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 2011-2016 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #if defined(MDB_VL32) || defined(__WIN64__)
39 #define _FILE_OFFSET_BITS 64
44 #include <wchar.h> /* get wcscpy() */
46 /* We use native NT APIs to setup the memory map, so that we can
47 * let the DB file grow incrementally instead of always preallocating
48 * the full size. These APIs are defined in <wdm.h> and <ntifs.h>
49 * but those headers are meant for driver-level development and
50 * conflict with the regular user-level headers, so we explicitly
51 * declare them here. Using these APIs also means we must link to
52 * ntdll.dll, which is not linked by default in user code.
55 NtCreateSection(OUT PHANDLE sh, IN ACCESS_MASK acc,
56 IN void * oa OPTIONAL,
57 IN PLARGE_INTEGER ms OPTIONAL,
58 IN ULONG pp, IN ULONG aa, IN HANDLE fh OPTIONAL);
60 typedef enum _SECTION_INHERIT {
66 NtMapViewOfSection(IN PHANDLE sh, IN HANDLE ph,
67 IN OUT PVOID *addr, IN ULONG_PTR zbits,
68 IN SIZE_T cs, IN OUT PLARGE_INTEGER off OPTIONAL,
69 IN OUT PSIZE_T vs, IN SECTION_INHERIT ih,
70 IN ULONG at, IN ULONG pp);
75 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
76 * as int64 which is wrong. MSVC doesn't define it at all, so just
80 #define MDB_THR_T DWORD
81 #include <sys/types.h>
84 # include <sys/param.h>
86 # define LITTLE_ENDIAN 1234
87 # define BIG_ENDIAN 4321
88 # define BYTE_ORDER LITTLE_ENDIAN
90 # define SSIZE_MAX INT_MAX
94 #include <sys/types.h>
96 #define MDB_PID_T pid_t
97 #define MDB_THR_T pthread_t
98 #include <sys/param.h>
100 #include <sys/mman.h>
101 #ifdef HAVE_SYS_FILE_H
102 #include <sys/file.h>
107 #if defined(__mips) && defined(__linux)
108 /* MIPS has cache coherency issues, requires explicit cache control */
109 #include <asm/cachectl.h>
110 extern int cacheflush(char *addr, int nbytes, int cache);
111 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
113 #define CACHEFLUSH(addr, bytes, cache)
116 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
117 /** fdatasync is broken on ext3/ext4fs on older kernels, see
118 * description in #mdb_env_open2 comments. You can safely
119 * define MDB_FDATASYNC_WORKS if this code will only be run
120 * on kernels 3.6 and newer.
122 #define BROKEN_FDATASYNC
128 #include <inttypes.h>
136 typedef SSIZE_T ssize_t;
141 #if defined(__sun) || defined(ANDROID)
142 /* Most platforms have posix_memalign, older may only have memalign */
143 #define HAVE_MEMALIGN 1
147 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
148 #include <netinet/in.h>
149 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
152 #if defined(__APPLE__) || defined (BSD)
153 # if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
154 # define MDB_USE_SYSV_SEM 1
156 # define MDB_FDATASYNC fsync
157 #elif defined(ANDROID)
158 # define MDB_FDATASYNC fsync
163 #ifdef MDB_USE_POSIX_SEM
164 # define MDB_USE_HASH 1
165 #include <semaphore.h>
166 #elif defined(MDB_USE_SYSV_SEM)
169 #ifdef _SEM_SEMUN_UNDEFINED
172 struct semid_ds *buf;
173 unsigned short *array;
175 #endif /* _SEM_SEMUN_UNDEFINED */
177 #define MDB_USE_POSIX_MUTEX 1
178 #endif /* MDB_USE_POSIX_SEM */
181 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
182 + defined(MDB_USE_POSIX_MUTEX) != 1
183 # error "Ambiguous shared-lock implementation"
187 #include <valgrind/memcheck.h>
188 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
189 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
190 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
191 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
192 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
194 #define VGMEMP_CREATE(h,r,z)
195 #define VGMEMP_ALLOC(h,a,s)
196 #define VGMEMP_FREE(h,a)
197 #define VGMEMP_DESTROY(h)
198 #define VGMEMP_DEFINED(a,s)
202 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
203 /* Solaris just defines one or the other */
204 # define LITTLE_ENDIAN 1234
205 # define BIG_ENDIAN 4321
206 # ifdef _LITTLE_ENDIAN
207 # define BYTE_ORDER LITTLE_ENDIAN
209 # define BYTE_ORDER BIG_ENDIAN
212 # define BYTE_ORDER __BYTE_ORDER
216 #ifndef LITTLE_ENDIAN
217 #define LITTLE_ENDIAN __LITTLE_ENDIAN
220 #define BIG_ENDIAN __BIG_ENDIAN
223 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
224 #define MISALIGNED_OK 1
230 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
231 # error "Unknown or unsupported endianness (BYTE_ORDER)"
232 #elif (-6 & 5) || CHAR_BIT!=8 || UINT_MAX!=0xffffffff || MDB_SIZE_MAX%UINT_MAX
233 # error "Two's complement, reasonably sized integer types, please"
237 /** Put infrequently used env functions in separate section */
239 # define ESECT __attribute__ ((section("__TEXT,text_env")))
241 # define ESECT __attribute__ ((section("text_env")))
248 #define CALL_CONV WINAPI
253 /** @defgroup internal LMDB Internals
256 /** @defgroup compat Compatibility Macros
257 * A bunch of macros to minimize the amount of platform-specific ifdefs
258 * needed throughout the rest of the code. When the features this library
259 * needs are similar enough to POSIX to be hidden in a one-or-two line
260 * replacement, this macro approach is used.
264 /** Features under development */
269 /** Wrapper around __func__, which is a C99 feature */
270 #if __STDC_VERSION__ >= 199901L
271 # define mdb_func_ __func__
272 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
273 # define mdb_func_ __FUNCTION__
275 /* If a debug message says <mdb_unknown>(), update the #if statements above */
276 # define mdb_func_ "<mdb_unknown>"
279 /* Internal error codes, not exposed outside liblmdb */
280 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
282 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
283 #elif defined MDB_USE_SYSV_SEM
284 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
285 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
286 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
290 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
292 /** Some platforms define the EOWNERDEAD error code
293 * even though they don't support Robust Mutexes.
294 * Compile with -DMDB_USE_ROBUST=0, or use some other
295 * mechanism like -DMDB_USE_SYSV_SEM instead of
296 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
297 * also Robust, but some systems don't support them
300 #ifndef MDB_USE_ROBUST
301 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
302 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
303 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
304 # define MDB_USE_ROBUST 0
306 # define MDB_USE_ROBUST 1
308 #endif /* !MDB_USE_ROBUST */
310 #if defined(MDB_USE_POSIX_MUTEX) && (MDB_USE_ROBUST)
311 /* glibc < 2.12 only provided _np API */
312 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
313 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
314 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
315 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
316 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
318 #endif /* MDB_USE_POSIX_MUTEX && MDB_USE_ROBUST */
320 #if defined(MDB_OWNERDEAD) && (MDB_USE_ROBUST)
321 #define MDB_ROBUST_SUPPORTED 1
325 #define MDB_USE_HASH 1
326 #define MDB_PIDLOCK 0
327 #define THREAD_RET DWORD
328 #define pthread_t HANDLE
329 #define pthread_mutex_t HANDLE
330 #define pthread_cond_t HANDLE
331 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
332 #define pthread_key_t DWORD
333 #define pthread_self() GetCurrentThreadId()
334 #define pthread_key_create(x,y) \
335 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
336 #define pthread_key_delete(x) TlsFree(x)
337 #define pthread_getspecific(x) TlsGetValue(x)
338 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
339 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
340 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
341 #define pthread_cond_signal(x) SetEvent(*x)
342 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
343 #define THREAD_CREATE(thr,start,arg) \
344 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
345 #define THREAD_FINISH(thr) \
346 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
347 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
348 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
349 #define mdb_mutex_consistent(mutex) 0
350 #define getpid() GetCurrentProcessId()
351 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
352 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
353 #define ErrCode() GetLastError()
354 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
355 #define close(fd) (CloseHandle(fd) ? 0 : -1)
356 #define munmap(ptr,len) UnmapViewOfFile(ptr)
357 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
358 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
360 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
363 #define THREAD_RET void *
364 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
365 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
367 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
368 #define MDB_PIDLOCK 1
370 #ifdef MDB_USE_POSIX_SEM
372 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
373 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
374 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
377 mdb_sem_wait(sem_t *sem)
380 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
384 #elif defined MDB_USE_SYSV_SEM
386 typedef struct mdb_mutex {
390 } mdb_mutex_t[1], *mdb_mutexref_t;
392 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
393 #define UNLOCK_MUTEX(mutex) do { \
394 struct sembuf sb = { 0, 1, SEM_UNDO }; \
395 sb.sem_num = (mutex)->semnum; \
396 *(mutex)->locked = 0; \
397 semop((mutex)->semid, &sb, 1); \
401 mdb_sem_wait(mdb_mutexref_t sem)
403 int rc, *locked = sem->locked;
404 struct sembuf sb = { 0, -1, SEM_UNDO };
405 sb.sem_num = sem->semnum;
407 if (!semop(sem->semid, &sb, 1)) {
408 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
412 } while ((rc = errno) == EINTR);
416 #define mdb_mutex_consistent(mutex) 0
418 #else /* MDB_USE_POSIX_MUTEX: */
419 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
420 * local variables keep it (mdb_mutexref_t).
422 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
423 * be the same, or an array[size 1] and a pointer.
426 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
428 /** Lock the reader or writer mutex.
429 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
431 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
432 /** Unlock the reader or writer mutex.
434 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
435 /** Mark mutex-protected data as repaired, after death of previous owner.
437 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
438 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
440 /** Get the error code for the last failed system function.
442 #define ErrCode() errno
444 /** An abstraction for a file handle.
445 * On POSIX systems file handles are small integers. On Windows
446 * they're opaque pointers.
450 /** A value for an invalid file handle.
451 * Mainly used to initialize file variables and signify that they are
454 #define INVALID_HANDLE_VALUE (-1)
456 /** Get the size of a memory page for the system.
457 * This is the basic size that the platform's memory manager uses, and is
458 * fundamental to the use of memory-mapped files.
460 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
463 #define Z MDB_FMT_Z /**< printf/scanf format modifier for size_t */
464 #define Yu MDB_PRIy(u) /**< printf format for #mdb_size_t */
465 #define Yd MDB_PRIy(d) /**< printf format for "signed #mdb_size_t" */
467 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
469 #elif defined(MDB_USE_SYSV_SEM)
470 #define MNAME_LEN (sizeof(int))
472 #define MNAME_LEN (sizeof(pthread_mutex_t))
475 #ifdef MDB_USE_SYSV_SEM
476 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
478 #define SYSV_SEM_FLAG 0
483 #ifdef MDB_ROBUST_SUPPORTED
484 /** Lock mutex, handle any error, set rc = result.
485 * Return 0 on success, nonzero (not rc) on error.
487 #define LOCK_MUTEX(rc, env, mutex) \
488 (((rc) = LOCK_MUTEX0(mutex)) && \
489 ((rc) = mdb_mutex_failed(env, mutex, rc)))
490 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
492 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
493 #define mdb_mutex_failed(env, mutex, rc) (rc)
497 /** A flag for opening a file and requesting synchronous data writes.
498 * This is only used when writing a meta page. It's not strictly needed;
499 * we could just do a normal write and then immediately perform a flush.
500 * But if this flag is available it saves us an extra system call.
502 * @note If O_DSYNC is undefined but exists in /usr/include,
503 * preferably set some compiler flag to get the definition.
507 # define MDB_DSYNC O_DSYNC
509 # define MDB_DSYNC O_SYNC
514 /** Function for flushing the data of a file. Define this to fsync
515 * if fdatasync() is not supported.
517 #ifndef MDB_FDATASYNC
518 # define MDB_FDATASYNC fdatasync
522 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
533 /** A page number in the database.
534 * Note that 64 bit page numbers are overkill, since pages themselves
535 * already represent 12-13 bits of addressable memory, and the OS will
536 * always limit applications to a maximum of 63 bits of address space.
538 * @note In the #MDB_node structure, we only store 48 bits of this value,
539 * which thus limits us to only 60 bits of addressable data.
541 typedef MDB_ID pgno_t;
543 /** A transaction ID.
544 * See struct MDB_txn.mt_txnid for details.
546 typedef MDB_ID txnid_t;
548 /** @defgroup debug Debug Macros
552 /** Enable debug output. Needs variable argument macros (a C99 feature).
553 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
554 * read from and written to the database (used for free space management).
560 static int mdb_debug;
561 static txnid_t mdb_debug_start;
563 /** Print a debug message with printf formatting.
564 * Requires double parenthesis around 2 or more args.
566 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
567 # define DPRINTF0(fmt, ...) \
568 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
570 # define DPRINTF(args) ((void) 0)
572 /** Print a debug string.
573 * The string is printed literally, with no format processing.
575 #define DPUTS(arg) DPRINTF(("%s", arg))
576 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
578 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
581 /** @brief The maximum size of a database page.
583 * It is 32k or 64k, since value-PAGEBASE must fit in
584 * #MDB_page.%mp_upper.
586 * LMDB will use database pages < OS pages if needed.
587 * That causes more I/O in write transactions: The OS must
588 * know (read) the whole page before writing a partial page.
590 * Note that we don't currently support Huge pages. On Linux,
591 * regular data files cannot use Huge pages, and in general
592 * Huge pages aren't actually pageable. We rely on the OS
593 * demand-pager to read our data and page it out when memory
594 * pressure from other processes is high. So until OSs have
595 * actual paging support for Huge pages, they're not viable.
597 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
599 /** The minimum number of keys required in a database page.
600 * Setting this to a larger value will place a smaller bound on the
601 * maximum size of a data item. Data items larger than this size will
602 * be pushed into overflow pages instead of being stored directly in
603 * the B-tree node. This value used to default to 4. With a page size
604 * of 4096 bytes that meant that any item larger than 1024 bytes would
605 * go into an overflow page. That also meant that on average 2-3KB of
606 * each overflow page was wasted space. The value cannot be lower than
607 * 2 because then there would no longer be a tree structure. With this
608 * value, items larger than 2KB will go into overflow pages, and on
609 * average only 1KB will be wasted.
611 #define MDB_MINKEYS 2
613 /** A stamp that identifies a file as an LMDB file.
614 * There's nothing special about this value other than that it is easily
615 * recognizable, and it will reflect any byte order mismatches.
617 #define MDB_MAGIC 0xBEEFC0DE
619 /** The version number for a database's datafile format. */
620 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
621 /** The version number for a database's lockfile format. */
622 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
624 /** @brief The max size of a key we can write, or 0 for computed max.
626 * This macro should normally be left alone or set to 0.
627 * Note that a database with big keys or dupsort data cannot be
628 * reliably modified by a liblmdb which uses a smaller max.
629 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
631 * Other values are allowed, for backwards compat. However:
632 * A value bigger than the computed max can break if you do not
633 * know what you are doing, and liblmdb <= 0.9.10 can break when
634 * modifying a DB with keys/dupsort data bigger than its max.
636 * Data items in an #MDB_DUPSORT database are also limited to
637 * this size, since they're actually keys of a sub-DB. Keys and
638 * #MDB_DUPSORT data items must fit on a node in a regular page.
640 #ifndef MDB_MAXKEYSIZE
641 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
644 /** The maximum size of a key we can write to the environment. */
646 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
648 #define ENV_MAXKEY(env) ((env)->me_maxkey)
651 /** @brief The maximum size of a data item.
653 * We only store a 32 bit value for node sizes.
655 #define MAXDATASIZE 0xffffffffUL
658 /** Key size which fits in a #DKBUF.
661 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
664 * This is used for printing a hex dump of a key's contents.
666 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
667 /** Display a key in hex.
669 * Invoke a function to display a key in hex.
671 #define DKEY(x) mdb_dkey(x, kbuf)
677 /** An invalid page number.
678 * Mainly used to denote an empty tree.
680 #define P_INVALID (~(pgno_t)0)
682 /** Test if the flags \b f are set in a flag word \b w. */
683 #define F_ISSET(w, f) (((w) & (f)) == (f))
685 /** Round \b n up to an even number. */
686 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
688 /** Used for offsets within a single page.
689 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
692 typedef uint16_t indx_t;
694 /** Default size of memory map.
695 * This is certainly too small for any actual applications. Apps should always set
696 * the size explicitly using #mdb_env_set_mapsize().
698 #define DEFAULT_MAPSIZE 1048576
700 /** @defgroup readers Reader Lock Table
701 * Readers don't acquire any locks for their data access. Instead, they
702 * simply record their transaction ID in the reader table. The reader
703 * mutex is needed just to find an empty slot in the reader table. The
704 * slot's address is saved in thread-specific data so that subsequent read
705 * transactions started by the same thread need no further locking to proceed.
707 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
709 * No reader table is used if the database is on a read-only filesystem, or
710 * if #MDB_NOLOCK is set.
712 * Since the database uses multi-version concurrency control, readers don't
713 * actually need any locking. This table is used to keep track of which
714 * readers are using data from which old transactions, so that we'll know
715 * when a particular old transaction is no longer in use. Old transactions
716 * that have discarded any data pages can then have those pages reclaimed
717 * for use by a later write transaction.
719 * The lock table is constructed such that reader slots are aligned with the
720 * processor's cache line size. Any slot is only ever used by one thread.
721 * This alignment guarantees that there will be no contention or cache
722 * thrashing as threads update their own slot info, and also eliminates
723 * any need for locking when accessing a slot.
725 * A writer thread will scan every slot in the table to determine the oldest
726 * outstanding reader transaction. Any freed pages older than this will be
727 * reclaimed by the writer. The writer doesn't use any locks when scanning
728 * this table. This means that there's no guarantee that the writer will
729 * see the most up-to-date reader info, but that's not required for correct
730 * operation - all we need is to know the upper bound on the oldest reader,
731 * we don't care at all about the newest reader. So the only consequence of
732 * reading stale information here is that old pages might hang around a
733 * while longer before being reclaimed. That's actually good anyway, because
734 * the longer we delay reclaiming old pages, the more likely it is that a
735 * string of contiguous pages can be found after coalescing old pages from
736 * many old transactions together.
739 /** Number of slots in the reader table.
740 * This value was chosen somewhat arbitrarily. 126 readers plus a
741 * couple mutexes fit exactly into 8KB on my development machine.
742 * Applications should set the table size using #mdb_env_set_maxreaders().
744 #define DEFAULT_READERS 126
746 /** The size of a CPU cache line in bytes. We want our lock structures
747 * aligned to this size to avoid false cache line sharing in the
749 * This value works for most CPUs. For Itanium this should be 128.
755 /** The information we store in a single slot of the reader table.
756 * In addition to a transaction ID, we also record the process and
757 * thread ID that owns a slot, so that we can detect stale information,
758 * e.g. threads or processes that went away without cleaning up.
759 * @note We currently don't check for stale records. We simply re-init
760 * the table when we know that we're the only process opening the
763 typedef struct MDB_rxbody {
764 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
765 * Multiple readers that start at the same time will probably have the
766 * same ID here. Again, it's not important to exclude them from
767 * anything; all we need to know is which version of the DB they
768 * started from so we can avoid overwriting any data used in that
769 * particular version.
771 volatile txnid_t mrb_txnid;
772 /** The process ID of the process owning this reader txn. */
773 volatile MDB_PID_T mrb_pid;
774 /** The thread ID of the thread owning this txn. */
775 volatile MDB_THR_T mrb_tid;
778 /** The actual reader record, with cacheline padding. */
779 typedef struct MDB_reader {
782 /** shorthand for mrb_txnid */
783 #define mr_txnid mru.mrx.mrb_txnid
784 #define mr_pid mru.mrx.mrb_pid
785 #define mr_tid mru.mrx.mrb_tid
786 /** cache line alignment */
787 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
791 /** The header for the reader table.
792 * The table resides in a memory-mapped file. (This is a different file
793 * than is used for the main database.)
795 * For POSIX the actual mutexes reside in the shared memory of this
796 * mapped file. On Windows, mutexes are named objects allocated by the
797 * kernel; we store the mutex names in this mapped file so that other
798 * processes can grab them. This same approach is also used on
799 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
800 * process-shared POSIX mutexes. For these cases where a named object
801 * is used, the object name is derived from a 64 bit FNV hash of the
802 * environment pathname. As such, naming collisions are extremely
803 * unlikely. If a collision occurs, the results are unpredictable.
805 typedef struct MDB_txbody {
806 /** Stamp identifying this as an LMDB file. It must be set
809 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
811 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
812 char mtb_rmname[MNAME_LEN];
813 #elif defined(MDB_USE_SYSV_SEM)
817 /** Mutex protecting access to this table.
818 * This is the reader table lock used with LOCK_MUTEX().
820 mdb_mutex_t mtb_rmutex;
822 /** The ID of the last transaction committed to the database.
823 * This is recorded here only for convenience; the value can always
824 * be determined by reading the main database meta pages.
826 volatile txnid_t mtb_txnid;
827 /** The number of slots that have been used in the reader table.
828 * This always records the maximum count, it is not decremented
829 * when readers release their slots.
831 volatile unsigned mtb_numreaders;
834 /** The actual reader table definition. */
835 typedef struct MDB_txninfo {
838 #define mti_magic mt1.mtb.mtb_magic
839 #define mti_format mt1.mtb.mtb_format
840 #define mti_rmutex mt1.mtb.mtb_rmutex
841 #define mti_rmname mt1.mtb.mtb_rmname
842 #define mti_txnid mt1.mtb.mtb_txnid
843 #define mti_numreaders mt1.mtb.mtb_numreaders
844 #ifdef MDB_USE_SYSV_SEM
845 #define mti_semid mt1.mtb.mtb_semid
846 #define mti_rlocked mt1.mtb.mtb_rlocked
848 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
851 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
852 char mt2_wmname[MNAME_LEN];
853 #define mti_wmname mt2.mt2_wmname
854 #elif defined MDB_USE_SYSV_SEM
856 #define mti_wlocked mt2.mt2_wlocked
858 mdb_mutex_t mt2_wmutex;
859 #define mti_wmutex mt2.mt2_wmutex
861 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
863 MDB_reader mti_readers[1];
866 /** Lockfile format signature: version, features and field layout */
867 #define MDB_LOCK_FORMAT \
869 ((MDB_LOCK_VERSION) \
870 /* Flags which describe functionality */ \
871 + (SYSV_SEM_FLAG << 18) \
872 + (((MDB_PIDLOCK) != 0) << 16)))
875 /** Common header for all page types. The page type depends on #mp_flags.
877 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with
878 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages
879 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header.
881 * #P_OVERFLOW records occupy one or more contiguous pages where only the
882 * first has a page header. They hold the real data of #F_BIGDATA nodes.
884 * #P_SUBP sub-pages are small leaf "pages" with duplicate data.
885 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page.
886 * (Duplicate data can also go in sub-databases, which use normal pages.)
888 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot.
890 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once
891 * in the snapshot: Either used by a database or listed in a freeDB record.
893 typedef struct MDB_page {
894 #define mp_pgno mp_p.p_pgno
895 #define mp_next mp_p.p_next
897 pgno_t p_pgno; /**< page number */
898 struct MDB_page *p_next; /**< for in-memory list of freed pages */
900 uint16_t mp_pad; /**< key size if this is a LEAF2 page */
901 /** @defgroup mdb_page Page Flags
903 * Flags for the page headers.
906 #define P_BRANCH 0x01 /**< branch page */
907 #define P_LEAF 0x02 /**< leaf page */
908 #define P_OVERFLOW 0x04 /**< overflow page */
909 #define P_META 0x08 /**< meta page */
910 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
911 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
912 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
913 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
914 #define P_KEEP 0x8000 /**< leave this page alone during spill */
916 uint16_t mp_flags; /**< @ref mdb_page */
917 #define mp_lower mp_pb.pb.pb_lower
918 #define mp_upper mp_pb.pb.pb_upper
919 #define mp_pages mp_pb.pb_pages
922 indx_t pb_lower; /**< lower bound of free space */
923 indx_t pb_upper; /**< upper bound of free space */
925 uint32_t pb_pages; /**< number of overflow pages */
927 indx_t mp_ptrs[1]; /**< dynamic size */
930 /** Size of the page header, excluding dynamic data at the end */
931 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
933 /** Address of first usable data byte in a page, after the header */
934 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
936 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
937 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
939 /** Number of nodes on a page */
940 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
942 /** The amount of space remaining in the page */
943 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
945 /** The percentage of space used in the page, in tenths of a percent. */
946 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
947 ((env)->me_psize - PAGEHDRSZ))
948 /** The minimum page fill factor, in tenths of a percent.
949 * Pages emptier than this are candidates for merging.
951 #define FILL_THRESHOLD 250
953 /** Test if a page is a leaf page */
954 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
955 /** Test if a page is a LEAF2 page */
956 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
957 /** Test if a page is a branch page */
958 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
959 /** Test if a page is an overflow page */
960 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
961 /** Test if a page is a sub page */
962 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
964 /** The number of overflow pages needed to store the given size. */
965 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
967 /** Link in #MDB_txn.%mt_loose_pgs list.
968 * Kept outside the page header, which is needed when reusing the page.
970 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
972 /** Header for a single key/data pair within a page.
973 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
974 * We guarantee 2-byte alignment for 'MDB_node's.
976 typedef struct MDB_node {
977 /** lo and hi are used for data size on leaf nodes and for
978 * child pgno on branch nodes. On 64 bit platforms, flags
979 * is also used for pgno. (Branch nodes have no flags).
980 * They are in host byte order in case that lets some
981 * accesses be optimized into a 32-bit word access.
983 #if BYTE_ORDER == LITTLE_ENDIAN
984 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
986 unsigned short mn_hi, mn_lo;
988 /** @defgroup mdb_node Node Flags
990 * Flags for node headers.
993 #define F_BIGDATA 0x01 /**< data put on overflow page */
994 #define F_SUBDATA 0x02 /**< data is a sub-database */
995 #define F_DUPDATA 0x04 /**< data has duplicates */
997 /** valid flags for #mdb_node_add() */
998 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
1001 unsigned short mn_flags; /**< @ref mdb_node */
1002 unsigned short mn_ksize; /**< key size */
1003 char mn_data[1]; /**< key and data are appended here */
1006 /** Size of the node header, excluding dynamic data at the end */
1007 #define NODESIZE offsetof(MDB_node, mn_data)
1009 /** Bit position of top word in page number, for shifting mn_flags */
1010 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
1012 /** Size of a node in a branch page with a given key.
1013 * This is just the node header plus the key, there is no data.
1015 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
1017 /** Size of a node in a leaf page with a given key and data.
1018 * This is node header plus key plus data size.
1020 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
1022 /** Address of node \b i in page \b p */
1023 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
1025 /** Address of the key for the node */
1026 #define NODEKEY(node) (void *)((node)->mn_data)
1028 /** Address of the data for a node */
1029 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
1031 /** Get the page number pointed to by a branch node */
1032 #define NODEPGNO(node) \
1033 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
1034 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
1035 /** Set the page number in a branch node */
1036 #define SETPGNO(node,pgno) do { \
1037 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
1038 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
1040 /** Get the size of the data in a leaf node */
1041 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
1042 /** Set the size of the data for a leaf node */
1043 #define SETDSZ(node,size) do { \
1044 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
1045 /** The size of a key in a node */
1046 #define NODEKSZ(node) ((node)->mn_ksize)
1048 /** Copy a page number from src to dst */
1049 #ifdef MISALIGNED_OK
1050 #define COPY_PGNO(dst,src) dst = src
1052 #if MDB_SIZE_MAX > 0xffffffffU
1053 #define COPY_PGNO(dst,src) do { \
1054 unsigned short *s, *d; \
1055 s = (unsigned short *)&(src); \
1056 d = (unsigned short *)&(dst); \
1063 #define COPY_PGNO(dst,src) do { \
1064 unsigned short *s, *d; \
1065 s = (unsigned short *)&(src); \
1066 d = (unsigned short *)&(dst); \
1072 /** The address of a key in a LEAF2 page.
1073 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
1074 * There are no node headers, keys are stored contiguously.
1076 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
1078 /** Set the \b node's key into \b keyptr, if requested. */
1079 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
1080 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
1082 /** Set the \b node's key into \b key. */
1083 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
1085 /** Information about a single database in the environment. */
1086 typedef struct MDB_db {
1087 uint32_t md_pad; /**< also ksize for LEAF2 pages */
1088 uint16_t md_flags; /**< @ref mdb_dbi_open */
1089 uint16_t md_depth; /**< depth of this tree */
1090 pgno_t md_branch_pages; /**< number of internal pages */
1091 pgno_t md_leaf_pages; /**< number of leaf pages */
1092 pgno_t md_overflow_pages; /**< number of overflow pages */
1093 mdb_size_t md_entries; /**< number of data items */
1094 pgno_t md_root; /**< the root page of this tree */
1097 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1098 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1099 /** #mdb_dbi_open() flags */
1100 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1101 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1103 /** Handle for the DB used to track free pages. */
1105 /** Handle for the default DB. */
1107 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1110 /** Number of meta pages - also hardcoded elsewhere */
1113 /** Meta page content.
1114 * A meta page is the start point for accessing a database snapshot.
1115 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1117 typedef struct MDB_meta {
1118 /** Stamp identifying this as an LMDB file. It must be set
1121 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1122 uint32_t mm_version;
1124 union { /* always zero since we don't support fixed mapping in MDB_VL32 */
1128 #define mm_address mm_un.mmun_address
1130 void *mm_address; /**< address for fixed mapping */
1132 pgno_t mm_mapsize; /**< size of mmap region */
1133 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1134 /** The size of pages used in this DB */
1135 #define mm_psize mm_dbs[FREE_DBI].md_pad
1136 /** Any persistent environment flags. @ref mdb_env */
1137 #define mm_flags mm_dbs[FREE_DBI].md_flags
1138 /** Last used page in the datafile.
1139 * Actually the file may be shorter if the freeDB lists the final pages.
1142 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1145 /** Buffer for a stack-allocated meta page.
1146 * The members define size and alignment, and silence type
1147 * aliasing warnings. They are not used directly; that could
1148 * mean incorrectly using several union members in parallel.
1150 typedef union MDB_metabuf {
1153 char mm_pad[PAGEHDRSZ];
1158 /** Auxiliary DB info.
1159 * The information here is mostly static/read-only. There is
1160 * only a single copy of this record in the environment.
1162 typedef struct MDB_dbx {
1163 MDB_val md_name; /**< name of the database */
1164 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1165 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1166 MDB_rel_func *md_rel; /**< user relocate function */
1167 void *md_relctx; /**< user-provided context for md_rel */
1170 /** A database transaction.
1171 * Every operation requires a transaction handle.
1174 MDB_txn *mt_parent; /**< parent of a nested txn */
1175 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1177 pgno_t mt_next_pgno; /**< next unallocated page */
1179 pgno_t mt_last_pgno; /**< last written page */
1181 /** The ID of this transaction. IDs are integers incrementing from 1.
1182 * Only committed write transactions increment the ID. If a transaction
1183 * aborts, the ID may be re-used by the next writer.
1186 MDB_env *mt_env; /**< the DB environment */
1187 /** The list of pages that became unused during this transaction.
1189 MDB_IDL mt_free_pgs;
1190 /** The list of loose pages that became unused and may be reused
1191 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1193 MDB_page *mt_loose_pgs;
1194 /** Number of loose pages (#mt_loose_pgs) */
1196 /** The sorted list of dirty pages we temporarily wrote to disk
1197 * because the dirty list was full. page numbers in here are
1198 * shifted left by 1, deleted slots have the LSB set.
1200 MDB_IDL mt_spill_pgs;
1202 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1203 MDB_ID2L dirty_list;
1204 /** For read txns: This thread/txn's reader table slot, or NULL. */
1207 /** Array of records for each DB known in the environment. */
1209 /** Array of MDB_db records for each known DB */
1211 /** Array of sequence numbers for each DB handle */
1212 unsigned int *mt_dbiseqs;
1213 /** @defgroup mt_dbflag Transaction DB Flags
1217 #define DB_DIRTY 0x01 /**< DB was written in this txn */
1218 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1219 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1220 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1221 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1222 #define DB_DUPDATA 0x20 /**< DB is #MDB_DUPSORT data */
1224 /** In write txns, array of cursors for each DB */
1225 MDB_cursor **mt_cursors;
1226 /** Array of flags for each DB */
1227 unsigned char *mt_dbflags;
1229 /** List of read-only pages (actually chunks) */
1231 /** We map chunks of 16 pages. Even though Windows uses 4KB pages, all
1232 * mappings must begin on 64KB boundaries. So we round off all pgnos to
1233 * a chunk boundary. We do the same on Linux for symmetry, and also to
1234 * reduce the frequency of mmap/munmap calls.
