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 modified or is DUPSORT data */
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 */
1223 /** In write txns, array of cursors for each DB */
1224 MDB_cursor **mt_cursors;
1225 /** Array of flags for each DB */
1226 unsigned char *mt_dbflags;
1228 /** List of read-only pages (actually chunks) */
1230 /** We map chunks of 16 pages. Even though Windows uses 4KB pages, all
1231 * mappings must begin on 64KB boundaries. So we round off all pgnos to
1232 * a chunk boundary. We do the same on Linux for symmetry, and also to
1233 * reduce the frequency of mmap/munmap calls.
1235 #define MDB_RPAGE_CHUNK 16
1236 #define MDB_TRPAGE_SIZE 4096 /**< size of #mt_rpages array of chunks */
1237 #define MDB_TRPAGE_MAX (MDB_TRPAGE_SIZE-1) /**< maximum chunk index */
1238 unsigned int mt_rpcheck; /**< threshold for reclaiming unref'd chunks */
1240 /** Number of DB records in use, or 0 when the txn is finished.
1241 * This number only ever increments until the txn finishes; we
1242 * don't decrement it when individual DB handles are closed.
1246 /** @defgroup mdb_txn Transaction Flags
1250 /** #mdb_txn_begin() flags */
1251 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1252 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1253 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1254 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1255 /* internal txn flags */
1256 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1257 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1258 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1259 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1260 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1261 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1262 /** most operations on the txn are currently illegal */
1263 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1265 unsigned int mt_flags; /**< @ref mdb_txn */
1266 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1267 * Includes ancestor txns' dirty pages not hidden by other txns'
1268 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1269 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1271 unsigned int mt_dirty_room;
1274 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1275 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1276 * raise this on a 64 bit machine.
1278 #define CURSOR_STACK 32
1282 /** Cursors are used for all DB operations.
1283 * A cursor holds a path of (page pointer, key index) from the DB
1284 * root to a position in the DB, plus other state. #MDB_DUPSORT
1285 * cursors include an xcursor to the current data item. Write txns
1286 * track their cursors and keep them up to date when data moves.
1287 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1288 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1291 /** Next cursor on this DB in this txn */
1292 MDB_cursor *mc_next;
1293 /** Backup of the original cursor if this cursor is a shadow */
1294 MDB_cursor *mc_backup;
1295 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1296 struct MDB_xcursor *mc_xcursor;
1297 /** The transaction that owns this cursor */
1299 /** The database handle this cursor operates on */
1301 /** The database record for this cursor */
1303 /** The database auxiliary record for this cursor */
1305 /** The @ref mt_dbflag for this database */
1306 unsigned char *mc_dbflag;
1307 unsigned short mc_snum; /**< number of pushed pages */
1308 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1309 /** @defgroup mdb_cursor Cursor Flags
1311 * Cursor state flags.
1314 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1315 #define C_EOF 0x02 /**< No more data */
1316 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1317 #define C_DEL 0x08 /**< last op was a cursor_del */
1318 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1319 #define C_WRITEMAP MDB_TXN_WRITEMAP /**< Copy of txn flag */
1320 /** Read-only cursor into the txn's original snapshot in the map.
1321 * Set for read-only txns, and in #mdb_page_alloc() for #FREE_DBI when
1322 * #MDB_DEVEL & 2. Only implements code which is necessary for this.
1324 #define C_ORIG_RDONLY MDB_TXN_RDONLY
1326 unsigned int mc_flags; /**< @ref mdb_cursor */
1327 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1328 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1330 MDB_page *mc_ovpg; /**< a referenced overflow page */
1331 # define MC_OVPG(mc) ((mc)->mc_ovpg)
1332 # define MC_SET_OVPG(mc, pg) ((mc)->mc_ovpg = (pg))
1334 # define MC_OVPG(mc) ((MDB_page *)0)
1335 # define MC_SET_OVPG(mc, pg) ((void)0)
1339 /** Context for sorted-dup records.
1340 * We could have gone to a fully recursive design, with arbitrarily
1341 * deep nesting of sub-databases. But for now we only handle these
1342 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1344 typedef struct MDB_xcursor {
1345 /** A sub-cursor for traversing the Dup DB */
1346 MDB_cursor mx_cursor;
1347 /** The database record for this Dup DB */
1349 /** The auxiliary DB record for this Dup DB */
1351 /** The @ref mt_dbflag for this Dup DB */
1352 unsigned char mx_dbflag;
1355 /** State of FreeDB old pages, stored in the MDB_env */
1356 typedef struct MDB_pgstate {
1357 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1358 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1361 /** The database environment. */
1363 HANDLE me_fd; /**< The main data file */
1364 HANDLE me_lfd; /**< The lock file */
1365 HANDLE me_mfd; /**< For writing and syncing the meta pages */
1366 #if defined(MDB_VL32) && defined(_WIN32)
1367 HANDLE me_fmh; /**< File Mapping handle */
1369 /** Failed to update the meta page. Probably an I/O error. */
1370 #define MDB_FATAL_ERROR 0x80000000U
1371 /** Some fields are initialized. */
1372 #define MDB_ENV_ACTIVE 0x20000000U
1373 /** me_txkey is set */
1374 #define MDB_ENV_TXKEY 0x10000000U
1375 /** fdatasync is unreliable */
1376 #define MDB_FSYNCONLY 0x08000000U
1377 uint32_t me_flags; /**< @ref mdb_env */
1378 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1379 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1380 unsigned int me_maxreaders; /**< size of the reader table */
1381 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1382 volatile int me_close_readers;
1383 MDB_dbi me_numdbs; /**< number of DBs opened */
1384 MDB_dbi me_maxdbs; /**< size of the DB table */
1385 MDB_PID_T me_pid; /**< process ID of this env */
1386 char *me_path; /**< path to the DB files */
1387 char *me_map; /**< the memory map of the data file */
1388 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1389 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1390 void *me_pbuf; /**< scratch area for DUPSORT put() */
1391 MDB_txn *me_txn; /**< current write transaction */
1392 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1393 mdb_size_t me_mapsize; /**< size of the data memory map */
1394 off_t me_size; /**< current file size */
1395 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1396 MDB_dbx *me_dbxs; /**< array of static DB info */
1397 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1398 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1399 pthread_key_t me_txkey; /**< thread-key for readers */
1400 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1401 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1402 # define me_pglast me_pgstate.mf_pglast
1403 # define me_pghead me_pgstate.mf_pghead
1404 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1405 /** IDL of pages that became unused in a write txn */
1406 MDB_IDL me_free_pgs;
1407 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1408 MDB_ID2L me_dirty_list;
1409 /** Max number of freelist items that can fit in a single overflow page */
1411 /** Max size of a node on a page */
1412 unsigned int me_nodemax;
1413 #if !(MDB_MAXKEYSIZE)
1414 unsigned int me_maxkey; /**< max size of a key */
1416 int me_live_reader; /**< have liveness lock in reader table */
1418 int me_pidquery; /**< Used in OpenProcess */
1420 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1421 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1422 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1424 mdb_mutex_t me_rmutex;
1425 mdb_mutex_t me_wmutex;
1428 MDB_ID3L me_rpages; /**< like #mt_rpages, but global to env */
1429 pthread_mutex_t me_rpmutex; /**< control access to #me_rpages */
1430 #define MDB_ERPAGE_SIZE 16384
1431 #define MDB_ERPAGE_MAX (MDB_ERPAGE_SIZE-1)
1432 unsigned int me_rpcheck;
1434 void *me_userctx; /**< User-settable context */
1435 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1438 /** Nested transaction */
1439 typedef struct MDB_ntxn {
1440 MDB_txn mnt_txn; /**< the transaction */
1441 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1444 /** max number of pages to commit in one writev() call */
1445 #define MDB_COMMIT_PAGES 64
1446 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1447 #undef MDB_COMMIT_PAGES
1448 #define MDB_COMMIT_PAGES IOV_MAX
1451 /** max bytes to write in one call */
1452 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1454 /** Check \b txn and \b dbi arguments to a function */
1455 #define TXN_DBI_EXIST(txn, dbi, validity) \
1456 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1458 /** Check for misused \b dbi handles */
1459 #define TXN_DBI_CHANGED(txn, dbi) \
1460 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1462 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1463 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1464 static int mdb_page_touch(MDB_cursor *mc);
1466 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1467 "reset-tmp", "fail-begin", "fail-beginchild"}
1469 /* mdb_txn_end operation number, for logging */
1470 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1471 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1473 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1474 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1475 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1476 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1477 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1479 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1480 static int mdb_page_search_root(MDB_cursor *mc,
1481 MDB_val *key, int modify);
1482 #define MDB_PS_MODIFY 1
1483 #define MDB_PS_ROOTONLY 2
1484 #define MDB_PS_FIRST 4
1485 #define MDB_PS_LAST 8
1486 static int mdb_page_search(MDB_cursor *mc,
1487 MDB_val *key, int flags);
1488 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1490 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1491 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1492 pgno_t newpgno, unsigned int nflags);
1494 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1495 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1496 static int mdb_env_write_meta(MDB_txn *txn);
1497 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1498 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1500 static void mdb_env_close0(MDB_env *env, int excl);
1502 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1503 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1504 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1505 static void mdb_node_del(MDB_cursor *mc, int ksize);
1506 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1507 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1508 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1509 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1510 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1512 static int mdb_rebalance(MDB_cursor *mc);
1513 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1515 static void mdb_cursor_pop(MDB_cursor *mc);
1516 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1518 static int mdb_cursor_del0(MDB_cursor *mc);
1519 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1520 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1521 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1522 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1523 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1525 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1526 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1528 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1529 static void mdb_xcursor_init0(MDB_cursor *mc);
1530 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1531 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1533 static int mdb_drop0(MDB_cursor *mc, int subs);
1534 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1535 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1538 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1541 /** Compare two items pointing at '#mdb_size_t's of unknown alignment. */
1542 #ifdef MISALIGNED_OK
1543 # define mdb_cmp_clong mdb_cmp_long
1545 # define mdb_cmp_clong mdb_cmp_cint
1548 /** True if we need #mdb_cmp_clong() instead of \b cmp for #MDB_INTEGERDUP */
1549 #define NEED_CMP_CLONG(cmp, ksize) \
1550 (UINT_MAX < MDB_SIZE_MAX && \
1551 (cmp) == mdb_cmp_int && (ksize) == sizeof(mdb_size_t))
1554 static SECURITY_DESCRIPTOR mdb_null_sd;
1555 static SECURITY_ATTRIBUTES mdb_all_sa;
1556 static int mdb_sec_inited;
1559 static int utf8_to_utf16(const char *src, struct MDB_name *dst, int xtra);
1562 /** Return the library version info. */
1564 mdb_version(int *major, int *minor, int *patch)
1566 if (major) *major = MDB_VERSION_MAJOR;
1567 if (minor) *minor = MDB_VERSION_MINOR;
1568 if (patch) *patch = MDB_VERSION_PATCH;
1569 return MDB_VERSION_STRING;
1572 /** Table of descriptions for LMDB @ref errors */
1573 static char *const mdb_errstr[] = {
1574 "MDB_KEYEXIST: Key/data pair already exists",
1575 "MDB_NOTFOUND: No matching key/data pair found",
1576 "MDB_PAGE_NOTFOUND: Requested page not found",
1577 "MDB_CORRUPTED: Located page was wrong type",
1578 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1579 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1580 "MDB_INVALID: File is not an LMDB file",
1581 "MDB_MAP_FULL: Environment mapsize limit reached",
1582 "MDB_DBS_FULL: Environment maxdbs limit reached",
1583 "MDB_READERS_FULL: Environment maxreaders limit reached",
1584 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1585 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1586 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1587 "MDB_PAGE_FULL: Internal error - page has no more space",
1588 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1589 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1590 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1591 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1592 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1593 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1594 "MDB_PROBLEM: Unexpected problem - txn should abort",
1598 mdb_strerror(int err)
1601 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1602 * This works as long as no function between the call to mdb_strerror
1603 * and the actual use of the message uses more than 4K of stack.
1605 #define MSGSIZE 1024
1606 #define PADSIZE 4096
1607 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1611 return ("Successful return: 0");
1613 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1614 i = err - MDB_KEYEXIST;
1615 return mdb_errstr[i];
1619 /* These are the C-runtime error codes we use. The comment indicates
1620 * their numeric value, and the Win32 error they would correspond to
1621 * if the error actually came from a Win32 API. A major mess, we should
1622 * have used LMDB-specific error codes for everything.
1625 case ENOENT: /* 2, FILE_NOT_FOUND */
1626 case EIO: /* 5, ACCESS_DENIED */
1627 case ENOMEM: /* 12, INVALID_ACCESS */
1628 case EACCES: /* 13, INVALID_DATA */
1629 case EBUSY: /* 16, CURRENT_DIRECTORY */
1630 case EINVAL: /* 22, BAD_COMMAND */
1631 case ENOSPC: /* 28, OUT_OF_PAPER */
1632 return strerror(err);
1637 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1638 FORMAT_MESSAGE_IGNORE_INSERTS,
1639 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1642 return strerror(err);
1646 /** assert(3) variant in cursor context */
1647 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1648 /** assert(3) variant in transaction context */
1649 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1650 /** assert(3) variant in environment context */
1651 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1654 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1655 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1658 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1659 const char *func, const char *file, int line)
1662 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1663 file, line, expr_txt, func);
1664 if (env->me_assert_func)
1665 env->me_assert_func(env, buf);
1666 fprintf(stderr, "%s\n", buf);
1670 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1674 /** Return the page number of \b mp which may be sub-page, for debug output */
1676 mdb_dbg_pgno(MDB_page *mp)
1679 COPY_PGNO(ret, mp->mp_pgno);
1683 /** Display a key in hexadecimal and return the address of the result.
1684 * @param[in] key the key to display
1685 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1686 * @return The key in hexadecimal form.
1689 mdb_dkey(MDB_val *key, char *buf)
1692 unsigned char *c = key->mv_data;
1698 if (key->mv_size > DKBUF_MAXKEYSIZE)
1699 return "MDB_MAXKEYSIZE";
1700 /* may want to make this a dynamic check: if the key is mostly
1701 * printable characters, print it as-is instead of converting to hex.
1705 for (i=0; i<key->mv_size; i++)
1706 ptr += sprintf(ptr, "%02x", *c++);
1708 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1714 mdb_leafnode_type(MDB_node *n)
1716 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1717 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1718 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1721 /** Display all the keys in the page. */
1723 mdb_page_list(MDB_page *mp)
1725 pgno_t pgno = mdb_dbg_pgno(mp);
1726 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1728 unsigned int i, nkeys, nsize, total = 0;
1732 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1733 case P_BRANCH: type = "Branch page"; break;
1734 case P_LEAF: type = "Leaf page"; break;
1735 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1736 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1737 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1739 fprintf(stderr, "Overflow page %"Yu" pages %u%s\n",
1740 pgno, mp->mp_pages, state);
1743 fprintf(stderr, "Meta-page %"Yu" txnid %"Yu"\n",
1744 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1747 fprintf(stderr, "Bad page %"Yu" flags 0x%X\n", pgno, mp->mp_flags);
1751 nkeys = NUMKEYS(mp);
1752 fprintf(stderr, "%s %"Yu" numkeys %d%s\n", type, pgno, nkeys, state);
1754 for (i=0; i<nkeys; i++) {
1755 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1756 key.mv_size = nsize = mp->mp_pad;
1757 key.mv_data = LEAF2KEY(mp, i, nsize);
1759 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1762 node = NODEPTR(mp, i);
1763 key.mv_size = node->mn_ksize;
1764 key.mv_data = node->mn_data;
1765 nsize = NODESIZE + key.mv_size;
1766 if (IS_BRANCH(mp)) {
1767 fprintf(stderr, "key %d: page %"Yu", %s\n", i, NODEPGNO(node),
1771 if (F_ISSET(node->mn_flags, F_BIGDATA))
1772 nsize += sizeof(pgno_t);
1774 nsize += NODEDSZ(node);
1776 nsize += sizeof(indx_t);
1777 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1778 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1780 total = EVEN(total);
1782 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1783 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1787 mdb_cursor_chk(MDB_cursor *mc)
1793 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1794 for (i=0; i<mc->mc_top; i++) {
1796 node = NODEPTR(mp, mc->mc_ki[i]);
1797 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1800 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1802 if (mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
1803 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1804 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1805 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1813 /** Count all the pages in each DB and in the freelist
1814 * and make sure it matches the actual number of pages
1816 * All named DBs must be open for a correct count.
1818 static void mdb_audit(MDB_txn *txn)
1822 MDB_ID freecount, count;
1827 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1828 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1829 freecount += *(MDB_ID *)data.mv_data;
1830 mdb_tassert(txn, rc == MDB_NOTFOUND);
1833 for (i = 0; i<txn->mt_numdbs; i++) {
1835 if (!(txn->mt_dbflags[i] & DB_VALID))
1837 mdb_cursor_init(&mc, txn, i, &mx);
1838 if (txn->mt_dbs[i].md_root == P_INVALID)
1840 count += txn->mt_dbs[i].md_branch_pages +
1841 txn->mt_dbs[i].md_leaf_pages +
1842 txn->mt_dbs[i].md_overflow_pages;
1843 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1844 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1845 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1848 mp = mc.mc_pg[mc.mc_top];
1849 for (j=0; j<NUMKEYS(mp); j++) {
1850 MDB_node *leaf = NODEPTR(mp, j);
1851 if (leaf->mn_flags & F_SUBDATA) {
1853 memcpy(&db, NODEDATA(leaf), sizeof(db));
1854 count += db.md_branch_pages + db.md_leaf_pages +
1855 db.md_overflow_pages;
1859 mdb_tassert(txn, rc == MDB_NOTFOUND);
1862 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1863 fprintf(stderr, "audit: %"Yu" freecount: %"Yu" count: %"Yu" total: %"Yu" next_pgno: %"Yu"\n",
1864 txn->mt_txnid, freecount, count+NUM_METAS,
1865 freecount+count+NUM_METAS, txn->mt_next_pgno);
1871 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1873 return txn->mt_dbxs[dbi].md_cmp(a, b);
1877 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1879 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1880 if (NEED_CMP_CLONG(dcmp, a->mv_size))
1881 dcmp = mdb_cmp_clong;
1885 /** Allocate memory for a page.
1886 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1889 mdb_page_malloc(MDB_txn *txn, unsigned num)
1891 MDB_env *env = txn->mt_env;
1892 MDB_page *ret = env->me_dpages;
1893 size_t psize = env->me_psize, sz = psize, off;
1894 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1895 * For a single page alloc, we init everything after the page header.
1896 * For multi-page, we init the final page; if the caller needed that
1897 * many pages they will be filling in at least up to the last page.
1901 VGMEMP_ALLOC(env, ret, sz);
1902 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1903 env->me_dpages = ret->mp_next;
1906 psize -= off = PAGEHDRSZ;
1911 if ((ret = malloc(sz)) != NULL) {
1912 VGMEMP_ALLOC(env, ret, sz);
1913 if (!(env->me_flags & MDB_NOMEMINIT)) {
1914 memset((char *)ret + off, 0, psize);
1918 txn->mt_flags |= MDB_TXN_ERROR;
1922 /** Free a single page.
1923 * Saves single pages to a list, for future reuse.
1924 * (This is not used for multi-page overflow pages.)
1927 mdb_page_free(MDB_env *env, MDB_page *mp)
1929 mp->mp_next = env->me_dpages;
1930 VGMEMP_FREE(env, mp);
1931 env->me_dpages = mp;
1934 /** Free a dirty page */
1936 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1938 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1939 mdb_page_free(env, dp);
1941 /* large pages just get freed directly */
1942 VGMEMP_FREE(env, dp);
1947 /** Return all dirty pages to dpage list */
1949 mdb_dlist_free(MDB_txn *txn)
1951 MDB_env *env = txn->mt_env;
1952 MDB_ID2L dl = txn->mt_u.dirty_list;
1953 unsigned i, n = dl[0].mid;
1955 for (i = 1; i <= n; i++) {
1956 mdb_dpage_free(env, dl[i].mptr);
1963 mdb_page_unref(MDB_txn *txn, MDB_page *mp)
1966 MDB_ID3L tl = txn->mt_rpages;
1968 if (mp->mp_flags & (P_SUBP|P_DIRTY))
1970 rem = mp->mp_pgno & (MDB_RPAGE_CHUNK-1);
1971 pgno = mp->mp_pgno ^ rem;
1972 x = mdb_mid3l_search(tl, pgno);
1973 if (x != tl[0].mid && tl[x+1].mid == mp->mp_pgno)
1978 #define MDB_PAGE_UNREF(txn, mp) mdb_page_unref(txn, mp)
1981 mdb_cursor_unref(MDB_cursor *mc)
1984 if (!mc->mc_snum || !mc->mc_pg[0] || IS_SUBP(mc->mc_pg[0]))
1986 for (i=0; i<mc->mc_snum; i++)
1987 mdb_page_unref(mc->mc_txn, mc->mc_pg[i]);
1989 mdb_page_unref(mc->mc_txn, mc->mc_ovpg);
1992 mc->mc_snum = mc->mc_top = 0;
1993 mc->mc_pg[0] = NULL;
1994 mc->mc_flags &= ~C_INITIALIZED;
1996 #define MDB_CURSOR_UNREF(mc, force) \
1997 (((force) || ((mc)->mc_flags & C_INITIALIZED)) \
1998 ? mdb_cursor_unref(mc) \
2002 #define MDB_PAGE_UNREF(txn, mp)
2003 #define MDB_CURSOR_UNREF(mc, force) ((void)0)
2004 #endif /* MDB_VL32 */
2006 /** Loosen or free a single page.
2007 * Saves single pages to a list for future reuse
2008 * in this same txn. It has been pulled from the freeDB
2009 * and already resides on the dirty list, but has been
2010 * deleted. Use these pages first before pulling again
2013 * If the page wasn't dirtied in this txn, just add it
2014 * to this txn's free list.
2017 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
2020 pgno_t pgno = mp->mp_pgno;
2021 MDB_txn *txn = mc->mc_txn;
2023 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
2024 if (txn->mt_parent) {
2025 MDB_ID2 *dl = txn->mt_u.dirty_list;
2026 /* If txn has a parent, make sure the page is in our
2030 unsigned x = mdb_mid2l_search(dl, pgno);
2031 if (x <= dl[0].mid && dl[x].mid == pgno) {
2032 if (mp != dl[x].mptr) { /* bad cursor? */
2033 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2034 txn->mt_flags |= MDB_TXN_ERROR;
2042 /* no parent txn, so it's just ours */
2047 DPRINTF(("loosen db %d page %"Yu, DDBI(mc), mp->mp_pgno));
2048 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
2049 txn->mt_loose_pgs = mp;
2050 txn->mt_loose_count++;
2051 mp->mp_flags |= P_LOOSE;
2053 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
2061 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
2062 * @param[in] mc A cursor handle for the current operation.
2063 * @param[in] pflags Flags of the pages to update:
2064 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
2065 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
2066 * @return 0 on success, non-zero on failure.
2069 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
2071 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
2072 MDB_txn *txn = mc->mc_txn;
2073 MDB_cursor *m3, *m0 = mc;
2078 int rc = MDB_SUCCESS, level;
2080 /* Mark pages seen by cursors */
2081 if (mc->mc_flags & C_UNTRACK)
2082 mc = NULL; /* will find mc in mt_cursors */
2083 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
2084 for (; mc; mc=mc->mc_next) {
2085 if (!(mc->mc_flags & C_INITIALIZED))
2087 for (m3 = mc;; m3 = &mx->mx_cursor) {
2089 for (j=0; j<m3->mc_snum; j++) {
2091 if ((mp->mp_flags & Mask) == pflags)
2092 mp->mp_flags ^= P_KEEP;
2094 mx = m3->mc_xcursor;
2095 /* Proceed to mx if it is at a sub-database */
2096 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
2098 if (! (mp && (mp->mp_flags & P_LEAF)))
2100 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
2101 if (!(leaf->mn_flags & F_SUBDATA))
2110 /* Mark dirty root pages */
2111 for (i=0; i<txn->mt_numdbs; i++) {
2112 if (txn->mt_dbflags[i] & DB_DIRTY) {
2113 pgno_t pgno = txn->mt_dbs[i].md_root;
2114 if (pgno == P_INVALID)
2116 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
2118 if ((dp->mp_flags & Mask) == pflags && level <= 1)
2119 dp->mp_flags ^= P_KEEP;
2127 static int mdb_page_flush(MDB_txn *txn, int keep);
2129 /** Spill pages from the dirty list back to disk.
2130 * This is intended to prevent running into #MDB_TXN_FULL situations,
2131 * but note that they may still occur in a few cases:
2132 * 1) our estimate of the txn size could be too small. Currently this
2133 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
2134 * 2) child txns may run out of space if their parents dirtied a
2135 * lot of pages and never spilled them. TODO: we probably should do
2136 * a preemptive spill during #mdb_txn_begin() of a child txn, if
2137 * the parent's dirty_room is below a given threshold.
2139 * Otherwise, if not using nested txns, it is expected that apps will
2140 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
2141 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
2142 * If the txn never references them again, they can be left alone.
2143 * If the txn only reads them, they can be used without any fuss.
2144 * If the txn writes them again, they can be dirtied immediately without
2145 * going thru all of the work of #mdb_page_touch(). Such references are
2146 * handled by #mdb_page_unspill().
2148 * Also note, we never spill DB root pages, nor pages of active cursors,
2149 * because we'll need these back again soon anyway. And in nested txns,
2150 * we can't spill a page in a child txn if it was already spilled in a
2151 * parent txn. That would alter the parent txns' data even though
2152 * the child hasn't committed yet, and we'd have no way to undo it if
2153 * the child aborted.
2155 * @param[in] m0 cursor A cursor handle identifying the transaction and
2156 * database for which we are checking space.
2157 * @param[in] key For a put operation, the key being stored.
2158 * @param[in] data For a put operation, the data being stored.
2159 * @return 0 on success, non-zero on failure.
2162 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
2164 MDB_txn *txn = m0->mc_txn;
2166 MDB_ID2L dl = txn->mt_u.dirty_list;
2167 unsigned int i, j, need;
2170 if (m0->mc_flags & C_SUB)
2173 /* Estimate how much space this op will take */
2174 i = m0->mc_db->md_depth;
2175 /* Named DBs also dirty the main DB */
2176 if (m0->mc_dbi >= CORE_DBS)
2177 i += txn->mt_dbs[MAIN_DBI].md_depth;
2178 /* For puts, roughly factor in the key+data size */
2180 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2181 i += i; /* double it for good measure */
2184 if (txn->mt_dirty_room > i)
2187 if (!txn->mt_spill_pgs) {
2188 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2189 if (!txn->mt_spill_pgs)
2192 /* purge deleted slots */
2193 MDB_IDL sl = txn->mt_spill_pgs;
2194 unsigned int num = sl[0];
2196 for (i=1; i<=num; i++) {
2203 /* Preserve pages which may soon be dirtied again */
2204 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2207 /* Less aggressive spill - we originally spilled the entire dirty list,
2208 * with a few exceptions for cursor pages and DB root pages. But this
2209 * turns out to be a lot of wasted effort because in a large txn many
2210 * of those pages will need to be used again. So now we spill only 1/8th
2211 * of the dirty pages. Testing revealed this to be a good tradeoff,
2212 * better than 1/2, 1/4, or 1/10.
2214 if (need < MDB_IDL_UM_MAX / 8)
2215 need = MDB_IDL_UM_MAX / 8;
2217 /* Save the page IDs of all the pages we're flushing */
2218 /* flush from the tail forward, this saves a lot of shifting later on. */
2219 for (i=dl[0].mid; i && need; i--) {
2220 MDB_ID pn = dl[i].mid << 1;
2222 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2224 /* Can't spill twice, make sure it's not already in a parent's
2227 if (txn->mt_parent) {
2229 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2230 if (tx2->mt_spill_pgs) {
2231 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2232 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2233 dp->mp_flags |= P_KEEP;
2241 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2245 mdb_midl_sort(txn->mt_spill_pgs);
2247 /* Flush the spilled part of dirty list */
2248 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2251 /* Reset any dirty pages we kept that page_flush didn't see */
2252 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2255 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2259 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2261 mdb_find_oldest(MDB_txn *txn)
2264 txnid_t mr, oldest = txn->mt_txnid - 1;
2265 if (txn->mt_env->me_txns) {
2266 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2267 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2278 /** Add a page to the txn's dirty list */
2280 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2283 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2285 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2286 insert = mdb_mid2l_append;
2288 insert = mdb_mid2l_insert;
2290 mid.mid = mp->mp_pgno;
2292 rc = insert(txn->mt_u.dirty_list, &mid);
2293 mdb_tassert(txn, rc == 0);
2294 txn->mt_dirty_room--;
2297 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2298 * me_pghead and mt_next_pgno.
