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 < 2.10 only provided _np API */
304 # if defined(__GLIBC__) && GLIBC_VER < 0x02000a
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");
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;
4008 if (flags & MDB_WRITEMAP) {
4009 access |= SECTION_MAP_WRITE;
4010 pageprot = PAGE_READWRITE;
4013 rc = NtCreateSection(&mh, access, NULL, NULL, PAGE_READWRITE, SEC_RESERVE, env->me_fd);
4017 msize = env->me_mapsize;
4018 rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, MEM_RESERVE, pageprot);
4024 int prot = PROT_READ;
4025 if (flags & MDB_WRITEMAP) {
4027 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
4030 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
4032 if (env->me_map == MAP_FAILED) {
4037 if (flags & MDB_NORDAHEAD) {
4038 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
4040 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
4042 #ifdef POSIX_MADV_RANDOM
4043 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
4044 #endif /* POSIX_MADV_RANDOM */
4045 #endif /* MADV_RANDOM */
4049 /* Can happen because the address argument to mmap() is just a
4050 * hint. mmap() can pick another, e.g. if the range is in use.
4051 * The MAP_FIXED flag would prevent that, but then mmap could
4052 * instead unmap existing pages to make room for the new map.
4054 if (addr && env->me_map != addr)
4055 return EBUSY; /* TODO: Make a new MDB_* error code? */
4057 p = (MDB_page *)env->me_map;
4058 env->me_metas[0] = METADATA(p);
4059 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4065 mdb_env_set_mapsize(MDB_env *env, size_t size)
4067 /* If env is already open, caller is responsible for making
4068 * sure there are no active txns.
4076 meta = mdb_env_pick_meta(env);
4078 size = meta->mm_mapsize;
4080 /* Silently round up to minimum if the size is too small */
4081 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4085 munmap(env->me_map, env->me_mapsize);
4086 env->me_mapsize = size;
4087 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4088 rc = mdb_env_map(env, old);
4092 env->me_mapsize = size;
4094 env->me_maxpg = env->me_mapsize / env->me_psize;
4099 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4103 env->me_maxdbs = dbs + CORE_DBS;
4108 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4110 if (env->me_map || readers < 1)
4112 env->me_maxreaders = readers;
4117 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4119 if (!env || !readers)
4121 *readers = env->me_maxreaders;
4126 mdb_fsize(HANDLE fd, size_t *size)
4129 LARGE_INTEGER fsize;
4131 if (!GetFileSizeEx(fd, &fsize))
4134 *size = fsize.QuadPart;
4146 #ifdef BROKEN_FDATASYNC
4147 #include <sys/utsname.h>
4148 #include <sys/vfs.h>
4151 /** Further setup required for opening an LMDB environment
4154 mdb_env_open2(MDB_env *env)
4156 unsigned int flags = env->me_flags;
4157 int i, newenv = 0, rc;
4161 /* See if we should use QueryLimited */
4163 if ((rc & 0xff) > 5)
4164 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4166 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4169 #ifdef BROKEN_FDATASYNC
4170 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4171 * https://lkml.org/lkml/2012/9/3/83
4172 * Kernels after 3.6-rc6 are known good.
4173 * https://lkml.org/lkml/2012/9/10/556
4174 * See if the DB is on ext3/ext4, then check for new enough kernel
4175 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4180 fstatfs(env->me_fd, &st);
4181 while (st.f_type == 0xEF53) {
4185 if (uts.release[0] < '3') {
4186 if (!strncmp(uts.release, "2.6.32.", 7)) {
4187 i = atoi(uts.release+7);
4189 break; /* 2.6.32.60 and newer is OK */
4190 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4191 i = atoi(uts.release+7);
4193 break; /* 2.6.34.15 and newer is OK */
4195 } else if (uts.release[0] == '3') {
4196 i = atoi(uts.release+2);
4198 break; /* 3.6 and newer is OK */
4200 i = atoi(uts.release+4);
4202 break; /* 3.5.4 and newer is OK */
4203 } else if (i == 2) {
4204 i = atoi(uts.release+4);
4206 break; /* 3.2.30 and newer is OK */
4208 } else { /* 4.x and newer is OK */
4211 env->me_flags |= MDB_FSYNCONLY;
4217 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4220 DPUTS("new mdbenv");
4222 env->me_psize = env->me_os_psize;
4223 if (env->me_psize > MAX_PAGESIZE)
4224 env->me_psize = MAX_PAGESIZE;
4225 memset(&meta, 0, sizeof(meta));
4226 mdb_env_init_meta0(env, &meta);
4227 meta.mm_mapsize = DEFAULT_MAPSIZE;
4229 env->me_psize = meta.mm_psize;
4232 /* Was a mapsize configured? */
4233 if (!env->me_mapsize) {
4234 env->me_mapsize = meta.mm_mapsize;
4237 /* Make sure mapsize >= committed data size. Even when using
4238 * mm_mapsize, which could be broken in old files (ITS#7789).
4240 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4241 if (env->me_mapsize < minsize)
4242 env->me_mapsize = minsize;
4244 meta.mm_mapsize = env->me_mapsize;
4246 if (newenv && !(flags & MDB_FIXEDMAP)) {
4247 /* mdb_env_map() may grow the datafile. Write the metapages
4248 * first, so the file will be valid if initialization fails.
4249 * Except with FIXEDMAP, since we do not yet know mm_address.
4250 * We could fill in mm_address later, but then a different
4251 * program might end up doing that - one with a memory layout
4252 * and map address which does not suit the main program.
4254 rc = mdb_env_init_meta(env, &meta);
4260 /* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
4263 rc = WriteFile(env->me_fd, &dummy, 1, NULL, NULL);
4271 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4276 if (flags & MDB_FIXEDMAP)
4277 meta.mm_address = env->me_map;
4278 i = mdb_env_init_meta(env, &meta);
4279 if (i != MDB_SUCCESS) {
4284 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4285 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4287 #if !(MDB_MAXKEYSIZE)
4288 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4290 env->me_maxpg = env->me_mapsize / env->me_psize;
4294 MDB_meta *meta = mdb_env_pick_meta(env);
4295 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4297 DPRINTF(("opened database version %u, pagesize %u",
4298 meta->mm_version, env->me_psize));
4299 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4300 DPRINTF(("depth: %u", db->md_depth));
4301 DPRINTF(("entries: %"Z"u", db->md_entries));
4302 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4303 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4304 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4305 DPRINTF(("root: %"Z"u", db->md_root));
4313 /** Release a reader thread's slot in the reader lock table.
4314 * This function is called automatically when a thread exits.
4315 * @param[in] ptr This points to the slot in the reader lock table.
4318 mdb_env_reader_dest(void *ptr)
4320 MDB_reader *reader = ptr;
4326 /** Junk for arranging thread-specific callbacks on Windows. This is
4327 * necessarily platform and compiler-specific. Windows supports up
4328 * to 1088 keys. Let's assume nobody opens more than 64 environments
4329 * in a single process, for now. They can override this if needed.
4331 #ifndef MAX_TLS_KEYS
4332 #define MAX_TLS_KEYS 64
4334 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4335 static int mdb_tls_nkeys;
4337 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4341 case DLL_PROCESS_ATTACH: break;
4342 case DLL_THREAD_ATTACH: break;
4343 case DLL_THREAD_DETACH:
4344 for (i=0; i<mdb_tls_nkeys; i++) {
4345 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4347 mdb_env_reader_dest(r);
4351 case DLL_PROCESS_DETACH: break;
4356 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4358 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4362 /* Force some symbol references.
4363 * _tls_used forces the linker to create the TLS directory if not already done
4364 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4366 #pragma comment(linker, "/INCLUDE:_tls_used")
4367 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4368 #pragma const_seg(".CRT$XLB")
4369 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4370 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4373 #pragma comment(linker, "/INCLUDE:__tls_used")
4374 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4375 #pragma data_seg(".CRT$XLB")
4376 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4378 #endif /* WIN 32/64 */
4379 #endif /* !__GNUC__ */
4382 /** Downgrade the exclusive lock on the region back to shared */
4384 mdb_env_share_locks(MDB_env *env, int *excl)
4387 MDB_meta *meta = mdb_env_pick_meta(env);
4389 env->me_txns->mti_txnid = meta->mm_txnid;
4394 /* First acquire a shared lock. The Unlock will
4395 * then release the existing exclusive lock.
4397 memset(&ov, 0, sizeof(ov));
4398 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4401 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4407 struct flock lock_info;
4408 /* The shared lock replaces the existing lock */
4409 memset((void *)&lock_info, 0, sizeof(lock_info));
4410 lock_info.l_type = F_RDLCK;
4411 lock_info.l_whence = SEEK_SET;
4412 lock_info.l_start = 0;
4413 lock_info.l_len = 1;
4414 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4415 (rc = ErrCode()) == EINTR) ;
4416 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4423 /** Try to get exclusive lock, otherwise shared.
4424 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4427 mdb_env_excl_lock(MDB_env *env, int *excl)
4431 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4435 memset(&ov, 0, sizeof(ov));
4436 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4443 struct flock lock_info;
4444 memset((void *)&lock_info, 0, sizeof(lock_info));
4445 lock_info.l_type = F_WRLCK;
4446 lock_info.l_whence = SEEK_SET;
4447 lock_info.l_start = 0;
4448 lock_info.l_len = 1;
4449 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4450 (rc = ErrCode()) == EINTR) ;
4454 # ifndef MDB_USE_POSIX_MUTEX
4455 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4458 lock_info.l_type = F_RDLCK;
4459 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4460 (rc = ErrCode()) == EINTR) ;
4470 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4472 * @(#) $Revision: 5.1 $
4473 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4474 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4476 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4480 * Please do not copyright this code. This code is in the public domain.
4482 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4483 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4484 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4485 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4486 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4487 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4488 * PERFORMANCE OF THIS SOFTWARE.
4491 * chongo <Landon Curt Noll> /\oo/\
4492 * http://www.isthe.com/chongo/
4494 * Share and Enjoy! :-)
4497 typedef unsigned long long mdb_hash_t;
4498 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4500 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4501 * @param[in] val value to hash
4502 * @param[in] hval initial value for hash
4503 * @return 64 bit hash
4505 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4506 * hval arg on the first call.
4509 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4511 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4512 unsigned char *end = s + val->mv_size;
4514 * FNV-1a hash each octet of the string
4517 /* xor the bottom with the current octet */
4518 hval ^= (mdb_hash_t)*s++;
4520 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4521 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4522 (hval << 7) + (hval << 8) + (hval << 40);
4524 /* return our new hash value */
4528 /** Hash the string and output the encoded hash.
4529 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4530 * very short name limits. We don't care about the encoding being reversible,
4531 * we just want to preserve as many bits of the input as possible in a
4532 * small printable string.
4533 * @param[in] str string to hash
4534 * @param[out] encbuf an array of 11 chars to hold the hash
4536 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4539 mdb_pack85(unsigned long l, char *out)
4543 for (i=0; i<5; i++) {
4544 *out++ = mdb_a85[l % 85];
4550 mdb_hash_enc(MDB_val *val, char *encbuf)
4552 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4554 mdb_pack85(h, encbuf);
4555 mdb_pack85(h>>32, encbuf+5);
4560 /** Open and/or initialize the lock region for the environment.
4561 * @param[in] env The LMDB environment.
4562 * @param[in] lpath The pathname of the file used for the lock region.
4563 * @param[in] mode The Unix permissions for the file, if we create it.
4564 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4565 * @return 0 on success, non-zero on failure.
4568 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4571 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4573 # define MDB_ERRCODE_ROFS EROFS
4574 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4575 # define MDB_CLOEXEC O_CLOEXEC
4578 # define MDB_CLOEXEC 0
4581 #ifdef MDB_USE_SYSV_SEM
4590 utf8_to_utf16(lpath, -1, &wlpath, NULL);
4591 env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
4592 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4593 FILE_ATTRIBUTE_NORMAL, NULL);
4596 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4598 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4600 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4605 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4606 /* Lose record locks when exec*() */
4607 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4608 fcntl(env->me_lfd, F_SETFD, fdflags);
4611 if (!(env->me_flags & MDB_NOTLS)) {
4612 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4615 env->me_flags |= MDB_ENV_TXKEY;
4617 /* Windows TLS callbacks need help finding their TLS info. */
4618 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4622 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4626 /* Try to get exclusive lock. If we succeed, then
4627 * nobody is using the lock region and we should initialize it.
4629 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4632 size = GetFileSize(env->me_lfd, NULL);
4634 size = lseek(env->me_lfd, 0, SEEK_END);
4635 if (size == -1) goto fail_errno;
4637 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4638 if (size < rsize && *excl > 0) {
4640 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4641 || !SetEndOfFile(env->me_lfd))
4644 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4648 size = rsize - sizeof(MDB_txninfo);
4649 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4654 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4656 if (!mh) goto fail_errno;
4657 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4659 if (!env->me_txns) goto fail_errno;
4661 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4663 if (m == MAP_FAILED) goto fail_errno;
4669 BY_HANDLE_FILE_INFORMATION stbuf;
4678 if (!mdb_sec_inited) {
4679 InitializeSecurityDescriptor(&mdb_null_sd,
4680 SECURITY_DESCRIPTOR_REVISION);
4681 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4682 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4683 mdb_all_sa.bInheritHandle = FALSE;
4684 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4687 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4688 idbuf.volume = stbuf.dwVolumeSerialNumber;
4689 idbuf.nhigh = stbuf.nFileIndexHigh;
4690 idbuf.nlow = stbuf.nFileIndexLow;
4691 val.mv_data = &idbuf;
4692 val.mv_size = sizeof(idbuf);
4693 mdb_hash_enc(&val, encbuf);
4694 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4695 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4696 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4697 if (!env->me_rmutex) goto fail_errno;
4698 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4699 if (!env->me_wmutex) goto fail_errno;
4700 #elif defined(MDB_USE_POSIX_SEM)
4709 #if defined(__NetBSD__)
4710 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4712 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4713 idbuf.dev = stbuf.st_dev;
4714 idbuf.ino = stbuf.st_ino;
4715 val.mv_data = &idbuf;
4716 val.mv_size = sizeof(idbuf);
4717 mdb_hash_enc(&val, encbuf);
4718 #ifdef MDB_SHORT_SEMNAMES
4719 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4721 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4722 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4723 /* Clean up after a previous run, if needed: Try to
4724 * remove both semaphores before doing anything else.
4726 sem_unlink(env->me_txns->mti_rmname);
4727 sem_unlink(env->me_txns->mti_wmname);
4728 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4729 O_CREAT|O_EXCL, mode, 1);
4730 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4731 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4732 O_CREAT|O_EXCL, mode, 1);
4733 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4734 #elif defined(MDB_USE_SYSV_SEM)
4735 unsigned short vals[2] = {1, 1};
4736 key_t key = ftok(lpath, 'M');
4739 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4743 if (semctl(semid, 0, SETALL, semu) < 0)
4745 env->me_txns->mti_semid = semid;
4746 #else /* MDB_USE_POSIX_MUTEX: */
4747 pthread_mutexattr_t mattr;
4749 if ((rc = pthread_mutexattr_init(&mattr))
4750 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4751 #ifdef MDB_ROBUST_SUPPORTED
4752 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4754 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4755 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4757 pthread_mutexattr_destroy(&mattr);
4758 #endif /* _WIN32 || ... */
4760 env->me_txns->mti_magic = MDB_MAGIC;
4761 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4762 env->me_txns->mti_txnid = 0;
4763 env->me_txns->mti_numreaders = 0;
4766 #ifdef MDB_USE_SYSV_SEM
4767 struct semid_ds buf;
4769 if (env->me_txns->mti_magic != MDB_MAGIC) {
4770 DPUTS("lock region has invalid magic");
4774 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4775 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4776 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4777 rc = MDB_VERSION_MISMATCH;
4781 if (rc && rc != EACCES && rc != EAGAIN) {
4785 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4786 if (!env->me_rmutex) goto fail_errno;
4787 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4788 if (!env->me_wmutex) goto fail_errno;
4789 #elif defined(MDB_USE_POSIX_SEM)
4790 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4791 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4792 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4793 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4794 #elif defined(MDB_USE_SYSV_SEM)
4795 semid = env->me_txns->mti_semid;
4797 /* check for read access */
4798 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4800 /* check for write access */
4801 if (semctl(semid, 0, IPC_SET, semu) < 0)
4805 #ifdef MDB_USE_SYSV_SEM
4806 env->me_rmutex->semid = semid;
4807 env->me_wmutex->semid = semid;
4808 env->me_rmutex->semnum = 0;
4809 env->me_wmutex->semnum = 1;
4810 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4811 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4822 /** The name of the lock file in the DB environment */
4823 #define LOCKNAME "/lock.mdb"
4824 /** The name of the data file in the DB environment */
4825 #define DATANAME "/data.mdb"
4826 /** The suffix of the lock file when no subdir is used */
4827 #define LOCKSUFF "-lock"
4828 /** Only a subset of the @ref mdb_env flags can be changed
4829 * at runtime. Changing other flags requires closing the
4830 * environment and re-opening it with the new flags.
