8342 mc->mc_ki[mc->mc_top] = oldki;
8346 /** Complete a delete operation started by #mdb_cursor_del(). */
8348 mdb_cursor_del0(MDB_cursor *mc)
8354 MDB_cursor *m2, *m3;
8355 MDB_dbi dbi = mc->mc_dbi;
8357 ki = mc->mc_ki[mc->mc_top];
8358 mp = mc->mc_pg[mc->mc_top];
8359 mdb_node_del(mc, mc->mc_db->md_pad);
8360 mc->mc_db->md_entries--;
8362 /* Adjust other cursors pointing to mp */
8363 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8364 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8365 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8367 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8369 if (m3->mc_pg[mc->mc_top] == mp) {
8370 if (m3->mc_ki[mc->mc_top] == ki) {
8371 m3->mc_flags |= C_DEL;
8372 if (mc->mc_db->md_flags & MDB_DUPSORT)
8373 m3->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
8374 } else if (m3->mc_ki[mc->mc_top] > ki) {
8375 m3->mc_ki[mc->mc_top]--;
8377 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
8378 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8379 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8380 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8385 rc = mdb_rebalance(mc);
8387 if (rc == MDB_SUCCESS) {
8388 /* DB is totally empty now, just bail out.
8389 * Other cursors adjustments were already done
8390 * by mdb_rebalance and aren't needed here.
8395 mp = mc->mc_pg[mc->mc_top];
8396 nkeys = NUMKEYS(mp);
8398 /* Adjust other cursors pointing to mp */
8399 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8400 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8401 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8403 if (m3->mc_snum < mc->mc_snum)
8405 if (m3->mc_pg[mc->mc_top] == mp) {
8406 /* if m3 points past last node in page, find next sibling */
8407 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8408 rc = mdb_cursor_sibling(m3, 1);
8409 if (rc == MDB_NOTFOUND) {
8410 m3->mc_flags |= C_EOF;
8416 mc->mc_flags |= C_DEL;
8420 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8425 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8426 MDB_val *key, MDB_val *data)
8428 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8431 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8432 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8434 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8435 /* must ignore any data */
8439 return mdb_del0(txn, dbi, key, data, 0);
8443 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8444 MDB_val *key, MDB_val *data, unsigned flags)
8449 MDB_val rdata, *xdata;
8453 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8455 mdb_cursor_init(&mc, txn, dbi, &mx);
8464 flags |= MDB_NODUPDATA;
8466 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8468 /* let mdb_page_split know about this cursor if needed:
8469 * delete will trigger a rebalance; if it needs to move
8470 * a node from one page to another, it will have to
8471 * update the parent's separator key(s). If the new sepkey
8472 * is larger than the current one, the parent page may
8473 * run out of space, triggering a split. We need this
8474 * cursor to be consistent until the end of the rebalance.
8476 mc.mc_flags |= C_UNTRACK;
8477 mc.mc_next = txn->mt_cursors[dbi];
8478 txn->mt_cursors[dbi] = &mc;
8479 rc = mdb_cursor_del(&mc, flags);
8480 txn->mt_cursors[dbi] = mc.mc_next;
8485 /** Split a page and insert a new node.
8486 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8487 * The cursor will be updated to point to the actual page and index where
8488 * the node got inserted after the split.
8489 * @param[in] newkey The key for the newly inserted node.
8490 * @param[in] newdata The data for the newly inserted node.
8491 * @param[in] newpgno The page number, if the new node is a branch node.
8492 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8493 * @return 0 on success, non-zero on failure.
8496 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8497 unsigned int nflags)
8500 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8503 int i, j, split_indx, nkeys, pmax;
8504 MDB_env *env = mc->mc_txn->mt_env;
8506 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8507 MDB_page *copy = NULL;
8508 MDB_page *mp, *rp, *pp;
8513 mp = mc->mc_pg[mc->mc_top];
8514 newindx = mc->mc_ki[mc->mc_top];
8515 nkeys = NUMKEYS(mp);
8517 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8518 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8519 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8521 /* Create a right sibling. */
8522 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8524 rp->mp_pad = mp->mp_pad;
8525 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8527 /* Usually when splitting the root page, the cursor
8528 * height is 1. But when called from mdb_update_key,
8529 * the cursor height may be greater because it walks
8530 * up the stack while finding the branch slot to update.