2300 * If there are free pages available from older transactions, they
2301 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2302 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2303 * and move me_pglast to say which records were consumed. Only this
2304 * function can create me_pghead and move me_pglast/mt_next_pgno.
2305 * When #MDB_DEVEL & 2, it is not affected by #mdb_freelist_save(): it
2306 * then uses the transaction's original snapshot of the freeDB.
2307 * @param[in] mc cursor A cursor handle identifying the transaction and
2308 * database for which we are allocating.
2309 * @param[in] num the number of pages to allocate.
2310 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2311 * will always be satisfied by a single contiguous chunk of memory.
2312 * @return 0 on success, non-zero on failure.
2315 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2317 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2318 /* Get at most <Max_retries> more freeDB records once me_pghead
2319 * has enough pages. If not enough, use new pages from the map.
2320 * If <Paranoid> and mc is updating the freeDB, only get new
2321 * records if me_pghead is empty. Then the freelist cannot play
2322 * catch-up with itself by growing while trying to save it.
2324 enum { Paranoid = 1, Max_retries = 500 };
2326 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2328 int rc, retry = num * 60;
2329 MDB_txn *txn = mc->mc_txn;
2330 MDB_env *env = txn->mt_env;
2331 pgno_t pgno, *mop = env->me_pghead;
2332 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2334 txnid_t oldest = 0, last;
2339 /* If there are any loose pages, just use them */
2340 if (num == 1 && txn->mt_loose_pgs) {
2341 np = txn->mt_loose_pgs;
2342 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2343 txn->mt_loose_count--;
2344 DPRINTF(("db %d use loose page %"Yu, DDBI(mc), np->mp_pgno));
2351 /* If our dirty list is already full, we can't do anything */
2352 if (txn->mt_dirty_room == 0) {
2357 for (op = MDB_FIRST;; op = MDB_NEXT) {
2362 /* Seek a big enough contiguous page range. Prefer
2363 * pages at the tail, just truncating the list.
2369 if (mop[i-n2] == pgno+n2)
2376 if (op == MDB_FIRST) { /* 1st iteration */
2377 /* Prepare to fetch more and coalesce */
2378 last = env->me_pglast;
2379 oldest = env->me_pgoldest;
2380 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2381 #if (MDB_DEVEL) & 2 /* "& 2" so MDB_DEVEL=1 won't hide bugs breaking freeDB */
2382 /* Use original snapshot. TODO: Should need less care in code
2383 * which modifies the database. Maybe we can delete some code?
2385 m2.mc_flags |= C_ORIG_RDONLY;
2386 m2.mc_db = &env->me_metas[(txn->mt_txnid-1) & 1]->mm_dbs[FREE_DBI];
2387 m2.mc_dbflag = (unsigned char *)""; /* probably unnecessary */
2391 key.mv_data = &last; /* will look up last+1 */
2392 key.mv_size = sizeof(last);
2394 if (Paranoid && mc->mc_dbi == FREE_DBI)
2397 if (Paranoid && retry < 0 && mop_len)
2401 /* Do not fetch more if the record will be too recent */
2402 if (oldest <= last) {
2404 oldest = mdb_find_oldest(txn);
2405 env->me_pgoldest = oldest;
2411 rc = mdb_cursor_get(&m2, &key, NULL, op);
2413 if (rc == MDB_NOTFOUND)
2417 last = *(txnid_t*)key.mv_data;
2418 if (oldest <= last) {
2420 oldest = mdb_find_oldest(txn);
2421 env->me_pgoldest = oldest;
2427 np = m2.mc_pg[m2.mc_top];
2428 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2429 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2432 idl = (MDB_ID *) data.mv_data;
2435 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2440 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2442 mop = env->me_pghead;
2444 env->me_pglast = last;
2446 DPRINTF(("IDL read txn %"Yu" root %"Yu" num %u",
2447 last, txn->mt_dbs[FREE_DBI].md_root, i));
2449 DPRINTF(("IDL %"Yu, idl[j]));
2451 /* Merge in descending sorted order */
2452 mdb_midl_xmerge(mop, idl);
2456 /* Use new pages from the map when nothing suitable in the freeDB */
2458 pgno = txn->mt_next_pgno;
2459 if (pgno + num >= env->me_maxpg) {
2460 DPUTS("DB size maxed out");
2464 #if defined(_WIN32) && !defined(MDB_VL32)
2465 if (!(env->me_flags & MDB_RDONLY)) {
2467 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
2468 p = VirtualAlloc(p, env->me_psize * num, MEM_COMMIT,
2469 (env->me_flags & MDB_WRITEMAP) ? PAGE_READWRITE:
2472 DPUTS("VirtualAlloc failed");
2480 if (env->me_flags & MDB_WRITEMAP) {
2481 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2483 if (!(np = mdb_page_malloc(txn, num))) {
2489 mop[0] = mop_len -= num;
2490 /* Move any stragglers down */
2491 for (j = i-num; j < mop_len; )
2492 mop[++j] = mop[++i];
2494 txn->mt_next_pgno = pgno + num;
2497 mdb_page_dirty(txn, np);
2503 txn->mt_flags |= MDB_TXN_ERROR;
2507 /** Copy the used portions of a non-overflow page.
2508 * @param[in] dst page to copy into
2509 * @param[in] src page to copy from
2510 * @param[in] psize size of a page
2513 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2515 enum { Align = sizeof(pgno_t) };
2516 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2518 /* If page isn't full, just copy the used portion. Adjust
2519 * alignment so memcpy may copy words instead of bytes.
2521 if ((unused &= -Align) && !IS_LEAF2(src)) {
2522 upper = (upper + PAGEBASE) & -Align;
2523 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2524 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2527 memcpy(dst, src, psize - unused);
2531 /** Pull a page off the txn's spill list, if present.
2532 * If a page being referenced was spilled to disk in this txn, bring
2533 * it back and make it dirty/writable again.
2534 * @param[in] txn the transaction handle.
2535 * @param[in] mp the page being referenced. It must not be dirty.
2536 * @param[out] ret the writable page, if any. ret is unchanged if
2537 * mp wasn't spilled.
2540 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2542 MDB_env *env = txn->mt_env;
2545 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2547 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2548 if (!tx2->mt_spill_pgs)
2550 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2551 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2554 if (txn->mt_dirty_room == 0)
2555 return MDB_TXN_FULL;
2556 if (IS_OVERFLOW(mp))
2560 if (env->me_flags & MDB_WRITEMAP) {
2563 np = mdb_page_malloc(txn, num);
2567 memcpy(np, mp, num * env->me_psize);
2569 mdb_page_copy(np, mp, env->me_psize);
2572 /* If in current txn, this page is no longer spilled.
2573 * If it happens to be the last page, truncate the spill list.
2574 * Otherwise mark it as deleted by setting the LSB.
2576 if (x == txn->mt_spill_pgs[0])
2577 txn->mt_spill_pgs[0]--;
2579 txn->mt_spill_pgs[x] |= 1;
2580 } /* otherwise, if belonging to a parent txn, the
2581 * page remains spilled until child commits
2584 mdb_page_dirty(txn, np);
2585 np->mp_flags |= P_DIRTY;
2593 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2594 * @param[in] mc cursor pointing to the page to be touched
2595 * @return 0 on success, non-zero on failure.
2598 mdb_page_touch(MDB_cursor *mc)
2600 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2601 MDB_txn *txn = mc->mc_txn;
2602 MDB_cursor *m2, *m3;
2606 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2607 if (txn->mt_flags & MDB_TXN_SPILLS) {
2609 rc = mdb_page_unspill(txn, mp, &np);
2615 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2616 (rc = mdb_page_alloc(mc, 1, &np)))
2619 DPRINTF(("touched db %d page %"Yu" -> %"Yu, DDBI(mc),
2620 mp->mp_pgno, pgno));
2621 mdb_cassert(mc, mp->mp_pgno != pgno);
2622 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2623 /* Update the parent page, if any, to point to the new page */
2625 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2626 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2627 SETPGNO(node, pgno);
2629 mc->mc_db->md_root = pgno;
2631 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2632 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2634 /* If txn has a parent, make sure the page is in our
2638 unsigned x = mdb_mid2l_search(dl, pgno);
2639 if (x <= dl[0].mid && dl[x].mid == pgno) {
2640 if (mp != dl[x].mptr) { /* bad cursor? */
2641 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2642 txn->mt_flags |= MDB_TXN_ERROR;
2648 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2650 np = mdb_page_malloc(txn, 1);
2655 rc = mdb_mid2l_insert(dl, &mid);
2656 mdb_cassert(mc, rc == 0);
2661 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2663 np->mp_flags |= P_DIRTY;
2666 /* Adjust cursors pointing to mp */
2667 mc->mc_pg[mc->mc_top] = np;
2668 m2 = txn->mt_cursors[mc->mc_dbi];
2669 if (mc->mc_flags & C_SUB) {
2670 for (; m2; m2=m2->mc_next) {
2671 m3 = &m2->mc_xcursor->mx_cursor;
2672 if (m3->mc_snum < mc->mc_snum) continue;
2673 if (m3->mc_pg[mc->mc_top] == mp)
2674 m3->mc_pg[mc->mc_top] = np;
2677 for (; m2; m2=m2->mc_next) {
2678 if (m2->mc_snum < mc->mc_snum) continue;
2679 if (m2 == mc) continue;
2680 if (m2->mc_pg[mc->mc_top] == mp) {
2681 m2->mc_pg[mc->mc_top] = np;
2682 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2684 (m2->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
2686 MDB_node *leaf = NODEPTR(np, m2->mc_ki[mc->mc_top]);
2687 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
2688 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2693 MDB_PAGE_UNREF(mc->mc_txn, mp);
2697 txn->mt_flags |= MDB_TXN_ERROR;
2702 mdb_env_sync0(MDB_env *env, int force, pgno_t numpgs)
2705 if (env->me_flags & MDB_RDONLY)
2707 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2708 if (env->me_flags & MDB_WRITEMAP) {
2709 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2710 ? MS_ASYNC : MS_SYNC;
2711 if (MDB_MSYNC(env->me_map, env->me_psize * numpgs, flags))
2714 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2718 #ifdef BROKEN_FDATASYNC
2719 if (env->me_flags & MDB_FSYNCONLY) {
2720 if (fsync(env->me_fd))
2724 if (MDB_FDATASYNC(env->me_fd))
2732 mdb_env_sync(MDB_env *env, int force)
2734 MDB_meta *m = mdb_env_pick_meta(env);
2735 return mdb_env_sync0(env, force, m->mm_last_pg+1);
2738 /** Back up parent txn's cursors, then grab the originals for tracking */
2740 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2742 MDB_cursor *mc, *bk;
2747 for (i = src->mt_numdbs; --i >= 0; ) {
2748 if ((mc = src->mt_cursors[i]) != NULL) {
2749 size = sizeof(MDB_cursor);
2751 size += sizeof(MDB_xcursor);
2752 for (; mc; mc = bk->mc_next) {
2758 mc->mc_db = &dst->mt_dbs[i];
2759 /* Kill pointers into src to reduce abuse: The
2760 * user may not use mc until dst ends. But we need a valid
2761 * txn pointer here for cursor fixups to keep working.
2764 mc->mc_dbflag = &dst->mt_dbflags[i];
2765 if ((mx = mc->mc_xcursor) != NULL) {
2766 *(MDB_xcursor *)(bk+1) = *mx;
2767 mx->mx_cursor.mc_txn = dst;
2769 mc->mc_next = dst->mt_cursors[i];
2770 dst->mt_cursors[i] = mc;
2777 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2778 * @param[in] txn the transaction handle.
2779 * @param[in] merge true to keep changes to parent cursors, false to revert.
2780 * @return 0 on success, non-zero on failure.
2783 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2785 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2789 for (i = txn->mt_numdbs; --i >= 0; ) {
2790 for (mc = cursors[i]; mc; mc = next) {
2792 if ((bk = mc->mc_backup) != NULL) {
2794 /* Commit changes to parent txn */
2795 mc->mc_next = bk->mc_next;
2796 mc->mc_backup = bk->mc_backup;
2797 mc->mc_txn = bk->mc_txn;
2798 mc->mc_db = bk->mc_db;
2799 mc->mc_dbflag = bk->mc_dbflag;
2800 if ((mx = mc->mc_xcursor) != NULL)
2801 mx->mx_cursor.mc_txn = bk->mc_txn;
2803 /* Abort nested txn */
2805 if ((mx = mc->mc_xcursor) != NULL)
2806 *mx = *(MDB_xcursor *)(bk+1);
2810 /* Only malloced cursors are permanently tracked. */
2817 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2823 Pidset = F_SETLK, Pidcheck = F_GETLK
2827 /** Set or check a pid lock. Set returns 0 on success.
2828 * Check returns 0 if the process is certainly dead, nonzero if it may
2829 * be alive (the lock exists or an error happened so we do not know).
2831 * On Windows Pidset is a no-op, we merely check for the existence
2832 * of the process with the given pid. On POSIX we use a single byte
2833 * lock on the lockfile, set at an offset equal to the pid.
2836 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2838 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2841 if (op == Pidcheck) {
2842 h = OpenProcess(env->me_pidquery, FALSE, pid);
2843 /* No documented "no such process" code, but other program use this: */
2845 return ErrCode() != ERROR_INVALID_PARAMETER;
2846 /* A process exists until all handles to it close. Has it exited? */
2847 ret = WaitForSingleObject(h, 0) != 0;
2854 struct flock lock_info;
2855 memset(&lock_info, 0, sizeof(lock_info));
2856 lock_info.l_type = F_WRLCK;
2857 lock_info.l_whence = SEEK_SET;
2858 lock_info.l_start = pid;
2859 lock_info.l_len = 1;
2860 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2861 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2863 } else if ((rc = ErrCode()) == EINTR) {
2871 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2872 * @param[in] txn the transaction handle to initialize
2873 * @return 0 on success, non-zero on failure.
2876 mdb_txn_renew0(MDB_txn *txn)
2878 MDB_env *env = txn->mt_env;
2879 MDB_txninfo *ti = env->me_txns;
2881 unsigned int i, nr, flags = txn->mt_flags;
2883 int rc, new_notls = 0;
2885 if ((flags &= MDB_TXN_RDONLY) != 0) {
2887 meta = mdb_env_pick_meta(env);
2888 txn->mt_txnid = meta->mm_txnid;
2889 txn->mt_u.reader = NULL;
2891 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2892 pthread_getspecific(env->me_txkey);
2894 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2895 return MDB_BAD_RSLOT;
2897 MDB_PID_T pid = env->me_pid;
2898 MDB_THR_T tid = pthread_self();
2899 mdb_mutexref_t rmutex = env->me_rmutex;
2901 if (!env->me_live_reader) {
2902 rc = mdb_reader_pid(env, Pidset, pid);
2905 env->me_live_reader = 1;
2908 if (LOCK_MUTEX(rc, env, rmutex))
2910 nr = ti->mti_numreaders;
2911 for (i=0; i<nr; i++)
2912 if (ti->mti_readers[i].mr_pid == 0)
2914 if (i == env->me_maxreaders) {
2915 UNLOCK_MUTEX(rmutex);
2916 return MDB_READERS_FULL;
2918 r = &ti->mti_readers[i];
2919 /* Claim the reader slot, carefully since other code
2920 * uses the reader table un-mutexed: First reset the
2921 * slot, next publish it in mti_numreaders. After
2922 * that, it is safe for mdb_env_close() to touch it.
2923 * When it will be closed, we can finally claim it.
2926 r->mr_txnid = (txnid_t)-1;
2929 ti->mti_numreaders = ++nr;
2930 env->me_close_readers = nr;
2932 UNLOCK_MUTEX(rmutex);
2934 new_notls = (env->me_flags & MDB_NOTLS);
2935 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2940 do /* LY: Retry on a race, ITS#7970. */
2941 r->mr_txnid = ti->mti_txnid;
2942 while(r->mr_txnid != ti->mti_txnid);
2943 txn->mt_txnid = r->mr_txnid;
2944 txn->mt_u.reader = r;
2945 meta = env->me_metas[txn->mt_txnid & 1];
2949 /* Not yet touching txn == env->me_txn0, it may be active */
2951 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2953 txn->mt_txnid = ti->mti_txnid;
2954 meta = env->me_metas[txn->mt_txnid & 1];
2956 meta = mdb_env_pick_meta(env);
2957 txn->mt_txnid = meta->mm_txnid;
2961 if (txn->mt_txnid == mdb_debug_start)
2964 txn->mt_child = NULL;
2965 txn->mt_loose_pgs = NULL;
2966 txn->mt_loose_count = 0;
2967 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2968 txn->mt_u.dirty_list = env->me_dirty_list;
2969 txn->mt_u.dirty_list[0].mid = 0;
2970 txn->mt_free_pgs = env->me_free_pgs;
2971 txn->mt_free_pgs[0] = 0;
2972 txn->mt_spill_pgs = NULL;
2974 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2977 /* Copy the DB info and flags */
2978 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2980 /* Moved to here to avoid a data race in read TXNs */
2981 txn->mt_next_pgno = meta->mm_last_pg+1;
2983 txn->mt_last_pgno = txn->mt_next_pgno - 1;
2986 txn->mt_flags = flags;
2989 txn->mt_numdbs = env->me_numdbs;
2990 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2991 x = env->me_dbflags[i];
2992 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2993 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2995 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2996 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2998 if (env->me_flags & MDB_FATAL_ERROR) {
2999 DPUTS("environment had fatal error, must shutdown!");
3001 } else if (env->me_maxpg < txn->mt_next_pgno) {
3002 rc = MDB_MAP_RESIZED;
3006 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
3011 mdb_txn_renew(MDB_txn *txn)
3015 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
3018 rc = mdb_txn_renew0(txn);
3019 if (rc == MDB_SUCCESS) {
3020 DPRINTF(("renew txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3021 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3022 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
3028 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
3032 int rc, size, tsize;
3034 flags &= MDB_TXN_BEGIN_FLAGS;
3035 flags |= env->me_flags & MDB_WRITEMAP;
3037 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
3041 /* Nested transactions: Max 1 child, write txns only, no writemap */
3042 flags |= parent->mt_flags;
3043 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
3044 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
3046 /* Child txns save MDB_pgstate and use own copy of cursors */
3047 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
3048 size += tsize = sizeof(MDB_ntxn);
3049 } else if (flags & MDB_RDONLY) {
3050 size = env->me_maxdbs * (sizeof(MDB_db)+1);
3051 size += tsize = sizeof(MDB_txn);
3053 /* Reuse preallocated write txn. However, do not touch it until
3054 * mdb_txn_renew0() succeeds, since it currently may be active.
3059 if ((txn = calloc(1, size)) == NULL) {
3060 DPRINTF(("calloc: %s", strerror(errno)));
3065 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
3066 if (!txn->mt_rpages) {
3070 txn->mt_rpages[0].mid = 0;
3071 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
3074 txn->mt_dbxs = env->me_dbxs; /* static */
3075 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
3076 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
3077 txn->mt_flags = flags;
3082 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
3083 txn->mt_dbiseqs = parent->mt_dbiseqs;
3084 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
3085 if (!txn->mt_u.dirty_list ||
3086 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
3088 free(txn->mt_u.dirty_list);
3092 txn->mt_txnid = parent->mt_txnid;
3093 txn->mt_dirty_room = parent->mt_dirty_room;
3094 txn->mt_u.dirty_list[0].mid = 0;
3095 txn->mt_spill_pgs = NULL;
3096 txn->mt_next_pgno = parent->mt_next_pgno;
3097 parent->mt_flags |= MDB_TXN_HAS_CHILD;
3098 parent->mt_child = txn;
3099 txn->mt_parent = parent;
3100 txn->mt_numdbs = parent->mt_numdbs;
3102 txn->mt_rpages = parent->mt_rpages;
3104 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3105 /* Copy parent's mt_dbflags, but clear DB_NEW */
3106 for (i=0; i<txn->mt_numdbs; i++)
3107 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
3109 ntxn = (MDB_ntxn *)txn;
3110 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
3111 if (env->me_pghead) {
3112 size = MDB_IDL_SIZEOF(env->me_pghead);
3113 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
3115 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
3120 rc = mdb_cursor_shadow(parent, txn);
3122 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
3123 } else { /* MDB_RDONLY */
3124 txn->mt_dbiseqs = env->me_dbiseqs;
3126 rc = mdb_txn_renew0(txn);
3129 if (txn != env->me_txn0) {
3131 free(txn->mt_rpages);
3136 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
3138 DPRINTF(("begin txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3139 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
3140 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
3147 mdb_txn_env(MDB_txn *txn)
3149 if(!txn) return NULL;
3154 mdb_txn_id(MDB_txn *txn)
3157 return txn->mt_txnid;
3160 /** Export or close DBI handles opened in this txn. */
3162 mdb_dbis_update(MDB_txn *txn, int keep)
3165 MDB_dbi n = txn->mt_numdbs;
3166 MDB_env *env = txn->mt_env;
3167 unsigned char *tdbflags = txn->mt_dbflags;
3169 for (i = n; --i >= CORE_DBS;) {
3170 if (tdbflags[i] & DB_NEW) {
3172 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
3174 char *ptr = env->me_dbxs[i].md_name.mv_data;
3176 env->me_dbxs[i].md_name.mv_data = NULL;
3177 env->me_dbxs[i].md_name.mv_size = 0;
3178 env->me_dbflags[i] = 0;
3179 env->me_dbiseqs[i]++;
3185 if (keep && env->me_numdbs < n)
3189 /** End a transaction, except successful commit of a nested transaction.
3190 * May be called twice for readonly txns: First reset it, then abort.
3191 * @param[in] txn the transaction handle to end
3192 * @param[in] mode why and how to end the transaction
3195 mdb_txn_end(MDB_txn *txn, unsigned mode)
3197 MDB_env *env = txn->mt_env;
3199 static const char *const names[] = MDB_END_NAMES;
3202 /* Export or close DBI handles opened in this txn */
3203 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
3205 DPRINTF(("%s txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3206 names[mode & MDB_END_OPMASK],
3207 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3208 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
3210 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3211 if (txn->mt_u.reader) {
3212 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
3213 if (!(env->me_flags & MDB_NOTLS)) {
3214 txn->mt_u.reader = NULL; /* txn does not own reader */
3215 } else if (mode & MDB_END_SLOT) {
3216 txn->mt_u.reader->mr_pid = 0;
3217 txn->mt_u.reader = NULL;
3218 } /* else txn owns the slot until it does MDB_END_SLOT */
3220 txn->mt_numdbs = 0; /* prevent further DBI activity */
3221 txn->mt_flags |= MDB_TXN_FINISHED;
3223 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
3224 pgno_t *pghead = env->me_pghead;
3226 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
3227 mdb_cursors_close(txn, 0);
3228 if (!(env->me_flags & MDB_WRITEMAP)) {
3229 mdb_dlist_free(txn);
3233 txn->mt_flags = MDB_TXN_FINISHED;
3235 if (!txn->mt_parent) {
3236 mdb_midl_shrink(&txn->mt_free_pgs);
3237 env->me_free_pgs = txn->mt_free_pgs;
3239 env->me_pghead = NULL;
3243 mode = 0; /* txn == env->me_txn0, do not free() it */
3245 /* The writer mutex was locked in mdb_txn_begin. */
3247 UNLOCK_MUTEX(env->me_wmutex);
3249 txn->mt_parent->mt_child = NULL;
3250 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3251 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3252 mdb_midl_free(txn->mt_free_pgs);
3253 mdb_midl_free(txn->mt_spill_pgs);
3254 free(txn->mt_u.dirty_list);
3257 mdb_midl_free(pghead);
3260 if (!txn->mt_parent) {
3261 MDB_ID3L el = env->me_rpages, tl = txn->mt_rpages;
3262 unsigned i, x, n = tl[0].mid;
3263 pthread_mutex_lock(&env->me_rpmutex);
3264 for (i = 1; i <= n; i++) {
3265 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
3266 /* tmp overflow pages that we didn't share in env */
3267 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3269 x = mdb_mid3l_search(el, tl[i].mid);
3270 if (tl[i].mptr == el[x].mptr) {
3273 /* another tmp overflow page */
3274 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3278 pthread_mutex_unlock(&env->me_rpmutex);
3280 if (mode & MDB_END_FREE)
3284 if (mode & MDB_END_FREE)
3289 mdb_txn_reset(MDB_txn *txn)
3294 /* This call is only valid for read-only txns */
3295 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3298 mdb_txn_end(txn, MDB_END_RESET);
3302 mdb_txn_abort(MDB_txn *txn)
3308 mdb_txn_abort(txn->mt_child);
3310 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3313 /** Save the freelist as of this transaction to the freeDB.
3314 * This changes the freelist. Keep trying until it stabilizes.
3316 * When (MDB_DEVEL) & 2, the changes do not affect #mdb_page_alloc(),
3317 * it then uses the transaction's original snapshot of the freeDB.
3320 mdb_freelist_save(MDB_txn *txn)
3322 /* env->me_pghead[] can grow and shrink during this call.
3323 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3324 * Page numbers cannot disappear from txn->mt_free_pgs[].
3327 MDB_env *env = txn->mt_env;
3328 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3329 txnid_t pglast = 0, head_id = 0;
3330 pgno_t freecnt = 0, *free_pgs, *mop;
3331 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3333 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3335 if (env->me_pghead) {
3336 /* Make sure first page of freeDB is touched and on freelist */
3337 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3338 if (rc && rc != MDB_NOTFOUND)
3342 if (!env->me_pghead && txn->mt_loose_pgs) {
3343 /* Put loose page numbers in mt_free_pgs, since
3344 * we may be unable to return them to me_pghead.
3346 MDB_page *mp = txn->mt_loose_pgs;
3347 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3349 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3350 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3351 txn->mt_loose_pgs = NULL;
3352 txn->mt_loose_count = 0;
3355 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3356 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3357 ? SSIZE_MAX : maxfree_1pg;
3360 /* Come back here after each Put() in case freelist changed */
3365 /* If using records from freeDB which we have not yet
3366 * deleted, delete them and any we reserved for me_pghead.
3368 while (pglast < env->me_pglast) {
3369 rc = mdb_cursor_first(&mc, &key, NULL);
3372 pglast = head_id = *(txnid_t *)key.mv_data;
3373 total_room = head_room = 0;
3374 mdb_tassert(txn, pglast <= env->me_pglast);
3375 rc = mdb_cursor_del(&mc, 0);
3380 /* Save the IDL of pages freed by this txn, to a single record */
3381 if (freecnt < txn->mt_free_pgs[0]) {
3383 /* Make sure last page of freeDB is touched and on freelist */
3384 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3385 if (rc && rc != MDB_NOTFOUND)
3388 free_pgs = txn->mt_free_pgs;
3389 /* Write to last page of freeDB */
3390 key.mv_size = sizeof(txn->mt_txnid);
3391 key.mv_data = &txn->mt_txnid;
3393 freecnt = free_pgs[0];
3394 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3395 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3398 /* Retry if mt_free_pgs[] grew during the Put() */
3399 free_pgs = txn->mt_free_pgs;
3400 } while (freecnt < free_pgs[0]);
3401 mdb_midl_sort(free_pgs);
3402 memcpy(data.mv_data, free_pgs, data.mv_size);
3405 unsigned int i = free_pgs[0];
3406 DPRINTF(("IDL write txn %"Yu" root %"Yu" num %u",
3407 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3409 DPRINTF(("IDL %"Yu, free_pgs[i]));
3415 mop = env->me_pghead;
3416 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3418 /* Reserve records for me_pghead[]. Split it if multi-page,
3419 * to avoid searching freeDB for a page range. Use keys in
3420 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3422 if (total_room >= mop_len) {
3423 if (total_room == mop_len || --more < 0)
3425 } else if (head_room >= maxfree_1pg && head_id > 1) {
3426 /* Keep current record (overflow page), add a new one */
3430 /* (Re)write {key = head_id, IDL length = head_room} */
3431 total_room -= head_room;
3432 head_room = mop_len - total_room;
3433 if (head_room > maxfree_1pg && head_id > 1) {
3434 /* Overflow multi-page for part of me_pghead */
3435 head_room /= head_id; /* amortize page sizes */
3436 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3437 } else if (head_room < 0) {
3438 /* Rare case, not bothering to delete this record */
3441 key.mv_size = sizeof(head_id);
3442 key.mv_data = &head_id;
3443 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3444 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3447 /* IDL is initially empty, zero out at least the length */
3448 pgs = (pgno_t *)data.mv_data;
3449 j = head_room > clean_limit ? head_room : 0;
3453 total_room += head_room;
3456 /* Return loose page numbers to me_pghead, though usually none are
3457 * left at this point. The pages themselves remain in dirty_list.
