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 2011-2018 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. We get pointers to these functions from
52 * NTDLL.DLL at runtime, to avoid buildtime dependencies on any
53 * NTDLL import libraries.
55 typedef NTSTATUS WINAPI (NtCreateSectionFunc)
56 (OUT PHANDLE sh, IN ACCESS_MASK acc,
57 IN void * oa OPTIONAL,
58 IN PLARGE_INTEGER ms OPTIONAL,
59 IN ULONG pp, IN ULONG aa, IN HANDLE fh OPTIONAL);
61 static NtCreateSectionFunc *NtCreateSection;
63 typedef enum _SECTION_INHERIT {
68 typedef NTSTATUS WINAPI (NtMapViewOfSectionFunc)
69 (IN PHANDLE sh, IN HANDLE ph,
70 IN OUT PVOID *addr, IN ULONG_PTR zbits,
71 IN SIZE_T cs, IN OUT PLARGE_INTEGER off OPTIONAL,
72 IN OUT PSIZE_T vs, IN SECTION_INHERIT ih,
73 IN ULONG at, IN ULONG pp);
75 static NtMapViewOfSectionFunc *NtMapViewOfSection;
77 typedef NTSTATUS WINAPI (NtCloseFunc)(HANDLE h);
79 static NtCloseFunc *NtClose;
81 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
82 * as int64 which is wrong. MSVC doesn't define it at all, so just
86 #define MDB_THR_T DWORD
87 #include <sys/types.h>
90 # include <sys/param.h>
92 # define LITTLE_ENDIAN 1234
93 # define BIG_ENDIAN 4321
94 # define BYTE_ORDER LITTLE_ENDIAN
96 # define SSIZE_MAX INT_MAX
100 #include <sys/types.h>
101 #include <sys/stat.h>
102 #define MDB_PID_T pid_t
103 #define MDB_THR_T pthread_t
104 #include <sys/param.h>
106 #include <sys/mman.h>
107 #ifdef HAVE_SYS_FILE_H
108 #include <sys/file.h>
113 #if defined(__mips) && defined(__linux)
114 /* MIPS has cache coherency issues, requires explicit cache control */
115 #include <asm/cachectl.h>
116 extern int cacheflush(char *addr, int nbytes, int cache);
117 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
119 #define CACHEFLUSH(addr, bytes, cache)
122 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
123 /** fdatasync is broken on ext3/ext4fs on older kernels, see
124 * description in #mdb_env_open2 comments. You can safely
125 * define MDB_FDATASYNC_WORKS if this code will only be run
126 * on kernels 3.6 and newer.
128 #define BROKEN_FDATASYNC
134 #include <inttypes.h>
142 typedef SSIZE_T ssize_t;
147 #if defined(__sun) || defined(ANDROID)
148 /* Most platforms have posix_memalign, older may only have memalign */
149 #define HAVE_MEMALIGN 1
151 /* On Solaris, we need the POSIX sigwait function */
153 # define _POSIX_PTHREAD_SEMANTICS 1
157 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
158 #include <netinet/in.h>
159 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
162 #if defined(__APPLE__) || defined (BSD) || defined(__FreeBSD_kernel__)
163 # if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
164 # define MDB_USE_SYSV_SEM 1
166 # define MDB_FDATASYNC fsync
167 #elif defined(ANDROID)
168 # define MDB_FDATASYNC fsync
174 #ifdef MDB_USE_POSIX_SEM
175 # define MDB_USE_HASH 1
176 #include <semaphore.h>
177 #elif defined(MDB_USE_SYSV_SEM)
180 #ifdef _SEM_SEMUN_UNDEFINED
183 struct semid_ds *buf;
184 unsigned short *array;
186 #endif /* _SEM_SEMUN_UNDEFINED */
188 #define MDB_USE_POSIX_MUTEX 1
189 #endif /* MDB_USE_POSIX_SEM */
192 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
193 + defined(MDB_USE_POSIX_MUTEX) != 1
194 # error "Ambiguous shared-lock implementation"
198 #include <valgrind/memcheck.h>
199 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
200 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
201 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
202 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
203 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
205 #define VGMEMP_CREATE(h,r,z)
206 #define VGMEMP_ALLOC(h,a,s)
207 #define VGMEMP_FREE(h,a)
208 #define VGMEMP_DESTROY(h)
209 #define VGMEMP_DEFINED(a,s)
213 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
214 /* Solaris just defines one or the other */
215 # define LITTLE_ENDIAN 1234
216 # define BIG_ENDIAN 4321
217 # ifdef _LITTLE_ENDIAN
218 # define BYTE_ORDER LITTLE_ENDIAN
220 # define BYTE_ORDER BIG_ENDIAN
223 # define BYTE_ORDER __BYTE_ORDER
227 #ifndef LITTLE_ENDIAN
228 #define LITTLE_ENDIAN __LITTLE_ENDIAN
231 #define BIG_ENDIAN __BIG_ENDIAN
234 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
235 #define MISALIGNED_OK 1
241 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
242 # error "Unknown or unsupported endianness (BYTE_ORDER)"
243 #elif (-6 & 5) || CHAR_BIT!=8 || UINT_MAX!=0xffffffff || MDB_SIZE_MAX%UINT_MAX
244 # error "Two's complement, reasonably sized integer types, please"
248 /** Put infrequently used env functions in separate section */
250 # define ESECT __attribute__ ((section("__TEXT,text_env")))
252 # define ESECT __attribute__ ((section("text_env")))
259 #define CALL_CONV WINAPI
264 /** @defgroup internal LMDB Internals
267 /** @defgroup compat Compatibility Macros
268 * A bunch of macros to minimize the amount of platform-specific ifdefs
269 * needed throughout the rest of the code. When the features this library
270 * needs are similar enough to POSIX to be hidden in a one-or-two line
271 * replacement, this macro approach is used.
275 /** Features under development */
280 /** Wrapper around __func__, which is a C99 feature */
281 #if __STDC_VERSION__ >= 199901L
282 # define mdb_func_ __func__
283 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
284 # define mdb_func_ __FUNCTION__
286 /* If a debug message says <mdb_unknown>(), update the #if statements above */
287 # define mdb_func_ "<mdb_unknown>"
290 /* Internal error codes, not exposed outside liblmdb */
291 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
293 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
294 #elif defined MDB_USE_SYSV_SEM
295 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
296 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
297 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
301 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
303 /** Some platforms define the EOWNERDEAD error code
304 * even though they don't support Robust Mutexes.
305 * Compile with -DMDB_USE_ROBUST=0, or use some other
306 * mechanism like -DMDB_USE_SYSV_SEM instead of
307 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
308 * also Robust, but some systems don't support them
311 #ifndef MDB_USE_ROBUST
312 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
313 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
314 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
315 # define MDB_USE_ROBUST 0
317 # define MDB_USE_ROBUST 1
319 #endif /* !MDB_USE_ROBUST */
321 #if defined(MDB_USE_POSIX_MUTEX) && (MDB_USE_ROBUST)
322 /* glibc < 2.12 only provided _np API */
323 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
324 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
325 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
326 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
327 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
329 #endif /* MDB_USE_POSIX_MUTEX && MDB_USE_ROBUST */
331 #if defined(MDB_OWNERDEAD) && (MDB_USE_ROBUST)
332 #define MDB_ROBUST_SUPPORTED 1
336 #define MDB_USE_HASH 1
337 #define MDB_PIDLOCK 0
338 #define THREAD_RET DWORD
339 #define pthread_t HANDLE
340 #define pthread_mutex_t HANDLE
341 #define pthread_cond_t HANDLE
342 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
343 #define pthread_key_t DWORD
344 #define pthread_self() GetCurrentThreadId()
345 #define pthread_key_create(x,y) \
346 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
347 #define pthread_key_delete(x) TlsFree(x)
348 #define pthread_getspecific(x) TlsGetValue(x)
349 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
350 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
351 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
352 #define pthread_cond_signal(x) SetEvent(*x)
353 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
354 #define THREAD_CREATE(thr,start,arg) \
355 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
356 #define THREAD_FINISH(thr) \
357 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
358 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
359 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
360 #define mdb_mutex_consistent(mutex) 0
361 #define getpid() GetCurrentProcessId()
362 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
363 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
364 #define ErrCode() GetLastError()
365 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
366 #define close(fd) (CloseHandle(fd) ? 0 : -1)
367 #define munmap(ptr,len) UnmapViewOfFile(ptr)
368 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
369 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
371 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
374 #define THREAD_RET void *
375 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
376 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
378 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
379 #define MDB_PIDLOCK 1
381 #ifdef MDB_USE_POSIX_SEM
383 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
384 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
385 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
388 mdb_sem_wait(sem_t *sem)
391 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
395 #elif defined MDB_USE_SYSV_SEM
397 typedef struct mdb_mutex {
401 } mdb_mutex_t[1], *mdb_mutexref_t;
403 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
404 #define UNLOCK_MUTEX(mutex) do { \
405 struct sembuf sb = { 0, 1, SEM_UNDO }; \
406 sb.sem_num = (mutex)->semnum; \
407 *(mutex)->locked = 0; \
408 semop((mutex)->semid, &sb, 1); \
412 mdb_sem_wait(mdb_mutexref_t sem)
414 int rc, *locked = sem->locked;
415 struct sembuf sb = { 0, -1, SEM_UNDO };
416 sb.sem_num = sem->semnum;
418 if (!semop(sem->semid, &sb, 1)) {
419 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
423 } while ((rc = errno) == EINTR);
427 #define mdb_mutex_consistent(mutex) 0
429 #else /* MDB_USE_POSIX_MUTEX: */
430 /** Shared mutex/semaphore as the original is stored.
432 * Not for copies. Instead it can be assigned to an #mdb_mutexref_t.
433 * When mdb_mutexref_t is a pointer and mdb_mutex_t is not, then it
434 * is array[size 1] so it can be assigned to the pointer.
436 typedef pthread_mutex_t mdb_mutex_t[1];
437 /** Reference to an #mdb_mutex_t */
438 typedef pthread_mutex_t *mdb_mutexref_t;
439 /** Lock the reader or writer mutex.
440 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
442 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
443 /** Unlock the reader or writer mutex.
445 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
446 /** Mark mutex-protected data as repaired, after death of previous owner.
448 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
449 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
451 /** Get the error code for the last failed system function.
453 #define ErrCode() errno
455 /** An abstraction for a file handle.
456 * On POSIX systems file handles are small integers. On Windows
457 * they're opaque pointers.
461 /** A value for an invalid file handle.
462 * Mainly used to initialize file variables and signify that they are
465 #define INVALID_HANDLE_VALUE (-1)
467 /** Get the size of a memory page for the system.
468 * This is the basic size that the platform's memory manager uses, and is
469 * fundamental to the use of memory-mapped files.
471 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
474 #define Z MDB_FMT_Z /**< printf/scanf format modifier for size_t */
475 #define Yu MDB_PRIy(u) /**< printf format for #mdb_size_t */
476 #define Yd MDB_PRIy(d) /**< printf format for 'signed #mdb_size_t' */
478 #ifdef MDB_USE_SYSV_SEM
479 #define MNAME_LEN (sizeof(int))
481 #define MNAME_LEN (sizeof(pthread_mutex_t))
484 /** Initial part of #MDB_env.me_mutexname[].
485 * Changes to this code must be reflected in #MDB_LOCK_FORMAT.
488 #define MUTEXNAME_PREFIX "Global\\MDB"
489 #elif defined MDB_USE_POSIX_SEM
490 #define MUTEXNAME_PREFIX "/MDB"
495 #ifdef MDB_ROBUST_SUPPORTED
496 /** Lock mutex, handle any error, set rc = result.
497 * Return 0 on success, nonzero (not rc) on error.
499 #define LOCK_MUTEX(rc, env, mutex) \
500 (((rc) = LOCK_MUTEX0(mutex)) && \
501 ((rc) = mdb_mutex_failed(env, mutex, rc)))
502 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
504 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
505 #define mdb_mutex_failed(env, mutex, rc) (rc)
509 /** A flag for opening a file and requesting synchronous data writes.
510 * This is only used when writing a meta page. It's not strictly needed;
511 * we could just do a normal write and then immediately perform a flush.
512 * But if this flag is available it saves us an extra system call.
514 * @note If O_DSYNC is undefined but exists in /usr/include,
515 * preferably set some compiler flag to get the definition.
519 # define MDB_DSYNC O_DSYNC
521 # define MDB_DSYNC O_SYNC
526 /** Function for flushing the data of a file. Define this to fsync
527 * if fdatasync() is not supported.
529 #ifndef MDB_FDATASYNC
530 # define MDB_FDATASYNC fdatasync
534 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
545 /** A page number in the database.
546 * Note that 64 bit page numbers are overkill, since pages themselves
547 * already represent 12-13 bits of addressable memory, and the OS will
548 * always limit applications to a maximum of 63 bits of address space.
550 * @note In the #MDB_node structure, we only store 48 bits of this value,
551 * which thus limits us to only 60 bits of addressable data.
553 typedef MDB_ID pgno_t;
555 /** A transaction ID.
556 * See struct MDB_txn.mt_txnid for details.
558 typedef MDB_ID txnid_t;
560 /** @defgroup debug Debug Macros
564 /** Enable debug output. Needs variable argument macros (a C99 feature).
565 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
566 * read from and written to the database (used for free space management).
572 static int mdb_debug;
573 static txnid_t mdb_debug_start;
575 /** Print a debug message with printf formatting.
576 * Requires double parenthesis around 2 or more args.
578 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
579 # define DPRINTF0(fmt, ...) \
580 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
582 # define DPRINTF(args) ((void) 0)
584 /** Print a debug string.
585 * The string is printed literally, with no format processing.
587 #define DPUTS(arg) DPRINTF(("%s", arg))
588 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
590 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
593 /** @brief The maximum size of a database page.
595 * It is 32k or 64k, since value-PAGEBASE must fit in
596 * #MDB_page.%mp_upper.
598 * LMDB will use database pages < OS pages if needed.
599 * That causes more I/O in write transactions: The OS must
600 * know (read) the whole page before writing a partial page.
602 * Note that we don't currently support Huge pages. On Linux,
603 * regular data files cannot use Huge pages, and in general
604 * Huge pages aren't actually pageable. We rely on the OS
605 * demand-pager to read our data and page it out when memory
606 * pressure from other processes is high. So until OSs have
607 * actual paging support for Huge pages, they're not viable.
609 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
611 /** The minimum number of keys required in a database page.
612 * Setting this to a larger value will place a smaller bound on the
613 * maximum size of a data item. Data items larger than this size will
614 * be pushed into overflow pages instead of being stored directly in
615 * the B-tree node. This value used to default to 4. With a page size
616 * of 4096 bytes that meant that any item larger than 1024 bytes would
617 * go into an overflow page. That also meant that on average 2-3KB of
618 * each overflow page was wasted space. The value cannot be lower than
619 * 2 because then there would no longer be a tree structure. With this
620 * value, items larger than 2KB will go into overflow pages, and on
621 * average only 1KB will be wasted.
623 #define MDB_MINKEYS 2
625 /** A stamp that identifies a file as an LMDB file.
626 * There's nothing special about this value other than that it is easily
627 * recognizable, and it will reflect any byte order mismatches.
629 #define MDB_MAGIC 0xBEEFC0DE
631 /** The version number for a database's datafile format. */
632 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
633 /** The version number for a database's lockfile format. */
634 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 2)
635 /** Number of bits representing #MDB_LOCK_VERSION in #MDB_LOCK_FORMAT.