8532 if (mc->mc_top < 1) {
8533 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8535 /* shift current top to make room for new parent */
8536 for (i=mc->mc_snum; i>0; i--) {
8537 mc->mc_pg[i] = mc->mc_pg[i-1];
8538 mc->mc_ki[i] = mc->mc_ki[i-1];
8542 mc->mc_db->md_root = pp->mp_pgno;
8543 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8544 new_root = mc->mc_db->md_depth++;
8546 /* Add left (implicit) pointer. */
8547 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8548 /* undo the pre-push */
8549 mc->mc_pg[0] = mc->mc_pg[1];
8550 mc->mc_ki[0] = mc->mc_ki[1];
8551 mc->mc_db->md_root = mp->mp_pgno;
8552 mc->mc_db->md_depth--;
8559 ptop = mc->mc_top-1;
8560 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8563 mdb_cursor_copy(mc, &mn);
8564 mn.mc_xcursor = NULL;
8565 mn.mc_pg[mn.mc_top] = rp;
8566 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8568 if (nflags & MDB_APPEND) {
8569 mn.mc_ki[mn.mc_top] = 0;
8571 split_indx = newindx;
8575 split_indx = (nkeys+1) / 2;
8580 unsigned int lsize, rsize, ksize;
8581 /* Move half of the keys to the right sibling */
8582 x = mc->mc_ki[mc->mc_top] - split_indx;
8583 ksize = mc->mc_db->md_pad;
8584 split = LEAF2KEY(mp, split_indx, ksize);
8585 rsize = (nkeys - split_indx) * ksize;
8586 lsize = (nkeys - split_indx) * sizeof(indx_t);
8587 mp->mp_lower -= lsize;
8588 rp->mp_lower += lsize;
8589 mp->mp_upper += rsize - lsize;
8590 rp->mp_upper -= rsize - lsize;
8591 sepkey.mv_size = ksize;
8592 if (newindx == split_indx) {
8593 sepkey.mv_data = newkey->mv_data;
8595 sepkey.mv_data = split;
8598 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8599 memcpy(rp->mp_ptrs, split, rsize);
8600 sepkey.mv_data = rp->mp_ptrs;
8601 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8602 memcpy(ins, newkey->mv_data, ksize);
8603 mp->mp_lower += sizeof(indx_t);
8604 mp->mp_upper -= ksize - sizeof(indx_t);
8607 memcpy(rp->mp_ptrs, split, x * ksize);
8608 ins = LEAF2KEY(rp, x, ksize);
8609 memcpy(ins, newkey->mv_data, ksize);
8610 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8611 rp->mp_lower += sizeof(indx_t);
8612 rp->mp_upper -= ksize - sizeof(indx_t);
8613 mc->mc_ki[mc->mc_top] = x;
8616 int psize, nsize, k;
8617 /* Maximum free space in an empty page */
8618 pmax = env->me_psize - PAGEHDRSZ;
8620 nsize = mdb_leaf_size(env, newkey, newdata);
8622 nsize = mdb_branch_size(env, newkey);
8623 nsize = EVEN(nsize);
8625 /* grab a page to hold a temporary copy */
8626 copy = mdb_page_malloc(mc->mc_txn, 1);
8631 copy->mp_pgno = mp->mp_pgno;
8632 copy->mp_flags = mp->mp_flags;
8633 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8634 copy->mp_upper = env->me_psize - PAGEBASE;
8636 /* prepare to insert */
8637 for (i=0, j=0; i<nkeys; i++) {
8639 copy->mp_ptrs[j++] = 0;
8641 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8644 /* When items are relatively large the split point needs
8645 * to be checked, because being off-by-one will make the
8646 * difference between success or failure in mdb_node_add.
8648 * It's also relevant if a page happens to be laid out
8649 * such that one half of its nodes are all "small" and
8650 * the other half of its nodes are "large." If the new
8651 * item is also "large" and falls on the half with
8652 * "large" nodes, it also may not fit.
8654 * As a final tweak, if the new item goes on the last
8655 * spot on the page (and thus, onto the new page), bias
8656 * the split so the new page is emptier than the old page.
8657 * This yields better packing during sequential inserts.
8659 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8660 /* Find split point */
8662 if (newindx <= split_indx || newindx >= nkeys) {
8664 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8669 for (; i!=k; i+=j) {
8674 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8675 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8677 if (F_ISSET(node->mn_flags, F_BIGDATA))
8678 psize += sizeof(pgno_t);
8680 psize += NODEDSZ(node);
8682 psize = EVEN(psize);
8684 if (psize > pmax || i == k-j) {
8685 split_indx = i + (j<0);
8690 if (split_indx == newindx) {
8691 sepkey.mv_size = newkey->mv_size;
8692 sepkey.mv_data = newkey->mv_data;
8694 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8695 sepkey.mv_size = node->mn_ksize;
8696 sepkey.mv_data = NODEKEY(node);
8701 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8703 /* Copy separator key to the parent.
8705 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8706 int snum = mc->mc_snum;
8710 /* We want other splits to find mn when doing fixups */
8711 WITH_CURSOR_TRACKING(mn,
8712 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
8717 if (mc->mc_snum > snum) {
8720 /* Right page might now have changed parent.