1236 #define MDB_RPAGE_CHUNK 16
1237 #define MDB_TRPAGE_SIZE 4096 /**< size of #mt_rpages array of chunks */
1238 #define MDB_TRPAGE_MAX (MDB_TRPAGE_SIZE-1) /**< maximum chunk index */
1239 unsigned int mt_rpcheck; /**< threshold for reclaiming unref'd chunks */
1241 /** Number of DB records in use, or 0 when the txn is finished.
1242 * This number only ever increments until the txn finishes; we
1243 * don't decrement it when individual DB handles are closed.
1247 /** @defgroup mdb_txn Transaction Flags
1251 /** #mdb_txn_begin() flags */
1252 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1253 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1254 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1255 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1256 /* internal txn flags */
1257 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1258 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1259 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1260 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1261 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1262 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1263 /** most operations on the txn are currently illegal */
1264 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1266 unsigned int mt_flags; /**< @ref mdb_txn */
1267 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1268 * Includes ancestor txns' dirty pages not hidden by other txns'
1269 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1270 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1272 unsigned int mt_dirty_room;
1275 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1276 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1277 * raise this on a 64 bit machine.
1279 #define CURSOR_STACK 32
1283 /** Cursors are used for all DB operations.
1284 * A cursor holds a path of (page pointer, key index) from the DB
1285 * root to a position in the DB, plus other state. #MDB_DUPSORT
1286 * cursors include an xcursor to the current data item. Write txns
1287 * track their cursors and keep them up to date when data moves.
1288 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1289 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1292 /** Next cursor on this DB in this txn */
1293 MDB_cursor *mc_next;
1294 /** Backup of the original cursor if this cursor is a shadow */
1295 MDB_cursor *mc_backup;
1296 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1297 struct MDB_xcursor *mc_xcursor;
1298 /** The transaction that owns this cursor */
1300 /** The database handle this cursor operates on */
1302 /** The database record for this cursor */
1304 /** The database auxiliary record for this cursor */
1306 /** The @ref mt_dbflag for this database */
1307 unsigned char *mc_dbflag;
1308 unsigned short mc_snum; /**< number of pushed pages */
1309 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1310 /** @defgroup mdb_cursor Cursor Flags
1312 * Cursor state flags.
1315 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1316 #define C_EOF 0x02 /**< No more data */
1317 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1318 #define C_DEL 0x08 /**< last op was a cursor_del */
1319 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1320 #define C_WRITEMAP MDB_TXN_WRITEMAP /**< Copy of txn flag */
1321 /** Read-only cursor into the txn's original snapshot in the map.
1322 * Set for read-only txns, and in #mdb_page_alloc() for #FREE_DBI when
1323 * #MDB_DEVEL & 2. Only implements code which is necessary for this.
1325 #define C_ORIG_RDONLY MDB_TXN_RDONLY
1327 unsigned int mc_flags; /**< @ref mdb_cursor */
1328 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1329 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1331 MDB_page *mc_ovpg; /**< a referenced overflow page */
1332 # define MC_OVPG(mc) ((mc)->mc_ovpg)
1333 # define MC_SET_OVPG(mc, pg) ((mc)->mc_ovpg = (pg))
1335 # define MC_OVPG(mc) ((MDB_page *)0)
1336 # define MC_SET_OVPG(mc, pg) ((void)0)
1340 /** Context for sorted-dup records.
1341 * We could have gone to a fully recursive design, with arbitrarily
1342 * deep nesting of sub-databases. But for now we only handle these
1343 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1345 typedef struct MDB_xcursor {
1346 /** A sub-cursor for traversing the Dup DB */
1347 MDB_cursor mx_cursor;
1348 /** The database record for this Dup DB */
1350 /** The auxiliary DB record for this Dup DB */
1352 /** The @ref mt_dbflag for this Dup DB */
1353 unsigned char mx_dbflag;
1356 /** Check if there is an inited xcursor, so #XCURSOR_REFRESH() is proper */
1357 #define XCURSOR_INITED(mc) \
1358 ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
1360 /** Update sub-page pointer, if any, in \b mc->mc_xcursor. Needed
1361 * when the node which contains the sub-page may have moved. Called
1362 * with \b mp = mc->mc_pg[mc->mc_top], \b ki = mc->mc_ki[mc->mc_top].
1364 #define XCURSOR_REFRESH(mc, mp, ki) do { \
1365 MDB_page *xr_pg = (mp); \
1366 MDB_node *xr_node = NODEPTR(xr_pg, ki); \
1367 if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \
1368 (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \
1371 /** State of FreeDB old pages, stored in the MDB_env */
1372 typedef struct MDB_pgstate {
1373 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1374 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1377 /** The database environment. */
1379 HANDLE me_fd; /**< The main data file */
1380 HANDLE me_lfd; /**< The lock file */
1381 HANDLE me_mfd; /**< For writing and syncing the meta pages */
1382 #if defined(MDB_VL32) && defined(_WIN32)
1383 HANDLE me_fmh; /**< File Mapping handle */
1385 /** Failed to update the meta page. Probably an I/O error. */
1386 #define MDB_FATAL_ERROR 0x80000000U
1387 /** Some fields are initialized. */
1388 #define MDB_ENV_ACTIVE 0x20000000U
1389 /** me_txkey is set */
1390 #define MDB_ENV_TXKEY 0x10000000U
1391 /** fdatasync is unreliable */
1392 #define MDB_FSYNCONLY 0x08000000U
1393 uint32_t me_flags; /**< @ref mdb_env */
1394 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1395 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1396 unsigned int me_maxreaders; /**< size of the reader table */
1397 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1398 volatile int me_close_readers;
1399 MDB_dbi me_numdbs; /**< number of DBs opened */
1400 MDB_dbi me_maxdbs; /**< size of the DB table */
1401 MDB_PID_T me_pid; /**< process ID of this env */
1402 char *me_path; /**< path to the DB files */
1403 char *me_map; /**< the memory map of the data file */
1404 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1405 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1406 void *me_pbuf; /**< scratch area for DUPSORT put() */
1407 MDB_txn *me_txn; /**< current write transaction */
1408 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1409 mdb_size_t me_mapsize; /**< size of the data memory map */
1410 off_t me_size; /**< current file size */
1411 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1412 MDB_dbx *me_dbxs; /**< array of static DB info */
1413 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1414 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1415 pthread_key_t me_txkey; /**< thread-key for readers */
1416 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1417 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1418 # define me_pglast me_pgstate.mf_pglast
1419 # define me_pghead me_pgstate.mf_pghead
1420 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1421 /** IDL of pages that became unused in a write txn */
1422 MDB_IDL me_free_pgs;
1423 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1424 MDB_ID2L me_dirty_list;
1425 /** Max number of freelist items that can fit in a single overflow page */
1427 /** Max size of a node on a page */
1428 unsigned int me_nodemax;
1429 #if !(MDB_MAXKEYSIZE)
1430 unsigned int me_maxkey; /**< max size of a key */
1432 int me_live_reader; /**< have liveness lock in reader table */
1434 int me_pidquery; /**< Used in OpenProcess */
1436 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1437 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1438 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1440 mdb_mutex_t me_rmutex;
1441 mdb_mutex_t me_wmutex;
1444 MDB_ID3L me_rpages; /**< like #mt_rpages, but global to env */
1445 pthread_mutex_t me_rpmutex; /**< control access to #me_rpages */
1446 #define MDB_ERPAGE_SIZE 16384
1447 #define MDB_ERPAGE_MAX (MDB_ERPAGE_SIZE-1)
1448 unsigned int me_rpcheck;
1450 void *me_userctx; /**< User-settable context */
1451 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1454 /** Nested transaction */
1455 typedef struct MDB_ntxn {
1456 MDB_txn mnt_txn; /**< the transaction */
1457 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1460 /** max number of pages to commit in one writev() call */
1461 #define MDB_COMMIT_PAGES 64
1462 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1463 #undef MDB_COMMIT_PAGES
1464 #define MDB_COMMIT_PAGES IOV_MAX
1467 /** max bytes to write in one call */
1468 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1470 /** Check \b txn and \b dbi arguments to a function */
1471 #define TXN_DBI_EXIST(txn, dbi, validity) \
1472 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1474 /** Check for misused \b dbi handles */
1475 #define TXN_DBI_CHANGED(txn, dbi) \
1476 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1478 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1479 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1480 static int mdb_page_touch(MDB_cursor *mc);
1482 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1483 "reset-tmp", "fail-begin", "fail-beginchild"}
1485 /* mdb_txn_end operation number, for logging */
1486 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1487 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1489 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1490 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1491 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1492 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1493 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1495 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1496 static int mdb_page_search_root(MDB_cursor *mc,
1497 MDB_val *key, int modify);
1498 #define MDB_PS_MODIFY 1
1499 #define MDB_PS_ROOTONLY 2
1500 #define MDB_PS_FIRST 4
1501 #define MDB_PS_LAST 8
1502 static int mdb_page_search(MDB_cursor *mc,
1503 MDB_val *key, int flags);
1504 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1506 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1507 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1508 pgno_t newpgno, unsigned int nflags);
1510 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1511 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1512 static int mdb_env_write_meta(MDB_txn *txn);
1513 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1514 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1516 static void mdb_env_close0(MDB_env *env, int excl);
1518 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1519 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1520 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1521 static void mdb_node_del(MDB_cursor *mc, int ksize);
1522 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1523 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1524 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1525 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1526 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1528 static int mdb_rebalance(MDB_cursor *mc);
1529 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1531 static void mdb_cursor_pop(MDB_cursor *mc);
1532 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1534 static int mdb_cursor_del0(MDB_cursor *mc);
1535 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1536 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1537 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1538 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1539 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1541 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1542 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1544 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1545 static void mdb_xcursor_init0(MDB_cursor *mc);
1546 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1547 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1549 static int mdb_drop0(MDB_cursor *mc, int subs);
1550 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1551 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1554 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1557 /** Compare two items pointing at '#mdb_size_t's of unknown alignment. */
1558 #ifdef MISALIGNED_OK
1559 # define mdb_cmp_clong mdb_cmp_long
1561 # define mdb_cmp_clong mdb_cmp_cint
1564 /** True if we need #mdb_cmp_clong() instead of \b cmp for #MDB_INTEGERDUP */
1565 #define NEED_CMP_CLONG(cmp, ksize) \
1566 (UINT_MAX < MDB_SIZE_MAX && \
1567 (cmp) == mdb_cmp_int && (ksize) == sizeof(mdb_size_t))
1570 static SECURITY_DESCRIPTOR mdb_null_sd;
1571 static SECURITY_ATTRIBUTES mdb_all_sa;
1572 static int mdb_sec_inited;
1575 static int utf8_to_utf16(const char *src, struct MDB_name *dst, int xtra);
1578 /** Return the library version info. */
1580 mdb_version(int *major, int *minor, int *patch)
1582 if (major) *major = MDB_VERSION_MAJOR;
1583 if (minor) *minor = MDB_VERSION_MINOR;
1584 if (patch) *patch = MDB_VERSION_PATCH;
1585 return MDB_VERSION_STRING;
1588 /** Table of descriptions for LMDB @ref errors */
1589 static char *const mdb_errstr[] = {
1590 "MDB_KEYEXIST: Key/data pair already exists",
1591 "MDB_NOTFOUND: No matching key/data pair found",
1592 "MDB_PAGE_NOTFOUND: Requested page not found",
1593 "MDB_CORRUPTED: Located page was wrong type",
1594 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1595 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1596 "MDB_INVALID: File is not an LMDB file",
1597 "MDB_MAP_FULL: Environment mapsize limit reached",
1598 "MDB_DBS_FULL: Environment maxdbs limit reached",
1599 "MDB_READERS_FULL: Environment maxreaders limit reached",
1600 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1601 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1602 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1603 "MDB_PAGE_FULL: Internal error - page has no more space",
1604 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1605 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1606 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1607 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1608 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1609 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1610 "MDB_PROBLEM: Unexpected problem - txn should abort",
1614 mdb_strerror(int err)
1617 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1618 * This works as long as no function between the call to mdb_strerror
1619 * and the actual use of the message uses more than 4K of stack.
1621 #define MSGSIZE 1024
1622 #define PADSIZE 4096
1623 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1627 return ("Successful return: 0");
1629 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1630 i = err - MDB_KEYEXIST;
1631 return mdb_errstr[i];
1635 /* These are the C-runtime error codes we use. The comment indicates
1636 * their numeric value, and the Win32 error they would correspond to
1637 * if the error actually came from a Win32 API. A major mess, we should
1638 * have used LMDB-specific error codes for everything.
1641 case ENOENT: /* 2, FILE_NOT_FOUND */
1642 case EIO: /* 5, ACCESS_DENIED */
1643 case ENOMEM: /* 12, INVALID_ACCESS */
1644 case EACCES: /* 13, INVALID_DATA */
1645 case EBUSY: /* 16, CURRENT_DIRECTORY */
1646 case EINVAL: /* 22, BAD_COMMAND */
1647 case ENOSPC: /* 28, OUT_OF_PAPER */
1648 return strerror(err);
1653 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1654 FORMAT_MESSAGE_IGNORE_INSERTS,
1655 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1658 return strerror(err);
1662 /** assert(3) variant in cursor context */
1663 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1664 /** assert(3) variant in transaction context */
1665 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1666 /** assert(3) variant in environment context */
1667 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1670 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1671 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1674 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1675 const char *func, const char *file, int line)
1678 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1679 file, line, expr_txt, func);
1680 if (env->me_assert_func)
1681 env->me_assert_func(env, buf);
1682 fprintf(stderr, "%s\n", buf);
1686 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1690 /** Return the page number of \b mp which may be sub-page, for debug output */
1692 mdb_dbg_pgno(MDB_page *mp)
1695 COPY_PGNO(ret, mp->mp_pgno);
1699 /** Display a key in hexadecimal and return the address of the result.
1700 * @param[in] key the key to display
1701 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1702 * @return The key in hexadecimal form.
1705 mdb_dkey(MDB_val *key, char *buf)
1708 unsigned char *c = key->mv_data;
1714 if (key->mv_size > DKBUF_MAXKEYSIZE)
1715 return "MDB_MAXKEYSIZE";
1716 /* may want to make this a dynamic check: if the key is mostly
1717 * printable characters, print it as-is instead of converting to hex.
1721 for (i=0; i<key->mv_size; i++)
1722 ptr += sprintf(ptr, "%02x", *c++);
1724 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1730 mdb_leafnode_type(MDB_node *n)
1732 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1733 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1734 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1737 /** Display all the keys in the page. */
1739 mdb_page_list(MDB_page *mp)
1741 pgno_t pgno = mdb_dbg_pgno(mp);
1742 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1744 unsigned int i, nkeys, nsize, total = 0;
1748 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1749 case P_BRANCH: type = "Branch page"; break;
1750 case P_LEAF: type = "Leaf page"; break;
1751 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1752 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1753 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1755 fprintf(stderr, "Overflow page %"Yu" pages %u%s\n",
1756 pgno, mp->mp_pages, state);
1759 fprintf(stderr, "Meta-page %"Yu" txnid %"Yu"\n",
1760 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1763 fprintf(stderr, "Bad page %"Yu" flags 0x%X\n", pgno, mp->mp_flags);
1767 nkeys = NUMKEYS(mp);
1768 fprintf(stderr, "%s %"Yu" numkeys %d%s\n", type, pgno, nkeys, state);
1770 for (i=0; i<nkeys; i++) {
1771 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1772 key.mv_size = nsize = mp->mp_pad;
1773 key.mv_data = LEAF2KEY(mp, i, nsize);
1775 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1778 node = NODEPTR(mp, i);
1779 key.mv_size = node->mn_ksize;
1780 key.mv_data = node->mn_data;
1781 nsize = NODESIZE + key.mv_size;
1782 if (IS_BRANCH(mp)) {
1783 fprintf(stderr, "key %d: page %"Yu", %s\n", i, NODEPGNO(node),
1787 if (F_ISSET(node->mn_flags, F_BIGDATA))
1788 nsize += sizeof(pgno_t);
1790 nsize += NODEDSZ(node);
1792 nsize += sizeof(indx_t);
1793 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1794 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1796 total = EVEN(total);
1798 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1799 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1803 mdb_cursor_chk(MDB_cursor *mc)
1809 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1810 for (i=0; i<mc->mc_top; i++) {
1812 node = NODEPTR(mp, mc->mc_ki[i]);
1813 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1816 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1818 if (XCURSOR_INITED(mc)) {
1819 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1820 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1821 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1829 /** Count all the pages in each DB and in the freelist
1830 * and make sure it matches the actual number of pages
1832 * All named DBs must be open for a correct count.
1834 static void mdb_audit(MDB_txn *txn)
1838 MDB_ID freecount, count;
1843 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1844 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1845 freecount += *(MDB_ID *)data.mv_data;
1846 mdb_tassert(txn, rc == MDB_NOTFOUND);
1849 for (i = 0; i<txn->mt_numdbs; i++) {
1851 if (!(txn->mt_dbflags[i] & DB_VALID))
1853 mdb_cursor_init(&mc, txn, i, &mx);
1854 if (txn->mt_dbs[i].md_root == P_INVALID)
1856 count += txn->mt_dbs[i].md_branch_pages +
1857 txn->mt_dbs[i].md_leaf_pages +
1858 txn->mt_dbs[i].md_overflow_pages;
1859 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1860 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1861 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1864 mp = mc.mc_pg[mc.mc_top];
1865 for (j=0; j<NUMKEYS(mp); j++) {
1866 MDB_node *leaf = NODEPTR(mp, j);
1867 if (leaf->mn_flags & F_SUBDATA) {
1869 memcpy(&db, NODEDATA(leaf), sizeof(db));
1870 count += db.md_branch_pages + db.md_leaf_pages +
1871 db.md_overflow_pages;
1875 mdb_tassert(txn, rc == MDB_NOTFOUND);
1878 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1879 fprintf(stderr, "audit: %"Yu" freecount: %"Yu" count: %"Yu" total: %"Yu" next_pgno: %"Yu"\n",
1880 txn->mt_txnid, freecount, count+NUM_METAS,
1881 freecount+count+NUM_METAS, txn->mt_next_pgno);
1887 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1889 return txn->mt_dbxs[dbi].md_cmp(a, b);
1893 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1895 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1896 if (NEED_CMP_CLONG(dcmp, a->mv_size))
1897 dcmp = mdb_cmp_clong;
1901 /** Allocate memory for a page.
1902 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1903 * Set #MDB_TXN_ERROR on failure.
1906 mdb_page_malloc(MDB_txn *txn, unsigned num)
1908 MDB_env *env = txn->mt_env;
1909 MDB_page *ret = env->me_dpages;
1910 size_t psize = env->me_psize, sz = psize, off;
1911 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1912 * For a single page alloc, we init everything after the page header.
1913 * For multi-page, we init the final page; if the caller needed that
1914 * many pages they will be filling in at least up to the last page.
1918 VGMEMP_ALLOC(env, ret, sz);
1919 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1920 env->me_dpages = ret->mp_next;
1923 psize -= off = PAGEHDRSZ;
1928 if ((ret = malloc(sz)) != NULL) {
1929 VGMEMP_ALLOC(env, ret, sz);
1930 if (!(env->me_flags & MDB_NOMEMINIT)) {
1931 memset((char *)ret + off, 0, psize);
1935 txn->mt_flags |= MDB_TXN_ERROR;
1939 /** Free a single page.
1940 * Saves single pages to a list, for future reuse.
1941 * (This is not used for multi-page overflow pages.)
1944 mdb_page_free(MDB_env *env, MDB_page *mp)
1946 mp->mp_next = env->me_dpages;
1947 VGMEMP_FREE(env, mp);
1948 env->me_dpages = mp;
1951 /** Free a dirty page */
1953 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1955 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1956 mdb_page_free(env, dp);
1958 /* large pages just get freed directly */
1959 VGMEMP_FREE(env, dp);
1964 /** Return all dirty pages to dpage list */
1966 mdb_dlist_free(MDB_txn *txn)
1968 MDB_env *env = txn->mt_env;
1969 MDB_ID2L dl = txn->mt_u.dirty_list;
1970 unsigned i, n = dl[0].mid;
1972 for (i = 1; i <= n; i++) {
1973 mdb_dpage_free(env, dl[i].mptr);
1980 mdb_page_unref(MDB_txn *txn, MDB_page *mp)
1983 MDB_ID3L tl = txn->mt_rpages;
1985 if (mp->mp_flags & (P_SUBP|P_DIRTY))
1987 rem = mp->mp_pgno & (MDB_RPAGE_CHUNK-1);
1988 pgno = mp->mp_pgno ^ rem;
1989 x = mdb_mid3l_search(tl, pgno);
1990 if (x != tl[0].mid && tl[x+1].mid == mp->mp_pgno)
1995 #define MDB_PAGE_UNREF(txn, mp) mdb_page_unref(txn, mp)
1998 mdb_cursor_unref(MDB_cursor *mc)
2001 if (!mc->mc_snum || !mc->mc_pg[0] || IS_SUBP(mc->mc_pg[0]))
2003 for (i=0; i<mc->mc_snum; i++)
2004 mdb_page_unref(mc->mc_txn, mc->mc_pg[i]);
2006 mdb_page_unref(mc->mc_txn, mc->mc_ovpg);
2009 mc->mc_snum = mc->mc_top = 0;
2010 mc->mc_pg[0] = NULL;
2011 mc->mc_flags &= ~C_INITIALIZED;
2013 #define MDB_CURSOR_UNREF(mc, force) \
2014 (((force) || ((mc)->mc_flags & C_INITIALIZED)) \
2015 ? mdb_cursor_unref(mc) \
2019 #define MDB_PAGE_UNREF(txn, mp)
2020 #define MDB_CURSOR_UNREF(mc, force) ((void)0)
2021 #endif /* MDB_VL32 */
2023 /** Loosen or free a single page.
2024 * Saves single pages to a list for future reuse
2025 * in this same txn. It has been pulled from the freeDB
2026 * and already resides on the dirty list, but has been
2027 * deleted. Use these pages first before pulling again
2030 * If the page wasn't dirtied in this txn, just add it
2031 * to this txn's free list.
2034 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
2037 pgno_t pgno = mp->mp_pgno;
2038 MDB_txn *txn = mc->mc_txn;
2040 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
2041 if (txn->mt_parent) {
2042 MDB_ID2 *dl = txn->mt_u.dirty_list;
2043 /* If txn has a parent, make sure the page is in our
2047 unsigned x = mdb_mid2l_search(dl, pgno);
2048 if (x <= dl[0].mid && dl[x].mid == pgno) {
2049 if (mp != dl[x].mptr) { /* bad cursor? */
2050 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2051 txn->mt_flags |= MDB_TXN_ERROR;
2059 /* no parent txn, so it's just ours */
2064 DPRINTF(("loosen db %d page %"Yu, DDBI(mc), mp->mp_pgno));
2065 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
2066 txn->mt_loose_pgs = mp;
2067 txn->mt_loose_count++;
2068 mp->mp_flags |= P_LOOSE;
2070 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
2078 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
2079 * @param[in] mc A cursor handle for the current operation.
2080 * @param[in] pflags Flags of the pages to update:
2081 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
2082 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
2083 * @return 0 on success, non-zero on failure.
2086 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
2088 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
2089 MDB_txn *txn = mc->mc_txn;
2090 MDB_cursor *m3, *m0 = mc;
2095 int rc = MDB_SUCCESS, level;
2097 /* Mark pages seen by cursors */
2098 if (mc->mc_flags & C_UNTRACK)
2099 mc = NULL; /* will find mc in mt_cursors */
2100 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
2101 for (; mc; mc=mc->mc_next) {
2102 if (!(mc->mc_flags & C_INITIALIZED))
2104 for (m3 = mc;; m3 = &mx->mx_cursor) {
2106 for (j=0; j<m3->mc_snum; j++) {
2108 if ((mp->mp_flags & Mask) == pflags)
2109 mp->mp_flags ^= P_KEEP;
2111 mx = m3->mc_xcursor;
2112 /* Proceed to mx if it is at a sub-database */
2113 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
2115 if (! (mp && (mp->mp_flags & P_LEAF)))
2117 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
2118 if (!(leaf->mn_flags & F_SUBDATA))
2127 /* Mark dirty root pages */
2128 for (i=0; i<txn->mt_numdbs; i++) {
2129 if (txn->mt_dbflags[i] & DB_DIRTY) {
2130 pgno_t pgno = txn->mt_dbs[i].md_root;
2131 if (pgno == P_INVALID)
2133 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
2135 if ((dp->mp_flags & Mask) == pflags && level <= 1)
2136 dp->mp_flags ^= P_KEEP;
2144 static int mdb_page_flush(MDB_txn *txn, int keep);
2146 /** Spill pages from the dirty list back to disk.
2147 * This is intended to prevent running into #MDB_TXN_FULL situations,
2148 * but note that they may still occur in a few cases:
2149 * 1) our estimate of the txn size could be too small. Currently this
2150 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
2151 * 2) child txns may run out of space if their parents dirtied a
2152 * lot of pages and never spilled them. TODO: we probably should do
2153 * a preemptive spill during #mdb_txn_begin() of a child txn, if
2154 * the parent's dirty_room is below a given threshold.
2156 * Otherwise, if not using nested txns, it is expected that apps will
2157 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
2158 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
2159 * If the txn never references them again, they can be left alone.
2160 * If the txn only reads them, they can be used without any fuss.
2161 * If the txn writes them again, they can be dirtied immediately without
2162 * going thru all of the work of #mdb_page_touch(). Such references are
2163 * handled by #mdb_page_unspill().
2165 * Also note, we never spill DB root pages, nor pages of active cursors,
2166 * because we'll need these back again soon anyway. And in nested txns,
2167 * we can't spill a page in a child txn if it was already spilled in a
2168 * parent txn. That would alter the parent txns' data even though
2169 * the child hasn't committed yet, and we'd have no way to undo it if
2170 * the child aborted.
2172 * @param[in] m0 cursor A cursor handle identifying the transaction and
2173 * database for which we are checking space.
2174 * @param[in] key For a put operation, the key being stored.
2175 * @param[in] data For a put operation, the data being stored.
2176 * @return 0 on success, non-zero on failure.
2179 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
2181 MDB_txn *txn = m0->mc_txn;
2183 MDB_ID2L dl = txn->mt_u.dirty_list;
2184 unsigned int i, j, need;
2187 if (m0->mc_flags & C_SUB)
2190 /* Estimate how much space this op will take */
2191 i = m0->mc_db->md_depth;
2192 /* Named DBs also dirty the main DB */
2193 if (m0->mc_dbi >= CORE_DBS)
2194 i += txn->mt_dbs[MAIN_DBI].md_depth;
2195 /* For puts, roughly factor in the key+data size */
2197 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2198 i += i; /* double it for good measure */
2201 if (txn->mt_dirty_room > i)
2204 if (!txn->mt_spill_pgs) {
2205 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2206 if (!txn->mt_spill_pgs)
2209 /* purge deleted slots */
2210 MDB_IDL sl = txn->mt_spill_pgs;
2211 unsigned int num = sl[0];
2213 for (i=1; i<=num; i++) {
2220 /* Preserve pages which may soon be dirtied again */
2221 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2224 /* Less aggressive spill - we originally spilled the entire dirty list,
2225 * with a few exceptions for cursor pages and DB root pages. But this
2226 * turns out to be a lot of wasted effort because in a large txn many
2227 * of those pages will need to be used again. So now we spill only 1/8th
2228 * of the dirty pages. Testing revealed this to be a good tradeoff,
2229 * better than 1/2, 1/4, or 1/10.
2231 if (need < MDB_IDL_UM_MAX / 8)
2232 need = MDB_IDL_UM_MAX / 8;
2234 /* Save the page IDs of all the pages we're flushing */
2235 /* flush from the tail forward, this saves a lot of shifting later on. */
2236 for (i=dl[0].mid; i && need; i--) {
2237 MDB_ID pn = dl[i].mid << 1;
2239 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2241 /* Can't spill twice, make sure it's not already in a parent's
2244 if (txn->mt_parent) {
2246 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2247 if (tx2->mt_spill_pgs) {
2248 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2249 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2250 dp->mp_flags |= P_KEEP;
2258 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2262 mdb_midl_sort(txn->mt_spill_pgs);
2264 /* Flush the spilled part of dirty list */
2265 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2268 /* Reset any dirty pages we kept that page_flush didn't see */
2269 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2272 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2276 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2278 mdb_find_oldest(MDB_txn *txn)
2281 txnid_t mr, oldest = txn->mt_txnid - 1;
2282 if (txn->mt_env->me_txns) {
2283 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2284 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2295 /** Add a page to the txn's dirty list */
2297 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2300 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2302 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2303 insert = mdb_mid2l_append;
2305 insert = mdb_mid2l_insert;
2307 mid.mid = mp->mp_pgno;
2309 rc = insert(txn->mt_u.dirty_list, &mid);
2310 mdb_tassert(txn, rc == 0);
2311 txn->mt_dirty_room--;
2314 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2315 * me_pghead and mt_next_pgno. Set #MDB_TXN_ERROR on failure.
2317 * If there are free pages available from older transactions, they
2318 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2319 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2320 * and move me_pglast to say which records were consumed. Only this
2321 * function can create me_pghead and move me_pglast/mt_next_pgno.
2322 * When #MDB_DEVEL & 2, it is not affected by #mdb_freelist_save(): it
2323 * then uses the transaction's original snapshot of the freeDB.
2324 * @param[in] mc cursor A cursor handle identifying the transaction and
2325 * database for which we are allocating.
2326 * @param[in] num the number of pages to allocate.
2327 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2328 * will always be satisfied by a single contiguous chunk of memory.
2329 * @return 0 on success, non-zero on failure.
2332 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2334 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2335 /* Get at most <Max_retries> more freeDB records once me_pghead
2336 * has enough pages. If not enough, use new pages from the map.
2337 * If <Paranoid> and mc is updating the freeDB, only get new
2338 * records if me_pghead is empty. Then the freelist cannot play
2339 * catch-up with itself by growing while trying to save it.
2341 enum { Paranoid = 1, Max_retries = 500 };
2343 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2345 int rc, retry = num * 60;
2346 MDB_txn *txn = mc->mc_txn;
2347 MDB_env *env = txn->mt_env;
2348 pgno_t pgno, *mop = env->me_pghead;
2349 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2351 txnid_t oldest = 0, last;
2356 /* If there are any loose pages, just use them */
2357 if (num == 1 && txn->mt_loose_pgs) {
2358 np = txn->mt_loose_pgs;
2359 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2360 txn->mt_loose_count--;
2361 DPRINTF(("db %d use loose page %"Yu, DDBI(mc), np->mp_pgno));
2368 /* If our dirty list is already full, we can't do anything */
2369 if (txn->mt_dirty_room == 0) {
2374 for (op = MDB_FIRST;; op = MDB_NEXT) {
2379 /* Seek a big enough contiguous page range. Prefer
2380 * pages at the tail, just truncating the list.
2386 if (mop[i-n2] == pgno+n2)
2393 if (op == MDB_FIRST) { /* 1st iteration */
2394 /* Prepare to fetch more and coalesce */
2395 last = env->me_pglast;
2396 oldest = env->me_pgoldest;
2397 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2398 #if (MDB_DEVEL) & 2 /* "& 2" so MDB_DEVEL=1 won't hide bugs breaking freeDB */
2399 /* Use original snapshot. TODO: Should need less care in code
2400 * which modifies the database. Maybe we can delete some code?