636 * The remaining bits must leave room for #MDB_lock_desc.
638 #define MDB_LOCK_VERSION_BITS 12
640 /** @brief The max size of a key we can write, or 0 for computed max.
642 * This macro should normally be left alone or set to 0.
643 * Note that a database with big keys or dupsort data cannot be
644 * reliably modified by a liblmdb which uses a smaller max.
645 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
647 * Other values are allowed, for backwards compat. However:
648 * A value bigger than the computed max can break if you do not
649 * know what you are doing, and liblmdb <= 0.9.10 can break when
650 * modifying a DB with keys/dupsort data bigger than its max.
652 * Data items in an #MDB_DUPSORT database are also limited to
653 * this size, since they're actually keys of a sub-DB. Keys and
654 * #MDB_DUPSORT data items must fit on a node in a regular page.
656 #ifndef MDB_MAXKEYSIZE
657 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
660 /** The maximum size of a key we can write to the environment. */
662 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
664 #define ENV_MAXKEY(env) ((env)->me_maxkey)
667 /** @brief The maximum size of a data item.
669 * We only store a 32 bit value for node sizes.
671 #define MAXDATASIZE 0xffffffffUL
674 /** Key size which fits in a #DKBUF.
677 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
680 * This is used for printing a hex dump of a key's contents.
682 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
683 /** Display a key in hex.
685 * Invoke a function to display a key in hex.
687 #define DKEY(x) mdb_dkey(x, kbuf)
693 /** An invalid page number.
694 * Mainly used to denote an empty tree.
696 #define P_INVALID (~(pgno_t)0)
698 /** Test if the flags \b f are set in a flag word \b w. */
699 #define F_ISSET(w, f) (((w) & (f)) == (f))
701 /** Round \b n up to an even number. */
702 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
704 /** Least significant 1-bit of \b n. n must be of an unsigned type. */
705 #define LOW_BIT(n) ((n) & (-(n)))
707 /** (log2(\b p2) % \b n), for p2 = power of 2 and 0 < n < 8. */
708 #define LOG2_MOD(p2, n) (7 - 86 / ((p2) % ((1U<<(n))-1) + 11))
709 /* Explanation: Let p2 = 2**(n*y + x), x<n and M = (1U<<n)-1. Now p2 =
710 * (M+1)**y * 2**x = 2**x (mod M). Finally "/" "happens" to return 7-x.
713 /** Should be alignment of \b type. Ensure it is a power of 2. */
714 #define ALIGNOF2(type) \
715 LOW_BIT(offsetof(struct { char ch_; type align_; }, align_))
717 /** Used for offsets within a single page.
718 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
721 typedef uint16_t indx_t;
723 typedef unsigned long long mdb_hash_t;
725 /** Default size of memory map.
726 * This is certainly too small for any actual applications. Apps should always set
727 * the size explicitly using #mdb_env_set_mapsize().
729 #define DEFAULT_MAPSIZE 1048576
731 /** @defgroup readers Reader Lock Table
732 * Readers don't acquire any locks for their data access. Instead, they
733 * simply record their transaction ID in the reader table. The reader
734 * mutex is needed just to find an empty slot in the reader table. The
735 * slot's address is saved in thread-specific data so that subsequent read
736 * transactions started by the same thread need no further locking to proceed.
738 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
740 * No reader table is used if the database is on a read-only filesystem, or
741 * if #MDB_NOLOCK is set.
743 * Since the database uses multi-version concurrency control, readers don't
744 * actually need any locking. This table is used to keep track of which
745 * readers are using data from which old transactions, so that we'll know
746 * when a particular old transaction is no longer in use. Old transactions
747 * that have discarded any data pages can then have those pages reclaimed
748 * for use by a later write transaction.
750 * The lock table is constructed such that reader slots are aligned with the
751 * processor's cache line size. Any slot is only ever used by one thread.
752 * This alignment guarantees that there will be no contention or cache
753 * thrashing as threads update their own slot info, and also eliminates
754 * any need for locking when accessing a slot.
756 * A writer thread will scan every slot in the table to determine the oldest
757 * outstanding reader transaction. Any freed pages older than this will be
758 * reclaimed by the writer. The writer doesn't use any locks when scanning
759 * this table. This means that there's no guarantee that the writer will
760 * see the most up-to-date reader info, but that's not required for correct
761 * operation - all we need is to know the upper bound on the oldest reader,
762 * we don't care at all about the newest reader. So the only consequence of
763 * reading stale information here is that old pages might hang around a
764 * while longer before being reclaimed. That's actually good anyway, because
765 * the longer we delay reclaiming old pages, the more likely it is that a
766 * string of contiguous pages can be found after coalescing old pages from
767 * many old transactions together.
770 /** Number of slots in the reader table.
771 * This value was chosen somewhat arbitrarily. 126 readers plus a
772 * couple mutexes fit exactly into 8KB on my development machine.
773 * Applications should set the table size using #mdb_env_set_maxreaders().
775 #define DEFAULT_READERS 126
777 /** The size of a CPU cache line in bytes. We want our lock structures
778 * aligned to this size to avoid false cache line sharing in the
780 * This value works for most CPUs. For Itanium this should be 128.
786 /** The information we store in a single slot of the reader table.
787 * In addition to a transaction ID, we also record the process and
788 * thread ID that owns a slot, so that we can detect stale information,
789 * e.g. threads or processes that went away without cleaning up.
790 * @note We currently don't check for stale records. We simply re-init
791 * the table when we know that we're the only process opening the
794 typedef struct MDB_rxbody {
795 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
796 * Multiple readers that start at the same time will probably have the
797 * same ID here. Again, it's not important to exclude them from
798 * anything; all we need to know is which version of the DB they
799 * started from so we can avoid overwriting any data used in that
800 * particular version.
802 volatile txnid_t mrb_txnid;
803 /** The process ID of the process owning this reader txn. */
804 volatile MDB_PID_T mrb_pid;
805 /** The thread ID of the thread owning this txn. */
806 volatile MDB_THR_T mrb_tid;
809 /** The actual reader record, with cacheline padding. */
810 typedef struct MDB_reader {
813 /** shorthand for mrb_txnid */
814 #define mr_txnid mru.mrx.mrb_txnid
815 #define mr_pid mru.mrx.mrb_pid
816 #define mr_tid mru.mrx.mrb_tid
817 /** cache line alignment */
818 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
822 /** The header for the reader table.
823 * The table resides in a memory-mapped file. (This is a different file
824 * than is used for the main database.)
826 * For POSIX the actual mutexes reside in the shared memory of this
827 * mapped file. On Windows, mutexes are named objects allocated by the
828 * kernel; we store the mutex names in this mapped file so that other
829 * processes can grab them. This same approach is also used on
830 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
831 * process-shared POSIX mutexes. For these cases where a named object
832 * is used, the object name is derived from a 64 bit FNV hash of the
833 * environment pathname. As such, naming collisions are extremely
834 * unlikely. If a collision occurs, the results are unpredictable.
836 typedef struct MDB_txbody {
837 /** Stamp identifying this as an LMDB file. It must be set
840 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
842 /** The ID of the last transaction committed to the database.
843 * This is recorded here only for convenience; the value can always
844 * be determined by reading the main database meta pages.
846 volatile txnid_t mtb_txnid;
847 /** The number of slots that have been used in the reader table.
848 * This always records the maximum count, it is not decremented
849 * when readers release their slots.
851 volatile unsigned mtb_numreaders;
852 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
853 /** Binary form of names of the reader/writer locks */
854 mdb_hash_t mtb_mutexid;
855 #elif defined(MDB_USE_SYSV_SEM)
859 /** Mutex protecting access to this table.
860 * This is the reader table lock used with LOCK_MUTEX().
862 mdb_mutex_t mtb_rmutex;
866 /** The actual reader table definition. */
867 typedef struct MDB_txninfo {
870 #define mti_magic mt1.mtb.mtb_magic
871 #define mti_format mt1.mtb.mtb_format
872 #define mti_rmutex mt1.mtb.mtb_rmutex
873 #define mti_txnid mt1.mtb.mtb_txnid
874 #define mti_numreaders mt1.mtb.mtb_numreaders
875 #define mti_mutexid mt1.mtb.mtb_mutexid
876 #ifdef MDB_USE_SYSV_SEM
877 #define mti_semid mt1.mtb.mtb_semid
878 #define mti_rlocked mt1.mtb.mtb_rlocked
880 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
882 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM))
884 #ifdef MDB_USE_SYSV_SEM
886 #define mti_wlocked mt2.mt2_wlocked
888 mdb_mutex_t mt2_wmutex;
889 #define mti_wmutex mt2.mt2_wmutex
891 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
894 MDB_reader mti_readers[1];
897 /** Lockfile format signature: version, features and field layout */
898 #define MDB_LOCK_FORMAT \
900 (((MDB_LOCK_VERSION) % (1U << MDB_LOCK_VERSION_BITS)) \
901 + MDB_lock_desc * (1U << MDB_LOCK_VERSION_BITS)))
903 /** Lock type and layout. Values 0-119. _WIN32 implies #MDB_PIDLOCK.
904 * Some low values are reserved for future tweaks.
907 # define MDB_LOCK_TYPE (0 + ALIGNOF2(mdb_hash_t)/8 % 2)
908 #elif defined MDB_USE_POSIX_SEM
909 # define MDB_LOCK_TYPE (4 + ALIGNOF2(mdb_hash_t)/8 % 2)
910 #elif defined MDB_USE_SYSV_SEM
911 # define MDB_LOCK_TYPE (8)
912 #elif defined MDB_USE_POSIX_MUTEX
913 /* We do not know the inside of a POSIX mutex and how to check if mutexes
914 * used by two executables are compatible. Just check alignment and size.
916 # define MDB_LOCK_TYPE (10 + \
917 LOG2_MOD(ALIGNOF2(pthread_mutex_t), 5) + \
918 sizeof(pthread_mutex_t) / 4U % 22 * 5)
922 /** Magic number for lockfile layout and features.
924 * This *attempts* to stop liblmdb variants compiled with conflicting
925 * options from using the lockfile at the same time and thus breaking
926 * it. It describes locking types, and sizes and sometimes alignment
927 * of the various lockfile items.
929 * The detected ranges are mostly guesswork, or based simply on how
930 * big they could be without using more bits. So we can tweak them
931 * in good conscience when updating #MDB_LOCK_VERSION.
934 /* Default CACHELINE=64 vs. other values (have seen mention of 32-256) */
935 (CACHELINE==64 ? 0 : 1 + LOG2_MOD(CACHELINE >> (CACHELINE>64), 5))
936 + 6 * (sizeof(MDB_PID_T)/4 % 3) /* legacy(2) to word(4/8)? */
937 + 18 * (sizeof(pthread_t)/4 % 5) /* can be struct{id, active data} */
938 + 90 * (sizeof(MDB_txbody) / CACHELINE % 3)
939 + 270 * (MDB_LOCK_TYPE % 120)
940 /* The above is < 270*120 < 2**15 */
941 + ((sizeof(txnid_t) == 8) << 15) /* 32bit/64bit */
942 + ((sizeof(MDB_reader) > CACHELINE) << 16)
943 /* Not really needed - implied by MDB_LOCK_TYPE != (_WIN32 locking) */
944 + (((MDB_PIDLOCK) != 0) << 17)
945 /* 18 bits total: Must be <= (32 - MDB_LOCK_VERSION_BITS). */
949 /** Common header for all page types. The page type depends on #mp_flags.
951 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with
952 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages
953 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header.
955 * #P_OVERFLOW records occupy one or more contiguous pages where only the
956 * first has a page header. They hold the real data of #F_BIGDATA nodes.
958 * #P_SUBP sub-pages are small leaf "pages" with duplicate data.
959 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page.
960 * (Duplicate data can also go in sub-databases, which use normal pages.)
962 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot.
964 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once
965 * in the snapshot: Either used by a database or listed in a freeDB record.
967 typedef struct MDB_page {
968 #define mp_pgno mp_p.p_pgno
969 #define mp_next mp_p.p_next
971 pgno_t p_pgno; /**< page number */
972 struct MDB_page *p_next; /**< for in-memory list of freed pages */
974 uint16_t mp_pad; /**< key size if this is a LEAF2 page */
975 /** @defgroup mdb_page Page Flags
977 * Flags for the page headers.
980 #define P_BRANCH 0x01 /**< branch page */
981 #define P_LEAF 0x02 /**< leaf page */
982 #define P_OVERFLOW 0x04 /**< overflow page */
983 #define P_META 0x08 /**< meta page */
984 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
985 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
986 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
987 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
988 #define P_KEEP 0x8000 /**< leave this page alone during spill */
990 uint16_t mp_flags; /**< @ref mdb_page */
991 #define mp_lower mp_pb.pb.pb_lower
992 #define mp_upper mp_pb.pb.pb_upper
993 #define mp_pages mp_pb.pb_pages
996 indx_t pb_lower; /**< lower bound of free space */
997 indx_t pb_upper; /**< upper bound of free space */
999 uint32_t pb_pages; /**< number of overflow pages */
1001 indx_t mp_ptrs[1]; /**< dynamic size */
1004 /** Size of the page header, excluding dynamic data at the end */
1005 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
1007 /** Address of first usable data byte in a page, after the header */
1008 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
1010 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
1011 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
1013 /** Number of nodes on a page */
1014 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
1016 /** The amount of space remaining in the page */
1017 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
1019 /** The percentage of space used in the page, in tenths of a percent. */
1020 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
1021 ((env)->me_psize - PAGEHDRSZ))
1022 /** The minimum page fill factor, in tenths of a percent.
1023 * Pages emptier than this are candidates for merging.
1025 #define FILL_THRESHOLD 250
1027 /** Test if a page is a leaf page */
1028 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
1029 /** Test if a page is a LEAF2 page */
1030 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
1031 /** Test if a page is a branch page */
1032 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
1033 /** Test if a page is an overflow page */
1034 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
1035 /** Test if a page is a sub page */
1036 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
1038 /** The number of overflow pages needed to store the given size. */
1039 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
1041 /** Link in #MDB_txn.%mt_loose_pgs list.
1042 * Kept outside the page header, which is needed when reusing the page.
1044 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
1046 /** Header for a single key/data pair within a page.
1047 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
1048 * We guarantee 2-byte alignment for 'MDB_node's.
1050 * #mn_lo and #mn_hi are used for data size on leaf nodes, and for child
1051 * pgno on branch nodes. On 64 bit platforms, #mn_flags is also used
1052 * for pgno. (Branch nodes have no flags). Lo and hi are in host byte
1053 * order in case some accesses can be optimized to 32-bit word access.
1055 * Leaf node flags describe node contents. #F_BIGDATA says the node's
1056 * data part is the page number of an overflow page with actual data.
1057 * #F_DUPDATA and #F_SUBDATA can be combined giving duplicate data in
1058 * a sub-page/sub-database, and named databases (just #F_SUBDATA).
1060 typedef struct MDB_node {
1061 /** part of data size or pgno
1063 #if BYTE_ORDER == LITTLE_ENDIAN
1064 unsigned short mn_lo, mn_hi;
1066 unsigned short mn_hi, mn_lo;
1069 /** @defgroup mdb_node Node Flags
1071 * Flags for node headers.