4832 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4833 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4834 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4836 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4837 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4841 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4843 int oflags, rc, len, excl = -1;
4844 char *lpath, *dpath;
4849 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4853 if (flags & MDB_NOSUBDIR) {
4854 rc = len + sizeof(LOCKSUFF) + len + 1;
4856 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4861 if (flags & MDB_NOSUBDIR) {
4862 dpath = lpath + len + sizeof(LOCKSUFF);
4863 sprintf(lpath, "%s" LOCKSUFF, path);
4864 strcpy(dpath, path);
4866 dpath = lpath + len + sizeof(LOCKNAME);
4867 sprintf(lpath, "%s" LOCKNAME, path);
4868 sprintf(dpath, "%s" DATANAME, path);
4872 flags |= env->me_flags;
4873 if (flags & MDB_RDONLY) {
4874 /* silently ignore WRITEMAP when we're only getting read access */
4875 flags &= ~MDB_WRITEMAP;
4877 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4878 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4881 env->me_flags = flags |= MDB_ENV_ACTIVE;
4885 env->me_path = strdup(path);
4886 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4887 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4888 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4889 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4893 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4895 /* For RDONLY, get lockfile after we know datafile exists */
4896 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4897 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4903 if (F_ISSET(flags, MDB_RDONLY)) {
4904 oflags = GENERIC_READ;
4905 len = OPEN_EXISTING;
4907 oflags = GENERIC_READ|GENERIC_WRITE;
4910 mode = FILE_ATTRIBUTE_NORMAL;
4911 utf8_to_utf16(dpath, -1, &wpath, NULL);
4912 env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4913 NULL, len, mode, NULL);
4916 if (F_ISSET(flags, MDB_RDONLY))
4919 oflags = O_RDWR | O_CREAT;
4921 env->me_fd = open(dpath, oflags, mode);
4923 if (env->me_fd == INVALID_HANDLE_VALUE) {
4928 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4929 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4934 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4935 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4936 env->me_mfd = env->me_fd;
4938 /* Synchronous fd for meta writes. Needed even with
4939 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4942 len = OPEN_EXISTING;
4943 utf8_to_utf16(dpath, -1, &wpath, NULL);
4944 env->me_mfd = CreateFileW(wpath, oflags,
4945 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4946 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4950 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4952 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4957 DPRINTF(("opened dbenv %p", (void *) env));
4959 rc = mdb_env_share_locks(env, &excl);
4963 if (!(flags & MDB_RDONLY)) {
4965 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4966 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4967 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4968 (txn = calloc(1, size)))
4970 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4971 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4972 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4973 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4975 txn->mt_dbxs = env->me_dbxs;
4976 txn->mt_flags = MDB_TXN_FINISHED;
4986 mdb_env_close0(env, excl);
4992 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4994 mdb_env_close0(MDB_env *env, int excl)
4998 if (!(env->me_flags & MDB_ENV_ACTIVE))
5001 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5003 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5004 free(env->me_dbxs[i].md_name.mv_data);
5009 free(env->me_dbiseqs);
5010 free(env->me_dbflags);
5012 free(env->me_dirty_list);
5014 mdb_midl_free(env->me_free_pgs);
5016 if (env->me_flags & MDB_ENV_TXKEY) {
5017 pthread_key_delete(env->me_txkey);
5019 /* Delete our key from the global list */
5020 for (i=0; i<mdb_tls_nkeys; i++)
5021 if (mdb_tls_keys[i] == env->me_txkey) {
5022 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5030 munmap(env->me_map, env->me_mapsize);
5032 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
5033 (void) close(env->me_mfd);
5034 if (env->me_fd != INVALID_HANDLE_VALUE)
5035 (void) close(env->me_fd);
5037 MDB_PID_T pid = env->me_pid;
5038 /* Clearing readers is done in this function because
5039 * me_txkey with its destructor must be disabled first.
5041 * We skip the the reader mutex, so we touch only
5042 * data owned by this process (me_close_readers and
5043 * our readers), and clear each reader atomically.
5045 for (i = env->me_close_readers; --i >= 0; )
5046 if (env->me_txns->mti_readers[i].mr_pid == pid)
5047 env->me_txns->mti_readers[i].mr_pid = 0;
5049 if (env->me_rmutex) {
5050 CloseHandle(env->me_rmutex);
5051 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5053 /* Windows automatically destroys the mutexes when
5054 * the last handle closes.
5056 #elif defined(MDB_USE_POSIX_SEM)
5057 if (env->me_rmutex != SEM_FAILED) {
5058 sem_close(env->me_rmutex);
5059 if (env->me_wmutex != SEM_FAILED)
5060 sem_close(env->me_wmutex);
5061 /* If we have the filelock: If we are the
5062 * only remaining user, clean up semaphores.
5065 mdb_env_excl_lock(env, &excl);
5067 sem_unlink(env->me_txns->mti_rmname);
5068 sem_unlink(env->me_txns->mti_wmname);
5071 #elif defined(MDB_USE_SYSV_SEM)
5072 if (env->me_rmutex->semid != -1) {
5073 /* If we have the filelock: If we are the
5074 * only remaining user, clean up semaphores.
5077 mdb_env_excl_lock(env, &excl);
5079 semctl(env->me_rmutex->semid, 0, IPC_RMID);
5082 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5084 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5087 /* Unlock the lockfile. Windows would have unlocked it
5088 * after closing anyway, but not necessarily at once.
5090 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5093 (void) close(env->me_lfd);
5096 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5100 mdb_env_close(MDB_env *env)
5107 VGMEMP_DESTROY(env);
5108 while ((dp = env->me_dpages) != NULL) {
5109 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5110 env->me_dpages = dp->mp_next;
5114 mdb_env_close0(env, 0);
5118 /** Compare two items pointing at aligned size_t's */
5120 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5122 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5123 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5126 /** Compare two items pointing at aligned unsigned int's.
5128 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5129 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5132 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5134 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5135 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5138 /** Compare two items pointing at unsigned ints of unknown alignment.
5139 * Nodes and keys are guaranteed to be 2-byte aligned.
5142 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5144 #if BYTE_ORDER == LITTLE_ENDIAN
5145 unsigned short *u, *c;
5148 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5149 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5152 } while(!x && u > (unsigned short *)a->mv_data);
5155 unsigned short *u, *c, *end;
5158 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5159 u = (unsigned short *)a->mv_data;
5160 c = (unsigned short *)b->mv_data;
5163 } while(!x && u < end);
5168 /** Compare two items lexically */
5170 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5177 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5183 diff = memcmp(a->mv_data, b->mv_data, len);
5184 return diff ? diff : len_diff<0 ? -1 : len_diff;
5187 /** Compare two items in reverse byte order */
5189 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5191 const unsigned char *p1, *p2, *p1_lim;
5195 p1_lim = (const unsigned char *)a->mv_data;
5196 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5197 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5199 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5205 while (p1 > p1_lim) {
5206 diff = *--p1 - *--p2;
5210 return len_diff<0 ? -1 : len_diff;
5213 /** Search for key within a page, using binary search.
5214 * Returns the smallest entry larger or equal to the key.
5215 * If exactp is non-null, stores whether the found entry was an exact match
5216 * in *exactp (1 or 0).
5217 * Updates the cursor index with the index of the found entry.
5218 * If no entry larger or equal to the key is found, returns NULL.
5221 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5223 unsigned int i = 0, nkeys;
5226 MDB_page *mp = mc->mc_pg[mc->mc_top];
5227 MDB_node *node = NULL;
5232 nkeys = NUMKEYS(mp);
5234 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5235 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5238 low = IS_LEAF(mp) ? 0 : 1;
5240 cmp = mc->mc_dbx->md_cmp;
5242 /* Branch pages have no data, so if using integer keys,
5243 * alignment is guaranteed. Use faster mdb_cmp_int.
5245 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5246 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5253 nodekey.mv_size = mc->mc_db->md_pad;
5254 node = NODEPTR(mp, 0); /* fake */
5255 while (low <= high) {
5256 i = (low + high) >> 1;
5257 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5258 rc = cmp(key, &nodekey);
5259 DPRINTF(("found leaf index %u [%s], rc = %i",
5260 i, DKEY(&nodekey), rc));
5269 while (low <= high) {
5270 i = (low + high) >> 1;
5272 node = NODEPTR(mp, i);
5273 nodekey.mv_size = NODEKSZ(node);
5274 nodekey.mv_data = NODEKEY(node);
5276 rc = cmp(key, &nodekey);
5279 DPRINTF(("found leaf index %u [%s], rc = %i",
5280 i, DKEY(&nodekey), rc));
5282 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5283 i, DKEY(&nodekey), NODEPGNO(node), rc));
5294 if (rc > 0) { /* Found entry is less than the key. */
5295 i++; /* Skip to get the smallest entry larger than key. */
5297 node = NODEPTR(mp, i);
5300 *exactp = (rc == 0 && nkeys > 0);
5301 /* store the key index */
5302 mc->mc_ki[mc->mc_top] = i;
5304 /* There is no entry larger or equal to the key. */
5307 /* nodeptr is fake for LEAF2 */
5313 mdb_cursor_adjust(MDB_cursor *mc, func)
5317 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5318 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5325 /** Pop a page off the top of the cursor's stack. */
5327 mdb_cursor_pop(MDB_cursor *mc)
5330 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5331 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5337 mc->mc_flags &= ~C_INITIALIZED;
5342 /** Push a page onto the top of the cursor's stack. */
5344 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5346 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5347 DDBI(mc), (void *) mc));
5349 if (mc->mc_snum >= CURSOR_STACK) {
5350 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5351 return MDB_CURSOR_FULL;
5354 mc->mc_top = mc->mc_snum++;
5355 mc->mc_pg[mc->mc_top] = mp;
5356 mc->mc_ki[mc->mc_top] = 0;
5361 /** Find the address of the page corresponding to a given page number.
5362 * @param[in] txn the transaction for this access.
5363 * @param[in] pgno the page number for the page to retrieve.
5364 * @param[out] ret address of a pointer where the page's address will be stored.
5365 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5366 * @return 0 on success, non-zero on failure.
5369 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5371 MDB_env *env = txn->mt_env;
5375 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5379 MDB_ID2L dl = tx2->mt_u.dirty_list;
5381 /* Spilled pages were dirtied in this txn and flushed
5382 * because the dirty list got full. Bring this page
5383 * back in from the map (but don't unspill it here,
5384 * leave that unless page_touch happens again).
5386 if (tx2->mt_spill_pgs) {
5387 MDB_ID pn = pgno << 1;
5388 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5389 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5390 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5395 unsigned x = mdb_mid2l_search(dl, pgno);
5396 if (x <= dl[0].mid && dl[x].mid == pgno) {
5402 } while ((tx2 = tx2->mt_parent) != NULL);
5405 if (pgno < txn->mt_next_pgno) {
5407 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5409 DPRINTF(("page %"Z"u not found", pgno));
5410 txn->mt_flags |= MDB_TXN_ERROR;
5411 return MDB_PAGE_NOTFOUND;
5421 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5422 * The cursor is at the root page, set up the rest of it.
5425 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5427 MDB_page *mp = mc->mc_pg[mc->mc_top];
5431 while (IS_BRANCH(mp)) {
5435 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5436 mdb_cassert(mc, NUMKEYS(mp) > 1);
5437 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5439 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5441 if (flags & MDB_PS_LAST)
5442 i = NUMKEYS(mp) - 1;
5445 node = mdb_node_search(mc, key, &exact);
5447 i = NUMKEYS(mp) - 1;
5449 i = mc->mc_ki[mc->mc_top];
5451 mdb_cassert(mc, i > 0);
5455 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5458 mdb_cassert(mc, i < NUMKEYS(mp));
5459 node = NODEPTR(mp, i);
5461 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5464 mc->mc_ki[mc->mc_top] = i;
5465 if ((rc = mdb_cursor_push(mc, mp)))
5468 if (flags & MDB_PS_MODIFY) {
5469 if ((rc = mdb_page_touch(mc)) != 0)
5471 mp = mc->mc_pg[mc->mc_top];
5476 DPRINTF(("internal error, index points to a %02X page!?",
5478 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5479 return MDB_CORRUPTED;
5482 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5483 key ? DKEY(key) : "null"));
5484 mc->mc_flags |= C_INITIALIZED;
5485 mc->mc_flags &= ~C_EOF;
5490 /** Search for the lowest key under the current branch page.
5491 * This just bypasses a NUMKEYS check in the current page
5492 * before calling mdb_page_search_root(), because the callers
5493 * are all in situations where the current page is known to
5497 mdb_page_search_lowest(MDB_cursor *mc)
5499 MDB_page *mp = mc->mc_pg[mc->mc_top];
5500 MDB_node *node = NODEPTR(mp, 0);
5503 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5506 mc->mc_ki[mc->mc_top] = 0;
5507 if ((rc = mdb_cursor_push(mc, mp)))
5509 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5512 /** Search for the page a given key should be in.
5513 * Push it and its parent pages on the cursor stack.
5514 * @param[in,out] mc the cursor for this operation.
5515 * @param[in] key the key to search for, or NULL for first/last page.
5516 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5517 * are touched (updated with new page numbers).
5518 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5519 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5520 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5521 * @return 0 on success, non-zero on failure.
5524 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5529 /* Make sure the txn is still viable, then find the root from
5530 * the txn's db table and set it as the root of the cursor's stack.
5532 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5533 DPUTS("transaction may not be used now");
5536 /* Make sure we're using an up-to-date root */
5537 if (*mc->mc_dbflag & DB_STALE) {
5539 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5541 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5542 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5549 MDB_node *leaf = mdb_node_search(&mc2,
5550 &mc->mc_dbx->md_name, &exact);
5552 return MDB_NOTFOUND;
5553 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5554 return MDB_INCOMPATIBLE; /* not a named DB */
5555 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5558 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5560 /* The txn may not know this DBI, or another process may
5561 * have dropped and recreated the DB with other flags.
5563 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5564 return MDB_INCOMPATIBLE;
5565 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5567 *mc->mc_dbflag &= ~DB_STALE;
5569 root = mc->mc_db->md_root;
5571 if (root == P_INVALID) { /* Tree is empty. */
5572 DPUTS("tree is empty");
5573 return MDB_NOTFOUND;
5577 mdb_cassert(mc, root > 1);
5578 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5579 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5585 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5586 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5588 if (flags & MDB_PS_MODIFY) {
5589 if ((rc = mdb_page_touch(mc)))
5593 if (flags & MDB_PS_ROOTONLY)
5596 return mdb_page_search_root(mc, key, flags);
5600 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5602 MDB_txn *txn = mc->mc_txn;
5603 pgno_t pg = mp->mp_pgno;
5604 unsigned x = 0, ovpages = mp->mp_pages;
5605 MDB_env *env = txn->mt_env;
5606 MDB_IDL sl = txn->mt_spill_pgs;
5607 MDB_ID pn = pg << 1;
5610 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5611 /* If the page is dirty or on the spill list we just acquired it,
5612 * so we should give it back to our current free list, if any.