2402 m2.mc_flags |= C_ORIG_RDONLY;
2403 m2.mc_db = &env->me_metas[(txn->mt_txnid-1) & 1]->mm_dbs[FREE_DBI];
2404 m2.mc_dbflag = (unsigned char *)""; /* probably unnecessary */
2408 key.mv_data = &last; /* will look up last+1 */
2409 key.mv_size = sizeof(last);
2411 if (Paranoid && mc->mc_dbi == FREE_DBI)
2414 if (Paranoid && retry < 0 && mop_len)
2418 /* Do not fetch more if the record will be too recent */
2419 if (oldest <= last) {
2421 oldest = mdb_find_oldest(txn);
2422 env->me_pgoldest = oldest;
2428 rc = mdb_cursor_get(&m2, &key, NULL, op);
2430 if (rc == MDB_NOTFOUND)
2434 last = *(txnid_t*)key.mv_data;
2435 if (oldest <= last) {
2437 oldest = mdb_find_oldest(txn);
2438 env->me_pgoldest = oldest;
2444 np = m2.mc_pg[m2.mc_top];
2445 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2446 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2449 idl = (MDB_ID *) data.mv_data;
2452 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2457 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2459 mop = env->me_pghead;
2461 env->me_pglast = last;
2463 DPRINTF(("IDL read txn %"Yu" root %"Yu" num %u",
2464 last, txn->mt_dbs[FREE_DBI].md_root, i));
2466 DPRINTF(("IDL %"Yu, idl[j]));
2468 /* Merge in descending sorted order */
2469 mdb_midl_xmerge(mop, idl);
2473 /* Use new pages from the map when nothing suitable in the freeDB */
2475 pgno = txn->mt_next_pgno;
2476 if (pgno + num >= env->me_maxpg) {
2477 DPUTS("DB size maxed out");
2481 #if defined(_WIN32) && !defined(MDB_VL32)
2482 if (!(env->me_flags & MDB_RDONLY)) {
2484 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
2485 p = VirtualAlloc(p, env->me_psize * num, MEM_COMMIT,
2486 (env->me_flags & MDB_WRITEMAP) ? PAGE_READWRITE:
2489 DPUTS("VirtualAlloc failed");
2497 if (env->me_flags & MDB_WRITEMAP) {
2498 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2500 if (!(np = mdb_page_malloc(txn, num))) {
2506 mop[0] = mop_len -= num;
2507 /* Move any stragglers down */
2508 for (j = i-num; j < mop_len; )
2509 mop[++j] = mop[++i];
2511 txn->mt_next_pgno = pgno + num;
2514 mdb_page_dirty(txn, np);
2520 txn->mt_flags |= MDB_TXN_ERROR;
2524 /** Copy the used portions of a non-overflow page.
2525 * @param[in] dst page to copy into
2526 * @param[in] src page to copy from
2527 * @param[in] psize size of a page
2530 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2532 enum { Align = sizeof(pgno_t) };
2533 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2535 /* If page isn't full, just copy the used portion. Adjust
2536 * alignment so memcpy may copy words instead of bytes.
2538 if ((unused &= -Align) && !IS_LEAF2(src)) {
2539 upper = (upper + PAGEBASE) & -Align;
2540 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2541 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2544 memcpy(dst, src, psize - unused);
2548 /** Pull a page off the txn's spill list, if present.
2549 * If a page being referenced was spilled to disk in this txn, bring
2550 * it back and make it dirty/writable again.
2551 * @param[in] txn the transaction handle.
2552 * @param[in] mp the page being referenced. It must not be dirty.
2553 * @param[out] ret the writable page, if any. ret is unchanged if
2554 * mp wasn't spilled.
2557 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2559 MDB_env *env = txn->mt_env;
2562 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2564 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2565 if (!tx2->mt_spill_pgs)
2567 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2568 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2571 if (txn->mt_dirty_room == 0)
2572 return MDB_TXN_FULL;
2573 if (IS_OVERFLOW(mp))
2577 if (env->me_flags & MDB_WRITEMAP) {
2580 np = mdb_page_malloc(txn, num);
2584 memcpy(np, mp, num * env->me_psize);
2586 mdb_page_copy(np, mp, env->me_psize);
2589 /* If in current txn, this page is no longer spilled.
2590 * If it happens to be the last page, truncate the spill list.
2591 * Otherwise mark it as deleted by setting the LSB.
2593 if (x == txn->mt_spill_pgs[0])
2594 txn->mt_spill_pgs[0]--;
2596 txn->mt_spill_pgs[x] |= 1;
2597 } /* otherwise, if belonging to a parent txn, the
2598 * page remains spilled until child commits
2601 mdb_page_dirty(txn, np);
2602 np->mp_flags |= P_DIRTY;
2610 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2611 * Set #MDB_TXN_ERROR on failure.
2612 * @param[in] mc cursor pointing to the page to be touched
2613 * @return 0 on success, non-zero on failure.
2616 mdb_page_touch(MDB_cursor *mc)
2618 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2619 MDB_txn *txn = mc->mc_txn;
2620 MDB_cursor *m2, *m3;
2624 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2625 if (txn->mt_flags & MDB_TXN_SPILLS) {
2627 rc = mdb_page_unspill(txn, mp, &np);
2633 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2634 (rc = mdb_page_alloc(mc, 1, &np)))
2637 DPRINTF(("touched db %d page %"Yu" -> %"Yu, DDBI(mc),
2638 mp->mp_pgno, pgno));
2639 mdb_cassert(mc, mp->mp_pgno != pgno);
2640 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2641 /* Update the parent page, if any, to point to the new page */
2643 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2644 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2645 SETPGNO(node, pgno);
2647 mc->mc_db->md_root = pgno;
2649 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2650 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2652 /* If txn has a parent, make sure the page is in our
2656 unsigned x = mdb_mid2l_search(dl, pgno);
2657 if (x <= dl[0].mid && dl[x].mid == pgno) {
2658 if (mp != dl[x].mptr) { /* bad cursor? */
2659 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2660 txn->mt_flags |= MDB_TXN_ERROR;
2666 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2668 np = mdb_page_malloc(txn, 1);
2673 rc = mdb_mid2l_insert(dl, &mid);
2674 mdb_cassert(mc, rc == 0);
2679 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2681 np->mp_flags |= P_DIRTY;
2684 /* Adjust cursors pointing to mp */
2685 mc->mc_pg[mc->mc_top] = np;
2686 m2 = txn->mt_cursors[mc->mc_dbi];
2687 if (mc->mc_flags & C_SUB) {
2688 for (; m2; m2=m2->mc_next) {
2689 m3 = &m2->mc_xcursor->mx_cursor;
2690 if (m3->mc_snum < mc->mc_snum) continue;
2691 if (m3->mc_pg[mc->mc_top] == mp)
2692 m3->mc_pg[mc->mc_top] = np;
2695 for (; m2; m2=m2->mc_next) {
2696 if (m2->mc_snum < mc->mc_snum) continue;
2697 if (m2 == mc) continue;
2698 if (m2->mc_pg[mc->mc_top] == mp) {
2699 m2->mc_pg[mc->mc_top] = np;
2700 if (XCURSOR_INITED(m2) && IS_LEAF(np))
2701 XCURSOR_REFRESH(m2, np, m2->mc_ki[mc->mc_top]);
2705 MDB_PAGE_UNREF(mc->mc_txn, mp);
2709 txn->mt_flags |= MDB_TXN_ERROR;
2714 mdb_env_sync0(MDB_env *env, int force, pgno_t numpgs)
2717 if (env->me_flags & MDB_RDONLY)
2719 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2720 if (env->me_flags & MDB_WRITEMAP) {
2721 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2722 ? MS_ASYNC : MS_SYNC;
2723 if (MDB_MSYNC(env->me_map, env->me_psize * numpgs, flags))
2726 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2730 #ifdef BROKEN_FDATASYNC
2731 if (env->me_flags & MDB_FSYNCONLY) {
2732 if (fsync(env->me_fd))
2736 if (MDB_FDATASYNC(env->me_fd))
2744 mdb_env_sync(MDB_env *env, int force)
2746 MDB_meta *m = mdb_env_pick_meta(env);
2747 return mdb_env_sync0(env, force, m->mm_last_pg+1);
2750 /** Back up parent txn's cursors, then grab the originals for tracking */
2752 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2754 MDB_cursor *mc, *bk;
2759 for (i = src->mt_numdbs; --i >= 0; ) {
2760 if ((mc = src->mt_cursors[i]) != NULL) {
2761 size = sizeof(MDB_cursor);
2763 size += sizeof(MDB_xcursor);
2764 for (; mc; mc = bk->mc_next) {
2770 mc->mc_db = &dst->mt_dbs[i];
2771 /* Kill pointers into src to reduce abuse: The
2772 * user may not use mc until dst ends. But we need a valid
2773 * txn pointer here for cursor fixups to keep working.
2776 mc->mc_dbflag = &dst->mt_dbflags[i];
2777 if ((mx = mc->mc_xcursor) != NULL) {
2778 *(MDB_xcursor *)(bk+1) = *mx;
2779 mx->mx_cursor.mc_txn = dst;
2781 mc->mc_next = dst->mt_cursors[i];
2782 dst->mt_cursors[i] = mc;
2789 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2790 * @param[in] txn the transaction handle.
2791 * @param[in] merge true to keep changes to parent cursors, false to revert.
2792 * @return 0 on success, non-zero on failure.
2795 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2797 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2801 for (i = txn->mt_numdbs; --i >= 0; ) {
2802 for (mc = cursors[i]; mc; mc = next) {
2804 if ((bk = mc->mc_backup) != NULL) {
2806 /* Commit changes to parent txn */
2807 mc->mc_next = bk->mc_next;
2808 mc->mc_backup = bk->mc_backup;
2809 mc->mc_txn = bk->mc_txn;
2810 mc->mc_db = bk->mc_db;
2811 mc->mc_dbflag = bk->mc_dbflag;
2812 if ((mx = mc->mc_xcursor) != NULL)
2813 mx->mx_cursor.mc_txn = bk->mc_txn;
2815 /* Abort nested txn */
2817 if ((mx = mc->mc_xcursor) != NULL)
2818 *mx = *(MDB_xcursor *)(bk+1);
2822 /* Only malloced cursors are permanently tracked. */
2829 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2835 Pidset = F_SETLK, Pidcheck = F_GETLK
2839 /** Set or check a pid lock. Set returns 0 on success.
2840 * Check returns 0 if the process is certainly dead, nonzero if it may
2841 * be alive (the lock exists or an error happened so we do not know).
2843 * On Windows Pidset is a no-op, we merely check for the existence
2844 * of the process with the given pid. On POSIX we use a single byte
2845 * lock on the lockfile, set at an offset equal to the pid.
2848 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2850 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2853 if (op == Pidcheck) {
2854 h = OpenProcess(env->me_pidquery, FALSE, pid);
2855 /* No documented "no such process" code, but other program use this: */
2857 return ErrCode() != ERROR_INVALID_PARAMETER;
2858 /* A process exists until all handles to it close. Has it exited? */
2859 ret = WaitForSingleObject(h, 0) != 0;
2866 struct flock lock_info;
2867 memset(&lock_info, 0, sizeof(lock_info));
2868 lock_info.l_type = F_WRLCK;
2869 lock_info.l_whence = SEEK_SET;
2870 lock_info.l_start = pid;
2871 lock_info.l_len = 1;
2872 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2873 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2875 } else if ((rc = ErrCode()) == EINTR) {
2883 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2884 * @param[in] txn the transaction handle to initialize
2885 * @return 0 on success, non-zero on failure.
2888 mdb_txn_renew0(MDB_txn *txn)
2890 MDB_env *env = txn->mt_env;
2891 MDB_txninfo *ti = env->me_txns;
2893 unsigned int i, nr, flags = txn->mt_flags;
2895 int rc, new_notls = 0;
2897 if ((flags &= MDB_TXN_RDONLY) != 0) {
2899 meta = mdb_env_pick_meta(env);
2900 txn->mt_txnid = meta->mm_txnid;
2901 txn->mt_u.reader = NULL;
2903 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2904 pthread_getspecific(env->me_txkey);
2906 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2907 return MDB_BAD_RSLOT;
2909 MDB_PID_T pid = env->me_pid;
2910 MDB_THR_T tid = pthread_self();
2911 mdb_mutexref_t rmutex = env->me_rmutex;
2913 if (!env->me_live_reader) {
2914 rc = mdb_reader_pid(env, Pidset, pid);
2917 env->me_live_reader = 1;
2920 if (LOCK_MUTEX(rc, env, rmutex))
2922 nr = ti->mti_numreaders;
2923 for (i=0; i<nr; i++)
2924 if (ti->mti_readers[i].mr_pid == 0)
2926 if (i == env->me_maxreaders) {
2927 UNLOCK_MUTEX(rmutex);
2928 return MDB_READERS_FULL;
2930 r = &ti->mti_readers[i];
2931 /* Claim the reader slot, carefully since other code
2932 * uses the reader table un-mutexed: First reset the
2933 * slot, next publish it in mti_numreaders. After
2934 * that, it is safe for mdb_env_close() to touch it.
2935 * When it will be closed, we can finally claim it.
2938 r->mr_txnid = (txnid_t)-1;
2941 ti->mti_numreaders = ++nr;
2942 env->me_close_readers = nr;
2944 UNLOCK_MUTEX(rmutex);
2946 new_notls = (env->me_flags & MDB_NOTLS);
2947 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2952 do /* LY: Retry on a race, ITS#7970. */
2953 r->mr_txnid = ti->mti_txnid;
2954 while(r->mr_txnid != ti->mti_txnid);
2955 txn->mt_txnid = r->mr_txnid;
2956 txn->mt_u.reader = r;
2957 meta = env->me_metas[txn->mt_txnid & 1];
2961 /* Not yet touching txn == env->me_txn0, it may be active */
2963 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2965 txn->mt_txnid = ti->mti_txnid;
2966 meta = env->me_metas[txn->mt_txnid & 1];
2968 meta = mdb_env_pick_meta(env);
2969 txn->mt_txnid = meta->mm_txnid;
2973 if (txn->mt_txnid == mdb_debug_start)
2976 txn->mt_child = NULL;
2977 txn->mt_loose_pgs = NULL;
2978 txn->mt_loose_count = 0;
2979 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2980 txn->mt_u.dirty_list = env->me_dirty_list;
2981 txn->mt_u.dirty_list[0].mid = 0;
2982 txn->mt_free_pgs = env->me_free_pgs;
2983 txn->mt_free_pgs[0] = 0;
2984 txn->mt_spill_pgs = NULL;
2986 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2989 /* Copy the DB info and flags */
2990 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2992 /* Moved to here to avoid a data race in read TXNs */
2993 txn->mt_next_pgno = meta->mm_last_pg+1;
2995 txn->mt_last_pgno = txn->mt_next_pgno - 1;
2998 txn->mt_flags = flags;
3001 txn->mt_numdbs = env->me_numdbs;
3002 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3003 x = env->me_dbflags[i];
3004 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
3005 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
3007 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
3008 txn->mt_dbflags[FREE_DBI] = DB_VALID;
3010 if (env->me_flags & MDB_FATAL_ERROR) {
3011 DPUTS("environment had fatal error, must shutdown!");
3013 } else if (env->me_maxpg < txn->mt_next_pgno) {
3014 rc = MDB_MAP_RESIZED;
3018 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
3023 mdb_txn_renew(MDB_txn *txn)
3027 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
3030 rc = mdb_txn_renew0(txn);
3031 if (rc == MDB_SUCCESS) {
3032 DPRINTF(("renew txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3033 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3034 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
3040 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
3044 int rc, size, tsize;
3046 flags &= MDB_TXN_BEGIN_FLAGS;
3047 flags |= env->me_flags & MDB_WRITEMAP;
3049 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
3053 /* Nested transactions: Max 1 child, write txns only, no writemap */
3054 flags |= parent->mt_flags;
3055 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
3056 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
3058 /* Child txns save MDB_pgstate and use own copy of cursors */
3059 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
3060 size += tsize = sizeof(MDB_ntxn);
3061 } else if (flags & MDB_RDONLY) {
3062 size = env->me_maxdbs * (sizeof(MDB_db)+1);
3063 size += tsize = sizeof(MDB_txn);
3065 /* Reuse preallocated write txn. However, do not touch it until
3066 * mdb_txn_renew0() succeeds, since it currently may be active.
3071 if ((txn = calloc(1, size)) == NULL) {
3072 DPRINTF(("calloc: %s", strerror(errno)));
3077 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
3078 if (!txn->mt_rpages) {
3082 txn->mt_rpages[0].mid = 0;
3083 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
3086 txn->mt_dbxs = env->me_dbxs; /* static */
3087 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
3088 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
3089 txn->mt_flags = flags;
3094 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
3095 txn->mt_dbiseqs = parent->mt_dbiseqs;
3096 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
3097 if (!txn->mt_u.dirty_list ||
3098 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
3100 free(txn->mt_u.dirty_list);
3104 txn->mt_txnid = parent->mt_txnid;
3105 txn->mt_dirty_room = parent->mt_dirty_room;
3106 txn->mt_u.dirty_list[0].mid = 0;
3107 txn->mt_spill_pgs = NULL;
3108 txn->mt_next_pgno = parent->mt_next_pgno;
3109 parent->mt_flags |= MDB_TXN_HAS_CHILD;
3110 parent->mt_child = txn;
3111 txn->mt_parent = parent;
3112 txn->mt_numdbs = parent->mt_numdbs;
3114 txn->mt_rpages = parent->mt_rpages;
3116 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3117 /* Copy parent's mt_dbflags, but clear DB_NEW */
3118 for (i=0; i<txn->mt_numdbs; i++)
3119 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
3121 ntxn = (MDB_ntxn *)txn;
3122 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
3123 if (env->me_pghead) {
3124 size = MDB_IDL_SIZEOF(env->me_pghead);
3125 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
3127 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
3132 rc = mdb_cursor_shadow(parent, txn);
3134 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
3135 } else { /* MDB_RDONLY */
3136 txn->mt_dbiseqs = env->me_dbiseqs;
3138 rc = mdb_txn_renew0(txn);
3141 if (txn != env->me_txn0) {
3143 free(txn->mt_rpages);
3148 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
3150 DPRINTF(("begin txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3151 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
3152 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
3159 mdb_txn_env(MDB_txn *txn)
3161 if(!txn) return NULL;
3166 mdb_txn_id(MDB_txn *txn)
3169 return txn->mt_txnid;
3172 /** Export or close DBI handles opened in this txn. */
3174 mdb_dbis_update(MDB_txn *txn, int keep)
3177 MDB_dbi n = txn->mt_numdbs;
3178 MDB_env *env = txn->mt_env;
3179 unsigned char *tdbflags = txn->mt_dbflags;
3181 for (i = n; --i >= CORE_DBS;) {
3182 if (tdbflags[i] & DB_NEW) {
3184 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
3186 char *ptr = env->me_dbxs[i].md_name.mv_data;
3188 env->me_dbxs[i].md_name.mv_data = NULL;
3189 env->me_dbxs[i].md_name.mv_size = 0;
3190 env->me_dbflags[i] = 0;
3191 env->me_dbiseqs[i]++;
3197 if (keep && env->me_numdbs < n)
3201 /** End a transaction, except successful commit of a nested transaction.
3202 * May be called twice for readonly txns: First reset it, then abort.
3203 * @param[in] txn the transaction handle to end
3204 * @param[in] mode why and how to end the transaction
3207 mdb_txn_end(MDB_txn *txn, unsigned mode)
3209 MDB_env *env = txn->mt_env;
3211 static const char *const names[] = MDB_END_NAMES;
3214 /* Export or close DBI handles opened in this txn */
3215 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
3217 DPRINTF(("%s txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3218 names[mode & MDB_END_OPMASK],
3219 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3220 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
3222 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3223 if (txn->mt_u.reader) {
3224 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
3225 if (!(env->me_flags & MDB_NOTLS)) {
3226 txn->mt_u.reader = NULL; /* txn does not own reader */
3227 } else if (mode & MDB_END_SLOT) {
3228 txn->mt_u.reader->mr_pid = 0;
3229 txn->mt_u.reader = NULL;
3230 } /* else txn owns the slot until it does MDB_END_SLOT */
3232 txn->mt_numdbs = 0; /* prevent further DBI activity */
3233 txn->mt_flags |= MDB_TXN_FINISHED;
3235 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
3236 pgno_t *pghead = env->me_pghead;
3238 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
3239 mdb_cursors_close(txn, 0);
3240 if (!(env->me_flags & MDB_WRITEMAP)) {
3241 mdb_dlist_free(txn);
3245 txn->mt_flags = MDB_TXN_FINISHED;
3247 if (!txn->mt_parent) {
3248 mdb_midl_shrink(&txn->mt_free_pgs);
3249 env->me_free_pgs = txn->mt_free_pgs;
3251 env->me_pghead = NULL;
3255 mode = 0; /* txn == env->me_txn0, do not free() it */
3257 /* The writer mutex was locked in mdb_txn_begin. */
3259 UNLOCK_MUTEX(env->me_wmutex);
3261 txn->mt_parent->mt_child = NULL;
3262 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3263 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3264 mdb_midl_free(txn->mt_free_pgs);
3265 mdb_midl_free(txn->mt_spill_pgs);
3266 free(txn->mt_u.dirty_list);
3269 mdb_midl_free(pghead);
3272 if (!txn->mt_parent) {
3273 MDB_ID3L el = env->me_rpages, tl = txn->mt_rpages;
3274 unsigned i, x, n = tl[0].mid;
3275 pthread_mutex_lock(&env->me_rpmutex);
3276 for (i = 1; i <= n; i++) {
3277 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
3278 /* tmp overflow pages that we didn't share in env */
3279 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3281 x = mdb_mid3l_search(el, tl[i].mid);
3282 if (tl[i].mptr == el[x].mptr) {
3285 /* another tmp overflow page */
3286 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3290 pthread_mutex_unlock(&env->me_rpmutex);
3292 if (mode & MDB_END_FREE)
3296 if (mode & MDB_END_FREE)
3301 mdb_txn_reset(MDB_txn *txn)
3306 /* This call is only valid for read-only txns */
3307 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3310 mdb_txn_end(txn, MDB_END_RESET);
3314 mdb_txn_abort(MDB_txn *txn)
3320 mdb_txn_abort(txn->mt_child);
3322 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3325 /** Save the freelist as of this transaction to the freeDB.
3326 * This changes the freelist. Keep trying until it stabilizes.
3328 * When (MDB_DEVEL) & 2, the changes do not affect #mdb_page_alloc(),
3329 * it then uses the transaction's original snapshot of the freeDB.
3332 mdb_freelist_save(MDB_txn *txn)
3334 /* env->me_pghead[] can grow and shrink during this call.
3335 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3336 * Page numbers cannot disappear from txn->mt_free_pgs[].
3339 MDB_env *env = txn->mt_env;
3340 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3341 txnid_t pglast = 0, head_id = 0;
3342 pgno_t freecnt = 0, *free_pgs, *mop;
3343 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3345 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3347 if (env->me_pghead) {
3348 /* Make sure first page of freeDB is touched and on freelist */
3349 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3350 if (rc && rc != MDB_NOTFOUND)
3354 if (!env->me_pghead && txn->mt_loose_pgs) {
3355 /* Put loose page numbers in mt_free_pgs, since
3356 * we may be unable to return them to me_pghead.
3358 MDB_page *mp = txn->mt_loose_pgs;
3359 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3361 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3362 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3363 txn->mt_loose_pgs = NULL;
3364 txn->mt_loose_count = 0;
3367 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3368 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3369 ? SSIZE_MAX : maxfree_1pg;
3372 /* Come back here after each Put() in case freelist changed */
3377 /* If using records from freeDB which we have not yet
3378 * deleted, delete them and any we reserved for me_pghead.
3380 while (pglast < env->me_pglast) {
3381 rc = mdb_cursor_first(&mc, &key, NULL);
3384 pglast = head_id = *(txnid_t *)key.mv_data;
3385 total_room = head_room = 0;
3386 mdb_tassert(txn, pglast <= env->me_pglast);
3387 rc = mdb_cursor_del(&mc, 0);
3392 /* Save the IDL of pages freed by this txn, to a single record */
3393 if (freecnt < txn->mt_free_pgs[0]) {
3395 /* Make sure last page of freeDB is touched and on freelist */
3396 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3397 if (rc && rc != MDB_NOTFOUND)
3400 free_pgs = txn->mt_free_pgs;
3401 /* Write to last page of freeDB */
3402 key.mv_size = sizeof(txn->mt_txnid);
3403 key.mv_data = &txn->mt_txnid;
3405 freecnt = free_pgs[0];
3406 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3407 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3410 /* Retry if mt_free_pgs[] grew during the Put() */
3411 free_pgs = txn->mt_free_pgs;
3412 } while (freecnt < free_pgs[0]);
3413 mdb_midl_sort(free_pgs);
3414 memcpy(data.mv_data, free_pgs, data.mv_size);
3417 unsigned int i = free_pgs[0];
3418 DPRINTF(("IDL write txn %"Yu" root %"Yu" num %u",
3419 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3421 DPRINTF(("IDL %"Yu, free_pgs[i]));
3427 mop = env->me_pghead;
3428 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3430 /* Reserve records for me_pghead[]. Split it if multi-page,
3431 * to avoid searching freeDB for a page range. Use keys in
3432 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3434 if (total_room >= mop_len) {
3435 if (total_room == mop_len || --more < 0)
3437 } else if (head_room >= maxfree_1pg && head_id > 1) {
3438 /* Keep current record (overflow page), add a new one */
3442 /* (Re)write {key = head_id, IDL length = head_room} */
3443 total_room -= head_room;
3444 head_room = mop_len - total_room;
3445 if (head_room > maxfree_1pg && head_id > 1) {
3446 /* Overflow multi-page for part of me_pghead */
3447 head_room /= head_id; /* amortize page sizes */
3448 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3449 } else if (head_room < 0) {
3450 /* Rare case, not bothering to delete this record */
3453 key.mv_size = sizeof(head_id);
3454 key.mv_data = &head_id;
3455 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3456 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3459 /* IDL is initially empty, zero out at least the length */
3460 pgs = (pgno_t *)data.mv_data;
3461 j = head_room > clean_limit ? head_room : 0;
3465 total_room += head_room;
3468 /* Return loose page numbers to me_pghead, though usually none are
3469 * left at this point. The pages themselves remain in dirty_list.
3471 if (txn->mt_loose_pgs) {
3472 MDB_page *mp = txn->mt_loose_pgs;
3473 unsigned count = txn->mt_loose_count;
3475 /* Room for loose pages + temp IDL with same */
3476 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3478 mop = env->me_pghead;
3479 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3480 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3481 loose[ ++count ] = mp->mp_pgno;
3483 mdb_midl_sort(loose);
3484 mdb_midl_xmerge(mop, loose);
3485 txn->mt_loose_pgs = NULL;
3486 txn->mt_loose_count = 0;
3490 /* Fill in the reserved me_pghead records */
3496 rc = mdb_cursor_first(&mc, &key, &data);
3497 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3498 txnid_t id = *(txnid_t *)key.mv_data;
3499 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3502 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3504 if (len > mop_len) {
3506 data.mv_size = (len + 1) * sizeof(MDB_ID);
3508 data.mv_data = mop -= len;
3511 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3513 if (rc || !(mop_len -= len))
3520 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3521 * @param[in] txn the transaction that's being committed
3522 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3523 * @return 0 on success, non-zero on failure.
3526 mdb_page_flush(MDB_txn *txn, int keep)
3528 MDB_env *env = txn->mt_env;
3529 MDB_ID2L dl = txn->mt_u.dirty_list;
3530 unsigned psize = env->me_psize, j;
3531 int i, pagecount = dl[0].mid, rc;
3535 MDB_page *dp = NULL;
3539 struct iovec iov[MDB_COMMIT_PAGES];
3540 ssize_t wsize = 0, wres;
3541 off_t wpos = 0, next_pos = 1; /* impossible pos, so pos != next_pos */
3547 if (env->me_flags & MDB_WRITEMAP) {
3548 /* Clear dirty flags */
3549 while (++i <= pagecount) {
3551 /* Don't flush this page yet */
3552 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3553 dp->mp_flags &= ~P_KEEP;
3557 dp->mp_flags &= ~P_DIRTY;
3562 /* Write the pages */
3564 if (++i <= pagecount) {
3566 /* Don't flush this page yet */
3567 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3568 dp->mp_flags &= ~P_KEEP;
3573 /* clear dirty flag */
3574 dp->mp_flags &= ~P_DIRTY;
3577 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3582 /* Windows actually supports scatter/gather I/O, but only on
3583 * unbuffered file handles. Since we're relying on the OS page
3584 * cache for all our data, that's self-defeating. So we just
3585 * write pages one at a time. We use the ov structure to set
3586 * the write offset, to at least save the overhead of a Seek
3589 DPRINTF(("committing page %"Yu, pgno));
3590 memset(&ov, 0, sizeof(ov));
3591 ov.Offset = pos & 0xffffffff;
3592 ov.OffsetHigh = pos >> 16 >> 16;
3593 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3595 DPRINTF(("WriteFile: %d", rc));
3599 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3600 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3603 /* Write previous page(s) */
3604 #ifdef MDB_USE_PWRITEV
3605 wres = pwritev(env->me_fd, iov, n, wpos);
3608 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3611 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3615 DPRINTF(("lseek: %s", strerror(rc)));
3618 wres = writev(env->me_fd, iov, n);
3621 if (wres != wsize) {
3626 DPRINTF(("Write error: %s", strerror(rc)));
3628 rc = EIO; /* TODO: Use which error code? */
3629 DPUTS("short write, filesystem full?");
3640 DPRINTF(("committing page %"Yu, pgno));
3641 next_pos = pos + size;
3642 iov[n].iov_len = size;
3643 iov[n].iov_base = (char *)dp;
3649 if (pgno > txn->mt_last_pgno)
3650 txn->mt_last_pgno = pgno;
3653 /* MIPS has cache coherency issues, this is a no-op everywhere else
3654 * Note: for any size >= on-chip cache size, entire on-chip cache is
3657 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3659 for (i = keep; ++i <= pagecount; ) {
3661 /* This is a page we skipped above */
3664 dl[j].mid = dp->mp_pgno;
3667 mdb_dpage_free(env, dp);
3672 txn->mt_dirty_room += i - j;
3678 mdb_txn_commit(MDB_txn *txn)
3681 unsigned int i, end_mode;
3687 /* mdb_txn_end() mode for a commit which writes nothing */
3688 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3690 if (txn->mt_child) {
3691 rc = mdb_txn_commit(txn->mt_child);
3698 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3702 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3703 DPUTS("txn has failed/finished, can't commit");
3705 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3710 if (txn->mt_parent) {
3711 MDB_txn *parent = txn->mt_parent;
3715 unsigned x, y, len, ps_len;
3717 /* Append our free list to parent's */
3718 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3721 mdb_midl_free(txn->mt_free_pgs);
3722 /* Failures after this must either undo the changes
3723 * to the parent or set MDB_TXN_ERROR in the parent.