1074 #define F_BIGDATA 0x01 /**< data put on overflow page */
1075 #define F_SUBDATA 0x02 /**< data is a sub-database */
1076 #define F_DUPDATA 0x04 /**< data has duplicates */
1078 /** valid flags for #mdb_node_add() */
1079 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
1082 unsigned short mn_flags; /**< @ref mdb_node */
1083 unsigned short mn_ksize; /**< key size */
1084 char mn_data[1]; /**< key and data are appended here */
1087 /** Size of the node header, excluding dynamic data at the end */
1088 #define NODESIZE offsetof(MDB_node, mn_data)
1090 /** Bit position of top word in page number, for shifting mn_flags */
1091 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
1093 /** Size of a node in a branch page with a given key.
1094 * This is just the node header plus the key, there is no data.
1096 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
1098 /** Size of a node in a leaf page with a given key and data.
1099 * This is node header plus key plus data size.
1101 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
1103 /** Address of node \b i in page \b p */
1104 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
1106 /** Address of the key for the node */
1107 #define NODEKEY(node) (void *)((node)->mn_data)
1109 /** Address of the data for a node */
1110 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
1112 /** Get the page number pointed to by a branch node */
1113 #define NODEPGNO(node) \
1114 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
1115 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
1116 /** Set the page number in a branch node */
1117 #define SETPGNO(node,pgno) do { \
1118 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
1119 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
1121 /** Get the size of the data in a leaf node */
1122 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
1123 /** Set the size of the data for a leaf node */
1124 #define SETDSZ(node,size) do { \
1125 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
1126 /** The size of a key in a node */
1127 #define NODEKSZ(node) ((node)->mn_ksize)
1129 /** Copy a page number from src to dst */
1130 #ifdef MISALIGNED_OK
1131 #define COPY_PGNO(dst,src) dst = src
1133 #if MDB_SIZE_MAX > 0xffffffffU
1134 #define COPY_PGNO(dst,src) do { \
1135 unsigned short *s, *d; \
1136 s = (unsigned short *)&(src); \
1137 d = (unsigned short *)&(dst); \
1144 #define COPY_PGNO(dst,src) do { \
1145 unsigned short *s, *d; \
1146 s = (unsigned short *)&(src); \
1147 d = (unsigned short *)&(dst); \
1153 /** The address of a key in a LEAF2 page.
1154 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
1155 * There are no node headers, keys are stored contiguously.
1157 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
1159 /** Set the \b node's key into \b keyptr, if requested. */
1160 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
1161 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
1163 /** Set the \b node's key into \b key. */
1164 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
1166 /** Information about a single database in the environment. */
1167 typedef struct MDB_db {
1168 uint32_t md_pad; /**< also ksize for LEAF2 pages */
1169 uint16_t md_flags; /**< @ref mdb_dbi_open */
1170 uint16_t md_depth; /**< depth of this tree */
1171 pgno_t md_branch_pages; /**< number of internal pages */
1172 pgno_t md_leaf_pages; /**< number of leaf pages */
1173 pgno_t md_overflow_pages; /**< number of overflow pages */
1174 mdb_size_t md_entries; /**< number of data items */
1175 pgno_t md_root; /**< the root page of this tree */
1178 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1179 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1180 /** #mdb_dbi_open() flags */
1181 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1182 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1184 /** Handle for the DB used to track free pages. */
1186 /** Handle for the default DB. */
1188 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1191 /** Number of meta pages - also hardcoded elsewhere */
1194 /** Meta page content.
1195 * A meta page is the start point for accessing a database snapshot.
1196 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1198 typedef struct MDB_meta {
1199 /** Stamp identifying this as an LMDB file. It must be set
1202 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1203 uint32_t mm_version;
1205 union { /* always zero since we don't support fixed mapping in MDB_VL32 */
1209 #define mm_address mm_un.mmun_address
1211 void *mm_address; /**< address for fixed mapping */
1213 mdb_size_t mm_mapsize; /**< size of mmap region */
1214 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1215 /** The size of pages used in this DB */
1216 #define mm_psize mm_dbs[FREE_DBI].md_pad
1217 /** Any persistent environment flags. @ref mdb_env */
1218 #define mm_flags mm_dbs[FREE_DBI].md_flags
1219 /** Last used page in the datafile.
1220 * Actually the file may be shorter if the freeDB lists the final pages.
1223 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1226 /** Buffer for a stack-allocated meta page.
1227 * The members define size and alignment, and silence type
1228 * aliasing warnings. They are not used directly; that could
1229 * mean incorrectly using several union members in parallel.
1231 typedef union MDB_metabuf {
1234 char mm_pad[PAGEHDRSZ];
1239 /** Auxiliary DB info.
1240 * The information here is mostly static/read-only. There is
1241 * only a single copy of this record in the environment.
1243 typedef struct MDB_dbx {
1244 MDB_val md_name; /**< name of the database */
1245 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1246 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1247 MDB_rel_func *md_rel; /**< user relocate function */
1248 void *md_relctx; /**< user-provided context for md_rel */
1251 /** A database transaction.
1252 * Every operation requires a transaction handle.
1255 MDB_txn *mt_parent; /**< parent of a nested txn */
1256 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1258 pgno_t mt_next_pgno; /**< next unallocated page */
1260 pgno_t mt_last_pgno; /**< last written page */
1262 /** The ID of this transaction. IDs are integers incrementing from 1.
1263 * Only committed write transactions increment the ID. If a transaction
1264 * aborts, the ID may be re-used by the next writer.
1267 MDB_env *mt_env; /**< the DB environment */
1268 /** The list of pages that became unused during this transaction.
1270 MDB_IDL mt_free_pgs;
1271 /** The list of loose pages that became unused and may be reused
1272 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1274 MDB_page *mt_loose_pgs;
1275 /** Number of loose pages (#mt_loose_pgs) */
1277 /** The sorted list of dirty pages we temporarily wrote to disk
1278 * because the dirty list was full. page numbers in here are
1279 * shifted left by 1, deleted slots have the LSB set.
1281 MDB_IDL mt_spill_pgs;
1283 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1284 MDB_ID2L dirty_list;
1285 /** For read txns: This thread/txn's reader table slot, or NULL. */
1288 /** Array of records for each DB known in the environment. */
1290 /** Array of MDB_db records for each known DB */
1292 /** Array of sequence numbers for each DB handle */
1293 unsigned int *mt_dbiseqs;
1294 /** @defgroup mt_dbflag Transaction DB Flags
1298 #define DB_DIRTY 0x01 /**< DB was written in this txn */
1299 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1300 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1301 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1302 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1303 #define DB_DUPDATA 0x20 /**< DB is #MDB_DUPSORT data */
1305 /** In write txns, array of cursors for each DB */
1306 MDB_cursor **mt_cursors;
1307 /** Array of flags for each DB */
1308 unsigned char *mt_dbflags;
1310 /** List of read-only pages (actually chunks) */
1312 /** We map chunks of 16 pages. Even though Windows uses 4KB pages, all
1313 * mappings must begin on 64KB boundaries. So we round off all pgnos to
1314 * a chunk boundary. We do the same on Linux for symmetry, and also to
1315 * reduce the frequency of mmap/munmap calls.
1317 #define MDB_RPAGE_CHUNK 16
1318 #define MDB_TRPAGE_SIZE 4096 /**< size of #mt_rpages array of chunks */
1319 #define MDB_TRPAGE_MAX (MDB_TRPAGE_SIZE-1) /**< maximum chunk index */
1320 unsigned int mt_rpcheck; /**< threshold for reclaiming unref'd chunks */
1322 /** Number of DB records in use, or 0 when the txn is finished.
1323 * This number only ever increments until the txn finishes; we
1324 * don't decrement it when individual DB handles are closed.
1328 /** @defgroup mdb_txn Transaction Flags
1332 /** #mdb_txn_begin() flags */
1333 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1334 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1335 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1336 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1337 /* internal txn flags */
1338 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1339 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1340 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1341 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1342 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1343 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1344 /** most operations on the txn are currently illegal */
1345 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1347 unsigned int mt_flags; /**< @ref mdb_txn */
1348 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1349 * Includes ancestor txns' dirty pages not hidden by other txns'
1350 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1351 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1353 unsigned int mt_dirty_room;
1356 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1357 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1358 * raise this on a 64 bit machine.
1360 #define CURSOR_STACK 32
1364 /** Cursors are used for all DB operations.
1365 * A cursor holds a path of (page pointer, key index) from the DB
1366 * root to a position in the DB, plus other state. #MDB_DUPSORT
1367 * cursors include an xcursor to the current data item. Write txns
1368 * track their cursors and keep them up to date when data moves.
1369 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1370 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1373 /** Next cursor on this DB in this txn */
1374 MDB_cursor *mc_next;
1375 /** Backup of the original cursor if this cursor is a shadow */
1376 MDB_cursor *mc_backup;
1377 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1378 struct MDB_xcursor *mc_xcursor;
1379 /** The transaction that owns this cursor */
1381 /** The database handle this cursor operates on */
1383 /** The database record for this cursor */
1385 /** The database auxiliary record for this cursor */
1387 /** The @ref mt_dbflag for this database */
1388 unsigned char *mc_dbflag;
1389 unsigned short mc_snum; /**< number of pushed pages */
1390 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1391 /** @defgroup mdb_cursor Cursor Flags
1393 * Cursor state flags.
1396 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1397 #define C_EOF 0x02 /**< No more data */
1398 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1399 #define C_DEL 0x08 /**< last op was a cursor_del */
1400 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1401 #define C_WRITEMAP MDB_TXN_WRITEMAP /**< Copy of txn flag */
1402 /** Read-only cursor into the txn's original snapshot in the map.
1403 * Set for read-only txns, and in #mdb_page_alloc() for #FREE_DBI when
1404 * #MDB_DEVEL & 2. Only implements code which is necessary for this.
1406 #define C_ORIG_RDONLY MDB_TXN_RDONLY
1408 unsigned int mc_flags; /**< @ref mdb_cursor */
1409 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1410 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1412 MDB_page *mc_ovpg; /**< a referenced overflow page */
1413 # define MC_OVPG(mc) ((mc)->mc_ovpg)
1414 # define MC_SET_OVPG(mc, pg) ((mc)->mc_ovpg = (pg))
1416 # define MC_OVPG(mc) ((MDB_page *)0)
1417 # define MC_SET_OVPG(mc, pg) ((void)0)
1421 /** Context for sorted-dup records.
1422 * We could have gone to a fully recursive design, with arbitrarily
1423 * deep nesting of sub-databases. But for now we only handle these
1424 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1426 typedef struct MDB_xcursor {
1427 /** A sub-cursor for traversing the Dup DB */
1428 MDB_cursor mx_cursor;
1429 /** The database record for this Dup DB */
1431 /** The auxiliary DB record for this Dup DB */
1433 /** The @ref mt_dbflag for this Dup DB */
1434 unsigned char mx_dbflag;
1437 /** Check if there is an inited xcursor */
1438 #define XCURSOR_INITED(mc) \
1439 ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
1441 /** Update the xcursor's sub-page pointer, if any, in \b mc. Needed
1442 * when the node which contains the sub-page may have moved. Called
1443 * with leaf page \b mp = mc->mc_pg[\b top].
1445 #define XCURSOR_REFRESH(mc, top, mp) do { \
1446 MDB_page *xr_pg = (mp); \
1447 MDB_node *xr_node; \
1448 if (!XCURSOR_INITED(mc) || (mc)->mc_ki[top] >= NUMKEYS(xr_pg)) break; \
1449 xr_node = NODEPTR(xr_pg, (mc)->mc_ki[top]); \
1450 if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \
1451 (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \
1454 /** State of FreeDB old pages, stored in the MDB_env */
1455 typedef struct MDB_pgstate {
1456 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1457 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1460 /** The database environment. */
1462 HANDLE me_fd; /**< The main data file */
1463 HANDLE me_lfd; /**< The lock file */
1464 HANDLE me_mfd; /**< For writing and syncing the meta pages */
1465 #if defined(MDB_VL32) && defined(_WIN32)
1466 HANDLE me_fmh; /**< File Mapping handle */
1468 /** Failed to update the meta page. Probably an I/O error. */
1469 #define MDB_FATAL_ERROR 0x80000000U
1470 /** Some fields are initialized. */
1471 #define MDB_ENV_ACTIVE 0x20000000U
1472 /** me_txkey is set */
1473 #define MDB_ENV_TXKEY 0x10000000U
1474 /** fdatasync is unreliable */
1475 #define MDB_FSYNCONLY 0x08000000U
1476 uint32_t me_flags; /**< @ref mdb_env */
1477 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1478 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1479 unsigned int me_maxreaders; /**< size of the reader table */
1480 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1481 volatile int me_close_readers;
1482 MDB_dbi me_numdbs; /**< number of DBs opened */
1483 MDB_dbi me_maxdbs; /**< size of the DB table */
1484 MDB_PID_T me_pid; /**< process ID of this env */
1485 char *me_path; /**< path to the DB files */
1486 char *me_map; /**< the memory map of the data file */
1487 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1488 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1489 void *me_pbuf; /**< scratch area for DUPSORT put() */
1490 MDB_txn *me_txn; /**< current write transaction */
1491 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1492 mdb_size_t me_mapsize; /**< size of the data memory map */
1493 off_t me_size; /**< current file size */
1494 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1495 MDB_dbx *me_dbxs; /**< array of static DB info */
1496 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1497 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1498 pthread_key_t me_txkey; /**< thread-key for readers */
1499 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1500 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1501 # define me_pglast me_pgstate.mf_pglast
1502 # define me_pghead me_pgstate.mf_pghead
1503 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1504 /** IDL of pages that became unused in a write txn */
1505 MDB_IDL me_free_pgs;
1506 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1507 MDB_ID2L me_dirty_list;
1508 /** Max number of freelist items that can fit in a single overflow page */
1510 /** Max size of a node on a page */
1511 unsigned int me_nodemax;
1512 #if !(MDB_MAXKEYSIZE)
1513 unsigned int me_maxkey; /**< max size of a key */
1515 int me_live_reader; /**< have liveness lock in reader table */
1517 int me_pidquery; /**< Used in OpenProcess */
1519 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1520 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1521 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1523 mdb_mutex_t me_rmutex;
1524 mdb_mutex_t me_wmutex;
1525 # if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
1526 /** Half-initialized name of mutexes, to be completed by #MUTEXNAME() */
1527 char me_mutexname[sizeof(MUTEXNAME_PREFIX) + 11];
1531 MDB_ID3L me_rpages; /**< like #mt_rpages, but global to env */
1532 pthread_mutex_t me_rpmutex; /**< control access to #me_rpages */
1533 #define MDB_ERPAGE_SIZE 16384
1534 #define MDB_ERPAGE_MAX (MDB_ERPAGE_SIZE-1)
1535 unsigned int me_rpcheck;
1537 void *me_userctx; /**< User-settable context */
1538 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1541 /** Nested transaction */
1542 typedef struct MDB_ntxn {
1543 MDB_txn mnt_txn; /**< the transaction */
1544 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1547 /** max number of pages to commit in one writev() call */
1548 #define MDB_COMMIT_PAGES 64
1549 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1550 #undef MDB_COMMIT_PAGES
1551 #define MDB_COMMIT_PAGES IOV_MAX
1554 /** max bytes to write in one call */
1555 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1557 /** Check \b txn and \b dbi arguments to a function */
1558 #define TXN_DBI_EXIST(txn, dbi, validity) \
1559 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1561 /** Check for misused \b dbi handles */
1562 #define TXN_DBI_CHANGED(txn, dbi) \
1563 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1565 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1566 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1567 static int mdb_page_touch(MDB_cursor *mc);
1569 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1570 "reset-tmp", "fail-begin", "fail-beginchild"}
1572 /* mdb_txn_end operation number, for logging */
1573 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1574 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1576 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1577 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1578 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1579 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1580 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1582 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1583 static int mdb_page_search_root(MDB_cursor *mc,
1584 MDB_val *key, int modify);
1585 #define MDB_PS_MODIFY 1
1586 #define MDB_PS_ROOTONLY 2
1587 #define MDB_PS_FIRST 4
1588 #define MDB_PS_LAST 8
1589 static int mdb_page_search(MDB_cursor *mc,
1590 MDB_val *key, int flags);
1591 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1593 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1594 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1595 pgno_t newpgno, unsigned int nflags);
1597 static int mdb_env_read_header(MDB_env *env, int prev, MDB_meta *meta);
1598 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1599 static int mdb_env_write_meta(MDB_txn *txn);
1600 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1601 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1603 static void mdb_env_close0(MDB_env *env, int excl);
1605 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1606 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1607 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1608 static void mdb_node_del(MDB_cursor *mc, int ksize);
1609 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1610 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1611 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1612 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1613 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1615 static int mdb_rebalance(MDB_cursor *mc);
1616 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1618 static void mdb_cursor_pop(MDB_cursor *mc);
1619 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1621 static int mdb_cursor_del0(MDB_cursor *mc);
1622 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1623 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1624 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1625 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1626 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1628 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1629 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1631 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1632 static void mdb_xcursor_init0(MDB_cursor *mc);
1633 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1634 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1636 static int mdb_drop0(MDB_cursor *mc, int subs);
1637 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1638 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1641 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1644 /** Compare two items pointing at '#mdb_size_t's of unknown alignment. */
1645 #ifdef MISALIGNED_OK
1646 # define mdb_cmp_clong mdb_cmp_long
1648 # define mdb_cmp_clong mdb_cmp_cint
1651 /** True if we need #mdb_cmp_clong() instead of \b cmp for #MDB_INTEGERDUP */
1652 #define NEED_CMP_CLONG(cmp, ksize) \
1653 (UINT_MAX < MDB_SIZE_MAX && \
1654 (cmp) == mdb_cmp_int && (ksize) == sizeof(mdb_size_t))
1657 static SECURITY_DESCRIPTOR mdb_null_sd;
1658 static SECURITY_ATTRIBUTES mdb_all_sa;
1659 static int mdb_sec_inited;
1662 static int utf8_to_utf16(const char *src, struct MDB_name *dst, int xtra);
1665 /** Return the library version info. */
1667 mdb_version(int *major, int *minor, int *patch)
1669 if (major) *major = MDB_VERSION_MAJOR;
1670 if (minor) *minor = MDB_VERSION_MINOR;
1671 if (patch) *patch = MDB_VERSION_PATCH;
1672 return MDB_VERSION_STRING;
1675 /** Table of descriptions for LMDB @ref errors */
1676 static char *const mdb_errstr[] = {
1677 "MDB_KEYEXIST: Key/data pair already exists",
1678 "MDB_NOTFOUND: No matching key/data pair found",
1679 "MDB_PAGE_NOTFOUND: Requested page not found",
1680 "MDB_CORRUPTED: Located page was wrong type",
1681 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1682 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1683 "MDB_INVALID: File is not an LMDB file",
1684 "MDB_MAP_FULL: Environment mapsize limit reached",
1685 "MDB_DBS_FULL: Environment maxdbs limit reached",
1686 "MDB_READERS_FULL: Environment maxreaders limit reached",
1687 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1688 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1689 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1690 "MDB_PAGE_FULL: Internal error - page has no more space",
1691 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1692 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1693 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1694 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1695 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1696 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1697 "MDB_PROBLEM: Unexpected problem - txn should abort",
1701 mdb_strerror(int err)
1704 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1705 * This works as long as no function between the call to mdb_strerror
1706 * and the actual use of the message uses more than 4K of stack.