3459 if (txn->mt_loose_pgs) {
3460 MDB_page *mp = txn->mt_loose_pgs;
3461 unsigned count = txn->mt_loose_count;
3463 /* Room for loose pages + temp IDL with same */
3464 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3466 mop = env->me_pghead;
3467 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3468 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3469 loose[ ++count ] = mp->mp_pgno;
3471 mdb_midl_sort(loose);
3472 mdb_midl_xmerge(mop, loose);
3473 txn->mt_loose_pgs = NULL;
3474 txn->mt_loose_count = 0;
3478 /* Fill in the reserved me_pghead records */
3484 rc = mdb_cursor_first(&mc, &key, &data);
3485 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3486 txnid_t id = *(txnid_t *)key.mv_data;
3487 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3490 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3492 if (len > mop_len) {
3494 data.mv_size = (len + 1) * sizeof(MDB_ID);
3496 data.mv_data = mop -= len;
3499 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3501 if (rc || !(mop_len -= len))
3508 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3509 * @param[in] txn the transaction that's being committed
3510 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3511 * @return 0 on success, non-zero on failure.
3514 mdb_page_flush(MDB_txn *txn, int keep)
3516 MDB_env *env = txn->mt_env;
3517 MDB_ID2L dl = txn->mt_u.dirty_list;
3518 unsigned psize = env->me_psize, j;
3519 int i, pagecount = dl[0].mid, rc;
3523 MDB_page *dp = NULL;
3527 struct iovec iov[MDB_COMMIT_PAGES];
3528 ssize_t wsize = 0, wres;
3529 off_t wpos = 0, next_pos = 1; /* impossible pos, so pos != next_pos */
3535 if (env->me_flags & MDB_WRITEMAP) {
3536 /* Clear dirty flags */
3537 while (++i <= pagecount) {
3539 /* Don't flush this page yet */
3540 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3541 dp->mp_flags &= ~P_KEEP;
3545 dp->mp_flags &= ~P_DIRTY;
3550 /* Write the pages */
3552 if (++i <= pagecount) {
3554 /* Don't flush this page yet */
3555 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3556 dp->mp_flags &= ~P_KEEP;
3561 /* clear dirty flag */
3562 dp->mp_flags &= ~P_DIRTY;
3565 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3570 /* Windows actually supports scatter/gather I/O, but only on
3571 * unbuffered file handles. Since we're relying on the OS page
3572 * cache for all our data, that's self-defeating. So we just
3573 * write pages one at a time. We use the ov structure to set
3574 * the write offset, to at least save the overhead of a Seek
3577 DPRINTF(("committing page %"Yu, pgno));
3578 memset(&ov, 0, sizeof(ov));
3579 ov.Offset = pos & 0xffffffff;
3580 ov.OffsetHigh = pos >> 16 >> 16;
3581 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3583 DPRINTF(("WriteFile: %d", rc));
3587 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3588 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3591 /* Write previous page(s) */
3592 #ifdef MDB_USE_PWRITEV
3593 wres = pwritev(env->me_fd, iov, n, wpos);
3596 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3599 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3603 DPRINTF(("lseek: %s", strerror(rc)));
3606 wres = writev(env->me_fd, iov, n);
3609 if (wres != wsize) {
3614 DPRINTF(("Write error: %s", strerror(rc)));
3616 rc = EIO; /* TODO: Use which error code? */
3617 DPUTS("short write, filesystem full?");
3628 DPRINTF(("committing page %"Yu, pgno));
3629 next_pos = pos + size;
3630 iov[n].iov_len = size;
3631 iov[n].iov_base = (char *)dp;
3637 if (pgno > txn->mt_last_pgno)
3638 txn->mt_last_pgno = pgno;
3641 /* MIPS has cache coherency issues, this is a no-op everywhere else
3642 * Note: for any size >= on-chip cache size, entire on-chip cache is
3645 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3647 for (i = keep; ++i <= pagecount; ) {
3649 /* This is a page we skipped above */
3652 dl[j].mid = dp->mp_pgno;
3655 mdb_dpage_free(env, dp);
3660 txn->mt_dirty_room += i - j;
3666 mdb_txn_commit(MDB_txn *txn)
3669 unsigned int i, end_mode;
3675 /* mdb_txn_end() mode for a commit which writes nothing */
3676 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3678 if (txn->mt_child) {
3679 rc = mdb_txn_commit(txn->mt_child);
3686 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3690 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3691 DPUTS("txn has failed/finished, can't commit");
3693 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3698 if (txn->mt_parent) {
3699 MDB_txn *parent = txn->mt_parent;
3703 unsigned x, y, len, ps_len;
3705 /* Append our free list to parent's */
3706 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3709 mdb_midl_free(txn->mt_free_pgs);
3710 /* Failures after this must either undo the changes
3711 * to the parent or set MDB_TXN_ERROR in the parent.
3714 parent->mt_next_pgno = txn->mt_next_pgno;
3715 parent->mt_flags = txn->mt_flags;
3717 /* Merge our cursors into parent's and close them */
3718 mdb_cursors_close(txn, 1);
3720 /* Update parent's DB table. */
3721 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3722 parent->mt_numdbs = txn->mt_numdbs;
3723 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3724 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3725 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3726 /* preserve parent's DB_NEW status */
3727 x = parent->mt_dbflags[i] & DB_NEW;
3728 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3731 dst = parent->mt_u.dirty_list;
3732 src = txn->mt_u.dirty_list;
3733 /* Remove anything in our dirty list from parent's spill list */
3734 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3736 pspill[0] = (pgno_t)-1;
3737 /* Mark our dirty pages as deleted in parent spill list */
3738 for (i=0, len=src[0].mid; ++i <= len; ) {
3739 MDB_ID pn = src[i].mid << 1;
3740 while (pn > pspill[x])
3742 if (pn == pspill[x]) {
3747 /* Squash deleted pagenums if we deleted any */
3748 for (x=y; ++x <= ps_len; )
3749 if (!(pspill[x] & 1))
3750 pspill[++y] = pspill[x];
3754 /* Remove anything in our spill list from parent's dirty list */
3755 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3756 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3757 MDB_ID pn = txn->mt_spill_pgs[i];
3759 continue; /* deleted spillpg */
3761 y = mdb_mid2l_search(dst, pn);
3762 if (y <= dst[0].mid && dst[y].mid == pn) {
3764 while (y < dst[0].mid) {
3773 /* Find len = length of merging our dirty list with parent's */
3775 dst[0].mid = 0; /* simplify loops */
3776 if (parent->mt_parent) {
3777 len = x + src[0].mid;
3778 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3779 for (i = x; y && i; y--) {
3780 pgno_t yp = src[y].mid;
3781 while (yp < dst[i].mid)
3783 if (yp == dst[i].mid) {
3788 } else { /* Simplify the above for single-ancestor case */
3789 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3791 /* Merge our dirty list with parent's */
3793 for (i = len; y; dst[i--] = src[y--]) {
3794 pgno_t yp = src[y].mid;
3795 while (yp < dst[x].mid)
3796 dst[i--] = dst[x--];
3797 if (yp == dst[x].mid)
3798 free(dst[x--].mptr);
3800 mdb_tassert(txn, i == x);
3802 free(txn->mt_u.dirty_list);
3803 parent->mt_dirty_room = txn->mt_dirty_room;
3804 if (txn->mt_spill_pgs) {
3805 if (parent->mt_spill_pgs) {
3806 /* TODO: Prevent failure here, so parent does not fail */
3807 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3809 parent->mt_flags |= MDB_TXN_ERROR;
3810 mdb_midl_free(txn->mt_spill_pgs);
3811 mdb_midl_sort(parent->mt_spill_pgs);
3813 parent->mt_spill_pgs = txn->mt_spill_pgs;
3817 /* Append our loose page list to parent's */
3818 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3820 *lp = txn->mt_loose_pgs;
3821 parent->mt_loose_count += txn->mt_loose_count;
3823 parent->mt_child = NULL;
3824 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3829 if (txn != env->me_txn) {
3830 DPUTS("attempt to commit unknown transaction");
3835 mdb_cursors_close(txn, 0);
3837 if (!txn->mt_u.dirty_list[0].mid &&
3838 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3841 DPRINTF(("committing txn %"Yu" %p on mdbenv %p, root page %"Yu,
3842 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3844 /* Update DB root pointers */
3845 if (txn->mt_numdbs > CORE_DBS) {
3849 data.mv_size = sizeof(MDB_db);
3851 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3852 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3853 if (txn->mt_dbflags[i] & DB_DIRTY) {
3854 if (TXN_DBI_CHANGED(txn, i)) {
3858 data.mv_data = &txn->mt_dbs[i];
3859 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3867 rc = mdb_freelist_save(txn);
3871 mdb_midl_free(env->me_pghead);
3872 env->me_pghead = NULL;
3873 mdb_midl_shrink(&txn->mt_free_pgs);
3879 if ((rc = mdb_page_flush(txn, 0)))
3881 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3882 (rc = mdb_env_sync0(env, 0, txn->mt_next_pgno)))
3884 if ((rc = mdb_env_write_meta(txn)))
3886 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3889 mdb_txn_end(txn, end_mode);
3897 /** Read the environment parameters of a DB environment before
3898 * mapping it into memory.
3899 * @param[in] env the environment handle
3900 * @param[out] meta address of where to store the meta information
3901 * @return 0 on success, non-zero on failure.
3904 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3910 enum { Size = sizeof(pbuf) };
3912 /* We don't know the page size yet, so use a minimum value.
3913 * Read both meta pages so we can use the latest one.
3916 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3920 memset(&ov, 0, sizeof(ov));
3922 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3923 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3926 rc = pread(env->me_fd, &pbuf, Size, off);
3929 if (rc == 0 && off == 0)
3931 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3932 DPRINTF(("read: %s", mdb_strerror(rc)));
3936 p = (MDB_page *)&pbuf;
3938 if (!F_ISSET(p->mp_flags, P_META)) {
3939 DPRINTF(("page %"Yu" not a meta page", p->mp_pgno));
3944 if (m->mm_magic != MDB_MAGIC) {
3945 DPUTS("meta has invalid magic");
3949 if (m->mm_version != MDB_DATA_VERSION) {
3950 DPRINTF(("database is version %u, expected version %u",
3951 m->mm_version, MDB_DATA_VERSION));
3952 return MDB_VERSION_MISMATCH;
3955 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3961 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3963 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3965 meta->mm_magic = MDB_MAGIC;
3966 meta->mm_version = MDB_DATA_VERSION;
3967 meta->mm_mapsize = env->me_mapsize;
3968 meta->mm_psize = env->me_psize;
3969 meta->mm_last_pg = NUM_METAS-1;
3970 meta->mm_flags = env->me_flags & 0xffff;
3971 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3972 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3973 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3976 /** Write the environment parameters of a freshly created DB environment.
3977 * @param[in] env the environment handle
3978 * @param[in] meta the #MDB_meta to write
3979 * @return 0 on success, non-zero on failure.
3982 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3990 memset(&ov, 0, sizeof(ov));
3991 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3993 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3996 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3997 len = pwrite(fd, ptr, size, pos); \
3998 if (len == -1 && ErrCode() == EINTR) continue; \
3999 rc = (len >= 0); break; } while(1)
4002 DPUTS("writing new meta page");
4004 psize = env->me_psize;
4006 p = calloc(NUM_METAS, psize);
4010 p->mp_flags = P_META;
4011 *(MDB_meta *)METADATA(p) = *meta;
4013 q = (MDB_page *)((char *)p + psize);
4015 q->mp_flags = P_META;
4016 *(MDB_meta *)METADATA(q) = *meta;
4018 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
4021 else if ((unsigned) len == psize * NUM_METAS)
4029 /** Update the environment info to commit a transaction.
4030 * @param[in] txn the transaction that's being committed
4031 * @return 0 on success, non-zero on failure.
4034 mdb_env_write_meta(MDB_txn *txn)
4037 MDB_meta meta, metab, *mp;
4041 int rc, len, toggle;
4050 toggle = txn->mt_txnid & 1;
4051 DPRINTF(("writing meta page %d for root page %"Yu,
4052 toggle, txn->mt_dbs[MAIN_DBI].md_root));
4055 flags = txn->mt_flags | env->me_flags;
4056 mp = env->me_metas[toggle];
4057 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
4058 /* Persist any increases of mapsize config */
4059 if (mapsize < env->me_mapsize)
4060 mapsize = env->me_mapsize;
4062 if (flags & MDB_WRITEMAP) {
4063 mp->mm_mapsize = mapsize;
4064 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4065 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4066 mp->mm_last_pg = txn->mt_next_pgno - 1;
4067 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
4068 !(defined(__i386__) || defined(__x86_64__))
4069 /* LY: issue a memory barrier, if not x86. ITS#7969 */
4070 __sync_synchronize();
4072 mp->mm_txnid = txn->mt_txnid;
4073 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
4074 unsigned meta_size = env->me_psize;
4075 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
4076 ptr = (char *)mp - PAGEHDRSZ;
4077 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
4078 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
4082 if (MDB_MSYNC(ptr, meta_size, rc)) {
4089 metab.mm_txnid = mp->mm_txnid;
4090 metab.mm_last_pg = mp->mm_last_pg;
4092 meta.mm_mapsize = mapsize;
4093 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4094 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4095 meta.mm_last_pg = txn->mt_next_pgno - 1;
4096 meta.mm_txnid = txn->mt_txnid;
4098 off = offsetof(MDB_meta, mm_mapsize);
4099 ptr = (char *)&meta + off;
4100 len = sizeof(MDB_meta) - off;
4101 off += (char *)mp - env->me_map;
4103 /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC.
4104 * (me_mfd goes to the same file as me_fd, but writing to it
4105 * also syncs to disk. Avoids a separate fdatasync() call.)
4107 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
4110 memset(&ov, 0, sizeof(ov));
4112 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
4117 rc = pwrite(mfd, ptr, len, off);
4120 rc = rc < 0 ? ErrCode() : EIO;
4125 DPUTS("write failed, disk error?");
4126 /* On a failure, the pagecache still contains the new data.
4127 * Write some old data back, to prevent it from being used.
4128 * Use the non-SYNC fd; we know it will fail anyway.
4130 meta.mm_last_pg = metab.mm_last_pg;
4131 meta.mm_txnid = metab.mm_txnid;
4133 memset(&ov, 0, sizeof(ov));
4135 WriteFile(env->me_fd, ptr, len, NULL, &ov);
4137 r2 = pwrite(env->me_fd, ptr, len, off);
4138 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
4141 env->me_flags |= MDB_FATAL_ERROR;
4144 /* MIPS has cache coherency issues, this is a no-op everywhere else */
4145 CACHEFLUSH(env->me_map + off, len, DCACHE);
4147 /* Memory ordering issues are irrelevant; since the entire writer
4148 * is wrapped by wmutex, all of these changes will become visible
4149 * after the wmutex is unlocked. Since the DB is multi-version,
4150 * readers will get consistent data regardless of how fresh or
4151 * how stale their view of these values is.
4154 env->me_txns->mti_txnid = txn->mt_txnid;
4159 /** Check both meta pages to see which one is newer.
4160 * @param[in] env the environment handle
4161 * @return newest #MDB_meta.
4164 mdb_env_pick_meta(const MDB_env *env)
4166 MDB_meta *const *metas = env->me_metas;
4167 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
4171 mdb_env_create(MDB_env **env)
4175 e = calloc(1, sizeof(MDB_env));
4179 e->me_maxreaders = DEFAULT_READERS;
4180 e->me_maxdbs = e->me_numdbs = CORE_DBS;
4181 e->me_fd = INVALID_HANDLE_VALUE;
4182 e->me_lfd = INVALID_HANDLE_VALUE;
4183 e->me_mfd = INVALID_HANDLE_VALUE;
4184 #ifdef MDB_USE_POSIX_SEM
4185 e->me_rmutex = SEM_FAILED;
4186 e->me_wmutex = SEM_FAILED;
4187 #elif defined MDB_USE_SYSV_SEM
4188 e->me_rmutex->semid = -1;
4189 e->me_wmutex->semid = -1;
4191 e->me_pid = getpid();
4192 GET_PAGESIZE(e->me_os_psize);
4193 VGMEMP_CREATE(e,0,0);
4199 /** @brief Map a result from an NTAPI call to WIN32. */
4201 mdb_nt2win32(NTSTATUS st)
4206 GetOverlappedResult(NULL, &o, &br, FALSE);
4207 return GetLastError();
4212 mdb_env_map(MDB_env *env, void *addr)
4215 unsigned int flags = env->me_flags;
4218 int access = SECTION_MAP_READ;
4222 ULONG pageprot = PAGE_READONLY, secprot, alloctype;
4224 if (flags & MDB_WRITEMAP) {
4225 access |= SECTION_MAP_WRITE;
4226 pageprot = PAGE_READWRITE;
4228 if (flags & MDB_RDONLY) {
4229 secprot = PAGE_READONLY;
4233 secprot = PAGE_READWRITE;
4234 msize = env->me_mapsize;
4235 alloctype = MEM_RESERVE;
4238 rc = NtCreateSection(&mh, access, NULL, NULL, secprot, SEC_RESERVE, env->me_fd);
4240 return mdb_nt2win32(rc);
4243 msize = NUM_METAS * env->me_psize;
4245 rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
4252 return mdb_nt2win32(rc);
4257 env->me_map = mmap(addr, NUM_METAS * env->me_psize, PROT_READ, MAP_SHARED,
4259 if (env->me_map == MAP_FAILED) {
4264 int prot = PROT_READ;
4265 if (flags & MDB_WRITEMAP) {
4267 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
4270 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
4272 if (env->me_map == MAP_FAILED) {
4277 if (flags & MDB_NORDAHEAD) {
4278 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
4280 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
4282 #ifdef POSIX_MADV_RANDOM
4283 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
4284 #endif /* POSIX_MADV_RANDOM */
4285 #endif /* MADV_RANDOM */
4289 /* Can happen because the address argument to mmap() is just a
4290 * hint. mmap() can pick another, e.g. if the range is in use.
4291 * The MAP_FIXED flag would prevent that, but then mmap could
4292 * instead unmap existing pages to make room for the new map.
4294 if (addr && env->me_map != addr)
4295 return EBUSY; /* TODO: Make a new MDB_* error code? */
4298 p = (MDB_page *)env->me_map;
4299 env->me_metas[0] = METADATA(p);
4300 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4306 mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
4308 /* If env is already open, caller is responsible for making
4309 * sure there are no active txns.
4319 meta = mdb_env_pick_meta(env);
4321 size = meta->mm_mapsize;
4323 /* Silently round up to minimum if the size is too small */
4324 mdb_size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4329 /* For MDB_VL32 this bit is a noop since we dynamically remap
4330 * chunks of the DB anyway.
4332 munmap(env->me_map, env->me_mapsize);
4333 env->me_mapsize = size;
4334 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4335 rc = mdb_env_map(env, old);
4338 #endif /* !MDB_VL32 */
4340 env->me_mapsize = size;
4342 env->me_maxpg = env->me_mapsize / env->me_psize;
4347 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4351 env->me_maxdbs = dbs + CORE_DBS;
4356 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4358 if (env->me_map || readers < 1)
4360 env->me_maxreaders = readers;
4365 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4367 if (!env || !readers)
4369 *readers = env->me_maxreaders;
4374 mdb_fsize(HANDLE fd, mdb_size_t *size)
4377 LARGE_INTEGER fsize;
4379 if (!GetFileSizeEx(fd, &fsize))
4382 *size = fsize.QuadPart;
4396 typedef wchar_t mdb_nchar_t;
4397 # define MDB_NAME(str) L##str
4398 # define mdb_name_cpy wcscpy
4400 /** Character type for file names: char on Unix, wchar_t on Windows */
4401 typedef char mdb_nchar_t;
4402 # define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */
4403 # define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */
4406 /** Filename - string of #mdb_nchar_t[] */
4407 typedef struct MDB_name {
4408 int mn_len; /**< Length */
4409 int mn_alloced; /**< True if #mn_val was malloced */
4410 mdb_nchar_t *mn_val; /**< Contents */
4413 /** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */
4414 static const mdb_nchar_t *const mdb_suffixes[2][2] = {
4415 { MDB_NAME("/data.mdb"), MDB_NAME("") },
4416 { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") }
4419 #define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */
4421 /** Set up filename + scratch area for filename suffix, for opening files.
4422 * It should be freed with #mdb_fname_destroy().
4423 * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16.
4425 * @param[in] path Pathname for #mdb_env_open().
4426 * @param[in] envflags Whether a subdir and/or lockfile will be used.
4427 * @param[out] fname Resulting filename, with room for a suffix if necessary.
4430 mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname)
4432 int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK);
4433 fname->mn_alloced = 0;
4435 return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN);
4437 fname->mn_len = strlen(path);
4439 fname->mn_val = (char *) path;
4440 else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) {
4441 fname->mn_alloced = 1;
4442 strcpy(fname->mn_val, path);
4450 /** Destroy \b fname from #mdb_fname_init() */
4451 #define mdb_fname_destroy(fname) \
4452 do { if ((fname).mn_alloced) free((fname).mn_val); } while (0)
4454 #ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */
4455 # define MDB_CLOEXEC O_CLOEXEC
4457 # define MDB_CLOEXEC 0
4460 /** File type, access mode etc. for #mdb_fopen() */
4461 enum mdb_fopen_type {
4463 MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS
4465 /* A comment in mdb_fopen() explains some O_* flag choices. */
4466 MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */
4467 MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */
4468 MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */
4469 MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */
4470 /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits
4471 * distinguish otherwise-equal MDB_O_* constants from each other.
4473 MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY,
4474 MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */
4478 /** Open an LMDB file.
4479 * @param[in] env The LMDB environment.
4480 * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is
4481 * appended if necessary to create the filename, without changing mn_len.
4482 * @param[in] which Determines file type, access mode, etc.
4483 * @param[in] mode The Unix permissions for the file, if we create it.
4484 * @param[out] res Resulting file handle.
4485 * @return 0 on success, non-zero on failure.
4488 mdb_fopen(const MDB_env *env, MDB_name *fname,
4489 enum mdb_fopen_type which, mdb_mode_t mode,
4492 int rc = MDB_SUCCESS;
4495 DWORD acc, share, disp, attrs;
4500 if (fname->mn_alloced) /* modifiable copy */
4501 mdb_name_cpy(fname->mn_val + fname->mn_len,
4502 mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]);
4504 /* The directory must already exist. Usually the file need not.
4505 * MDB_O_META requires the file because we already created it using
4506 * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file.
4508 * With MDB_O_COPY we do not want the OS to cache the writes, since
4509 * the source data is already in the OS cache.
4511 * The lockfile needs FD_CLOEXEC (close file descriptor on exec*())
4512 * to avoid the flock() issues noted under Caveats in lmdb.h.
4513 * Also set it for other filehandles which the user cannot get at
4514 * and close himself, which he may need after fork(). I.e. all but
4515 * me_fd, which programs do use via mdb_env_get_fd().
4519 acc = GENERIC_READ|GENERIC_WRITE;
4520 share = FILE_SHARE_READ|FILE_SHARE_WRITE;
4522 attrs = FILE_ATTRIBUTE_NORMAL;
4524 case MDB_O_RDONLY: /* read-only datafile */
4526 disp = OPEN_EXISTING;
4528 case MDB_O_META: /* for writing metapages */
4529 acc = GENERIC_WRITE;
4530 disp = OPEN_EXISTING;
4531 attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH;
4533 case MDB_O_COPY: /* mdb_env_copy() & co */
4534 acc = GENERIC_WRITE;
4537 attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH;
4539 default: break; /* silence gcc -Wswitch (not all enum values handled) */
4541 fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL);
4543 fd = open(fname->mn_val, which & MDB_O_MASK, mode);
4546 if (fd == INVALID_HANDLE_VALUE)
4550 if (which != MDB_O_RDONLY && which != MDB_O_RDWR) {
4551 /* Set CLOEXEC if we could not pass it to open() */
4552 if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1)
4553 (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
4555 if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) {
4556 /* This may require buffer alignment. There is no portable
4557 * way to ask how much, so we require OS pagesize alignment.
4559 # ifdef F_NOCACHE /* __APPLE__ */
4560 (void) fcntl(fd, F_NOCACHE, 1);
4561 # elif defined O_DIRECT
4562 /* open(...O_DIRECT...) would break on filesystems without
4563 * O_DIRECT support (ITS#7682). Try to set it here instead.
4565 if ((flags = fcntl(fd, F_GETFL)) != -1)
4566 (void) fcntl(fd, F_SETFL, flags | O_DIRECT);
4570 #endif /* !_WIN32 */
4577 #ifdef BROKEN_FDATASYNC
4578 #include <sys/utsname.h>
4579 #include <sys/vfs.h>
4582 /** Further setup required for opening an LMDB environment
4585 mdb_env_open2(MDB_env *env)
4587 unsigned int flags = env->me_flags;
4588 int i, newenv = 0, rc;
4592 /* See if we should use QueryLimited */
4594 if ((rc & 0xff) > 5)
4595 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4597 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4600 #ifdef BROKEN_FDATASYNC
4601 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4602 * https://lkml.org/lkml/2012/9/3/83
4603 * Kernels after 3.6-rc6 are known good.
4604 * https://lkml.org/lkml/2012/9/10/556
4605 * See if the DB is on ext3/ext4, then check for new enough kernel
4606 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4611 fstatfs(env->me_fd, &st);
4612 while (st.f_type == 0xEF53) {
4616 if (uts.release[0] < '3') {
4617 if (!strncmp(uts.release, "2.6.32.", 7)) {
4618 i = atoi(uts.release+7);
4620 break; /* 2.6.32.60 and newer is OK */
4621 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4622 i = atoi(uts.release+7);
4624 break; /* 2.6.34.15 and newer is OK */
4626 } else if (uts.release[0] == '3') {
4627 i = atoi(uts.release+2);
4629 break; /* 3.6 and newer is OK */
4631 i = atoi(uts.release+4);
4633 break; /* 3.5.4 and newer is OK */
4634 } else if (i == 2) {
4635 i = atoi(uts.release+4);
4637 break; /* 3.2.30 and newer is OK */
4639 } else { /* 4.x and newer is OK */
4642 env->me_flags |= MDB_FSYNCONLY;
4648 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4651 DPUTS("new mdbenv");
4653 env->me_psize = env->me_os_psize;
4654 if (env->me_psize > MAX_PAGESIZE)
4655 env->me_psize = MAX_PAGESIZE;
4656 memset(&meta, 0, sizeof(meta));
4657 mdb_env_init_meta0(env, &meta);
4658 meta.mm_mapsize = DEFAULT_MAPSIZE;
4660 env->me_psize = meta.mm_psize;
4663 /* Was a mapsize configured? */
4664 if (!env->me_mapsize) {
4665 env->me_mapsize = meta.mm_mapsize;
4668 /* Make sure mapsize >= committed data size. Even when using
4669 * mm_mapsize, which could be broken in old files (ITS#7789).