8721 * Check if left page also changed parent.
8723 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8724 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8725 for (i=0; i<ptop; i++) {
8726 mc->mc_pg[i] = mn.mc_pg[i];
8727 mc->mc_ki[i] = mn.mc_ki[i];
8729 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8730 if (mn.mc_ki[ptop]) {
8731 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8733 /* find right page's left sibling */
8734 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8735 mdb_cursor_sibling(mc, 0);
8740 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8743 if (rc != MDB_SUCCESS) {
8746 if (nflags & MDB_APPEND) {
8747 mc->mc_pg[mc->mc_top] = rp;
8748 mc->mc_ki[mc->mc_top] = 0;
8749 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8752 for (i=0; i<mc->mc_top; i++)
8753 mc->mc_ki[i] = mn.mc_ki[i];
8754 } else if (!IS_LEAF2(mp)) {
8756 mc->mc_pg[mc->mc_top] = rp;
8761 rkey.mv_data = newkey->mv_data;
8762 rkey.mv_size = newkey->mv_size;
8768 /* Update index for the new key. */
8769 mc->mc_ki[mc->mc_top] = j;
8771 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8772 rkey.mv_data = NODEKEY(node);
8773 rkey.mv_size = node->mn_ksize;
8775 xdata.mv_data = NODEDATA(node);
8776 xdata.mv_size = NODEDSZ(node);
8779 pgno = NODEPGNO(node);
8780 flags = node->mn_flags;
8783 if (!IS_LEAF(mp) && j == 0) {
8784 /* First branch index doesn't need key data. */
8788 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8794 mc->mc_pg[mc->mc_top] = copy;
8799 } while (i != split_indx);
8801 nkeys = NUMKEYS(copy);
8802 for (i=0; i<nkeys; i++)
8803 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8804 mp->mp_lower = copy->mp_lower;
8805 mp->mp_upper = copy->mp_upper;
8806 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8807 env->me_psize - copy->mp_upper - PAGEBASE);
8809 /* reset back to original page */
8810 if (newindx < split_indx) {
8811 mc->mc_pg[mc->mc_top] = mp;
8813 mc->mc_pg[mc->mc_top] = rp;
8815 /* Make sure mc_ki is still valid.
8817 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8818 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8819 for (i=0; i<=ptop; i++) {
8820 mc->mc_pg[i] = mn.mc_pg[i];
8821 mc->mc_ki[i] = mn.mc_ki[i];
8825 if (nflags & MDB_RESERVE) {
8826 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8827 if (!(node->mn_flags & F_BIGDATA))
8828 newdata->mv_data = NODEDATA(node);
8831 if (newindx >= split_indx) {
8832 mc->mc_pg[mc->mc_top] = rp;
8834 /* Make sure mc_ki is still valid.
8836 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8837 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8838 for (i=0; i<=ptop; i++) {
8839 mc->mc_pg[i] = mn.mc_pg[i];
8840 mc->mc_ki[i] = mn.mc_ki[i];
8847 /* Adjust other cursors pointing to mp */
8848 MDB_cursor *m2, *m3;
8849 MDB_dbi dbi = mc->mc_dbi;
8850 nkeys = NUMKEYS(mp);
8852 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8853 if (mc->mc_flags & C_SUB)
8854 m3 = &m2->mc_xcursor->mx_cursor;
8859 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8863 /* sub cursors may be on different DB */
8864 if (m3->mc_pg[0] != mp)
8867 for (k=new_root; k>=0; k--) {
8868 m3->mc_ki[k+1] = m3->mc_ki[k];
8869 m3->mc_pg[k+1] = m3->mc_pg[k];
8871 if (m3->mc_ki[0] >= nkeys) {
8876 m3->mc_pg[0] = mc->mc_pg[0];
8880 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8881 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8882 m3->mc_ki[mc->mc_top]++;
8883 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8884 m3->mc_pg[mc->mc_top] = rp;
8885 m3->mc_ki[mc->mc_top] -= nkeys;
8886 for (i=0; i<mc->mc_top; i++) {
8887 m3->mc_ki[i] = mn.mc_ki[i];
8888 m3->mc_pg[i] = mn.mc_pg[i];
8891 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8892 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8895 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8897 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8898 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8899 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8903 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8906 if (copy) /* tmp page */
8907 mdb_page_free(env, copy);
8909 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8914 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8915 MDB_val *key, MDB_val *data, unsigned int flags)
8921 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8924 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8927 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8928 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8930 mdb_cursor_init(&mc, txn, dbi, &mx);