3726 parent->mt_next_pgno = txn->mt_next_pgno;
3727 parent->mt_flags = txn->mt_flags;
3729 /* Merge our cursors into parent's and close them */
3730 mdb_cursors_close(txn, 1);
3732 /* Update parent's DB table. */
3733 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3734 parent->mt_numdbs = txn->mt_numdbs;
3735 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3736 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3737 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3738 /* preserve parent's DB_NEW status */
3739 x = parent->mt_dbflags[i] & DB_NEW;
3740 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3743 dst = parent->mt_u.dirty_list;
3744 src = txn->mt_u.dirty_list;
3745 /* Remove anything in our dirty list from parent's spill list */
3746 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3748 pspill[0] = (pgno_t)-1;
3749 /* Mark our dirty pages as deleted in parent spill list */
3750 for (i=0, len=src[0].mid; ++i <= len; ) {
3751 MDB_ID pn = src[i].mid << 1;
3752 while (pn > pspill[x])
3754 if (pn == pspill[x]) {
3759 /* Squash deleted pagenums if we deleted any */
3760 for (x=y; ++x <= ps_len; )
3761 if (!(pspill[x] & 1))
3762 pspill[++y] = pspill[x];
3766 /* Remove anything in our spill list from parent's dirty list */
3767 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3768 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3769 MDB_ID pn = txn->mt_spill_pgs[i];
3771 continue; /* deleted spillpg */
3773 y = mdb_mid2l_search(dst, pn);
3774 if (y <= dst[0].mid && dst[y].mid == pn) {
3776 while (y < dst[0].mid) {
3785 /* Find len = length of merging our dirty list with parent's */
3787 dst[0].mid = 0; /* simplify loops */
3788 if (parent->mt_parent) {
3789 len = x + src[0].mid;
3790 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3791 for (i = x; y && i; y--) {
3792 pgno_t yp = src[y].mid;
3793 while (yp < dst[i].mid)
3795 if (yp == dst[i].mid) {
3800 } else { /* Simplify the above for single-ancestor case */
3801 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3803 /* Merge our dirty list with parent's */
3805 for (i = len; y; dst[i--] = src[y--]) {
3806 pgno_t yp = src[y].mid;
3807 while (yp < dst[x].mid)
3808 dst[i--] = dst[x--];
3809 if (yp == dst[x].mid)
3810 free(dst[x--].mptr);
3812 mdb_tassert(txn, i == x);
3814 free(txn->mt_u.dirty_list);
3815 parent->mt_dirty_room = txn->mt_dirty_room;
3816 if (txn->mt_spill_pgs) {
3817 if (parent->mt_spill_pgs) {
3818 /* TODO: Prevent failure here, so parent does not fail */
3819 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3821 parent->mt_flags |= MDB_TXN_ERROR;
3822 mdb_midl_free(txn->mt_spill_pgs);
3823 mdb_midl_sort(parent->mt_spill_pgs);
3825 parent->mt_spill_pgs = txn->mt_spill_pgs;
3829 /* Append our loose page list to parent's */
3830 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3832 *lp = txn->mt_loose_pgs;
3833 parent->mt_loose_count += txn->mt_loose_count;
3835 parent->mt_child = NULL;
3836 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3841 if (txn != env->me_txn) {
3842 DPUTS("attempt to commit unknown transaction");
3847 mdb_cursors_close(txn, 0);
3849 if (!txn->mt_u.dirty_list[0].mid &&
3850 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3853 DPRINTF(("committing txn %"Yu" %p on mdbenv %p, root page %"Yu,
3854 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3856 /* Update DB root pointers */
3857 if (txn->mt_numdbs > CORE_DBS) {
3861 data.mv_size = sizeof(MDB_db);
3863 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3864 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3865 if (txn->mt_dbflags[i] & DB_DIRTY) {
3866 if (TXN_DBI_CHANGED(txn, i)) {
3870 data.mv_data = &txn->mt_dbs[i];
3871 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3879 rc = mdb_freelist_save(txn);
3883 mdb_midl_free(env->me_pghead);
3884 env->me_pghead = NULL;
3885 mdb_midl_shrink(&txn->mt_free_pgs);
3891 if ((rc = mdb_page_flush(txn, 0)))
3893 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3894 (rc = mdb_env_sync0(env, 0, txn->mt_next_pgno)))
3896 if ((rc = mdb_env_write_meta(txn)))
3898 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3901 mdb_txn_end(txn, end_mode);
3909 /** Read the environment parameters of a DB environment before
3910 * mapping it into memory.
3911 * @param[in] env the environment handle
3912 * @param[out] meta address of where to store the meta information
3913 * @return 0 on success, non-zero on failure.
3916 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3922 enum { Size = sizeof(pbuf) };
3924 /* We don't know the page size yet, so use a minimum value.
3925 * Read both meta pages so we can use the latest one.
3928 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3932 memset(&ov, 0, sizeof(ov));
3934 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3935 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3938 rc = pread(env->me_fd, &pbuf, Size, off);
3941 if (rc == 0 && off == 0)
3943 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3944 DPRINTF(("read: %s", mdb_strerror(rc)));
3948 p = (MDB_page *)&pbuf;
3950 if (!F_ISSET(p->mp_flags, P_META)) {
3951 DPRINTF(("page %"Yu" not a meta page", p->mp_pgno));
3956 if (m->mm_magic != MDB_MAGIC) {
3957 DPUTS("meta has invalid magic");
3961 if (m->mm_version != MDB_DATA_VERSION) {
3962 DPRINTF(("database is version %u, expected version %u",
3963 m->mm_version, MDB_DATA_VERSION));
3964 return MDB_VERSION_MISMATCH;
3967 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3973 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3975 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3977 meta->mm_magic = MDB_MAGIC;
3978 meta->mm_version = MDB_DATA_VERSION;
3979 meta->mm_mapsize = env->me_mapsize;
3980 meta->mm_psize = env->me_psize;
3981 meta->mm_last_pg = NUM_METAS-1;
3982 meta->mm_flags = env->me_flags & 0xffff;
3983 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3984 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3985 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3988 /** Write the environment parameters of a freshly created DB environment.
3989 * @param[in] env the environment handle
3990 * @param[in] meta the #MDB_meta to write
3991 * @return 0 on success, non-zero on failure.
3994 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
4002 memset(&ov, 0, sizeof(ov));
4003 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
4005 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
4008 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
4009 len = pwrite(fd, ptr, size, pos); \
4010 if (len == -1 && ErrCode() == EINTR) continue; \
4011 rc = (len >= 0); break; } while(1)
4014 DPUTS("writing new meta page");
4016 psize = env->me_psize;
4018 p = calloc(NUM_METAS, psize);
4022 p->mp_flags = P_META;
4023 *(MDB_meta *)METADATA(p) = *meta;
4025 q = (MDB_page *)((char *)p + psize);
4027 q->mp_flags = P_META;
4028 *(MDB_meta *)METADATA(q) = *meta;
4030 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
4033 else if ((unsigned) len == psize * NUM_METAS)
4041 /** Update the environment info to commit a transaction.
4042 * @param[in] txn the transaction that's being committed
4043 * @return 0 on success, non-zero on failure.
4046 mdb_env_write_meta(MDB_txn *txn)
4049 MDB_meta meta, metab, *mp;
4053 int rc, len, toggle;
4062 toggle = txn->mt_txnid & 1;
4063 DPRINTF(("writing meta page %d for root page %"Yu,
4064 toggle, txn->mt_dbs[MAIN_DBI].md_root));
4067 flags = txn->mt_flags | env->me_flags;
4068 mp = env->me_metas[toggle];
4069 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
4070 /* Persist any increases of mapsize config */
4071 if (mapsize < env->me_mapsize)
4072 mapsize = env->me_mapsize;
4074 if (flags & MDB_WRITEMAP) {
4075 mp->mm_mapsize = mapsize;
4076 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4077 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4078 mp->mm_last_pg = txn->mt_next_pgno - 1;
4079 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
4080 !(defined(__i386__) || defined(__x86_64__))
4081 /* LY: issue a memory barrier, if not x86. ITS#7969 */
4082 __sync_synchronize();
4084 mp->mm_txnid = txn->mt_txnid;
4085 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
4086 unsigned meta_size = env->me_psize;
4087 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
4088 ptr = (char *)mp - PAGEHDRSZ;
4089 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
4090 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
4094 if (MDB_MSYNC(ptr, meta_size, rc)) {
4101 metab.mm_txnid = mp->mm_txnid;
4102 metab.mm_last_pg = mp->mm_last_pg;
4104 meta.mm_mapsize = mapsize;
4105 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4106 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4107 meta.mm_last_pg = txn->mt_next_pgno - 1;
4108 meta.mm_txnid = txn->mt_txnid;
4110 off = offsetof(MDB_meta, mm_mapsize);
4111 ptr = (char *)&meta + off;
4112 len = sizeof(MDB_meta) - off;
4113 off += (char *)mp - env->me_map;
4115 /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC.
4116 * (me_mfd goes to the same file as me_fd, but writing to it
4117 * also syncs to disk. Avoids a separate fdatasync() call.)
4119 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
4122 memset(&ov, 0, sizeof(ov));
4124 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
4129 rc = pwrite(mfd, ptr, len, off);
4132 rc = rc < 0 ? ErrCode() : EIO;
4137 DPUTS("write failed, disk error?");
4138 /* On a failure, the pagecache still contains the new data.
4139 * Write some old data back, to prevent it from being used.
4140 * Use the non-SYNC fd; we know it will fail anyway.
4142 meta.mm_last_pg = metab.mm_last_pg;
4143 meta.mm_txnid = metab.mm_txnid;
4145 memset(&ov, 0, sizeof(ov));
4147 WriteFile(env->me_fd, ptr, len, NULL, &ov);
4149 r2 = pwrite(env->me_fd, ptr, len, off);
4150 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
4153 env->me_flags |= MDB_FATAL_ERROR;
4156 /* MIPS has cache coherency issues, this is a no-op everywhere else */
4157 CACHEFLUSH(env->me_map + off, len, DCACHE);
4159 /* Memory ordering issues are irrelevant; since the entire writer
4160 * is wrapped by wmutex, all of these changes will become visible
4161 * after the wmutex is unlocked. Since the DB is multi-version,
4162 * readers will get consistent data regardless of how fresh or
4163 * how stale their view of these values is.
4166 env->me_txns->mti_txnid = txn->mt_txnid;
4171 /** Check both meta pages to see which one is newer.
4172 * @param[in] env the environment handle
4173 * @return newest #MDB_meta.
4176 mdb_env_pick_meta(const MDB_env *env)
4178 MDB_meta *const *metas = env->me_metas;
4179 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
4183 mdb_env_create(MDB_env **env)
4187 e = calloc(1, sizeof(MDB_env));
4191 e->me_maxreaders = DEFAULT_READERS;
4192 e->me_maxdbs = e->me_numdbs = CORE_DBS;
4193 e->me_fd = INVALID_HANDLE_VALUE;
4194 e->me_lfd = INVALID_HANDLE_VALUE;
4195 e->me_mfd = INVALID_HANDLE_VALUE;
4196 #ifdef MDB_USE_POSIX_SEM
4197 e->me_rmutex = SEM_FAILED;
4198 e->me_wmutex = SEM_FAILED;
4199 #elif defined MDB_USE_SYSV_SEM
4200 e->me_rmutex->semid = -1;
4201 e->me_wmutex->semid = -1;
4203 e->me_pid = getpid();
4204 GET_PAGESIZE(e->me_os_psize);
4205 VGMEMP_CREATE(e,0,0);
4211 /** @brief Map a result from an NTAPI call to WIN32. */
4213 mdb_nt2win32(NTSTATUS st)
4218 GetOverlappedResult(NULL, &o, &br, FALSE);
4219 return GetLastError();
4224 mdb_env_map(MDB_env *env, void *addr)
4227 unsigned int flags = env->me_flags;
4230 int access = SECTION_MAP_READ;
4234 ULONG pageprot = PAGE_READONLY, secprot, alloctype;
4236 if (flags & MDB_WRITEMAP) {
4237 access |= SECTION_MAP_WRITE;
4238 pageprot = PAGE_READWRITE;
4240 if (flags & MDB_RDONLY) {
4241 secprot = PAGE_READONLY;
4245 secprot = PAGE_READWRITE;
4246 msize = env->me_mapsize;
4247 alloctype = MEM_RESERVE;
4250 rc = NtCreateSection(&mh, access, NULL, NULL, secprot, SEC_RESERVE, env->me_fd);
4252 return mdb_nt2win32(rc);
4255 msize = NUM_METAS * env->me_psize;
4257 rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
4264 return mdb_nt2win32(rc);
4269 env->me_map = mmap(addr, NUM_METAS * env->me_psize, PROT_READ, MAP_SHARED,
4271 if (env->me_map == MAP_FAILED) {
4276 int prot = PROT_READ;
4277 if (flags & MDB_WRITEMAP) {
4279 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
4282 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
4284 if (env->me_map == MAP_FAILED) {
4289 if (flags & MDB_NORDAHEAD) {
4290 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
4292 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
4294 #ifdef POSIX_MADV_RANDOM
4295 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
4296 #endif /* POSIX_MADV_RANDOM */
4297 #endif /* MADV_RANDOM */
4301 /* Can happen because the address argument to mmap() is just a
4302 * hint. mmap() can pick another, e.g. if the range is in use.
4303 * The MAP_FIXED flag would prevent that, but then mmap could
4304 * instead unmap existing pages to make room for the new map.
4306 if (addr && env->me_map != addr)
4307 return EBUSY; /* TODO: Make a new MDB_* error code? */
4310 p = (MDB_page *)env->me_map;
4311 env->me_metas[0] = METADATA(p);
4312 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4318 mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
4320 /* If env is already open, caller is responsible for making
4321 * sure there are no active txns.
4331 meta = mdb_env_pick_meta(env);
4333 size = meta->mm_mapsize;
4335 /* Silently round up to minimum if the size is too small */
4336 mdb_size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4341 /* For MDB_VL32 this bit is a noop since we dynamically remap
4342 * chunks of the DB anyway.
4344 munmap(env->me_map, env->me_mapsize);
4345 env->me_mapsize = size;
4346 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4347 rc = mdb_env_map(env, old);
4350 #endif /* !MDB_VL32 */
4352 env->me_mapsize = size;
4354 env->me_maxpg = env->me_mapsize / env->me_psize;
4359 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4363 env->me_maxdbs = dbs + CORE_DBS;
4368 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4370 if (env->me_map || readers < 1)
4372 env->me_maxreaders = readers;
4377 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4379 if (!env || !readers)
4381 *readers = env->me_maxreaders;
4386 mdb_fsize(HANDLE fd, mdb_size_t *size)
4389 LARGE_INTEGER fsize;
4391 if (!GetFileSizeEx(fd, &fsize))
4394 *size = fsize.QuadPart;
4408 typedef wchar_t mdb_nchar_t;
4409 # define MDB_NAME(str) L##str
4410 # define mdb_name_cpy wcscpy
4412 /** Character type for file names: char on Unix, wchar_t on Windows */
4413 typedef char mdb_nchar_t;
4414 # define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */
4415 # define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */
4418 /** Filename - string of #mdb_nchar_t[] */
4419 typedef struct MDB_name {
4420 int mn_len; /**< Length */
4421 int mn_alloced; /**< True if #mn_val was malloced */
4422 mdb_nchar_t *mn_val; /**< Contents */
4425 /** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */
4426 static const mdb_nchar_t *const mdb_suffixes[2][2] = {
4427 { MDB_NAME("/data.mdb"), MDB_NAME("") },
4428 { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") }
4431 #define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */
4433 /** Set up filename + scratch area for filename suffix, for opening files.
4434 * It should be freed with #mdb_fname_destroy().
4435 * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16.
4437 * @param[in] path Pathname for #mdb_env_open().
4438 * @param[in] envflags Whether a subdir and/or lockfile will be used.
4439 * @param[out] fname Resulting filename, with room for a suffix if necessary.
4442 mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname)
4444 int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK);
4445 fname->mn_alloced = 0;
4447 return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN);
4449 fname->mn_len = strlen(path);
4451 fname->mn_val = (char *) path;
4452 else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) {
4453 fname->mn_alloced = 1;
4454 strcpy(fname->mn_val, path);
4462 /** Destroy \b fname from #mdb_fname_init() */
4463 #define mdb_fname_destroy(fname) \
4464 do { if ((fname).mn_alloced) free((fname).mn_val); } while (0)
4466 #ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */
4467 # define MDB_CLOEXEC O_CLOEXEC
4469 # define MDB_CLOEXEC 0
4472 /** File type, access mode etc. for #mdb_fopen() */
4473 enum mdb_fopen_type {
4475 MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS
4477 /* A comment in mdb_fopen() explains some O_* flag choices. */
4478 MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */
4479 MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */
4480 MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */
4481 MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */
4482 /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits
4483 * distinguish otherwise-equal MDB_O_* constants from each other.
4485 MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY,
4486 MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */
4490 /** Open an LMDB file.
4491 * @param[in] env The LMDB environment.
4492 * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is
4493 * appended if necessary to create the filename, without changing mn_len.
4494 * @param[in] which Determines file type, access mode, etc.
4495 * @param[in] mode The Unix permissions for the file, if we create it.
4496 * @param[out] res Resulting file handle.
4497 * @return 0 on success, non-zero on failure.
4500 mdb_fopen(const MDB_env *env, MDB_name *fname,
4501 enum mdb_fopen_type which, mdb_mode_t mode,
4504 int rc = MDB_SUCCESS;
4507 DWORD acc, share, disp, attrs;
4512 if (fname->mn_alloced) /* modifiable copy */
4513 mdb_name_cpy(fname->mn_val + fname->mn_len,
4514 mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]);
4516 /* The directory must already exist. Usually the file need not.
4517 * MDB_O_META requires the file because we already created it using
4518 * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file.
4520 * With MDB_O_COPY we do not want the OS to cache the writes, since
4521 * the source data is already in the OS cache.
4523 * The lockfile needs FD_CLOEXEC (close file descriptor on exec*())
4524 * to avoid the flock() issues noted under Caveats in lmdb.h.
4525 * Also set it for other filehandles which the user cannot get at
4526 * and close himself, which he may need after fork(). I.e. all but
4527 * me_fd, which programs do use via mdb_env_get_fd().
4531 acc = GENERIC_READ|GENERIC_WRITE;
4532 share = FILE_SHARE_READ|FILE_SHARE_WRITE;
4534 attrs = FILE_ATTRIBUTE_NORMAL;
4536 case MDB_O_RDONLY: /* read-only datafile */
4538 disp = OPEN_EXISTING;
4540 case MDB_O_META: /* for writing metapages */
4541 acc = GENERIC_WRITE;
4542 disp = OPEN_EXISTING;
4543 attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH;
4545 case MDB_O_COPY: /* mdb_env_copy() & co */
4546 acc = GENERIC_WRITE;
4549 attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH;
4551 default: break; /* silence gcc -Wswitch (not all enum values handled) */
4553 fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL);
4555 fd = open(fname->mn_val, which & MDB_O_MASK, mode);
4558 if (fd == INVALID_HANDLE_VALUE)
4562 if (which != MDB_O_RDONLY && which != MDB_O_RDWR) {
4563 /* Set CLOEXEC if we could not pass it to open() */
4564 if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1)
4565 (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
4567 if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) {
4568 /* This may require buffer alignment. There is no portable
4569 * way to ask how much, so we require OS pagesize alignment.
4571 # ifdef F_NOCACHE /* __APPLE__ */
4572 (void) fcntl(fd, F_NOCACHE, 1);
4573 # elif defined O_DIRECT
4574 /* open(...O_DIRECT...) would break on filesystems without
4575 * O_DIRECT support (ITS#7682). Try to set it here instead.
4577 if ((flags = fcntl(fd, F_GETFL)) != -1)
4578 (void) fcntl(fd, F_SETFL, flags | O_DIRECT);
4582 #endif /* !_WIN32 */
4589 #ifdef BROKEN_FDATASYNC
4590 #include <sys/utsname.h>
4591 #include <sys/vfs.h>
4594 /** Further setup required for opening an LMDB environment
4597 mdb_env_open2(MDB_env *env)
4599 unsigned int flags = env->me_flags;
4600 int i, newenv = 0, rc;
4604 /* See if we should use QueryLimited */
4606 if ((rc & 0xff) > 5)
4607 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4609 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4612 #ifdef BROKEN_FDATASYNC
4613 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4614 * https://lkml.org/lkml/2012/9/3/83
4615 * Kernels after 3.6-rc6 are known good.
4616 * https://lkml.org/lkml/2012/9/10/556
4617 * See if the DB is on ext3/ext4, then check for new enough kernel
4618 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4623 fstatfs(env->me_fd, &st);
4624 while (st.f_type == 0xEF53) {
4628 if (uts.release[0] < '3') {
4629 if (!strncmp(uts.release, "2.6.32.", 7)) {
4630 i = atoi(uts.release+7);
4632 break; /* 2.6.32.60 and newer is OK */
4633 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4634 i = atoi(uts.release+7);
4636 break; /* 2.6.34.15 and newer is OK */
4638 } else if (uts.release[0] == '3') {
4639 i = atoi(uts.release+2);
4641 break; /* 3.6 and newer is OK */
4643 i = atoi(uts.release+4);
4645 break; /* 3.5.4 and newer is OK */
4646 } else if (i == 2) {
4647 i = atoi(uts.release+4);
4649 break; /* 3.2.30 and newer is OK */
4651 } else { /* 4.x and newer is OK */
4654 env->me_flags |= MDB_FSYNCONLY;
4660 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4663 DPUTS("new mdbenv");
4665 env->me_psize = env->me_os_psize;
4666 if (env->me_psize > MAX_PAGESIZE)
4667 env->me_psize = MAX_PAGESIZE;
4668 memset(&meta, 0, sizeof(meta));
4669 mdb_env_init_meta0(env, &meta);
4670 meta.mm_mapsize = DEFAULT_MAPSIZE;
4672 env->me_psize = meta.mm_psize;
4675 /* Was a mapsize configured? */
4676 if (!env->me_mapsize) {
4677 env->me_mapsize = meta.mm_mapsize;
4680 /* Make sure mapsize >= committed data size. Even when using
4681 * mm_mapsize, which could be broken in old files (ITS#7789).
4683 mdb_size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4684 if (env->me_mapsize < minsize)
4685 env->me_mapsize = minsize;
4687 meta.mm_mapsize = env->me_mapsize;
4689 if (newenv && !(flags & MDB_FIXEDMAP)) {
4690 /* mdb_env_map() may grow the datafile. Write the metapages
4691 * first, so the file will be valid if initialization fails.
4692 * Except with FIXEDMAP, since we do not yet know mm_address.
4693 * We could fill in mm_address later, but then a different
4694 * program might end up doing that - one with a memory layout
4695 * and map address which does not suit the main program.
4697 rc = mdb_env_init_meta(env, &meta);
4703 /* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
4707 rc = WriteFile(env->me_fd, &dummy, 1, &len, NULL);
4715 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4720 if (flags & MDB_FIXEDMAP)
4721 meta.mm_address = env->me_map;
4722 i = mdb_env_init_meta(env, &meta);
4723 if (i != MDB_SUCCESS) {
4728 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4729 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4731 #if !(MDB_MAXKEYSIZE)
4732 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4734 env->me_maxpg = env->me_mapsize / env->me_psize;
4738 MDB_meta *meta = mdb_env_pick_meta(env);
4739 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4741 DPRINTF(("opened database version %u, pagesize %u",
4742 meta->mm_version, env->me_psize));
4743 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4744 DPRINTF(("depth: %u", db->md_depth));
4745 DPRINTF(("entries: %"Yu, db->md_entries));
4746 DPRINTF(("branch pages: %"Yu, db->md_branch_pages));
4747 DPRINTF(("leaf pages: %"Yu, db->md_leaf_pages));
4748 DPRINTF(("overflow pages: %"Yu, db->md_overflow_pages));
4749 DPRINTF(("root: %"Yu, db->md_root));
4757 /** Release a reader thread's slot in the reader lock table.
4758 * This function is called automatically when a thread exits.
4759 * @param[in] ptr This points to the slot in the reader lock table.
4762 mdb_env_reader_dest(void *ptr)
4764 MDB_reader *reader = ptr;
4767 if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */
4769 /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */
4774 /** Junk for arranging thread-specific callbacks on Windows. This is
4775 * necessarily platform and compiler-specific. Windows supports up
4776 * to 1088 keys. Let's assume nobody opens more than 64 environments
4777 * in a single process, for now. They can override this if needed.
4779 #ifndef MAX_TLS_KEYS
4780 #define MAX_TLS_KEYS 64
4782 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4783 static int mdb_tls_nkeys;
4785 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4789 case DLL_PROCESS_ATTACH: break;
4790 case DLL_THREAD_ATTACH: break;
4791 case DLL_THREAD_DETACH:
4792 for (i=0; i<mdb_tls_nkeys; i++) {
4793 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4795 mdb_env_reader_dest(r);
4799 case DLL_PROCESS_DETACH: break;
4804 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4806 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4810 /* Force some symbol references.
4811 * _tls_used forces the linker to create the TLS directory if not already done
4812 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4814 #pragma comment(linker, "/INCLUDE:_tls_used")
4815 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4816 #pragma const_seg(".CRT$XLB")
4817 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4818 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4821 #pragma comment(linker, "/INCLUDE:__tls_used")
4822 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4823 #pragma data_seg(".CRT$XLB")
4824 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4826 #endif /* WIN 32/64 */
4827 #endif /* !__GNUC__ */
4830 /** Downgrade the exclusive lock on the region back to shared */
4832 mdb_env_share_locks(MDB_env *env, int *excl)
4835 MDB_meta *meta = mdb_env_pick_meta(env);
4837 env->me_txns->mti_txnid = meta->mm_txnid;
4842 /* First acquire a shared lock. The Unlock will
4843 * then release the existing exclusive lock.
4845 memset(&ov, 0, sizeof(ov));
4846 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4849 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4855 struct flock lock_info;
4856 /* The shared lock replaces the existing lock */
4857 memset((void *)&lock_info, 0, sizeof(lock_info));
4858 lock_info.l_type = F_RDLCK;
4859 lock_info.l_whence = SEEK_SET;
4860 lock_info.l_start = 0;
4861 lock_info.l_len = 1;
4862 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4863 (rc = ErrCode()) == EINTR) ;
4864 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4871 /** Try to get exclusive lock, otherwise shared.
4872 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4875 mdb_env_excl_lock(MDB_env *env, int *excl)
4879 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4883 memset(&ov, 0, sizeof(ov));
4884 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4891 struct flock lock_info;
4892 memset((void *)&lock_info, 0, sizeof(lock_info));
4893 lock_info.l_type = F_WRLCK;
4894 lock_info.l_whence = SEEK_SET;
4895 lock_info.l_start = 0;
4896 lock_info.l_len = 1;
4897 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4898 (rc = ErrCode()) == EINTR) ;
4902 # ifndef MDB_USE_POSIX_MUTEX
4903 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4906 lock_info.l_type = F_RDLCK;
4907 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4908 (rc = ErrCode()) == EINTR) ;
4918 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4920 * @(#) $Revision: 5.1 $
4921 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4922 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4924 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4928 * Please do not copyright this code. This code is in the public domain.
4930 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4931 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4932 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4933 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4934 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4935 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4936 * PERFORMANCE OF THIS SOFTWARE.
4939 * chongo <Landon Curt Noll> /\oo/\
4940 * http://www.isthe.com/chongo/
4942 * Share and Enjoy! :-)
4945 typedef unsigned long long mdb_hash_t;
4946 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4948 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4949 * @param[in] val value to hash
4950 * @param[in] hval initial value for hash
4951 * @return 64 bit hash
4953 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4954 * hval arg on the first call.
4957 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4959 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4960 unsigned char *end = s + val->mv_size;
4962 * FNV-1a hash each octet of the string
4965 /* xor the bottom with the current octet */
4966 hval ^= (mdb_hash_t)*s++;
4968 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4969 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4970 (hval << 7) + (hval << 8) + (hval << 40);
4972 /* return our new hash value */
4976 /** Hash the string and output the encoded hash.
4977 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4978 * very short name limits. We don't care about the encoding being reversible,
4979 * we just want to preserve as many bits of the input as possible in a
4980 * small printable string.
4981 * @param[in] str string to hash
4982 * @param[out] encbuf an array of 11 chars to hold the hash
4984 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4987 mdb_pack85(unsigned long l, char *out)
4991 for (i=0; i<5; i++) {
4992 *out++ = mdb_a85[l % 85];
4998 mdb_hash_enc(MDB_val *val, char *encbuf)
5000 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
5002 mdb_pack85(h, encbuf);
5003 mdb_pack85(h>>32, encbuf+5);
5008 /** Open and/or initialize the lock region for the environment.
5009 * @param[in] env The LMDB environment.
5010 * @param[in] fname Filename + scratch area, from #mdb_fname_init().
5011 * @param[in] mode The Unix permissions for the file, if we create it.
5012 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
5013 * @return 0 on success, non-zero on failure.
5016 mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl)
5019 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
5021 # define MDB_ERRCODE_ROFS EROFS
5023 #ifdef MDB_USE_SYSV_SEM
5030 rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd);
5032 /* Omit lockfile if read-only env on read-only filesystem */
5033 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
5039 if (!(env->me_flags & MDB_NOTLS)) {
5040 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
5043 env->me_flags |= MDB_ENV_TXKEY;
5045 /* Windows TLS callbacks need help finding their TLS info. */
5046 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
5050 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
5054 /* Try to get exclusive lock. If we succeed, then
5055 * nobody is using the lock region and we should initialize it.
5057 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
5060 size = GetFileSize(env->me_lfd, NULL);
5062 size = lseek(env->me_lfd, 0, SEEK_END);
5063 if (size == -1) goto fail_errno;
5065 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
5066 if (size < rsize && *excl > 0) {
5068 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
5069 || !SetEndOfFile(env->me_lfd))
5072 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
5076 size = rsize - sizeof(MDB_txninfo);
5077 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
5082 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
5084 if (!mh) goto fail_errno;
5085 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
5087 if (!env->me_txns) goto fail_errno;
5089 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
5091 if (m == MAP_FAILED) goto fail_errno;
5097 BY_HANDLE_FILE_INFORMATION stbuf;
5106 if (!mdb_sec_inited) {
5107 InitializeSecurityDescriptor(&mdb_null_sd,
5108 SECURITY_DESCRIPTOR_REVISION);
5109 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
5110 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
5111 mdb_all_sa.bInheritHandle = FALSE;
5112 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
5115 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
5116 idbuf.volume = stbuf.dwVolumeSerialNumber;
5117 idbuf.nhigh = stbuf.nFileIndexHigh;
5118 idbuf.nlow = stbuf.nFileIndexLow;
5119 val.mv_data = &idbuf;
5120 val.mv_size = sizeof(idbuf);
5121 mdb_hash_enc(&val, encbuf);
5122 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
5123 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
5124 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
5125 if (!env->me_rmutex) goto fail_errno;
5126 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
5127 if (!env->me_wmutex) goto fail_errno;
5128 #elif defined(MDB_USE_POSIX_SEM)
5137 #if defined(__NetBSD__)
5138 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
5140 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
5141 idbuf.dev = stbuf.st_dev;
5142 idbuf.ino = stbuf.st_ino;
5143 val.mv_data = &idbuf;
5144 val.mv_size = sizeof(idbuf);
5145 mdb_hash_enc(&val, encbuf);
5146 #ifdef MDB_SHORT_SEMNAMES
5147 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
5149 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
5150 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
5151 /* Clean up after a previous run, if needed: Try to
5152 * remove both semaphores before doing anything else.
5154 sem_unlink(env->me_txns->mti_rmname);
5155 sem_unlink(env->me_txns->mti_wmname);
5156 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
5157 O_CREAT|O_EXCL, mode, 1);
5158 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
5159 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
5160 O_CREAT|O_EXCL, mode, 1);
5161 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
5162 #elif defined(MDB_USE_SYSV_SEM)
5163 unsigned short vals[2] = {1, 1};
5164 key_t key = ftok(fname->mn_val, 'M'); /* fname is lockfile path now */
5167 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
5171 if (semctl(semid, 0, SETALL, semu) < 0)
5173 env->me_txns->mti_semid = semid;
5174 env->me_txns->mti_rlocked = 0;
5175 env->me_txns->mti_wlocked = 0;
5176 #else /* MDB_USE_POSIX_MUTEX: */
5177 pthread_mutexattr_t mattr;
5179 /* Solaris needs this before initing a robust mutex. Otherwise
5180 * it may skip the init and return EBUSY "seems someone already
5181 * inited" or EINVAL "it was inited differently".
5183 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
5184 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
5186 if ((rc = pthread_mutexattr_init(&mattr)) != 0)
5188 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
5189 #ifdef MDB_ROBUST_SUPPORTED
5190 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
5192 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
5193 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
5194 pthread_mutexattr_destroy(&mattr);
5197 #endif /* _WIN32 || ... */
5199 env->me_txns->mti_magic = MDB_MAGIC;
5200 env->me_txns->mti_format = MDB_LOCK_FORMAT;
5201 env->me_txns->mti_txnid = 0;
5202 env->me_txns->mti_numreaders = 0;
5205 #ifdef MDB_USE_SYSV_SEM
5206 struct semid_ds buf;
5208 if (env->me_txns->mti_magic != MDB_MAGIC) {
5209 DPUTS("lock region has invalid magic");
5213 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
5214 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
5215 env->me_txns->mti_format, MDB_LOCK_FORMAT));
5216 rc = MDB_VERSION_MISMATCH;
5220 if (rc && rc != EACCES && rc != EAGAIN) {
5224 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
5225 if (!env->me_rmutex) goto fail_errno;
5226 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
5227 if (!env->me_wmutex) goto fail_errno;
5228 #elif defined(MDB_USE_POSIX_SEM)
5229 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
5230 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
5231 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
5232 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
5233 #elif defined(MDB_USE_SYSV_SEM)
5234 semid = env->me_txns->mti_semid;
5236 /* check for read access */
5237 if (semctl(semid, 0, IPC_STAT, semu) < 0)
5239 /* check for write access */
5240 if (semctl(semid, 0, IPC_SET, semu) < 0)
5244 #ifdef MDB_USE_SYSV_SEM
5245 env->me_rmutex->semid = semid;
5246 env->me_wmutex->semid = semid;
5247 env->me_rmutex->semnum = 0;
5248 env->me_wmutex->semnum = 1;
5249 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
5250 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
5261 /** Only a subset of the @ref mdb_env flags can be changed
5262 * at runtime. Changing other flags requires closing the
5263 * environment and re-opening it with the new flags.