1708 #define MSGSIZE 1024
1709 #define PADSIZE 4096
1710 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1714 return ("Successful return: 0");
1716 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1717 i = err - MDB_KEYEXIST;
1718 return mdb_errstr[i];
1722 /* These are the C-runtime error codes we use. The comment indicates
1723 * their numeric value, and the Win32 error they would correspond to
1724 * if the error actually came from a Win32 API. A major mess, we should
1725 * have used LMDB-specific error codes for everything.
1728 case ENOENT: /* 2, FILE_NOT_FOUND */
1729 case EIO: /* 5, ACCESS_DENIED */
1730 case ENOMEM: /* 12, INVALID_ACCESS */
1731 case EACCES: /* 13, INVALID_DATA */
1732 case EBUSY: /* 16, CURRENT_DIRECTORY */
1733 case EINVAL: /* 22, BAD_COMMAND */
1734 case ENOSPC: /* 28, OUT_OF_PAPER */
1735 return strerror(err);
1740 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1741 FORMAT_MESSAGE_IGNORE_INSERTS,
1742 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1745 return strerror(err);
1749 /** assert(3) variant in cursor context */
1750 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1751 /** assert(3) variant in transaction context */
1752 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1753 /** assert(3) variant in environment context */
1754 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1757 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1758 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1761 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1762 const char *func, const char *file, int line)
1765 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1766 file, line, expr_txt, func);
1767 if (env->me_assert_func)
1768 env->me_assert_func(env, buf);
1769 fprintf(stderr, "%s\n", buf);
1773 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1777 /** Return the page number of \b mp which may be sub-page, for debug output */
1779 mdb_dbg_pgno(MDB_page *mp)
1782 COPY_PGNO(ret, mp->mp_pgno);
1786 /** Display a key in hexadecimal and return the address of the result.
1787 * @param[in] key the key to display
1788 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1789 * @return The key in hexadecimal form.
1792 mdb_dkey(MDB_val *key, char *buf)
1795 unsigned char *c = key->mv_data;
1801 if (key->mv_size > DKBUF_MAXKEYSIZE)
1802 return "MDB_MAXKEYSIZE";
1803 /* may want to make this a dynamic check: if the key is mostly
1804 * printable characters, print it as-is instead of converting to hex.
1808 for (i=0; i<key->mv_size; i++)
1809 ptr += sprintf(ptr, "%02x", *c++);
1811 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1817 mdb_leafnode_type(MDB_node *n)
1819 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1820 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1821 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1824 /** Display all the keys in the page. */
1826 mdb_page_list(MDB_page *mp)
1828 pgno_t pgno = mdb_dbg_pgno(mp);
1829 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1831 unsigned int i, nkeys, nsize, total = 0;
1835 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1836 case P_BRANCH: type = "Branch page"; break;
1837 case P_LEAF: type = "Leaf page"; break;
1838 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1839 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1840 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1842 fprintf(stderr, "Overflow page %"Yu" pages %u%s\n",
1843 pgno, mp->mp_pages, state);
1846 fprintf(stderr, "Meta-page %"Yu" txnid %"Yu"\n",
1847 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1850 fprintf(stderr, "Bad page %"Yu" flags 0x%X\n", pgno, mp->mp_flags);
1854 nkeys = NUMKEYS(mp);
1855 fprintf(stderr, "%s %"Yu" numkeys %d%s\n", type, pgno, nkeys, state);
1857 for (i=0; i<nkeys; i++) {
1858 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1859 key.mv_size = nsize = mp->mp_pad;
1860 key.mv_data = LEAF2KEY(mp, i, nsize);
1862 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1865 node = NODEPTR(mp, i);
1866 key.mv_size = node->mn_ksize;
1867 key.mv_data = node->mn_data;
1868 nsize = NODESIZE + key.mv_size;
1869 if (IS_BRANCH(mp)) {
1870 fprintf(stderr, "key %d: page %"Yu", %s\n", i, NODEPGNO(node),
1874 if (F_ISSET(node->mn_flags, F_BIGDATA))
1875 nsize += sizeof(pgno_t);
1877 nsize += NODEDSZ(node);
1879 nsize += sizeof(indx_t);
1880 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1881 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1883 total = EVEN(total);
1885 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1886 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1890 mdb_cursor_chk(MDB_cursor *mc)
1896 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1897 for (i=0; i<mc->mc_top; i++) {
1899 node = NODEPTR(mp, mc->mc_ki[i]);
1900 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1903 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1905 if (XCURSOR_INITED(mc)) {
1906 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1907 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1908 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1916 /** Count all the pages in each DB and in the freelist
1917 * and make sure it matches the actual number of pages
1919 * All named DBs must be open for a correct count.
1921 static void mdb_audit(MDB_txn *txn)
1925 MDB_ID freecount, count;
1930 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1931 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1932 freecount += *(MDB_ID *)data.mv_data;
1933 mdb_tassert(txn, rc == MDB_NOTFOUND);
1936 for (i = 0; i<txn->mt_numdbs; i++) {
1938 if (!(txn->mt_dbflags[i] & DB_VALID))
1940 mdb_cursor_init(&mc, txn, i, &mx);
1941 if (txn->mt_dbs[i].md_root == P_INVALID)
1943 count += txn->mt_dbs[i].md_branch_pages +
1944 txn->mt_dbs[i].md_leaf_pages +
1945 txn->mt_dbs[i].md_overflow_pages;
1946 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1947 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1948 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1951 mp = mc.mc_pg[mc.mc_top];
1952 for (j=0; j<NUMKEYS(mp); j++) {
1953 MDB_node *leaf = NODEPTR(mp, j);
1954 if (leaf->mn_flags & F_SUBDATA) {
1956 memcpy(&db, NODEDATA(leaf), sizeof(db));
1957 count += db.md_branch_pages + db.md_leaf_pages +
1958 db.md_overflow_pages;
1962 mdb_tassert(txn, rc == MDB_NOTFOUND);
1965 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1966 fprintf(stderr, "audit: %"Yu" freecount: %"Yu" count: %"Yu" total: %"Yu" next_pgno: %"Yu"\n",
1967 txn->mt_txnid, freecount, count+NUM_METAS,
1968 freecount+count+NUM_METAS, txn->mt_next_pgno);
1974 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1976 return txn->mt_dbxs[dbi].md_cmp(a, b);
1980 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1982 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1983 if (NEED_CMP_CLONG(dcmp, a->mv_size))
1984 dcmp = mdb_cmp_clong;
1988 /** Allocate memory for a page.
1989 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1990 * Set #MDB_TXN_ERROR on failure.
1993 mdb_page_malloc(MDB_txn *txn, unsigned num)
1995 MDB_env *env = txn->mt_env;
1996 MDB_page *ret = env->me_dpages;
1997 size_t psize = env->me_psize, sz = psize, off;
1998 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1999 * For a single page alloc, we init everything after the page header.
2000 * For multi-page, we init the final page; if the caller needed that
2001 * many pages they will be filling in at least up to the last page.
2005 VGMEMP_ALLOC(env, ret, sz);
2006 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
2007 env->me_dpages = ret->mp_next;
2010 psize -= off = PAGEHDRSZ;
2015 if ((ret = malloc(sz)) != NULL) {
2016 VGMEMP_ALLOC(env, ret, sz);
2017 if (!(env->me_flags & MDB_NOMEMINIT)) {
2018 memset((char *)ret + off, 0, psize);
2022 txn->mt_flags |= MDB_TXN_ERROR;
2026 /** Free a single page.
2027 * Saves single pages to a list, for future reuse.
2028 * (This is not used for multi-page overflow pages.)
2031 mdb_page_free(MDB_env *env, MDB_page *mp)
2033 mp->mp_next = env->me_dpages;
2034 VGMEMP_FREE(env, mp);
2035 env->me_dpages = mp;
2038 /** Free a dirty page */
2040 mdb_dpage_free(MDB_env *env, MDB_page *dp)
2042 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2043 mdb_page_free(env, dp);
2045 /* large pages just get freed directly */
2046 VGMEMP_FREE(env, dp);
2051 /** Return all dirty pages to dpage list */
2053 mdb_dlist_free(MDB_txn *txn)
2055 MDB_env *env = txn->mt_env;
2056 MDB_ID2L dl = txn->mt_u.dirty_list;
2057 unsigned i, n = dl[0].mid;
2059 for (i = 1; i <= n; i++) {
2060 mdb_dpage_free(env, dl[i].mptr);
2067 mdb_page_unref(MDB_txn *txn, MDB_page *mp)
2070 MDB_ID3L tl = txn->mt_rpages;
2072 if (mp->mp_flags & (P_SUBP|P_DIRTY))
2074 rem = mp->mp_pgno & (MDB_RPAGE_CHUNK-1);
2075 pgno = mp->mp_pgno ^ rem;
2076 x = mdb_mid3l_search(tl, pgno);
2077 if (x != tl[0].mid && tl[x+1].mid == mp->mp_pgno)
2082 #define MDB_PAGE_UNREF(txn, mp) mdb_page_unref(txn, mp)
2085 mdb_cursor_unref(MDB_cursor *mc)
2088 if (mc->mc_txn->mt_rpages[0].mid) {
2089 if (!mc->mc_snum || !mc->mc_pg[0] || IS_SUBP(mc->mc_pg[0]))
2091 for (i=0; i<mc->mc_snum; i++)
2092 mdb_page_unref(mc->mc_txn, mc->mc_pg[i]);
2094 mdb_page_unref(mc->mc_txn, mc->mc_ovpg);
2098 mc->mc_snum = mc->mc_top = 0;
2099 mc->mc_pg[0] = NULL;
2100 mc->mc_flags &= ~C_INITIALIZED;
2102 #define MDB_CURSOR_UNREF(mc, force) \
2103 (((force) || ((mc)->mc_flags & C_INITIALIZED)) \
2104 ? mdb_cursor_unref(mc) \
2108 #define MDB_PAGE_UNREF(txn, mp)
2109 #define MDB_CURSOR_UNREF(mc, force) ((void)0)
2110 #endif /* MDB_VL32 */
2112 /** Loosen or free a single page.
2113 * Saves single pages to a list for future reuse
2114 * in this same txn. It has been pulled from the freeDB
2115 * and already resides on the dirty list, but has been
2116 * deleted. Use these pages first before pulling again
2119 * If the page wasn't dirtied in this txn, just add it
2120 * to this txn's free list.
2123 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
2126 pgno_t pgno = mp->mp_pgno;
2127 MDB_txn *txn = mc->mc_txn;
2129 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
2130 if (txn->mt_parent) {
2131 MDB_ID2 *dl = txn->mt_u.dirty_list;
2132 /* If txn has a parent, make sure the page is in our
2136 unsigned x = mdb_mid2l_search(dl, pgno);
2137 if (x <= dl[0].mid && dl[x].mid == pgno) {
2138 if (mp != dl[x].mptr) { /* bad cursor? */
2139 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2140 txn->mt_flags |= MDB_TXN_ERROR;
2148 /* no parent txn, so it's just ours */
2153 DPRINTF(("loosen db %d page %"Yu, DDBI(mc), mp->mp_pgno));
2154 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
2155 txn->mt_loose_pgs = mp;
2156 txn->mt_loose_count++;
2157 mp->mp_flags |= P_LOOSE;
2159 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
2167 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
2168 * @param[in] mc A cursor handle for the current operation.