8931 mc.mc_next = txn->mt_cursors[dbi];
8932 txn->mt_cursors[dbi] = &mc;
8933 rc = mdb_cursor_put(&mc, key, data, flags);
8934 txn->mt_cursors[dbi] = mc.mc_next;
8939 #define MDB_WBUF (1024*1024)
8942 /** State needed for a compacting copy. */
8943 typedef struct mdb_copy {
8944 pthread_mutex_t mc_mutex;
8945 pthread_cond_t mc_cond;
8952 pgno_t mc_next_pgno;
8955 volatile int mc_new;
8960 /** Dedicated writer thread for compacting copy. */
8961 static THREAD_RET ESECT CALL_CONV
8962 mdb_env_copythr(void *arg)
8966 int toggle = 0, wsize, rc;
8969 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8972 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8975 pthread_mutex_lock(&my->mc_mutex);
8977 pthread_cond_signal(&my->mc_cond);
8980 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8981 if (my->mc_new < 0) {
8986 wsize = my->mc_wlen[toggle];
8987 ptr = my->mc_wbuf[toggle];
8990 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8994 } else if (len > 0) {
9008 /* If there's an overflow page tail, write it too */
9009 if (my->mc_olen[toggle]) {
9010 wsize = my->mc_olen[toggle];
9011 ptr = my->mc_over[toggle];
9012 my->mc_olen[toggle] = 0;
9015 my->mc_wlen[toggle] = 0;
9017 pthread_cond_signal(&my->mc_cond);
9019 pthread_cond_signal(&my->mc_cond);
9020 pthread_mutex_unlock(&my->mc_mutex);
9021 return (THREAD_RET)0;
9025 /** Tell the writer thread there's a buffer ready to write */
9027 mdb_env_cthr_toggle(mdb_copy *my, int st)
9029 int toggle = my->mc_toggle ^ 1;
9030 pthread_mutex_lock(&my->mc_mutex);
9031 if (my->mc_status) {
9032 pthread_mutex_unlock(&my->mc_mutex);
9033 return my->mc_status;
9035 while (my->mc_new == 1)
9036 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9038 my->mc_toggle = toggle;
9039 pthread_cond_signal(&my->mc_cond);
9040 pthread_mutex_unlock(&my->mc_mutex);
9044 /** Depth-first tree traversal for compacting copy. */
9046 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9049 MDB_txn *txn = my->mc_txn;
9051 MDB_page *mo, *mp, *leaf;
9056 /* Empty DB, nothing to do */
9057 if (*pg == P_INVALID)
9064 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
9067 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9071 /* Make cursor pages writable */
9072 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9076 for (i=0; i<mc.mc_top; i++) {
9077 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9078 mc.mc_pg[i] = (MDB_page *)ptr;
9079 ptr += my->mc_env->me_psize;
9082 /* This is writable space for a leaf page. Usually not needed. */
9083 leaf = (MDB_page *)ptr;
9085 toggle = my->mc_toggle;
9086 while (mc.mc_snum > 0) {
9088 mp = mc.mc_pg[mc.mc_top];
9092 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9093 for (i=0; i<n; i++) {
9094 ni = NODEPTR(mp, i);
9095 if (ni->mn_flags & F_BIGDATA) {
9099 /* Need writable leaf */
9101 mc.mc_pg[mc.mc_top] = leaf;
9102 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9104 ni = NODEPTR(mp, i);
9107 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9108 rc = mdb_page_get(txn, pg, &omp, NULL);
9111 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9112 rc = mdb_env_cthr_toggle(my, 1);
9115 toggle = my->mc_toggle;
9117 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9118 memcpy(mo, omp, my->mc_env->me_psize);
9119 mo->mp_pgno = my->mc_next_pgno;
9120 my->mc_next_pgno += omp->mp_pages;
9121 my->mc_wlen[toggle] += my->mc_env->me_psize;
9122 if (omp->mp_pages > 1) {
9123 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9124 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9125 rc = mdb_env_cthr_toggle(my, 1);
9128 toggle = my->mc_toggle;
9130 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
9131 } else if (ni->mn_flags & F_SUBDATA) {
9134 /* Need writable leaf */
9136 mc.mc_pg[mc.mc_top] = leaf;
9137 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9139 ni = NODEPTR(mp, i);
9142 memcpy(&db, NODEDATA(ni), sizeof(db));
9143 my->mc_toggle = toggle;
9144 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9147 toggle = my->mc_toggle;
9148 memcpy(NODEDATA(ni), &db, sizeof(db));
9153 mc.mc_ki[mc.mc_top]++;
9154 if (mc.mc_ki[mc.mc_top] < n) {
9157 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9159 rc = mdb_page_get(txn, pg, &mp, NULL);
9164 mc.mc_ki[mc.mc_top] = 0;
9165 if (IS_BRANCH(mp)) {
9166 /* Whenever we advance to a sibling branch page,
9167 * we must proceed all the way down to its first leaf.