5265 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
5266 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
5267 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
5269 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
5270 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
5274 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
5279 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
5283 if (flags & MDB_WRITEMAP) {
5284 /* silently ignore WRITEMAP in 32 bit mode */
5285 flags ^= MDB_WRITEMAP;
5287 if (flags & MDB_FIXEDMAP) {
5288 /* cannot support FIXEDMAP */
5292 flags |= env->me_flags;
5294 rc = mdb_fname_init(path, flags, &fname);
5300 env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
5301 if (!env->me_rpmutex) {
5306 rc = pthread_mutex_init(&env->me_rpmutex, NULL);
5311 flags |= MDB_ENV_ACTIVE; /* tell mdb_env_close0() to clean up */
5313 if (flags & MDB_RDONLY) {
5314 /* silently ignore WRITEMAP when we're only getting read access */
5315 flags &= ~MDB_WRITEMAP;
5317 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
5318 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
5322 env->me_flags = flags;
5328 env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
5329 if (!env->me_rpages) {
5333 env->me_rpages[0].mid = 0;
5334 env->me_rpcheck = MDB_ERPAGE_SIZE/2;
5338 env->me_path = strdup(path);
5339 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
5340 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
5341 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
5342 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
5346 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
5348 /* For RDONLY, get lockfile after we know datafile exists */
5349 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
5350 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
5355 rc = mdb_fopen(env, &fname,
5356 (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR,
5361 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
5362 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
5367 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
5368 if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) {
5369 /* Synchronous fd for meta writes. Needed even with
5370 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
5372 rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd);
5376 DPRINTF(("opened dbenv %p", (void *) env));
5378 rc = mdb_env_share_locks(env, &excl);
5382 if (!(flags & MDB_RDONLY)) {
5384 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
5385 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
5386 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
5387 (txn = calloc(1, size)))
5389 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
5390 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
5391 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
5392 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
5395 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
5396 if (!txn->mt_rpages) {
5401 txn->mt_rpages[0].mid = 0;
5402 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
5404 txn->mt_dbxs = env->me_dbxs;
5405 txn->mt_flags = MDB_TXN_FINISHED;
5415 mdb_env_close0(env, excl);
5417 mdb_fname_destroy(fname);
5421 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5423 mdb_env_close0(MDB_env *env, int excl)
5427 if (!(env->me_flags & MDB_ENV_ACTIVE))
5430 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5432 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5433 free(env->me_dbxs[i].md_name.mv_data);
5438 free(env->me_dbiseqs);
5439 free(env->me_dbflags);
5441 free(env->me_dirty_list);
5443 if (env->me_txn0 && env->me_txn0->mt_rpages)
5444 free(env->me_txn0->mt_rpages);
5445 if (env->me_rpages) {
5446 MDB_ID3L el = env->me_rpages;
5448 for (x=1; x<=el[0].mid; x++)
5449 munmap(el[x].mptr, el[x].mcnt * env->me_psize);
5454 mdb_midl_free(env->me_free_pgs);
5456 if (env->me_flags & MDB_ENV_TXKEY) {
5457 pthread_key_delete(env->me_txkey);
5459 /* Delete our key from the global list */
5460 for (i=0; i<mdb_tls_nkeys; i++)
5461 if (mdb_tls_keys[i] == env->me_txkey) {
5462 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5471 munmap(env->me_map, NUM_METAS*env->me_psize);
5473 munmap(env->me_map, env->me_mapsize);
5476 if (env->me_mfd != INVALID_HANDLE_VALUE)
5477 (void) close(env->me_mfd);
5478 if (env->me_fd != INVALID_HANDLE_VALUE)
5479 (void) close(env->me_fd);
5481 MDB_PID_T pid = env->me_pid;
5482 /* Clearing readers is done in this function because
5483 * me_txkey with its destructor must be disabled first.
5485 * We skip the the reader mutex, so we touch only
5486 * data owned by this process (me_close_readers and
5487 * our readers), and clear each reader atomically.
5489 for (i = env->me_close_readers; --i >= 0; )
5490 if (env->me_txns->mti_readers[i].mr_pid == pid)
5491 env->me_txns->mti_readers[i].mr_pid = 0;
5493 if (env->me_rmutex) {
5494 CloseHandle(env->me_rmutex);
5495 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5497 /* Windows automatically destroys the mutexes when
5498 * the last handle closes.
5500 #elif defined(MDB_USE_POSIX_SEM)
5501 if (env->me_rmutex != SEM_FAILED) {
5502 sem_close(env->me_rmutex);
5503 if (env->me_wmutex != SEM_FAILED)
5504 sem_close(env->me_wmutex);
5505 /* If we have the filelock: If we are the
5506 * only remaining user, clean up semaphores.
5509 mdb_env_excl_lock(env, &excl);
5511 sem_unlink(env->me_txns->mti_rmname);
5512 sem_unlink(env->me_txns->mti_wmname);
5515 #elif defined(MDB_USE_SYSV_SEM)
5516 if (env->me_rmutex->semid != -1) {
5517 /* If we have the filelock: If we are the
5518 * only remaining user, clean up semaphores.
5521 mdb_env_excl_lock(env, &excl);
5523 semctl(env->me_rmutex->semid, 0, IPC_RMID);
5526 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5528 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5531 /* Unlock the lockfile. Windows would have unlocked it
5532 * after closing anyway, but not necessarily at once.
5534 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5537 (void) close(env->me_lfd);
5541 if (env->me_fmh) CloseHandle(env->me_fmh);
5542 if (env->me_rpmutex) CloseHandle(env->me_rpmutex);
5544 pthread_mutex_destroy(&env->me_rpmutex);
5548 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5552 mdb_env_close(MDB_env *env)
5559 VGMEMP_DESTROY(env);
5560 while ((dp = env->me_dpages) != NULL) {
5561 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5562 env->me_dpages = dp->mp_next;
5566 mdb_env_close0(env, 0);
5570 /** Compare two items pointing at aligned #mdb_size_t's */
5572 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5574 return (*(mdb_size_t *)a->mv_data < *(mdb_size_t *)b->mv_data) ? -1 :
5575 *(mdb_size_t *)a->mv_data > *(mdb_size_t *)b->mv_data;
5578 /** Compare two items pointing at aligned unsigned int's.
5580 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5581 * but #mdb_cmp_clong() is called instead if the data type is #mdb_size_t.
5584 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5586 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5587 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5590 /** Compare two items pointing at unsigned ints of unknown alignment.
5591 * Nodes and keys are guaranteed to be 2-byte aligned.
5594 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5596 #if BYTE_ORDER == LITTLE_ENDIAN
5597 unsigned short *u, *c;
5600 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5601 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5604 } while(!x && u > (unsigned short *)a->mv_data);
5607 unsigned short *u, *c, *end;
5610 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5611 u = (unsigned short *)a->mv_data;
5612 c = (unsigned short *)b->mv_data;
5615 } while(!x && u < end);
5620 /** Compare two items lexically */
5622 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5629 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5635 diff = memcmp(a->mv_data, b->mv_data, len);
5636 return diff ? diff : len_diff<0 ? -1 : len_diff;
5639 /** Compare two items in reverse byte order */
5641 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5643 const unsigned char *p1, *p2, *p1_lim;
5647 p1_lim = (const unsigned char *)a->mv_data;
5648 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5649 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5651 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5657 while (p1 > p1_lim) {
5658 diff = *--p1 - *--p2;
5662 return len_diff<0 ? -1 : len_diff;
5665 /** Search for key within a page, using binary search.
5666 * Returns the smallest entry larger or equal to the key.
5667 * If exactp is non-null, stores whether the found entry was an exact match
5668 * in *exactp (1 or 0).
5669 * Updates the cursor index with the index of the found entry.
5670 * If no entry larger or equal to the key is found, returns NULL.
5673 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5675 unsigned int i = 0, nkeys;
5678 MDB_page *mp = mc->mc_pg[mc->mc_top];
5679 MDB_node *node = NULL;
5684 nkeys = NUMKEYS(mp);
5686 DPRINTF(("searching %u keys in %s %spage %"Yu,
5687 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5690 low = IS_LEAF(mp) ? 0 : 1;
5692 cmp = mc->mc_dbx->md_cmp;
5694 /* Branch pages have no data, so if using integer keys,
5695 * alignment is guaranteed. Use faster mdb_cmp_int.
5697 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5698 if (NODEPTR(mp, 1)->mn_ksize == sizeof(mdb_size_t))
5705 nodekey.mv_size = mc->mc_db->md_pad;
5706 node = NODEPTR(mp, 0); /* fake */
5707 while (low <= high) {
5708 i = (low + high) >> 1;
5709 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5710 rc = cmp(key, &nodekey);
5711 DPRINTF(("found leaf index %u [%s], rc = %i",
5712 i, DKEY(&nodekey), rc));
5721 while (low <= high) {
5722 i = (low + high) >> 1;
5724 node = NODEPTR(mp, i);
5725 nodekey.mv_size = NODEKSZ(node);
5726 nodekey.mv_data = NODEKEY(node);
5728 rc = cmp(key, &nodekey);
5731 DPRINTF(("found leaf index %u [%s], rc = %i",
5732 i, DKEY(&nodekey), rc));
5734 DPRINTF(("found branch index %u [%s -> %"Yu"], rc = %i",
5735 i, DKEY(&nodekey), NODEPGNO(node), rc));
5746 if (rc > 0) { /* Found entry is less than the key. */
5747 i++; /* Skip to get the smallest entry larger than key. */
5749 node = NODEPTR(mp, i);
5752 *exactp = (rc == 0 && nkeys > 0);
5753 /* store the key index */
5754 mc->mc_ki[mc->mc_top] = i;
5756 /* There is no entry larger or equal to the key. */
5759 /* nodeptr is fake for LEAF2 */
5765 mdb_cursor_adjust(MDB_cursor *mc, func)
5769 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5770 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5777 /** Pop a page off the top of the cursor's stack. */
5779 mdb_cursor_pop(MDB_cursor *mc)
5782 DPRINTF(("popping page %"Yu" off db %d cursor %p",
5783 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5789 mc->mc_flags &= ~C_INITIALIZED;
5794 /** Push a page onto the top of the cursor's stack.
5795 * Set #MDB_TXN_ERROR on failure.
5798 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5800 DPRINTF(("pushing page %"Yu" on db %d cursor %p", mp->mp_pgno,
5801 DDBI(mc), (void *) mc));
5803 if (mc->mc_snum >= CURSOR_STACK) {
5804 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5805 return MDB_CURSOR_FULL;
5808 mc->mc_top = mc->mc_snum++;
5809 mc->mc_pg[mc->mc_top] = mp;
5810 mc->mc_ki[mc->mc_top] = 0;
5816 /** Map a read-only page.
5817 * There are two levels of tracking in use, a per-txn list and a per-env list.
5818 * ref'ing and unref'ing the per-txn list is faster since it requires no
5819 * locking. Pages are cached in the per-env list for global reuse, and a lock
5820 * is required. Pages are not immediately unmapped when their refcnt goes to
5821 * zero; they hang around in case they will be reused again soon.
5823 * When the per-txn list gets full, all pages with refcnt=0 are purged from the
5824 * list and their refcnts in the per-env list are decremented.
5826 * When the per-env list gets full, all pages with refcnt=0 are purged from the
5827 * list and their pages are unmapped.
5829 * @note "full" means the list has reached its respective rpcheck threshold.
5830 * This threshold slowly raises if no pages could be purged on a given check,
5831 * and returns to its original value when enough pages were purged.
5833 * If purging doesn't free any slots, filling the per-txn list will return
5834 * MDB_TXN_FULL, and filling the per-env list returns MDB_MAP_FULL.
5836 * Reference tracking in a txn is imperfect, pages can linger with non-zero
5837 * refcnt even without active references. It was deemed to be too invasive
5838 * to add unrefs in every required location. However, all pages are unref'd
5839 * at the end of the transaction. This guarantees that no stale references
5840 * linger in the per-env list.
5842 * Usually we map chunks of 16 pages at a time, but if an overflow page begins
5843 * at the tail of the chunk we extend the chunk to include the entire overflow
5844 * page. Unfortunately, pages can be turned into overflow pages after their
5845 * chunk was already mapped. In that case we must remap the chunk if the
5846 * overflow page is referenced. If the chunk's refcnt is 0 we can just remap
5847 * it, otherwise we temporarily map a new chunk just for the overflow page.
5849 * @note this chunk handling means we cannot guarantee that a data item
5850 * returned from the DB will stay alive for the duration of the transaction:
5851 * We unref pages as soon as a cursor moves away from the page
5852 * A subsequent op may cause a purge, which may unmap any unref'd chunks
5853 * The caller must copy the data if it must be used later in the same txn.
5855 * Also - our reference counting revolves around cursors, but overflow pages
5856 * aren't pointed to by a cursor's page stack. We have to remember them
5857 * explicitly, in the added mc_ovpg field. A single cursor can only hold a
5858 * reference to one overflow page at a time.
5860 * @param[in] txn the transaction for this access.
5861 * @param[in] pgno the page number for the page to retrieve.
5862 * @param[out] ret address of a pointer where the page's address will be stored.
5863 * @return 0 on success, non-zero on failure.
5866 mdb_rpage_get(MDB_txn *txn, pgno_t pg0, MDB_page **ret)
5868 MDB_env *env = txn->mt_env;
5870 MDB_ID3L tl = txn->mt_rpages;
5871 MDB_ID3L el = env->me_rpages;
5875 int rc, retries = 1;
5879 #define SET_OFF(off,val) off.QuadPart = val
5880 #define MAP(rc,env,addr,len,off) \
5882 rc = NtMapViewOfSection(env->me_fmh, GetCurrentProcess(), &addr, 0, \
5883 len, &off, &len, ViewUnmap, (env->me_flags & MDB_RDONLY) ? 0 : MEM_RESERVE, PAGE_READONLY); \
5884 if (rc) rc = mdb_nt2win32(rc)
5888 #define SET_OFF(off,val) off = val
5889 #define MAP(rc,env,addr,len,off) \
5890 addr = mmap(NULL, len, PROT_READ, MAP_SHARED, env->me_fd, off); \
5891 rc = (addr == MAP_FAILED) ? errno : 0
5894 /* remember the offset of the actual page number, so we can
5895 * return the correct pointer at the end.
5897 rem = pg0 & (MDB_RPAGE_CHUNK-1);
5901 x = mdb_mid3l_search(tl, pgno);
5902 if (x <= tl[0].mid && tl[x].mid == pgno) {
5903 if (x != tl[0].mid && tl[x+1].mid == pg0)
5905 /* check for overflow size */
5906 p = (MDB_page *)((char *)tl[x].mptr + rem * env->me_psize);
5907 if (IS_OVERFLOW(p) && p->mp_pages + rem > tl[x].mcnt) {
5908 id3.mcnt = p->mp_pages + rem;
5909 len = id3.mcnt * env->me_psize;
5910 SET_OFF(off, pgno * env->me_psize);
5911 MAP(rc, env, id3.mptr, len, off);
5914 /* check for local-only page */
5916 mdb_tassert(txn, tl[x].mid != pg0);
5917 /* hope there's room to insert this locally.
5918 * setting mid here tells later code to just insert
5919 * this id3 instead of searching for a match.
5924 /* ignore the mapping we got from env, use new one */
5925 tl[x].mptr = id3.mptr;
5926 tl[x].mcnt = id3.mcnt;
5927 /* if no active ref, see if we can replace in env */
5930 pthread_mutex_lock(&env->me_rpmutex);
5931 i = mdb_mid3l_search(el, tl[x].mid);
5932 if (el[i].mref == 1) {
5933 /* just us, replace it */
5934 munmap(el[i].mptr, el[i].mcnt * env->me_psize);
5935 el[i].mptr = tl[x].mptr;
5936 el[i].mcnt = tl[x].mcnt;
5938 /* there are others, remove ourself */
5941 pthread_mutex_unlock(&env->me_rpmutex);
5945 id3.mptr = tl[x].mptr;
5946 id3.mcnt = tl[x].mcnt;
5952 if (tl[0].mid >= MDB_TRPAGE_MAX - txn->mt_rpcheck) {
5954 /* purge unref'd pages from our list and unref in env */
5955 pthread_mutex_lock(&env->me_rpmutex);
5958 for (i=1; i<=tl[0].mid; i++) {
5961 /* tmp overflow pages don't go to env */
5962 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
5963 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
5966 x = mdb_mid3l_search(el, tl[i].mid);
5970 pthread_mutex_unlock(&env->me_rpmutex);
5972 /* we didn't find any unref'd chunks.
5973 * if we're out of room, fail.
5975 if (tl[0].mid >= MDB_TRPAGE_MAX)
5976 return MDB_TXN_FULL;
5977 /* otherwise, raise threshold for next time around
5980 txn->mt_rpcheck /= 2;
5982 /* we found some unused; consolidate the list */
5983 for (i=y+1; i<= tl[0].mid; i++)
5987 /* decrease the check threshold toward its original value */
5988 if (!txn->mt_rpcheck)
5989 txn->mt_rpcheck = 1;
5990 while (txn->mt_rpcheck < tl[0].mid && txn->mt_rpcheck < MDB_TRPAGE_SIZE/2)
5991 txn->mt_rpcheck *= 2;
5994 if (tl[0].mid < MDB_TRPAGE_SIZE) {
5998 /* don't map past last written page in read-only envs */
5999 if ((env->me_flags & MDB_RDONLY) && pgno + MDB_RPAGE_CHUNK-1 > txn->mt_last_pgno)
6000 id3.mcnt = txn->mt_last_pgno + 1 - pgno;
6002 id3.mcnt = MDB_RPAGE_CHUNK;
6003 len = id3.mcnt * env->me_psize;
6006 /* search for page in env */
6007 pthread_mutex_lock(&env->me_rpmutex);
6008 x = mdb_mid3l_search(el, pgno);
6009 if (x <= el[0].mid && el[x].mid == pgno) {
6010 id3.mptr = el[x].mptr;
6011 id3.mcnt = el[x].mcnt;
6012 /* check for overflow size */
6013 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
6014 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
6015 id3.mcnt = p->mp_pages + rem;
6016 len = id3.mcnt * env->me_psize;
6017 SET_OFF(off, pgno * env->me_psize);
6018 MAP(rc, env, id3.mptr, len, off);
6022 munmap(el[x].mptr, env->me_psize * el[x].mcnt);
6023 el[x].mptr = id3.mptr;
6024 el[x].mcnt = id3.mcnt;
6027 pthread_mutex_unlock(&env->me_rpmutex);
6032 pthread_mutex_unlock(&env->me_rpmutex);
6035 if (el[0].mid >= MDB_ERPAGE_MAX - env->me_rpcheck) {
6036 /* purge unref'd pages */
6038 for (i=1; i<=el[0].mid; i++) {
6041 munmap(el[i].mptr, env->me_psize * el[i].mcnt);
6046 /* see if we can unref some local pages */
6051 if (el[0].mid >= MDB_ERPAGE_MAX) {
6052 pthread_mutex_unlock(&env->me_rpmutex);
6053 return MDB_MAP_FULL;
6055 env->me_rpcheck /= 2;
6057 for (i=y+1; i<= el[0].mid; i++)
6061 if (!env->me_rpcheck)
6062 env->me_rpcheck = 1;
6063 while (env->me_rpcheck < el[0].mid && env->me_rpcheck < MDB_ERPAGE_SIZE/2)
6064 env->me_rpcheck *= 2;
6067 SET_OFF(off, pgno * env->me_psize);
6068 MAP(rc, env, id3.mptr, len, off);
6071 pthread_mutex_unlock(&env->me_rpmutex);
6074 /* check for overflow size */
6075 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
6076 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
6077 id3.mcnt = p->mp_pages + rem;
6078 munmap(id3.mptr, len);
6079 len = id3.mcnt * env->me_psize;
6080 MAP(rc, env, id3.mptr, len, off);
6084 mdb_mid3l_insert(el, &id3);
6085 pthread_mutex_unlock(&env->me_rpmutex);
6087 mdb_mid3l_insert(tl, &id3);
6089 return MDB_TXN_FULL;
6092 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
6093 #if MDB_DEBUG /* we don't need this check any more */
6094 if (IS_OVERFLOW(p)) {
6095 mdb_tassert(txn, p->mp_pages + rem <= id3.mcnt);
6103 /** Find the address of the page corresponding to a given page number.
6104 * Set #MDB_TXN_ERROR on failure.
6105 * @param[in] mc the cursor accessing the page.
6106 * @param[in] pgno the page number for the page to retrieve.
6107 * @param[out] ret address of a pointer where the page's address will be stored.
6108 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
6109 * @return 0 on success, non-zero on failure.
6112 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
6114 MDB_txn *txn = mc->mc_txn;
6118 if (! (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP))) {
6122 MDB_ID2L dl = tx2->mt_u.dirty_list;
6124 /* Spilled pages were dirtied in this txn and flushed
6125 * because the dirty list got full. Bring this page
6126 * back in from the map (but don't unspill it here,
6127 * leave that unless page_touch happens again).
6129 if (tx2->mt_spill_pgs) {
6130 MDB_ID pn = pgno << 1;
6131 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
6132 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
6137 unsigned x = mdb_mid2l_search(dl, pgno);
6138 if (x <= dl[0].mid && dl[x].mid == pgno) {
6144 } while ((tx2 = tx2->mt_parent) != NULL);
6147 if (pgno >= txn->mt_next_pgno) {
6148 DPRINTF(("page %"Yu" not found", pgno));
6149 txn->mt_flags |= MDB_TXN_ERROR;
6150 return MDB_PAGE_NOTFOUND;
6158 int rc = mdb_rpage_get(txn, pgno, &p);
6160 txn->mt_flags |= MDB_TXN_ERROR;
6164 MDB_env *env = txn->mt_env;
6165 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
6176 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
6177 * The cursor is at the root page, set up the rest of it.
6180 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
6182 MDB_page *mp = mc->mc_pg[mc->mc_top];
6186 while (IS_BRANCH(mp)) {
6190 DPRINTF(("branch page %"Yu" has %u keys", mp->mp_pgno, NUMKEYS(mp)));
6191 /* Don't assert on branch pages in the FreeDB. We can get here
6192 * while in the process of rebalancing a FreeDB branch page; we must
6193 * let that proceed. ITS#8336
6195 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
6196 DPRINTF(("found index 0 to page %"Yu, NODEPGNO(NODEPTR(mp, 0))));
6198 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
6200 if (flags & MDB_PS_LAST)
6201 i = NUMKEYS(mp) - 1;
6204 node = mdb_node_search(mc, key, &exact);
6206 i = NUMKEYS(mp) - 1;
6208 i = mc->mc_ki[mc->mc_top];
6210 mdb_cassert(mc, i > 0);
6214 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
6217 mdb_cassert(mc, i < NUMKEYS(mp));
6218 node = NODEPTR(mp, i);
6220 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6223 mc->mc_ki[mc->mc_top] = i;
6224 if ((rc = mdb_cursor_push(mc, mp)))
6227 if (flags & MDB_PS_MODIFY) {
6228 if ((rc = mdb_page_touch(mc)) != 0)
6230 mp = mc->mc_pg[mc->mc_top];
6235 DPRINTF(("internal error, index points to a %02X page!?",
6237 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6238 return MDB_CORRUPTED;
6241 DPRINTF(("found leaf page %"Yu" for key [%s]", mp->mp_pgno,
6242 key ? DKEY(key) : "null"));
6243 mc->mc_flags |= C_INITIALIZED;
6244 mc->mc_flags &= ~C_EOF;
6249 /** Search for the lowest key under the current branch page.
6250 * This just bypasses a NUMKEYS check in the current page
6251 * before calling mdb_page_search_root(), because the callers
6252 * are all in situations where the current page is known to
6256 mdb_page_search_lowest(MDB_cursor *mc)
6258 MDB_page *mp = mc->mc_pg[mc->mc_top];
6259 MDB_node *node = NODEPTR(mp, 0);
6262 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6265 mc->mc_ki[mc->mc_top] = 0;
6266 if ((rc = mdb_cursor_push(mc, mp)))
6268 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
6271 /** Search for the page a given key should be in.
6272 * Push it and its parent pages on the cursor stack.
6273 * @param[in,out] mc the cursor for this operation.
6274 * @param[in] key the key to search for, or NULL for first/last page.
6275 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
6276 * are touched (updated with new page numbers).
6277 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
6278 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
6279 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
6280 * @return 0 on success, non-zero on failure.
6283 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
6288 /* Make sure the txn is still viable, then find the root from
6289 * the txn's db table and set it as the root of the cursor's stack.
6291 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
6292 DPUTS("transaction may not be used now");
6295 /* Make sure we're using an up-to-date root */
6296 if (*mc->mc_dbflag & DB_STALE) {
6298 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6300 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
6301 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
6308 MDB_node *leaf = mdb_node_search(&mc2,
6309 &mc->mc_dbx->md_name, &exact);
6311 return MDB_NOTFOUND;
6312 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
6313 return MDB_INCOMPATIBLE; /* not a named DB */
6314 rc = mdb_node_read(&mc2, leaf, &data);
6317 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
6319 /* The txn may not know this DBI, or another process may
6320 * have dropped and recreated the DB with other flags.
6322 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
6323 return MDB_INCOMPATIBLE;
6324 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
6326 *mc->mc_dbflag &= ~DB_STALE;
6328 root = mc->mc_db->md_root;
6330 if (root == P_INVALID) { /* Tree is empty. */
6331 DPUTS("tree is empty");
6332 return MDB_NOTFOUND;
6336 mdb_cassert(mc, root > 1);
6337 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) {
6340 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[0]);
6342 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
6349 for (i=1; i<mc->mc_snum; i++)
6350 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[i]);
6356 DPRINTF(("db %d root page %"Yu" has flags 0x%X",
6357 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
6359 if (flags & MDB_PS_MODIFY) {
6360 if ((rc = mdb_page_touch(mc)))
6364 if (flags & MDB_PS_ROOTONLY)
6367 return mdb_page_search_root(mc, key, flags);
6371 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
6373 MDB_txn *txn = mc->mc_txn;
6374 pgno_t pg = mp->mp_pgno;
6375 unsigned x = 0, ovpages = mp->mp_pages;
6376 MDB_env *env = txn->mt_env;
6377 MDB_IDL sl = txn->mt_spill_pgs;
6378 MDB_ID pn = pg << 1;
6381 DPRINTF(("free ov page %"Yu" (%d)", pg, ovpages));
6382 /* If the page is dirty or on the spill list we just acquired it,
6383 * so we should give it back to our current free list, if any.
6384 * Otherwise put it onto the list of pages we freed in this txn.
6386 * Won't create me_pghead: me_pglast must be inited along with it.
6387 * Unsupported in nested txns: They would need to hide the page
6388 * range in ancestor txns' dirty and spilled lists.
6390 if (env->me_pghead &&
6392 ((mp->mp_flags & P_DIRTY) ||
6393 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
6397 MDB_ID2 *dl, ix, iy;
6398 rc = mdb_midl_need(&env->me_pghead, ovpages);
6401 if (!(mp->mp_flags & P_DIRTY)) {
6402 /* This page is no longer spilled */
6409 /* Remove from dirty list */
6410 dl = txn->mt_u.dirty_list;
6412 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
6418 mdb_cassert(mc, x > 1);
6420 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
6421 txn->mt_flags |= MDB_TXN_ERROR;
6425 txn->mt_dirty_room++;
6426 if (!(env->me_flags & MDB_WRITEMAP))
6427 mdb_dpage_free(env, mp);
6429 /* Insert in me_pghead */
6430 mop = env->me_pghead;
6431 j = mop[0] + ovpages;
6432 for (i = mop[0]; i && mop[i] < pg; i--)
6438 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
6442 mc->mc_db->md_overflow_pages -= ovpages;
6446 /** Return the data associated with a given node.
6447 * @param[in] mc The cursor for this operation.
6448 * @param[in] leaf The node being read.
6449 * @param[out] data Updated to point to the node's data.
6450 * @return 0 on success, non-zero on failure.
6453 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
6455 MDB_page *omp; /* overflow page */
6460 MDB_PAGE_UNREF(mc->mc_txn, MC_OVPG(mc));
6461 MC_SET_OVPG(mc, NULL);
6463 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6464 data->mv_size = NODEDSZ(leaf);
6465 data->mv_data = NODEDATA(leaf);
6469 /* Read overflow data.
6471 data->mv_size = NODEDSZ(leaf);
6472 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
6473 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
6474 DPRINTF(("read overflow page %"Yu" failed", pgno));
6477 data->mv_data = METADATA(omp);
6478 MC_SET_OVPG(mc, omp);
6484 mdb_get(MDB_txn *txn, MDB_dbi dbi,
6485 MDB_val *key, MDB_val *data)
6492 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
6494 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
6497 if (txn->mt_flags & MDB_TXN_BLOCKED)
6500 mdb_cursor_init(&mc, txn, dbi, &mx);
6501 rc = mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
6502 /* unref all the pages when MDB_VL32 - caller must copy the data
6503 * before doing anything else
6505 MDB_CURSOR_UNREF(&mc, 1);
6509 /** Find a sibling for a page.
6510 * Replaces the page at the top of the cursor's stack with the
6511 * specified sibling, if one exists.
6512 * @param[in] mc The cursor for this operation.
6513 * @param[in] move_right Non-zero if the right sibling is requested,
6514 * otherwise the left sibling.
6515 * @return 0 on success, non-zero on failure.
6518 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
6527 if (mc->mc_snum < 2) {
6528 return MDB_NOTFOUND; /* root has no siblings */
6532 op = mc->mc_pg[mc->mc_top];
6535 DPRINTF(("parent page is page %"Yu", index %u",
6536 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
6538 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6539 : (mc->mc_ki[mc->mc_top] == 0)) {
6540 DPRINTF(("no more keys left, moving to %s sibling",
6541 move_right ? "right" : "left"));
6542 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
6543 /* undo cursor_pop before returning */
6550 mc->mc_ki[mc->mc_top]++;
6552 mc->mc_ki[mc->mc_top]--;
6553 DPRINTF(("just moving to %s index key %u",
6554 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
6556 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
6558 MDB_PAGE_UNREF(mc->mc_txn, op);
6560 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6561 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
6562 /* mc will be inconsistent if caller does mc_snum++ as above */
6563 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
6567 mdb_cursor_push(mc, mp);
6569 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
6574 /** Move the cursor to the next data item. */
6576 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6582 if ((mc->mc_flags & C_EOF) ||
6583 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
6584 return MDB_NOTFOUND;
6586 if (!(mc->mc_flags & C_INITIALIZED))
6587 return mdb_cursor_first(mc, key, data);
6589 mp = mc->mc_pg[mc->mc_top];
6591 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6592 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6593 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6594 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
6595 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
6596 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
6597 if (rc == MDB_SUCCESS)
6598 MDB_GET_KEY(leaf, key);
6603 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6606 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6607 if (op == MDB_NEXT_DUP)
6608 return MDB_NOTFOUND;
6612 DPRINTF(("cursor_next: top page is %"Yu" in cursor %p",
6613 mdb_dbg_pgno(mp), (void *) mc));
6614 if (mc->mc_flags & C_DEL) {
6615 mc->mc_flags ^= C_DEL;
6619 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
6620 DPUTS("=====> move to next sibling page");
6621 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6622 mc->mc_flags |= C_EOF;
6625 mp = mc->mc_pg[mc->mc_top];
6626 DPRINTF(("next page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6628 mc->mc_ki[mc->mc_top]++;
6631 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6632 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6635 key->mv_size = mc->mc_db->md_pad;
6636 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6640 mdb_cassert(mc, IS_LEAF(mp));
6641 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6643 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6644 mdb_xcursor_init1(mc, leaf);
6647 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6650 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6651 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6652 if (rc != MDB_SUCCESS)
6657 MDB_GET_KEY(leaf, key);
6661 /** Move the cursor to the previous data item. */
6663 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6669 if (!(mc->mc_flags & C_INITIALIZED)) {
6670 rc = mdb_cursor_last(mc, key, data);
6673 mc->mc_ki[mc->mc_top]++;
6676 mp = mc->mc_pg[mc->mc_top];
6678 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6679 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6680 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6681 if (op == MDB_PREV || op == MDB_PREV_DUP) {
6682 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
6683 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
6684 if (rc == MDB_SUCCESS) {
6685 MDB_GET_KEY(leaf, key);
6686 mc->mc_flags &= ~C_EOF;
6692 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6695 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6696 if (op == MDB_PREV_DUP)
6697 return MDB_NOTFOUND;
6701 DPRINTF(("cursor_prev: top page is %"Yu" in cursor %p",
6702 mdb_dbg_pgno(mp), (void *) mc));
6704 mc->mc_flags &= ~(C_EOF|C_DEL);
6706 if (mc->mc_ki[mc->mc_top] == 0) {
6707 DPUTS("=====> move to prev sibling page");
6708 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
6711 mp = mc->mc_pg[mc->mc_top];
6712 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
6713 DPRINTF(("prev page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6715 mc->mc_ki[mc->mc_top]--;
6717 mc->mc_flags &= ~C_EOF;
6719 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6720 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6723 key->mv_size = mc->mc_db->md_pad;
6724 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6728 mdb_cassert(mc, IS_LEAF(mp));
6729 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6731 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6732 mdb_xcursor_init1(mc, leaf);
6735 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6738 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6739 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6740 if (rc != MDB_SUCCESS)
6745 MDB_GET_KEY(leaf, key);
6749 /** Set the cursor on a specific data item. */
6751 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6752 MDB_cursor_op op, int *exactp)
6756 MDB_node *leaf = NULL;
6759 if (key->mv_size == 0)
6760 return MDB_BAD_VALSIZE;
6762 if (mc->mc_xcursor) {
6763 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6764 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6767 /* See if we're already on the right page */
6768 if (mc->mc_flags & C_INITIALIZED) {
6771 mp = mc->mc_pg[mc->mc_top];
6773 mc->mc_ki[mc->mc_top] = 0;
6774 return MDB_NOTFOUND;
6776 if (mp->mp_flags & P_LEAF2) {
6777 nodekey.mv_size = mc->mc_db->md_pad;
6778 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
6780 leaf = NODEPTR(mp, 0);
6781 MDB_GET_KEY2(leaf, nodekey);
6783 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6785 /* Probably happens rarely, but first node on the page
6786 * was the one we wanted.