4671 mdb_size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4672 if (env->me_mapsize < minsize)
4673 env->me_mapsize = minsize;
4675 meta.mm_mapsize = env->me_mapsize;
4677 if (newenv && !(flags & MDB_FIXEDMAP)) {
4678 /* mdb_env_map() may grow the datafile. Write the metapages
4679 * first, so the file will be valid if initialization fails.
4680 * Except with FIXEDMAP, since we do not yet know mm_address.
4681 * We could fill in mm_address later, but then a different
4682 * program might end up doing that - one with a memory layout
4683 * and map address which does not suit the main program.
4685 rc = mdb_env_init_meta(env, &meta);
4691 /* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
4695 rc = WriteFile(env->me_fd, &dummy, 1, &len, NULL);
4703 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4708 if (flags & MDB_FIXEDMAP)
4709 meta.mm_address = env->me_map;
4710 i = mdb_env_init_meta(env, &meta);
4711 if (i != MDB_SUCCESS) {
4716 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4717 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4719 #if !(MDB_MAXKEYSIZE)
4720 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4722 env->me_maxpg = env->me_mapsize / env->me_psize;
4726 MDB_meta *meta = mdb_env_pick_meta(env);
4727 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4729 DPRINTF(("opened database version %u, pagesize %u",
4730 meta->mm_version, env->me_psize));
4731 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4732 DPRINTF(("depth: %u", db->md_depth));
4733 DPRINTF(("entries: %"Yu, db->md_entries));
4734 DPRINTF(("branch pages: %"Yu, db->md_branch_pages));
4735 DPRINTF(("leaf pages: %"Yu, db->md_leaf_pages));
4736 DPRINTF(("overflow pages: %"Yu, db->md_overflow_pages));
4737 DPRINTF(("root: %"Yu, db->md_root));
4745 /** Release a reader thread's slot in the reader lock table.
4746 * This function is called automatically when a thread exits.
4747 * @param[in] ptr This points to the slot in the reader lock table.
4750 mdb_env_reader_dest(void *ptr)
4752 MDB_reader *reader = ptr;
4755 if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */
4757 /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */
4762 /** Junk for arranging thread-specific callbacks on Windows. This is
4763 * necessarily platform and compiler-specific. Windows supports up
4764 * to 1088 keys. Let's assume nobody opens more than 64 environments
4765 * in a single process, for now. They can override this if needed.
4767 #ifndef MAX_TLS_KEYS
4768 #define MAX_TLS_KEYS 64
4770 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4771 static int mdb_tls_nkeys;
4773 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4777 case DLL_PROCESS_ATTACH: break;
4778 case DLL_THREAD_ATTACH: break;
4779 case DLL_THREAD_DETACH:
4780 for (i=0; i<mdb_tls_nkeys; i++) {
4781 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4783 mdb_env_reader_dest(r);
4787 case DLL_PROCESS_DETACH: break;
4792 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4794 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4798 /* Force some symbol references.
4799 * _tls_used forces the linker to create the TLS directory if not already done
4800 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4802 #pragma comment(linker, "/INCLUDE:_tls_used")
4803 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4804 #pragma const_seg(".CRT$XLB")
4805 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4806 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4809 #pragma comment(linker, "/INCLUDE:__tls_used")
4810 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4811 #pragma data_seg(".CRT$XLB")
4812 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4814 #endif /* WIN 32/64 */
4815 #endif /* !__GNUC__ */
4818 /** Downgrade the exclusive lock on the region back to shared */
4820 mdb_env_share_locks(MDB_env *env, int *excl)
4823 MDB_meta *meta = mdb_env_pick_meta(env);
4825 env->me_txns->mti_txnid = meta->mm_txnid;
4830 /* First acquire a shared lock. The Unlock will
4831 * then release the existing exclusive lock.
4833 memset(&ov, 0, sizeof(ov));
4834 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4837 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4843 struct flock lock_info;
4844 /* The shared lock replaces the existing lock */
4845 memset((void *)&lock_info, 0, sizeof(lock_info));
4846 lock_info.l_type = F_RDLCK;
4847 lock_info.l_whence = SEEK_SET;
4848 lock_info.l_start = 0;
4849 lock_info.l_len = 1;
4850 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4851 (rc = ErrCode()) == EINTR) ;
4852 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4859 /** Try to get exclusive lock, otherwise shared.
4860 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4863 mdb_env_excl_lock(MDB_env *env, int *excl)
4867 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4871 memset(&ov, 0, sizeof(ov));
4872 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4879 struct flock lock_info;
4880 memset((void *)&lock_info, 0, sizeof(lock_info));
4881 lock_info.l_type = F_WRLCK;
4882 lock_info.l_whence = SEEK_SET;
4883 lock_info.l_start = 0;
4884 lock_info.l_len = 1;
4885 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4886 (rc = ErrCode()) == EINTR) ;
4890 # ifndef MDB_USE_POSIX_MUTEX
4891 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4894 lock_info.l_type = F_RDLCK;
4895 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4896 (rc = ErrCode()) == EINTR) ;
4906 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4908 * @(#) $Revision: 5.1 $
4909 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4910 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4912 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4916 * Please do not copyright this code. This code is in the public domain.
4918 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4919 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4920 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4921 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4922 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4923 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4924 * PERFORMANCE OF THIS SOFTWARE.
4927 * chongo <Landon Curt Noll> /\oo/\
4928 * http://www.isthe.com/chongo/
4930 * Share and Enjoy! :-)
4933 typedef unsigned long long mdb_hash_t;
4934 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4936 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4937 * @param[in] val value to hash
4938 * @param[in] hval initial value for hash
4939 * @return 64 bit hash
4941 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4942 * hval arg on the first call.
4945 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4947 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4948 unsigned char *end = s + val->mv_size;
4950 * FNV-1a hash each octet of the string
4953 /* xor the bottom with the current octet */
4954 hval ^= (mdb_hash_t)*s++;
4956 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4957 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4958 (hval << 7) + (hval << 8) + (hval << 40);
4960 /* return our new hash value */
4964 /** Hash the string and output the encoded hash.
4965 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4966 * very short name limits. We don't care about the encoding being reversible,
4967 * we just want to preserve as many bits of the input as possible in a
4968 * small printable string.
4969 * @param[in] str string to hash
4970 * @param[out] encbuf an array of 11 chars to hold the hash
4972 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4975 mdb_pack85(unsigned long l, char *out)
4979 for (i=0; i<5; i++) {
4980 *out++ = mdb_a85[l % 85];
4986 mdb_hash_enc(MDB_val *val, char *encbuf)
4988 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4990 mdb_pack85(h, encbuf);
4991 mdb_pack85(h>>32, encbuf+5);
4996 /** Open and/or initialize the lock region for the environment.
4997 * @param[in] env The LMDB environment.
4998 * @param[in] fname Filename + scratch area, from #mdb_fname_init().
4999 * @param[in] mode The Unix permissions for the file, if we create it.
5000 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
5001 * @return 0 on success, non-zero on failure.
5004 mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl)
5007 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
5009 # define MDB_ERRCODE_ROFS EROFS
5011 #ifdef MDB_USE_SYSV_SEM
5018 rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd);
5020 /* Omit lockfile if read-only env on read-only filesystem */
5021 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
5027 if (!(env->me_flags & MDB_NOTLS)) {
5028 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
5031 env->me_flags |= MDB_ENV_TXKEY;
5033 /* Windows TLS callbacks need help finding their TLS info. */
5034 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
5038 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
5042 /* Try to get exclusive lock. If we succeed, then
5043 * nobody is using the lock region and we should initialize it.
5045 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
5048 size = GetFileSize(env->me_lfd, NULL);
5050 size = lseek(env->me_lfd, 0, SEEK_END);
5051 if (size == -1) goto fail_errno;
5053 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
5054 if (size < rsize && *excl > 0) {
5056 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
5057 || !SetEndOfFile(env->me_lfd))
5060 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
5064 size = rsize - sizeof(MDB_txninfo);
5065 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
5070 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
5072 if (!mh) goto fail_errno;
5073 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
5075 if (!env->me_txns) goto fail_errno;
5077 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
5079 if (m == MAP_FAILED) goto fail_errno;
5085 BY_HANDLE_FILE_INFORMATION stbuf;
5094 if (!mdb_sec_inited) {
5095 InitializeSecurityDescriptor(&mdb_null_sd,
5096 SECURITY_DESCRIPTOR_REVISION);
5097 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
5098 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
5099 mdb_all_sa.bInheritHandle = FALSE;
5100 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
5103 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
5104 idbuf.volume = stbuf.dwVolumeSerialNumber;
5105 idbuf.nhigh = stbuf.nFileIndexHigh;
5106 idbuf.nlow = stbuf.nFileIndexLow;
5107 val.mv_data = &idbuf;
5108 val.mv_size = sizeof(idbuf);
5109 mdb_hash_enc(&val, encbuf);
5110 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
5111 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
5112 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
5113 if (!env->me_rmutex) goto fail_errno;
5114 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
5115 if (!env->me_wmutex) goto fail_errno;
5116 #elif defined(MDB_USE_POSIX_SEM)
5125 #if defined(__NetBSD__)
5126 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
5128 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
5129 idbuf.dev = stbuf.st_dev;
5130 idbuf.ino = stbuf.st_ino;
5131 val.mv_data = &idbuf;
5132 val.mv_size = sizeof(idbuf);
5133 mdb_hash_enc(&val, encbuf);
5134 #ifdef MDB_SHORT_SEMNAMES
5135 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
5137 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
5138 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
5139 /* Clean up after a previous run, if needed: Try to
5140 * remove both semaphores before doing anything else.
5142 sem_unlink(env->me_txns->mti_rmname);
5143 sem_unlink(env->me_txns->mti_wmname);
5144 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
5145 O_CREAT|O_EXCL, mode, 1);
5146 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
5147 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
5148 O_CREAT|O_EXCL, mode, 1);
5149 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
5150 #elif defined(MDB_USE_SYSV_SEM)
5151 unsigned short vals[2] = {1, 1};
5152 key_t key = ftok(fname->mn_val, 'M'); /* fname is lockfile path now */
5155 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
5159 if (semctl(semid, 0, SETALL, semu) < 0)
5161 env->me_txns->mti_semid = semid;
5162 env->me_txns->mti_rlocked = 0;
5163 env->me_txns->mti_wlocked = 0;
5164 #else /* MDB_USE_POSIX_MUTEX: */
5165 pthread_mutexattr_t mattr;
5167 /* Solaris needs this before initing a robust mutex. Otherwise
5168 * it may skip the init and return EBUSY "seems someone already
5169 * inited" or EINVAL "it was inited differently".
5171 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
5172 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
5174 if ((rc = pthread_mutexattr_init(&mattr)) != 0)
5176 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
5177 #ifdef MDB_ROBUST_SUPPORTED
5178 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
5180 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
5181 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
5182 pthread_mutexattr_destroy(&mattr);
5185 #endif /* _WIN32 || ... */
5187 env->me_txns->mti_magic = MDB_MAGIC;
5188 env->me_txns->mti_format = MDB_LOCK_FORMAT;
5189 env->me_txns->mti_txnid = 0;
5190 env->me_txns->mti_numreaders = 0;
5193 #ifdef MDB_USE_SYSV_SEM
5194 struct semid_ds buf;
5196 if (env->me_txns->mti_magic != MDB_MAGIC) {
5197 DPUTS("lock region has invalid magic");
5201 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
5202 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
5203 env->me_txns->mti_format, MDB_LOCK_FORMAT));
5204 rc = MDB_VERSION_MISMATCH;
5208 if (rc && rc != EACCES && rc != EAGAIN) {
5212 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
5213 if (!env->me_rmutex) goto fail_errno;
5214 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
5215 if (!env->me_wmutex) goto fail_errno;
5216 #elif defined(MDB_USE_POSIX_SEM)
5217 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
5218 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
5219 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
5220 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
5221 #elif defined(MDB_USE_SYSV_SEM)
5222 semid = env->me_txns->mti_semid;
5224 /* check for read access */
5225 if (semctl(semid, 0, IPC_STAT, semu) < 0)
5227 /* check for write access */
5228 if (semctl(semid, 0, IPC_SET, semu) < 0)
5232 #ifdef MDB_USE_SYSV_SEM
5233 env->me_rmutex->semid = semid;
5234 env->me_wmutex->semid = semid;
5235 env->me_rmutex->semnum = 0;
5236 env->me_wmutex->semnum = 1;
5237 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
5238 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
5242 env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
5244 pthread_mutex_init(&env->me_rpmutex, NULL);
5256 /** Only a subset of the @ref mdb_env flags can be changed
5257 * at runtime. Changing other flags requires closing the
5258 * environment and re-opening it with the new flags.
5260 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
5261 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
5262 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
5264 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
5265 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
5269 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
5274 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
5278 if (flags & MDB_WRITEMAP) {
5279 /* silently ignore WRITEMAP in 32 bit mode */
5280 flags ^= MDB_WRITEMAP;
5282 if (flags & MDB_FIXEDMAP) {
5283 /* cannot support FIXEDMAP */
5287 flags |= env->me_flags;
5289 rc = mdb_fname_init(path, flags, &fname);
5293 if (flags & MDB_RDONLY) {
5294 /* silently ignore WRITEMAP when we're only getting read access */
5295 flags &= ~MDB_WRITEMAP;
5297 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
5298 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
5303 env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
5304 if (!env->me_rpages) {
5308 env->me_rpages[0].mid = 0;
5309 env->me_rpcheck = MDB_ERPAGE_SIZE/2;
5312 env->me_flags = flags |= MDB_ENV_ACTIVE;
5316 env->me_path = strdup(path);
5317 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
5318 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
5319 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
5320 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
5324 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
5326 /* For RDONLY, get lockfile after we know datafile exists */
5327 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
5328 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
5333 rc = mdb_fopen(env, &fname,
5334 (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR,
5339 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
5340 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
5345 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
5346 if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) {
5347 /* Synchronous fd for meta writes. Needed even with
5348 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
5350 rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd);
5354 DPRINTF(("opened dbenv %p", (void *) env));
5356 rc = mdb_env_share_locks(env, &excl);
5360 if (!(flags & MDB_RDONLY)) {
5362 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
5363 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
5364 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
5365 (txn = calloc(1, size)))
5367 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
5368 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
5369 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
5370 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
5373 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
5374 if (!txn->mt_rpages) {
5379 txn->mt_rpages[0].mid = 0;
5380 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
5382 txn->mt_dbxs = env->me_dbxs;
5383 txn->mt_flags = MDB_TXN_FINISHED;
5393 mdb_env_close0(env, excl);
5395 mdb_fname_destroy(fname);
5399 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5401 mdb_env_close0(MDB_env *env, int excl)
5405 if (!(env->me_flags & MDB_ENV_ACTIVE))
5408 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5410 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5411 free(env->me_dbxs[i].md_name.mv_data);
5416 free(env->me_dbiseqs);
5417 free(env->me_dbflags);
5419 free(env->me_dirty_list);
5421 if (env->me_txn0 && env->me_txn0->mt_rpages)
5422 free(env->me_txn0->mt_rpages);
5424 for (x=1; x<=env->me_rpages[0].mid; x++)
5425 munmap(env->me_rpages[x].mptr, env->me_rpages[x].mcnt * env->me_psize);
5427 free(env->me_rpages);
5430 mdb_midl_free(env->me_free_pgs);
5432 if (env->me_flags & MDB_ENV_TXKEY) {
5433 pthread_key_delete(env->me_txkey);
5435 /* Delete our key from the global list */
5436 for (i=0; i<mdb_tls_nkeys; i++)
5437 if (mdb_tls_keys[i] == env->me_txkey) {
5438 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5447 munmap(env->me_map, NUM_METAS*env->me_psize);
5449 munmap(env->me_map, env->me_mapsize);
5452 if (env->me_mfd != INVALID_HANDLE_VALUE)
5453 (void) close(env->me_mfd);
5454 if (env->me_fd != INVALID_HANDLE_VALUE)
5455 (void) close(env->me_fd);
5457 MDB_PID_T pid = env->me_pid;
5458 /* Clearing readers is done in this function because
5459 * me_txkey with its destructor must be disabled first.
5461 * We skip the the reader mutex, so we touch only
5462 * data owned by this process (me_close_readers and
5463 * our readers), and clear each reader atomically.
5465 for (i = env->me_close_readers; --i >= 0; )
5466 if (env->me_txns->mti_readers[i].mr_pid == pid)
5467 env->me_txns->mti_readers[i].mr_pid = 0;
5469 if (env->me_rmutex) {
5470 CloseHandle(env->me_rmutex);
5471 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5473 /* Windows automatically destroys the mutexes when
5474 * the last handle closes.
5476 #elif defined(MDB_USE_POSIX_SEM)
5477 if (env->me_rmutex != SEM_FAILED) {
5478 sem_close(env->me_rmutex);
5479 if (env->me_wmutex != SEM_FAILED)
5480 sem_close(env->me_wmutex);
5481 /* If we have the filelock: If we are the
5482 * only remaining user, clean up semaphores.
5485 mdb_env_excl_lock(env, &excl);
5487 sem_unlink(env->me_txns->mti_rmname);
5488 sem_unlink(env->me_txns->mti_wmname);
5491 #elif defined(MDB_USE_SYSV_SEM)
5492 if (env->me_rmutex->semid != -1) {
5493 /* If we have the filelock: If we are the
5494 * only remaining user, clean up semaphores.
5497 mdb_env_excl_lock(env, &excl);
5499 semctl(env->me_rmutex->semid, 0, IPC_RMID);
5502 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5504 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5507 /* Unlock the lockfile. Windows would have unlocked it
5508 * after closing anyway, but not necessarily at once.
5510 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5513 (void) close(env->me_lfd);
5517 if (env->me_fmh) CloseHandle(env->me_fmh);
5518 if (env->me_rpmutex) CloseHandle(env->me_rpmutex);
5520 pthread_mutex_destroy(&env->me_rpmutex);
5524 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5528 mdb_env_close(MDB_env *env)
5535 VGMEMP_DESTROY(env);
5536 while ((dp = env->me_dpages) != NULL) {
5537 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5538 env->me_dpages = dp->mp_next;
5542 mdb_env_close0(env, 0);
5546 /** Compare two items pointing at aligned #mdb_size_t's */
5548 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5550 return (*(mdb_size_t *)a->mv_data < *(mdb_size_t *)b->mv_data) ? -1 :
5551 *(mdb_size_t *)a->mv_data > *(mdb_size_t *)b->mv_data;
5554 /** Compare two items pointing at aligned unsigned int's.
5556 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5557 * but #mdb_cmp_clong() is called instead if the data type is #mdb_size_t.
5560 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5562 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5563 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5566 /** Compare two items pointing at unsigned ints of unknown alignment.
5567 * Nodes and keys are guaranteed to be 2-byte aligned.
5570 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5572 #if BYTE_ORDER == LITTLE_ENDIAN
5573 unsigned short *u, *c;
5576 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5577 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5580 } while(!x && u > (unsigned short *)a->mv_data);
5583 unsigned short *u, *c, *end;
5586 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5587 u = (unsigned short *)a->mv_data;
5588 c = (unsigned short *)b->mv_data;
5591 } while(!x && u < end);
5596 /** Compare two items lexically */
5598 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5605 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5611 diff = memcmp(a->mv_data, b->mv_data, len);
5612 return diff ? diff : len_diff<0 ? -1 : len_diff;
5615 /** Compare two items in reverse byte order */
5617 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5619 const unsigned char *p1, *p2, *p1_lim;
5623 p1_lim = (const unsigned char *)a->mv_data;
5624 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5625 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5627 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5633 while (p1 > p1_lim) {
5634 diff = *--p1 - *--p2;
5638 return len_diff<0 ? -1 : len_diff;
5641 /** Search for key within a page, using binary search.
5642 * Returns the smallest entry larger or equal to the key.
5643 * If exactp is non-null, stores whether the found entry was an exact match
5644 * in *exactp (1 or 0).
5645 * Updates the cursor index with the index of the found entry.
5646 * If no entry larger or equal to the key is found, returns NULL.
5649 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5651 unsigned int i = 0, nkeys;
5654 MDB_page *mp = mc->mc_pg[mc->mc_top];
5655 MDB_node *node = NULL;
5660 nkeys = NUMKEYS(mp);
5662 DPRINTF(("searching %u keys in %s %spage %"Yu,
5663 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5666 low = IS_LEAF(mp) ? 0 : 1;
5668 cmp = mc->mc_dbx->md_cmp;
5670 /* Branch pages have no data, so if using integer keys,
5671 * alignment is guaranteed. Use faster mdb_cmp_int.
5673 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5674 if (NODEPTR(mp, 1)->mn_ksize == sizeof(mdb_size_t))
5681 nodekey.mv_size = mc->mc_db->md_pad;
5682 node = NODEPTR(mp, 0); /* fake */
5683 while (low <= high) {
5684 i = (low + high) >> 1;
5685 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5686 rc = cmp(key, &nodekey);
5687 DPRINTF(("found leaf index %u [%s], rc = %i",
5688 i, DKEY(&nodekey), rc));
5697 while (low <= high) {
5698 i = (low + high) >> 1;
5700 node = NODEPTR(mp, i);
5701 nodekey.mv_size = NODEKSZ(node);
5702 nodekey.mv_data = NODEKEY(node);
5704 rc = cmp(key, &nodekey);
5707 DPRINTF(("found leaf index %u [%s], rc = %i",
5708 i, DKEY(&nodekey), rc));
5710 DPRINTF(("found branch index %u [%s -> %"Yu"], rc = %i",
5711 i, DKEY(&nodekey), NODEPGNO(node), rc));
5722 if (rc > 0) { /* Found entry is less than the key. */
5723 i++; /* Skip to get the smallest entry larger than key. */
5725 node = NODEPTR(mp, i);
5728 *exactp = (rc == 0 && nkeys > 0);
5729 /* store the key index */
5730 mc->mc_ki[mc->mc_top] = i;
5732 /* There is no entry larger or equal to the key. */
5735 /* nodeptr is fake for LEAF2 */
5741 mdb_cursor_adjust(MDB_cursor *mc, func)
5745 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5746 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5753 /** Pop a page off the top of the cursor's stack. */
5755 mdb_cursor_pop(MDB_cursor *mc)
5758 DPRINTF(("popping page %"Yu" off db %d cursor %p",
5759 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5765 mc->mc_flags &= ~C_INITIALIZED;
5770 /** Push a page onto the top of the cursor's stack. */
5772 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5774 DPRINTF(("pushing page %"Yu" on db %d cursor %p", mp->mp_pgno,
5775 DDBI(mc), (void *) mc));
5777 if (mc->mc_snum >= CURSOR_STACK) {
5778 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5779 return MDB_CURSOR_FULL;
5782 mc->mc_top = mc->mc_snum++;
5783 mc->mc_pg[mc->mc_top] = mp;
5784 mc->mc_ki[mc->mc_top] = 0;
5790 /** Map a read-only page.
5791 * There are two levels of tracking in use, a per-txn list and a per-env list.
5792 * ref'ing and unref'ing the per-txn list is faster since it requires no
5793 * locking. Pages are cached in the per-env list for global reuse, and a lock
5794 * is required. Pages are not immediately unmapped when their refcnt goes to
5795 * zero; they hang around in case they will be reused again soon.
5797 * When the per-txn list gets full, all pages with refcnt=0 are purged from the
5798 * list and their refcnts in the per-env list are decremented.
5800 * When the per-env list gets full, all pages with refcnt=0 are purged from the
5801 * list and their pages are unmapped.
5803 * @note "full" means the list has reached its respective rpcheck threshold.
5804 * This threshold slowly raises if no pages could be purged on a given check,
5805 * and returns to its original value when enough pages were purged.
5807 * If purging doesn't free any slots, filling the per-txn list will return
5808 * MDB_TXN_FULL, and filling the per-env list returns MDB_MAP_FULL.
5810 * Reference tracking in a txn is imperfect, pages can linger with non-zero
5811 * refcnt even without active references. It was deemed to be too invasive
5812 * to add unrefs in every required location. However, all pages are unref'd
5813 * at the end of the transaction. This guarantees that no stale references
5814 * linger in the per-env list.
5816 * Usually we map chunks of 16 pages at a time, but if an overflow page begins
5817 * at the tail of the chunk we extend the chunk to include the entire overflow
5818 * page. Unfortunately, pages can be turned into overflow pages after their
5819 * chunk was already mapped. In that case we must remap the chunk if the
5820 * overflow page is referenced. If the chunk's refcnt is 0 we can just remap
5821 * it, otherwise we temporarily map a new chunk just for the overflow page.
5823 * @note this chunk handling means we cannot guarantee that a data item
5824 * returned from the DB will stay alive for the duration of the transaction:
5825 * We unref pages as soon as a cursor moves away from the page
5826 * A subsequent op may cause a purge, which may unmap any unref'd chunks
5827 * The caller must copy the data if it must be used later in the same txn.
5829 * Also - our reference counting revolves around cursors, but overflow pages
5830 * aren't pointed to by a cursor's page stack. We have to remember them
5831 * explicitly, in the added mc_ovpg field. A single cursor can only hold a
5832 * reference to one overflow page at a time.
5834 * @param[in] txn the transaction for this access.
5835 * @param[in] pgno the page number for the page to retrieve.
5836 * @param[out] ret address of a pointer where the page's address will be stored.
5837 * @return 0 on success, non-zero on failure.
5840 mdb_rpage_get(MDB_txn *txn, pgno_t pg0, MDB_page **ret)
5842 MDB_env *env = txn->mt_env;
5844 MDB_ID3L tl = txn->mt_rpages;
5845 MDB_ID3L el = env->me_rpages;
5849 int rc, retries = 1;
5853 #define SET_OFF(off,val) off.QuadPart = val
5854 #define MAP(rc,env,addr,len,off) \
5856 rc = NtMapViewOfSection(env->me_fmh, GetCurrentProcess(), &addr, 0, \
5857 len, &off, &len, ViewUnmap, (env->me_flags & MDB_RDONLY) ? 0 : MEM_RESERVE, PAGE_READONLY); \
5858 if (rc) rc = mdb_nt2win32(rc)
5862 #define SET_OFF(off,val) off = val
5863 #define MAP(rc,env,addr,len,off) \
5864 addr = mmap(NULL, len, PROT_READ, MAP_SHARED, env->me_fd, off); \
5865 rc = (addr == MAP_FAILED) ? errno : 0
5868 /* remember the offset of the actual page number, so we can
5869 * return the correct pointer at the end.
5871 rem = pg0 & (MDB_RPAGE_CHUNK-1);
5875 x = mdb_mid3l_search(tl, pgno);
5876 if (x <= tl[0].mid && tl[x].mid == pgno) {
5877 if (x != tl[0].mid && tl[x+1].mid == pg0)
5879 /* check for overflow size */
5880 p = (MDB_page *)((char *)tl[x].mptr + rem * env->me_psize);
5881 if (IS_OVERFLOW(p) && p->mp_pages + rem > tl[x].mcnt) {
5882 id3.mcnt = p->mp_pages + rem;
5883 len = id3.mcnt * env->me_psize;
5884 SET_OFF(off, pgno * env->me_psize);
5885 MAP(rc, env, id3.mptr, len, off);
5888 /* check for local-only page */
5890 mdb_tassert(txn, tl[x].mid != pg0);
5891 /* hope there's room to insert this locally.
5892 * setting mid here tells later code to just insert
5893 * this id3 instead of searching for a match.
5898 /* ignore the mapping we got from env, use new one */
5899 tl[x].mptr = id3.mptr;
5900 tl[x].mcnt = id3.mcnt;
5901 /* if no active ref, see if we can replace in env */
5904 pthread_mutex_lock(&env->me_rpmutex);
5905 i = mdb_mid3l_search(el, tl[x].mid);
5906 if (el[i].mref == 1) {
5907 /* just us, replace it */
5908 munmap(el[i].mptr, el[i].mcnt * env->me_psize);
5909 el[i].mptr = tl[x].mptr;
5910 el[i].mcnt = tl[x].mcnt;
5912 /* there are others, remove ourself */
5915 pthread_mutex_unlock(&env->me_rpmutex);
5919 id3.mptr = tl[x].mptr;
5920 id3.mcnt = tl[x].mcnt;
5926 if (tl[0].mid >= MDB_TRPAGE_MAX - txn->mt_rpcheck) {
5928 /* purge unref'd pages from our list and unref in env */
5929 pthread_mutex_lock(&env->me_rpmutex);
5932 for (i=1; i<=tl[0].mid; i++) {
5935 /* tmp overflow pages don't go to env */
5936 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
5937 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
5940 x = mdb_mid3l_search(el, tl[i].mid);
5944 pthread_mutex_unlock(&env->me_rpmutex);
5946 /* we didn't find any unref'd chunks.