9169 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9172 mc.mc_pg[mc.mc_top] = mp;
9176 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9177 rc = mdb_env_cthr_toggle(my, 1);
9180 toggle = my->mc_toggle;
9182 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9183 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9184 mo->mp_pgno = my->mc_next_pgno++;
9185 my->mc_wlen[toggle] += my->mc_env->me_psize;
9187 /* Update parent if there is one */
9188 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9189 SETPGNO(ni, mo->mp_pgno);
9190 mdb_cursor_pop(&mc);
9192 /* Otherwise we're done */
9202 /** Copy environment with compaction. */
9204 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9209 MDB_txn *txn = NULL;
9214 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
9215 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
9216 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9217 if (my.mc_wbuf[0] == NULL)
9220 pthread_mutex_init(&my.mc_mutex, NULL);
9221 pthread_cond_init(&my.mc_cond, NULL);
9222 #ifdef HAVE_MEMALIGN
9223 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9224 if (my.mc_wbuf[0] == NULL)
9227 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
9232 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9233 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9238 my.mc_next_pgno = NUM_METAS;
9244 THREAD_CREATE(thr, mdb_env_copythr, &my);
9246 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9250 mp = (MDB_page *)my.mc_wbuf[0];
9251 memset(mp, 0, NUM_METAS * env->me_psize);
9253 mp->mp_flags = P_META;
9254 mm = (MDB_meta *)METADATA(mp);
9255 mdb_env_init_meta0(env, mm);
9256 mm->mm_address = env->me_metas[0]->mm_address;
9258 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9260 mp->mp_flags = P_META;
9261 *(MDB_meta *)METADATA(mp) = *mm;
9262 mm = (MDB_meta *)METADATA(mp);
9264 /* Count the number of free pages, subtract from lastpg to find
9265 * number of active pages
9268 MDB_ID freecount = 0;
9271 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9272 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9273 freecount += *(MDB_ID *)data.mv_data;
9274 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9275 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9276 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9278 /* Set metapage 1 */
9279 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
9280 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9281 if (mm->mm_last_pg > NUM_METAS-1) {
9282 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
9285 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
9288 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9290 pthread_mutex_lock(&my.mc_mutex);
9292 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9293 pthread_mutex_unlock(&my.mc_mutex);
9294 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
9295 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
9296 rc = mdb_env_cthr_toggle(&my, 1);
9297 mdb_env_cthr_toggle(&my, -1);
9298 pthread_mutex_lock(&my.mc_mutex);
9300 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9301 pthread_mutex_unlock(&my.mc_mutex);
9306 CloseHandle(my.mc_cond);
9307 CloseHandle(my.mc_mutex);
9308 _aligned_free(my.mc_wbuf[0]);
9310 pthread_cond_destroy(&my.mc_cond);
9311 pthread_mutex_destroy(&my.mc_mutex);
9312 free(my.mc_wbuf[0]);
9317 /** Copy environment as-is. */
9319 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9321 MDB_txn *txn = NULL;
9322 mdb_mutexref_t wmutex = NULL;
9328 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9332 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9335 /* Do the lock/unlock of the reader mutex before starting the
9336 * write txn. Otherwise other read txns could block writers.
9338 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9343 /* We must start the actual read txn after blocking writers */
9344 mdb_txn_end(txn, MDB_END_RESET_TMP);
9346 /* Temporarily block writers until we snapshot the meta pages */
9347 wmutex = env->me_wmutex;
9348 if (LOCK_MUTEX(rc, env, wmutex))
9351 rc = mdb_txn_renew0(txn);
9353 UNLOCK_MUTEX(wmutex);
9358 wsize = env->me_psize * NUM_METAS;
9362 DO_WRITE(rc, fd, ptr, w2, len);
9366 } else if (len > 0) {
9372 /* Non-blocking or async handles are not supported */
9378 UNLOCK_MUTEX(wmutex);
9383 w2 = txn->mt_next_pgno * env->me_psize;
9386 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9393 if (wsize > MAX_WRITE)
9397 DO_WRITE(rc, fd, ptr, w2, len);
9401 } else if (len > 0) {
9418 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9420 if (flags & MDB_CP_COMPACT)
9421 return mdb_env_copyfd1(env, fd);
9423 return mdb_env_copyfd0(env, fd);
9427 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9429 return mdb_env_copyfd2(env, fd, 0);
9433 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9437 HANDLE newfd = INVALID_HANDLE_VALUE;
9442 if (env->me_flags & MDB_NOSUBDIR) {
9443 lpath = (char *)path;
9446 len += sizeof(DATANAME);
9447 lpath = malloc(len);
9450 sprintf(lpath, "%s" DATANAME, path);
9453 /* The destination path must exist, but the destination file must not.