6788 mc->mc_ki[mc->mc_top] = 0;
6795 unsigned int nkeys = NUMKEYS(mp);
6797 if (mp->mp_flags & P_LEAF2) {
6798 nodekey.mv_data = LEAF2KEY(mp,
6799 nkeys-1, nodekey.mv_size);
6801 leaf = NODEPTR(mp, nkeys-1);
6802 MDB_GET_KEY2(leaf, nodekey);
6804 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6806 /* last node was the one we wanted */
6807 mc->mc_ki[mc->mc_top] = nkeys-1;
6813 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6814 /* This is definitely the right page, skip search_page */
6815 if (mp->mp_flags & P_LEAF2) {
6816 nodekey.mv_data = LEAF2KEY(mp,
6817 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6819 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6820 MDB_GET_KEY2(leaf, nodekey);
6822 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6824 /* current node was the one we wanted */
6831 mc->mc_flags &= ~C_EOF;
6835 /* If any parents have right-sibs, search.
6836 * Otherwise, there's nothing further.
6838 for (i=0; i<mc->mc_top; i++)
6840 NUMKEYS(mc->mc_pg[i])-1)
6842 if (i == mc->mc_top) {
6843 /* There are no other pages */
6844 mc->mc_ki[mc->mc_top] = nkeys;
6845 return MDB_NOTFOUND;
6849 /* There are no other pages */
6850 mc->mc_ki[mc->mc_top] = 0;
6851 if (op == MDB_SET_RANGE && !exactp) {
6855 return MDB_NOTFOUND;
6861 rc = mdb_page_search(mc, key, 0);
6862 if (rc != MDB_SUCCESS)
6865 mp = mc->mc_pg[mc->mc_top];
6866 mdb_cassert(mc, IS_LEAF(mp));
6869 leaf = mdb_node_search(mc, key, exactp);
6870 if (exactp != NULL && !*exactp) {
6871 /* MDB_SET specified and not an exact match. */
6872 return MDB_NOTFOUND;
6876 DPUTS("===> inexact leaf not found, goto sibling");
6877 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6878 mc->mc_flags |= C_EOF;
6879 return rc; /* no entries matched */
6881 mp = mc->mc_pg[mc->mc_top];
6882 mdb_cassert(mc, IS_LEAF(mp));
6883 leaf = NODEPTR(mp, 0);
6887 mc->mc_flags |= C_INITIALIZED;
6888 mc->mc_flags &= ~C_EOF;
6891 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6892 key->mv_size = mc->mc_db->md_pad;
6893 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6898 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6899 mdb_xcursor_init1(mc, leaf);
6902 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6903 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6904 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6907 if (op == MDB_GET_BOTH) {
6913 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6914 if (rc != MDB_SUCCESS)
6917 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6920 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6922 dcmp = mc->mc_dbx->md_dcmp;
6923 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
6924 dcmp = mdb_cmp_clong;
6925 rc = dcmp(data, &olddata);
6927 if (op == MDB_GET_BOTH || rc > 0)
6928 return MDB_NOTFOUND;
6935 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6936 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6941 /* The key already matches in all other cases */
6942 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6943 MDB_GET_KEY(leaf, key);
6944 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6949 /** Move the cursor to the first item in the database. */
6951 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6956 if (mc->mc_xcursor) {
6957 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6958 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6961 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6962 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6963 if (rc != MDB_SUCCESS)
6966 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6968 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6969 mc->mc_flags |= C_INITIALIZED;
6970 mc->mc_flags &= ~C_EOF;
6972 mc->mc_ki[mc->mc_top] = 0;
6974 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6975 key->mv_size = mc->mc_db->md_pad;
6976 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6981 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6982 mdb_xcursor_init1(mc, leaf);
6983 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6987 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6991 MDB_GET_KEY(leaf, key);
6995 /** Move the cursor to the last item in the database. */
6997 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
7002 if (mc->mc_xcursor) {
7003 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
7004 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
7007 if (!(mc->mc_flags & C_EOF)) {
7009 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
7010 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
7011 if (rc != MDB_SUCCESS)
7014 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
7017 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
7018 mc->mc_flags |= C_INITIALIZED|C_EOF;
7019 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7021 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
7022 key->mv_size = mc->mc_db->md_pad;
7023 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
7028 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7029 mdb_xcursor_init1(mc, leaf);
7030 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
7034 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
7039 MDB_GET_KEY(leaf, key);
7044 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
7049 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
7054 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7058 case MDB_GET_CURRENT:
7059 if (!(mc->mc_flags & C_INITIALIZED)) {
7062 MDB_page *mp = mc->mc_pg[mc->mc_top];
7063 int nkeys = NUMKEYS(mp);
7064 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
7065 mc->mc_ki[mc->mc_top] = nkeys;
7071 key->mv_size = mc->mc_db->md_pad;
7072 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
7074 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7075 MDB_GET_KEY(leaf, key);
7077 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7078 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
7080 rc = mdb_node_read(mc, leaf, data);
7087 case MDB_GET_BOTH_RANGE:
7092 if (mc->mc_xcursor == NULL) {
7093 rc = MDB_INCOMPATIBLE;
7103 rc = mdb_cursor_set(mc, key, data, op,
7104 op == MDB_SET_RANGE ? NULL : &exact);
7107 case MDB_GET_MULTIPLE:
7108 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
7112 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7113 rc = MDB_INCOMPATIBLE;
7117 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
7118 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
7121 case MDB_NEXT_MULTIPLE:
7126 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7127 rc = MDB_INCOMPATIBLE;
7130 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
7131 if (rc == MDB_SUCCESS) {
7132 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
7135 mx = &mc->mc_xcursor->mx_cursor;
7136 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
7138 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
7139 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
7145 case MDB_PREV_MULTIPLE:
7150 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7151 rc = MDB_INCOMPATIBLE;
7154 if (!(mc->mc_flags & C_INITIALIZED))
7155 rc = mdb_cursor_last(mc, key, data);
7158 if (rc == MDB_SUCCESS) {
7159 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
7160 if (mx->mc_flags & C_INITIALIZED) {
7161 rc = mdb_cursor_sibling(mx, 0);
7162 if (rc == MDB_SUCCESS)
7171 case MDB_NEXT_NODUP:
7172 rc = mdb_cursor_next(mc, key, data, op);
7176 case MDB_PREV_NODUP:
7177 rc = mdb_cursor_prev(mc, key, data, op);
7180 rc = mdb_cursor_first(mc, key, data);
7183 mfunc = mdb_cursor_first;
7185 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
7189 if (mc->mc_xcursor == NULL) {
7190 rc = MDB_INCOMPATIBLE;
7194 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7195 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7196 MDB_GET_KEY(leaf, key);
7197 rc = mdb_node_read(mc, leaf, data);
7201 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7205 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
7208 rc = mdb_cursor_last(mc, key, data);
7211 mfunc = mdb_cursor_last;
7214 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
7219 if (mc->mc_flags & C_DEL)
7220 mc->mc_flags ^= C_DEL;
7225 /** Touch all the pages in the cursor stack. Set mc_top.
7226 * Makes sure all the pages are writable, before attempting a write operation.
7227 * @param[in] mc The cursor to operate on.
7230 mdb_cursor_touch(MDB_cursor *mc)
7232 int rc = MDB_SUCCESS;
7234 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & (DB_DIRTY|DB_DUPDATA))) {
7235 /* Touch DB record of named DB */
7238 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
7240 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
7241 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
7244 *mc->mc_dbflag |= DB_DIRTY;
7249 rc = mdb_page_touch(mc);
7250 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
7251 mc->mc_top = mc->mc_snum-1;
7256 /** Do not spill pages to disk if txn is getting full, may fail instead */
7257 #define MDB_NOSPILL 0x8000
7260 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
7264 MDB_node *leaf = NULL;
7265 MDB_page *fp, *mp, *sub_root = NULL;
7267 MDB_val xdata, *rdata, dkey, olddata;
7269 int do_sub = 0, insert_key, insert_data;
7270 unsigned int mcount = 0, dcount = 0, nospill;
7273 unsigned int nflags;
7276 if (mc == NULL || key == NULL)
7279 env = mc->mc_txn->mt_env;
7281 /* Check this first so counter will always be zero on any
7284 if (flags & MDB_MULTIPLE) {
7285 dcount = data[1].mv_size;
7286 data[1].mv_size = 0;
7287 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
7288 return MDB_INCOMPATIBLE;
7291 nospill = flags & MDB_NOSPILL;
7292 flags &= ~MDB_NOSPILL;
7294 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7295 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7297 if (key->mv_size-1 >= ENV_MAXKEY(env))
7298 return MDB_BAD_VALSIZE;
7300 #if SIZE_MAX > MAXDATASIZE
7301 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
7302 return MDB_BAD_VALSIZE;
7304 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
7305 return MDB_BAD_VALSIZE;
7308 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
7309 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
7313 if (flags == MDB_CURRENT) {
7314 if (!(mc->mc_flags & C_INITIALIZED))
7317 } else if (mc->mc_db->md_root == P_INVALID) {
7318 /* new database, cursor has nothing to point to */
7321 mc->mc_flags &= ~C_INITIALIZED;
7326 if (flags & MDB_APPEND) {
7328 rc = mdb_cursor_last(mc, &k2, &d2);
7330 rc = mc->mc_dbx->md_cmp(key, &k2);
7333 mc->mc_ki[mc->mc_top]++;
7335 /* new key is <= last key */
7340 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
7342 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
7343 DPRINTF(("duplicate key [%s]", DKEY(key)));
7345 return MDB_KEYEXIST;
7347 if (rc && rc != MDB_NOTFOUND)
7351 if (mc->mc_flags & C_DEL)
7352 mc->mc_flags ^= C_DEL;
7354 /* Cursor is positioned, check for room in the dirty list */
7356 if (flags & MDB_MULTIPLE) {
7358 xdata.mv_size = data->mv_size * dcount;
7362 if ((rc2 = mdb_page_spill(mc, key, rdata)))
7366 if (rc == MDB_NO_ROOT) {
7368 /* new database, write a root leaf page */
7369 DPUTS("allocating new root leaf page");
7370 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
7373 mdb_cursor_push(mc, np);
7374 mc->mc_db->md_root = np->mp_pgno;
7375 mc->mc_db->md_depth++;
7376 *mc->mc_dbflag |= DB_DIRTY;
7377 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
7379 np->mp_flags |= P_LEAF2;
7380 mc->mc_flags |= C_INITIALIZED;
7382 /* make sure all cursor pages are writable */
7383 rc2 = mdb_cursor_touch(mc);
7388 insert_key = insert_data = rc;
7390 /* The key does not exist */
7391 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
7392 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
7393 LEAFSIZE(key, data) > env->me_nodemax)
7395 /* Too big for a node, insert in sub-DB. Set up an empty
7396 * "old sub-page" for prep_subDB to expand to a full page.
7398 fp_flags = P_LEAF|P_DIRTY;
7400 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
7401 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
7402 olddata.mv_size = PAGEHDRSZ;
7406 /* there's only a key anyway, so this is a no-op */
7407 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
7409 unsigned int ksize = mc->mc_db->md_pad;
7410 if (key->mv_size != ksize)
7411 return MDB_BAD_VALSIZE;
7412 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
7413 memcpy(ptr, key->mv_data, ksize);
7415 /* if overwriting slot 0 of leaf, need to
7416 * update branch key if there is a parent page
7418 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7419 unsigned short dtop = 1;
7421 /* slot 0 is always an empty key, find real slot */
7422 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7426 if (mc->mc_ki[mc->mc_top])
7427 rc2 = mdb_update_key(mc, key);
7438 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7439 olddata.mv_size = NODEDSZ(leaf);
7440 olddata.mv_data = NODEDATA(leaf);
7443 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
7444 /* Prepare (sub-)page/sub-DB to accept the new item,
7445 * if needed. fp: old sub-page or a header faking
7446 * it. mp: new (sub-)page. offset: growth in page
7447 * size. xdata: node data with new page or DB.
7449 unsigned i, offset = 0;
7450 mp = fp = xdata.mv_data = env->me_pbuf;
7451 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
7453 /* Was a single item before, must convert now */
7454 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7456 /* Just overwrite the current item */
7457 if (flags == MDB_CURRENT)
7459 dcmp = mc->mc_dbx->md_dcmp;
7460 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
7461 dcmp = mdb_cmp_clong;
7462 /* does data match? */
7463 if (!dcmp(data, &olddata)) {
7464 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
7465 return MDB_KEYEXIST;
7470 /* Back up original data item */
7471 dkey.mv_size = olddata.mv_size;
7472 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
7474 /* Make sub-page header for the dup items, with dummy body */
7475 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
7476 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
7477 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
7478 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7479 fp->mp_flags |= P_LEAF2;
7480 fp->mp_pad = data->mv_size;
7481 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
7483 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
7484 (dkey.mv_size & 1) + (data->mv_size & 1);
7486 fp->mp_upper = xdata.mv_size - PAGEBASE;
7487 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
7488 } else if (leaf->mn_flags & F_SUBDATA) {
7489 /* Data is on sub-DB, just store it */
7490 flags |= F_DUPDATA|F_SUBDATA;
7493 /* Data is on sub-page */
7494 fp = olddata.mv_data;
7497 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7498 offset = EVEN(NODESIZE + sizeof(indx_t) +
7502 offset = fp->mp_pad;
7503 if (SIZELEFT(fp) < offset) {
7504 offset *= 4; /* space for 4 more */
7507 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
7509 fp->mp_flags |= P_DIRTY;
7510 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
7511 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
7515 xdata.mv_size = olddata.mv_size + offset;
7518 fp_flags = fp->mp_flags;
7519 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
7520 /* Too big for a sub-page, convert to sub-DB */
7521 fp_flags &= ~P_SUBP;
7523 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7524 fp_flags |= P_LEAF2;
7525 dummy.md_pad = fp->mp_pad;
7526 dummy.md_flags = MDB_DUPFIXED;
7527 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7528 dummy.md_flags |= MDB_INTEGERKEY;
7534 dummy.md_branch_pages = 0;
7535 dummy.md_leaf_pages = 1;
7536 dummy.md_overflow_pages = 0;
7537 dummy.md_entries = NUMKEYS(fp);
7538 xdata.mv_size = sizeof(MDB_db);
7539 xdata.mv_data = &dummy;
7540 if ((rc = mdb_page_alloc(mc, 1, &mp)))
7542 offset = env->me_psize - olddata.mv_size;
7543 flags |= F_DUPDATA|F_SUBDATA;
7544 dummy.md_root = mp->mp_pgno;
7548 mp->mp_flags = fp_flags | P_DIRTY;
7549 mp->mp_pad = fp->mp_pad;
7550 mp->mp_lower = fp->mp_lower;
7551 mp->mp_upper = fp->mp_upper + offset;
7552 if (fp_flags & P_LEAF2) {
7553 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
7555 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
7556 olddata.mv_size - fp->mp_upper - PAGEBASE);
7557 for (i=0; i<NUMKEYS(fp); i++)
7558 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
7566 mdb_node_del(mc, 0);
7570 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
7571 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
7572 return MDB_INCOMPATIBLE;
7573 /* overflow page overwrites need special handling */
7574 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7577 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
7579 memcpy(&pg, olddata.mv_data, sizeof(pg));
7580 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
7582 ovpages = omp->mp_pages;
7584 /* Is the ov page large enough? */
7585 if (ovpages >= dpages) {
7586 if (!(omp->mp_flags & P_DIRTY) &&
7587 (level || (env->me_flags & MDB_WRITEMAP)))
7589 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
7592 level = 0; /* dirty in this txn or clean */
7595 if (omp->mp_flags & P_DIRTY) {
7596 /* yes, overwrite it. Note in this case we don't
7597 * bother to try shrinking the page if the new data
7598 * is smaller than the overflow threshold.
7601 /* It is writable only in a parent txn */
7602 size_t sz = (size_t) env->me_psize * ovpages, off;
7603 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
7609 /* Note - this page is already counted in parent's dirty_room */
7610 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
7611 mdb_cassert(mc, rc2 == 0);
7612 /* Currently we make the page look as with put() in the
7613 * parent txn, in case the user peeks at MDB_RESERVEd
7614 * or unused parts. Some users treat ovpages specially.
7616 if (!(flags & MDB_RESERVE)) {
7617 /* Skip the part where LMDB will put *data.
7618 * Copy end of page, adjusting alignment so
7619 * compiler may copy words instead of bytes.
7621 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
7622 memcpy((size_t *)((char *)np + off),
7623 (size_t *)((char *)omp + off), sz - off);
7626 memcpy(np, omp, sz); /* Copy beginning of page */
7629 SETDSZ(leaf, data->mv_size);
7630 if (F_ISSET(flags, MDB_RESERVE))
7631 data->mv_data = METADATA(omp);
7633 memcpy(METADATA(omp), data->mv_data, data->mv_size);
7637 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
7639 } else if (data->mv_size == olddata.mv_size) {
7640 /* same size, just replace it. Note that we could
7641 * also reuse this node if the new data is smaller,
7642 * but instead we opt to shrink the node in that case.
7644 if (F_ISSET(flags, MDB_RESERVE))
7645 data->mv_data = olddata.mv_data;
7646 else if (!(mc->mc_flags & C_SUB))
7647 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
7649 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
7654 mdb_node_del(mc, 0);
7660 nflags = flags & NODE_ADD_FLAGS;
7661 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
7662 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
7663 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
7664 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
7666 nflags |= MDB_SPLIT_REPLACE;
7667 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
7669 /* There is room already in this leaf page. */
7670 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
7672 /* Adjust other cursors pointing to mp */
7673 MDB_cursor *m2, *m3;
7674 MDB_dbi dbi = mc->mc_dbi;
7675 unsigned i = mc->mc_top;
7676 MDB_page *mp = mc->mc_pg[i];
7678 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7679 if (mc->mc_flags & C_SUB)
7680 m3 = &m2->mc_xcursor->mx_cursor;
7683 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
7684 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
7687 if (XCURSOR_INITED(m3))
7688 XCURSOR_REFRESH(m3, mp, m3->mc_ki[i]);
7693 if (rc == MDB_SUCCESS) {
7694 /* Now store the actual data in the child DB. Note that we're
7695 * storing the user data in the keys field, so there are strict
7696 * size limits on dupdata. The actual data fields of the child
7697 * DB are all zero size.
7700 int xflags, new_dupdata;
7705 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7706 if (flags & MDB_CURRENT) {
7707 xflags = MDB_CURRENT|MDB_NOSPILL;
7709 mdb_xcursor_init1(mc, leaf);
7710 xflags = (flags & MDB_NODUPDATA) ?
7711 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
7714 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
7715 new_dupdata = (int)dkey.mv_size;
7716 /* converted, write the original data first */
7718 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
7721 /* we've done our job */
7724 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
7725 /* Adjust other cursors pointing to mp */
7727 MDB_xcursor *mx = mc->mc_xcursor;
7728 unsigned i = mc->mc_top;
7729 MDB_page *mp = mc->mc_pg[i];
7730 int nkeys = NUMKEYS(mp);
7732 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7733 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7734 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7735 if (m2->mc_pg[i] == mp) {
7736 if (m2->mc_ki[i] == mc->mc_ki[i]) {
7737 mdb_xcursor_init2(m2, mx, new_dupdata);
7738 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
7739 XCURSOR_REFRESH(m2, mp, m2->mc_ki[i]);
7744 ecount = mc->mc_xcursor->mx_db.md_entries;
7745 if (flags & MDB_APPENDDUP)
7746 xflags |= MDB_APPEND;
7747 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
7748 if (flags & F_SUBDATA) {
7749 void *db = NODEDATA(leaf);
7750 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7752 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
7754 /* Increment count unless we just replaced an existing item. */
7756 mc->mc_db->md_entries++;
7758 /* Invalidate txn if we created an empty sub-DB */
7761 /* If we succeeded and the key didn't exist before,
7762 * make sure the cursor is marked valid.
7764 mc->mc_flags |= C_INITIALIZED;
7766 if (flags & MDB_MULTIPLE) {
7769 /* let caller know how many succeeded, if any */
7770 data[1].mv_size = mcount;
7771 if (mcount < dcount) {
7772 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
7773 insert_key = insert_data = 0;
7780 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
7783 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7788 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7794 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7795 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7797 if (!(mc->mc_flags & C_INITIALIZED))
7800 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7801 return MDB_NOTFOUND;
7803 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7806 rc = mdb_cursor_touch(mc);
7810 mp = mc->mc_pg[mc->mc_top];
7813 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7815 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7816 if (flags & MDB_NODUPDATA) {
7817 /* mdb_cursor_del0() will subtract the final entry */
7818 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7819 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7821 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7822 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7824 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7827 /* If sub-DB still has entries, we're done */
7828 if (mc->mc_xcursor->mx_db.md_entries) {
7829 if (leaf->mn_flags & F_SUBDATA) {
7830 /* update subDB info */
7831 void *db = NODEDATA(leaf);
7832 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7835 /* shrink fake page */
7836 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7837 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7838 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7839 /* fix other sub-DB cursors pointed at fake pages on this page */
7840 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7841 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7842 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7843 if (m2->mc_pg[mc->mc_top] == mp) {
7844 MDB_node *n2 = leaf;
7845 if (m2->mc_ki[mc->mc_top] != mc->mc_ki[mc->mc_top]) {
7846 n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7847 if (n2->mn_flags & F_SUBDATA) continue;
7849 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7853 mc->mc_db->md_entries--;
7856 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7858 /* otherwise fall thru and delete the sub-DB */
7861 if (leaf->mn_flags & F_SUBDATA) {
7862 /* add all the child DB's pages to the free list */
7863 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7868 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7869 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7870 rc = MDB_INCOMPATIBLE;
7874 /* add overflow pages to free list */
7875 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7879 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7880 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7881 (rc = mdb_ovpage_free(mc, omp)))
7886 return mdb_cursor_del0(mc);
7889 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7893 /** Allocate and initialize new pages for a database.
7894 * Set #MDB_TXN_ERROR on failure.
7895 * @param[in] mc a cursor on the database being added to.
7896 * @param[in] flags flags defining what type of page is being allocated.
7897 * @param[in] num the number of pages to allocate. This is usually 1,
7898 * unless allocating overflow pages for a large record.
7899 * @param[out] mp Address of a page, or NULL on failure.
7900 * @return 0 on success, non-zero on failure.
7903 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7908 if ((rc = mdb_page_alloc(mc, num, &np)))
7910 DPRINTF(("allocated new mpage %"Yu", page size %u",
7911 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7912 np->mp_flags = flags | P_DIRTY;
7913 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7914 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7917 mc->mc_db->md_branch_pages++;
7918 else if (IS_LEAF(np))
7919 mc->mc_db->md_leaf_pages++;
7920 else if (IS_OVERFLOW(np)) {
7921 mc->mc_db->md_overflow_pages += num;
7929 /** Calculate the size of a leaf node.
7930 * The size depends on the environment's page size; if a data item
7931 * is too large it will be put onto an overflow page and the node
7932 * size will only include the key and not the data. Sizes are always
7933 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7934 * of the #MDB_node headers.
7935 * @param[in] env The environment handle.
7936 * @param[in] key The key for the node.
7937 * @param[in] data The data for the node.
7938 * @return The number of bytes needed to store the node.
7941 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7945 sz = LEAFSIZE(key, data);
7946 if (sz > env->me_nodemax) {
7947 /* put on overflow page */
7948 sz -= data->mv_size - sizeof(pgno_t);
7951 return EVEN(sz + sizeof(indx_t));
7954 /** Calculate the size of a branch node.
7955 * The size should depend on the environment's page size but since
7956 * we currently don't support spilling large keys onto overflow
7957 * pages, it's simply the size of the #MDB_node header plus the
7958 * size of the key. Sizes are always rounded up to an even number
7959 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7960 * @param[in] env The environment handle.
7961 * @param[in] key The key for the node.
7962 * @return The number of bytes needed to store the node.
7965 mdb_branch_size(MDB_env *env, MDB_val *key)
7970 if (sz > env->me_nodemax) {
7971 /* put on overflow page */
7972 /* not implemented */
7973 /* sz -= key->size - sizeof(pgno_t); */
7976 return sz + sizeof(indx_t);
7979 /** Add a node to the page pointed to by the cursor.
7980 * Set #MDB_TXN_ERROR on failure.
7981 * @param[in] mc The cursor for this operation.
7982 * @param[in] indx The index on the page where the new node should be added.
7983 * @param[in] key The key for the new node.
7984 * @param[in] data The data for the new node, if any.
7985 * @param[in] pgno The page number, if adding a branch node.
7986 * @param[in] flags Flags for the node.
7987 * @return 0 on success, non-zero on failure. Possible errors are:
7989 * <li>ENOMEM - failed to allocate overflow pages for the node.
7990 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7991 * should never happen since all callers already calculate the
7992 * page's free space before calling this function.
7996 mdb_node_add(MDB_cursor *mc, indx_t indx,
7997 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
8000 size_t node_size = NODESIZE;
8004 MDB_page *mp = mc->mc_pg[mc->mc_top];
8005 MDB_page *ofp = NULL; /* overflow page */
8009 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
8011 DPRINTF(("add to %s %spage %"Yu" index %i, data size %"Z"u key size %"Z"u [%s]",
8012 IS_LEAF(mp) ? "leaf" : "branch",
8013 IS_SUBP(mp) ? "sub-" : "",
8014 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
8015 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
8018 /* Move higher keys up one slot. */
8019 int ksize = mc->mc_db->md_pad, dif;
8020 char *ptr = LEAF2KEY(mp, indx, ksize);
8021 dif = NUMKEYS(mp) - indx;
8023 memmove(ptr+ksize, ptr, dif*ksize);
8024 /* insert new key */
8025 memcpy(ptr, key->mv_data, ksize);
8027 /* Just using these for counting */
8028 mp->mp_lower += sizeof(indx_t);
8029 mp->mp_upper -= ksize - sizeof(indx_t);
8033 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
8035 node_size += key->mv_size;
8037 mdb_cassert(mc, key && data);
8038 if (F_ISSET(flags, F_BIGDATA)) {
8039 /* Data already on overflow page. */
8040 node_size += sizeof(pgno_t);
8041 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
8042 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
8044 /* Put data on overflow page. */
8045 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
8046 data->mv_size, node_size+data->mv_size));
8047 node_size = EVEN(node_size + sizeof(pgno_t));
8048 if ((ssize_t)node_size > room)
8050 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
8052 DPRINTF(("allocated overflow page %"Yu, ofp->mp_pgno));
8056 node_size += data->mv_size;
8059 node_size = EVEN(node_size);
8060 if ((ssize_t)node_size > room)
8064 /* Move higher pointers up one slot. */
8065 for (i = NUMKEYS(mp); i > indx; i--)
8066 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
8068 /* Adjust free space offsets. */
8069 ofs = mp->mp_upper - node_size;
8070 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
8071 mp->mp_ptrs[indx] = ofs;
8073 mp->mp_lower += sizeof(indx_t);
8075 /* Write the node data. */
8076 node = NODEPTR(mp, indx);
8077 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
8078 node->mn_flags = flags;
8080 SETDSZ(node,data->mv_size);
8085 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8088 ndata = NODEDATA(node);
8090 if (F_ISSET(flags, F_BIGDATA))
8091 memcpy(ndata, data->mv_data, sizeof(pgno_t));
8092 else if (F_ISSET(flags, MDB_RESERVE))
8093 data->mv_data = ndata;
8095 memcpy(ndata, data->mv_data, data->mv_size);
8097 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
8098 ndata = METADATA(ofp);
8099 if (F_ISSET(flags, MDB_RESERVE))
8100 data->mv_data = ndata;
8102 memcpy(ndata, data->mv_data, data->mv_size);
8109 DPRINTF(("not enough room in page %"Yu", got %u ptrs",
8110 mdb_dbg_pgno(mp), NUMKEYS(mp)));
8111 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
8112 DPRINTF(("node size = %"Z"u", node_size));
8113 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8114 return MDB_PAGE_FULL;
8117 /** Delete the specified node from a page.
8118 * @param[in] mc Cursor pointing to the node to delete.
8119 * @param[in] ksize The size of a node. Only used if the page is
8120 * part of a #MDB_DUPFIXED database.
8123 mdb_node_del(MDB_cursor *mc, int ksize)
8125 MDB_page *mp = mc->mc_pg[mc->mc_top];
8126 indx_t indx = mc->mc_ki[mc->mc_top];
8128 indx_t i, j, numkeys, ptr;
8132 DPRINTF(("delete node %u on %s page %"Yu, indx,
8133 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
8134 numkeys = NUMKEYS(mp);
8135 mdb_cassert(mc, indx < numkeys);
8138 int x = numkeys - 1 - indx;
8139 base = LEAF2KEY(mp, indx, ksize);
8141 memmove(base, base + ksize, x * ksize);
8142 mp->mp_lower -= sizeof(indx_t);
8143 mp->mp_upper += ksize - sizeof(indx_t);
8147 node = NODEPTR(mp, indx);
8148 sz = NODESIZE + node->mn_ksize;
8150 if (F_ISSET(node->mn_flags, F_BIGDATA))
8151 sz += sizeof(pgno_t);
8153 sz += NODEDSZ(node);
8157 ptr = mp->mp_ptrs[indx];
8158 for (i = j = 0; i < numkeys; i++) {
8160 mp->mp_ptrs[j] = mp->mp_ptrs[i];
8161 if (mp->mp_ptrs[i] < ptr)
8162 mp->mp_ptrs[j] += sz;
8167 base = (char *)mp + mp->mp_upper + PAGEBASE;
8168 memmove(base + sz, base, ptr - mp->mp_upper);
8170 mp->mp_lower -= sizeof(indx_t);
8174 /** Compact the main page after deleting a node on a subpage.
8175 * @param[in] mp The main page to operate on.
8176 * @param[in] indx The index of the subpage on the main page.
8179 mdb_node_shrink(MDB_page *mp, indx_t indx)
8184 indx_t delta, nsize, len, ptr;
8187 node = NODEPTR(mp, indx);
8188 sp = (MDB_page *)NODEDATA(node);
8189 delta = SIZELEFT(sp);
8190 nsize = NODEDSZ(node) - delta;
8192 /* Prepare to shift upward, set len = length(subpage part to shift) */
8196 return; /* do not make the node uneven-sized */
8198 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
8199 for (i = NUMKEYS(sp); --i >= 0; )
8200 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
8203 sp->mp_upper = sp->mp_lower;
8204 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
8205 SETDSZ(node, nsize);
8207 /* Shift <lower nodes...initial part of subpage> upward */
8208 base = (char *)mp + mp->mp_upper + PAGEBASE;
8209 memmove(base + delta, base, (char *)sp + len - base);
8211 ptr = mp->mp_ptrs[indx];
8212 for (i = NUMKEYS(mp); --i >= 0; ) {
8213 if (mp->mp_ptrs[i] <= ptr)
8214 mp->mp_ptrs[i] += delta;
8216 mp->mp_upper += delta;
8219 /** Initial setup of a sorted-dups cursor.