5613 * Otherwise put it onto the list of pages we freed in this txn.
5615 * Won't create me_pghead: me_pglast must be inited along with it.
5616 * Unsupported in nested txns: They would need to hide the page
5617 * range in ancestor txns' dirty and spilled lists.
5619 if (env->me_pghead &&
5621 ((mp->mp_flags & P_DIRTY) ||
5622 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5626 MDB_ID2 *dl, ix, iy;
5627 rc = mdb_midl_need(&env->me_pghead, ovpages);
5630 if (!(mp->mp_flags & P_DIRTY)) {
5631 /* This page is no longer spilled */
5638 /* Remove from dirty list */
5639 dl = txn->mt_u.dirty_list;
5641 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5647 mdb_cassert(mc, x > 1);
5649 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5650 txn->mt_flags |= MDB_TXN_ERROR;
5651 return MDB_CORRUPTED;
5654 txn->mt_dirty_room++;
5655 if (!(env->me_flags & MDB_WRITEMAP))
5656 mdb_dpage_free(env, mp);
5658 /* Insert in me_pghead */
5659 mop = env->me_pghead;
5660 j = mop[0] + ovpages;
5661 for (i = mop[0]; i && mop[i] < pg; i--)
5667 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5671 mc->mc_db->md_overflow_pages -= ovpages;
5675 /** Return the data associated with a given node.
5676 * @param[in] txn The transaction for this operation.
5677 * @param[in] leaf The node being read.
5678 * @param[out] data Updated to point to the node's data.
5679 * @return 0 on success, non-zero on failure.
5682 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5684 MDB_page *omp; /* overflow page */
5688 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5689 data->mv_size = NODEDSZ(leaf);
5690 data->mv_data = NODEDATA(leaf);
5694 /* Read overflow data.
5696 data->mv_size = NODEDSZ(leaf);
5697 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5698 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5699 DPRINTF(("read overflow page %"Z"u failed", pgno));
5702 data->mv_data = METADATA(omp);
5708 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5709 MDB_val *key, MDB_val *data)
5716 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5718 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5721 if (txn->mt_flags & MDB_TXN_BLOCKED)
5724 mdb_cursor_init(&mc, txn, dbi, &mx);
5725 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5728 /** Find a sibling for a page.
5729 * Replaces the page at the top of the cursor's stack with the
5730 * specified sibling, if one exists.
5731 * @param[in] mc The cursor for this operation.
5732 * @param[in] move_right Non-zero if the right sibling is requested,
5733 * otherwise the left sibling.
5734 * @return 0 on success, non-zero on failure.
5737 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5743 if (mc->mc_snum < 2) {
5744 return MDB_NOTFOUND; /* root has no siblings */
5748 DPRINTF(("parent page is page %"Z"u, index %u",
5749 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5751 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5752 : (mc->mc_ki[mc->mc_top] == 0)) {
5753 DPRINTF(("no more keys left, moving to %s sibling",
5754 move_right ? "right" : "left"));
5755 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5756 /* undo cursor_pop before returning */
5763 mc->mc_ki[mc->mc_top]++;
5765 mc->mc_ki[mc->mc_top]--;
5766 DPRINTF(("just moving to %s index key %u",
5767 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5769 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5771 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5772 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5773 /* mc will be inconsistent if caller does mc_snum++ as above */
5774 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5778 mdb_cursor_push(mc, mp);
5780 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5785 /** Move the cursor to the next data item. */
5787 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5793 if (mc->mc_flags & C_EOF) {
5794 return MDB_NOTFOUND;
5797 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5799 mp = mc->mc_pg[mc->mc_top];
5801 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5802 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5803 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5804 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5805 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5806 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5807 if (rc == MDB_SUCCESS)
5808 MDB_GET_KEY(leaf, key);
5813 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5814 if (op == MDB_NEXT_DUP)
5815 return MDB_NOTFOUND;
5819 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5820 mdb_dbg_pgno(mp), (void *) mc));
5821 if (mc->mc_flags & C_DEL)
5824 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5825 DPUTS("=====> move to next sibling page");
5826 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5827 mc->mc_flags |= C_EOF;
5830 mp = mc->mc_pg[mc->mc_top];
5831 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5833 mc->mc_ki[mc->mc_top]++;
5836 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5837 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5840 key->mv_size = mc->mc_db->md_pad;
5841 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5845 mdb_cassert(mc, IS_LEAF(mp));
5846 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5848 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5849 mdb_xcursor_init1(mc, leaf);
5852 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5855 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5856 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5857 if (rc != MDB_SUCCESS)
5862 MDB_GET_KEY(leaf, key);
5866 /** Move the cursor to the previous data item. */
5868 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5874 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5876 mp = mc->mc_pg[mc->mc_top];
5878 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5879 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5880 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5881 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5882 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5883 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5884 if (rc == MDB_SUCCESS) {
5885 MDB_GET_KEY(leaf, key);
5886 mc->mc_flags &= ~C_EOF;
5892 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5893 if (op == MDB_PREV_DUP)
5894 return MDB_NOTFOUND;
5898 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5899 mdb_dbg_pgno(mp), (void *) mc));
5901 if (mc->mc_ki[mc->mc_top] == 0) {
5902 DPUTS("=====> move to prev sibling page");
5903 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5906 mp = mc->mc_pg[mc->mc_top];
5907 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5908 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5910 mc->mc_ki[mc->mc_top]--;
5912 mc->mc_flags &= ~C_EOF;
5914 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5915 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5918 key->mv_size = mc->mc_db->md_pad;
5919 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5923 mdb_cassert(mc, IS_LEAF(mp));
5924 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5926 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5927 mdb_xcursor_init1(mc, leaf);
5930 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5933 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5934 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5935 if (rc != MDB_SUCCESS)
5940 MDB_GET_KEY(leaf, key);
5944 /** Set the cursor on a specific data item. */
5946 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5947 MDB_cursor_op op, int *exactp)
5951 MDB_node *leaf = NULL;
5954 if (key->mv_size == 0)
5955 return MDB_BAD_VALSIZE;
5958 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5960 /* See if we're already on the right page */
5961 if (mc->mc_flags & C_INITIALIZED) {
5964 mp = mc->mc_pg[mc->mc_top];
5966 mc->mc_ki[mc->mc_top] = 0;
5967 return MDB_NOTFOUND;
5969 if (mp->mp_flags & P_LEAF2) {
5970 nodekey.mv_size = mc->mc_db->md_pad;
5971 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5973 leaf = NODEPTR(mp, 0);
5974 MDB_GET_KEY2(leaf, nodekey);
5976 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5978 /* Probably happens rarely, but first node on the page
5979 * was the one we wanted.
5981 mc->mc_ki[mc->mc_top] = 0;
5988 unsigned int nkeys = NUMKEYS(mp);
5990 if (mp->mp_flags & P_LEAF2) {
5991 nodekey.mv_data = LEAF2KEY(mp,
5992 nkeys-1, nodekey.mv_size);
5994 leaf = NODEPTR(mp, nkeys-1);
5995 MDB_GET_KEY2(leaf, nodekey);
5997 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5999 /* last node was the one we wanted */
6000 mc->mc_ki[mc->mc_top] = nkeys-1;
6006 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6007 /* This is definitely the right page, skip search_page */
6008 if (mp->mp_flags & P_LEAF2) {
6009 nodekey.mv_data = LEAF2KEY(mp,
6010 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6012 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6013 MDB_GET_KEY2(leaf, nodekey);
6015 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6017 /* current node was the one we wanted */
6027 /* If any parents have right-sibs, search.
6028 * Otherwise, there's nothing further.
6030 for (i=0; i<mc->mc_top; i++)
6032 NUMKEYS(mc->mc_pg[i])-1)
6034 if (i == mc->mc_top) {
6035 /* There are no other pages */
6036 mc->mc_ki[mc->mc_top] = nkeys;
6037 return MDB_NOTFOUND;
6041 /* There are no other pages */
6042 mc->mc_ki[mc->mc_top] = 0;
6043 if (op == MDB_SET_RANGE && !exactp) {
6047 return MDB_NOTFOUND;
6053 rc = mdb_page_search(mc, key, 0);
6054 if (rc != MDB_SUCCESS)
6057 mp = mc->mc_pg[mc->mc_top];
6058 mdb_cassert(mc, IS_LEAF(mp));
6061 leaf = mdb_node_search(mc, key, exactp);
6062 if (exactp != NULL && !*exactp) {
6063 /* MDB_SET specified and not an exact match. */
6064 return MDB_NOTFOUND;
6068 DPUTS("===> inexact leaf not found, goto sibling");
6069 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6070 mc->mc_flags |= C_EOF;
6071 return rc; /* no entries matched */
6073 mp = mc->mc_pg[mc->mc_top];
6074 mdb_cassert(mc, IS_LEAF(mp));
6075 leaf = NODEPTR(mp, 0);
6079 mc->mc_flags |= C_INITIALIZED;
6080 mc->mc_flags &= ~C_EOF;
6083 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6084 key->mv_size = mc->mc_db->md_pad;
6085 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6090 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6091 mdb_xcursor_init1(mc, leaf);
6094 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6095 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6096 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6099 if (op == MDB_GET_BOTH) {
6105 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6106 if (rc != MDB_SUCCESS)
6109 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6112 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
6114 dcmp = mc->mc_dbx->md_dcmp;
6115 #if UINT_MAX < SIZE_MAX
6116 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6117 dcmp = mdb_cmp_clong;
6119 rc = dcmp(data, &olddata);
6121 if (op == MDB_GET_BOTH || rc > 0)
6122 return MDB_NOTFOUND;
6129 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6130 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6135 /* The key already matches in all other cases */
6136 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6137 MDB_GET_KEY(leaf, key);
6138 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6143 /** Move the cursor to the first item in the database. */
6145 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6151 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6153 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6154 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6155 if (rc != MDB_SUCCESS)
6158 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6160 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6161 mc->mc_flags |= C_INITIALIZED;
6162 mc->mc_flags &= ~C_EOF;
6164 mc->mc_ki[mc->mc_top] = 0;
6166 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6167 key->mv_size = mc->mc_db->md_pad;
6168 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6173 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6174 mdb_xcursor_init1(mc, leaf);
6175 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6179 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6183 MDB_GET_KEY(leaf, key);
6187 /** Move the cursor to the last item in the database. */
6189 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6195 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6197 if (!(mc->mc_flags & C_EOF)) {
6199 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6200 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6201 if (rc != MDB_SUCCESS)
6204 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6207 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6208 mc->mc_flags |= C_INITIALIZED|C_EOF;
6209 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6211 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6212 key->mv_size = mc->mc_db->md_pad;
6213 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6218 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6219 mdb_xcursor_init1(mc, leaf);
6220 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6224 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6229 MDB_GET_KEY(leaf, key);
6234 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6239 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6244 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6248 case MDB_GET_CURRENT:
6249 if (!(mc->mc_flags & C_INITIALIZED)) {
6252 MDB_page *mp = mc->mc_pg[mc->mc_top];
6253 int nkeys = NUMKEYS(mp);
6254 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6255 mc->mc_ki[mc->mc_top] = nkeys;
6261 key->mv_size = mc->mc_db->md_pad;
6262 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6264 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6265 MDB_GET_KEY(leaf, key);
6267 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6268 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6270 rc = mdb_node_read(mc->mc_txn, leaf, data);
6277 case MDB_GET_BOTH_RANGE:
6282 if (mc->mc_xcursor == NULL) {
6283 rc = MDB_INCOMPATIBLE;
6293 rc = mdb_cursor_set(mc, key, data, op,
6294 op == MDB_SET_RANGE ? NULL : &exact);
6297 case MDB_GET_MULTIPLE:
6298 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6302 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6303 rc = MDB_INCOMPATIBLE;
6307 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6308 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6311 case MDB_NEXT_MULTIPLE:
6316 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6317 rc = MDB_INCOMPATIBLE;
6320 if (!(mc->mc_flags & C_INITIALIZED))
6321 rc = mdb_cursor_first(mc, key, data);
6323 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6324 if (rc == MDB_SUCCESS) {
6325 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6328 mx = &mc->mc_xcursor->mx_cursor;
6329 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6331 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6332 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6340 case MDB_NEXT_NODUP:
6341 if (!(mc->mc_flags & C_INITIALIZED))
6342 rc = mdb_cursor_first(mc, key, data);
6344 rc = mdb_cursor_next(mc, key, data, op);
6348 case MDB_PREV_NODUP:
6349 if (!(mc->mc_flags & C_INITIALIZED)) {
6350 rc = mdb_cursor_last(mc, key, data);
6353 mc->mc_flags |= C_INITIALIZED;
6354 mc->mc_ki[mc->mc_top]++;
6356 rc = mdb_cursor_prev(mc, key, data, op);
6359 rc = mdb_cursor_first(mc, key, data);
6362 mfunc = mdb_cursor_first;
6364 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6368 if (mc->mc_xcursor == NULL) {
6369 rc = MDB_INCOMPATIBLE;
6373 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6374 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6375 MDB_GET_KEY(leaf, key);
6376 rc = mdb_node_read(mc->mc_txn, leaf, data);
6380 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6384 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6387 rc = mdb_cursor_last(mc, key, data);
6390 mfunc = mdb_cursor_last;
6393 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6398 if (mc->mc_flags & C_DEL)
6399 mc->mc_flags ^= C_DEL;
6404 /** Touch all the pages in the cursor stack. Set mc_top.
6405 * Makes sure all the pages are writable, before attempting a write operation.
6406 * @param[in] mc The cursor to operate on.
6409 mdb_cursor_touch(MDB_cursor *mc)
6411 int rc = MDB_SUCCESS;
6413 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6416 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6418 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6419 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6422 *mc->mc_dbflag |= DB_DIRTY;
6427 rc = mdb_page_touch(mc);
6428 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6429 mc->mc_top = mc->mc_snum-1;
6434 /** Do not spill pages to disk if txn is getting full, may fail instead */
6435 #define MDB_NOSPILL 0x8000
6438 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6442 MDB_node *leaf = NULL;
6443 MDB_page *fp, *mp, *sub_root = NULL;
6445 MDB_val xdata, *rdata, dkey, olddata;
6447 int do_sub = 0, insert_key, insert_data;
6448 unsigned int mcount = 0, dcount = 0, nospill;
6451 unsigned int nflags;
6454 if (mc == NULL || key == NULL)
6457 env = mc->mc_txn->mt_env;
6459 /* Check this first so counter will always be zero on any
6462 if (flags & MDB_MULTIPLE) {
6463 dcount = data[1].mv_size;
6464 data[1].mv_size = 0;
6465 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6466 return MDB_INCOMPATIBLE;
6469 nospill = flags & MDB_NOSPILL;
6470 flags &= ~MDB_NOSPILL;
6472 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6473 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6475 if (key->mv_size-1 >= ENV_MAXKEY(env))
6476 return MDB_BAD_VALSIZE;
6478 #if SIZE_MAX > MAXDATASIZE
6479 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6480 return MDB_BAD_VALSIZE;
6482 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6483 return MDB_BAD_VALSIZE;
6486 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6487 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6491 if (flags == MDB_CURRENT) {
6492 if (!(mc->mc_flags & C_INITIALIZED))
6495 } else if (mc->mc_db->md_root == P_INVALID) {
6496 /* new database, cursor has nothing to point to */
6499 mc->mc_flags &= ~C_INITIALIZED;
6504 if (flags & MDB_APPEND) {
6506 rc = mdb_cursor_last(mc, &k2, &d2);
6508 rc = mc->mc_dbx->md_cmp(key, &k2);
6511 mc->mc_ki[mc->mc_top]++;
6513 /* new key is <= last key */
6518 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6520 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6521 DPRINTF(("duplicate key [%s]", DKEY(key)));
6523 return MDB_KEYEXIST;
6525 if (rc && rc != MDB_NOTFOUND)
6529 if (mc->mc_flags & C_DEL)
6530 mc->mc_flags ^= C_DEL;
6532 /* Cursor is positioned, check for room in the dirty list */
6534 if (flags & MDB_MULTIPLE) {
6536 xdata.mv_size = data->mv_size * dcount;
6540 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6544 if (rc == MDB_NO_ROOT) {
6546 /* new database, write a root leaf page */
6547 DPUTS("allocating new root leaf page");
6548 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6551 mdb_cursor_push(mc, np);
6552 mc->mc_db->md_root = np->mp_pgno;
6553 mc->mc_db->md_depth++;
6554 *mc->mc_dbflag |= DB_DIRTY;
6555 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6557 np->mp_flags |= P_LEAF2;
6558 mc->mc_flags |= C_INITIALIZED;
6560 /* make sure all cursor pages are writable */
6561 rc2 = mdb_cursor_touch(mc);
6566 insert_key = insert_data = rc;
6568 /* The key does not exist */
6569 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6570 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6571 LEAFSIZE(key, data) > env->me_nodemax)
6573 /* Too big for a node, insert in sub-DB. Set up an empty
6574 * "old sub-page" for prep_subDB to expand to a full page.