5947 * if we're out of room, fail.
5949 if (tl[0].mid >= MDB_TRPAGE_MAX)
5950 return MDB_TXN_FULL;
5951 /* otherwise, raise threshold for next time around
5954 txn->mt_rpcheck /= 2;
5956 /* we found some unused; consolidate the list */
5957 for (i=y+1; i<= tl[0].mid; i++)
5961 /* decrease the check threshold toward its original value */
5962 if (!txn->mt_rpcheck)
5963 txn->mt_rpcheck = 1;
5964 while (txn->mt_rpcheck < tl[0].mid && txn->mt_rpcheck < MDB_TRPAGE_SIZE/2)
5965 txn->mt_rpcheck *= 2;
5968 if (tl[0].mid < MDB_TRPAGE_SIZE) {
5972 /* don't map past last written page in read-only envs */
5973 if ((env->me_flags & MDB_RDONLY) && pgno + MDB_RPAGE_CHUNK-1 > txn->mt_last_pgno)
5974 id3.mcnt = txn->mt_last_pgno + 1 - pgno;
5976 id3.mcnt = MDB_RPAGE_CHUNK;
5977 len = id3.mcnt * env->me_psize;
5980 /* search for page in env */
5981 pthread_mutex_lock(&env->me_rpmutex);
5982 x = mdb_mid3l_search(el, pgno);
5983 if (x <= el[0].mid && el[x].mid == pgno) {
5984 id3.mptr = el[x].mptr;
5985 id3.mcnt = el[x].mcnt;
5986 /* check for overflow size */
5987 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5988 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
5989 id3.mcnt = p->mp_pages + rem;
5990 len = id3.mcnt * env->me_psize;
5991 SET_OFF(off, pgno * env->me_psize);
5992 MAP(rc, env, id3.mptr, len, off);
5996 munmap(el[x].mptr, env->me_psize * el[x].mcnt);
5997 el[x].mptr = id3.mptr;
5998 el[x].mcnt = id3.mcnt;
6001 pthread_mutex_unlock(&env->me_rpmutex);
6006 pthread_mutex_unlock(&env->me_rpmutex);
6009 if (el[0].mid >= MDB_ERPAGE_MAX - env->me_rpcheck) {
6010 /* purge unref'd pages */
6012 for (i=1; i<=el[0].mid; i++) {
6015 munmap(el[i].mptr, env->me_psize * el[i].mcnt);
6020 /* see if we can unref some local pages */
6025 if (el[0].mid >= MDB_ERPAGE_MAX) {
6026 pthread_mutex_unlock(&env->me_rpmutex);
6027 return MDB_MAP_FULL;
6029 env->me_rpcheck /= 2;
6031 for (i=y+1; i<= el[0].mid; i++)
6035 if (!env->me_rpcheck)
6036 env->me_rpcheck = 1;
6037 while (env->me_rpcheck < el[0].mid && env->me_rpcheck < MDB_ERPAGE_SIZE/2)
6038 env->me_rpcheck *= 2;
6041 SET_OFF(off, pgno * env->me_psize);
6042 MAP(rc, env, id3.mptr, len, off);
6045 pthread_mutex_unlock(&env->me_rpmutex);
6048 /* check for overflow size */
6049 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
6050 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
6051 id3.mcnt = p->mp_pages + rem;
6052 munmap(id3.mptr, len);
6053 len = id3.mcnt * env->me_psize;
6054 MAP(rc, env, id3.mptr, len, off);
6058 mdb_mid3l_insert(el, &id3);
6059 pthread_mutex_unlock(&env->me_rpmutex);
6061 mdb_mid3l_insert(tl, &id3);
6063 return MDB_TXN_FULL;
6066 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
6067 #if MDB_DEBUG /* we don't need this check any more */
6068 if (IS_OVERFLOW(p)) {
6069 mdb_tassert(txn, p->mp_pages + rem <= id3.mcnt);
6077 /** Find the address of the page corresponding to a given page number.
6078 * @param[in] mc the cursor accessing the page.
6079 * @param[in] pgno the page number for the page to retrieve.
6080 * @param[out] ret address of a pointer where the page's address will be stored.
6081 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
6082 * @return 0 on success, non-zero on failure.
6085 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
6087 MDB_txn *txn = mc->mc_txn;
6091 if (! (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP))) {
6095 MDB_ID2L dl = tx2->mt_u.dirty_list;
6097 /* Spilled pages were dirtied in this txn and flushed
6098 * because the dirty list got full. Bring this page
6099 * back in from the map (but don't unspill it here,
6100 * leave that unless page_touch happens again).
6102 if (tx2->mt_spill_pgs) {
6103 MDB_ID pn = pgno << 1;
6104 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
6105 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
6110 unsigned x = mdb_mid2l_search(dl, pgno);
6111 if (x <= dl[0].mid && dl[x].mid == pgno) {
6117 } while ((tx2 = tx2->mt_parent) != NULL);
6120 if (pgno >= txn->mt_next_pgno) {
6121 DPRINTF(("page %"Yu" not found", pgno));
6122 txn->mt_flags |= MDB_TXN_ERROR;
6123 return MDB_PAGE_NOTFOUND;
6131 int rc = mdb_rpage_get(txn, pgno, &p);
6135 MDB_env *env = txn->mt_env;
6136 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
6147 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
6148 * The cursor is at the root page, set up the rest of it.
6151 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
6153 MDB_page *mp = mc->mc_pg[mc->mc_top];
6157 while (IS_BRANCH(mp)) {
6161 DPRINTF(("branch page %"Yu" has %u keys", mp->mp_pgno, NUMKEYS(mp)));
6162 /* Don't assert on branch pages in the FreeDB. We can get here
6163 * while in the process of rebalancing a FreeDB branch page; we must
6164 * let that proceed. ITS#8336
6166 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
6167 DPRINTF(("found index 0 to page %"Yu, NODEPGNO(NODEPTR(mp, 0))));
6169 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
6171 if (flags & MDB_PS_LAST)
6172 i = NUMKEYS(mp) - 1;
6175 node = mdb_node_search(mc, key, &exact);
6177 i = NUMKEYS(mp) - 1;
6179 i = mc->mc_ki[mc->mc_top];
6181 mdb_cassert(mc, i > 0);
6185 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
6188 mdb_cassert(mc, i < NUMKEYS(mp));
6189 node = NODEPTR(mp, i);
6191 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6194 mc->mc_ki[mc->mc_top] = i;
6195 if ((rc = mdb_cursor_push(mc, mp)))
6198 if (flags & MDB_PS_MODIFY) {
6199 if ((rc = mdb_page_touch(mc)) != 0)
6201 mp = mc->mc_pg[mc->mc_top];
6206 DPRINTF(("internal error, index points to a %02X page!?",
6208 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6209 return MDB_CORRUPTED;
6212 DPRINTF(("found leaf page %"Yu" for key [%s]", mp->mp_pgno,
6213 key ? DKEY(key) : "null"));
6214 mc->mc_flags |= C_INITIALIZED;
6215 mc->mc_flags &= ~C_EOF;
6220 /** Search for the lowest key under the current branch page.
6221 * This just bypasses a NUMKEYS check in the current page
6222 * before calling mdb_page_search_root(), because the callers
6223 * are all in situations where the current page is known to
6227 mdb_page_search_lowest(MDB_cursor *mc)
6229 MDB_page *mp = mc->mc_pg[mc->mc_top];
6230 MDB_node *node = NODEPTR(mp, 0);
6233 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6236 mc->mc_ki[mc->mc_top] = 0;
6237 if ((rc = mdb_cursor_push(mc, mp)))
6239 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
6242 /** Search for the page a given key should be in.
6243 * Push it and its parent pages on the cursor stack.
6244 * @param[in,out] mc the cursor for this operation.
6245 * @param[in] key the key to search for, or NULL for first/last page.
6246 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
6247 * are touched (updated with new page numbers).
6248 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
6249 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
6250 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
6251 * @return 0 on success, non-zero on failure.
6254 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
6259 /* Make sure the txn is still viable, then find the root from
6260 * the txn's db table and set it as the root of the cursor's stack.
6262 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
6263 DPUTS("transaction may not be used now");
6266 /* Make sure we're using an up-to-date root */
6267 if (*mc->mc_dbflag & DB_STALE) {
6269 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6271 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
6272 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
6279 MDB_node *leaf = mdb_node_search(&mc2,
6280 &mc->mc_dbx->md_name, &exact);
6282 return MDB_NOTFOUND;
6283 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
6284 return MDB_INCOMPATIBLE; /* not a named DB */
6285 rc = mdb_node_read(&mc2, leaf, &data);
6288 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
6290 /* The txn may not know this DBI, or another process may
6291 * have dropped and recreated the DB with other flags.
6293 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
6294 return MDB_INCOMPATIBLE;
6295 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
6297 *mc->mc_dbflag &= ~DB_STALE;
6299 root = mc->mc_db->md_root;
6301 if (root == P_INVALID) { /* Tree is empty. */
6302 DPUTS("tree is empty");
6303 return MDB_NOTFOUND;
6307 mdb_cassert(mc, root > 1);
6308 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) {
6311 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[0]);
6313 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
6320 for (i=1; i<mc->mc_snum; i++)
6321 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[i]);
6327 DPRINTF(("db %d root page %"Yu" has flags 0x%X",
6328 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
6330 if (flags & MDB_PS_MODIFY) {
6331 if ((rc = mdb_page_touch(mc)))
6335 if (flags & MDB_PS_ROOTONLY)
6338 return mdb_page_search_root(mc, key, flags);
6342 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
6344 MDB_txn *txn = mc->mc_txn;
6345 pgno_t pg = mp->mp_pgno;
6346 unsigned x = 0, ovpages = mp->mp_pages;
6347 MDB_env *env = txn->mt_env;
6348 MDB_IDL sl = txn->mt_spill_pgs;
6349 MDB_ID pn = pg << 1;
6352 DPRINTF(("free ov page %"Yu" (%d)", pg, ovpages));
6353 /* If the page is dirty or on the spill list we just acquired it,
6354 * so we should give it back to our current free list, if any.
6355 * Otherwise put it onto the list of pages we freed in this txn.
6357 * Won't create me_pghead: me_pglast must be inited along with it.
6358 * Unsupported in nested txns: They would need to hide the page
6359 * range in ancestor txns' dirty and spilled lists.
6361 if (env->me_pghead &&
6363 ((mp->mp_flags & P_DIRTY) ||
6364 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
6368 MDB_ID2 *dl, ix, iy;
6369 rc = mdb_midl_need(&env->me_pghead, ovpages);
6372 if (!(mp->mp_flags & P_DIRTY)) {
6373 /* This page is no longer spilled */
6380 /* Remove from dirty list */
6381 dl = txn->mt_u.dirty_list;
6383 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
6389 mdb_cassert(mc, x > 1);
6391 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
6392 txn->mt_flags |= MDB_TXN_ERROR;
6396 txn->mt_dirty_room++;
6397 if (!(env->me_flags & MDB_WRITEMAP))
6398 mdb_dpage_free(env, mp);
6400 /* Insert in me_pghead */
6401 mop = env->me_pghead;
6402 j = mop[0] + ovpages;
6403 for (i = mop[0]; i && mop[i] < pg; i--)
6409 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
6413 mc->mc_db->md_overflow_pages -= ovpages;
6417 /** Return the data associated with a given node.
6418 * @param[in] mc The cursor for this operation.
6419 * @param[in] leaf The node being read.
6420 * @param[out] data Updated to point to the node's data.
6421 * @return 0 on success, non-zero on failure.
6424 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
6426 MDB_page *omp; /* overflow page */
6431 MDB_PAGE_UNREF(mc->mc_txn, MC_OVPG(mc));
6432 MC_SET_OVPG(mc, NULL);
6434 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6435 data->mv_size = NODEDSZ(leaf);
6436 data->mv_data = NODEDATA(leaf);
6440 /* Read overflow data.
6442 data->mv_size = NODEDSZ(leaf);
6443 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
6444 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
6445 DPRINTF(("read overflow page %"Yu" failed", pgno));
6448 data->mv_data = METADATA(omp);
6449 MC_SET_OVPG(mc, omp);
6455 mdb_get(MDB_txn *txn, MDB_dbi dbi,
6456 MDB_val *key, MDB_val *data)
6463 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
6465 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
6468 if (txn->mt_flags & MDB_TXN_BLOCKED)
6471 mdb_cursor_init(&mc, txn, dbi, &mx);
6472 rc = mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
6473 /* unref all the pages when MDB_VL32 - caller must copy the data
6474 * before doing anything else
6476 MDB_CURSOR_UNREF(&mc, 1);
6480 /** Find a sibling for a page.
6481 * Replaces the page at the top of the cursor's stack with the
6482 * specified sibling, if one exists.
6483 * @param[in] mc The cursor for this operation.
6484 * @param[in] move_right Non-zero if the right sibling is requested,
6485 * otherwise the left sibling.
6486 * @return 0 on success, non-zero on failure.
6489 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
6498 if (mc->mc_snum < 2) {
6499 return MDB_NOTFOUND; /* root has no siblings */
6503 op = mc->mc_pg[mc->mc_top];
6506 DPRINTF(("parent page is page %"Yu", index %u",
6507 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
6509 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6510 : (mc->mc_ki[mc->mc_top] == 0)) {
6511 DPRINTF(("no more keys left, moving to %s sibling",
6512 move_right ? "right" : "left"));
6513 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
6514 /* undo cursor_pop before returning */
6521 mc->mc_ki[mc->mc_top]++;
6523 mc->mc_ki[mc->mc_top]--;
6524 DPRINTF(("just moving to %s index key %u",
6525 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
6527 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
6529 MDB_PAGE_UNREF(mc->mc_txn, op);
6531 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6532 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
6533 /* mc will be inconsistent if caller does mc_snum++ as above */
6534 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
6538 mdb_cursor_push(mc, mp);
6540 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
6545 /** Move the cursor to the next data item. */
6547 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6553 if ((mc->mc_flags & C_EOF) ||
6554 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
6555 return MDB_NOTFOUND;
6557 if (!(mc->mc_flags & C_INITIALIZED))
6558 return mdb_cursor_first(mc, key, data);
6560 mp = mc->mc_pg[mc->mc_top];
6562 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6563 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6564 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6565 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
6566 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
6567 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
6568 if (rc == MDB_SUCCESS)
6569 MDB_GET_KEY(leaf, key);
6574 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6577 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6578 if (op == MDB_NEXT_DUP)
6579 return MDB_NOTFOUND;
6583 DPRINTF(("cursor_next: top page is %"Yu" in cursor %p",
6584 mdb_dbg_pgno(mp), (void *) mc));
6585 if (mc->mc_flags & C_DEL) {
6586 mc->mc_flags ^= C_DEL;
6590 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
6591 DPUTS("=====> move to next sibling page");
6592 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6593 mc->mc_flags |= C_EOF;
6596 mp = mc->mc_pg[mc->mc_top];
6597 DPRINTF(("next page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6599 mc->mc_ki[mc->mc_top]++;
6602 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6603 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6606 key->mv_size = mc->mc_db->md_pad;
6607 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6611 mdb_cassert(mc, IS_LEAF(mp));
6612 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6614 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6615 mdb_xcursor_init1(mc, leaf);
6618 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6621 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6622 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6623 if (rc != MDB_SUCCESS)
6628 MDB_GET_KEY(leaf, key);
6632 /** Move the cursor to the previous data item. */
6634 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6640 if (!(mc->mc_flags & C_INITIALIZED)) {
6641 rc = mdb_cursor_last(mc, key, data);
6644 mc->mc_ki[mc->mc_top]++;
6647 mp = mc->mc_pg[mc->mc_top];
6649 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6650 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6651 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6652 if (op == MDB_PREV || op == MDB_PREV_DUP) {
6653 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
6654 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
6655 if (rc == MDB_SUCCESS) {
6656 MDB_GET_KEY(leaf, key);
6657 mc->mc_flags &= ~C_EOF;
6663 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6666 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6667 if (op == MDB_PREV_DUP)
6668 return MDB_NOTFOUND;
6672 DPRINTF(("cursor_prev: top page is %"Yu" in cursor %p",
6673 mdb_dbg_pgno(mp), (void *) mc));
6675 mc->mc_flags &= ~(C_EOF|C_DEL);
6677 if (mc->mc_ki[mc->mc_top] == 0) {
6678 DPUTS("=====> move to prev sibling page");
6679 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
6682 mp = mc->mc_pg[mc->mc_top];
6683 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
6684 DPRINTF(("prev page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6686 mc->mc_ki[mc->mc_top]--;
6688 mc->mc_flags &= ~C_EOF;
6690 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6691 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6694 key->mv_size = mc->mc_db->md_pad;
6695 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6699 mdb_cassert(mc, IS_LEAF(mp));
6700 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6702 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6703 mdb_xcursor_init1(mc, leaf);
6706 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6709 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6710 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6711 if (rc != MDB_SUCCESS)
6716 MDB_GET_KEY(leaf, key);
6720 /** Set the cursor on a specific data item. */
6722 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6723 MDB_cursor_op op, int *exactp)
6727 MDB_node *leaf = NULL;
6730 if (key->mv_size == 0)
6731 return MDB_BAD_VALSIZE;
6733 if (mc->mc_xcursor) {
6734 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6735 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6738 /* See if we're already on the right page */
6739 if (mc->mc_flags & C_INITIALIZED) {
6742 mp = mc->mc_pg[mc->mc_top];
6744 mc->mc_ki[mc->mc_top] = 0;
6745 return MDB_NOTFOUND;
6747 if (mp->mp_flags & P_LEAF2) {
6748 nodekey.mv_size = mc->mc_db->md_pad;
6749 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
6751 leaf = NODEPTR(mp, 0);
6752 MDB_GET_KEY2(leaf, nodekey);
6754 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6756 /* Probably happens rarely, but first node on the page
6757 * was the one we wanted.
6759 mc->mc_ki[mc->mc_top] = 0;
6766 unsigned int nkeys = NUMKEYS(mp);
6768 if (mp->mp_flags & P_LEAF2) {
6769 nodekey.mv_data = LEAF2KEY(mp,
6770 nkeys-1, nodekey.mv_size);
6772 leaf = NODEPTR(mp, nkeys-1);
6773 MDB_GET_KEY2(leaf, nodekey);
6775 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6777 /* last node was the one we wanted */
6778 mc->mc_ki[mc->mc_top] = nkeys-1;
6784 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6785 /* This is definitely the right page, skip search_page */
6786 if (mp->mp_flags & P_LEAF2) {
6787 nodekey.mv_data = LEAF2KEY(mp,
6788 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6790 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6791 MDB_GET_KEY2(leaf, nodekey);
6793 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6795 /* current node was the one we wanted */
6802 mc->mc_flags &= ~C_EOF;
6806 /* If any parents have right-sibs, search.
6807 * Otherwise, there's nothing further.
6809 for (i=0; i<mc->mc_top; i++)
6811 NUMKEYS(mc->mc_pg[i])-1)
6813 if (i == mc->mc_top) {
6814 /* There are no other pages */
6815 mc->mc_ki[mc->mc_top] = nkeys;
6816 return MDB_NOTFOUND;
6820 /* There are no other pages */
6821 mc->mc_ki[mc->mc_top] = 0;
6822 if (op == MDB_SET_RANGE && !exactp) {
6826 return MDB_NOTFOUND;
6832 rc = mdb_page_search(mc, key, 0);
6833 if (rc != MDB_SUCCESS)
6836 mp = mc->mc_pg[mc->mc_top];
6837 mdb_cassert(mc, IS_LEAF(mp));
6840 leaf = mdb_node_search(mc, key, exactp);
6841 if (exactp != NULL && !*exactp) {
6842 /* MDB_SET specified and not an exact match. */
6843 return MDB_NOTFOUND;
6847 DPUTS("===> inexact leaf not found, goto sibling");
6848 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6849 mc->mc_flags |= C_EOF;
6850 return rc; /* no entries matched */
6852 mp = mc->mc_pg[mc->mc_top];
6853 mdb_cassert(mc, IS_LEAF(mp));
6854 leaf = NODEPTR(mp, 0);
6858 mc->mc_flags |= C_INITIALIZED;
6859 mc->mc_flags &= ~C_EOF;
6862 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6863 key->mv_size = mc->mc_db->md_pad;
6864 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6869 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6870 mdb_xcursor_init1(mc, leaf);
6873 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6874 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6875 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6878 if (op == MDB_GET_BOTH) {
6884 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6885 if (rc != MDB_SUCCESS)
6888 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6891 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6893 dcmp = mc->mc_dbx->md_dcmp;
6894 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
6895 dcmp = mdb_cmp_clong;
6896 rc = dcmp(data, &olddata);
6898 if (op == MDB_GET_BOTH || rc > 0)
6899 return MDB_NOTFOUND;
6906 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6907 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6912 /* The key already matches in all other cases */
6913 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6914 MDB_GET_KEY(leaf, key);
6915 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6920 /** Move the cursor to the first item in the database. */
6922 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6927 if (mc->mc_xcursor) {
6928 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6929 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6932 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6933 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6934 if (rc != MDB_SUCCESS)
6937 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6939 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6940 mc->mc_flags |= C_INITIALIZED;
6941 mc->mc_flags &= ~C_EOF;
6943 mc->mc_ki[mc->mc_top] = 0;
6945 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6946 key->mv_size = mc->mc_db->md_pad;
6947 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6952 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6953 mdb_xcursor_init1(mc, leaf);
6954 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6958 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6962 MDB_GET_KEY(leaf, key);
6966 /** Move the cursor to the last item in the database. */
6968 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6973 if (mc->mc_xcursor) {
6974 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6975 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6978 if (!(mc->mc_flags & C_EOF)) {
6980 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6981 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6982 if (rc != MDB_SUCCESS)
6985 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6988 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6989 mc->mc_flags |= C_INITIALIZED|C_EOF;
6990 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6992 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6993 key->mv_size = mc->mc_db->md_pad;
6994 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6999 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7000 mdb_xcursor_init1(mc, leaf);
7001 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
7005 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
7010 MDB_GET_KEY(leaf, key);
7015 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
7020 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
7025 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7029 case MDB_GET_CURRENT:
7030 if (!(mc->mc_flags & C_INITIALIZED)) {
7033 MDB_page *mp = mc->mc_pg[mc->mc_top];
7034 int nkeys = NUMKEYS(mp);
7035 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
7036 mc->mc_ki[mc->mc_top] = nkeys;
7042 key->mv_size = mc->mc_db->md_pad;
7043 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
7045 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7046 MDB_GET_KEY(leaf, key);
7048 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7049 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
7051 rc = mdb_node_read(mc, leaf, data);
7058 case MDB_GET_BOTH_RANGE:
7063 if (mc->mc_xcursor == NULL) {
7064 rc = MDB_INCOMPATIBLE;
7074 rc = mdb_cursor_set(mc, key, data, op,
7075 op == MDB_SET_RANGE ? NULL : &exact);
7078 case MDB_GET_MULTIPLE:
7079 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
7083 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7084 rc = MDB_INCOMPATIBLE;
7088 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
7089 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
7092 case MDB_NEXT_MULTIPLE:
7097 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7098 rc = MDB_INCOMPATIBLE;
7101 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
7102 if (rc == MDB_SUCCESS) {
7103 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
7106 mx = &mc->mc_xcursor->mx_cursor;
7107 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
7109 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
7110 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
7116 case MDB_PREV_MULTIPLE:
7121 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7122 rc = MDB_INCOMPATIBLE;
7125 if (!(mc->mc_flags & C_INITIALIZED))
7126 rc = mdb_cursor_last(mc, key, data);
7129 if (rc == MDB_SUCCESS) {
7130 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
7131 if (mx->mc_flags & C_INITIALIZED) {
7132 rc = mdb_cursor_sibling(mx, 0);
7133 if (rc == MDB_SUCCESS)
7142 case MDB_NEXT_NODUP:
7143 rc = mdb_cursor_next(mc, key, data, op);
7147 case MDB_PREV_NODUP:
7148 rc = mdb_cursor_prev(mc, key, data, op);
7151 rc = mdb_cursor_first(mc, key, data);
7154 mfunc = mdb_cursor_first;
7156 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
7160 if (mc->mc_xcursor == NULL) {
7161 rc = MDB_INCOMPATIBLE;
7165 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7166 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7167 MDB_GET_KEY(leaf, key);
7168 rc = mdb_node_read(mc, leaf, data);
7172 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7176 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
7179 rc = mdb_cursor_last(mc, key, data);
7182 mfunc = mdb_cursor_last;
7185 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
7190 if (mc->mc_flags & C_DEL)
7191 mc->mc_flags ^= C_DEL;
7196 /** Touch all the pages in the cursor stack. Set mc_top.
7197 * Makes sure all the pages are writable, before attempting a write operation.
7198 * @param[in] mc The cursor to operate on.
7201 mdb_cursor_touch(MDB_cursor *mc)
7203 int rc = MDB_SUCCESS;
7205 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
7208 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
7210 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
7211 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
7214 *mc->mc_dbflag |= DB_DIRTY;
7219 rc = mdb_page_touch(mc);
7220 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
7221 mc->mc_top = mc->mc_snum-1;
7226 /** Do not spill pages to disk if txn is getting full, may fail instead */
7227 #define MDB_NOSPILL 0x8000
7230 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
7234 MDB_node *leaf = NULL;
7235 MDB_page *fp, *mp, *sub_root = NULL;
7237 MDB_val xdata, *rdata, dkey, olddata;
7239 int do_sub = 0, insert_key, insert_data;
7240 unsigned int mcount = 0, dcount = 0, nospill;
7243 unsigned int nflags;
7246 if (mc == NULL || key == NULL)
7249 env = mc->mc_txn->mt_env;
7251 /* Check this first so counter will always be zero on any
7254 if (flags & MDB_MULTIPLE) {
7255 dcount = data[1].mv_size;
7256 data[1].mv_size = 0;
7257 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
7258 return MDB_INCOMPATIBLE;
7261 nospill = flags & MDB_NOSPILL;
7262 flags &= ~MDB_NOSPILL;
7264 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7265 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7267 if (key->mv_size-1 >= ENV_MAXKEY(env))
7268 return MDB_BAD_VALSIZE;
7270 #if SIZE_MAX > MAXDATASIZE
7271 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
7272 return MDB_BAD_VALSIZE;
7274 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
7275 return MDB_BAD_VALSIZE;
7278 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
7279 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
7283 if (flags == MDB_CURRENT) {
7284 if (!(mc->mc_flags & C_INITIALIZED))
7287 } else if (mc->mc_db->md_root == P_INVALID) {
7288 /* new database, cursor has nothing to point to */
7291 mc->mc_flags &= ~C_INITIALIZED;
7296 if (flags & MDB_APPEND) {
7298 rc = mdb_cursor_last(mc, &k2, &d2);
7300 rc = mc->mc_dbx->md_cmp(key, &k2);
7303 mc->mc_ki[mc->mc_top]++;
7305 /* new key is <= last key */
7310 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
7312 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
7313 DPRINTF(("duplicate key [%s]", DKEY(key)));
7315 return MDB_KEYEXIST;
7317 if (rc && rc != MDB_NOTFOUND)
7321 if (mc->mc_flags & C_DEL)
7322 mc->mc_flags ^= C_DEL;
7324 /* Cursor is positioned, check for room in the dirty list */
7326 if (flags & MDB_MULTIPLE) {
7328 xdata.mv_size = data->mv_size * dcount;
7332 if ((rc2 = mdb_page_spill(mc, key, rdata)))
7336 if (rc == MDB_NO_ROOT) {
7338 /* new database, write a root leaf page */
7339 DPUTS("allocating new root leaf page");
7340 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
7343 mdb_cursor_push(mc, np);
7344 mc->mc_db->md_root = np->mp_pgno;
7345 mc->mc_db->md_depth++;
7346 *mc->mc_dbflag |= DB_DIRTY;
7347 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
7349 np->mp_flags |= P_LEAF2;
7350 mc->mc_flags |= C_INITIALIZED;
7352 /* make sure all cursor pages are writable */
7353 rc2 = mdb_cursor_touch(mc);
7358 insert_key = insert_data = rc;
7360 /* The key does not exist */
7361 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
7362 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
7363 LEAFSIZE(key, data) > env->me_nodemax)
7365 /* Too big for a node, insert in sub-DB. Set up an empty
7366 * "old sub-page" for prep_subDB to expand to a full page.