2169 * @param[in] pflags Flags of the pages to update:
2170 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
2171 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
2172 * @return 0 on success, non-zero on failure.
2175 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
2177 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
2178 MDB_txn *txn = mc->mc_txn;
2179 MDB_cursor *m3, *m0 = mc;
2184 int rc = MDB_SUCCESS, level;
2186 /* Mark pages seen by cursors: First m0, then tracked cursors */
2187 for (i = txn->mt_numdbs;; ) {
2188 if (mc->mc_flags & C_INITIALIZED) {
2189 for (m3 = mc;; m3 = &mx->mx_cursor) {
2191 for (j=0; j<m3->mc_snum; j++) {
2193 if ((mp->mp_flags & Mask) == pflags)
2194 mp->mp_flags ^= P_KEEP;
2196 mx = m3->mc_xcursor;
2197 /* Proceed to mx if it is at a sub-database */
2198 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
2200 if (! (mp && (mp->mp_flags & P_LEAF)))
2202 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
2203 if (!(leaf->mn_flags & F_SUBDATA))
2208 for (; !mc || mc == m0; mc = txn->mt_cursors[--i])
2215 /* Mark dirty root pages */
2216 for (i=0; i<txn->mt_numdbs; i++) {
2217 if (txn->mt_dbflags[i] & DB_DIRTY) {
2218 pgno_t pgno = txn->mt_dbs[i].md_root;
2219 if (pgno == P_INVALID)
2221 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
2223 if ((dp->mp_flags & Mask) == pflags && level <= 1)
2224 dp->mp_flags ^= P_KEEP;
2232 static int mdb_page_flush(MDB_txn *txn, int keep);
2234 /** Spill pages from the dirty list back to disk.
2235 * This is intended to prevent running into #MDB_TXN_FULL situations,
2236 * but note that they may still occur in a few cases:
2237 * 1) our estimate of the txn size could be too small. Currently this
2238 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
2239 * 2) child txns may run out of space if their parents dirtied a
2240 * lot of pages and never spilled them. TODO: we probably should do
2241 * a preemptive spill during #mdb_txn_begin() of a child txn, if
2242 * the parent's dirty_room is below a given threshold.
2244 * Otherwise, if not using nested txns, it is expected that apps will
2245 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
2246 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
2247 * If the txn never references them again, they can be left alone.
2248 * If the txn only reads them, they can be used without any fuss.
2249 * If the txn writes them again, they can be dirtied immediately without
2250 * going thru all of the work of #mdb_page_touch(). Such references are
2251 * handled by #mdb_page_unspill().
2253 * Also note, we never spill DB root pages, nor pages of active cursors,
2254 * because we'll need these back again soon anyway. And in nested txns,
2255 * we can't spill a page in a child txn if it was already spilled in a
2256 * parent txn. That would alter the parent txns' data even though
2257 * the child hasn't committed yet, and we'd have no way to undo it if
2258 * the child aborted.
2260 * @param[in] m0 cursor A cursor handle identifying the transaction and
2261 * database for which we are checking space.
2262 * @param[in] key For a put operation, the key being stored.
2263 * @param[in] data For a put operation, the data being stored.
2264 * @return 0 on success, non-zero on failure.
2267 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
2269 MDB_txn *txn = m0->mc_txn;
2271 MDB_ID2L dl = txn->mt_u.dirty_list;
2272 unsigned int i, j, need;
2275 if (m0->mc_flags & C_SUB)
2278 /* Estimate how much space this op will take */
2279 i = m0->mc_db->md_depth;
2280 /* Named DBs also dirty the main DB */
2281 if (m0->mc_dbi >= CORE_DBS)
2282 i += txn->mt_dbs[MAIN_DBI].md_depth;
2283 /* For puts, roughly factor in the key+data size */
2285 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2286 i += i; /* double it for good measure */
2289 if (txn->mt_dirty_room > i)
2292 if (!txn->mt_spill_pgs) {
2293 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2294 if (!txn->mt_spill_pgs)
2297 /* purge deleted slots */
2298 MDB_IDL sl = txn->mt_spill_pgs;
2299 unsigned int num = sl[0];
2301 for (i=1; i<=num; i++) {
2308 /* Preserve pages which may soon be dirtied again */
2309 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2312 /* Less aggressive spill - we originally spilled the entire dirty list,
2313 * with a few exceptions for cursor pages and DB root pages. But this
2314 * turns out to be a lot of wasted effort because in a large txn many
2315 * of those pages will need to be used again. So now we spill only 1/8th
2316 * of the dirty pages. Testing revealed this to be a good tradeoff,
2317 * better than 1/2, 1/4, or 1/10.
2319 if (need < MDB_IDL_UM_MAX / 8)
2320 need = MDB_IDL_UM_MAX / 8;
2322 /* Save the page IDs of all the pages we're flushing */
2323 /* flush from the tail forward, this saves a lot of shifting later on. */
2324 for (i=dl[0].mid; i && need; i--) {
2325 MDB_ID pn = dl[i].mid << 1;
2327 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2329 /* Can't spill twice, make sure it's not already in a parent's
2332 if (txn->mt_parent) {
2334 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2335 if (tx2->mt_spill_pgs) {
2336 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2337 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2338 dp->mp_flags |= P_KEEP;
2346 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2350 mdb_midl_sort(txn->mt_spill_pgs);
2352 /* Flush the spilled part of dirty list */
2353 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2356 /* Reset any dirty pages we kept that page_flush didn't see */
2357 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2360 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2364 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2366 mdb_find_oldest(MDB_txn *txn)
2369 txnid_t mr, oldest = txn->mt_txnid - 1;
2370 if (txn->mt_env->me_txns) {
2371 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2372 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2383 /** Add a page to the txn's dirty list */
2385 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2388 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2390 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2391 insert = mdb_mid2l_append;
2393 insert = mdb_mid2l_insert;
2395 mid.mid = mp->mp_pgno;
2397 rc = insert(txn->mt_u.dirty_list, &mid);
2398 mdb_tassert(txn, rc == 0);
2399 txn->mt_dirty_room--;
2402 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2403 * me_pghead and mt_next_pgno. Set #MDB_TXN_ERROR on failure.
2405 * If there are free pages available from older transactions, they
2406 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2407 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2408 * and move me_pglast to say which records were consumed. Only this
2409 * function can create me_pghead and move me_pglast/mt_next_pgno.
2410 * When #MDB_DEVEL & 2, it is not affected by #mdb_freelist_save(): it
2411 * then uses the transaction's original snapshot of the freeDB.
2412 * @param[in] mc cursor A cursor handle identifying the transaction and
2413 * database for which we are allocating.
2414 * @param[in] num the number of pages to allocate.
2415 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2416 * will always be satisfied by a single contiguous chunk of memory.
2417 * @return 0 on success, non-zero on failure.
2420 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2422 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2423 /* Get at most <Max_retries> more freeDB records once me_pghead
2424 * has enough pages. If not enough, use new pages from the map.
2425 * If <Paranoid> and mc is updating the freeDB, only get new
2426 * records if me_pghead is empty. Then the freelist cannot play
2427 * catch-up with itself by growing while trying to save it.
2429 enum { Paranoid = 1, Max_retries = 500 };
2431 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2433 int rc, retry = num * 60;
2434 MDB_txn *txn = mc->mc_txn;
2435 MDB_env *env = txn->mt_env;
2436 pgno_t pgno, *mop = env->me_pghead;
2437 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2439 txnid_t oldest = 0, last;
2444 /* If there are any loose pages, just use them */
2445 if (num == 1 && txn->mt_loose_pgs) {
2446 np = txn->mt_loose_pgs;
2447 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2448 txn->mt_loose_count--;
2449 DPRINTF(("db %d use loose page %"Yu, DDBI(mc), np->mp_pgno));
2456 /* If our dirty list is already full, we can't do anything */
2457 if (txn->mt_dirty_room == 0) {
2462 for (op = MDB_FIRST;; op = MDB_NEXT) {
2467 /* Seek a big enough contiguous page range. Prefer
2468 * pages at the tail, just truncating the list.
2474 if (mop[i-n2] == pgno+n2)
2481 if (op == MDB_FIRST) { /* 1st iteration */
2482 /* Prepare to fetch more and coalesce */
2483 last = env->me_pglast;
2484 oldest = env->me_pgoldest;
2485 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2486 #if (MDB_DEVEL) & 2 /* "& 2" so MDB_DEVEL=1 won't hide bugs breaking freeDB */
2487 /* Use original snapshot. TODO: Should need less care in code
2488 * which modifies the database. Maybe we can delete some code?
2490 m2.mc_flags |= C_ORIG_RDONLY;
2491 m2.mc_db = &env->me_metas[(txn->mt_txnid-1) & 1]->mm_dbs[FREE_DBI];
2492 m2.mc_dbflag = (unsigned char *)""; /* probably unnecessary */
2496 key.mv_data = &last; /* will look up last+1 */
2497 key.mv_size = sizeof(last);
2499 if (Paranoid && mc->mc_dbi == FREE_DBI)
2502 if (Paranoid && retry < 0 && mop_len)
2506 /* Do not fetch more if the record will be too recent */
2507 if (oldest <= last) {
2509 oldest = mdb_find_oldest(txn);
2510 env->me_pgoldest = oldest;
2516 rc = mdb_cursor_get(&m2, &key, NULL, op);
2518 if (rc == MDB_NOTFOUND)
2522 last = *(txnid_t*)key.mv_data;
2523 if (oldest <= last) {
2525 oldest = mdb_find_oldest(txn);
2526 env->me_pgoldest = oldest;
2532 np = m2.mc_pg[m2.mc_top];
2533 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2534 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2537 idl = (MDB_ID *) data.mv_data;
2540 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2545 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2547 mop = env->me_pghead;
2549 env->me_pglast = last;
2551 DPRINTF(("IDL read txn %"Yu" root %"Yu" num %u",
2552 last, txn->mt_dbs[FREE_DBI].md_root, i));
2554 DPRINTF(("IDL %"Yu, idl[j]));
2556 /* Merge in descending sorted order */
2557 mdb_midl_xmerge(mop, idl);
2561 /* Use new pages from the map when nothing suitable in the freeDB */
2563 pgno = txn->mt_next_pgno;
2564 if (pgno + num >= env->me_maxpg) {
2565 DPUTS("DB size maxed out");
2569 #if defined(_WIN32) && !defined(MDB_VL32)
2570 if (!(env->me_flags & MDB_RDONLY)) {
2572 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
2573 p = VirtualAlloc(p, env->me_psize * num, MEM_COMMIT,
2574 (env->me_flags & MDB_WRITEMAP) ? PAGE_READWRITE:
2577 DPUTS("VirtualAlloc failed");
2585 if (env->me_flags & MDB_WRITEMAP) {
2586 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2588 if (!(np = mdb_page_malloc(txn, num))) {
2594 mop[0] = mop_len -= num;
2595 /* Move any stragglers down */
2596 for (j = i-num; j < mop_len; )
2597 mop[++j] = mop[++i];
2599 txn->mt_next_pgno = pgno + num;
2602 mdb_page_dirty(txn, np);
2608 txn->mt_flags |= MDB_TXN_ERROR;
2612 /** Copy the used portions of a non-overflow page.
2613 * @param[in] dst page to copy into
2614 * @param[in] src page to copy from
2615 * @param[in] psize size of a page
2618 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2620 enum { Align = sizeof(pgno_t) };
2621 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2623 /* If page isn't full, just copy the used portion. Adjust
2624 * alignment so memcpy may copy words instead of bytes.
2626 if ((unused &= -Align) && !IS_LEAF2(src)) {
2627 upper = (upper + PAGEBASE) & -Align;
2628 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2629 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2632 memcpy(dst, src, psize - unused);
2636 /** Pull a page off the txn's spill list, if present.
2637 * If a page being referenced was spilled to disk in this txn, bring
2638 * it back and make it dirty/writable again.
2639 * @param[in] txn the transaction handle.
2640 * @param[in] mp the page being referenced. It must not be dirty.
2641 * @param[out] ret the writable page, if any. ret is unchanged if
2642 * mp wasn't spilled.
2645 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2647 MDB_env *env = txn->mt_env;
2650 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2652 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2653 if (!tx2->mt_spill_pgs)
2655 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2656 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2659 if (txn->mt_dirty_room == 0)
2660 return MDB_TXN_FULL;
2661 if (IS_OVERFLOW(mp))
2665 if (env->me_flags & MDB_WRITEMAP) {
2668 np = mdb_page_malloc(txn, num);
2672 memcpy(np, mp, num * env->me_psize);
2674 mdb_page_copy(np, mp, env->me_psize);
2677 /* If in current txn, this page is no longer spilled.
2678 * If it happens to be the last page, truncate the spill list.
2679 * Otherwise mark it as deleted by setting the LSB.
2681 if (x == txn->mt_spill_pgs[0])
2682 txn->mt_spill_pgs[0]--;
2684 txn->mt_spill_pgs[x] |= 1;
2685 } /* otherwise, if belonging to a parent txn, the
2686 * page remains spilled until child commits
2689 mdb_page_dirty(txn, np);
2690 np->mp_flags |= P_DIRTY;
2698 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2699 * Set #MDB_TXN_ERROR on failure.
2700 * @param[in] mc cursor pointing to the page to be touched
2701 * @return 0 on success, non-zero on failure.
2704 mdb_page_touch(MDB_cursor *mc)
2706 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2707 MDB_txn *txn = mc->mc_txn;
2708 MDB_cursor *m2, *m3;
2712 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2713 if (txn->mt_flags & MDB_TXN_SPILLS) {
2715 rc = mdb_page_unspill(txn, mp, &np);
2721 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2722 (rc = mdb_page_alloc(mc, 1, &np)))
2725 DPRINTF(("touched db %d page %"Yu" -> %"Yu, DDBI(mc),
2726 mp->mp_pgno, pgno));
2727 mdb_cassert(mc, mp->mp_pgno != pgno);
2728 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2729 /* Update the parent page, if any, to point to the new page */
2731 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2732 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2733 SETPGNO(node, pgno);
2735 mc->mc_db->md_root = pgno;
2737 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2738 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2740 /* If txn has a parent, make sure the page is in our
2744 unsigned x = mdb_mid2l_search(dl, pgno);
2745 if (x <= dl[0].mid && dl[x].mid == pgno) {
2746 if (mp != dl[x].mptr) { /* bad cursor? */
2747 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2748 txn->mt_flags |= MDB_TXN_ERROR;
2754 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2756 np = mdb_page_malloc(txn, 1);
2761 rc = mdb_mid2l_insert(dl, &mid);
2762 mdb_cassert(mc, rc == 0);
2767 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2769 np->mp_flags |= P_DIRTY;
2772 /* Adjust cursors pointing to mp */
2773 mc->mc_pg[mc->mc_top] = np;
2774 m2 = txn->mt_cursors[mc->mc_dbi];
2775 if (mc->mc_flags & C_SUB) {
2776 for (; m2; m2=m2->mc_next) {
2777 m3 = &m2->mc_xcursor->mx_cursor;
2778 if (m3->mc_snum < mc->mc_snum) continue;
2779 if (m3->mc_pg[mc->mc_top] == mp)
2780 m3->mc_pg[mc->mc_top] = np;
2783 for (; m2; m2=m2->mc_next) {
2784 if (m2->mc_snum < mc->mc_snum) continue;
2785 if (m2 == mc) continue;
2786 if (m2->mc_pg[mc->mc_top] == mp) {
2787 m2->mc_pg[mc->mc_top] = np;
2789 XCURSOR_REFRESH(m2, mc->mc_top, np);
2793 MDB_PAGE_UNREF(mc->mc_txn, mp);
2797 txn->mt_flags |= MDB_TXN_ERROR;
2802 mdb_env_sync0(MDB_env *env, int force, pgno_t numpgs)
2805 if (env->me_flags & MDB_RDONLY)
2807 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2808 if (env->me_flags & MDB_WRITEMAP) {
2809 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2810 ? MS_ASYNC : MS_SYNC;
2811 if (MDB_MSYNC(env->me_map, env->me_psize * numpgs, flags))
2814 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2818 #ifdef BROKEN_FDATASYNC
2819 if (env->me_flags & MDB_FSYNCONLY) {
2820 if (fsync(env->me_fd))
2824 if (MDB_FDATASYNC(env->me_fd))
2832 mdb_env_sync(MDB_env *env, int force)
2834 MDB_meta *m = mdb_env_pick_meta(env);
2835 return mdb_env_sync0(env, force, m->mm_last_pg+1);
2838 /** Back up parent txn's cursors, then grab the originals for tracking */
2840 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2842 MDB_cursor *mc, *bk;
2847 for (i = src->mt_numdbs; --i >= 0; ) {
2848 if ((mc = src->mt_cursors[i]) != NULL) {
2849 size = sizeof(MDB_cursor);
2851 size += sizeof(MDB_xcursor);
2852 for (; mc; mc = bk->mc_next) {
2858 mc->mc_db = &dst->mt_dbs[i];
2859 /* Kill pointers into src to reduce abuse: The
2860 * user may not use mc until dst ends. But we need a valid
2861 * txn pointer here for cursor fixups to keep working.