6576 fp_flags = P_LEAF|P_DIRTY;
6578 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6579 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6580 olddata.mv_size = PAGEHDRSZ;
6584 /* there's only a key anyway, so this is a no-op */
6585 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6587 unsigned int ksize = mc->mc_db->md_pad;
6588 if (key->mv_size != ksize)
6589 return MDB_BAD_VALSIZE;
6590 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6591 memcpy(ptr, key->mv_data, ksize);
6593 /* if overwriting slot 0 of leaf, need to
6594 * update branch key if there is a parent page
6596 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6597 unsigned short dtop = 1;
6599 /* slot 0 is always an empty key, find real slot */
6600 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6604 if (mc->mc_ki[mc->mc_top])
6605 rc2 = mdb_update_key(mc, key);
6616 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6617 olddata.mv_size = NODEDSZ(leaf);
6618 olddata.mv_data = NODEDATA(leaf);
6621 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6622 /* Prepare (sub-)page/sub-DB to accept the new item,
6623 * if needed. fp: old sub-page or a header faking
6624 * it. mp: new (sub-)page. offset: growth in page
6625 * size. xdata: node data with new page or DB.
6627 unsigned i, offset = 0;
6628 mp = fp = xdata.mv_data = env->me_pbuf;
6629 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6631 /* Was a single item before, must convert now */
6632 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6634 /* Just overwrite the current item */
6635 if (flags == MDB_CURRENT)
6637 dcmp = mc->mc_dbx->md_dcmp;
6638 #if UINT_MAX < SIZE_MAX
6639 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6640 dcmp = mdb_cmp_clong;
6642 /* does data match? */
6643 if (!dcmp(data, &olddata)) {
6644 if (flags & MDB_NODUPDATA)
6645 return MDB_KEYEXIST;
6650 /* Back up original data item */
6651 dkey.mv_size = olddata.mv_size;
6652 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6654 /* Make sub-page header for the dup items, with dummy body */
6655 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6656 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6657 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6658 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6659 fp->mp_flags |= P_LEAF2;
6660 fp->mp_pad = data->mv_size;
6661 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6663 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6664 (dkey.mv_size & 1) + (data->mv_size & 1);
6666 fp->mp_upper = xdata.mv_size - PAGEBASE;
6667 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6668 } else if (leaf->mn_flags & F_SUBDATA) {
6669 /* Data is on sub-DB, just store it */
6670 flags |= F_DUPDATA|F_SUBDATA;
6673 /* Data is on sub-page */
6674 fp = olddata.mv_data;
6677 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6678 offset = EVEN(NODESIZE + sizeof(indx_t) +
6682 offset = fp->mp_pad;
6683 if (SIZELEFT(fp) < offset) {
6684 offset *= 4; /* space for 4 more */
6687 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6689 fp->mp_flags |= P_DIRTY;
6690 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6691 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6695 xdata.mv_size = olddata.mv_size + offset;
6698 fp_flags = fp->mp_flags;
6699 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6700 /* Too big for a sub-page, convert to sub-DB */
6701 fp_flags &= ~P_SUBP;
6703 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6704 fp_flags |= P_LEAF2;
6705 dummy.md_pad = fp->mp_pad;
6706 dummy.md_flags = MDB_DUPFIXED;
6707 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6708 dummy.md_flags |= MDB_INTEGERKEY;
6714 dummy.md_branch_pages = 0;
6715 dummy.md_leaf_pages = 1;
6716 dummy.md_overflow_pages = 0;
6717 dummy.md_entries = NUMKEYS(fp);
6718 xdata.mv_size = sizeof(MDB_db);
6719 xdata.mv_data = &dummy;
6720 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6722 offset = env->me_psize - olddata.mv_size;
6723 flags |= F_DUPDATA|F_SUBDATA;
6724 dummy.md_root = mp->mp_pgno;
6728 mp->mp_flags = fp_flags | P_DIRTY;
6729 mp->mp_pad = fp->mp_pad;
6730 mp->mp_lower = fp->mp_lower;
6731 mp->mp_upper = fp->mp_upper + offset;
6732 if (fp_flags & P_LEAF2) {
6733 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6735 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6736 olddata.mv_size - fp->mp_upper - PAGEBASE);
6737 for (i=0; i<NUMKEYS(fp); i++)
6738 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6746 mdb_node_del(mc, 0);
6750 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6751 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6752 return MDB_INCOMPATIBLE;
6753 /* overflow page overwrites need special handling */
6754 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6757 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6759 memcpy(&pg, olddata.mv_data, sizeof(pg));
6760 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6762 ovpages = omp->mp_pages;
6764 /* Is the ov page large enough? */
6765 if (ovpages >= dpages) {
6766 if (!(omp->mp_flags & P_DIRTY) &&
6767 (level || (env->me_flags & MDB_WRITEMAP)))
6769 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6772 level = 0; /* dirty in this txn or clean */
6775 if (omp->mp_flags & P_DIRTY) {
6776 /* yes, overwrite it. Note in this case we don't
6777 * bother to try shrinking the page if the new data
6778 * is smaller than the overflow threshold.
6781 /* It is writable only in a parent txn */
6782 size_t sz = (size_t) env->me_psize * ovpages, off;
6783 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6789 /* Note - this page is already counted in parent's dirty_room */
6790 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6791 mdb_cassert(mc, rc2 == 0);
6792 if (!(flags & MDB_RESERVE)) {
6793 /* Copy end of page, adjusting alignment so
6794 * compiler may copy words instead of bytes.
6796 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6797 memcpy((size_t *)((char *)np + off),
6798 (size_t *)((char *)omp + off), sz - off);
6801 memcpy(np, omp, sz); /* Copy beginning of page */
6804 SETDSZ(leaf, data->mv_size);
6805 if (F_ISSET(flags, MDB_RESERVE))
6806 data->mv_data = METADATA(omp);
6808 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6812 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6814 } else if (data->mv_size == olddata.mv_size) {
6815 /* same size, just replace it. Note that we could
6816 * also reuse this node if the new data is smaller,
6817 * but instead we opt to shrink the node in that case.
6819 if (F_ISSET(flags, MDB_RESERVE))
6820 data->mv_data = olddata.mv_data;
6821 else if (!(mc->mc_flags & C_SUB))
6822 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6824 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6829 mdb_node_del(mc, 0);
6835 nflags = flags & NODE_ADD_FLAGS;
6836 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6837 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6838 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6839 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6841 nflags |= MDB_SPLIT_REPLACE;
6842 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6844 /* There is room already in this leaf page. */
6845 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6847 /* Adjust other cursors pointing to mp */
6848 MDB_cursor *m2, *m3;
6849 MDB_dbi dbi = mc->mc_dbi;
6850 unsigned i = mc->mc_top;
6851 MDB_page *mp = mc->mc_pg[i];
6853 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6854 if (mc->mc_flags & C_SUB)
6855 m3 = &m2->mc_xcursor->mx_cursor;
6858 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
6859 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
6862 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6863 MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
6864 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
6865 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6871 if (rc == MDB_SUCCESS) {
6872 /* Now store the actual data in the child DB. Note that we're
6873 * storing the user data in the keys field, so there are strict
6874 * size limits on dupdata. The actual data fields of the child
6875 * DB are all zero size.
6878 int xflags, new_dupdata;
6883 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6884 if (flags & MDB_CURRENT) {
6885 xflags = MDB_CURRENT|MDB_NOSPILL;
6887 mdb_xcursor_init1(mc, leaf);
6888 xflags = (flags & MDB_NODUPDATA) ?
6889 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6892 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6893 new_dupdata = (int)dkey.mv_size;
6894 /* converted, write the original data first */
6896 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6899 /* we've done our job */
6902 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6903 /* Adjust other cursors pointing to mp */
6905 MDB_xcursor *mx = mc->mc_xcursor;
6906 unsigned i = mc->mc_top;
6907 MDB_page *mp = mc->mc_pg[i];
6908 int nkeys = NUMKEYS(mp);
6910 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6911 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6912 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6913 if (m2->mc_pg[i] == mp) {
6914 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6915 mdb_xcursor_init2(m2, mx, new_dupdata);
6916 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
6917 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6918 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
6919 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6924 ecount = mc->mc_xcursor->mx_db.md_entries;
6925 if (flags & MDB_APPENDDUP)
6926 xflags |= MDB_APPEND;
6927 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6928 if (flags & F_SUBDATA) {
6929 void *db = NODEDATA(leaf);
6930 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6932 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6934 /* Increment count unless we just replaced an existing item. */
6936 mc->mc_db->md_entries++;
6938 /* Invalidate txn if we created an empty sub-DB */
6941 /* If we succeeded and the key didn't exist before,
6942 * make sure the cursor is marked valid.
6944 mc->mc_flags |= C_INITIALIZED;
6946 if (flags & MDB_MULTIPLE) {
6949 /* let caller know how many succeeded, if any */
6950 data[1].mv_size = mcount;
6951 if (mcount < dcount) {
6952 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6953 insert_key = insert_data = 0;
6960 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6963 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6968 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6974 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6975 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6977 if (!(mc->mc_flags & C_INITIALIZED))
6980 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6981 return MDB_NOTFOUND;
6983 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6986 rc = mdb_cursor_touch(mc);
6990 mp = mc->mc_pg[mc->mc_top];
6993 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6995 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6996 if (flags & MDB_NODUPDATA) {
6997 /* mdb_cursor_del0() will subtract the final entry */
6998 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6999 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7001 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7002 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7004 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7007 /* If sub-DB still has entries, we're done */
7008 if (mc->mc_xcursor->mx_db.md_entries) {
7009 if (leaf->mn_flags & F_SUBDATA) {
7010 /* update subDB info */
7011 void *db = NODEDATA(leaf);
7012 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7015 /* shrink fake page */
7016 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7017 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7018 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7019 /* fix other sub-DB cursors pointed at fake pages on this page */
7020 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7021 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7022 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7023 if (m2->mc_pg[mc->mc_top] == mp) {
7024 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
7025 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7027 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7028 if (!(n2->mn_flags & F_SUBDATA))
7029 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7034 mc->mc_db->md_entries--;
7037 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7039 /* otherwise fall thru and delete the sub-DB */
7042 if (leaf->mn_flags & F_SUBDATA) {
7043 /* add all the child DB's pages to the free list */
7044 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7049 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7050 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7051 rc = MDB_INCOMPATIBLE;
7055 /* add overflow pages to free list */
7056 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7060 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7061 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7062 (rc = mdb_ovpage_free(mc, omp)))
7067 return mdb_cursor_del0(mc);
7070 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7074 /** Allocate and initialize new pages for a database.
7075 * @param[in] mc a cursor on the database being added to.
7076 * @param[in] flags flags defining what type of page is being allocated.
7077 * @param[in] num the number of pages to allocate. This is usually 1,
7078 * unless allocating overflow pages for a large record.
7079 * @param[out] mp Address of a page, or NULL on failure.
7080 * @return 0 on success, non-zero on failure.
7083 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7088 if ((rc = mdb_page_alloc(mc, num, &np)))
7090 DPRINTF(("allocated new mpage %"Z"u, page size %u",
7091 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7092 np->mp_flags = flags | P_DIRTY;
7093 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7094 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7097 mc->mc_db->md_branch_pages++;
7098 else if (IS_LEAF(np))
7099 mc->mc_db->md_leaf_pages++;
7100 else if (IS_OVERFLOW(np)) {
7101 mc->mc_db->md_overflow_pages += num;
7109 /** Calculate the size of a leaf node.
7110 * The size depends on the environment's page size; if a data item
7111 * is too large it will be put onto an overflow page and the node
7112 * size will only include the key and not the data. Sizes are always
7113 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7114 * of the #MDB_node headers.
7115 * @param[in] env The environment handle.
7116 * @param[in] key The key for the node.
7117 * @param[in] data The data for the node.
7118 * @return The number of bytes needed to store the node.
7121 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7125 sz = LEAFSIZE(key, data);
7126 if (sz > env->me_nodemax) {
7127 /* put on overflow page */
7128 sz -= data->mv_size - sizeof(pgno_t);
7131 return EVEN(sz + sizeof(indx_t));
7134 /** Calculate the size of a branch node.
7135 * The size should depend on the environment's page size but since
7136 * we currently don't support spilling large keys onto overflow
7137 * pages, it's simply the size of the #MDB_node header plus the
7138 * size of the key. Sizes are always rounded up to an even number
7139 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7140 * @param[in] env The environment handle.
7141 * @param[in] key The key for the node.
7142 * @return The number of bytes needed to store the node.
7145 mdb_branch_size(MDB_env *env, MDB_val *key)
7150 if (sz > env->me_nodemax) {
7151 /* put on overflow page */
7152 /* not implemented */
7153 /* sz -= key->size - sizeof(pgno_t); */
7156 return sz + sizeof(indx_t);
7159 /** Add a node to the page pointed to by the cursor.
7160 * @param[in] mc The cursor for this operation.
7161 * @param[in] indx The index on the page where the new node should be added.
7162 * @param[in] key The key for the new node.
7163 * @param[in] data The data for the new node, if any.
7164 * @param[in] pgno The page number, if adding a branch node.
7165 * @param[in] flags Flags for the node.
7166 * @return 0 on success, non-zero on failure. Possible errors are:
7168 * <li>ENOMEM - failed to allocate overflow pages for the node.
7169 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7170 * should never happen since all callers already calculate the
7171 * page's free space before calling this function.