9454 * We don't want the OS to cache the writes, since the source data is
9455 * already in the OS cache.
9458 utf8_to_utf16(lpath, -1, &wpath, NULL);
9459 newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9460 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9463 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9465 if (newfd == INVALID_HANDLE_VALUE) {
9470 if (env->me_psize >= env->me_os_psize) {
9472 /* Set O_DIRECT if the file system supports it */
9473 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9474 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9476 #ifdef F_NOCACHE /* __APPLE__ */
9477 rc = fcntl(newfd, F_NOCACHE, 1);
9485 rc = mdb_env_copyfd2(env, newfd, flags);
9488 if (!(env->me_flags & MDB_NOSUBDIR))
9490 if (newfd != INVALID_HANDLE_VALUE)
9491 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9498 mdb_env_copy(MDB_env *env, const char *path)
9500 return mdb_env_copy2(env, path, 0);
9504 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9506 if (flag & ~CHANGEABLE)
9509 env->me_flags |= flag;
9511 env->me_flags &= ~flag;
9516 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9521 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9526 mdb_env_set_userctx(MDB_env *env, void *ctx)
9530 env->me_userctx = ctx;
9535 mdb_env_get_userctx(MDB_env *env)
9537 return env ? env->me_userctx : NULL;
9541 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9546 env->me_assert_func = func;
9552 mdb_env_get_path(MDB_env *env, const char **arg)
9557 *arg = env->me_path;
9562 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9571 /** Common code for #mdb_stat() and #mdb_env_stat().
9572 * @param[in] env the environment to operate in.
9573 * @param[in] db the #MDB_db record containing the stats to return.
9574 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9575 * @return 0, this function always succeeds.
9578 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9580 arg->ms_psize = env->me_psize;
9581 arg->ms_depth = db->md_depth;
9582 arg->ms_branch_pages = db->md_branch_pages;
9583 arg->ms_leaf_pages = db->md_leaf_pages;
9584 arg->ms_overflow_pages = db->md_overflow_pages;
9585 arg->ms_entries = db->md_entries;
9591 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9595 if (env == NULL || arg == NULL)
9598 meta = mdb_env_pick_meta(env);
9600 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9604 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9608 if (env == NULL || arg == NULL)
9611 meta = mdb_env_pick_meta(env);
9612 arg->me_mapaddr = meta->mm_address;
9613 arg->me_last_pgno = meta->mm_last_pg;
9614 arg->me_last_txnid = meta->mm_txnid;
9616 arg->me_mapsize = env->me_mapsize;
9617 arg->me_maxreaders = env->me_maxreaders;
9618 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9622 /** Set the default comparison functions for a database.
9623 * Called immediately after a database is opened to set the defaults.
9624 * The user can then override them with #mdb_set_compare() or
9625 * #mdb_set_dupsort().
9626 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9627 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9630 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9632 uint16_t f = txn->mt_dbs[dbi].md_flags;
9634 txn->mt_dbxs[dbi].md_cmp =
9635 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9636 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9638 txn->mt_dbxs[dbi].md_dcmp =
9639 !(f & MDB_DUPSORT) ? 0 :
9640 ((f & MDB_INTEGERDUP)
9641 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9642 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9645 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9651 int rc, dbflag, exact;
9652 unsigned int unused = 0, seq;
9656 if (flags & ~VALID_FLAGS)
9658 if (txn->mt_flags & MDB_TXN_BLOCKED)
9664 if (flags & PERSISTENT_FLAGS) {
9665 uint16_t f2 = flags & PERSISTENT_FLAGS;
9666 /* make sure flag changes get committed */
9667 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9668 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9669 txn->mt_flags |= MDB_TXN_DIRTY;
9672 mdb_default_cmp(txn, MAIN_DBI);
9676 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9677 mdb_default_cmp(txn, MAIN_DBI);
9680 /* Is the DB already open? */
9682 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9683 if (!txn->mt_dbxs[i].md_name.mv_size) {
9684 /* Remember this free slot */