2864 mc->mc_dbflag = &dst->mt_dbflags[i];
2865 if ((mx = mc->mc_xcursor) != NULL) {
2866 *(MDB_xcursor *)(bk+1) = *mx;
2867 mx->mx_cursor.mc_txn = dst;
2869 mc->mc_next = dst->mt_cursors[i];
2870 dst->mt_cursors[i] = mc;
2877 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2878 * @param[in] txn the transaction handle.
2879 * @param[in] merge true to keep changes to parent cursors, false to revert.
2880 * @return 0 on success, non-zero on failure.
2883 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2885 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2889 for (i = txn->mt_numdbs; --i >= 0; ) {
2890 for (mc = cursors[i]; mc; mc = next) {
2892 if ((bk = mc->mc_backup) != NULL) {
2894 /* Commit changes to parent txn */
2895 mc->mc_next = bk->mc_next;
2896 mc->mc_backup = bk->mc_backup;
2897 mc->mc_txn = bk->mc_txn;
2898 mc->mc_db = bk->mc_db;
2899 mc->mc_dbflag = bk->mc_dbflag;
2900 if ((mx = mc->mc_xcursor) != NULL)
2901 mx->mx_cursor.mc_txn = bk->mc_txn;
2903 /* Abort nested txn */
2905 if ((mx = mc->mc_xcursor) != NULL)
2906 *mx = *(MDB_xcursor *)(bk+1);
2910 /* Only malloced cursors are permanently tracked. */
2917 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2923 Pidset = F_SETLK, Pidcheck = F_GETLK
2927 /** Set or check a pid lock. Set returns 0 on success.
2928 * Check returns 0 if the process is certainly dead, nonzero if it may
2929 * be alive (the lock exists or an error happened so we do not know).
2931 * On Windows Pidset is a no-op, we merely check for the existence
2932 * of the process with the given pid. On POSIX we use a single byte
2933 * lock on the lockfile, set at an offset equal to the pid.
2936 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2938 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2941 if (op == Pidcheck) {
2942 h = OpenProcess(env->me_pidquery, FALSE, pid);
2943 /* No documented "no such process" code, but other program use this: */
2945 return ErrCode() != ERROR_INVALID_PARAMETER;
2946 /* A process exists until all handles to it close. Has it exited? */
2947 ret = WaitForSingleObject(h, 0) != 0;
2954 struct flock lock_info;
2955 memset(&lock_info, 0, sizeof(lock_info));
2956 lock_info.l_type = F_WRLCK;
2957 lock_info.l_whence = SEEK_SET;
2958 lock_info.l_start = pid;
2959 lock_info.l_len = 1;
2960 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2961 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2963 } else if ((rc = ErrCode()) == EINTR) {
2971 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2972 * @param[in] txn the transaction handle to initialize
2973 * @return 0 on success, non-zero on failure.
2976 mdb_txn_renew0(MDB_txn *txn)
2978 MDB_env *env = txn->mt_env;
2979 MDB_txninfo *ti = env->me_txns;
2981 unsigned int i, nr, flags = txn->mt_flags;
2983 int rc, new_notls = 0;
2985 if ((flags &= MDB_TXN_RDONLY) != 0) {
2987 meta = mdb_env_pick_meta(env);
2988 txn->mt_txnid = meta->mm_txnid;
2989 txn->mt_u.reader = NULL;
2991 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2992 pthread_getspecific(env->me_txkey);
2994 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2995 return MDB_BAD_RSLOT;
2997 MDB_PID_T pid = env->me_pid;
2998 MDB_THR_T tid = pthread_self();
2999 mdb_mutexref_t rmutex = env->me_rmutex;
3001 if (!env->me_live_reader) {
3002 rc = mdb_reader_pid(env, Pidset, pid);
3005 env->me_live_reader = 1;
3008 if (LOCK_MUTEX(rc, env, rmutex))
3010 nr = ti->mti_numreaders;
3011 for (i=0; i<nr; i++)
3012 if (ti->mti_readers[i].mr_pid == 0)
3014 if (i == env->me_maxreaders) {
3015 UNLOCK_MUTEX(rmutex);
3016 return MDB_READERS_FULL;
3018 r = &ti->mti_readers[i];
3019 /* Claim the reader slot, carefully since other code
3020 * uses the reader table un-mutexed: First reset the
3021 * slot, next publish it in mti_numreaders. After
3022 * that, it is safe for mdb_env_close() to touch it.
3023 * When it will be closed, we can finally claim it.
3026 r->mr_txnid = (txnid_t)-1;
3029 ti->mti_numreaders = ++nr;
3030 env->me_close_readers = nr;
3032 UNLOCK_MUTEX(rmutex);
3034 new_notls = (env->me_flags & MDB_NOTLS);
3035 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
3040 do /* LY: Retry on a race, ITS#7970. */
3041 r->mr_txnid = ti->mti_txnid;
3042 while(r->mr_txnid != ti->mti_txnid);
3043 txn->mt_txnid = r->mr_txnid;
3044 txn->mt_u.reader = r;
3045 meta = env->me_metas[txn->mt_txnid & 1];
3049 /* Not yet touching txn == env->me_txn0, it may be active */
3051 if (LOCK_MUTEX(rc, env, env->me_wmutex))
3053 txn->mt_txnid = ti->mti_txnid;
3054 meta = env->me_metas[txn->mt_txnid & 1];
3056 meta = mdb_env_pick_meta(env);
3057 txn->mt_txnid = meta->mm_txnid;
3061 if (txn->mt_txnid == mdb_debug_start)
3064 txn->mt_child = NULL;
3065 txn->mt_loose_pgs = NULL;
3066 txn->mt_loose_count = 0;
3067 txn->mt_dirty_room = MDB_IDL_UM_MAX;
3068 txn->mt_u.dirty_list = env->me_dirty_list;
3069 txn->mt_u.dirty_list[0].mid = 0;
3070 txn->mt_free_pgs = env->me_free_pgs;
3071 txn->mt_free_pgs[0] = 0;
3072 txn->mt_spill_pgs = NULL;
3074 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
3077 /* Copy the DB info and flags */
3078 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
3080 /* Moved to here to avoid a data race in read TXNs */
3081 txn->mt_next_pgno = meta->mm_last_pg+1;
3083 txn->mt_last_pgno = txn->mt_next_pgno - 1;
3086 txn->mt_flags = flags;
3089 txn->mt_numdbs = env->me_numdbs;
3090 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3091 x = env->me_dbflags[i];
3092 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
3093 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
3095 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
3096 txn->mt_dbflags[FREE_DBI] = DB_VALID;
3098 if (env->me_flags & MDB_FATAL_ERROR) {
3099 DPUTS("environment had fatal error, must shutdown!");
3101 } else if (env->me_maxpg < txn->mt_next_pgno) {
3102 rc = MDB_MAP_RESIZED;
3106 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
3111 mdb_txn_renew(MDB_txn *txn)
3115 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
3118 rc = mdb_txn_renew0(txn);
3119 if (rc == MDB_SUCCESS) {
3120 DPRINTF(("renew txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3121 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3122 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
3128 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
3132 int rc, size, tsize;
3134 flags &= MDB_TXN_BEGIN_FLAGS;
3135 flags |= env->me_flags & MDB_WRITEMAP;
3137 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
3141 /* Nested transactions: Max 1 child, write txns only, no writemap */
3142 flags |= parent->mt_flags;
3143 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
3144 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
3146 /* Child txns save MDB_pgstate and use own copy of cursors */
3147 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
3148 size += tsize = sizeof(MDB_ntxn);
3149 } else if (flags & MDB_RDONLY) {
3150 size = env->me_maxdbs * (sizeof(MDB_db)+1);
3151 size += tsize = sizeof(MDB_txn);
3153 /* Reuse preallocated write txn. However, do not touch it until
3154 * mdb_txn_renew0() succeeds, since it currently may be active.
3159 if ((txn = calloc(1, size)) == NULL) {
3160 DPRINTF(("calloc: %s", strerror(errno)));
3165 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
3166 if (!txn->mt_rpages) {
3170 txn->mt_rpages[0].mid = 0;
3171 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
3174 txn->mt_dbxs = env->me_dbxs; /* static */
3175 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
3176 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
3177 txn->mt_flags = flags;
3182 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
3183 txn->mt_dbiseqs = parent->mt_dbiseqs;
3184 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
3185 if (!txn->mt_u.dirty_list ||
3186 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
3188 free(txn->mt_u.dirty_list);
3192 txn->mt_txnid = parent->mt_txnid;
3193 txn->mt_dirty_room = parent->mt_dirty_room;
3194 txn->mt_u.dirty_list[0].mid = 0;
3195 txn->mt_spill_pgs = NULL;
3196 txn->mt_next_pgno = parent->mt_next_pgno;
3197 parent->mt_flags |= MDB_TXN_HAS_CHILD;
3198 parent->mt_child = txn;
3199 txn->mt_parent = parent;
3200 txn->mt_numdbs = parent->mt_numdbs;
3202 txn->mt_rpages = parent->mt_rpages;
3204 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3205 /* Copy parent's mt_dbflags, but clear DB_NEW */
3206 for (i=0; i<txn->mt_numdbs; i++)
3207 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
3209 ntxn = (MDB_ntxn *)txn;
3210 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
3211 if (env->me_pghead) {
3212 size = MDB_IDL_SIZEOF(env->me_pghead);
3213 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
3215 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
3220 rc = mdb_cursor_shadow(parent, txn);
3222 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
3223 } else { /* MDB_RDONLY */
3224 txn->mt_dbiseqs = env->me_dbiseqs;
3226 rc = mdb_txn_renew0(txn);
3229 if (txn != env->me_txn0) {
3231 free(txn->mt_rpages);
3236 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
3238 DPRINTF(("begin txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3239 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
3240 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
3247 mdb_txn_env(MDB_txn *txn)
3249 if(!txn) return NULL;
3254 mdb_txn_id(MDB_txn *txn)
3257 return txn->mt_txnid;
3260 /** Export or close DBI handles opened in this txn. */
3262 mdb_dbis_update(MDB_txn *txn, int keep)
3265 MDB_dbi n = txn->mt_numdbs;
3266 MDB_env *env = txn->mt_env;
3267 unsigned char *tdbflags = txn->mt_dbflags;
3269 for (i = n; --i >= CORE_DBS;) {
3270 if (tdbflags[i] & DB_NEW) {
3272 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
3274 char *ptr = env->me_dbxs[i].md_name.mv_data;
3276 env->me_dbxs[i].md_name.mv_data = NULL;
3277 env->me_dbxs[i].md_name.mv_size = 0;
3278 env->me_dbflags[i] = 0;
3279 env->me_dbiseqs[i]++;
3285 if (keep && env->me_numdbs < n)
3289 /** End a transaction, except successful commit of a nested transaction.
3290 * May be called twice for readonly txns: First reset it, then abort.
3291 * @param[in] txn the transaction handle to end
3292 * @param[in] mode why and how to end the transaction
3295 mdb_txn_end(MDB_txn *txn, unsigned mode)
3297 MDB_env *env = txn->mt_env;
3299 static const char *const names[] = MDB_END_NAMES;
3302 /* Export or close DBI handles opened in this txn */
3303 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
3305 DPRINTF(("%s txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3306 names[mode & MDB_END_OPMASK],
3307 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3308 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
3310 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3311 if (txn->mt_u.reader) {
3312 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
3313 if (!(env->me_flags & MDB_NOTLS)) {
3314 txn->mt_u.reader = NULL; /* txn does not own reader */
3315 } else if (mode & MDB_END_SLOT) {
3316 txn->mt_u.reader->mr_pid = 0;
3317 txn->mt_u.reader = NULL;
3318 } /* else txn owns the slot until it does MDB_END_SLOT */
3320 txn->mt_numdbs = 0; /* prevent further DBI activity */
3321 txn->mt_flags |= MDB_TXN_FINISHED;
3323 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
3324 pgno_t *pghead = env->me_pghead;
3326 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
3327 mdb_cursors_close(txn, 0);
3328 if (!(env->me_flags & MDB_WRITEMAP)) {
3329 mdb_dlist_free(txn);
3333 txn->mt_flags = MDB_TXN_FINISHED;
3335 if (!txn->mt_parent) {
3336 mdb_midl_shrink(&txn->mt_free_pgs);
3337 env->me_free_pgs = txn->mt_free_pgs;
3339 env->me_pghead = NULL;
3343 mode = 0; /* txn == env->me_txn0, do not free() it */
3345 /* The writer mutex was locked in mdb_txn_begin. */
3347 UNLOCK_MUTEX(env->me_wmutex);
3349 txn->mt_parent->mt_child = NULL;
3350 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3351 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3352 mdb_midl_free(txn->mt_free_pgs);
3353 mdb_midl_free(txn->mt_spill_pgs);
3354 free(txn->mt_u.dirty_list);
3357 mdb_midl_free(pghead);
3360 if (!txn->mt_parent) {
3361 MDB_ID3L el = env->me_rpages, tl = txn->mt_rpages;
3362 unsigned i, x, n = tl[0].mid;
3363 pthread_mutex_lock(&env->me_rpmutex);
3364 for (i = 1; i <= n; i++) {
3365 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
3366 /* tmp overflow pages that we didn't share in env */
3367 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3369 x = mdb_mid3l_search(el, tl[i].mid);
3370 if (tl[i].mptr == el[x].mptr) {
3373 /* another tmp overflow page */
3374 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3378 pthread_mutex_unlock(&env->me_rpmutex);
3380 if (mode & MDB_END_FREE)
3384 if (mode & MDB_END_FREE)
3389 mdb_txn_reset(MDB_txn *txn)
3394 /* This call is only valid for read-only txns */
3395 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3398 mdb_txn_end(txn, MDB_END_RESET);
3402 mdb_txn_abort(MDB_txn *txn)
3408 mdb_txn_abort(txn->mt_child);
3410 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3413 /** Save the freelist as of this transaction to the freeDB.