7368 fp_flags = P_LEAF|P_DIRTY;
7370 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
7371 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
7372 olddata.mv_size = PAGEHDRSZ;
7376 /* there's only a key anyway, so this is a no-op */
7377 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
7379 unsigned int ksize = mc->mc_db->md_pad;
7380 if (key->mv_size != ksize)
7381 return MDB_BAD_VALSIZE;
7382 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
7383 memcpy(ptr, key->mv_data, ksize);
7385 /* if overwriting slot 0 of leaf, need to
7386 * update branch key if there is a parent page
7388 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7389 unsigned short dtop = 1;
7391 /* slot 0 is always an empty key, find real slot */
7392 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7396 if (mc->mc_ki[mc->mc_top])
7397 rc2 = mdb_update_key(mc, key);
7408 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7409 olddata.mv_size = NODEDSZ(leaf);
7410 olddata.mv_data = NODEDATA(leaf);
7413 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
7414 /* Prepare (sub-)page/sub-DB to accept the new item,
7415 * if needed. fp: old sub-page or a header faking
7416 * it. mp: new (sub-)page. offset: growth in page
7417 * size. xdata: node data with new page or DB.
7419 unsigned i, offset = 0;
7420 mp = fp = xdata.mv_data = env->me_pbuf;
7421 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
7423 /* Was a single item before, must convert now */
7424 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7426 /* Just overwrite the current item */
7427 if (flags == MDB_CURRENT)
7429 dcmp = mc->mc_dbx->md_dcmp;
7430 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
7431 dcmp = mdb_cmp_clong;
7432 /* does data match? */
7433 if (!dcmp(data, &olddata)) {
7434 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
7435 return MDB_KEYEXIST;
7440 /* Back up original data item */
7441 dkey.mv_size = olddata.mv_size;
7442 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
7444 /* Make sub-page header for the dup items, with dummy body */
7445 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
7446 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
7447 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
7448 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7449 fp->mp_flags |= P_LEAF2;
7450 fp->mp_pad = data->mv_size;
7451 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
7453 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
7454 (dkey.mv_size & 1) + (data->mv_size & 1);
7456 fp->mp_upper = xdata.mv_size - PAGEBASE;
7457 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
7458 } else if (leaf->mn_flags & F_SUBDATA) {
7459 /* Data is on sub-DB, just store it */
7460 flags |= F_DUPDATA|F_SUBDATA;
7463 /* Data is on sub-page */
7464 fp = olddata.mv_data;
7467 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7468 offset = EVEN(NODESIZE + sizeof(indx_t) +
7472 offset = fp->mp_pad;
7473 if (SIZELEFT(fp) < offset) {
7474 offset *= 4; /* space for 4 more */
7477 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
7479 fp->mp_flags |= P_DIRTY;
7480 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
7481 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
7485 xdata.mv_size = olddata.mv_size + offset;
7488 fp_flags = fp->mp_flags;
7489 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
7490 /* Too big for a sub-page, convert to sub-DB */
7491 fp_flags &= ~P_SUBP;
7493 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7494 fp_flags |= P_LEAF2;
7495 dummy.md_pad = fp->mp_pad;
7496 dummy.md_flags = MDB_DUPFIXED;
7497 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7498 dummy.md_flags |= MDB_INTEGERKEY;
7504 dummy.md_branch_pages = 0;
7505 dummy.md_leaf_pages = 1;
7506 dummy.md_overflow_pages = 0;
7507 dummy.md_entries = NUMKEYS(fp);
7508 xdata.mv_size = sizeof(MDB_db);
7509 xdata.mv_data = &dummy;
7510 if ((rc = mdb_page_alloc(mc, 1, &mp)))
7512 offset = env->me_psize - olddata.mv_size;
7513 flags |= F_DUPDATA|F_SUBDATA;
7514 dummy.md_root = mp->mp_pgno;
7518 mp->mp_flags = fp_flags | P_DIRTY;
7519 mp->mp_pad = fp->mp_pad;
7520 mp->mp_lower = fp->mp_lower;
7521 mp->mp_upper = fp->mp_upper + offset;
7522 if (fp_flags & P_LEAF2) {
7523 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
7525 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
7526 olddata.mv_size - fp->mp_upper - PAGEBASE);
7527 for (i=0; i<NUMKEYS(fp); i++)
7528 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
7536 mdb_node_del(mc, 0);
7540 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
7541 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
7542 return MDB_INCOMPATIBLE;
7543 /* overflow page overwrites need special handling */
7544 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7547 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
7549 memcpy(&pg, olddata.mv_data, sizeof(pg));
7550 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
7552 ovpages = omp->mp_pages;
7554 /* Is the ov page large enough? */
7555 if (ovpages >= dpages) {
7556 if (!(omp->mp_flags & P_DIRTY) &&
7557 (level || (env->me_flags & MDB_WRITEMAP)))
7559 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
7562 level = 0; /* dirty in this txn or clean */
7565 if (omp->mp_flags & P_DIRTY) {
7566 /* yes, overwrite it. Note in this case we don't
7567 * bother to try shrinking the page if the new data
7568 * is smaller than the overflow threshold.
7571 /* It is writable only in a parent txn */
7572 size_t sz = (size_t) env->me_psize * ovpages, off;
7573 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
7579 /* Note - this page is already counted in parent's dirty_room */
7580 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
7581 mdb_cassert(mc, rc2 == 0);
7582 /* Currently we make the page look as with put() in the
7583 * parent txn, in case the user peeks at MDB_RESERVEd
7584 * or unused parts. Some users treat ovpages specially.
7586 if (!(flags & MDB_RESERVE)) {
7587 /* Skip the part where LMDB will put *data.
7588 * Copy end of page, adjusting alignment so
7589 * compiler may copy words instead of bytes.
7591 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
7592 memcpy((size_t *)((char *)np + off),
7593 (size_t *)((char *)omp + off), sz - off);
7596 memcpy(np, omp, sz); /* Copy beginning of page */
7599 SETDSZ(leaf, data->mv_size);
7600 if (F_ISSET(flags, MDB_RESERVE))
7601 data->mv_data = METADATA(omp);
7603 memcpy(METADATA(omp), data->mv_data, data->mv_size);
7607 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
7609 } else if (data->mv_size == olddata.mv_size) {
7610 /* same size, just replace it. Note that we could
7611 * also reuse this node if the new data is smaller,
7612 * but instead we opt to shrink the node in that case.
7614 if (F_ISSET(flags, MDB_RESERVE))
7615 data->mv_data = olddata.mv_data;
7616 else if (!(mc->mc_flags & C_SUB))
7617 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
7619 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
7624 mdb_node_del(mc, 0);
7630 nflags = flags & NODE_ADD_FLAGS;
7631 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
7632 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
7633 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
7634 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
7636 nflags |= MDB_SPLIT_REPLACE;
7637 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
7639 /* There is room already in this leaf page. */
7640 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
7642 /* Adjust other cursors pointing to mp */
7643 MDB_cursor *m2, *m3;
7644 MDB_dbi dbi = mc->mc_dbi;
7645 unsigned i = mc->mc_top;
7646 MDB_page *mp = mc->mc_pg[i];
7648 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7649 if (mc->mc_flags & C_SUB)
7650 m3 = &m2->mc_xcursor->mx_cursor;
7653 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
7654 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
7657 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7658 MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
7659 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7660 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7666 if (rc == MDB_SUCCESS) {
7667 /* Now store the actual data in the child DB. Note that we're
7668 * storing the user data in the keys field, so there are strict
7669 * size limits on dupdata. The actual data fields of the child
7670 * DB are all zero size.
7673 int xflags, new_dupdata;
7678 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7679 if (flags & MDB_CURRENT) {
7680 xflags = MDB_CURRENT|MDB_NOSPILL;
7682 mdb_xcursor_init1(mc, leaf);
7683 xflags = (flags & MDB_NODUPDATA) ?
7684 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
7687 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
7688 new_dupdata = (int)dkey.mv_size;
7689 /* converted, write the original data first */
7691 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
7694 /* we've done our job */
7697 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
7698 /* Adjust other cursors pointing to mp */
7700 MDB_xcursor *mx = mc->mc_xcursor;
7701 unsigned i = mc->mc_top;
7702 MDB_page *mp = mc->mc_pg[i];
7703 int nkeys = NUMKEYS(mp);
7705 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7706 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7707 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7708 if (m2->mc_pg[i] == mp) {
7709 if (m2->mc_ki[i] == mc->mc_ki[i]) {
7710 mdb_xcursor_init2(m2, mx, new_dupdata);
7711 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
7712 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
7713 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7714 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7719 ecount = mc->mc_xcursor->mx_db.md_entries;
7720 if (flags & MDB_APPENDDUP)
7721 xflags |= MDB_APPEND;
7722 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
7723 if (flags & F_SUBDATA) {
7724 void *db = NODEDATA(leaf);
7725 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7727 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
7729 /* Increment count unless we just replaced an existing item. */
7731 mc->mc_db->md_entries++;
7733 /* Invalidate txn if we created an empty sub-DB */
7736 /* If we succeeded and the key didn't exist before,
7737 * make sure the cursor is marked valid.
7739 mc->mc_flags |= C_INITIALIZED;
7741 if (flags & MDB_MULTIPLE) {
7744 /* let caller know how many succeeded, if any */
7745 data[1].mv_size = mcount;
7746 if (mcount < dcount) {
7747 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
7748 insert_key = insert_data = 0;
7755 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
7758 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7763 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7769 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7770 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7772 if (!(mc->mc_flags & C_INITIALIZED))
7775 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7776 return MDB_NOTFOUND;
7778 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7781 rc = mdb_cursor_touch(mc);
7785 mp = mc->mc_pg[mc->mc_top];
7788 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7790 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7791 if (flags & MDB_NODUPDATA) {
7792 /* mdb_cursor_del0() will subtract the final entry */
7793 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7794 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7796 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7797 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7799 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7802 /* If sub-DB still has entries, we're done */
7803 if (mc->mc_xcursor->mx_db.md_entries) {
7804 if (leaf->mn_flags & F_SUBDATA) {
7805 /* update subDB info */
7806 void *db = NODEDATA(leaf);
7807 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7810 /* shrink fake page */
7811 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7812 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7813 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7814 /* fix other sub-DB cursors pointed at fake pages on this page */
7815 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7816 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7817 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7818 if (m2->mc_pg[mc->mc_top] == mp) {
7819 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
7820 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7822 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7823 if (!(n2->mn_flags & F_SUBDATA))
7824 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7829 mc->mc_db->md_entries--;
7832 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7834 /* otherwise fall thru and delete the sub-DB */
7837 if (leaf->mn_flags & F_SUBDATA) {
7838 /* add all the child DB's pages to the free list */
7839 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7844 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7845 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7846 rc = MDB_INCOMPATIBLE;
7850 /* add overflow pages to free list */
7851 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7855 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7856 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7857 (rc = mdb_ovpage_free(mc, omp)))
7862 return mdb_cursor_del0(mc);
7865 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7869 /** Allocate and initialize new pages for a database.
7870 * @param[in] mc a cursor on the database being added to.
7871 * @param[in] flags flags defining what type of page is being allocated.
7872 * @param[in] num the number of pages to allocate. This is usually 1,
7873 * unless allocating overflow pages for a large record.
7874 * @param[out] mp Address of a page, or NULL on failure.
7875 * @return 0 on success, non-zero on failure.
7878 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7883 if ((rc = mdb_page_alloc(mc, num, &np)))
7885 DPRINTF(("allocated new mpage %"Yu", page size %u",
7886 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7887 np->mp_flags = flags | P_DIRTY;
7888 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7889 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7892 mc->mc_db->md_branch_pages++;
7893 else if (IS_LEAF(np))
7894 mc->mc_db->md_leaf_pages++;
7895 else if (IS_OVERFLOW(np)) {
7896 mc->mc_db->md_overflow_pages += num;
7904 /** Calculate the size of a leaf node.
7905 * The size depends on the environment's page size; if a data item
7906 * is too large it will be put onto an overflow page and the node
7907 * size will only include the key and not the data. Sizes are always
7908 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7909 * of the #MDB_node headers.
7910 * @param[in] env The environment handle.
7911 * @param[in] key The key for the node.
7912 * @param[in] data The data for the node.
7913 * @return The number of bytes needed to store the node.
7916 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7920 sz = LEAFSIZE(key, data);
7921 if (sz > env->me_nodemax) {
7922 /* put on overflow page */
7923 sz -= data->mv_size - sizeof(pgno_t);
7926 return EVEN(sz + sizeof(indx_t));
7929 /** Calculate the size of a branch node.
7930 * The size should depend on the environment's page size but since
7931 * we currently don't support spilling large keys onto overflow
7932 * pages, it's simply the size of the #MDB_node header plus the
7933 * size of the key. Sizes are always rounded up to an even number
7934 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7935 * @param[in] env The environment handle.
7936 * @param[in] key The key for the node.
7937 * @return The number of bytes needed to store the node.
7940 mdb_branch_size(MDB_env *env, MDB_val *key)
7945 if (sz > env->me_nodemax) {
7946 /* put on overflow page */
7947 /* not implemented */
7948 /* sz -= key->size - sizeof(pgno_t); */
7951 return sz + sizeof(indx_t);
7954 /** Add a node to the page pointed to by the cursor.
7955 * @param[in] mc The cursor for this operation.
7956 * @param[in] indx The index on the page where the new node should be added.
7957 * @param[in] key The key for the new node.
7958 * @param[in] data The data for the new node, if any.
7959 * @param[in] pgno The page number, if adding a branch node.
7960 * @param[in] flags Flags for the node.
7961 * @return 0 on success, non-zero on failure. Possible errors are:
7963 * <li>ENOMEM - failed to allocate overflow pages for the node.
7964 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7965 * should never happen since all callers already calculate the
7966 * page's free space before calling this function.
7970 mdb_node_add(MDB_cursor *mc, indx_t indx,
7971 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7974 size_t node_size = NODESIZE;
7978 MDB_page *mp = mc->mc_pg[mc->mc_top];
7979 MDB_page *ofp = NULL; /* overflow page */
7983 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7985 DPRINTF(("add to %s %spage %"Yu" index %i, data size %"Z"u key size %"Z"u [%s]",
7986 IS_LEAF(mp) ? "leaf" : "branch",
7987 IS_SUBP(mp) ? "sub-" : "",
7988 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7989 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7992 /* Move higher keys up one slot. */
7993 int ksize = mc->mc_db->md_pad, dif;
7994 char *ptr = LEAF2KEY(mp, indx, ksize);
7995 dif = NUMKEYS(mp) - indx;
7997 memmove(ptr+ksize, ptr, dif*ksize);
7998 /* insert new key */
7999 memcpy(ptr, key->mv_data, ksize);
8001 /* Just using these for counting */
8002 mp->mp_lower += sizeof(indx_t);
8003 mp->mp_upper -= ksize - sizeof(indx_t);
8007 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
8009 node_size += key->mv_size;
8011 mdb_cassert(mc, key && data);
8012 if (F_ISSET(flags, F_BIGDATA)) {
8013 /* Data already on overflow page. */
8014 node_size += sizeof(pgno_t);
8015 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
8016 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
8018 /* Put data on overflow page. */
8019 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
8020 data->mv_size, node_size+data->mv_size));
8021 node_size = EVEN(node_size + sizeof(pgno_t));
8022 if ((ssize_t)node_size > room)
8024 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
8026 DPRINTF(("allocated overflow page %"Yu, ofp->mp_pgno));
8030 node_size += data->mv_size;
8033 node_size = EVEN(node_size);
8034 if ((ssize_t)node_size > room)
8038 /* Move higher pointers up one slot. */
8039 for (i = NUMKEYS(mp); i > indx; i--)
8040 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
8042 /* Adjust free space offsets. */
8043 ofs = mp->mp_upper - node_size;
8044 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
8045 mp->mp_ptrs[indx] = ofs;
8047 mp->mp_lower += sizeof(indx_t);
8049 /* Write the node data. */
8050 node = NODEPTR(mp, indx);
8051 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
8052 node->mn_flags = flags;
8054 SETDSZ(node,data->mv_size);
8059 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8062 ndata = NODEDATA(node);
8064 if (F_ISSET(flags, F_BIGDATA))
8065 memcpy(ndata, data->mv_data, sizeof(pgno_t));
8066 else if (F_ISSET(flags, MDB_RESERVE))
8067 data->mv_data = ndata;
8069 memcpy(ndata, data->mv_data, data->mv_size);
8071 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
8072 ndata = METADATA(ofp);
8073 if (F_ISSET(flags, MDB_RESERVE))
8074 data->mv_data = ndata;
8076 memcpy(ndata, data->mv_data, data->mv_size);
8083 DPRINTF(("not enough room in page %"Yu", got %u ptrs",
8084 mdb_dbg_pgno(mp), NUMKEYS(mp)));
8085 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
8086 DPRINTF(("node size = %"Z"u", node_size));
8087 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8088 return MDB_PAGE_FULL;
8091 /** Delete the specified node from a page.
8092 * @param[in] mc Cursor pointing to the node to delete.
8093 * @param[in] ksize The size of a node. Only used if the page is
8094 * part of a #MDB_DUPFIXED database.
8097 mdb_node_del(MDB_cursor *mc, int ksize)
8099 MDB_page *mp = mc->mc_pg[mc->mc_top];
8100 indx_t indx = mc->mc_ki[mc->mc_top];
8102 indx_t i, j, numkeys, ptr;
8106 DPRINTF(("delete node %u on %s page %"Yu, indx,
8107 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
8108 numkeys = NUMKEYS(mp);
8109 mdb_cassert(mc, indx < numkeys);
8112 int x = numkeys - 1 - indx;
8113 base = LEAF2KEY(mp, indx, ksize);
8115 memmove(base, base + ksize, x * ksize);
8116 mp->mp_lower -= sizeof(indx_t);
8117 mp->mp_upper += ksize - sizeof(indx_t);
8121 node = NODEPTR(mp, indx);
8122 sz = NODESIZE + node->mn_ksize;
8124 if (F_ISSET(node->mn_flags, F_BIGDATA))
8125 sz += sizeof(pgno_t);
8127 sz += NODEDSZ(node);
8131 ptr = mp->mp_ptrs[indx];
8132 for (i = j = 0; i < numkeys; i++) {
8134 mp->mp_ptrs[j] = mp->mp_ptrs[i];
8135 if (mp->mp_ptrs[i] < ptr)
8136 mp->mp_ptrs[j] += sz;
8141 base = (char *)mp + mp->mp_upper + PAGEBASE;
8142 memmove(base + sz, base, ptr - mp->mp_upper);
8144 mp->mp_lower -= sizeof(indx_t);
8148 /** Compact the main page after deleting a node on a subpage.
8149 * @param[in] mp The main page to operate on.
8150 * @param[in] indx The index of the subpage on the main page.
8153 mdb_node_shrink(MDB_page *mp, indx_t indx)
8158 indx_t delta, nsize, len, ptr;
8161 node = NODEPTR(mp, indx);
8162 sp = (MDB_page *)NODEDATA(node);
8163 delta = SIZELEFT(sp);
8164 nsize = NODEDSZ(node) - delta;
8166 /* Prepare to shift upward, set len = length(subpage part to shift) */
8170 return; /* do not make the node uneven-sized */
8172 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
8173 for (i = NUMKEYS(sp); --i >= 0; )
8174 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
8177 sp->mp_upper = sp->mp_lower;
8178 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
8179 SETDSZ(node, nsize);
8181 /* Shift <lower nodes...initial part of subpage> upward */
8182 base = (char *)mp + mp->mp_upper + PAGEBASE;
8183 memmove(base + delta, base, (char *)sp + len - base);
8185 ptr = mp->mp_ptrs[indx];
8186 for (i = NUMKEYS(mp); --i >= 0; ) {
8187 if (mp->mp_ptrs[i] <= ptr)
8188 mp->mp_ptrs[i] += delta;
8190 mp->mp_upper += delta;
8193 /** Initial setup of a sorted-dups cursor.
8194 * Sorted duplicates are implemented as a sub-database for the given key.
8195 * The duplicate data items are actually keys of the sub-database.
8196 * Operations on the duplicate data items are performed using a sub-cursor
8197 * initialized when the sub-database is first accessed. This function does
8198 * the preliminary setup of the sub-cursor, filling in the fields that
8199 * depend only on the parent DB.
8200 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8203 mdb_xcursor_init0(MDB_cursor *mc)
8205 MDB_xcursor *mx = mc->mc_xcursor;
8207 mx->mx_cursor.mc_xcursor = NULL;
8208 mx->mx_cursor.mc_txn = mc->mc_txn;
8209 mx->mx_cursor.mc_db = &mx->mx_db;
8210 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
8211 mx->mx_cursor.mc_dbi = mc->mc_dbi;
8212 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
8213 mx->mx_cursor.mc_snum = 0;
8214 mx->mx_cursor.mc_top = 0;
8215 MC_SET_OVPG(&mx->mx_cursor, NULL);
8216 mx->mx_cursor.mc_flags = C_SUB | (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP));
8217 mx->mx_dbx.md_name.mv_size = 0;
8218 mx->mx_dbx.md_name.mv_data = NULL;
8219 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
8220 mx->mx_dbx.md_dcmp = NULL;
8221 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
8224 /** Final setup of a sorted-dups cursor.
8225 * Sets up the fields that depend on the data from the main cursor.
8226 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8227 * @param[in] node The data containing the #MDB_db record for the
8228 * sorted-dup database.
8231 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
8233 MDB_xcursor *mx = mc->mc_xcursor;
8235 mx->mx_cursor.mc_flags &= C_SUB|C_ORIG_RDONLY|C_WRITEMAP;
8236 if (node->mn_flags & F_SUBDATA) {
8237 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
8238 mx->mx_cursor.mc_pg[0] = 0;
8239 mx->mx_cursor.mc_snum = 0;
8240 mx->mx_cursor.mc_top = 0;
8242 MDB_page *fp = NODEDATA(node);
8243 mx->mx_db.md_pad = 0;
8244 mx->mx_db.md_flags = 0;
8245 mx->mx_db.md_depth = 1;
8246 mx->mx_db.md_branch_pages = 0;
8247 mx->mx_db.md_leaf_pages = 1;
8248 mx->mx_db.md_overflow_pages = 0;
8249 mx->mx_db.md_entries = NUMKEYS(fp);
8250 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
8251 mx->mx_cursor.mc_snum = 1;
8252 mx->mx_cursor.mc_top = 0;
8253 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8254 mx->mx_cursor.mc_pg[0] = fp;
8255 mx->mx_cursor.mc_ki[0] = 0;
8256 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
8257 mx->mx_db.md_flags = MDB_DUPFIXED;
8258 mx->mx_db.md_pad = fp->mp_pad;
8259 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
8260 mx->mx_db.md_flags |= MDB_INTEGERKEY;
8263 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8264 mx->mx_db.md_root));
8265 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8266 if (NEED_CMP_CLONG(mx->mx_dbx.md_cmp, mx->mx_db.md_pad))
8267 mx->mx_dbx.md_cmp = mdb_cmp_clong;
8271 /** Fixup a sorted-dups cursor due to underlying update.
8272 * Sets up some fields that depend on the data from the main cursor.
8273 * Almost the same as init1, but skips initialization steps if the
8274 * xcursor had already been used.
8275 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
8276 * @param[in] src_mx The xcursor of an up-to-date cursor.
8277 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
8280 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
8282 MDB_xcursor *mx = mc->mc_xcursor;
8285 mx->mx_cursor.mc_snum = 1;
8286 mx->mx_cursor.mc_top = 0;
8287 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8288 mx->mx_cursor.mc_ki[0] = 0;
8289 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8290 #if UINT_MAX < MDB_SIZE_MAX /* matches mdb_xcursor_init1:NEED_CMP_CLONG() */
8291 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
8293 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
8296 mx->mx_db = src_mx->mx_db;
8297 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
8298 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8299 mx->mx_db.md_root));
8302 /** Initialize a cursor for a given transaction and database. */
8304 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
8307 mc->mc_backup = NULL;
8310 mc->mc_db = &txn->mt_dbs[dbi];
8311 mc->mc_dbx = &txn->mt_dbxs[dbi];
8312 mc->mc_dbflag = &txn->mt_dbflags[dbi];
8317 MC_SET_OVPG(mc, NULL);
8318 mc->mc_flags = txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
8319 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
8320 mdb_tassert(txn, mx != NULL);
8321 mc->mc_xcursor = mx;
8322 mdb_xcursor_init0(mc);
8324 mc->mc_xcursor = NULL;
8326 if (*mc->mc_dbflag & DB_STALE) {
8327 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
8332 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
8335 size_t size = sizeof(MDB_cursor);
8337 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
8340 if (txn->mt_flags & MDB_TXN_BLOCKED)
8343 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8346 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
8347 size += sizeof(MDB_xcursor);
8349 if ((mc = malloc(size)) != NULL) {
8350 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
8351 if (txn->mt_cursors) {
8352 mc->mc_next = txn->mt_cursors[dbi];
8353 txn->mt_cursors[dbi] = mc;
8354 mc->mc_flags |= C_UNTRACK;
8366 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
8368 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
8371 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
8374 if (txn->mt_flags & MDB_TXN_BLOCKED)
8377 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
8381 /* Return the count of duplicate data items for the current key */
8383 mdb_cursor_count(MDB_cursor *mc, mdb_size_t *countp)
8387 if (mc == NULL || countp == NULL)
8390 if (mc->mc_xcursor == NULL)
8391 return MDB_INCOMPATIBLE;
8393 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
8396 if (!(mc->mc_flags & C_INITIALIZED))
8399 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
8400 return MDB_NOTFOUND;
8402 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8403 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
8406 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
8409 *countp = mc->mc_xcursor->mx_db.md_entries;
8415 mdb_cursor_close(MDB_cursor *mc)
8418 MDB_CURSOR_UNREF(mc, 0);
8420 if (mc && !mc->mc_backup) {
8421 /* remove from txn, if tracked */
8422 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
8423 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
8424 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
8426 *prev = mc->mc_next;
8433 mdb_cursor_txn(MDB_cursor *mc)
8435 if (!mc) return NULL;
8440 mdb_cursor_dbi(MDB_cursor *mc)
8445 /** Replace the key for a branch node with a new key.
8446 * @param[in] mc Cursor pointing to the node to operate on.
8447 * @param[in] key The new key to use.
8448 * @return 0 on success, non-zero on failure.
8451 mdb_update_key(MDB_cursor *mc, MDB_val *key)
8457 int delta, ksize, oksize;
8458 indx_t ptr, i, numkeys, indx;
8461 indx = mc->mc_ki[mc->mc_top];
8462 mp = mc->mc_pg[mc->mc_top];
8463 node = NODEPTR(mp, indx);
8464 ptr = mp->mp_ptrs[indx];
8468 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
8469 k2.mv_data = NODEKEY(node);
8470 k2.mv_size = node->mn_ksize;
8471 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Yu,
8473 mdb_dkey(&k2, kbuf2),
8479 /* Sizes must be 2-byte aligned. */
8480 ksize = EVEN(key->mv_size);
8481 oksize = EVEN(node->mn_ksize);
8482 delta = ksize - oksize;
8484 /* Shift node contents if EVEN(key length) changed. */
8486 if (delta > 0 && SIZELEFT(mp) < delta) {
8488 /* not enough space left, do a delete and split */
8489 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
8490 pgno = NODEPGNO(node);
8491 mdb_node_del(mc, 0);
8492 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
8495 numkeys = NUMKEYS(mp);
8496 for (i = 0; i < numkeys; i++) {
8497 if (mp->mp_ptrs[i] <= ptr)
8498 mp->mp_ptrs[i] -= delta;
8501 base = (char *)mp + mp->mp_upper + PAGEBASE;
8502 len = ptr - mp->mp_upper + NODESIZE;
8503 memmove(base - delta, base, len);
8504 mp->mp_upper -= delta;
8506 node = NODEPTR(mp, indx);
8509 /* But even if no shift was needed, update ksize */
8510 if (node->mn_ksize != key->mv_size)
8511 node->mn_ksize = key->mv_size;
8514 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8520 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
8522 /** Perform \b act while tracking temporary cursor \b mn */
8523 #define WITH_CURSOR_TRACKING(mn, act) do { \
8524 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
8525 if ((mn).mc_flags & C_SUB) { \
8526 dummy.mc_flags = C_INITIALIZED; \
8527 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
8532 tracked->mc_next = *tp; \
8535 *tp = tracked->mc_next; \
8538 /** Move a node from csrc to cdst.