8220 * Sorted duplicates are implemented as a sub-database for the given key.
8221 * The duplicate data items are actually keys of the sub-database.
8222 * Operations on the duplicate data items are performed using a sub-cursor
8223 * initialized when the sub-database is first accessed. This function does
8224 * the preliminary setup of the sub-cursor, filling in the fields that
8225 * depend only on the parent DB.
8226 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8229 mdb_xcursor_init0(MDB_cursor *mc)
8231 MDB_xcursor *mx = mc->mc_xcursor;
8233 mx->mx_cursor.mc_xcursor = NULL;
8234 mx->mx_cursor.mc_txn = mc->mc_txn;
8235 mx->mx_cursor.mc_db = &mx->mx_db;
8236 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
8237 mx->mx_cursor.mc_dbi = mc->mc_dbi;
8238 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
8239 mx->mx_cursor.mc_snum = 0;
8240 mx->mx_cursor.mc_top = 0;
8241 MC_SET_OVPG(&mx->mx_cursor, NULL);
8242 mx->mx_cursor.mc_flags = C_SUB | (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP));
8243 mx->mx_dbx.md_name.mv_size = 0;
8244 mx->mx_dbx.md_name.mv_data = NULL;
8245 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
8246 mx->mx_dbx.md_dcmp = NULL;
8247 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
8250 /** Final setup of a sorted-dups cursor.
8251 * Sets up the fields that depend on the data from the main cursor.
8252 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8253 * @param[in] node The data containing the #MDB_db record for the
8254 * sorted-dup database.
8257 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
8259 MDB_xcursor *mx = mc->mc_xcursor;
8261 mx->mx_cursor.mc_flags &= C_SUB|C_ORIG_RDONLY|C_WRITEMAP;
8262 if (node->mn_flags & F_SUBDATA) {
8263 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
8264 mx->mx_cursor.mc_pg[0] = 0;
8265 mx->mx_cursor.mc_snum = 0;
8266 mx->mx_cursor.mc_top = 0;
8268 MDB_page *fp = NODEDATA(node);
8269 mx->mx_db.md_pad = 0;
8270 mx->mx_db.md_flags = 0;
8271 mx->mx_db.md_depth = 1;
8272 mx->mx_db.md_branch_pages = 0;
8273 mx->mx_db.md_leaf_pages = 1;
8274 mx->mx_db.md_overflow_pages = 0;
8275 mx->mx_db.md_entries = NUMKEYS(fp);
8276 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
8277 mx->mx_cursor.mc_snum = 1;
8278 mx->mx_cursor.mc_top = 0;
8279 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8280 mx->mx_cursor.mc_pg[0] = fp;
8281 mx->mx_cursor.mc_ki[0] = 0;
8282 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
8283 mx->mx_db.md_flags = MDB_DUPFIXED;
8284 mx->mx_db.md_pad = fp->mp_pad;
8285 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
8286 mx->mx_db.md_flags |= MDB_INTEGERKEY;
8289 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8290 mx->mx_db.md_root));
8291 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
8292 if (NEED_CMP_CLONG(mx->mx_dbx.md_cmp, mx->mx_db.md_pad))
8293 mx->mx_dbx.md_cmp = mdb_cmp_clong;
8297 /** Fixup a sorted-dups cursor due to underlying update.
8298 * Sets up some fields that depend on the data from the main cursor.
8299 * Almost the same as init1, but skips initialization steps if the
8300 * xcursor had already been used.
8301 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
8302 * @param[in] src_mx The xcursor of an up-to-date cursor.
8303 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
8306 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
8308 MDB_xcursor *mx = mc->mc_xcursor;
8311 mx->mx_cursor.mc_snum = 1;
8312 mx->mx_cursor.mc_top = 0;
8313 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8314 mx->mx_cursor.mc_ki[0] = 0;
8315 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
8316 #if UINT_MAX < MDB_SIZE_MAX /* matches mdb_xcursor_init1:NEED_CMP_CLONG() */
8317 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
8319 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
8322 mx->mx_db = src_mx->mx_db;
8323 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
8324 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8325 mx->mx_db.md_root));
8328 /** Initialize a cursor for a given transaction and database. */
8330 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
8333 mc->mc_backup = NULL;
8336 mc->mc_db = &txn->mt_dbs[dbi];
8337 mc->mc_dbx = &txn->mt_dbxs[dbi];
8338 mc->mc_dbflag = &txn->mt_dbflags[dbi];
8343 MC_SET_OVPG(mc, NULL);
8344 mc->mc_flags = txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
8345 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
8346 mdb_tassert(txn, mx != NULL);
8347 mc->mc_xcursor = mx;
8348 mdb_xcursor_init0(mc);
8350 mc->mc_xcursor = NULL;
8352 if (*mc->mc_dbflag & DB_STALE) {
8353 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
8358 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
8361 size_t size = sizeof(MDB_cursor);
8363 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
8366 if (txn->mt_flags & MDB_TXN_BLOCKED)
8369 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8372 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
8373 size += sizeof(MDB_xcursor);
8375 if ((mc = malloc(size)) != NULL) {
8376 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
8377 if (txn->mt_cursors) {
8378 mc->mc_next = txn->mt_cursors[dbi];
8379 txn->mt_cursors[dbi] = mc;
8380 mc->mc_flags |= C_UNTRACK;
8392 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
8394 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
8397 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
8400 if (txn->mt_flags & MDB_TXN_BLOCKED)
8403 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
8407 /* Return the count of duplicate data items for the current key */
8409 mdb_cursor_count(MDB_cursor *mc, mdb_size_t *countp)
8413 if (mc == NULL || countp == NULL)
8416 if (mc->mc_xcursor == NULL)
8417 return MDB_INCOMPATIBLE;
8419 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
8422 if (!(mc->mc_flags & C_INITIALIZED))
8425 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
8426 return MDB_NOTFOUND;
8428 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8429 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
8432 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
8435 *countp = mc->mc_xcursor->mx_db.md_entries;
8441 mdb_cursor_close(MDB_cursor *mc)
8444 MDB_CURSOR_UNREF(mc, 0);
8446 if (mc && !mc->mc_backup) {
8447 /* remove from txn, if tracked */
8448 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
8449 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
8450 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
8452 *prev = mc->mc_next;
8459 mdb_cursor_txn(MDB_cursor *mc)
8461 if (!mc) return NULL;
8466 mdb_cursor_dbi(MDB_cursor *mc)
8471 /** Replace the key for a branch node with a new key.
8472 * Set #MDB_TXN_ERROR on failure.
8473 * @param[in] mc Cursor pointing to the node to operate on.
8474 * @param[in] key The new key to use.
8475 * @return 0 on success, non-zero on failure.
8478 mdb_update_key(MDB_cursor *mc, MDB_val *key)
8484 int delta, ksize, oksize;
8485 indx_t ptr, i, numkeys, indx;
8488 indx = mc->mc_ki[mc->mc_top];
8489 mp = mc->mc_pg[mc->mc_top];
8490 node = NODEPTR(mp, indx);
8491 ptr = mp->mp_ptrs[indx];
8495 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
8496 k2.mv_data = NODEKEY(node);
8497 k2.mv_size = node->mn_ksize;
8498 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Yu,
8500 mdb_dkey(&k2, kbuf2),
8506 /* Sizes must be 2-byte aligned. */
8507 ksize = EVEN(key->mv_size);
8508 oksize = EVEN(node->mn_ksize);
8509 delta = ksize - oksize;
8511 /* Shift node contents if EVEN(key length) changed. */
8513 if (delta > 0 && SIZELEFT(mp) < delta) {
8515 /* not enough space left, do a delete and split */
8516 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
8517 pgno = NODEPGNO(node);
8518 mdb_node_del(mc, 0);
8519 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
8522 numkeys = NUMKEYS(mp);
8523 for (i = 0; i < numkeys; i++) {
8524 if (mp->mp_ptrs[i] <= ptr)
8525 mp->mp_ptrs[i] -= delta;
8528 base = (char *)mp + mp->mp_upper + PAGEBASE;
8529 len = ptr - mp->mp_upper + NODESIZE;
8530 memmove(base - delta, base, len);
8531 mp->mp_upper -= delta;
8533 node = NODEPTR(mp, indx);
8536 /* But even if no shift was needed, update ksize */
8537 if (node->mn_ksize != key->mv_size)
8538 node->mn_ksize = key->mv_size;
8541 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8547 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
8549 /** Perform \b act while tracking temporary cursor \b mn */
8550 #define WITH_CURSOR_TRACKING(mn, act) do { \
8551 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
8552 if ((mn).mc_flags & C_SUB) { \
8553 dummy.mc_flags = C_INITIALIZED; \
8554 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
8559 tracked->mc_next = *tp; \
8562 *tp = tracked->mc_next; \
8565 /** Move a node from csrc to cdst.
8568 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
8575 unsigned short flags;
8579 /* Mark src and dst as dirty. */
8580 if ((rc = mdb_page_touch(csrc)) ||
8581 (rc = mdb_page_touch(cdst)))
8584 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8585 key.mv_size = csrc->mc_db->md_pad;
8586 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
8588 data.mv_data = NULL;
8592 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
8593 mdb_cassert(csrc, !((size_t)srcnode & 1));
8594 srcpg = NODEPGNO(srcnode);
8595 flags = srcnode->mn_flags;
8596 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8597 unsigned int snum = csrc->mc_snum;
8599 /* must find the lowest key below src */
8600 rc = mdb_page_search_lowest(csrc);
8603 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8604 key.mv_size = csrc->mc_db->md_pad;
8605 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8607 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8608 key.mv_size = NODEKSZ(s2);
8609 key.mv_data = NODEKEY(s2);
8611 csrc->mc_snum = snum--;
8612 csrc->mc_top = snum;
8614 key.mv_size = NODEKSZ(srcnode);
8615 key.mv_data = NODEKEY(srcnode);
8617 data.mv_size = NODEDSZ(srcnode);
8618 data.mv_data = NODEDATA(srcnode);
8620 mn.mc_xcursor = NULL;
8621 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
8622 unsigned int snum = cdst->mc_snum;
8625 /* must find the lowest key below dst */
8626 mdb_cursor_copy(cdst, &mn);
8627 rc = mdb_page_search_lowest(&mn);
8630 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8631 bkey.mv_size = mn.mc_db->md_pad;
8632 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
8634 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8635 bkey.mv_size = NODEKSZ(s2);
8636 bkey.mv_data = NODEKEY(s2);
8638 mn.mc_snum = snum--;
8641 rc = mdb_update_key(&mn, &bkey);
8646 DPRINTF(("moving %s node %u [%s] on page %"Yu" to node %u on page %"Yu,
8647 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
8648 csrc->mc_ki[csrc->mc_top],
8650 csrc->mc_pg[csrc->mc_top]->mp_pgno,
8651 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
8653 /* Add the node to the destination page.
8655 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
8656 if (rc != MDB_SUCCESS)
8659 /* Delete the node from the source page.
8661 mdb_node_del(csrc, key.mv_size);
8664 /* Adjust other cursors pointing to mp */
8665 MDB_cursor *m2, *m3;
8666 MDB_dbi dbi = csrc->mc_dbi;
8667 MDB_page *mpd, *mps;
8669 mps = csrc->mc_pg[csrc->mc_top];
8670 /* If we're adding on the left, bump others up */
8672 mpd = cdst->mc_pg[csrc->mc_top];
8673 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8674 if (csrc->mc_flags & C_SUB)
8675 m3 = &m2->mc_xcursor->mx_cursor;
8678 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8681 m3->mc_pg[csrc->mc_top] == mpd &&
8682 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
8683 m3->mc_ki[csrc->mc_top]++;
8686 m3->mc_pg[csrc->mc_top] == mps &&
8687 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
8688 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8689 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8690 m3->mc_ki[csrc->mc_top-1]++;
8692 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
8693 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8696 /* Adding on the right, bump others down */
8698 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8699 if (csrc->mc_flags & C_SUB)
8700 m3 = &m2->mc_xcursor->mx_cursor;
8703 if (m3 == csrc) continue;
8704 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8706 if (m3->mc_pg[csrc->mc_top] == mps) {
8707 if (!m3->mc_ki[csrc->mc_top]) {
8708 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8709 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8710 m3->mc_ki[csrc->mc_top-1]--;
8712 m3->mc_ki[csrc->mc_top]--;
8714 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
8715 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8721 /* Update the parent separators.
8723 if (csrc->mc_ki[csrc->mc_top] == 0) {
8724 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
8725 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8726 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8728 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8729 key.mv_size = NODEKSZ(srcnode);
8730 key.mv_data = NODEKEY(srcnode);
8732 DPRINTF(("update separator for source page %"Yu" to [%s]",
8733 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
8734 mdb_cursor_copy(csrc, &mn);
8737 /* We want mdb_rebalance to find mn when doing fixups */
8738 WITH_CURSOR_TRACKING(mn,
8739 rc = mdb_update_key(&mn, &key));
8743 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8745 indx_t ix = csrc->mc_ki[csrc->mc_top];
8746 nullkey.mv_size = 0;
8747 csrc->mc_ki[csrc->mc_top] = 0;
8748 rc = mdb_update_key(csrc, &nullkey);
8749 csrc->mc_ki[csrc->mc_top] = ix;
8750 mdb_cassert(csrc, rc == MDB_SUCCESS);
8754 if (cdst->mc_ki[cdst->mc_top] == 0) {
8755 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
8756 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8757 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
8759 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
8760 key.mv_size = NODEKSZ(srcnode);
8761 key.mv_data = NODEKEY(srcnode);
8763 DPRINTF(("update separator for destination page %"Yu" to [%s]",
8764 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
8765 mdb_cursor_copy(cdst, &mn);
8768 /* We want mdb_rebalance to find mn when doing fixups */
8769 WITH_CURSOR_TRACKING(mn,
8770 rc = mdb_update_key(&mn, &key));
8774 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
8776 indx_t ix = cdst->mc_ki[cdst->mc_top];
8777 nullkey.mv_size = 0;
8778 cdst->mc_ki[cdst->mc_top] = 0;
8779 rc = mdb_update_key(cdst, &nullkey);
8780 cdst->mc_ki[cdst->mc_top] = ix;
8781 mdb_cassert(cdst, rc == MDB_SUCCESS);
8788 /** Merge one page into another.
8789 * The nodes from the page pointed to by \b csrc will
8790 * be copied to the page pointed to by \b cdst and then
8791 * the \b csrc page will be freed.
8792 * @param[in] csrc Cursor pointing to the source page.
8793 * @param[in] cdst Cursor pointing to the destination page.
8794 * @return 0 on success, non-zero on failure.
8797 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8799 MDB_page *psrc, *pdst;
8806 psrc = csrc->mc_pg[csrc->mc_top];
8807 pdst = cdst->mc_pg[cdst->mc_top];
8809 DPRINTF(("merging page %"Yu" into %"Yu, psrc->mp_pgno, pdst->mp_pgno));
8811 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8812 mdb_cassert(csrc, cdst->mc_snum > 1);
8814 /* Mark dst as dirty. */
8815 if ((rc = mdb_page_touch(cdst)))
8818 /* get dst page again now that we've touched it. */
8819 pdst = cdst->mc_pg[cdst->mc_top];
8821 /* Move all nodes from src to dst.
8823 j = nkeys = NUMKEYS(pdst);
8824 if (IS_LEAF2(psrc)) {
8825 key.mv_size = csrc->mc_db->md_pad;
8826 key.mv_data = METADATA(psrc);
8827 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8828 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8829 if (rc != MDB_SUCCESS)
8831 key.mv_data = (char *)key.mv_data + key.mv_size;
8834 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8835 srcnode = NODEPTR(psrc, i);
8836 if (i == 0 && IS_BRANCH(psrc)) {
8839 mdb_cursor_copy(csrc, &mn);
8840 mn.mc_xcursor = NULL;
8841 /* must find the lowest key below src */
8842 rc = mdb_page_search_lowest(&mn);
8845 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8846 key.mv_size = mn.mc_db->md_pad;
8847 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8849 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8850 key.mv_size = NODEKSZ(s2);
8851 key.mv_data = NODEKEY(s2);
8854 key.mv_size = srcnode->mn_ksize;
8855 key.mv_data = NODEKEY(srcnode);
8858 data.mv_size = NODEDSZ(srcnode);
8859 data.mv_data = NODEDATA(srcnode);
8860 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8861 if (rc != MDB_SUCCESS)
8866 DPRINTF(("dst page %"Yu" now has %u keys (%.1f%% filled)",
8867 pdst->mp_pgno, NUMKEYS(pdst),
8868 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8870 /* Unlink the src page from parent and add to free list.
8873 mdb_node_del(csrc, 0);
8874 if (csrc->mc_ki[csrc->mc_top] == 0) {
8876 rc = mdb_update_key(csrc, &key);
8884 psrc = csrc->mc_pg[csrc->mc_top];
8885 /* If not operating on FreeDB, allow this page to be reused
8886 * in this txn. Otherwise just add to free list.
8888 rc = mdb_page_loose(csrc, psrc);
8892 csrc->mc_db->md_leaf_pages--;
8894 csrc->mc_db->md_branch_pages--;
8896 /* Adjust other cursors pointing to mp */
8897 MDB_cursor *m2, *m3;
8898 MDB_dbi dbi = csrc->mc_dbi;
8899 unsigned int top = csrc->mc_top;
8901 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8902 if (csrc->mc_flags & C_SUB)
8903 m3 = &m2->mc_xcursor->mx_cursor;
8906 if (m3 == csrc) continue;
8907 if (m3->mc_snum < csrc->mc_snum) continue;
8908 if (m3->mc_pg[top] == psrc) {
8909 m3->mc_pg[top] = pdst;
8910 m3->mc_ki[top] += nkeys;
8911 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8912 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8913 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8916 if (XCURSOR_INITED(m3) && IS_LEAF(psrc))
8917 XCURSOR_REFRESH(m3, m3->mc_pg[top], m3->mc_ki[top]);
8921 unsigned int snum = cdst->mc_snum;
8922 uint16_t depth = cdst->mc_db->md_depth;
8923 mdb_cursor_pop(cdst);
8924 rc = mdb_rebalance(cdst);
8925 /* Did the tree height change? */
8926 if (depth != cdst->mc_db->md_depth)
8927 snum += cdst->mc_db->md_depth - depth;
8928 cdst->mc_snum = snum;
8929 cdst->mc_top = snum-1;
8934 /** Copy the contents of a cursor.
8935 * @param[in] csrc The cursor to copy from.
8936 * @param[out] cdst The cursor to copy to.
8939 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8943 cdst->mc_txn = csrc->mc_txn;
8944 cdst->mc_dbi = csrc->mc_dbi;
8945 cdst->mc_db = csrc->mc_db;
8946 cdst->mc_dbx = csrc->mc_dbx;
8947 cdst->mc_snum = csrc->mc_snum;
8948 cdst->mc_top = csrc->mc_top;
8949 cdst->mc_flags = csrc->mc_flags;
8950 MC_SET_OVPG(cdst, MC_OVPG(csrc));
8952 for (i=0; i<csrc->mc_snum; i++) {
8953 cdst->mc_pg[i] = csrc->mc_pg[i];
8954 cdst->mc_ki[i] = csrc->mc_ki[i];
8958 /** Rebalance the tree after a delete operation.
8959 * @param[in] mc Cursor pointing to the page where rebalancing
8961 * @return 0 on success, non-zero on failure.
8964 mdb_rebalance(MDB_cursor *mc)
8968 unsigned int ptop, minkeys, thresh;
8972 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8977 thresh = FILL_THRESHOLD;
8979 DPRINTF(("rebalancing %s page %"Yu" (has %u keys, %.1f%% full)",
8980 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8981 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8982 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8984 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8985 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8986 DPRINTF(("no need to rebalance page %"Yu", above fill threshold",
8987 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8991 if (mc->mc_snum < 2) {
8992 MDB_page *mp = mc->mc_pg[0];
8994 DPUTS("Can't rebalance a subpage, ignoring");
8997 if (NUMKEYS(mp) == 0) {
8998 DPUTS("tree is completely empty");
8999 mc->mc_db->md_root = P_INVALID;
9000 mc->mc_db->md_depth = 0;
9001 mc->mc_db->md_leaf_pages = 0;
9002 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
9005 /* Adjust cursors pointing to mp */
9008 mc->mc_flags &= ~C_INITIALIZED;
9010 MDB_cursor *m2, *m3;
9011 MDB_dbi dbi = mc->mc_dbi;
9013 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9014 if (mc->mc_flags & C_SUB)
9015 m3 = &m2->mc_xcursor->mx_cursor;
9018 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
9020 if (m3->mc_pg[0] == mp) {
9023 m3->mc_flags &= ~C_INITIALIZED;
9027 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
9029 DPUTS("collapsing root page!");
9030 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
9033 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
9034 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
9037 mc->mc_db->md_depth--;
9038 mc->mc_db->md_branch_pages--;
9039 mc->mc_ki[0] = mc->mc_ki[1];
9040 for (i = 1; i<mc->mc_db->md_depth; i++) {
9041 mc->mc_pg[i] = mc->mc_pg[i+1];
9042 mc->mc_ki[i] = mc->mc_ki[i+1];
9045 /* Adjust other cursors pointing to mp */
9046 MDB_cursor *m2, *m3;
9047 MDB_dbi dbi = mc->mc_dbi;
9049 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9050 if (mc->mc_flags & C_SUB)
9051 m3 = &m2->mc_xcursor->mx_cursor;
9054 if (m3 == mc) continue;
9055 if (!(m3->mc_flags & C_INITIALIZED))
9057 if (m3->mc_pg[0] == mp) {
9058 for (i=0; i<mc->mc_db->md_depth; i++) {
9059 m3->mc_pg[i] = m3->mc_pg[i+1];
9060 m3->mc_ki[i] = m3->mc_ki[i+1];
9068 DPUTS("root page doesn't need rebalancing");
9072 /* The parent (branch page) must have at least 2 pointers,
9073 * otherwise the tree is invalid.
9075 ptop = mc->mc_top-1;
9076 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
9078 /* Leaf page fill factor is below the threshold.
9079 * Try to move keys from left or right neighbor, or
9080 * merge with a neighbor page.
9085 mdb_cursor_copy(mc, &mn);
9086 mn.mc_xcursor = NULL;
9088 oldki = mc->mc_ki[mc->mc_top];
9089 if (mc->mc_ki[ptop] == 0) {
9090 /* We're the leftmost leaf in our parent.
9092 DPUTS("reading right neighbor");
9094 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
9095 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
9098 mn.mc_ki[mn.mc_top] = 0;
9099 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
9102 /* There is at least one neighbor to the left.
9104 DPUTS("reading left neighbor");
9106 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
9107 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
9110 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
9111 mc->mc_ki[mc->mc_top] = 0;
9115 DPRINTF(("found neighbor page %"Yu" (%u keys, %.1f%% full)",
9116 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
9117 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
9119 /* If the neighbor page is above threshold and has enough keys,
9120 * move one key from it. Otherwise we should try to merge them.
9121 * (A branch page must never have less than 2 keys.)
9123 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
9124 rc = mdb_node_move(&mn, mc, fromleft);
9126 /* if we inserted on left, bump position up */
9131 rc = mdb_page_merge(&mn, mc);
9133 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
9134 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
9135 /* We want mdb_rebalance to find mn when doing fixups */
9136 WITH_CURSOR_TRACKING(mn,
9137 rc = mdb_page_merge(mc, &mn));
9138 mdb_cursor_copy(&mn, mc);
9140 mc->mc_flags &= ~C_EOF;
9142 mc->mc_ki[mc->mc_top] = oldki;
9146 /** Complete a delete operation started by #mdb_cursor_del(). */
9148 mdb_cursor_del0(MDB_cursor *mc)
9154 MDB_cursor *m2, *m3;
9155 MDB_dbi dbi = mc->mc_dbi;
9157 ki = mc->mc_ki[mc->mc_top];
9158 mp = mc->mc_pg[mc->mc_top];
9159 mdb_node_del(mc, mc->mc_db->md_pad);
9160 mc->mc_db->md_entries--;
9162 /* Adjust other cursors pointing to mp */
9163 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9164 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9165 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9167 if (m3 == mc || m3->mc_snum < mc->mc_snum)
9169 if (m3->mc_pg[mc->mc_top] == mp) {
9170 if (m3->mc_ki[mc->mc_top] == ki) {
9171 m3->mc_flags |= C_DEL;
9172 if (mc->mc_db->md_flags & MDB_DUPSORT) {
9173 /* Sub-cursor referred into dataset which is gone */
9174 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
9177 } else if (m3->mc_ki[mc->mc_top] > ki) {
9178 m3->mc_ki[mc->mc_top]--;
9180 if (XCURSOR_INITED(m3))
9181 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9185 rc = mdb_rebalance(mc);
9187 if (rc == MDB_SUCCESS) {
9188 /* DB is totally empty now, just bail out.
9189 * Other cursors adjustments were already done
9190 * by mdb_rebalance and aren't needed here.
9195 mp = mc->mc_pg[mc->mc_top];
9196 nkeys = NUMKEYS(mp);
9198 /* Adjust other cursors pointing to mp */
9199 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
9200 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9201 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9203 if (m3->mc_snum < mc->mc_snum)
9205 if (m3->mc_pg[mc->mc_top] == mp) {
9206 /* if m3 points past last node in page, find next sibling */
9207 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
9208 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9209 rc = mdb_cursor_sibling(m3, 1);
9210 if (rc == MDB_NOTFOUND) {
9211 m3->mc_flags |= C_EOF;
9216 if (mc->mc_db->md_flags & MDB_DUPSORT) {
9217 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
9218 /* If this node is a fake page, it needs to be reinited
9219 * because its data has moved. But just reset mc_pg[0]
9220 * if the xcursor is already live.
9222 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) {
9223 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)
9224 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9226 mdb_xcursor_init1(m3, node);
9232 mc->mc_flags |= C_DEL;
9236 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9241 mdb_del(MDB_txn *txn, MDB_dbi dbi,
9242 MDB_val *key, MDB_val *data)
9244 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9247 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9248 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9250 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
9251 /* must ignore any data */
9255 return mdb_del0(txn, dbi, key, data, 0);
9259 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
9260 MDB_val *key, MDB_val *data, unsigned flags)
9265 MDB_val rdata, *xdata;
9269 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
9271 mdb_cursor_init(&mc, txn, dbi, &mx);
9280 flags |= MDB_NODUPDATA;
9282 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
9284 /* let mdb_page_split know about this cursor if needed:
9285 * delete will trigger a rebalance; if it needs to move
9286 * a node from one page to another, it will have to
9287 * update the parent's separator key(s). If the new sepkey
9288 * is larger than the current one, the parent page may
9289 * run out of space, triggering a split. We need this
9290 * cursor to be consistent until the end of the rebalance.
9292 mc.mc_flags |= C_UNTRACK;
9293 mc.mc_next = txn->mt_cursors[dbi];
9294 txn->mt_cursors[dbi] = &mc;
9295 rc = mdb_cursor_del(&mc, flags);
9296 txn->mt_cursors[dbi] = mc.mc_next;
9301 /** Split a page and insert a new node.
9302 * Set #MDB_TXN_ERROR on failure.
9303 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
9304 * The cursor will be updated to point to the actual page and index where
9305 * the node got inserted after the split.
9306 * @param[in] newkey The key for the newly inserted node.
9307 * @param[in] newdata The data for the newly inserted node.
9308 * @param[in] newpgno The page number, if the new node is a branch node.
9309 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
9310 * @return 0 on success, non-zero on failure.
9313 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
9314 unsigned int nflags)
9317 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
9320 int i, j, split_indx, nkeys, pmax;
9321 MDB_env *env = mc->mc_txn->mt_env;
9323 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
9324 MDB_page *copy = NULL;
9325 MDB_page *mp, *rp, *pp;
9330 mp = mc->mc_pg[mc->mc_top];
9331 newindx = mc->mc_ki[mc->mc_top];
9332 nkeys = NUMKEYS(mp);
9334 DPRINTF(("-----> splitting %s page %"Yu" and adding [%s] at index %i/%i",
9335 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
9336 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
9338 /* Create a right sibling. */
9339 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
9341 rp->mp_pad = mp->mp_pad;
9342 DPRINTF(("new right sibling: page %"Yu, rp->mp_pgno));
9344 /* Usually when splitting the root page, the cursor
9345 * height is 1. But when called from mdb_update_key,
9346 * the cursor height may be greater because it walks
9347 * up the stack while finding the branch slot to update.
9349 if (mc->mc_top < 1) {
9350 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
9352 /* shift current top to make room for new parent */
9353 for (i=mc->mc_snum; i>0; i--) {
9354 mc->mc_pg[i] = mc->mc_pg[i-1];
9355 mc->mc_ki[i] = mc->mc_ki[i-1];
9359 mc->mc_db->md_root = pp->mp_pgno;
9360 DPRINTF(("root split! new root = %"Yu, pp->mp_pgno));
9361 new_root = mc->mc_db->md_depth++;
9363 /* Add left (implicit) pointer. */
9364 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
9365 /* undo the pre-push */
9366 mc->mc_pg[0] = mc->mc_pg[1];
9367 mc->mc_ki[0] = mc->mc_ki[1];
9368 mc->mc_db->md_root = mp->mp_pgno;
9369 mc->mc_db->md_depth--;
9376 ptop = mc->mc_top-1;
9377 DPRINTF(("parent branch page is %"Yu, mc->mc_pg[ptop]->mp_pgno));
9380 mdb_cursor_copy(mc, &mn);
9381 mn.mc_xcursor = NULL;
9382 mn.mc_pg[mn.mc_top] = rp;
9383 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
9385 if (nflags & MDB_APPEND) {
9386 mn.mc_ki[mn.mc_top] = 0;
9388 split_indx = newindx;
9392 split_indx = (nkeys+1) / 2;
9397 unsigned int lsize, rsize, ksize;
9398 /* Move half of the keys to the right sibling */
9399 x = mc->mc_ki[mc->mc_top] - split_indx;
9400 ksize = mc->mc_db->md_pad;
9401 split = LEAF2KEY(mp, split_indx, ksize);
9402 rsize = (nkeys - split_indx) * ksize;
9403 lsize = (nkeys - split_indx) * sizeof(indx_t);
9404 mp->mp_lower -= lsize;
9405 rp->mp_lower += lsize;
9406 mp->mp_upper += rsize - lsize;
9407 rp->mp_upper -= rsize - lsize;
9408 sepkey.mv_size = ksize;
9409 if (newindx == split_indx) {
9410 sepkey.mv_data = newkey->mv_data;
9412 sepkey.mv_data = split;
9415 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
9416 memcpy(rp->mp_ptrs, split, rsize);
9417 sepkey.mv_data = rp->mp_ptrs;
9418 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
9419 memcpy(ins, newkey->mv_data, ksize);
9420 mp->mp_lower += sizeof(indx_t);
9421 mp->mp_upper -= ksize - sizeof(indx_t);
9424 memcpy(rp->mp_ptrs, split, x * ksize);
9425 ins = LEAF2KEY(rp, x, ksize);
9426 memcpy(ins, newkey->mv_data, ksize);
9427 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
9428 rp->mp_lower += sizeof(indx_t);
9429 rp->mp_upper -= ksize - sizeof(indx_t);
9430 mc->mc_ki[mc->mc_top] = x;
9433 int psize, nsize, k;
9434 /* Maximum free space in an empty page */
9435 pmax = env->me_psize - PAGEHDRSZ;
9437 nsize = mdb_leaf_size(env, newkey, newdata);
9439 nsize = mdb_branch_size(env, newkey);
9440 nsize = EVEN(nsize);
9442 /* grab a page to hold a temporary copy */
9443 copy = mdb_page_malloc(mc->mc_txn, 1);
9448 copy->mp_pgno = mp->mp_pgno;
9449 copy->mp_flags = mp->mp_flags;
9450 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
9451 copy->mp_upper = env->me_psize - PAGEBASE;
9453 /* prepare to insert */
9454 for (i=0, j=0; i<nkeys; i++) {
9456 copy->mp_ptrs[j++] = 0;
9458 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
9461 /* When items are relatively large the split point needs
9462 * to be checked, because being off-by-one will make the
9463 * difference between success or failure in mdb_node_add.