3414 * This changes the freelist. Keep trying until it stabilizes.
3416 * When (MDB_DEVEL) & 2, the changes do not affect #mdb_page_alloc(),
3417 * it then uses the transaction's original snapshot of the freeDB.
3420 mdb_freelist_save(MDB_txn *txn)
3422 /* env->me_pghead[] can grow and shrink during this call.
3423 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3424 * Page numbers cannot disappear from txn->mt_free_pgs[].
3427 MDB_env *env = txn->mt_env;
3428 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3429 txnid_t pglast = 0, head_id = 0;
3430 pgno_t freecnt = 0, *free_pgs, *mop;
3431 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3433 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3435 if (env->me_pghead) {
3436 /* Make sure first page of freeDB is touched and on freelist */
3437 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3438 if (rc && rc != MDB_NOTFOUND)
3442 if (!env->me_pghead && txn->mt_loose_pgs) {
3443 /* Put loose page numbers in mt_free_pgs, since
3444 * we may be unable to return them to me_pghead.
3446 MDB_page *mp = txn->mt_loose_pgs;
3447 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3449 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3450 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3451 txn->mt_loose_pgs = NULL;
3452 txn->mt_loose_count = 0;
3455 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3456 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3457 ? SSIZE_MAX : maxfree_1pg;
3460 /* Come back here after each Put() in case freelist changed */
3465 /* If using records from freeDB which we have not yet
3466 * deleted, delete them and any we reserved for me_pghead.
3468 while (pglast < env->me_pglast) {
3469 rc = mdb_cursor_first(&mc, &key, NULL);
3472 pglast = head_id = *(txnid_t *)key.mv_data;
3473 total_room = head_room = 0;
3474 mdb_tassert(txn, pglast <= env->me_pglast);
3475 rc = mdb_cursor_del(&mc, 0);
3480 /* Save the IDL of pages freed by this txn, to a single record */
3481 if (freecnt < txn->mt_free_pgs[0]) {
3483 /* Make sure last page of freeDB is touched and on freelist */
3484 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3485 if (rc && rc != MDB_NOTFOUND)
3488 free_pgs = txn->mt_free_pgs;
3489 /* Write to last page of freeDB */
3490 key.mv_size = sizeof(txn->mt_txnid);
3491 key.mv_data = &txn->mt_txnid;
3493 freecnt = free_pgs[0];
3494 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3495 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3498 /* Retry if mt_free_pgs[] grew during the Put() */
3499 free_pgs = txn->mt_free_pgs;
3500 } while (freecnt < free_pgs[0]);
3501 mdb_midl_sort(free_pgs);
3502 memcpy(data.mv_data, free_pgs, data.mv_size);
3505 unsigned int i = free_pgs[0];
3506 DPRINTF(("IDL write txn %"Yu" root %"Yu" num %u",
3507 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3509 DPRINTF(("IDL %"Yu, free_pgs[i]));
3515 mop = env->me_pghead;
3516 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3518 /* Reserve records for me_pghead[]. Split it if multi-page,
3519 * to avoid searching freeDB for a page range. Use keys in
3520 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3522 if (total_room >= mop_len) {
3523 if (total_room == mop_len || --more < 0)
3525 } else if (head_room >= maxfree_1pg && head_id > 1) {
3526 /* Keep current record (overflow page), add a new one */
3530 /* (Re)write {key = head_id, IDL length = head_room} */
3531 total_room -= head_room;
3532 head_room = mop_len - total_room;
3533 if (head_room > maxfree_1pg && head_id > 1) {
3534 /* Overflow multi-page for part of me_pghead */
3535 head_room /= head_id; /* amortize page sizes */
3536 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3537 } else if (head_room < 0) {
3538 /* Rare case, not bothering to delete this record */
3541 key.mv_size = sizeof(head_id);
3542 key.mv_data = &head_id;
3543 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3544 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3547 /* IDL is initially empty, zero out at least the length */
3548 pgs = (pgno_t *)data.mv_data;
3549 j = head_room > clean_limit ? head_room : 0;
3553 total_room += head_room;
3556 /* Return loose page numbers to me_pghead, though usually none are
3557 * left at this point. The pages themselves remain in dirty_list.
3559 if (txn->mt_loose_pgs) {
3560 MDB_page *mp = txn->mt_loose_pgs;
3561 unsigned count = txn->mt_loose_count;
3563 /* Room for loose pages + temp IDL with same */
3564 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3566 mop = env->me_pghead;
3567 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3568 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3569 loose[ ++count ] = mp->mp_pgno;
3571 mdb_midl_sort(loose);
3572 mdb_midl_xmerge(mop, loose);
3573 txn->mt_loose_pgs = NULL;
3574 txn->mt_loose_count = 0;
3578 /* Fill in the reserved me_pghead records */
3584 rc = mdb_cursor_first(&mc, &key, &data);
3585 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3586 txnid_t id = *(txnid_t *)key.mv_data;
3587 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3590 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3592 if (len > mop_len) {
3594 data.mv_size = (len + 1) * sizeof(MDB_ID);
3596 data.mv_data = mop -= len;
3599 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3601 if (rc || !(mop_len -= len))
3608 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3609 * @param[in] txn the transaction that's being committed
3610 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3611 * @return 0 on success, non-zero on failure.
3614 mdb_page_flush(MDB_txn *txn, int keep)
3616 MDB_env *env = txn->mt_env;
3617 MDB_ID2L dl = txn->mt_u.dirty_list;
3618 unsigned psize = env->me_psize, j;
3619 int i, pagecount = dl[0].mid, rc;
3623 MDB_page *dp = NULL;
3627 struct iovec iov[MDB_COMMIT_PAGES];
3628 ssize_t wsize = 0, wres;
3629 off_t wpos = 0, next_pos = 1; /* impossible pos, so pos != next_pos */
3635 if (env->me_flags & MDB_WRITEMAP) {
3636 /* Clear dirty flags */
3637 while (++i <= pagecount) {
3639 /* Don't flush this page yet */
3640 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3641 dp->mp_flags &= ~P_KEEP;
3645 dp->mp_flags &= ~P_DIRTY;
3650 /* Write the pages */
3652 if (++i <= pagecount) {
3654 /* Don't flush this page yet */
3655 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3656 dp->mp_flags &= ~P_KEEP;
3661 /* clear dirty flag */
3662 dp->mp_flags &= ~P_DIRTY;
3665 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3670 /* Windows actually supports scatter/gather I/O, but only on
3671 * unbuffered file handles. Since we're relying on the OS page
3672 * cache for all our data, that's self-defeating. So we just
3673 * write pages one at a time. We use the ov structure to set
3674 * the write offset, to at least save the overhead of a Seek
3677 DPRINTF(("committing page %"Yu, pgno));
3678 memset(&ov, 0, sizeof(ov));
3679 ov.Offset = pos & 0xffffffff;
3680 ov.OffsetHigh = pos >> 16 >> 16;
3681 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3683 DPRINTF(("WriteFile: %d", rc));
3687 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3688 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3691 /* Write previous page(s) */
3692 #ifdef MDB_USE_PWRITEV
3693 wres = pwritev(env->me_fd, iov, n, wpos);
3696 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3699 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3703 DPRINTF(("lseek: %s", strerror(rc)));
3706 wres = writev(env->me_fd, iov, n);
3709 if (wres != wsize) {
3714 DPRINTF(("Write error: %s", strerror(rc)));
3716 rc = EIO; /* TODO: Use which error code? */
3717 DPUTS("short write, filesystem full?");
3728 DPRINTF(("committing page %"Yu, pgno));
3729 next_pos = pos + size;
3730 iov[n].iov_len = size;
3731 iov[n].iov_base = (char *)dp;
3737 if (pgno > txn->mt_last_pgno)
3738 txn->mt_last_pgno = pgno;
3741 /* MIPS has cache coherency issues, this is a no-op everywhere else
3742 * Note: for any size >= on-chip cache size, entire on-chip cache is
3745 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3747 for (i = keep; ++i <= pagecount; ) {
3749 /* This is a page we skipped above */
3752 dl[j].mid = dp->mp_pgno;
3755 mdb_dpage_free(env, dp);
3760 txn->mt_dirty_room += i - j;
3766 mdb_txn_commit(MDB_txn *txn)
3769 unsigned int i, end_mode;
3775 /* mdb_txn_end() mode for a commit which writes nothing */
3776 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3778 if (txn->mt_child) {
3779 rc = mdb_txn_commit(txn->mt_child);
3786 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3790 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3791 DPUTS("txn has failed/finished, can't commit");
3793 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3798 if (txn->mt_parent) {
3799 MDB_txn *parent = txn->mt_parent;
3803 unsigned x, y, len, ps_len;
3805 /* Append our free list to parent's */
3806 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3809 mdb_midl_free(txn->mt_free_pgs);
3810 /* Failures after this must either undo the changes
3811 * to the parent or set MDB_TXN_ERROR in the parent.
3814 parent->mt_next_pgno = txn->mt_next_pgno;
3815 parent->mt_flags = txn->mt_flags;
3817 /* Merge our cursors into parent's and close them */
3818 mdb_cursors_close(txn, 1);
3820 /* Update parent's DB table. */
3821 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3822 parent->mt_numdbs = txn->mt_numdbs;
3823 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3824 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3825 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3826 /* preserve parent's DB_NEW status */
3827 x = parent->mt_dbflags[i] & DB_NEW;
3828 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3831 dst = parent->mt_u.dirty_list;
3832 src = txn->mt_u.dirty_list;
3833 /* Remove anything in our dirty list from parent's spill list */
3834 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3836 pspill[0] = (pgno_t)-1;
3837 /* Mark our dirty pages as deleted in parent spill list */
3838 for (i=0, len=src[0].mid; ++i <= len; ) {
3839 MDB_ID pn = src[i].mid << 1;
3840 while (pn > pspill[x])
3842 if (pn == pspill[x]) {
3847 /* Squash deleted pagenums if we deleted any */
3848 for (x=y; ++x <= ps_len; )
3849 if (!(pspill[x] & 1))
3850 pspill[++y] = pspill[x];
3854 /* Remove anything in our spill list from parent's dirty list */
3855 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3856 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3857 MDB_ID pn = txn->mt_spill_pgs[i];
3859 continue; /* deleted spillpg */
3861 y = mdb_mid2l_search(dst, pn);
3862 if (y <= dst[0].mid && dst[y].mid == pn) {
3864 while (y < dst[0].mid) {
3873 /* Find len = length of merging our dirty list with parent's */
3875 dst[0].mid = 0; /* simplify loops */
3876 if (parent->mt_parent) {
3877 len = x + src[0].mid;
3878 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3879 for (i = x; y && i; y--) {
3880 pgno_t yp = src[y].mid;
3881 while (yp < dst[i].mid)
3883 if (yp == dst[i].mid) {
3888 } else { /* Simplify the above for single-ancestor case */
3889 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3891 /* Merge our dirty list with parent's */
3893 for (i = len; y; dst[i--] = src[y--]) {
3894 pgno_t yp = src[y].mid;
3895 while (yp < dst[x].mid)
3896 dst[i--] = dst[x--];
3897 if (yp == dst[x].mid)
3898 free(dst[x--].mptr);
3900 mdb_tassert(txn, i == x);
3902 free(txn->mt_u.dirty_list);
3903 parent->mt_dirty_room = txn->mt_dirty_room;
3904 if (txn->mt_spill_pgs) {
3905 if (parent->mt_spill_pgs) {
3906 /* TODO: Prevent failure here, so parent does not fail */
3907 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3909 parent->mt_flags |= MDB_TXN_ERROR;
3910 mdb_midl_free(txn->mt_spill_pgs);
3911 mdb_midl_sort(parent->mt_spill_pgs);
3913 parent->mt_spill_pgs = txn->mt_spill_pgs;
3917 /* Append our loose page list to parent's */
3918 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3920 *lp = txn->mt_loose_pgs;
3921 parent->mt_loose_count += txn->mt_loose_count;
3923 parent->mt_child = NULL;
3924 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3929 if (txn != env->me_txn) {
3930 DPUTS("attempt to commit unknown transaction");
3935 mdb_cursors_close(txn, 0);
3937 if (!txn->mt_u.dirty_list[0].mid &&
3938 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3941 DPRINTF(("committing txn %"Yu" %p on mdbenv %p, root page %"Yu,
3942 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3944 /* Update DB root pointers */
3945 if (txn->mt_numdbs > CORE_DBS) {
3949 data.mv_size = sizeof(MDB_db);
3951 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3952 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3953 if (txn->mt_dbflags[i] & DB_DIRTY) {
3954 if (TXN_DBI_CHANGED(txn, i)) {
3958 data.mv_data = &txn->mt_dbs[i];
3959 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3967 rc = mdb_freelist_save(txn);
3971 mdb_midl_free(env->me_pghead);
3972 env->me_pghead = NULL;
3973 mdb_midl_shrink(&txn->mt_free_pgs);
3979 if ((rc = mdb_page_flush(txn, 0)))
3981 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3982 (rc = mdb_env_sync0(env, 0, txn->mt_next_pgno)))
3984 if ((rc = mdb_env_write_meta(txn)))
3986 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3989 mdb_txn_end(txn, end_mode);
3997 /** Read the environment parameters of a DB environment before
3998 * mapping it into memory.
3999 * @param[in] env the environment handle
4000 * @param[in] prev whether to read the backup meta page
4001 * @param[out] meta address of where to store the meta information
4002 * @return 0 on success, non-zero on failure.
4005 mdb_env_read_header(MDB_env *env, int prev, MDB_meta *meta)
4011 enum { Size = sizeof(pbuf) };
4013 /* We don't know the page size yet, so use a minimum value.
4014 * Read both meta pages so we can use the latest one.
4017 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
4021 memset(&ov, 0, sizeof(ov));
4023 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
4024 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
4027 rc = pread(env->me_fd, &pbuf, Size, off);
4030 if (rc == 0 && off == 0)
4032 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
4033 DPRINTF(("read: %s", mdb_strerror(rc)));
4037 p = (MDB_page *)&pbuf;
4039 if (!F_ISSET(p->mp_flags, P_META)) {
4040 DPRINTF(("page %"Yu" not a meta page", p->mp_pgno));
4045 if (m->mm_magic != MDB_MAGIC) {
4046 DPUTS("meta has invalid magic");
4050 if (m->mm_version != MDB_DATA_VERSION) {
4051 DPRINTF(("database is version %u, expected version %u",
4052 m->mm_version, MDB_DATA_VERSION));
4053 return MDB_VERSION_MISMATCH;
4056 if (off == 0 || (prev ? m->mm_txnid < meta->mm_txnid : m->mm_txnid > meta->mm_txnid))
4062 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
4064 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
4066 meta->mm_magic = MDB_MAGIC;
4067 meta->mm_version = MDB_DATA_VERSION;
4068 meta->mm_mapsize = env->me_mapsize;
4069 meta->mm_psize = env->me_psize;
4070 meta->mm_last_pg = NUM_METAS-1;
4071 meta->mm_flags = env->me_flags & 0xffff;
4072 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
4073 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
4074 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
4077 /** Write the environment parameters of a freshly created DB environment.