8541 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
8548 unsigned short flags;
8552 /* Mark src and dst as dirty. */
8553 if ((rc = mdb_page_touch(csrc)) ||
8554 (rc = mdb_page_touch(cdst)))
8557 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8558 key.mv_size = csrc->mc_db->md_pad;
8559 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
8561 data.mv_data = NULL;
8565 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
8566 mdb_cassert(csrc, !((size_t)srcnode & 1));
8567 srcpg = NODEPGNO(srcnode);
8568 flags = srcnode->mn_flags;
8569 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8570 unsigned int snum = csrc->mc_snum;
8572 /* must find the lowest key below src */
8573 rc = mdb_page_search_lowest(csrc);
8576 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8577 key.mv_size = csrc->mc_db->md_pad;
8578 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8580 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8581 key.mv_size = NODEKSZ(s2);
8582 key.mv_data = NODEKEY(s2);
8584 csrc->mc_snum = snum--;
8585 csrc->mc_top = snum;
8587 key.mv_size = NODEKSZ(srcnode);
8588 key.mv_data = NODEKEY(srcnode);
8590 data.mv_size = NODEDSZ(srcnode);
8591 data.mv_data = NODEDATA(srcnode);
8593 mn.mc_xcursor = NULL;
8594 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
8595 unsigned int snum = cdst->mc_snum;
8598 /* must find the lowest key below dst */
8599 mdb_cursor_copy(cdst, &mn);
8600 rc = mdb_page_search_lowest(&mn);
8603 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8604 bkey.mv_size = mn.mc_db->md_pad;
8605 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
8607 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8608 bkey.mv_size = NODEKSZ(s2);
8609 bkey.mv_data = NODEKEY(s2);
8611 mn.mc_snum = snum--;
8614 rc = mdb_update_key(&mn, &bkey);
8619 DPRINTF(("moving %s node %u [%s] on page %"Yu" to node %u on page %"Yu,
8620 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
8621 csrc->mc_ki[csrc->mc_top],
8623 csrc->mc_pg[csrc->mc_top]->mp_pgno,
8624 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
8626 /* Add the node to the destination page.
8628 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
8629 if (rc != MDB_SUCCESS)
8632 /* Delete the node from the source page.
8634 mdb_node_del(csrc, key.mv_size);
8637 /* Adjust other cursors pointing to mp */
8638 MDB_cursor *m2, *m3;
8639 MDB_dbi dbi = csrc->mc_dbi;
8640 MDB_page *mpd, *mps;
8642 mps = csrc->mc_pg[csrc->mc_top];
8643 /* If we're adding on the left, bump others up */
8645 mpd = cdst->mc_pg[csrc->mc_top];
8646 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8647 if (csrc->mc_flags & C_SUB)
8648 m3 = &m2->mc_xcursor->mx_cursor;
8651 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8654 m3->mc_pg[csrc->mc_top] == mpd &&
8655 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
8656 m3->mc_ki[csrc->mc_top]++;
8659 m3->mc_pg[csrc->mc_top] == mps &&
8660 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
8661 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8662 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8663 m3->mc_ki[csrc->mc_top-1]++;
8665 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8667 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8668 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8669 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8673 /* Adding on the right, bump others down */
8675 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8676 if (csrc->mc_flags & C_SUB)
8677 m3 = &m2->mc_xcursor->mx_cursor;
8680 if (m3 == csrc) continue;
8681 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8683 if (m3->mc_pg[csrc->mc_top] == mps) {
8684 if (!m3->mc_ki[csrc->mc_top]) {
8685 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8686 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8687 m3->mc_ki[csrc->mc_top-1]--;
8689 m3->mc_ki[csrc->mc_top]--;
8691 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8693 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8694 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8695 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8702 /* Update the parent separators.
8704 if (csrc->mc_ki[csrc->mc_top] == 0) {
8705 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
8706 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8707 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8709 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8710 key.mv_size = NODEKSZ(srcnode);
8711 key.mv_data = NODEKEY(srcnode);
8713 DPRINTF(("update separator for source page %"Yu" to [%s]",
8714 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
8715 mdb_cursor_copy(csrc, &mn);
8718 /* We want mdb_rebalance to find mn when doing fixups */
8719 WITH_CURSOR_TRACKING(mn,
8720 rc = mdb_update_key(&mn, &key));
8724 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8726 indx_t ix = csrc->mc_ki[csrc->mc_top];
8727 nullkey.mv_size = 0;
8728 csrc->mc_ki[csrc->mc_top] = 0;
8729 rc = mdb_update_key(csrc, &nullkey);
8730 csrc->mc_ki[csrc->mc_top] = ix;
8731 mdb_cassert(csrc, rc == MDB_SUCCESS);
8735 if (cdst->mc_ki[cdst->mc_top] == 0) {
8736 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
8737 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8738 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
8740 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
8741 key.mv_size = NODEKSZ(srcnode);
8742 key.mv_data = NODEKEY(srcnode);
8744 DPRINTF(("update separator for destination page %"Yu" to [%s]",
8745 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
8746 mdb_cursor_copy(cdst, &mn);
8749 /* We want mdb_rebalance to find mn when doing fixups */
8750 WITH_CURSOR_TRACKING(mn,
8751 rc = mdb_update_key(&mn, &key));
8755 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
8757 indx_t ix = cdst->mc_ki[cdst->mc_top];
8758 nullkey.mv_size = 0;
8759 cdst->mc_ki[cdst->mc_top] = 0;
8760 rc = mdb_update_key(cdst, &nullkey);
8761 cdst->mc_ki[cdst->mc_top] = ix;
8762 mdb_cassert(cdst, rc == MDB_SUCCESS);
8769 /** Merge one page into another.
8770 * The nodes from the page pointed to by \b csrc will
8771 * be copied to the page pointed to by \b cdst and then
8772 * the \b csrc page will be freed.
8773 * @param[in] csrc Cursor pointing to the source page.
8774 * @param[in] cdst Cursor pointing to the destination page.
8775 * @return 0 on success, non-zero on failure.
8778 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8780 MDB_page *psrc, *pdst;
8787 psrc = csrc->mc_pg[csrc->mc_top];
8788 pdst = cdst->mc_pg[cdst->mc_top];
8790 DPRINTF(("merging page %"Yu" into %"Yu, psrc->mp_pgno, pdst->mp_pgno));
8792 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8793 mdb_cassert(csrc, cdst->mc_snum > 1);
8795 /* Mark dst as dirty. */
8796 if ((rc = mdb_page_touch(cdst)))
8799 /* get dst page again now that we've touched it. */
8800 pdst = cdst->mc_pg[cdst->mc_top];
8802 /* Move all nodes from src to dst.
8804 j = nkeys = NUMKEYS(pdst);
8805 if (IS_LEAF2(psrc)) {
8806 key.mv_size = csrc->mc_db->md_pad;
8807 key.mv_data = METADATA(psrc);
8808 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8809 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8810 if (rc != MDB_SUCCESS)
8812 key.mv_data = (char *)key.mv_data + key.mv_size;
8815 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8816 srcnode = NODEPTR(psrc, i);
8817 if (i == 0 && IS_BRANCH(psrc)) {
8820 mdb_cursor_copy(csrc, &mn);
8821 mn.mc_xcursor = NULL;
8822 /* must find the lowest key below src */
8823 rc = mdb_page_search_lowest(&mn);
8826 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8827 key.mv_size = mn.mc_db->md_pad;
8828 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8830 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8831 key.mv_size = NODEKSZ(s2);
8832 key.mv_data = NODEKEY(s2);
8835 key.mv_size = srcnode->mn_ksize;
8836 key.mv_data = NODEKEY(srcnode);
8839 data.mv_size = NODEDSZ(srcnode);
8840 data.mv_data = NODEDATA(srcnode);
8841 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8842 if (rc != MDB_SUCCESS)
8847 DPRINTF(("dst page %"Yu" now has %u keys (%.1f%% filled)",
8848 pdst->mp_pgno, NUMKEYS(pdst),
8849 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8851 /* Unlink the src page from parent and add to free list.
8854 mdb_node_del(csrc, 0);
8855 if (csrc->mc_ki[csrc->mc_top] == 0) {
8857 rc = mdb_update_key(csrc, &key);
8865 psrc = csrc->mc_pg[csrc->mc_top];
8866 /* If not operating on FreeDB, allow this page to be reused
8867 * in this txn. Otherwise just add to free list.
8869 rc = mdb_page_loose(csrc, psrc);
8873 csrc->mc_db->md_leaf_pages--;
8875 csrc->mc_db->md_branch_pages--;
8877 /* Adjust other cursors pointing to mp */
8878 MDB_cursor *m2, *m3;
8879 MDB_dbi dbi = csrc->mc_dbi;
8880 unsigned int top = csrc->mc_top;
8882 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8883 if (csrc->mc_flags & C_SUB)
8884 m3 = &m2->mc_xcursor->mx_cursor;
8887 if (m3 == csrc) continue;
8888 if (m3->mc_snum < csrc->mc_snum) continue;
8889 if (m3->mc_pg[top] == psrc) {
8890 m3->mc_pg[top] = pdst;
8891 m3->mc_ki[top] += nkeys;
8892 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8893 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8894 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8897 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8899 MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
8900 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8901 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8906 unsigned int snum = cdst->mc_snum;
8907 uint16_t depth = cdst->mc_db->md_depth;
8908 mdb_cursor_pop(cdst);
8909 rc = mdb_rebalance(cdst);
8910 /* Did the tree height change? */
8911 if (depth != cdst->mc_db->md_depth)
8912 snum += cdst->mc_db->md_depth - depth;
8913 cdst->mc_snum = snum;
8914 cdst->mc_top = snum-1;
8919 /** Copy the contents of a cursor.
8920 * @param[in] csrc The cursor to copy from.
8921 * @param[out] cdst The cursor to copy to.
8924 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8928 cdst->mc_txn = csrc->mc_txn;
8929 cdst->mc_dbi = csrc->mc_dbi;
8930 cdst->mc_db = csrc->mc_db;
8931 cdst->mc_dbx = csrc->mc_dbx;
8932 cdst->mc_snum = csrc->mc_snum;
8933 cdst->mc_top = csrc->mc_top;
8934 cdst->mc_flags = csrc->mc_flags;
8935 MC_SET_OVPG(cdst, MC_OVPG(csrc));
8937 for (i=0; i<csrc->mc_snum; i++) {
8938 cdst->mc_pg[i] = csrc->mc_pg[i];
8939 cdst->mc_ki[i] = csrc->mc_ki[i];
8943 /** Rebalance the tree after a delete operation.
8944 * @param[in] mc Cursor pointing to the page where rebalancing
8946 * @return 0 on success, non-zero on failure.
8949 mdb_rebalance(MDB_cursor *mc)
8953 unsigned int ptop, minkeys, thresh;
8957 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8962 thresh = FILL_THRESHOLD;
8964 DPRINTF(("rebalancing %s page %"Yu" (has %u keys, %.1f%% full)",
8965 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8966 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8967 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8969 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8970 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8971 DPRINTF(("no need to rebalance page %"Yu", above fill threshold",
8972 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8976 if (mc->mc_snum < 2) {
8977 MDB_page *mp = mc->mc_pg[0];
8979 DPUTS("Can't rebalance a subpage, ignoring");
8982 if (NUMKEYS(mp) == 0) {
8983 DPUTS("tree is completely empty");
8984 mc->mc_db->md_root = P_INVALID;
8985 mc->mc_db->md_depth = 0;
8986 mc->mc_db->md_leaf_pages = 0;
8987 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8990 /* Adjust cursors pointing to mp */
8993 mc->mc_flags &= ~C_INITIALIZED;
8995 MDB_cursor *m2, *m3;
8996 MDB_dbi dbi = mc->mc_dbi;
8998 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8999 if (mc->mc_flags & C_SUB)
9000 m3 = &m2->mc_xcursor->mx_cursor;
9003 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
9005 if (m3->mc_pg[0] == mp) {
9008 m3->mc_flags &= ~C_INITIALIZED;
9012 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
9014 DPUTS("collapsing root page!");
9015 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
9018 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
9019 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
9022 mc->mc_db->md_depth--;
9023 mc->mc_db->md_branch_pages--;
9024 mc->mc_ki[0] = mc->mc_ki[1];
9025 for (i = 1; i<mc->mc_db->md_depth; i++) {
9026 mc->mc_pg[i] = mc->mc_pg[i+1];
9027 mc->mc_ki[i] = mc->mc_ki[i+1];
9030 /* Adjust other cursors pointing to mp */
9031 MDB_cursor *m2, *m3;
9032 MDB_dbi dbi = mc->mc_dbi;
9034 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9035 if (mc->mc_flags & C_SUB)
9036 m3 = &m2->mc_xcursor->mx_cursor;
9039 if (m3 == mc) continue;
9040 if (!(m3->mc_flags & C_INITIALIZED))
9042 if (m3->mc_pg[0] == mp) {
9043 for (i=0; i<mc->mc_db->md_depth; i++) {
9044 m3->mc_pg[i] = m3->mc_pg[i+1];
9045 m3->mc_ki[i] = m3->mc_ki[i+1];
9053 DPUTS("root page doesn't need rebalancing");
9057 /* The parent (branch page) must have at least 2 pointers,
9058 * otherwise the tree is invalid.
9060 ptop = mc->mc_top-1;
9061 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
9063 /* Leaf page fill factor is below the threshold.
9064 * Try to move keys from left or right neighbor, or
9065 * merge with a neighbor page.
9070 mdb_cursor_copy(mc, &mn);
9071 mn.mc_xcursor = NULL;
9073 oldki = mc->mc_ki[mc->mc_top];
9074 if (mc->mc_ki[ptop] == 0) {
9075 /* We're the leftmost leaf in our parent.
9077 DPUTS("reading right neighbor");
9079 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
9080 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
9083 mn.mc_ki[mn.mc_top] = 0;
9084 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
9087 /* There is at least one neighbor to the left.
9089 DPUTS("reading left neighbor");
9091 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
9092 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
9095 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
9096 mc->mc_ki[mc->mc_top] = 0;
9100 DPRINTF(("found neighbor page %"Yu" (%u keys, %.1f%% full)",
9101 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
9102 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
9104 /* If the neighbor page is above threshold and has enough keys,
9105 * move one key from it. Otherwise we should try to merge them.
9106 * (A branch page must never have less than 2 keys.)
9108 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
9109 rc = mdb_node_move(&mn, mc, fromleft);
9111 /* if we inserted on left, bump position up */
9116 rc = mdb_page_merge(&mn, mc);
9118 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
9119 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
9120 /* We want mdb_rebalance to find mn when doing fixups */
9121 WITH_CURSOR_TRACKING(mn,
9122 rc = mdb_page_merge(mc, &mn));
9123 mdb_cursor_copy(&mn, mc);
9125 mc->mc_flags &= ~C_EOF;
9127 mc->mc_ki[mc->mc_top] = oldki;
9131 /** Complete a delete operation started by #mdb_cursor_del(). */
9133 mdb_cursor_del0(MDB_cursor *mc)
9139 MDB_cursor *m2, *m3;
9140 MDB_dbi dbi = mc->mc_dbi;
9142 ki = mc->mc_ki[mc->mc_top];
9143 mp = mc->mc_pg[mc->mc_top];
9144 mdb_node_del(mc, mc->mc_db->md_pad);
9145 mc->mc_db->md_entries--;
9147 /* Adjust other cursors pointing to mp */
9148 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9149 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9150 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9152 if (m3 == mc || m3->mc_snum < mc->mc_snum)
9154 if (m3->mc_pg[mc->mc_top] == mp) {
9155 if (m3->mc_ki[mc->mc_top] == ki) {
9156 m3->mc_flags |= C_DEL;
9157 if (mc->mc_db->md_flags & MDB_DUPSORT) {
9158 /* Sub-cursor referred into dataset which is gone */
9159 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
9162 } else if (m3->mc_ki[mc->mc_top] > ki) {
9163 m3->mc_ki[mc->mc_top]--;
9165 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
9166 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9167 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
9168 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9173 rc = mdb_rebalance(mc);
9175 if (rc == MDB_SUCCESS) {
9176 /* DB is totally empty now, just bail out.
9177 * Other cursors adjustments were already done
9178 * by mdb_rebalance and aren't needed here.
9183 mp = mc->mc_pg[mc->mc_top];
9184 nkeys = NUMKEYS(mp);
9186 /* Adjust other cursors pointing to mp */
9187 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
9188 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9189 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9191 if (m3->mc_snum < mc->mc_snum)
9193 if (m3->mc_pg[mc->mc_top] == mp) {
9194 /* if m3 points past last node in page, find next sibling */
9195 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
9196 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9197 rc = mdb_cursor_sibling(m3, 1);
9198 if (rc == MDB_NOTFOUND) {
9199 m3->mc_flags |= C_EOF;
9204 if (mc->mc_db->md_flags & MDB_DUPSORT) {
9205 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
9206 /* If this node is a fake page, it needs to be reinited
9207 * because its data has moved. But just reset mc_pg[0]
9208 * if the xcursor is already live.
9210 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) {
9211 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)
9212 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9214 mdb_xcursor_init1(m3, node);
9220 mc->mc_flags |= C_DEL;
9224 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9229 mdb_del(MDB_txn *txn, MDB_dbi dbi,
9230 MDB_val *key, MDB_val *data)
9232 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9235 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9236 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9238 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
9239 /* must ignore any data */
9243 return mdb_del0(txn, dbi, key, data, 0);
9247 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
9248 MDB_val *key, MDB_val *data, unsigned flags)
9253 MDB_val rdata, *xdata;
9257 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
9259 mdb_cursor_init(&mc, txn, dbi, &mx);
9268 flags |= MDB_NODUPDATA;
9270 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
9272 /* let mdb_page_split know about this cursor if needed:
9273 * delete will trigger a rebalance; if it needs to move
9274 * a node from one page to another, it will have to
9275 * update the parent's separator key(s). If the new sepkey
9276 * is larger than the current one, the parent page may
9277 * run out of space, triggering a split. We need this
9278 * cursor to be consistent until the end of the rebalance.
9280 mc.mc_flags |= C_UNTRACK;
9281 mc.mc_next = txn->mt_cursors[dbi];
9282 txn->mt_cursors[dbi] = &mc;
9283 rc = mdb_cursor_del(&mc, flags);
9284 txn->mt_cursors[dbi] = mc.mc_next;
9289 /** Split a page and insert a new node.
9290 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
9291 * The cursor will be updated to point to the actual page and index where
9292 * the node got inserted after the split.
9293 * @param[in] newkey The key for the newly inserted node.
9294 * @param[in] newdata The data for the newly inserted node.
9295 * @param[in] newpgno The page number, if the new node is a branch node.
9296 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
9297 * @return 0 on success, non-zero on failure.
9300 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
9301 unsigned int nflags)
9304 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
9307 int i, j, split_indx, nkeys, pmax;
9308 MDB_env *env = mc->mc_txn->mt_env;
9310 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
9311 MDB_page *copy = NULL;
9312 MDB_page *mp, *rp, *pp;
9317 mp = mc->mc_pg[mc->mc_top];
9318 newindx = mc->mc_ki[mc->mc_top];
9319 nkeys = NUMKEYS(mp);
9321 DPRINTF(("-----> splitting %s page %"Yu" and adding [%s] at index %i/%i",
9322 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
9323 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
9325 /* Create a right sibling. */
9326 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
9328 rp->mp_pad = mp->mp_pad;
9329 DPRINTF(("new right sibling: page %"Yu, rp->mp_pgno));
9331 /* Usually when splitting the root page, the cursor
9332 * height is 1. But when called from mdb_update_key,
9333 * the cursor height may be greater because it walks
9334 * up the stack while finding the branch slot to update.
9336 if (mc->mc_top < 1) {
9337 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
9339 /* shift current top to make room for new parent */
9340 for (i=mc->mc_snum; i>0; i--) {
9341 mc->mc_pg[i] = mc->mc_pg[i-1];
9342 mc->mc_ki[i] = mc->mc_ki[i-1];
9346 mc->mc_db->md_root = pp->mp_pgno;
9347 DPRINTF(("root split! new root = %"Yu, pp->mp_pgno));
9348 new_root = mc->mc_db->md_depth++;
9350 /* Add left (implicit) pointer. */
9351 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
9352 /* undo the pre-push */
9353 mc->mc_pg[0] = mc->mc_pg[1];
9354 mc->mc_ki[0] = mc->mc_ki[1];
9355 mc->mc_db->md_root = mp->mp_pgno;
9356 mc->mc_db->md_depth--;
9363 ptop = mc->mc_top-1;
9364 DPRINTF(("parent branch page is %"Yu, mc->mc_pg[ptop]->mp_pgno));
9367 mdb_cursor_copy(mc, &mn);
9368 mn.mc_xcursor = NULL;
9369 mn.mc_pg[mn.mc_top] = rp;
9370 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
9372 if (nflags & MDB_APPEND) {
9373 mn.mc_ki[mn.mc_top] = 0;
9375 split_indx = newindx;
9379 split_indx = (nkeys+1) / 2;
9384 unsigned int lsize, rsize, ksize;
9385 /* Move half of the keys to the right sibling */
9386 x = mc->mc_ki[mc->mc_top] - split_indx;
9387 ksize = mc->mc_db->md_pad;
9388 split = LEAF2KEY(mp, split_indx, ksize);
9389 rsize = (nkeys - split_indx) * ksize;
9390 lsize = (nkeys - split_indx) * sizeof(indx_t);
9391 mp->mp_lower -= lsize;
9392 rp->mp_lower += lsize;
9393 mp->mp_upper += rsize - lsize;
9394 rp->mp_upper -= rsize - lsize;
9395 sepkey.mv_size = ksize;
9396 if (newindx == split_indx) {
9397 sepkey.mv_data = newkey->mv_data;
9399 sepkey.mv_data = split;
9402 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
9403 memcpy(rp->mp_ptrs, split, rsize);
9404 sepkey.mv_data = rp->mp_ptrs;
9405 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
9406 memcpy(ins, newkey->mv_data, ksize);
9407 mp->mp_lower += sizeof(indx_t);
9408 mp->mp_upper -= ksize - sizeof(indx_t);
9411 memcpy(rp->mp_ptrs, split, x * ksize);
9412 ins = LEAF2KEY(rp, x, ksize);
9413 memcpy(ins, newkey->mv_data, ksize);
9414 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
9415 rp->mp_lower += sizeof(indx_t);
9416 rp->mp_upper -= ksize - sizeof(indx_t);
9417 mc->mc_ki[mc->mc_top] = x;
9420 int psize, nsize, k;
9421 /* Maximum free space in an empty page */
9422 pmax = env->me_psize - PAGEHDRSZ;
9424 nsize = mdb_leaf_size(env, newkey, newdata);
9426 nsize = mdb_branch_size(env, newkey);
9427 nsize = EVEN(nsize);
9429 /* grab a page to hold a temporary copy */
9430 copy = mdb_page_malloc(mc->mc_txn, 1);
9435 copy->mp_pgno = mp->mp_pgno;
9436 copy->mp_flags = mp->mp_flags;
9437 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
9438 copy->mp_upper = env->me_psize - PAGEBASE;
9440 /* prepare to insert */
9441 for (i=0, j=0; i<nkeys; i++) {
9443 copy->mp_ptrs[j++] = 0;
9445 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
9448 /* When items are relatively large the split point needs
9449 * to be checked, because being off-by-one will make the
9450 * difference between success or failure in mdb_node_add.
9452 * It's also relevant if a page happens to be laid out
9453 * such that one half of its nodes are all "small" and
9454 * the other half of its nodes are "large." If the new
9455 * item is also "large" and falls on the half with
9456 * "large" nodes, it also may not fit.
9458 * As a final tweak, if the new item goes on the last
9459 * spot on the page (and thus, onto the new page), bias
9460 * the split so the new page is emptier than the old page.
9461 * This yields better packing during sequential inserts.
9463 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
9464 /* Find split point */
9466 if (newindx <= split_indx || newindx >= nkeys) {
9468 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
9473 for (; i!=k; i+=j) {
9478 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9479 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
9481 if (F_ISSET(node->mn_flags, F_BIGDATA))
9482 psize += sizeof(pgno_t);
9484 psize += NODEDSZ(node);
9486 psize = EVEN(psize);
9488 if (psize > pmax || i == k-j) {
9489 split_indx = i + (j<0);
9494 if (split_indx == newindx) {
9495 sepkey.mv_size = newkey->mv_size;
9496 sepkey.mv_data = newkey->mv_data;
9498 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
9499 sepkey.mv_size = node->mn_ksize;
9500 sepkey.mv_data = NODEKEY(node);
9505 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
9507 /* Copy separator key to the parent.
9509 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
9510 int snum = mc->mc_snum;
9514 /* We want other splits to find mn when doing fixups */
9515 WITH_CURSOR_TRACKING(mn,
9516 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
9521 if (mc->mc_snum > snum) {
9524 /* Right page might now have changed parent.
9525 * Check if left page also changed parent.
9527 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9528 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9529 for (i=0; i<ptop; i++) {
9530 mc->mc_pg[i] = mn.mc_pg[i];
9531 mc->mc_ki[i] = mn.mc_ki[i];
9533 mc->mc_pg[ptop] = mn.mc_pg[ptop];
9534 if (mn.mc_ki[ptop]) {
9535 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
9537 /* find right page's left sibling */
9538 mc->mc_ki[ptop] = mn.mc_ki[ptop];
9539 mdb_cursor_sibling(mc, 0);
9544 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
9547 if (rc != MDB_SUCCESS) {
9550 if (nflags & MDB_APPEND) {
9551 mc->mc_pg[mc->mc_top] = rp;
9552 mc->mc_ki[mc->mc_top] = 0;
9553 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
9556 for (i=0; i<mc->mc_top; i++)
9557 mc->mc_ki[i] = mn.mc_ki[i];
9558 } else if (!IS_LEAF2(mp)) {
9560 mc->mc_pg[mc->mc_top] = rp;
9565 rkey.mv_data = newkey->mv_data;
9566 rkey.mv_size = newkey->mv_size;
9572 /* Update index for the new key. */
9573 mc->mc_ki[mc->mc_top] = j;
9575 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9576 rkey.mv_data = NODEKEY(node);
9577 rkey.mv_size = node->mn_ksize;
9579 xdata.mv_data = NODEDATA(node);
9580 xdata.mv_size = NODEDSZ(node);
9583 pgno = NODEPGNO(node);
9584 flags = node->mn_flags;
9587 if (!IS_LEAF(mp) && j == 0) {
9588 /* First branch index doesn't need key data. */
9592 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
9598 mc->mc_pg[mc->mc_top] = copy;
9603 } while (i != split_indx);
9605 nkeys = NUMKEYS(copy);
9606 for (i=0; i<nkeys; i++)
9607 mp->mp_ptrs[i] = copy->mp_ptrs[i];
9608 mp->mp_lower = copy->mp_lower;
9609 mp->mp_upper = copy->mp_upper;
9610 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
9611 env->me_psize - copy->mp_upper - PAGEBASE);
9613 /* reset back to original page */
9614 if (newindx < split_indx) {
9615 mc->mc_pg[mc->mc_top] = mp;
9617 mc->mc_pg[mc->mc_top] = rp;
9619 /* Make sure mc_ki is still valid.