9465 * It's also relevant if a page happens to be laid out
9466 * such that one half of its nodes are all "small" and
9467 * the other half of its nodes are "large." If the new
9468 * item is also "large" and falls on the half with
9469 * "large" nodes, it also may not fit.
9471 * As a final tweak, if the new item goes on the last
9472 * spot on the page (and thus, onto the new page), bias
9473 * the split so the new page is emptier than the old page.
9474 * This yields better packing during sequential inserts.
9476 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
9477 /* Find split point */
9479 if (newindx <= split_indx || newindx >= nkeys) {
9481 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
9486 for (; i!=k; i+=j) {
9491 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9492 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
9494 if (F_ISSET(node->mn_flags, F_BIGDATA))
9495 psize += sizeof(pgno_t);
9497 psize += NODEDSZ(node);
9499 psize = EVEN(psize);
9501 if (psize > pmax || i == k-j) {
9502 split_indx = i + (j<0);
9507 if (split_indx == newindx) {
9508 sepkey.mv_size = newkey->mv_size;
9509 sepkey.mv_data = newkey->mv_data;
9511 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
9512 sepkey.mv_size = node->mn_ksize;
9513 sepkey.mv_data = NODEKEY(node);
9518 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
9520 /* Copy separator key to the parent.
9522 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
9523 int snum = mc->mc_snum;
9527 /* We want other splits to find mn when doing fixups */
9528 WITH_CURSOR_TRACKING(mn,
9529 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
9534 if (mc->mc_snum > snum) {
9537 /* Right page might now have changed parent.
9538 * Check if left page also changed parent.
9540 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9541 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9542 for (i=0; i<ptop; i++) {
9543 mc->mc_pg[i] = mn.mc_pg[i];
9544 mc->mc_ki[i] = mn.mc_ki[i];
9546 mc->mc_pg[ptop] = mn.mc_pg[ptop];
9547 if (mn.mc_ki[ptop]) {
9548 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
9550 /* find right page's left sibling */
9551 mc->mc_ki[ptop] = mn.mc_ki[ptop];
9552 mdb_cursor_sibling(mc, 0);
9557 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
9560 if (rc != MDB_SUCCESS) {
9563 if (nflags & MDB_APPEND) {
9564 mc->mc_pg[mc->mc_top] = rp;
9565 mc->mc_ki[mc->mc_top] = 0;
9566 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
9569 for (i=0; i<mc->mc_top; i++)
9570 mc->mc_ki[i] = mn.mc_ki[i];
9571 } else if (!IS_LEAF2(mp)) {
9573 mc->mc_pg[mc->mc_top] = rp;
9578 rkey.mv_data = newkey->mv_data;
9579 rkey.mv_size = newkey->mv_size;
9585 /* Update index for the new key. */
9586 mc->mc_ki[mc->mc_top] = j;
9588 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9589 rkey.mv_data = NODEKEY(node);
9590 rkey.mv_size = node->mn_ksize;
9592 xdata.mv_data = NODEDATA(node);
9593 xdata.mv_size = NODEDSZ(node);
9596 pgno = NODEPGNO(node);
9597 flags = node->mn_flags;
9600 if (!IS_LEAF(mp) && j == 0) {
9601 /* First branch index doesn't need key data. */
9605 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
9611 mc->mc_pg[mc->mc_top] = copy;
9616 } while (i != split_indx);
9618 nkeys = NUMKEYS(copy);
9619 for (i=0; i<nkeys; i++)
9620 mp->mp_ptrs[i] = copy->mp_ptrs[i];
9621 mp->mp_lower = copy->mp_lower;
9622 mp->mp_upper = copy->mp_upper;
9623 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
9624 env->me_psize - copy->mp_upper - PAGEBASE);
9626 /* reset back to original page */
9627 if (newindx < split_indx) {
9628 mc->mc_pg[mc->mc_top] = mp;
9630 mc->mc_pg[mc->mc_top] = rp;
9632 /* Make sure mc_ki is still valid.
9634 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9635 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9636 for (i=0; i<=ptop; i++) {
9637 mc->mc_pg[i] = mn.mc_pg[i];
9638 mc->mc_ki[i] = mn.mc_ki[i];
9642 if (nflags & MDB_RESERVE) {
9643 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
9644 if (!(node->mn_flags & F_BIGDATA))
9645 newdata->mv_data = NODEDATA(node);
9648 if (newindx >= split_indx) {
9649 mc->mc_pg[mc->mc_top] = rp;
9651 /* Make sure mc_ki is still valid.
9653 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9654 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9655 for (i=0; i<=ptop; i++) {
9656 mc->mc_pg[i] = mn.mc_pg[i];
9657 mc->mc_ki[i] = mn.mc_ki[i];
9664 /* Adjust other cursors pointing to mp */
9665 MDB_cursor *m2, *m3;
9666 MDB_dbi dbi = mc->mc_dbi;
9667 nkeys = NUMKEYS(mp);
9669 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9670 if (mc->mc_flags & C_SUB)
9671 m3 = &m2->mc_xcursor->mx_cursor;
9676 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9680 /* sub cursors may be on different DB */
9681 if (m3->mc_pg[0] != mp)
9684 for (k=new_root; k>=0; k--) {
9685 m3->mc_ki[k+1] = m3->mc_ki[k];
9686 m3->mc_pg[k+1] = m3->mc_pg[k];
9688 if (m3->mc_ki[0] >= nkeys) {
9693 m3->mc_pg[0] = mc->mc_pg[0];
9697 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
9698 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
9699 m3->mc_ki[mc->mc_top]++;
9700 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9701 m3->mc_pg[mc->mc_top] = rp;
9702 m3->mc_ki[mc->mc_top] -= nkeys;
9703 for (i=0; i<mc->mc_top; i++) {
9704 m3->mc_ki[i] = mn.mc_ki[i];
9705 m3->mc_pg[i] = mn.mc_pg[i];
9708 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
9709 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
9712 if (XCURSOR_INITED(m3) && IS_LEAF(mp))
9713 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9716 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
9719 if (copy) /* tmp page */
9720 mdb_page_free(env, copy);
9722 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9727 mdb_put(MDB_txn *txn, MDB_dbi dbi,
9728 MDB_val *key, MDB_val *data, unsigned int flags)
9734 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9737 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
9740 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9741 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9743 mdb_cursor_init(&mc, txn, dbi, &mx);
9744 mc.mc_next = txn->mt_cursors[dbi];
9745 txn->mt_cursors[dbi] = &mc;
9746 rc = mdb_cursor_put(&mc, key, data, flags);
9747 txn->mt_cursors[dbi] = mc.mc_next;
9752 #define MDB_WBUF (1024*1024)
9754 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
9756 /** State needed for a double-buffering compacting copy. */
9757 typedef struct mdb_copy {
9760 pthread_mutex_t mc_mutex;
9761 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
9766 pgno_t mc_next_pgno;
9768 int mc_toggle; /**< Buffer number in provider */
9769 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
9770 /** Error code. Never cleared if set. Both threads can set nonzero
9771 * to fail the copy. Not mutex-protected, LMDB expects atomic int.
9773 volatile int mc_error;
9776 /** Dedicated writer thread for compacting copy. */
9777 static THREAD_RET ESECT CALL_CONV
9778 mdb_env_copythr(void *arg)
9782 int toggle = 0, wsize, rc;
9785 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9788 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9791 pthread_mutex_lock(&my->mc_mutex);
9794 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9795 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
9797 wsize = my->mc_wlen[toggle];
9798 ptr = my->mc_wbuf[toggle];
9801 while (wsize > 0 && !my->mc_error) {
9802 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9806 } else if (len > 0) {
9819 /* If there's an overflow page tail, write it too */
9820 if (my->mc_olen[toggle]) {
9821 wsize = my->mc_olen[toggle];
9822 ptr = my->mc_over[toggle];
9823 my->mc_olen[toggle] = 0;
9826 my->mc_wlen[toggle] = 0;
9828 /* Return the empty buffer to provider */
9830 pthread_cond_signal(&my->mc_cond);
9832 pthread_mutex_unlock(&my->mc_mutex);
9833 return (THREAD_RET)0;
9837 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9839 * @param[in] my control structure.
9840 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9843 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9845 pthread_mutex_lock(&my->mc_mutex);
9846 my->mc_new += adjust;
9847 pthread_cond_signal(&my->mc_cond);
9848 while (my->mc_new & 2) /* both buffers in use */
9849 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9850 pthread_mutex_unlock(&my->mc_mutex);
9852 my->mc_toggle ^= (adjust & 1);
9853 /* Both threads reset mc_wlen, to be safe from threading errors */
9854 my->mc_wlen[my->mc_toggle] = 0;
9855 return my->mc_error;
9858 /** Depth-first tree traversal for compacting copy.
9859 * @param[in] my control structure.
9860 * @param[in,out] pg database root.
9861 * @param[in] flags includes #F_DUPDATA if it is a sorted-duplicate sub-DB.
9864 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9866 MDB_cursor mc = {0};
9868 MDB_page *mo, *mp, *leaf;
9873 /* Empty DB, nothing to do */
9874 if (*pg == P_INVALID)
9878 mc.mc_txn = my->mc_txn;
9879 mc.mc_flags = my->mc_txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
9881 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9884 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9888 /* Make cursor pages writable */
9889 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9893 for (i=0; i<mc.mc_top; i++) {
9894 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9895 mc.mc_pg[i] = (MDB_page *)ptr;
9896 ptr += my->mc_env->me_psize;
9899 /* This is writable space for a leaf page. Usually not needed. */
9900 leaf = (MDB_page *)ptr;
9902 toggle = my->mc_toggle;
9903 while (mc.mc_snum > 0) {
9905 mp = mc.mc_pg[mc.mc_top];
9909 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9910 for (i=0; i<n; i++) {
9911 ni = NODEPTR(mp, i);
9912 if (ni->mn_flags & F_BIGDATA) {
9916 /* Need writable leaf */
9918 mc.mc_pg[mc.mc_top] = leaf;
9919 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9921 ni = NODEPTR(mp, i);
9924 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9925 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
9926 rc = mdb_page_get(&mc, pg, &omp, NULL);
9929 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9930 rc = mdb_env_cthr_toggle(my, 1);
9933 toggle = my->mc_toggle;
9935 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9936 memcpy(mo, omp, my->mc_env->me_psize);
9937 mo->mp_pgno = my->mc_next_pgno;
9938 my->mc_next_pgno += omp->mp_pages;
9939 my->mc_wlen[toggle] += my->mc_env->me_psize;
9940 if (omp->mp_pages > 1) {
9941 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9942 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9943 rc = mdb_env_cthr_toggle(my, 1);
9946 toggle = my->mc_toggle;
9948 } else if (ni->mn_flags & F_SUBDATA) {
9951 /* Need writable leaf */
9953 mc.mc_pg[mc.mc_top] = leaf;
9954 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9956 ni = NODEPTR(mp, i);
9959 memcpy(&db, NODEDATA(ni), sizeof(db));
9960 my->mc_toggle = toggle;
9961 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9964 toggle = my->mc_toggle;
9965 memcpy(NODEDATA(ni), &db, sizeof(db));
9970 mc.mc_ki[mc.mc_top]++;
9971 if (mc.mc_ki[mc.mc_top] < n) {
9974 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9976 rc = mdb_page_get(&mc, pg, &mp, NULL);
9981 mc.mc_ki[mc.mc_top] = 0;
9982 if (IS_BRANCH(mp)) {
9983 /* Whenever we advance to a sibling branch page,
9984 * we must proceed all the way down to its first leaf.
9986 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9989 mc.mc_pg[mc.mc_top] = mp;
9993 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9994 rc = mdb_env_cthr_toggle(my, 1);
9997 toggle = my->mc_toggle;
9999 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
10000 mdb_page_copy(mo, mp, my->mc_env->me_psize);
10001 mo->mp_pgno = my->mc_next_pgno++;
10002 my->mc_wlen[toggle] += my->mc_env->me_psize;
10004 /* Update parent if there is one */
10005 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
10006 SETPGNO(ni, mo->mp_pgno);
10007 mdb_cursor_pop(&mc);
10009 /* Otherwise we're done */
10019 /** Copy environment with compaction. */
10021 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
10026 MDB_txn *txn = NULL;
10028 pgno_t root, new_root;
10029 int rc = MDB_SUCCESS;
10032 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
10033 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
10037 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
10038 if (my.mc_wbuf[0] == NULL) {
10039 /* _aligned_malloc() sets errno, but we use Windows error codes */
10040 rc = ERROR_NOT_ENOUGH_MEMORY;
10044 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
10046 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
10048 #ifdef HAVE_MEMALIGN
10049 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
10050 if (my.mc_wbuf[0] == NULL) {
10057 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
10063 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
10064 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
10065 my.mc_next_pgno = NUM_METAS;
10068 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
10072 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
10076 mp = (MDB_page *)my.mc_wbuf[0];
10077 memset(mp, 0, NUM_METAS * env->me_psize);
10079 mp->mp_flags = P_META;
10080 mm = (MDB_meta *)METADATA(mp);
10081 mdb_env_init_meta0(env, mm);
10082 mm->mm_address = env->me_metas[0]->mm_address;
10084 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
10086 mp->mp_flags = P_META;
10087 *(MDB_meta *)METADATA(mp) = *mm;
10088 mm = (MDB_meta *)METADATA(mp);
10090 /* Set metapage 1 with current main DB */
10091 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
10092 if (root != P_INVALID) {
10093 /* Count free pages + freeDB pages. Subtract from last_pg
10094 * to find the new last_pg, which also becomes the new root.
10096 MDB_ID freecount = 0;
10099 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
10100 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
10101 freecount += *(MDB_ID *)data.mv_data;
10102 if (rc != MDB_NOTFOUND)
10104 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
10105 txn->mt_dbs[FREE_DBI].md_leaf_pages +
10106 txn->mt_dbs[FREE_DBI].md_overflow_pages;
10108 new_root = txn->mt_next_pgno - 1 - freecount;
10109 mm->mm_last_pg = new_root;
10110 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
10111 mm->mm_dbs[MAIN_DBI].md_root = new_root;
10113 /* When the DB is empty, handle it specially to
10114 * fix any breakage like page leaks from ITS#8174.
10116 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
10118 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
10119 mm->mm_txnid = 1; /* use metapage 1 */
10122 my.mc_wlen[0] = env->me_psize * NUM_METAS;
10124 rc = mdb_env_cwalk(&my, &root, 0);
10125 if (rc == MDB_SUCCESS && root != new_root) {
10126 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
10132 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
10133 rc = THREAD_FINISH(thr);
10134 mdb_txn_abort(txn);
10138 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
10139 if (my.mc_cond) CloseHandle(my.mc_cond);
10140 if (my.mc_mutex) CloseHandle(my.mc_mutex);
10142 free(my.mc_wbuf[0]);
10143 pthread_cond_destroy(&my.mc_cond);
10145 pthread_mutex_destroy(&my.mc_mutex);
10147 return rc ? rc : my.mc_error;
10150 /** Copy environment as-is. */
10152 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
10154 MDB_txn *txn = NULL;
10155 mdb_mutexref_t wmutex = NULL;
10157 mdb_size_t wsize, w3;
10161 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
10165 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
10168 /* Do the lock/unlock of the reader mutex before starting the
10169 * write txn. Otherwise other read txns could block writers.
10171 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
10175 if (env->me_txns) {
10176 /* We must start the actual read txn after blocking writers */
10177 mdb_txn_end(txn, MDB_END_RESET_TMP);
10179 /* Temporarily block writers until we snapshot the meta pages */
10180 wmutex = env->me_wmutex;
10181 if (LOCK_MUTEX(rc, env, wmutex))
10184 rc = mdb_txn_renew0(txn);
10186 UNLOCK_MUTEX(wmutex);
10191 wsize = env->me_psize * NUM_METAS;
10195 DO_WRITE(rc, fd, ptr, w2, len);
10199 } else if (len > 0) {
10205 /* Non-blocking or async handles are not supported */
10211 UNLOCK_MUTEX(wmutex);
10216 w3 = txn->mt_next_pgno * env->me_psize;
10218 mdb_size_t fsize = 0;
10219 if ((rc = mdb_fsize(env->me_fd, &fsize)))
10224 wsize = w3 - wsize;
10225 while (wsize > 0) {
10226 if (wsize > MAX_WRITE)
10230 DO_WRITE(rc, fd, ptr, w2, len);
10234 } else if (len > 0) {
10246 mdb_txn_abort(txn);
10251 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
10253 if (flags & MDB_CP_COMPACT)
10254 return mdb_env_copyfd1(env, fd);
10256 return mdb_env_copyfd0(env, fd);
10260 mdb_env_copyfd(MDB_env *env, HANDLE fd)
10262 return mdb_env_copyfd2(env, fd, 0);
10266 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
10270 HANDLE newfd = INVALID_HANDLE_VALUE;
10272 rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname);
10273 if (rc == MDB_SUCCESS) {
10274 rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd);
10275 mdb_fname_destroy(fname);
10277 if (rc == MDB_SUCCESS) {
10278 rc = mdb_env_copyfd2(env, newfd, flags);
10279 if (close(newfd) < 0 && rc == MDB_SUCCESS)
10286 mdb_env_copy(MDB_env *env, const char *path)
10288 return mdb_env_copy2(env, path, 0);
10292 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
10294 if (flag & ~CHANGEABLE)
10297 env->me_flags |= flag;
10299 env->me_flags &= ~flag;
10300 return MDB_SUCCESS;
10304 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
10309 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
10310 return MDB_SUCCESS;
10314 mdb_env_set_userctx(MDB_env *env, void *ctx)
10318 env->me_userctx = ctx;
10319 return MDB_SUCCESS;
10323 mdb_env_get_userctx(MDB_env *env)
10325 return env ? env->me_userctx : NULL;
10329 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
10334 env->me_assert_func = func;
10336 return MDB_SUCCESS;
10340 mdb_env_get_path(MDB_env *env, const char **arg)
10345 *arg = env->me_path;
10346 return MDB_SUCCESS;
10350 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
10356 return MDB_SUCCESS;
10359 /** Common code for #mdb_stat() and #mdb_env_stat().
10360 * @param[in] env the environment to operate in.
10361 * @param[in] db the #MDB_db record containing the stats to return.
10362 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
10363 * @return 0, this function always succeeds.
10366 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
10368 arg->ms_psize = env->me_psize;
10369 arg->ms_depth = db->md_depth;
10370 arg->ms_branch_pages = db->md_branch_pages;
10371 arg->ms_leaf_pages = db->md_leaf_pages;
10372 arg->ms_overflow_pages = db->md_overflow_pages;
10373 arg->ms_entries = db->md_entries;
10375 return MDB_SUCCESS;
10379 mdb_env_stat(MDB_env *env, MDB_stat *arg)
10383 if (env == NULL || arg == NULL)
10386 meta = mdb_env_pick_meta(env);
10388 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
10392 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
10396 if (env == NULL || arg == NULL)
10399 meta = mdb_env_pick_meta(env);
10400 arg->me_mapaddr = meta->mm_address;
10401 arg->me_last_pgno = meta->mm_last_pg;
10402 arg->me_last_txnid = meta->mm_txnid;
10404 arg->me_mapsize = env->me_mapsize;
10405 arg->me_maxreaders = env->me_maxreaders;
10406 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
10407 return MDB_SUCCESS;
10410 /** Set the default comparison functions for a database.
10411 * Called immediately after a database is opened to set the defaults.
10412 * The user can then override them with #mdb_set_compare() or
10413 * #mdb_set_dupsort().
10414 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
10415 * @param[in] dbi A database handle returned by #mdb_dbi_open()
10418 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
10420 uint16_t f = txn->mt_dbs[dbi].md_flags;
10422 txn->mt_dbxs[dbi].md_cmp =
10423 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
10424 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
10426 txn->mt_dbxs[dbi].md_dcmp =
10427 !(f & MDB_DUPSORT) ? 0 :
10428 ((f & MDB_INTEGERDUP)
10429 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
10430 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
10433 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
10439 int rc, dbflag, exact;
10440 unsigned int unused = 0, seq;
10444 if (flags & ~VALID_FLAGS)
10446 if (txn->mt_flags & MDB_TXN_BLOCKED)
10447 return MDB_BAD_TXN;
10452 if (flags & PERSISTENT_FLAGS) {
10453 uint16_t f2 = flags & PERSISTENT_FLAGS;
10454 /* make sure flag changes get committed */
10455 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
10456 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
10457 txn->mt_flags |= MDB_TXN_DIRTY;
10460 mdb_default_cmp(txn, MAIN_DBI);
10461 return MDB_SUCCESS;
10464 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
10465 mdb_default_cmp(txn, MAIN_DBI);
10468 /* Is the DB already open? */
10469 len = strlen(name);
10470 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
10471 if (!txn->mt_dbxs[i].md_name.mv_size) {
10472 /* Remember this free slot */
10473 if (!unused) unused = i;
10476 if (len == txn->mt_dbxs[i].md_name.mv_size &&
10477 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
10479 return MDB_SUCCESS;
10483 /* If no free slot and max hit, fail */
10484 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
10485 return MDB_DBS_FULL;
10487 /* Cannot mix named databases with some mainDB flags */
10488 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
10489 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
10491 /* Find the DB info */
10492 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
10495 key.mv_data = (void *)name;
10496 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
10497 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
10498 if (rc == MDB_SUCCESS) {
10499 /* make sure this is actually a DB */
10500 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
10501 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
10502 return MDB_INCOMPATIBLE;
10503 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
10507 /* Done here so we cannot fail after creating a new DB */
10508 if ((namedup = strdup(name)) == NULL)
10512 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
10513 data.mv_size = sizeof(MDB_db);
10514 data.mv_data = &dummy;
10515 memset(&dummy, 0, sizeof(dummy));
10516 dummy.md_root = P_INVALID;
10517 dummy.md_flags = flags & PERSISTENT_FLAGS;
10518 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
10519 dbflag |= DB_DIRTY;
10525 /* Got info, register DBI in this txn */
10526 unsigned int slot = unused ? unused : txn->mt_numdbs;
10527 txn->mt_dbxs[slot].md_name.mv_data = namedup;
10528 txn->mt_dbxs[slot].md_name.mv_size = len;
10529 txn->mt_dbxs[slot].md_rel = NULL;
10530 txn->mt_dbflags[slot] = dbflag;
10531 /* txn-> and env-> are the same in read txns, use
10532 * tmp variable to avoid undefined assignment
10534 seq = ++txn->mt_env->me_dbiseqs[slot];
10535 txn->mt_dbiseqs[slot] = seq;
10537 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
10539 mdb_default_cmp(txn, slot);
10549 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
10551 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
10554 if (txn->mt_flags & MDB_TXN_BLOCKED)
10555 return MDB_BAD_TXN;
10557 if (txn->mt_dbflags[dbi] & DB_STALE) {
10560 /* Stale, must read the DB's root. cursor_init does it for us. */
10561 mdb_cursor_init(&mc, txn, dbi, &mx);
10563 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
10566 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
10569 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
10571 ptr = env->me_dbxs[dbi].md_name.mv_data;
10572 /* If there was no name, this was already closed */
10574 env->me_dbxs[dbi].md_name.mv_data = NULL;
10575 env->me_dbxs[dbi].md_name.mv_size = 0;
10576 env->me_dbflags[dbi] = 0;
10577 env->me_dbiseqs[dbi]++;
10582 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
10584 /* We could return the flags for the FREE_DBI too but what's the point? */
10585 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10587 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
10588 return MDB_SUCCESS;
10591 /** Add all the DB's pages to the free list.
10592 * @param[in] mc Cursor on the DB to free.
10593 * @param[in] subs non-Zero to check for sub-DBs in this DB.
10594 * @return 0 on success, non-zero on failure.
10597 mdb_drop0(MDB_cursor *mc, int subs)
10601 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
10602 if (rc == MDB_SUCCESS) {
10603 MDB_txn *txn = mc->mc_txn;
10608 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
10609 * This also avoids any P_LEAF2 pages, which have no nodes.
10610 * Also if the DB doesn't have sub-DBs and has no overflow
10611 * pages, omit scanning leaves.
10613 if ((mc->mc_flags & C_SUB) ||
10614 (!subs && !mc->mc_db->md_overflow_pages))
10615 mdb_cursor_pop(mc);
10617 mdb_cursor_copy(mc, &mx);
10619 /* bump refcount for mx's pages */
10620 for (i=0; i<mc->mc_snum; i++)
10621 mdb_page_get(&mx, mc->mc_pg[i]->mp_pgno, &mx.mc_pg[i], NULL);
10623 while (mc->mc_snum > 0) {
10624 MDB_page *mp = mc->mc_pg[mc->mc_top];
10625 unsigned n = NUMKEYS(mp);
10627 for (i=0; i<n; i++) {
10628 ni = NODEPTR(mp, i);
10629 if (ni->mn_flags & F_BIGDATA) {
10632 memcpy(&pg, NODEDATA(ni), sizeof(pg));
10633 rc = mdb_page_get(mc, pg, &omp, NULL);
10636 mdb_cassert(mc, IS_OVERFLOW(omp));
10637 rc = mdb_midl_append_range(&txn->mt_free_pgs,
10638 pg, omp->mp_pages);
10641 mc->mc_db->md_overflow_pages -= omp->mp_pages;
10642 if (!mc->mc_db->md_overflow_pages && !subs)
10644 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
10645 mdb_xcursor_init1(mc, ni);
10646 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
10651 if (!subs && !mc->mc_db->md_overflow_pages)
10654 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
10656 for (i=0; i<n; i++) {
10658 ni = NODEPTR(mp, i);
10661 mdb_midl_xappend(txn->mt_free_pgs, pg);
10666 mc->mc_ki[mc->mc_top] = i;
10667 rc = mdb_cursor_sibling(mc, 1);
10669 if (rc != MDB_NOTFOUND)
10671 /* no more siblings, go back to beginning
10672 * of previous level.
10675 mdb_cursor_pop(mc);
10677 for (i=1; i<mc->mc_snum; i++) {
10679 mc->mc_pg[i] = mx.mc_pg[i];
10684 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
10687 txn->mt_flags |= MDB_TXN_ERROR;
10688 /* drop refcount for mx's pages */
10689 MDB_CURSOR_UNREF(&mx, 0);
10690 } else if (rc == MDB_NOTFOUND) {
10693 mc->mc_flags &= ~C_INITIALIZED;
10697 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
10699 MDB_cursor *mc, *m2;
10702 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10705 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
10708 if (TXN_DBI_CHANGED(txn, dbi))
10709 return MDB_BAD_DBI;
10711 rc = mdb_cursor_open(txn, dbi, &mc);
10715 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
10716 /* Invalidate the dropped DB's cursors */
10717 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
10718 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
10722 /* Can't delete the main DB */
10723 if (del && dbi >= CORE_DBS) {
10724 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
10726 txn->mt_dbflags[dbi] = DB_STALE;
10727 mdb_dbi_close(txn->mt_env, dbi);
10729 txn->mt_flags |= MDB_TXN_ERROR;
10732 /* reset the DB record, mark it dirty */
10733 txn->mt_dbflags[dbi] |= DB_DIRTY;
10734 txn->mt_dbs[dbi].md_depth = 0;
10735 txn->mt_dbs[dbi].md_branch_pages = 0;
10736 txn->mt_dbs[dbi].md_leaf_pages = 0;
10737 txn->mt_dbs[dbi].md_overflow_pages = 0;
10738 txn->mt_dbs[dbi].md_entries = 0;
10739 txn->mt_dbs[dbi].md_root = P_INVALID;
10741 txn->mt_flags |= MDB_TXN_DIRTY;
10744 mdb_cursor_close(mc);
10748 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10750 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10753 txn->mt_dbxs[dbi].md_cmp = cmp;
10754 return MDB_SUCCESS;
10757 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10759 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10762 txn->mt_dbxs[dbi].md_dcmp = cmp;
10763 return MDB_SUCCESS;
10766 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
10768 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10771 txn->mt_dbxs[dbi].md_rel = rel;
10772 return MDB_SUCCESS;
10775 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
10777 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10780 txn->mt_dbxs[dbi].md_relctx = ctx;
10781 return MDB_SUCCESS;
10785 mdb_env_get_maxkeysize(MDB_env *env)
10787 return ENV_MAXKEY(env);
10791 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
10793 unsigned int i, rdrs;
10796 int rc = 0, first = 1;
10800 if (!env->me_txns) {
10801 return func("(no reader locks)\n", ctx);
10803 rdrs = env->me_txns->mti_numreaders;
10804 mr = env->me_txns->mti_readers;
10805 for (i=0; i<rdrs; i++) {
10806 if (mr[i].mr_pid) {
10807 txnid_t txnid = mr[i].mr_txnid;
10808 sprintf(buf, txnid == (txnid_t)-1 ?
10809 "%10d %"Z"x -\n" : "%10d %"Z"x %"Yu"\n",
10810 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10813 rc = func(" pid thread txnid\n", ctx);
10817 rc = func(buf, ctx);
10823 rc = func("(no active readers)\n", ctx);
10828 /** Insert pid into list if not already present.
10829 * return -1 if already present.
10832 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10834 /* binary search of pid in list */
10836 unsigned cursor = 1;
10838 unsigned n = ids[0];
10841 unsigned pivot = n >> 1;
10842 cursor = base + pivot + 1;
10843 val = pid - ids[cursor];
10848 } else if ( val > 0 ) {
10853 /* found, so it's a duplicate */
10862 for (n = ids[0]; n > cursor; n--)
10869 mdb_reader_check(MDB_env *env, int *dead)
10875 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10878 /** As #mdb_reader_check(). \b rlocked is set if caller locked #me_rmutex. */
10880 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10882 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10883 unsigned int i, j, rdrs;
10885 MDB_PID_T *pids, pid;
10886 int rc = MDB_SUCCESS, count = 0;
10888 rdrs = env->me_txns->mti_numreaders;
10889 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10893 mr = env->me_txns->mti_readers;
10894 for (i=0; i<rdrs; i++) {
10895 pid = mr[i].mr_pid;
10896 if (pid && pid != env->me_pid) {
10897 if (mdb_pid_insert(pids, pid) == 0) {
10898 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10899 /* Stale reader found */
10902 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10903 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10905 rdrs = 0; /* the above checked all readers */
10907 /* Recheck, a new process may have reused pid */
10908 if (mdb_reader_pid(env, Pidcheck, pid))
10912 for (; j<rdrs; j++)
10913 if (mr[j].mr_pid == pid) {
10914 DPRINTF(("clear stale reader pid %u txn %"Yd,
10915 (unsigned) pid, mr[j].mr_txnid));
10920 UNLOCK_MUTEX(rmutex);
10931 #ifdef MDB_ROBUST_SUPPORTED
10932 /** Handle #LOCK_MUTEX0() failure.
10933 * Try to repair the lock file if the mutex owner died.
10934 * @param[in] env the environment handle
10935 * @param[in] mutex LOCK_MUTEX0() mutex
10936 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10937 * @return 0 on success with the mutex locked, or an error code on failure.
10940 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10945 if (rc == MDB_OWNERDEAD) {
10946 /* We own the mutex. Clean up after dead previous owner. */
10948 rlocked = (mutex == env->me_rmutex);
10950 /* Keep mti_txnid updated, otherwise next writer can
10951 * overwrite data which latest meta page refers to.
10953 meta = mdb_env_pick_meta(env);
10954 env->me_txns->mti_txnid = meta->mm_txnid;
10955 /* env is hosed if the dead thread was ours */
10957 env->me_flags |= MDB_FATAL_ERROR;
10958 env->me_txn = NULL;
10962 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10963 (rc ? "this process' env is hosed" : "recovering")));
10964 rc2 = mdb_reader_check0(env, rlocked, NULL);
10966 rc2 = mdb_mutex_consistent(mutex);
10967 if (rc || (rc = rc2)) {
10968 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10969 UNLOCK_MUTEX(mutex);
10975 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10980 #endif /* MDB_ROBUST_SUPPORTED */
10982 #if defined(_WIN32)
10983 /** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */
10985 utf8_to_utf16(const char *src, MDB_name *dst, int xtra)
10988 wchar_t *result = NULL;
10989 for (;;) { /* malloc result, then fill it in */
10990 need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need);
10997 result = malloc(sizeof(wchar_t) * (need + xtra));
11002 dst->mn_alloced = 1;
11003 dst->mn_len = need - 1;
11004 dst->mn_val = result;
11005 return MDB_SUCCESS;
11008 #endif /* defined(_WIN32) */