9685 if (!unused) unused = i;
9688 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9689 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9695 /* If no free slot and max hit, fail */
9696 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9697 return MDB_DBS_FULL;
9699 /* Cannot mix named databases with some mainDB flags */
9700 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9701 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9703 /* Find the DB info */
9704 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9707 key.mv_data = (void *)name;
9708 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9709 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9710 if (rc == MDB_SUCCESS) {
9711 /* make sure this is actually a DB */
9712 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9713 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9714 return MDB_INCOMPATIBLE;
9715 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
9719 /* Done here so we cannot fail after creating a new DB */
9720 if ((namedup = strdup(name)) == NULL)
9724 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
9725 data.mv_size = sizeof(MDB_db);
9726 data.mv_data = &dummy;
9727 memset(&dummy, 0, sizeof(dummy));
9728 dummy.md_root = P_INVALID;
9729 dummy.md_flags = flags & PERSISTENT_FLAGS;
9730 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9737 /* Got info, register DBI in this txn */
9738 unsigned int slot = unused ? unused : txn->mt_numdbs;
9739 txn->mt_dbxs[slot].md_name.mv_data = namedup;
9740 txn->mt_dbxs[slot].md_name.mv_size = len;
9741 txn->mt_dbxs[slot].md_rel = NULL;
9742 txn->mt_dbflags[slot] = dbflag;
9743 /* txn-> and env-> are the same in read txns, use
9744 * tmp variable to avoid undefined assignment
9746 seq = ++txn->mt_env->me_dbiseqs[slot];
9747 txn->mt_dbiseqs[slot] = seq;
9749 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9751 mdb_default_cmp(txn, slot);
9761 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9763 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9766 if (txn->mt_flags & MDB_TXN_BLOCKED)
9769 if (txn->mt_dbflags[dbi] & DB_STALE) {
9772 /* Stale, must read the DB's root. cursor_init does it for us. */
9773 mdb_cursor_init(&mc, txn, dbi, &mx);
9775 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9778 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9781 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9783 ptr = env->me_dbxs[dbi].md_name.mv_data;
9784 /* If there was no name, this was already closed */
9786 env->me_dbxs[dbi].md_name.mv_data = NULL;
9787 env->me_dbxs[dbi].md_name.mv_size = 0;
9788 env->me_dbflags[dbi] = 0;
9789 env->me_dbiseqs[dbi]++;
9794 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9796 /* We could return the flags for the FREE_DBI too but what's the point? */
9797 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9799 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9803 /** Add all the DB's pages to the free list.
9804 * @param[in] mc Cursor on the DB to free.
9805 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9806 * @return 0 on success, non-zero on failure.
9809 mdb_drop0(MDB_cursor *mc, int subs)
9813 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9814 if (rc == MDB_SUCCESS) {
9815 MDB_txn *txn = mc->mc_txn;
9820 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9821 * This also avoids any P_LEAF2 pages, which have no nodes.
9823 if (mc->mc_flags & C_SUB)
9826 mdb_cursor_copy(mc, &mx);
9827 while (mc->mc_snum > 0) {
9828 MDB_page *mp = mc->mc_pg[mc->mc_top];
9829 unsigned n = NUMKEYS(mp);
9831 for (i=0; i<n; i++) {
9832 ni = NODEPTR(mp, i);
9833 if (ni->mn_flags & F_BIGDATA) {
9836 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9837 rc = mdb_page_get(txn, pg, &omp, NULL);
9840 mdb_cassert(mc, IS_OVERFLOW(omp));
9841 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9845 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9846 mdb_xcursor_init1(mc, ni);
9847 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9853 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9855 for (i=0; i<n; i++) {
9857 ni = NODEPTR(mp, i);
9860 mdb_midl_xappend(txn->mt_free_pgs, pg);
9865 mc->mc_ki[mc->mc_top] = i;
9866 rc = mdb_cursor_sibling(mc, 1);
9868 if (rc != MDB_NOTFOUND)
9870 /* no more siblings, go back to beginning
9871 * of previous level.