9621 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9622 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9623 for (i=0; i<=ptop; i++) {
9624 mc->mc_pg[i] = mn.mc_pg[i];
9625 mc->mc_ki[i] = mn.mc_ki[i];
9629 if (nflags & MDB_RESERVE) {
9630 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
9631 if (!(node->mn_flags & F_BIGDATA))
9632 newdata->mv_data = NODEDATA(node);
9635 if (newindx >= split_indx) {
9636 mc->mc_pg[mc->mc_top] = rp;
9638 /* Make sure mc_ki is still valid.
9640 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9641 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9642 for (i=0; i<=ptop; i++) {
9643 mc->mc_pg[i] = mn.mc_pg[i];
9644 mc->mc_ki[i] = mn.mc_ki[i];
9651 /* Adjust other cursors pointing to mp */
9652 MDB_cursor *m2, *m3;
9653 MDB_dbi dbi = mc->mc_dbi;
9654 nkeys = NUMKEYS(mp);
9656 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9657 if (mc->mc_flags & C_SUB)
9658 m3 = &m2->mc_xcursor->mx_cursor;
9663 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9667 /* sub cursors may be on different DB */
9668 if (m3->mc_pg[0] != mp)
9671 for (k=new_root; k>=0; k--) {
9672 m3->mc_ki[k+1] = m3->mc_ki[k];
9673 m3->mc_pg[k+1] = m3->mc_pg[k];
9675 if (m3->mc_ki[0] >= nkeys) {
9680 m3->mc_pg[0] = mc->mc_pg[0];
9684 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
9685 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
9686 m3->mc_ki[mc->mc_top]++;
9687 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9688 m3->mc_pg[mc->mc_top] = rp;
9689 m3->mc_ki[mc->mc_top] -= nkeys;
9690 for (i=0; i<mc->mc_top; i++) {
9691 m3->mc_ki[i] = mn.mc_ki[i];
9692 m3->mc_pg[i] = mn.mc_pg[i];
9695 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
9696 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
9699 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
9701 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9702 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
9703 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9707 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
9710 if (copy) /* tmp page */
9711 mdb_page_free(env, copy);
9713 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9718 mdb_put(MDB_txn *txn, MDB_dbi dbi,
9719 MDB_val *key, MDB_val *data, unsigned int flags)
9725 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9728 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
9731 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9732 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9734 mdb_cursor_init(&mc, txn, dbi, &mx);
9735 mc.mc_next = txn->mt_cursors[dbi];
9736 txn->mt_cursors[dbi] = &mc;
9737 rc = mdb_cursor_put(&mc, key, data, flags);
9738 txn->mt_cursors[dbi] = mc.mc_next;
9743 #define MDB_WBUF (1024*1024)
9745 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
9747 /** State needed for a double-buffering compacting copy. */
9748 typedef struct mdb_copy {
9751 pthread_mutex_t mc_mutex;
9752 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
9757 pgno_t mc_next_pgno;
9759 int mc_toggle; /**< Buffer number in provider */
9760 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
9761 volatile int mc_error; /**< Error code, never cleared if set */
9764 /** Dedicated writer thread for compacting copy. */
9765 static THREAD_RET ESECT CALL_CONV
9766 mdb_env_copythr(void *arg)
9770 int toggle = 0, wsize, rc;
9773 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9776 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9779 pthread_mutex_lock(&my->mc_mutex);
9782 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9783 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
9785 wsize = my->mc_wlen[toggle];
9786 ptr = my->mc_wbuf[toggle];
9789 while (wsize > 0 && !my->mc_error) {
9790 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9794 } else if (len > 0) {
9807 /* If there's an overflow page tail, write it too */
9808 if (my->mc_olen[toggle]) {
9809 wsize = my->mc_olen[toggle];
9810 ptr = my->mc_over[toggle];
9811 my->mc_olen[toggle] = 0;
9814 my->mc_wlen[toggle] = 0;
9816 /* Return the empty buffer to provider */
9818 pthread_cond_signal(&my->mc_cond);
9820 pthread_mutex_unlock(&my->mc_mutex);
9821 return (THREAD_RET)0;
9825 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9827 * @param[in] my control structure.
9828 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9831 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9833 pthread_mutex_lock(&my->mc_mutex);
9834 my->mc_new += adjust;
9835 pthread_cond_signal(&my->mc_cond);
9836 while (my->mc_new & 2) /* both buffers in use */
9837 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9838 pthread_mutex_unlock(&my->mc_mutex);
9840 my->mc_toggle ^= (adjust & 1);
9841 /* Both threads reset mc_wlen, to be safe from threading errors */
9842 my->mc_wlen[my->mc_toggle] = 0;
9843 return my->mc_error;
9846 /** Depth-first tree traversal for compacting copy. */
9848 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9850 MDB_cursor mc = {0};
9852 MDB_page *mo, *mp, *leaf;
9857 /* Empty DB, nothing to do */
9858 if (*pg == P_INVALID)
9862 mc.mc_txn = my->mc_txn;
9863 mc.mc_flags = my->mc_txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
9865 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9868 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9872 /* Make cursor pages writable */
9873 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9877 for (i=0; i<mc.mc_top; i++) {
9878 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9879 mc.mc_pg[i] = (MDB_page *)ptr;
9880 ptr += my->mc_env->me_psize;
9883 /* This is writable space for a leaf page. Usually not needed. */
9884 leaf = (MDB_page *)ptr;
9886 toggle = my->mc_toggle;
9887 while (mc.mc_snum > 0) {
9889 mp = mc.mc_pg[mc.mc_top];
9893 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9894 for (i=0; i<n; i++) {
9895 ni = NODEPTR(mp, i);
9896 if (ni->mn_flags & F_BIGDATA) {
9900 /* Need writable leaf */
9902 mc.mc_pg[mc.mc_top] = leaf;
9903 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9905 ni = NODEPTR(mp, i);
9908 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9909 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
9910 rc = mdb_page_get(&mc, pg, &omp, NULL);
9913 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9914 rc = mdb_env_cthr_toggle(my, 1);
9917 toggle = my->mc_toggle;
9919 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9920 memcpy(mo, omp, my->mc_env->me_psize);
9921 mo->mp_pgno = my->mc_next_pgno;
9922 my->mc_next_pgno += omp->mp_pages;
9923 my->mc_wlen[toggle] += my->mc_env->me_psize;
9924 if (omp->mp_pages > 1) {
9925 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9926 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9927 rc = mdb_env_cthr_toggle(my, 1);
9930 toggle = my->mc_toggle;
9932 } else if (ni->mn_flags & F_SUBDATA) {
9935 /* Need writable leaf */
9937 mc.mc_pg[mc.mc_top] = leaf;
9938 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9940 ni = NODEPTR(mp, i);
9943 memcpy(&db, NODEDATA(ni), sizeof(db));
9944 my->mc_toggle = toggle;
9945 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9948 toggle = my->mc_toggle;
9949 memcpy(NODEDATA(ni), &db, sizeof(db));
9954 mc.mc_ki[mc.mc_top]++;
9955 if (mc.mc_ki[mc.mc_top] < n) {
9958 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9960 rc = mdb_page_get(&mc, pg, &mp, NULL);
9965 mc.mc_ki[mc.mc_top] = 0;
9966 if (IS_BRANCH(mp)) {
9967 /* Whenever we advance to a sibling branch page,
9968 * we must proceed all the way down to its first leaf.
9970 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9973 mc.mc_pg[mc.mc_top] = mp;
9977 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9978 rc = mdb_env_cthr_toggle(my, 1);
9981 toggle = my->mc_toggle;
9983 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9984 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9985 mo->mp_pgno = my->mc_next_pgno++;
9986 my->mc_wlen[toggle] += my->mc_env->me_psize;
9988 /* Update parent if there is one */
9989 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9990 SETPGNO(ni, mo->mp_pgno);
9991 mdb_cursor_pop(&mc);
9993 /* Otherwise we're done */
10003 /** Copy environment with compaction. */
10005 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
10010 MDB_txn *txn = NULL;
10012 pgno_t root, new_root;
10013 int rc = MDB_SUCCESS;
10016 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
10017 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
10021 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
10022 if (my.mc_wbuf[0] == NULL) {
10023 /* _aligned_malloc() sets errno, but we use Windows error codes */
10024 rc = ERROR_NOT_ENOUGH_MEMORY;
10028 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
10030 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
10032 #ifdef HAVE_MEMALIGN
10033 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
10034 if (my.mc_wbuf[0] == NULL) {
10041 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
10047 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
10048 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
10049 my.mc_next_pgno = NUM_METAS;
10052 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
10056 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
10060 mp = (MDB_page *)my.mc_wbuf[0];
10061 memset(mp, 0, NUM_METAS * env->me_psize);
10063 mp->mp_flags = P_META;
10064 mm = (MDB_meta *)METADATA(mp);
10065 mdb_env_init_meta0(env, mm);
10066 mm->mm_address = env->me_metas[0]->mm_address;
10068 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
10070 mp->mp_flags = P_META;
10071 *(MDB_meta *)METADATA(mp) = *mm;
10072 mm = (MDB_meta *)METADATA(mp);
10074 /* Set metapage 1 with current main DB */
10075 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
10076 if (root != P_INVALID) {
10077 /* Count free pages + freeDB pages. Subtract from last_pg
10078 * to find the new last_pg, which also becomes the new root.
10080 MDB_ID freecount = 0;
10083 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
10084 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
10085 freecount += *(MDB_ID *)data.mv_data;
10086 if (rc != MDB_NOTFOUND)
10088 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
10089 txn->mt_dbs[FREE_DBI].md_leaf_pages +
10090 txn->mt_dbs[FREE_DBI].md_overflow_pages;
10092 new_root = txn->mt_next_pgno - 1 - freecount;
10093 mm->mm_last_pg = new_root;
10094 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
10095 mm->mm_dbs[MAIN_DBI].md_root = new_root;
10097 /* When the DB is empty, handle it specially to
10098 * fix any breakage like page leaks from ITS#8174.
10100 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
10102 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
10103 mm->mm_txnid = 1; /* use metapage 1 */
10106 my.mc_wlen[0] = env->me_psize * NUM_METAS;
10108 rc = mdb_env_cwalk(&my, &root, 0);
10109 if (rc == MDB_SUCCESS && root != new_root) {
10110 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
10116 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
10117 rc = THREAD_FINISH(thr);
10118 mdb_txn_abort(txn);
10122 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
10123 if (my.mc_cond) CloseHandle(my.mc_cond);
10124 if (my.mc_mutex) CloseHandle(my.mc_mutex);
10126 free(my.mc_wbuf[0]);
10127 pthread_cond_destroy(&my.mc_cond);
10129 pthread_mutex_destroy(&my.mc_mutex);
10131 return rc ? rc : my.mc_error;
10134 /** Copy environment as-is. */
10136 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
10138 MDB_txn *txn = NULL;
10139 mdb_mutexref_t wmutex = NULL;
10141 mdb_size_t wsize, w3;
10145 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
10149 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
10152 /* Do the lock/unlock of the reader mutex before starting the
10153 * write txn. Otherwise other read txns could block writers.
10155 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
10159 if (env->me_txns) {
10160 /* We must start the actual read txn after blocking writers */
10161 mdb_txn_end(txn, MDB_END_RESET_TMP);
10163 /* Temporarily block writers until we snapshot the meta pages */
10164 wmutex = env->me_wmutex;
10165 if (LOCK_MUTEX(rc, env, wmutex))
10168 rc = mdb_txn_renew0(txn);
10170 UNLOCK_MUTEX(wmutex);
10175 wsize = env->me_psize * NUM_METAS;
10179 DO_WRITE(rc, fd, ptr, w2, len);
10183 } else if (len > 0) {
10189 /* Non-blocking or async handles are not supported */
10195 UNLOCK_MUTEX(wmutex);
10200 w3 = txn->mt_next_pgno * env->me_psize;
10202 mdb_size_t fsize = 0;
10203 if ((rc = mdb_fsize(env->me_fd, &fsize)))
10208 wsize = w3 - wsize;
10209 while (wsize > 0) {
10210 if (wsize > MAX_WRITE)
10214 DO_WRITE(rc, fd, ptr, w2, len);
10218 } else if (len > 0) {
10230 mdb_txn_abort(txn);
10235 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
10237 if (flags & MDB_CP_COMPACT)
10238 return mdb_env_copyfd1(env, fd);
10240 return mdb_env_copyfd0(env, fd);
10244 mdb_env_copyfd(MDB_env *env, HANDLE fd)
10246 return mdb_env_copyfd2(env, fd, 0);
10250 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
10254 HANDLE newfd = INVALID_HANDLE_VALUE;
10256 rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname);
10257 if (rc == MDB_SUCCESS) {
10258 rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd);
10259 mdb_fname_destroy(fname);
10261 if (rc == MDB_SUCCESS) {
10262 rc = mdb_env_copyfd2(env, newfd, flags);
10263 if (close(newfd) < 0 && rc == MDB_SUCCESS)
10270 mdb_env_copy(MDB_env *env, const char *path)
10272 return mdb_env_copy2(env, path, 0);
10276 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
10278 if (flag & ~CHANGEABLE)
10281 env->me_flags |= flag;
10283 env->me_flags &= ~flag;
10284 return MDB_SUCCESS;
10288 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
10293 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
10294 return MDB_SUCCESS;
10298 mdb_env_set_userctx(MDB_env *env, void *ctx)
10302 env->me_userctx = ctx;
10303 return MDB_SUCCESS;
10307 mdb_env_get_userctx(MDB_env *env)
10309 return env ? env->me_userctx : NULL;
10313 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
10318 env->me_assert_func = func;
10320 return MDB_SUCCESS;
10324 mdb_env_get_path(MDB_env *env, const char **arg)
10329 *arg = env->me_path;
10330 return MDB_SUCCESS;
10334 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
10340 return MDB_SUCCESS;
10343 /** Common code for #mdb_stat() and #mdb_env_stat().
10344 * @param[in] env the environment to operate in.
10345 * @param[in] db the #MDB_db record containing the stats to return.
10346 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
10347 * @return 0, this function always succeeds.
10350 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
10352 arg->ms_psize = env->me_psize;
10353 arg->ms_depth = db->md_depth;
10354 arg->ms_branch_pages = db->md_branch_pages;
10355 arg->ms_leaf_pages = db->md_leaf_pages;
10356 arg->ms_overflow_pages = db->md_overflow_pages;
10357 arg->ms_entries = db->md_entries;
10359 return MDB_SUCCESS;
10363 mdb_env_stat(MDB_env *env, MDB_stat *arg)
10367 if (env == NULL || arg == NULL)
10370 meta = mdb_env_pick_meta(env);
10372 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
10376 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
10380 if (env == NULL || arg == NULL)
10383 meta = mdb_env_pick_meta(env);
10384 arg->me_mapaddr = meta->mm_address;
10385 arg->me_last_pgno = meta->mm_last_pg;
10386 arg->me_last_txnid = meta->mm_txnid;
10388 arg->me_mapsize = env->me_mapsize;
10389 arg->me_maxreaders = env->me_maxreaders;
10390 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
10391 return MDB_SUCCESS;
10394 /** Set the default comparison functions for a database.
10395 * Called immediately after a database is opened to set the defaults.
10396 * The user can then override them with #mdb_set_compare() or
10397 * #mdb_set_dupsort().
10398 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
10399 * @param[in] dbi A database handle returned by #mdb_dbi_open()
10402 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
10404 uint16_t f = txn->mt_dbs[dbi].md_flags;
10406 txn->mt_dbxs[dbi].md_cmp =
10407 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
10408 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
10410 txn->mt_dbxs[dbi].md_dcmp =
10411 !(f & MDB_DUPSORT) ? 0 :
10412 ((f & MDB_INTEGERDUP)
10413 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
10414 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
10417 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
10423 int rc, dbflag, exact;
10424 unsigned int unused = 0, seq;
10428 if (flags & ~VALID_FLAGS)
10430 if (txn->mt_flags & MDB_TXN_BLOCKED)
10431 return MDB_BAD_TXN;
10436 if (flags & PERSISTENT_FLAGS) {
10437 uint16_t f2 = flags & PERSISTENT_FLAGS;
10438 /* make sure flag changes get committed */
10439 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
10440 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
10441 txn->mt_flags |= MDB_TXN_DIRTY;
10444 mdb_default_cmp(txn, MAIN_DBI);
10445 return MDB_SUCCESS;
10448 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
10449 mdb_default_cmp(txn, MAIN_DBI);
10452 /* Is the DB already open? */
10453 len = strlen(name);
10454 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
10455 if (!txn->mt_dbxs[i].md_name.mv_size) {
10456 /* Remember this free slot */
10457 if (!unused) unused = i;
10460 if (len == txn->mt_dbxs[i].md_name.mv_size &&
10461 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
10463 return MDB_SUCCESS;
10467 /* If no free slot and max hit, fail */
10468 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
10469 return MDB_DBS_FULL;
10471 /* Cannot mix named databases with some mainDB flags */
10472 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
10473 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
10475 /* Find the DB info */
10476 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
10479 key.mv_data = (void *)name;
10480 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
10481 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
10482 if (rc == MDB_SUCCESS) {
10483 /* make sure this is actually a DB */
10484 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
10485 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
10486 return MDB_INCOMPATIBLE;
10487 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
10491 /* Done here so we cannot fail after creating a new DB */
10492 if ((namedup = strdup(name)) == NULL)
10496 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
10497 data.mv_size = sizeof(MDB_db);
10498 data.mv_data = &dummy;
10499 memset(&dummy, 0, sizeof(dummy));
10500 dummy.md_root = P_INVALID;
10501 dummy.md_flags = flags & PERSISTENT_FLAGS;
10502 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
10503 dbflag |= DB_DIRTY;
10509 /* Got info, register DBI in this txn */
10510 unsigned int slot = unused ? unused : txn->mt_numdbs;
10511 txn->mt_dbxs[slot].md_name.mv_data = namedup;
10512 txn->mt_dbxs[slot].md_name.mv_size = len;
10513 txn->mt_dbxs[slot].md_rel = NULL;
10514 txn->mt_dbflags[slot] = dbflag;
10515 /* txn-> and env-> are the same in read txns, use
10516 * tmp variable to avoid undefined assignment
10518 seq = ++txn->mt_env->me_dbiseqs[slot];
10519 txn->mt_dbiseqs[slot] = seq;
10521 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
10523 mdb_default_cmp(txn, slot);
10533 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
10535 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
10538 if (txn->mt_flags & MDB_TXN_BLOCKED)
10539 return MDB_BAD_TXN;
10541 if (txn->mt_dbflags[dbi] & DB_STALE) {
10544 /* Stale, must read the DB's root. cursor_init does it for us. */
10545 mdb_cursor_init(&mc, txn, dbi, &mx);
10547 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
10550 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
10553 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
10555 ptr = env->me_dbxs[dbi].md_name.mv_data;
10556 /* If there was no name, this was already closed */
10558 env->me_dbxs[dbi].md_name.mv_data = NULL;
10559 env->me_dbxs[dbi].md_name.mv_size = 0;
10560 env->me_dbflags[dbi] = 0;
10561 env->me_dbiseqs[dbi]++;
10566 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
10568 /* We could return the flags for the FREE_DBI too but what's the point? */
10569 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10571 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
10572 return MDB_SUCCESS;
10575 /** Add all the DB's pages to the free list.
10576 * @param[in] mc Cursor on the DB to free.
10577 * @param[in] subs non-Zero to check for sub-DBs in this DB.
10578 * @return 0 on success, non-zero on failure.
10581 mdb_drop0(MDB_cursor *mc, int subs)
10585 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
10586 if (rc == MDB_SUCCESS) {
10587 MDB_txn *txn = mc->mc_txn;
10592 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
10593 * This also avoids any P_LEAF2 pages, which have no nodes.
10594 * Also if the DB doesn't have sub-DBs and has no overflow
10595 * pages, omit scanning leaves.
10597 if ((mc->mc_flags & C_SUB) ||
10598 (!subs && !mc->mc_db->md_overflow_pages))
10599 mdb_cursor_pop(mc);
10601 mdb_cursor_copy(mc, &mx);
10603 /* bump refcount for mx's pages */
10604 for (i=0; i<mc->mc_snum; i++)
10605 mdb_page_get(&mx, mc->mc_pg[i]->mp_pgno, &mx.mc_pg[i], NULL);
10607 while (mc->mc_snum > 0) {
10608 MDB_page *mp = mc->mc_pg[mc->mc_top];
10609 unsigned n = NUMKEYS(mp);
10611 for (i=0; i<n; i++) {
10612 ni = NODEPTR(mp, i);
10613 if (ni->mn_flags & F_BIGDATA) {
10616 memcpy(&pg, NODEDATA(ni), sizeof(pg));
10617 rc = mdb_page_get(mc, pg, &omp, NULL);
10620 mdb_cassert(mc, IS_OVERFLOW(omp));
10621 rc = mdb_midl_append_range(&txn->mt_free_pgs,
10622 pg, omp->mp_pages);
10625 mc->mc_db->md_overflow_pages -= omp->mp_pages;
10626 if (!mc->mc_db->md_overflow_pages && !subs)
10628 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
10629 mdb_xcursor_init1(mc, ni);
10630 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
10635 if (!subs && !mc->mc_db->md_overflow_pages)
10638 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
10640 for (i=0; i<n; i++) {
10642 ni = NODEPTR(mp, i);
10645 mdb_midl_xappend(txn->mt_free_pgs, pg);
10650 mc->mc_ki[mc->mc_top] = i;
10651 rc = mdb_cursor_sibling(mc, 1);
10653 if (rc != MDB_NOTFOUND)
10655 /* no more siblings, go back to beginning
10656 * of previous level.
10659 mdb_cursor_pop(mc);
10661 for (i=1; i<mc->mc_snum; i++) {
10663 mc->mc_pg[i] = mx.mc_pg[i];
10668 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
10671 txn->mt_flags |= MDB_TXN_ERROR;
10672 /* drop refcount for mx's pages */
10673 MDB_CURSOR_UNREF(&mx, 0);
10674 } else if (rc == MDB_NOTFOUND) {
10677 mc->mc_flags &= ~C_INITIALIZED;
10681 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
10683 MDB_cursor *mc, *m2;
10686 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10689 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
10692 if (TXN_DBI_CHANGED(txn, dbi))
10693 return MDB_BAD_DBI;
10695 rc = mdb_cursor_open(txn, dbi, &mc);
10699 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
10700 /* Invalidate the dropped DB's cursors */
10701 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
10702 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
10706 /* Can't delete the main DB */
10707 if (del && dbi >= CORE_DBS) {
10708 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
10710 txn->mt_dbflags[dbi] = DB_STALE;
10711 mdb_dbi_close(txn->mt_env, dbi);
10713 txn->mt_flags |= MDB_TXN_ERROR;
10716 /* reset the DB record, mark it dirty */
10717 txn->mt_dbflags[dbi] |= DB_DIRTY;
10718 txn->mt_dbs[dbi].md_depth = 0;
10719 txn->mt_dbs[dbi].md_branch_pages = 0;
10720 txn->mt_dbs[dbi].md_leaf_pages = 0;
10721 txn->mt_dbs[dbi].md_overflow_pages = 0;
10722 txn->mt_dbs[dbi].md_entries = 0;
10723 txn->mt_dbs[dbi].md_root = P_INVALID;
10725 txn->mt_flags |= MDB_TXN_DIRTY;
10728 mdb_cursor_close(mc);
10732 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10734 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10737 txn->mt_dbxs[dbi].md_cmp = cmp;
10738 return MDB_SUCCESS;
10741 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10743 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10746 txn->mt_dbxs[dbi].md_dcmp = cmp;
10747 return MDB_SUCCESS;
10750 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
10752 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10755 txn->mt_dbxs[dbi].md_rel = rel;
10756 return MDB_SUCCESS;
10759 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
10761 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10764 txn->mt_dbxs[dbi].md_relctx = ctx;
10765 return MDB_SUCCESS;
10769 mdb_env_get_maxkeysize(MDB_env *env)
10771 return ENV_MAXKEY(env);
10775 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
10777 unsigned int i, rdrs;
10780 int rc = 0, first = 1;
10784 if (!env->me_txns) {
10785 return func("(no reader locks)\n", ctx);
10787 rdrs = env->me_txns->mti_numreaders;
10788 mr = env->me_txns->mti_readers;
10789 for (i=0; i<rdrs; i++) {
10790 if (mr[i].mr_pid) {
10791 txnid_t txnid = mr[i].mr_txnid;
10792 sprintf(buf, txnid == (txnid_t)-1 ?
10793 "%10d %"Z"x -\n" : "%10d %"Z"x %"Yu"\n",
10794 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10797 rc = func(" pid thread txnid\n", ctx);
10801 rc = func(buf, ctx);
10807 rc = func("(no active readers)\n", ctx);
10812 /** Insert pid into list if not already present.
10813 * return -1 if already present.
10816 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10818 /* binary search of pid in list */
10820 unsigned cursor = 1;
10822 unsigned n = ids[0];
10825 unsigned pivot = n >> 1;
10826 cursor = base + pivot + 1;
10827 val = pid - ids[cursor];
10832 } else if ( val > 0 ) {
10837 /* found, so it's a duplicate */
10846 for (n = ids[0]; n > cursor; n--)
10853 mdb_reader_check(MDB_env *env, int *dead)
10859 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10862 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
10864 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10866 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10867 unsigned int i, j, rdrs;
10869 MDB_PID_T *pids, pid;
10870 int rc = MDB_SUCCESS, count = 0;
10872 rdrs = env->me_txns->mti_numreaders;
10873 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10877 mr = env->me_txns->mti_readers;
10878 for (i=0; i<rdrs; i++) {
10879 pid = mr[i].mr_pid;
10880 if (pid && pid != env->me_pid) {
10881 if (mdb_pid_insert(pids, pid) == 0) {
10882 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10883 /* Stale reader found */
10886 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10887 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10889 rdrs = 0; /* the above checked all readers */
10891 /* Recheck, a new process may have reused pid */
10892 if (mdb_reader_pid(env, Pidcheck, pid))
10896 for (; j<rdrs; j++)
10897 if (mr[j].mr_pid == pid) {
10898 DPRINTF(("clear stale reader pid %u txn %"Yd,
10899 (unsigned) pid, mr[j].mr_txnid));
10904 UNLOCK_MUTEX(rmutex);
10915 #ifdef MDB_ROBUST_SUPPORTED
10916 /** Handle #LOCK_MUTEX0() failure.
10917 * Try to repair the lock file if the mutex owner died.
10918 * @param[in] env the environment handle
10919 * @param[in] mutex LOCK_MUTEX0() mutex
10920 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10921 * @return 0 on success with the mutex locked, or an error code on failure.
10924 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10929 if (rc == MDB_OWNERDEAD) {
10930 /* We own the mutex. Clean up after dead previous owner. */
10932 rlocked = (mutex == env->me_rmutex);
10934 /* Keep mti_txnid updated, otherwise next writer can
10935 * overwrite data which latest meta page refers to.
10937 meta = mdb_env_pick_meta(env);
10938 env->me_txns->mti_txnid = meta->mm_txnid;
10939 /* env is hosed if the dead thread was ours */
10941 env->me_flags |= MDB_FATAL_ERROR;
10942 env->me_txn = NULL;
10946 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10947 (rc ? "this process' env is hosed" : "recovering")));
10948 rc2 = mdb_reader_check0(env, rlocked, NULL);
10950 rc2 = mdb_mutex_consistent(mutex);
10951 if (rc || (rc = rc2)) {
10952 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10953 UNLOCK_MUTEX(mutex);
10959 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10964 #endif /* MDB_ROBUST_SUPPORTED */
10966 #if defined(_WIN32)
10967 /** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */
10969 utf8_to_utf16(const char *src, MDB_name *dst, int xtra)
10972 wchar_t *result = NULL;
10973 for (;;) { /* malloc result, then fill it in */
10974 need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need);
10981 result = malloc(sizeof(wchar_t) * (need + xtra));
10986 dst->mn_alloced = 1;
10987 dst->mn_len = need - 1;
10988 dst->mn_val = result;
10989 return MDB_SUCCESS;
10992 #endif /* defined(_WIN32) */
projects / platform / upstream / lmdb.git / blob