7175 mdb_node_add(MDB_cursor *mc, indx_t indx,
7176 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7179 size_t node_size = NODESIZE;
7183 MDB_page *mp = mc->mc_pg[mc->mc_top];
7184 MDB_page *ofp = NULL; /* overflow page */
7188 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7190 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7191 IS_LEAF(mp) ? "leaf" : "branch",
7192 IS_SUBP(mp) ? "sub-" : "",
7193 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7194 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7197 /* Move higher keys up one slot. */
7198 int ksize = mc->mc_db->md_pad, dif;
7199 char *ptr = LEAF2KEY(mp, indx, ksize);
7200 dif = NUMKEYS(mp) - indx;
7202 memmove(ptr+ksize, ptr, dif*ksize);
7203 /* insert new key */
7204 memcpy(ptr, key->mv_data, ksize);
7206 /* Just using these for counting */
7207 mp->mp_lower += sizeof(indx_t);
7208 mp->mp_upper -= ksize - sizeof(indx_t);
7212 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7214 node_size += key->mv_size;
7216 mdb_cassert(mc, key && data);
7217 if (F_ISSET(flags, F_BIGDATA)) {
7218 /* Data already on overflow page. */
7219 node_size += sizeof(pgno_t);
7220 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7221 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7223 /* Put data on overflow page. */
7224 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7225 data->mv_size, node_size+data->mv_size));
7226 node_size = EVEN(node_size + sizeof(pgno_t));
7227 if ((ssize_t)node_size > room)
7229 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7231 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7235 node_size += data->mv_size;
7238 node_size = EVEN(node_size);
7239 if ((ssize_t)node_size > room)
7243 /* Move higher pointers up one slot. */
7244 for (i = NUMKEYS(mp); i > indx; i--)
7245 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7247 /* Adjust free space offsets. */
7248 ofs = mp->mp_upper - node_size;
7249 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7250 mp->mp_ptrs[indx] = ofs;
7252 mp->mp_lower += sizeof(indx_t);
7254 /* Write the node data. */
7255 node = NODEPTR(mp, indx);
7256 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7257 node->mn_flags = flags;
7259 SETDSZ(node,data->mv_size);
7264 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7267 ndata = NODEDATA(node);
7269 if (F_ISSET(flags, F_BIGDATA))
7270 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7271 else if (F_ISSET(flags, MDB_RESERVE))
7272 data->mv_data = ndata;
7274 memcpy(ndata, data->mv_data, data->mv_size);
7276 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7277 ndata = METADATA(ofp);
7278 if (F_ISSET(flags, MDB_RESERVE))
7279 data->mv_data = ndata;
7281 memcpy(ndata, data->mv_data, data->mv_size);
7288 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7289 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7290 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7291 DPRINTF(("node size = %"Z"u", node_size));
7292 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7293 return MDB_PAGE_FULL;
7296 /** Delete the specified node from a page.
7297 * @param[in] mc Cursor pointing to the node to delete.
7298 * @param[in] ksize The size of a node. Only used if the page is
7299 * part of a #MDB_DUPFIXED database.
7302 mdb_node_del(MDB_cursor *mc, int ksize)
7304 MDB_page *mp = mc->mc_pg[mc->mc_top];
7305 indx_t indx = mc->mc_ki[mc->mc_top];
7307 indx_t i, j, numkeys, ptr;
7311 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7312 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7313 numkeys = NUMKEYS(mp);
7314 mdb_cassert(mc, indx < numkeys);
7317 int x = numkeys - 1 - indx;
7318 base = LEAF2KEY(mp, indx, ksize);
7320 memmove(base, base + ksize, x * ksize);
7321 mp->mp_lower -= sizeof(indx_t);
7322 mp->mp_upper += ksize - sizeof(indx_t);
7326 node = NODEPTR(mp, indx);
7327 sz = NODESIZE + node->mn_ksize;
7329 if (F_ISSET(node->mn_flags, F_BIGDATA))
7330 sz += sizeof(pgno_t);
7332 sz += NODEDSZ(node);
7336 ptr = mp->mp_ptrs[indx];
7337 for (i = j = 0; i < numkeys; i++) {
7339 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7340 if (mp->mp_ptrs[i] < ptr)
7341 mp->mp_ptrs[j] += sz;
7346 base = (char *)mp + mp->mp_upper + PAGEBASE;
7347 memmove(base + sz, base, ptr - mp->mp_upper);
7349 mp->mp_lower -= sizeof(indx_t);
7353 /** Compact the main page after deleting a node on a subpage.
7354 * @param[in] mp The main page to operate on.
7355 * @param[in] indx The index of the subpage on the main page.
7358 mdb_node_shrink(MDB_page *mp, indx_t indx)
7363 indx_t delta, nsize, len, ptr;
7366 node = NODEPTR(mp, indx);
7367 sp = (MDB_page *)NODEDATA(node);
7368 delta = SIZELEFT(sp);
7369 nsize = NODEDSZ(node) - delta;
7371 /* Prepare to shift upward, set len = length(subpage part to shift) */
7375 return; /* do not make the node uneven-sized */
7377 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7378 for (i = NUMKEYS(sp); --i >= 0; )
7379 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7382 sp->mp_upper = sp->mp_lower;
7383 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7384 SETDSZ(node, nsize);
7386 /* Shift <lower nodes...initial part of subpage> upward */
7387 base = (char *)mp + mp->mp_upper + PAGEBASE;
7388 memmove(base + delta, base, (char *)sp + len - base);
7390 ptr = mp->mp_ptrs[indx];
7391 for (i = NUMKEYS(mp); --i >= 0; ) {
7392 if (mp->mp_ptrs[i] <= ptr)
7393 mp->mp_ptrs[i] += delta;
7395 mp->mp_upper += delta;
7398 /** Initial setup of a sorted-dups cursor.
7399 * Sorted duplicates are implemented as a sub-database for the given key.
7400 * The duplicate data items are actually keys of the sub-database.
7401 * Operations on the duplicate data items are performed using a sub-cursor
7402 * initialized when the sub-database is first accessed. This function does
7403 * the preliminary setup of the sub-cursor, filling in the fields that
7404 * depend only on the parent DB.
7405 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7408 mdb_xcursor_init0(MDB_cursor *mc)
7410 MDB_xcursor *mx = mc->mc_xcursor;
7412 mx->mx_cursor.mc_xcursor = NULL;
7413 mx->mx_cursor.mc_txn = mc->mc_txn;
7414 mx->mx_cursor.mc_db = &mx->mx_db;
7415 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7416 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7417 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7418 mx->mx_cursor.mc_snum = 0;
7419 mx->mx_cursor.mc_top = 0;
7420 mx->mx_cursor.mc_flags = C_SUB;
7421 mx->mx_dbx.md_name.mv_size = 0;
7422 mx->mx_dbx.md_name.mv_data = NULL;
7423 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7424 mx->mx_dbx.md_dcmp = NULL;
7425 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7428 /** Final setup of a sorted-dups cursor.
7429 * Sets up the fields that depend on the data from the main cursor.
7430 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7431 * @param[in] node The data containing the #MDB_db record for the
7432 * sorted-dup database.
7435 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7437 MDB_xcursor *mx = mc->mc_xcursor;
7439 if (node->mn_flags & F_SUBDATA) {
7440 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7441 mx->mx_cursor.mc_pg[0] = 0;
7442 mx->mx_cursor.mc_snum = 0;
7443 mx->mx_cursor.mc_top = 0;
7444 mx->mx_cursor.mc_flags = C_SUB;
7446 MDB_page *fp = NODEDATA(node);
7447 mx->mx_db.md_pad = 0;
7448 mx->mx_db.md_flags = 0;
7449 mx->mx_db.md_depth = 1;
7450 mx->mx_db.md_branch_pages = 0;
7451 mx->mx_db.md_leaf_pages = 1;
7452 mx->mx_db.md_overflow_pages = 0;
7453 mx->mx_db.md_entries = NUMKEYS(fp);
7454 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7455 mx->mx_cursor.mc_snum = 1;
7456 mx->mx_cursor.mc_top = 0;
7457 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7458 mx->mx_cursor.mc_pg[0] = fp;
7459 mx->mx_cursor.mc_ki[0] = 0;
7460 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7461 mx->mx_db.md_flags = MDB_DUPFIXED;
7462 mx->mx_db.md_pad = fp->mp_pad;
7463 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7464 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7467 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7468 mx->mx_db.md_root));
7469 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7470 #if UINT_MAX < SIZE_MAX
7471 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7472 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7477 /** Fixup a sorted-dups cursor due to underlying update.
7478 * Sets up some fields that depend on the data from the main cursor.
7479 * Almost the same as init1, but skips initialization steps if the
7480 * xcursor had already been used.
7481 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7482 * @param[in] src_mx The xcursor of an up-to-date cursor.
7483 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7486 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7488 MDB_xcursor *mx = mc->mc_xcursor;
7491 mx->mx_cursor.mc_snum = 1;
7492 mx->mx_cursor.mc_top = 0;
7493 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7494 mx->mx_cursor.mc_ki[0] = 0;
7495 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7496 #if UINT_MAX < SIZE_MAX
7497 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7499 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7502 mx->mx_db = src_mx->mx_db;
7503 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7504 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7505 mx->mx_db.md_root));
7508 /** Initialize a cursor for a given transaction and database. */
7510 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7513 mc->mc_backup = NULL;
7516 mc->mc_db = &txn->mt_dbs[dbi];
7517 mc->mc_dbx = &txn->mt_dbxs[dbi];
7518 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7524 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7525 mdb_tassert(txn, mx != NULL);
7526 mc->mc_xcursor = mx;
7527 mdb_xcursor_init0(mc);
7529 mc->mc_xcursor = NULL;
7531 if (*mc->mc_dbflag & DB_STALE) {
7532 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7537 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7540 size_t size = sizeof(MDB_cursor);
7542 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7545 if (txn->mt_flags & MDB_TXN_BLOCKED)
7548 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7551 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7552 size += sizeof(MDB_xcursor);
7554 if ((mc = malloc(size)) != NULL) {
7555 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7556 if (txn->mt_cursors) {
7557 mc->mc_next = txn->mt_cursors[dbi];
7558 txn->mt_cursors[dbi] = mc;
7559 mc->mc_flags |= C_UNTRACK;
7571 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7573 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7576 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7579 if (txn->mt_flags & MDB_TXN_BLOCKED)
7582 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7586 /* Return the count of duplicate data items for the current key */
7588 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7592 if (mc == NULL || countp == NULL)
7595 if (mc->mc_xcursor == NULL)
7596 return MDB_INCOMPATIBLE;
7598 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7601 if (!(mc->mc_flags & C_INITIALIZED))
7604 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7605 return MDB_NOTFOUND;
7607 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7608 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7611 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7614 *countp = mc->mc_xcursor->mx_db.md_entries;
7620 mdb_cursor_close(MDB_cursor *mc)
7622 if (mc && !mc->mc_backup) {
7623 /* remove from txn, if tracked */
7624 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7625 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7626 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7628 *prev = mc->mc_next;
7635 mdb_cursor_txn(MDB_cursor *mc)
7637 if (!mc) return NULL;
7642 mdb_cursor_dbi(MDB_cursor *mc)
7647 /** Replace the key for a branch node with a new key.
7648 * @param[in] mc Cursor pointing to the node to operate on.
7649 * @param[in] key The new key to use.
7650 * @return 0 on success, non-zero on failure.
7653 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7659 int delta, ksize, oksize;
7660 indx_t ptr, i, numkeys, indx;
7663 indx = mc->mc_ki[mc->mc_top];
7664 mp = mc->mc_pg[mc->mc_top];
7665 node = NODEPTR(mp, indx);
7666 ptr = mp->mp_ptrs[indx];
7670 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7671 k2.mv_data = NODEKEY(node);
7672 k2.mv_size = node->mn_ksize;
7673 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7675 mdb_dkey(&k2, kbuf2),
7681 /* Sizes must be 2-byte aligned. */
7682 ksize = EVEN(key->mv_size);
7683 oksize = EVEN(node->mn_ksize);
7684 delta = ksize - oksize;
7686 /* Shift node contents if EVEN(key length) changed. */
7688 if (delta > 0 && SIZELEFT(mp) < delta) {
7690 /* not enough space left, do a delete and split */
7691 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7692 pgno = NODEPGNO(node);
7693 mdb_node_del(mc, 0);
7694 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7697 numkeys = NUMKEYS(mp);
7698 for (i = 0; i < numkeys; i++) {
7699 if (mp->mp_ptrs[i] <= ptr)
7700 mp->mp_ptrs[i] -= delta;
7703 base = (char *)mp + mp->mp_upper + PAGEBASE;
7704 len = ptr - mp->mp_upper + NODESIZE;
7705 memmove(base - delta, base, len);
7706 mp->mp_upper -= delta;
7708 node = NODEPTR(mp, indx);
7711 /* But even if no shift was needed, update ksize */
7712 if (node->mn_ksize != key->mv_size)
7713 node->mn_ksize = key->mv_size;
7716 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7722 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7724 /** Perform \b act while tracking temporary cursor \b mn */
7725 #define WITH_CURSOR_TRACKING(mn, act) do { \
7726 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
7727 if ((mn).mc_flags & C_SUB) { \
7728 dummy.mc_flags = C_INITIALIZED; \
7729 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
7734 tracked->mc_next = *tp; \
7737 *tp = tracked->mc_next; \
7740 /** Move a node from csrc to cdst.
7743 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
7750 unsigned short flags;
7754 /* Mark src and dst as dirty. */
7755 if ((rc = mdb_page_touch(csrc)) ||
7756 (rc = mdb_page_touch(cdst)))
7759 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7760 key.mv_size = csrc->mc_db->md_pad;
7761 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7763 data.mv_data = NULL;
7767 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7768 mdb_cassert(csrc, !((size_t)srcnode & 1));
7769 srcpg = NODEPGNO(srcnode);
7770 flags = srcnode->mn_flags;
7771 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7772 unsigned int snum = csrc->mc_snum;
7774 /* must find the lowest key below src */
7775 rc = mdb_page_search_lowest(csrc);
7778 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7779 key.mv_size = csrc->mc_db->md_pad;
7780 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7782 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7783 key.mv_size = NODEKSZ(s2);
7784 key.mv_data = NODEKEY(s2);
7786 csrc->mc_snum = snum--;
7787 csrc->mc_top = snum;
7789 key.mv_size = NODEKSZ(srcnode);
7790 key.mv_data = NODEKEY(srcnode);
7792 data.mv_size = NODEDSZ(srcnode);
7793 data.mv_data = NODEDATA(srcnode);
7795 mn.mc_xcursor = NULL;
7796 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7797 unsigned int snum = cdst->mc_snum;
7800 /* must find the lowest key below dst */
7801 mdb_cursor_copy(cdst, &mn);
7802 rc = mdb_page_search_lowest(&mn);
7805 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7806 bkey.mv_size = mn.mc_db->md_pad;
7807 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7809 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7810 bkey.mv_size = NODEKSZ(s2);
7811 bkey.mv_data = NODEKEY(s2);
7813 mn.mc_snum = snum--;
7816 rc = mdb_update_key(&mn, &bkey);
7821 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7822 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7823 csrc->mc_ki[csrc->mc_top],
7825 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7826 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7828 /* Add the node to the destination page.
7830 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7831 if (rc != MDB_SUCCESS)
7834 /* Delete the node from the source page.
7836 mdb_node_del(csrc, key.mv_size);
7839 /* Adjust other cursors pointing to mp */
7840 MDB_cursor *m2, *m3;
7841 MDB_dbi dbi = csrc->mc_dbi;
7842 MDB_page *mpd, *mps;
7844 mps = csrc->mc_pg[csrc->mc_top];
7845 /* If we're adding on the left, bump others up */
7847 mpd = cdst->mc_pg[csrc->mc_top];
7848 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7849 if (csrc->mc_flags & C_SUB)
7850 m3 = &m2->mc_xcursor->mx_cursor;
7853 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7856 m3->mc_pg[csrc->mc_top] == mpd &&
7857 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7858 m3->mc_ki[csrc->mc_top]++;
7861 m3->mc_pg[csrc->mc_top] == mps &&
7862 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7863 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7864 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7865 m3->mc_ki[csrc->mc_top-1]++;
7867 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
7869 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7870 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
7871 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
7875 /* Adding on the right, bump others down */
7877 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7878 if (csrc->mc_flags & C_SUB)
7879 m3 = &m2->mc_xcursor->mx_cursor;
7882 if (m3 == csrc) continue;
7883 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7885 if (m3->mc_pg[csrc->mc_top] == mps) {
7886 if (!m3->mc_ki[csrc->mc_top]) {
7887 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7888 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7889 m3->mc_ki[csrc->mc_top-1]--;
7891 m3->mc_ki[csrc->mc_top]--;
7893 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
7895 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7896 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
7897 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
7904 /* Update the parent separators.