9875 for (i=1; i<mc->mc_snum; i++) {
9877 mc->mc_pg[i] = mx.mc_pg[i];
9882 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9885 txn->mt_flags |= MDB_TXN_ERROR;
9886 } else if (rc == MDB_NOTFOUND) {
9889 mc->mc_flags &= ~C_INITIALIZED;
9893 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9895 MDB_cursor *mc, *m2;
9898 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9901 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9904 if (TXN_DBI_CHANGED(txn, dbi))
9907 rc = mdb_cursor_open(txn, dbi, &mc);
9911 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9912 /* Invalidate the dropped DB's cursors */
9913 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9914 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9918 /* Can't delete the main DB */
9919 if (del && dbi >= CORE_DBS) {
9920 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9922 txn->mt_dbflags[dbi] = DB_STALE;
9923 mdb_dbi_close(txn->mt_env, dbi);
9925 txn->mt_flags |= MDB_TXN_ERROR;
9928 /* reset the DB record, mark it dirty */
9929 txn->mt_dbflags[dbi] |= DB_DIRTY;
9930 txn->mt_dbs[dbi].md_depth = 0;
9931 txn->mt_dbs[dbi].md_branch_pages = 0;
9932 txn->mt_dbs[dbi].md_leaf_pages = 0;
9933 txn->mt_dbs[dbi].md_overflow_pages = 0;
9934 txn->mt_dbs[dbi].md_entries = 0;
9935 txn->mt_dbs[dbi].md_root = P_INVALID;
9937 txn->mt_flags |= MDB_TXN_DIRTY;
9940 mdb_cursor_close(mc);
9944 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9946 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9949 txn->mt_dbxs[dbi].md_cmp = cmp;
9953 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9955 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9958 txn->mt_dbxs[dbi].md_dcmp = cmp;
9962 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9964 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9967 txn->mt_dbxs[dbi].md_rel = rel;
9971 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9973 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9976 txn->mt_dbxs[dbi].md_relctx = ctx;
9981 mdb_env_get_maxkeysize(MDB_env *env)
9983 return ENV_MAXKEY(env);
9987 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9989 unsigned int i, rdrs;
9992 int rc = 0, first = 1;
9996 if (!env->me_txns) {
9997 return func("(no reader locks)\n", ctx);
9999 rdrs = env->me_txns->mti_numreaders;
10000 mr = env->me_txns->mti_readers;
10001 for (i=0; i<rdrs; i++) {
10002 if (mr[i].mr_pid) {
10003 txnid_t txnid = mr[i].mr_txnid;
10004 sprintf(buf, txnid == (txnid_t)-1 ?
10005 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
10006 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10009 rc = func(" pid thread txnid\n", ctx);
10013 rc = func(buf, ctx);
10019 rc = func("(no active readers)\n", ctx);
10024 /** Insert pid into list if not already present.
10025 * return -1 if already present.
10028 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10030 /* binary search of pid in list */
10032 unsigned cursor = 1;
10034 unsigned n = ids[0];
10037 unsigned pivot = n >> 1;
10038 cursor = base + pivot + 1;
10039 val = pid - ids[cursor];
10044 } else if ( val > 0 ) {
10049 /* found, so it's a duplicate */
10058 for (n = ids[0]; n > cursor; n--)
10065 mdb_reader_check(MDB_env *env, int *dead)
10071 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10074 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
10076 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10078 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10079 unsigned int i, j, rdrs;
10081 MDB_PID_T *pids, pid;
10082 int rc = MDB_SUCCESS, count = 0;
10084 rdrs = env->me_txns->mti_numreaders;
10085 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10089 mr = env->me_txns->mti_readers;
10090 for (i=0; i<rdrs; i++) {
10091 pid = mr[i].mr_pid;
10092 if (pid && pid != env->me_pid) {
10093 if (mdb_pid_insert(pids, pid) == 0) {
10094 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10095 /* Stale reader found */
10098 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10099 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10101 rdrs = 0; /* the above checked all readers */
10103 /* Recheck, a new process may have reused pid */
10104 if (mdb_reader_pid(env, Pidcheck, pid))
10108 for (; j<rdrs; j++)
10109 if (mr[j].mr_pid == pid) {
10110 DPRINTF(("clear stale reader pid %u txn %"Z"d",
10111 (unsigned) pid, mr[j].mr_txnid));
10116 UNLOCK_MUTEX(rmutex);
10127 #ifdef MDB_ROBUST_SUPPORTED
10128 /** Handle #LOCK_MUTEX0() failure.
10129 * Try to repair the lock file if the mutex owner died.
10130 * @param[in] env the environment handle
10131 * @param[in] mutex LOCK_MUTEX0() mutex
10132 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10133 * @return 0 on success with the mutex locked, or an error code on failure.
10136 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10141 if (rc == MDB_OWNERDEAD) {
10142 /* We own the mutex. Clean up after dead previous owner. */
10144 rlocked = (mutex == env->me_rmutex);
10146 /* Keep mti_txnid updated, otherwise next writer can
10147 * overwrite data which latest meta page refers to.
10149 meta = mdb_env_pick_meta(env);
10150 env->me_txns->mti_txnid = meta->mm_txnid;
10151 /* env is hosed if the dead thread was ours */
10153 env->me_flags |= MDB_FATAL_ERROR;
10154 env->me_txn = NULL;
10158 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10159 (rc ? "this process' env is hosed" : "recovering")));
10160 rc2 = mdb_reader_check0(env, rlocked, NULL);
10162 rc2 = mdb_mutex_consistent(mutex);
10163 if (rc || (rc = rc2)) {
10164 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10165 UNLOCK_MUTEX(mutex);
10171 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10176 #endif /* MDB_ROBUST_SUPPORTED */
10179 #if defined(_WIN32)
10180 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
10184 need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
10185 if (need == 0xFFFD)
10189 result = malloc(sizeof(wchar_t) * need);
10190 MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
10196 #endif /* defined(_WIN32) */