7906 if (csrc->mc_ki[csrc->mc_top] == 0) {
7907 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7908 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7909 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7911 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7912 key.mv_size = NODEKSZ(srcnode);
7913 key.mv_data = NODEKEY(srcnode);
7915 DPRINTF(("update separator for source page %"Z"u to [%s]",
7916 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7917 mdb_cursor_copy(csrc, &mn);
7920 /* We want mdb_rebalance to find mn when doing fixups */
7921 WITH_CURSOR_TRACKING(mn,
7922 rc = mdb_update_key(&mn, &key));
7926 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7928 indx_t ix = csrc->mc_ki[csrc->mc_top];
7929 nullkey.mv_size = 0;
7930 csrc->mc_ki[csrc->mc_top] = 0;
7931 rc = mdb_update_key(csrc, &nullkey);
7932 csrc->mc_ki[csrc->mc_top] = ix;
7933 mdb_cassert(csrc, rc == MDB_SUCCESS);
7937 if (cdst->mc_ki[cdst->mc_top] == 0) {
7938 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7939 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7940 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7942 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7943 key.mv_size = NODEKSZ(srcnode);
7944 key.mv_data = NODEKEY(srcnode);
7946 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7947 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7948 mdb_cursor_copy(cdst, &mn);
7951 /* We want mdb_rebalance to find mn when doing fixups */
7952 WITH_CURSOR_TRACKING(mn,
7953 rc = mdb_update_key(&mn, &key));
7957 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7959 indx_t ix = cdst->mc_ki[cdst->mc_top];
7960 nullkey.mv_size = 0;
7961 cdst->mc_ki[cdst->mc_top] = 0;
7962 rc = mdb_update_key(cdst, &nullkey);
7963 cdst->mc_ki[cdst->mc_top] = ix;
7964 mdb_cassert(cdst, rc == MDB_SUCCESS);
7971 /** Merge one page into another.
7972 * The nodes from the page pointed to by \b csrc will
7973 * be copied to the page pointed to by \b cdst and then
7974 * the \b csrc page will be freed.
7975 * @param[in] csrc Cursor pointing to the source page.
7976 * @param[in] cdst Cursor pointing to the destination page.
7977 * @return 0 on success, non-zero on failure.
7980 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7982 MDB_page *psrc, *pdst;
7989 psrc = csrc->mc_pg[csrc->mc_top];
7990 pdst = cdst->mc_pg[cdst->mc_top];
7992 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7994 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7995 mdb_cassert(csrc, cdst->mc_snum > 1);
7997 /* Mark dst as dirty. */
7998 if ((rc = mdb_page_touch(cdst)))
8001 /* get dst page again now that we've touched it. */
8002 pdst = cdst->mc_pg[cdst->mc_top];
8004 /* Move all nodes from src to dst.
8006 j = nkeys = NUMKEYS(pdst);
8007 if (IS_LEAF2(psrc)) {
8008 key.mv_size = csrc->mc_db->md_pad;
8009 key.mv_data = METADATA(psrc);
8010 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8011 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8012 if (rc != MDB_SUCCESS)
8014 key.mv_data = (char *)key.mv_data + key.mv_size;
8017 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8018 srcnode = NODEPTR(psrc, i);
8019 if (i == 0 && IS_BRANCH(psrc)) {
8022 mdb_cursor_copy(csrc, &mn);
8023 mn.mc_xcursor = NULL;
8024 /* must find the lowest key below src */
8025 rc = mdb_page_search_lowest(&mn);
8028 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8029 key.mv_size = mn.mc_db->md_pad;
8030 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8032 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8033 key.mv_size = NODEKSZ(s2);
8034 key.mv_data = NODEKEY(s2);
8037 key.mv_size = srcnode->mn_ksize;
8038 key.mv_data = NODEKEY(srcnode);
8041 data.mv_size = NODEDSZ(srcnode);
8042 data.mv_data = NODEDATA(srcnode);
8043 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8044 if (rc != MDB_SUCCESS)
8049 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
8050 pdst->mp_pgno, NUMKEYS(pdst),
8051 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8053 /* Unlink the src page from parent and add to free list.
8056 mdb_node_del(csrc, 0);
8057 if (csrc->mc_ki[csrc->mc_top] == 0) {
8059 rc = mdb_update_key(csrc, &key);
8067 psrc = csrc->mc_pg[csrc->mc_top];
8068 /* If not operating on FreeDB, allow this page to be reused
8069 * in this txn. Otherwise just add to free list.
8071 rc = mdb_page_loose(csrc, psrc);
8075 csrc->mc_db->md_leaf_pages--;
8077 csrc->mc_db->md_branch_pages--;
8079 /* Adjust other cursors pointing to mp */
8080 MDB_cursor *m2, *m3;
8081 MDB_dbi dbi = csrc->mc_dbi;
8082 unsigned int top = csrc->mc_top;
8084 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8085 if (csrc->mc_flags & C_SUB)
8086 m3 = &m2->mc_xcursor->mx_cursor;
8089 if (m3 == csrc) continue;
8090 if (m3->mc_snum < csrc->mc_snum) continue;
8091 if (m3->mc_pg[top] == psrc) {
8092 m3->mc_pg[top] = pdst;
8093 m3->mc_ki[top] += nkeys;
8094 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8095 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8096 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8099 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8101 MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
8102 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8103 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8108 unsigned int snum = cdst->mc_snum;
8109 uint16_t depth = cdst->mc_db->md_depth;
8110 mdb_cursor_pop(cdst);
8111 rc = mdb_rebalance(cdst);
8112 /* Did the tree height change? */
8113 if (depth != cdst->mc_db->md_depth)
8114 snum += cdst->mc_db->md_depth - depth;
8115 cdst->mc_snum = snum;
8116 cdst->mc_top = snum-1;
8121 /** Copy the contents of a cursor.
8122 * @param[in] csrc The cursor to copy from.
8123 * @param[out] cdst The cursor to copy to.
8126 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8130 cdst->mc_txn = csrc->mc_txn;
8131 cdst->mc_dbi = csrc->mc_dbi;
8132 cdst->mc_db = csrc->mc_db;
8133 cdst->mc_dbx = csrc->mc_dbx;
8134 cdst->mc_snum = csrc->mc_snum;
8135 cdst->mc_top = csrc->mc_top;
8136 cdst->mc_flags = csrc->mc_flags;
8138 for (i=0; i<csrc->mc_snum; i++) {
8139 cdst->mc_pg[i] = csrc->mc_pg[i];
8140 cdst->mc_ki[i] = csrc->mc_ki[i];
8144 /** Rebalance the tree after a delete operation.
8145 * @param[in] mc Cursor pointing to the page where rebalancing
8147 * @return 0 on success, non-zero on failure.
8150 mdb_rebalance(MDB_cursor *mc)
8154 unsigned int ptop, minkeys, thresh;
8158 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8163 thresh = FILL_THRESHOLD;
8165 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
8166 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8167 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8168 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8170 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8171 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8172 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
8173 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8177 if (mc->mc_snum < 2) {
8178 MDB_page *mp = mc->mc_pg[0];
8180 DPUTS("Can't rebalance a subpage, ignoring");
8183 if (NUMKEYS(mp) == 0) {
8184 DPUTS("tree is completely empty");
8185 mc->mc_db->md_root = P_INVALID;
8186 mc->mc_db->md_depth = 0;
8187 mc->mc_db->md_leaf_pages = 0;
8188 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8191 /* Adjust cursors pointing to mp */
8194 mc->mc_flags &= ~C_INITIALIZED;
8196 MDB_cursor *m2, *m3;
8197 MDB_dbi dbi = mc->mc_dbi;
8199 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8200 if (mc->mc_flags & C_SUB)
8201 m3 = &m2->mc_xcursor->mx_cursor;
8204 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8206 if (m3->mc_pg[0] == mp) {
8209 m3->mc_flags &= ~C_INITIALIZED;
8213 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8215 DPUTS("collapsing root page!");
8216 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8219 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8220 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
8223 mc->mc_db->md_depth--;
8224 mc->mc_db->md_branch_pages--;
8225 mc->mc_ki[0] = mc->mc_ki[1];
8226 for (i = 1; i<mc->mc_db->md_depth; i++) {
8227 mc->mc_pg[i] = mc->mc_pg[i+1];
8228 mc->mc_ki[i] = mc->mc_ki[i+1];
8231 /* Adjust other cursors pointing to mp */
8232 MDB_cursor *m2, *m3;
8233 MDB_dbi dbi = mc->mc_dbi;
8235 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8236 if (mc->mc_flags & C_SUB)
8237 m3 = &m2->mc_xcursor->mx_cursor;
8240 if (m3 == mc) continue;
8241 if (!(m3->mc_flags & C_INITIALIZED))
8243 if (m3->mc_pg[0] == mp) {
8244 for (i=0; i<mc->mc_db->md_depth; i++) {
8245 m3->mc_pg[i] = m3->mc_pg[i+1];
8246 m3->mc_ki[i] = m3->mc_ki[i+1];
8254 DPUTS("root page doesn't need rebalancing");
8258 /* The parent (branch page) must have at least 2 pointers,
8259 * otherwise the tree is invalid.
8261 ptop = mc->mc_top-1;
8262 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8264 /* Leaf page fill factor is below the threshold.
8265 * Try to move keys from left or right neighbor, or
8266 * merge with a neighbor page.
8271 mdb_cursor_copy(mc, &mn);
8272 mn.mc_xcursor = NULL;
8274 oldki = mc->mc_ki[mc->mc_top];
8275 if (mc->mc_ki[ptop] == 0) {
8276 /* We're the leftmost leaf in our parent.
8278 DPUTS("reading right neighbor");
8280 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8281 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8284 mn.mc_ki[mn.mc_top] = 0;
8285 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8288 /* There is at least one neighbor to the left.
8290 DPUTS("reading left neighbor");
8292 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8293 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8296 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8297 mc->mc_ki[mc->mc_top] = 0;
8301 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8302 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8303 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8305 /* If the neighbor page is above threshold and has enough keys,
8306 * move one key from it. Otherwise we should try to merge them.
8307 * (A branch page must never have less than 2 keys.)
8309 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8310 rc = mdb_node_move(&mn, mc, fromleft);
8312 /* if we inserted on left, bump position up */
8317 rc = mdb_page_merge(&mn, mc);
8319 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8320 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8321 /* We want mdb_rebalance to find mn when doing fixups */
8322 WITH_CURSOR_TRACKING(mn,
8323 rc = mdb_page_merge(mc, &mn));
8324 mdb_cursor_copy(&mn, mc);
8326 mc->mc_flags &= ~C_EOF;
8328 mc->mc_ki[mc->mc_top] = oldki;
8332 /** Complete a delete operation started by #mdb_cursor_del(). */
8334 mdb_cursor_del0(MDB_cursor *mc)
8340 MDB_cursor *m2, *m3;
8341 MDB_dbi dbi = mc->mc_dbi;
8343 ki = mc->mc_ki[mc->mc_top];
8344 mp = mc->mc_pg[mc->mc_top];
8345 mdb_node_del(mc, mc->mc_db->md_pad);
8346 mc->mc_db->md_entries--;
8348 /* Adjust other cursors pointing to mp */
8349 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8350 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8351 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8353 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8355 if (m3->mc_pg[mc->mc_top] == mp) {
8356 if (m3->mc_ki[mc->mc_top] == ki) {
8357 m3->mc_flags |= C_DEL;
8358 if (mc->mc_db->md_flags & MDB_DUPSORT)
8359 m3->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
8360 } else if (m3->mc_ki[mc->mc_top] > ki) {
8361 m3->mc_ki[mc->mc_top]--;
8363 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
8364 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8365 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8366 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8371 rc = mdb_rebalance(mc);
8373 if (rc == MDB_SUCCESS) {
8374 /* DB is totally empty now, just bail out.
8375 * Other cursors adjustments were already done
8376 * by mdb_rebalance and aren't needed here.
8381 mp = mc->mc_pg[mc->mc_top];
8382 nkeys = NUMKEYS(mp);
8384 /* Adjust other cursors pointing to mp */
8385 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8386 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8387 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8389 if (m3->mc_snum < mc->mc_snum)
8391 if (m3->mc_pg[mc->mc_top] == mp) {
8392 /* if m3 points past last node in page, find next sibling */
8393 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8394 rc = mdb_cursor_sibling(m3, 1);
8395 if (rc == MDB_NOTFOUND) {
8396 m3->mc_flags |= C_EOF;
8402 mc->mc_flags |= C_DEL;
8406 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8411 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8412 MDB_val *key, MDB_val *data)
8414 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8417 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8418 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8420 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8421 /* must ignore any data */
8425 return mdb_del0(txn, dbi, key, data, 0);
8429 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8430 MDB_val *key, MDB_val *data, unsigned flags)
8435 MDB_val rdata, *xdata;
8439 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8441 mdb_cursor_init(&mc, txn, dbi, &mx);
8450 flags |= MDB_NODUPDATA;
8452 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8454 /* let mdb_page_split know about this cursor if needed:
8455 * delete will trigger a rebalance; if it needs to move
8456 * a node from one page to another, it will have to
8457 * update the parent's separator key(s). If the new sepkey
8458 * is larger than the current one, the parent page may
8459 * run out of space, triggering a split. We need this
8460 * cursor to be consistent until the end of the rebalance.
8462 mc.mc_flags |= C_UNTRACK;
8463 mc.mc_next = txn->mt_cursors[dbi];
8464 txn->mt_cursors[dbi] = &mc;
8465 rc = mdb_cursor_del(&mc, flags);
8466 txn->mt_cursors[dbi] = mc.mc_next;
8471 /** Split a page and insert a new node.
8472 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8473 * The cursor will be updated to point to the actual page and index where
8474 * the node got inserted after the split.
8475 * @param[in] newkey The key for the newly inserted node.
8476 * @param[in] newdata The data for the newly inserted node.
8477 * @param[in] newpgno The page number, if the new node is a branch node.
8478 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8479 * @return 0 on success, non-zero on failure.
8482 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8483 unsigned int nflags)
8486 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8489 int i, j, split_indx, nkeys, pmax;
8490 MDB_env *env = mc->mc_txn->mt_env;
8492 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8493 MDB_page *copy = NULL;
8494 MDB_page *mp, *rp, *pp;
8499 mp = mc->mc_pg[mc->mc_top];
8500 newindx = mc->mc_ki[mc->mc_top];
8501 nkeys = NUMKEYS(mp);
8503 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8504 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8505 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8507 /* Create a right sibling. */
8508 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8510 rp->mp_pad = mp->mp_pad;
8511 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8513 /* Usually when splitting the root page, the cursor
8514 * height is 1. But when called from mdb_update_key,
8515 * the cursor height may be greater because it walks
8516 * up the stack while finding the branch slot to update.
8518 if (mc->mc_top < 1) {
8519 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8521 /* shift current top to make room for new parent */
8522 for (i=mc->mc_snum; i>0; i--) {
8523 mc->mc_pg[i] = mc->mc_pg[i-1];
8524 mc->mc_ki[i] = mc->mc_ki[i-1];
8528 mc->mc_db->md_root = pp->mp_pgno;
8529 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8530 new_root = mc->mc_db->md_depth++;
8532 /* Add left (implicit) pointer. */
8533 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8534 /* undo the pre-push */
8535 mc->mc_pg[0] = mc->mc_pg[1];
8536 mc->mc_ki[0] = mc->mc_ki[1];
8537 mc->mc_db->md_root = mp->mp_pgno;
8538 mc->mc_db->md_depth--;
8545 ptop = mc->mc_top-1;
8546 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8549 mdb_cursor_copy(mc, &mn);
8550 mn.mc_xcursor = NULL;
8551 mn.mc_pg[mn.mc_top] = rp;
8552 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8554 if (nflags & MDB_APPEND) {
8555 mn.mc_ki[mn.mc_top] = 0;
8557 split_indx = newindx;
8561 split_indx = (nkeys+1) / 2;
8566 unsigned int lsize, rsize, ksize;
8567 /* Move half of the keys to the right sibling */
8568 x = mc->mc_ki[mc->mc_top] - split_indx;
8569 ksize = mc->mc_db->md_pad;
8570 split = LEAF2KEY(mp, split_indx, ksize);
8571 rsize = (nkeys - split_indx) * ksize;
8572 lsize = (nkeys - split_indx) * sizeof(indx_t);
8573 mp->mp_lower -= lsize;
8574 rp->mp_lower += lsize;
8575 mp->mp_upper += rsize - lsize;
8576 rp->mp_upper -= rsize - lsize;
8577 sepkey.mv_size = ksize;
8578 if (newindx == split_indx) {
8579 sepkey.mv_data = newkey->mv_data;
8581 sepkey.mv_data = split;
8584 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8585 memcpy(rp->mp_ptrs, split, rsize);
8586 sepkey.mv_data = rp->mp_ptrs;
8587 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8588 memcpy(ins, newkey->mv_data, ksize);
8589 mp->mp_lower += sizeof(indx_t);
8590 mp->mp_upper -= ksize - sizeof(indx_t);
8593 memcpy(rp->mp_ptrs, split, x * ksize);
8594 ins = LEAF2KEY(rp, x, ksize);
8595 memcpy(ins, newkey->mv_data, ksize);
8596 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8597 rp->mp_lower += sizeof(indx_t);
8598 rp->mp_upper -= ksize - sizeof(indx_t);
8599 mc->mc_ki[mc->mc_top] = x;
8602 int psize, nsize, k;
8603 /* Maximum free space in an empty page */
8604 pmax = env->me_psize - PAGEHDRSZ;
8606 nsize = mdb_leaf_size(env, newkey, newdata);
8608 nsize = mdb_branch_size(env, newkey);
8609 nsize = EVEN(nsize);
8611 /* grab a page to hold a temporary copy */
8612 copy = mdb_page_malloc(mc->mc_txn, 1);
8617 copy->mp_pgno = mp->mp_pgno;
8618 copy->mp_flags = mp->mp_flags;
8619 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8620 copy->mp_upper = env->me_psize - PAGEBASE;
8622 /* prepare to insert */
8623 for (i=0, j=0; i<nkeys; i++) {
8625 copy->mp_ptrs[j++] = 0;
8627 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8630 /* When items are relatively large the split point needs
8631 * to be checked, because being off-by-one will make the
8632 * difference between success or failure in mdb_node_add.
8634 * It's also relevant if a page happens to be laid out
8635 * such that one half of its nodes are all "small" and
8636 * the other half of its nodes are "large." If the new
8637 * item is also "large" and falls on the half with
8638 * "large" nodes, it also may not fit.
8640 * As a final tweak, if the new item goes on the last
8641 * spot on the page (and thus, onto the new page), bias
8642 * the split so the new page is emptier than the old page.
8643 * This yields better packing during sequential inserts.
8645 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8646 /* Find split point */
8648 if (newindx <= split_indx || newindx >= nkeys) {
8650 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8655 for (; i!=k; i+=j) {
8660 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8661 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8663 if (F_ISSET(node->mn_flags, F_BIGDATA))
8664 psize += sizeof(pgno_t);
8666 psize += NODEDSZ(node);
8668 psize = EVEN(psize);
8670 if (psize > pmax || i == k-j) {
8671 split_indx = i + (j<0);
8676 if (split_indx == newindx) {
8677 sepkey.mv_size = newkey->mv_size;
8678 sepkey.mv_data = newkey->mv_data;
8680 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8681 sepkey.mv_size = node->mn_ksize;
8682 sepkey.mv_data = NODEKEY(node);
8687 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8689 /* Copy separator key to the parent.
8691 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8692 int snum = mc->mc_snum;
8696 /* We want other splits to find mn when doing fixups */
8697 WITH_CURSOR_TRACKING(mn,
8698 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
8703 if (mc->mc_snum > snum) {
8706 /* Right page might now have changed parent.
8707 * Check if left page also changed parent.
8709 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8710 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8711 for (i=0; i<ptop; i++) {
8712 mc->mc_pg[i] = mn.mc_pg[i];
8713 mc->mc_ki[i] = mn.mc_ki[i];
8715 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8716 if (mn.mc_ki[ptop]) {
8717 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8719 /* find right page's left sibling */
8720 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8721 mdb_cursor_sibling(mc, 0);
8726 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8729 if (rc != MDB_SUCCESS) {
8732 if (nflags & MDB_APPEND) {
8733 mc->mc_pg[mc->mc_top] = rp;
8734 mc->mc_ki[mc->mc_top] = 0;
8735 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8738 for (i=0; i<mc->mc_top; i++)
8739 mc->mc_ki[i] = mn.mc_ki[i];
8740 } else if (!IS_LEAF2(mp)) {
8742 mc->mc_pg[mc->mc_top] = rp;
8747 rkey.mv_data = newkey->mv_data;
8748 rkey.mv_size = newkey->mv_size;
8754 /* Update index for the new key. */
8755 mc->mc_ki[mc->mc_top] = j;
8757 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8758 rkey.mv_data = NODEKEY(node);
8759 rkey.mv_size = node->mn_ksize;
8761 xdata.mv_data = NODEDATA(node);
8762 xdata.mv_size = NODEDSZ(node);
8765 pgno = NODEPGNO(node);
8766 flags = node->mn_flags;
8769 if (!IS_LEAF(mp) && j == 0) {
8770 /* First branch index doesn't need key data. */
8774 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8780 mc->mc_pg[mc->mc_top] = copy;
8785 } while (i != split_indx);
8787 nkeys = NUMKEYS(copy);
8788 for (i=0; i<nkeys; i++)
8789 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8790 mp->mp_lower = copy->mp_lower;
8791 mp->mp_upper = copy->mp_upper;
8792 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8793 env->me_psize - copy->mp_upper - PAGEBASE);
8795 /* reset back to original page */
8796 if (newindx < split_indx) {
8797 mc->mc_pg[mc->mc_top] = mp;
8799 mc->mc_pg[mc->mc_top] = rp;
8801 /* Make sure mc_ki is still valid.
8803 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8804 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8805 for (i=0; i<=ptop; i++) {
8806 mc->mc_pg[i] = mn.mc_pg[i];
8807 mc->mc_ki[i] = mn.mc_ki[i];
8811 if (nflags & MDB_RESERVE) {
8812 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8813 if (!(node->mn_flags & F_BIGDATA))
8814 newdata->mv_data = NODEDATA(node);
8817 if (newindx >= split_indx) {
8818 mc->mc_pg[mc->mc_top] = rp;
8820 /* Make sure mc_ki is still valid.
8822 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8823 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8824 for (i=0; i<=ptop; i++) {
8825 mc->mc_pg[i] = mn.mc_pg[i];
8826 mc->mc_ki[i] = mn.mc_ki[i];
8833 /* Adjust other cursors pointing to mp */
8834 MDB_cursor *m2, *m3;
8835 MDB_dbi dbi = mc->mc_dbi;
8836 nkeys = NUMKEYS(mp);
8838 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8839 if (mc->mc_flags & C_SUB)
8840 m3 = &m2->mc_xcursor->mx_cursor;
8845 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8849 /* sub cursors may be on different DB */
8850 if (m3->mc_pg[0] != mp)
8853 for (k=new_root; k>=0; k--) {
8854 m3->mc_ki[k+1] = m3->mc_ki[k];
8855 m3->mc_pg[k+1] = m3->mc_pg[k];
8857 if (m3->mc_ki[0] >= nkeys) {
8862 m3->mc_pg[0] = mc->mc_pg[0];
8866 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8867 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8868 m3->mc_ki[mc->mc_top]++;
8869 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8870 m3->mc_pg[mc->mc_top] = rp;
8871 m3->mc_ki[mc->mc_top] -= nkeys;
8872 for (i=0; i<mc->mc_top; i++) {
8873 m3->mc_ki[i] = mn.mc_ki[i];
8874 m3->mc_pg[i] = mn.mc_pg[i];
8877 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8878 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8881 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8883 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8884 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8885 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8889 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8892 if (copy) /* tmp page */
8893 mdb_page_free(env, copy);
8895 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8900 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8901 MDB_val *key, MDB_val *data, unsigned int flags)
8907 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8910 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8913 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8914 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8916 mdb_cursor_init(&mc, txn, dbi, &mx);
8917 mc.mc_next = txn->mt_cursors[dbi];
8918 txn->mt_cursors[dbi] = &mc;
8919 rc = mdb_cursor_put(&mc, key, data, flags);
8920 txn->mt_cursors[dbi] = mc.mc_next;
8925 #define MDB_WBUF (1024*1024)
8928 /** State needed for a compacting copy. */
8929 typedef struct mdb_copy {
8930 pthread_mutex_t mc_mutex;
8931 pthread_cond_t mc_cond;
8938 pgno_t mc_next_pgno;
8941 volatile int mc_new;
8946 /** Dedicated writer thread for compacting copy. */
8947 static THREAD_RET ESECT CALL_CONV
8948 mdb_env_copythr(void *arg)
8952 int toggle = 0, wsize, rc;
8955 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8958 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8961 pthread_mutex_lock(&my->mc_mutex);
8963 pthread_cond_signal(&my->mc_cond);
8966 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8967 if (my->mc_new < 0) {
8972 wsize = my->mc_wlen[toggle];
8973 ptr = my->mc_wbuf[toggle];
8976 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8980 } else if (len > 0) {
8994 /* If there's an overflow page tail, write it too */
8995 if (my->mc_olen[toggle]) {
8996 wsize = my->mc_olen[toggle];
8997 ptr = my->mc_over[toggle];
8998 my->mc_olen[toggle] = 0;
9001 my->mc_wlen[toggle] = 0;
9003 pthread_cond_signal(&my->mc_cond);
9005 pthread_cond_signal(&my->mc_cond);
9006 pthread_mutex_unlock(&my->mc_mutex);
9007 return (THREAD_RET)0;
9011 /** Tell the writer thread there's a buffer ready to write */
9013 mdb_env_cthr_toggle(mdb_copy *my, int st)
9015 int toggle = my->mc_toggle ^ 1;
9016 pthread_mutex_lock(&my->mc_mutex);
9017 if (my->mc_status) {
9018 pthread_mutex_unlock(&my->mc_mutex);
9019 return my->mc_status;
9021 while (my->mc_new == 1)
9022 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9024 my->mc_toggle = toggle;
9025 pthread_cond_signal(&my->mc_cond);
9026 pthread_mutex_unlock(&my->mc_mutex);
9030 /** Depth-first tree traversal for compacting copy. */
9032 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9035 MDB_txn *txn = my->mc_txn;
9037 MDB_page *mo, *mp, *leaf;
9042 /* Empty DB, nothing to do */
9043 if (*pg == P_INVALID)
9050 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
9053 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9057 /* Make cursor pages writable */
9058 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9062 for (i=0; i<mc.mc_top; i++) {
9063 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9064 mc.mc_pg[i] = (MDB_page *)ptr;
9065 ptr += my->mc_env->me_psize;
9068 /* This is writable space for a leaf page. Usually not needed. */
9069 leaf = (MDB_page *)ptr;
9071 toggle = my->mc_toggle;
9072 while (mc.mc_snum > 0) {
9074 mp = mc.mc_pg[mc.mc_top];
9078 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9079 for (i=0; i<n; i++) {
9080 ni = NODEPTR(mp, i);
9081 if (ni->mn_flags & F_BIGDATA) {
9085 /* Need writable leaf */
9087 mc.mc_pg[mc.mc_top] = leaf;
9088 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9090 ni = NODEPTR(mp, i);
9093 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9094 rc = mdb_page_get(txn, pg, &omp, NULL);
9097 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9098 rc = mdb_env_cthr_toggle(my, 1);
9101 toggle = my->mc_toggle;
9103 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9104 memcpy(mo, omp, my->mc_env->me_psize);
9105 mo->mp_pgno = my->mc_next_pgno;
9106 my->mc_next_pgno += omp->mp_pages;
9107 my->mc_wlen[toggle] += my->mc_env->me_psize;
9108 if (omp->mp_pages > 1) {
9109 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9110 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9111 rc = mdb_env_cthr_toggle(my, 1);
9114 toggle = my->mc_toggle;
9116 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
9117 } else if (ni->mn_flags & F_SUBDATA) {
9120 /* Need writable leaf */
9122 mc.mc_pg[mc.mc_top] = leaf;
9123 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9125 ni = NODEPTR(mp, i);
9128 memcpy(&db, NODEDATA(ni), sizeof(db));
9129 my->mc_toggle = toggle;
9130 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9133 toggle = my->mc_toggle;
9134 memcpy(NODEDATA(ni), &db, sizeof(db));
9139 mc.mc_ki[mc.mc_top]++;
9140 if (mc.mc_ki[mc.mc_top] < n) {
9143 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9145 rc = mdb_page_get(txn, pg, &mp, NULL);
9150 mc.mc_ki[mc.mc_top] = 0;
9151 if (IS_BRANCH(mp)) {
9152 /* Whenever we advance to a sibling branch page,
9153 * we must proceed all the way down to its first leaf.
9155 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9158 mc.mc_pg[mc.mc_top] = mp;
9162 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9163 rc = mdb_env_cthr_toggle(my, 1);
9166 toggle = my->mc_toggle;
9168 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9169 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9170 mo->mp_pgno = my->mc_next_pgno++;
9171 my->mc_wlen[toggle] += my->mc_env->me_psize;
9173 /* Update parent if there is one */
9174 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9175 SETPGNO(ni, mo->mp_pgno);
9176 mdb_cursor_pop(&mc);
9178 /* Otherwise we're done */
9188 /** Copy environment with compaction. */
9190 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9195 MDB_txn *txn = NULL;
9200 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
9201 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
9202 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9203 if (my.mc_wbuf[0] == NULL)
9206 pthread_mutex_init(&my.mc_mutex, NULL);
9207 pthread_cond_init(&my.mc_cond, NULL);
9208 #ifdef HAVE_MEMALIGN
9209 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9210 if (my.mc_wbuf[0] == NULL)
9213 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
9218 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9219 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9224 my.mc_next_pgno = NUM_METAS;
9230 THREAD_CREATE(thr, mdb_env_copythr, &my);
9232 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9236 mp = (MDB_page *)my.mc_wbuf[0];
9237 memset(mp, 0, NUM_METAS * env->me_psize);
9239 mp->mp_flags = P_META;
9240 mm = (MDB_meta *)METADATA(mp);
9241 mdb_env_init_meta0(env, mm);
9242 mm->mm_address = env->me_metas[0]->mm_address;
9244 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9246 mp->mp_flags = P_META;
9247 *(MDB_meta *)METADATA(mp) = *mm;
9248 mm = (MDB_meta *)METADATA(mp);
9250 /* Count the number of free pages, subtract from lastpg to find
9251 * number of active pages
9254 MDB_ID freecount = 0;
9257 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9258 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9259 freecount += *(MDB_ID *)data.mv_data;
9260 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9261 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9262 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9264 /* Set metapage 1 */
9265 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
9266 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9267 if (mm->mm_last_pg > NUM_METAS-1) {
9268 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
9271 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
9274 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9276 pthread_mutex_lock(&my.mc_mutex);
9278 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9279 pthread_mutex_unlock(&my.mc_mutex);
9280 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
9281 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
9282 rc = mdb_env_cthr_toggle(&my, 1);
9283 mdb_env_cthr_toggle(&my, -1);
9284 pthread_mutex_lock(&my.mc_mutex);
9286 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9287 pthread_mutex_unlock(&my.mc_mutex);
9292 CloseHandle(my.mc_cond);
9293 CloseHandle(my.mc_mutex);
9294 _aligned_free(my.mc_wbuf[0]);
9296 pthread_cond_destroy(&my.mc_cond);
9297 pthread_mutex_destroy(&my.mc_mutex);
9298 free(my.mc_wbuf[0]);
9303 /** Copy environment as-is. */
9305 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9307 MDB_txn *txn = NULL;
9308 mdb_mutexref_t wmutex = NULL;
9314 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9318 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9321 /* Do the lock/unlock of the reader mutex before starting the
9322 * write txn. Otherwise other read txns could block writers.
9324 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9329 /* We must start the actual read txn after blocking writers */
9330 mdb_txn_end(txn, MDB_END_RESET_TMP);
9332 /* Temporarily block writers until we snapshot the meta pages */
9333 wmutex = env->me_wmutex;
9334 if (LOCK_MUTEX(rc, env, wmutex))
9337 rc = mdb_txn_renew0(txn);
9339 UNLOCK_MUTEX(wmutex);
9344 wsize = env->me_psize * NUM_METAS;
9348 DO_WRITE(rc, fd, ptr, w2, len);
9352 } else if (len > 0) {
9358 /* Non-blocking or async handles are not supported */
9364 UNLOCK_MUTEX(wmutex);
9369 w2 = txn->mt_next_pgno * env->me_psize;
9372 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9379 if (wsize > MAX_WRITE)
9383 DO_WRITE(rc, fd, ptr, w2, len);
9387 } else if (len > 0) {
9404 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9406 if (flags & MDB_CP_COMPACT)
9407 return mdb_env_copyfd1(env, fd);
9409 return mdb_env_copyfd0(env, fd);
9413 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9415 return mdb_env_copyfd2(env, fd, 0);
9419 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9423 HANDLE newfd = INVALID_HANDLE_VALUE;
9428 if (env->me_flags & MDB_NOSUBDIR) {
9429 lpath = (char *)path;
9432 len += sizeof(DATANAME);
9433 lpath = malloc(len);
9436 sprintf(lpath, "%s" DATANAME, path);
9439 /* The destination path must exist, but the destination file must not.
9440 * We don't want the OS to cache the writes, since the source data is
9441 * already in the OS cache.
9444 utf8_to_utf16(lpath, -1, &wpath, NULL);
9445 newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9446 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9449 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9451 if (newfd == INVALID_HANDLE_VALUE) {
9456 if (env->me_psize >= env->me_os_psize) {
9458 /* Set O_DIRECT if the file system supports it */
9459 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9460 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9462 #ifdef F_NOCACHE /* __APPLE__ */
9463 rc = fcntl(newfd, F_NOCACHE, 1);
9471 rc = mdb_env_copyfd2(env, newfd, flags);
9474 if (!(env->me_flags & MDB_NOSUBDIR))
9476 if (newfd != INVALID_HANDLE_VALUE)
9477 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9484 mdb_env_copy(MDB_env *env, const char *path)
9486 return mdb_env_copy2(env, path, 0);
9490 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9492 if (flag & ~CHANGEABLE)
9495 env->me_flags |= flag;
9497 env->me_flags &= ~flag;
9502 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9507 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9512 mdb_env_set_userctx(MDB_env *env, void *ctx)
9516 env->me_userctx = ctx;
9521 mdb_env_get_userctx(MDB_env *env)
9523 return env ? env->me_userctx : NULL;
9527 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9532 env->me_assert_func = func;
9538 mdb_env_get_path(MDB_env *env, const char **arg)
9543 *arg = env->me_path;
9548 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9557 /** Common code for #mdb_stat() and #mdb_env_stat().
9558 * @param[in] env the environment to operate in.
9559 * @param[in] db the #MDB_db record containing the stats to return.
9560 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9561 * @return 0, this function always succeeds.
9564 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9566 arg->ms_psize = env->me_psize;
9567 arg->ms_depth = db->md_depth;
9568 arg->ms_branch_pages = db->md_branch_pages;
9569 arg->ms_leaf_pages = db->md_leaf_pages;
9570 arg->ms_overflow_pages = db->md_overflow_pages;
9571 arg->ms_entries = db->md_entries;
9577 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9581 if (env == NULL || arg == NULL)
9584 meta = mdb_env_pick_meta(env);
9586 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9590 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9594 if (env == NULL || arg == NULL)
9597 meta = mdb_env_pick_meta(env);
9598 arg->me_mapaddr = meta->mm_address;
9599 arg->me_last_pgno = meta->mm_last_pg;
9600 arg->me_last_txnid = meta->mm_txnid;
9602 arg->me_mapsize = env->me_mapsize;
9603 arg->me_maxreaders = env->me_maxreaders;
9604 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9608 /** Set the default comparison functions for a database.
9609 * Called immediately after a database is opened to set the defaults.
9610 * The user can then override them with #mdb_set_compare() or
9611 * #mdb_set_dupsort().
9612 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9613 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9616 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9618 uint16_t f = txn->mt_dbs[dbi].md_flags;
9620 txn->mt_dbxs[dbi].md_cmp =
9621 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9622 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9624 txn->mt_dbxs[dbi].md_dcmp =
9625 !(f & MDB_DUPSORT) ? 0 :
9626 ((f & MDB_INTEGERDUP)
9627 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9628 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9631 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9637 int rc, dbflag, exact;
9638 unsigned int unused = 0, seq;
9641 if (flags & ~VALID_FLAGS)
9643 if (txn->mt_flags & MDB_TXN_BLOCKED)
9649 if (flags & PERSISTENT_FLAGS) {
9650 uint16_t f2 = flags & PERSISTENT_FLAGS;
9651 /* make sure flag changes get committed */
9652 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9653 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9654 txn->mt_flags |= MDB_TXN_DIRTY;
9657 mdb_default_cmp(txn, MAIN_DBI);
9661 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9662 mdb_default_cmp(txn, MAIN_DBI);
9665 /* Is the DB already open? */
9667 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9668 if (!txn->mt_dbxs[i].md_name.mv_size) {
9669 /* Remember this free slot */
9670 if (!unused) unused = i;
9673 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9674 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9680 /* If no free slot and max hit, fail */
9681 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9682 return MDB_DBS_FULL;
9684 /* Cannot mix named databases with some mainDB flags */
9685 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9686 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9688 /* Find the DB info */
9689 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9692 key.mv_data = (void *)name;
9693 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9694 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9695 if (rc == MDB_SUCCESS) {
9696 /* make sure this is actually a DB */
9697 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9698 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9699 return MDB_INCOMPATIBLE;
9700 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9701 /* Create if requested */
9702 data.mv_size = sizeof(MDB_db);
9703 data.mv_data = &dummy;
9704 memset(&dummy, 0, sizeof(dummy));
9705 dummy.md_root = P_INVALID;
9706 dummy.md_flags = flags & PERSISTENT_FLAGS;
9707 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9711 /* OK, got info, add to table */
9712 if (rc == MDB_SUCCESS) {
9713 unsigned int slot = unused ? unused : txn->mt_numdbs;
9714 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9715 txn->mt_dbxs[slot].md_name.mv_size = len;
9716 txn->mt_dbxs[slot].md_rel = NULL;
9717 txn->mt_dbflags[slot] = dbflag;
9718 /* txn-> and env-> are the same in read txns, use
9719 * tmp variable to avoid undefined assignment
9721 seq = ++txn->mt_env->me_dbiseqs[slot];
9722 txn->mt_dbiseqs[slot] = seq;
9724 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9726 mdb_default_cmp(txn, slot);
9736 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9738 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9741 if (txn->mt_flags & MDB_TXN_BLOCKED)
9744 if (txn->mt_dbflags[dbi] & DB_STALE) {
9747 /* Stale, must read the DB's root. cursor_init does it for us. */
9748 mdb_cursor_init(&mc, txn, dbi, &mx);
9750 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9753 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9756 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9758 ptr = env->me_dbxs[dbi].md_name.mv_data;
9759 /* If there was no name, this was already closed */
9761 env->me_dbxs[dbi].md_name.mv_data = NULL;
9762 env->me_dbxs[dbi].md_name.mv_size = 0;
9763 env->me_dbflags[dbi] = 0;
9764 env->me_dbiseqs[dbi]++;
9769 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9771 /* We could return the flags for the FREE_DBI too but what's the point? */
9772 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9774 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9778 /** Add all the DB's pages to the free list.
9779 * @param[in] mc Cursor on the DB to free.
9780 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9781 * @return 0 on success, non-zero on failure.
9784 mdb_drop0(MDB_cursor *mc, int subs)
9788 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9789 if (rc == MDB_SUCCESS) {
9790 MDB_txn *txn = mc->mc_txn;
9795 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9796 * This also avoids any P_LEAF2 pages, which have no nodes.
9798 if (mc->mc_flags & C_SUB)
9801 mdb_cursor_copy(mc, &mx);
9802 while (mc->mc_snum > 0) {
9803 MDB_page *mp = mc->mc_pg[mc->mc_top];
9804 unsigned n = NUMKEYS(mp);
9806 for (i=0; i<n; i++) {
9807 ni = NODEPTR(mp, i);
9808 if (ni->mn_flags & F_BIGDATA) {
9811 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9812 rc = mdb_page_get(txn, pg, &omp, NULL);
9815 mdb_cassert(mc, IS_OVERFLOW(omp));
9816 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9820 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9821 mdb_xcursor_init1(mc, ni);
9822 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9828 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9830 for (i=0; i<n; i++) {
9832 ni = NODEPTR(mp, i);
9835 mdb_midl_xappend(txn->mt_free_pgs, pg);
9840 mc->mc_ki[mc->mc_top] = i;
9841 rc = mdb_cursor_sibling(mc, 1);
9843 if (rc != MDB_NOTFOUND)
9845 /* no more siblings, go back to beginning
9846 * of previous level.
9850 for (i=1; i<mc->mc_snum; i++) {
9852 mc->mc_pg[i] = mx.mc_pg[i];
9857 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9860 txn->mt_flags |= MDB_TXN_ERROR;
9861 } else if (rc == MDB_NOTFOUND) {
9864 mc->mc_flags &= ~C_INITIALIZED;
9868 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9870 MDB_cursor *mc, *m2;
9873 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9876 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9879 if (TXN_DBI_CHANGED(txn, dbi))
9882 rc = mdb_cursor_open(txn, dbi, &mc);
9886 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9887 /* Invalidate the dropped DB's cursors */
9888 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9889 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9893 /* Can't delete the main DB */
9894 if (del && dbi >= CORE_DBS) {
9895 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9897 txn->mt_dbflags[dbi] = DB_STALE;
9898 mdb_dbi_close(txn->mt_env, dbi);
9900 txn->mt_flags |= MDB_TXN_ERROR;
9903 /* reset the DB record, mark it dirty */
9904 txn->mt_dbflags[dbi] |= DB_DIRTY;
9905 txn->mt_dbs[dbi].md_depth = 0;
9906 txn->mt_dbs[dbi].md_branch_pages = 0;
9907 txn->mt_dbs[dbi].md_leaf_pages = 0;
9908 txn->mt_dbs[dbi].md_overflow_pages = 0;
9909 txn->mt_dbs[dbi].md_entries = 0;
9910 txn->mt_dbs[dbi].md_root = P_INVALID;
9912 txn->mt_flags |= MDB_TXN_DIRTY;
9915 mdb_cursor_close(mc);
9919 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9921 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9924 txn->mt_dbxs[dbi].md_cmp = cmp;
9928 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9930 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9933 txn->mt_dbxs[dbi].md_dcmp = cmp;
9937 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9939 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9942 txn->mt_dbxs[dbi].md_rel = rel;
9946 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9948 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9951 txn->mt_dbxs[dbi].md_relctx = ctx;
9956 mdb_env_get_maxkeysize(MDB_env *env)
9958 return ENV_MAXKEY(env);
9962 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9964 unsigned int i, rdrs;
9967 int rc = 0, first = 1;
9971 if (!env->me_txns) {
9972 return func("(no reader locks)\n", ctx);
9974 rdrs = env->me_txns->mti_numreaders;
9975 mr = env->me_txns->mti_readers;
9976 for (i=0; i<rdrs; i++) {
9978 txnid_t txnid = mr[i].mr_txnid;
9979 sprintf(buf, txnid == (txnid_t)-1 ?
9980 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9981 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9984 rc = func(" pid thread txnid\n", ctx);
9988 rc = func(buf, ctx);
9994 rc = func("(no active readers)\n", ctx);
9999 /** Insert pid into list if not already present.
10000 * return -1 if already present.
10003 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10005 /* binary search of pid in list */
10007 unsigned cursor = 1;
10009 unsigned n = ids[0];
10012 unsigned pivot = n >> 1;
10013 cursor = base + pivot + 1;
10014 val = pid - ids[cursor];
10019 } else if ( val > 0 ) {
10024 /* found, so it's a duplicate */
10033 for (n = ids[0]; n > cursor; n--)
10040 mdb_reader_check(MDB_env *env, int *dead)
10046 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10049 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
10051 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10053 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10054 unsigned int i, j, rdrs;
10056 MDB_PID_T *pids, pid;
10057 int rc = MDB_SUCCESS, count = 0;
10059 rdrs = env->me_txns->mti_numreaders;
10060 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10064 mr = env->me_txns->mti_readers;
10065 for (i=0; i<rdrs; i++) {
10066 pid = mr[i].mr_pid;
10067 if (pid && pid != env->me_pid) {
10068 if (mdb_pid_insert(pids, pid) == 0) {
10069 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10070 /* Stale reader found */
10073 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10074 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10076 rdrs = 0; /* the above checked all readers */
10078 /* Recheck, a new process may have reused pid */
10079 if (mdb_reader_pid(env, Pidcheck, pid))
10083 for (; j<rdrs; j++)
10084 if (mr[j].mr_pid == pid) {
10085 DPRINTF(("clear stale reader pid %u txn %"Z"d",
10086 (unsigned) pid, mr[j].mr_txnid));
10091 UNLOCK_MUTEX(rmutex);
10102 #ifdef MDB_ROBUST_SUPPORTED
10103 /** Handle #LOCK_MUTEX0() failure.
10104 * Try to repair the lock file if the mutex owner died.
10105 * @param[in] env the environment handle
10106 * @param[in] mutex LOCK_MUTEX0() mutex
10107 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10108 * @return 0 on success with the mutex locked, or an error code on failure.
10111 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10116 if (rc == MDB_OWNERDEAD) {
10117 /* We own the mutex. Clean up after dead previous owner. */
10119 rlocked = (mutex == env->me_rmutex);
10121 /* Keep mti_txnid updated, otherwise next writer can
10122 * overwrite data which latest meta page refers to.
10124 meta = mdb_env_pick_meta(env);
10125 env->me_txns->mti_txnid = meta->mm_txnid;
10126 /* env is hosed if the dead thread was ours */
10128 env->me_flags |= MDB_FATAL_ERROR;
10129 env->me_txn = NULL;
10133 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10134 (rc ? "this process' env is hosed" : "recovering")));
10135 rc2 = mdb_reader_check0(env, rlocked, NULL);
10137 rc2 = mdb_mutex_consistent(mutex);
10138 if (rc || (rc = rc2)) {
10139 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10140 UNLOCK_MUTEX(mutex);
10146 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10151 #endif /* MDB_ROBUST_SUPPORTED */
10154 #if defined(_WIN32)
10155 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
10159 need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
10160 if (need == 0xFFFD)
10164 result = malloc(sizeof(wchar_t) * need);
10165 MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
10171 #endif /* defined(_WIN32) */