4078 * @param[in] env the environment handle
4079 * @param[in] meta the #MDB_meta to write
4080 * @return 0 on success, non-zero on failure.
4083 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
4091 memset(&ov, 0, sizeof(ov));
4092 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
4094 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
4097 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
4098 len = pwrite(fd, ptr, size, pos); \
4099 if (len == -1 && ErrCode() == EINTR) continue; \
4100 rc = (len >= 0); break; } while(1)
4103 DPUTS("writing new meta page");
4105 psize = env->me_psize;
4107 p = calloc(NUM_METAS, psize);
4111 p->mp_flags = P_META;
4112 *(MDB_meta *)METADATA(p) = *meta;
4114 q = (MDB_page *)((char *)p + psize);
4116 q->mp_flags = P_META;
4117 *(MDB_meta *)METADATA(q) = *meta;
4119 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
4122 else if ((unsigned) len == psize * NUM_METAS)
4130 /** Update the environment info to commit a transaction.
4131 * @param[in] txn the transaction that's being committed
4132 * @return 0 on success, non-zero on failure.
4135 mdb_env_write_meta(MDB_txn *txn)
4138 MDB_meta meta, metab, *mp;
4142 int rc, len, toggle;
4151 toggle = txn->mt_txnid & 1;
4152 DPRINTF(("writing meta page %d for root page %"Yu,
4153 toggle, txn->mt_dbs[MAIN_DBI].md_root));
4156 flags = txn->mt_flags | env->me_flags;
4157 mp = env->me_metas[toggle];
4158 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
4159 /* Persist any increases of mapsize config */
4160 if (mapsize < env->me_mapsize)
4161 mapsize = env->me_mapsize;
4163 if (flags & MDB_WRITEMAP) {
4164 mp->mm_mapsize = mapsize;
4165 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4166 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4167 mp->mm_last_pg = txn->mt_next_pgno - 1;
4168 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
4169 !(defined(__i386__) || defined(__x86_64__))
4170 /* LY: issue a memory barrier, if not x86. ITS#7969 */
4171 __sync_synchronize();
4173 mp->mm_txnid = txn->mt_txnid;
4174 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
4175 unsigned meta_size = env->me_psize;
4176 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
4177 ptr = (char *)mp - PAGEHDRSZ;
4178 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
4179 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
4183 if (MDB_MSYNC(ptr, meta_size, rc)) {
4190 metab.mm_txnid = mp->mm_txnid;
4191 metab.mm_last_pg = mp->mm_last_pg;
4193 meta.mm_mapsize = mapsize;
4194 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4195 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4196 meta.mm_last_pg = txn->mt_next_pgno - 1;
4197 meta.mm_txnid = txn->mt_txnid;
4199 off = offsetof(MDB_meta, mm_mapsize);
4200 ptr = (char *)&meta + off;
4201 len = sizeof(MDB_meta) - off;
4202 off += (char *)mp - env->me_map;
4204 /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC.
4205 * (me_mfd goes to the same file as me_fd, but writing to it
4206 * also syncs to disk. Avoids a separate fdatasync() call.)
4208 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
4211 memset(&ov, 0, sizeof(ov));
4213 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
4218 rc = pwrite(mfd, ptr, len, off);
4221 rc = rc < 0 ? ErrCode() : EIO;
4226 DPUTS("write failed, disk error?");
4227 /* On a failure, the pagecache still contains the new data.
4228 * Write some old data back, to prevent it from being used.
4229 * Use the non-SYNC fd; we know it will fail anyway.
4231 meta.mm_last_pg = metab.mm_last_pg;
4232 meta.mm_txnid = metab.mm_txnid;
4234 memset(&ov, 0, sizeof(ov));
4236 WriteFile(env->me_fd, ptr, len, NULL, &ov);
4238 r2 = pwrite(env->me_fd, ptr, len, off);
4239 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
4242 env->me_flags |= MDB_FATAL_ERROR;
4245 /* MIPS has cache coherency issues, this is a no-op everywhere else */
4246 CACHEFLUSH(env->me_map + off, len, DCACHE);
4248 /* Memory ordering issues are irrelevant; since the entire writer
4249 * is wrapped by wmutex, all of these changes will become visible
4250 * after the wmutex is unlocked. Since the DB is multi-version,
4251 * readers will get consistent data regardless of how fresh or
4252 * how stale their view of these values is.
4255 env->me_txns->mti_txnid = txn->mt_txnid;
4260 /** Check both meta pages to see which one is newer.
4261 * @param[in] env the environment handle
4262 * @return newest #MDB_meta.
4265 mdb_env_pick_meta(const MDB_env *env)
4267 MDB_meta *const *metas = env->me_metas;
4268 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
4272 mdb_env_create(MDB_env **env)
4276 e = calloc(1, sizeof(MDB_env));
4280 e->me_maxreaders = DEFAULT_READERS;
4281 e->me_maxdbs = e->me_numdbs = CORE_DBS;
4282 e->me_fd = INVALID_HANDLE_VALUE;
4283 e->me_lfd = INVALID_HANDLE_VALUE;
4284 e->me_mfd = INVALID_HANDLE_VALUE;
4285 #ifdef MDB_USE_POSIX_SEM
4286 e->me_rmutex = SEM_FAILED;
4287 e->me_wmutex = SEM_FAILED;
4288 #elif defined MDB_USE_SYSV_SEM
4289 e->me_rmutex->semid = -1;
4290 e->me_wmutex->semid = -1;
4292 e->me_pid = getpid();
4293 GET_PAGESIZE(e->me_os_psize);
4294 VGMEMP_CREATE(e,0,0);
4300 /** @brief Map a result from an NTAPI call to WIN32. */
4302 mdb_nt2win32(NTSTATUS st)
4307 GetOverlappedResult(NULL, &o, &br, FALSE);
4308 return GetLastError();
4313 mdb_env_map(MDB_env *env, void *addr)
4316 unsigned int flags = env->me_flags;
4319 int access = SECTION_MAP_READ;
4323 ULONG pageprot = PAGE_READONLY, secprot, alloctype;
4325 if (flags & MDB_WRITEMAP) {
4326 access |= SECTION_MAP_WRITE;
4327 pageprot = PAGE_READWRITE;
4329 if (flags & MDB_RDONLY) {
4330 secprot = PAGE_READONLY;
4334 secprot = PAGE_READWRITE;
4335 msize = env->me_mapsize;
4336 alloctype = MEM_RESERVE;
4339 rc = NtCreateSection(&mh, access, NULL, NULL, secprot, SEC_RESERVE, env->me_fd);
4341 return mdb_nt2win32(rc);
4344 msize = NUM_METAS * env->me_psize;
4346 rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
4353 return mdb_nt2win32(rc);
4358 env->me_map = mmap(addr, NUM_METAS * env->me_psize, PROT_READ, MAP_SHARED,
4360 if (env->me_map == MAP_FAILED) {
4365 int prot = PROT_READ;
4366 if (flags & MDB_WRITEMAP) {
4368 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
4371 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
4373 if (env->me_map == MAP_FAILED) {
4378 if (flags & MDB_NORDAHEAD) {
4379 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
4381 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
4383 #ifdef POSIX_MADV_RANDOM
4384 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
4385 #endif /* POSIX_MADV_RANDOM */
4386 #endif /* MADV_RANDOM */
4390 /* Can happen because the address argument to mmap() is just a
4391 * hint. mmap() can pick another, e.g. if the range is in use.
4392 * The MAP_FIXED flag would prevent that, but then mmap could
4393 * instead unmap existing pages to make room for the new map.
4395 if (addr && env->me_map != addr)
4396 return EBUSY; /* TODO: Make a new MDB_* error code? */
4399 p = (MDB_page *)env->me_map;
4400 env->me_metas[0] = METADATA(p);
4401 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4407 mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
4409 /* If env is already open, caller is responsible for making
4410 * sure there are no active txns.
4420 meta = mdb_env_pick_meta(env);
4422 size = meta->mm_mapsize;
4424 /* Silently round up to minimum if the size is too small */
4425 mdb_size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4430 /* For MDB_VL32 this bit is a noop since we dynamically remap
4431 * chunks of the DB anyway.
4433 munmap(env->me_map, env->me_mapsize);
4434 env->me_mapsize = size;
4435 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4436 rc = mdb_env_map(env, old);
4439 #endif /* !MDB_VL32 */
4441 env->me_mapsize = size;
4443 env->me_maxpg = env->me_mapsize / env->me_psize;
4448 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4452 env->me_maxdbs = dbs + CORE_DBS;
4457 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4459 if (env->me_map || readers < 1)
4461 env->me_maxreaders = readers;
4466 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4468 if (!env || !readers)
4470 *readers = env->me_maxreaders;
4475 mdb_fsize(HANDLE fd, mdb_size_t *size)
4478 LARGE_INTEGER fsize;
4480 if (!GetFileSizeEx(fd, &fsize))
4483 *size = fsize.QuadPart;
4497 typedef wchar_t mdb_nchar_t;
4498 # define MDB_NAME(str) L##str
4499 # define mdb_name_cpy wcscpy
4501 /** Character type for file names: char on Unix, wchar_t on Windows */
4502 typedef char mdb_nchar_t;
4503 # define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */
4504 # define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */
4507 /** Filename - string of #mdb_nchar_t[] */
4508 typedef struct MDB_name {
4509 int mn_len; /**< Length */
4510 int mn_alloced; /**< True if #mn_val was malloced */
4511 mdb_nchar_t *mn_val; /**< Contents */
4514 /** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */
4515 static const mdb_nchar_t *const mdb_suffixes[2][2] = {
4516 { MDB_NAME("/data.mdb"), MDB_NAME("") },
4517 { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") }
4520 #define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */
4522 /** Set up filename + scratch area for filename suffix, for opening files.
4523 * It should be freed with #mdb_fname_destroy().
4524 * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16.
4526 * @param[in] path Pathname for #mdb_env_open().
4527 * @param[in] envflags Whether a subdir and/or lockfile will be used.
4528 * @param[out] fname Resulting filename, with room for a suffix if necessary.
4531 mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname)
4533 int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK);
4534 fname->mn_alloced = 0;
4536 return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN);
4538 fname->mn_len = strlen(path);
4540 fname->mn_val = (char *) path;
4541 else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) {
4542 fname->mn_alloced = 1;
4543 strcpy(fname->mn_val, path);
4551 /** Destroy \b fname from #mdb_fname_init() */
4552 #define mdb_fname_destroy(fname) \
4553 do { if ((fname).mn_alloced) free((fname).mn_val); } while (0)
4555 #ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */
4556 # define MDB_CLOEXEC O_CLOEXEC
4558 # define MDB_CLOEXEC 0
4561 /** File type, access mode etc. for #mdb_fopen() */
4562 enum mdb_fopen_type {
4564 MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS
4566 /* A comment in mdb_fopen() explains some O_* flag choices. */
4567 MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */
4568 MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */
4569 MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */
4570 MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */
4571 /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits
4572 * distinguish otherwise-equal MDB_O_* constants from each other.
4574 MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY,
4575 MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */
4579 /** Open an LMDB file.
4580 * @param[in] env The LMDB environment.
4581 * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is
4582 * appended if necessary to create the filename, without changing mn_len.
4583 * @param[in] which Determines file type, access mode, etc.
4584 * @param[in] mode The Unix permissions for the file, if we create it.
4585 * @param[out] res Resulting file handle.
4586 * @return 0 on success, non-zero on failure.
4589 mdb_fopen(const MDB_env *env, MDB_name *fname,
4590 enum mdb_fopen_type which, mdb_mode_t mode,
4593 int rc = MDB_SUCCESS;
4596 DWORD acc, share, disp, attrs;
4601 if (fname->mn_alloced) /* modifiable copy */
4602 mdb_name_cpy(fname->mn_val + fname->mn_len,
4603 mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]);
4605 /* The directory must already exist. Usually the file need not.
4606 * MDB_O_META requires the file because we already created it using
4607 * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file.
4609 * With MDB_O_COPY we do not want the OS to cache the writes, since
4610 * the source data is already in the OS cache.
4612 * The lockfile needs FD_CLOEXEC (close file descriptor on exec*())
4613 * to avoid the flock() issues noted under Caveats in lmdb.h.
4614 * Also set it for other filehandles which the user cannot get at
4615 * and close himself, which he may need after fork(). I.e. all but
4616 * me_fd, which programs do use via mdb_env_get_fd().
4620 acc = GENERIC_READ|GENERIC_WRITE;
4621 share = FILE_SHARE_READ|FILE_SHARE_WRITE;
4623 attrs = FILE_ATTRIBUTE_NORMAL;
4625 case MDB_O_RDONLY: /* read-only datafile */
4627 disp = OPEN_EXISTING;
4629 case MDB_O_META: /* for writing metapages */
4630 acc = GENERIC_WRITE;
4631 disp = OPEN_EXISTING;
4632 attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH;
4634 case MDB_O_COPY: /* mdb_env_copy() & co */
4635 acc = GENERIC_WRITE;
4638 attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH;
4640 default: break; /* silence gcc -Wswitch (not all enum values handled) */
4642 fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL);
4644 fd = open(fname->mn_val, which & MDB_O_MASK, mode);
4647 if (fd == INVALID_HANDLE_VALUE)
4651 if (which != MDB_O_RDONLY && which != MDB_O_RDWR) {
4652 /* Set CLOEXEC if we could not pass it to open() */
4653 if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1)
4654 (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
4656 if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) {
4657 /* This may require buffer alignment. There is no portable
4658 * way to ask how much, so we require OS pagesize alignment.
4660 # ifdef F_NOCACHE /* __APPLE__ */
4661 (void) fcntl(fd, F_NOCACHE, 1);
4662 # elif defined O_DIRECT
4663 /* open(...O_DIRECT...) would break on filesystems without
4664 * O_DIRECT support (ITS#7682). Try to set it here instead.
4666 if ((flags = fcntl(fd, F_GETFL)) != -1)
4667 (void) fcntl(fd, F_SETFL, flags | O_DIRECT);
4671 #endif /* !_WIN32 */
4678 #ifdef BROKEN_FDATASYNC
4679 #include <sys/utsname.h>
4680 #include <sys/vfs.h>
4683 /** Further setup required for opening an LMDB environment
4686 mdb_env_open2(MDB_env *env, int prev)
4688 unsigned int flags = env->me_flags;
4689 int i, newenv = 0, rc;
4693 /* See if we should use QueryLimited */
4695 if ((rc & 0xff) > 5)
4696 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4698 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4699 /* Grab functions we need from NTDLL */
4700 if (!NtCreateSection) {
4701 HMODULE h = GetModuleHandle("NTDLL.DLL");
4704 NtClose = (NtCloseFunc *)GetProcAddress(h, "NtClose");
4707 NtMapViewOfSection = (NtMapViewOfSectionFunc *)GetProcAddress(h, "NtMapViewOfSection");
4708 if (!NtMapViewOfSection)
4710 NtCreateSection = (NtCreateSectionFunc *)GetProcAddress(h, "NtCreateSection");
4711 if (!NtCreateSection)
4716 #ifdef BROKEN_FDATASYNC
4717 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4718 * https://lkml.org/lkml/2012/9/3/83
4719 * Kernels after 3.6-rc6 are known good.
4720 * https://lkml.org/lkml/2012/9/10/556
4721 * See if the DB is on ext3/ext4, then check for new enough kernel
4722 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4727 fstatfs(env->me_fd, &st);
4728 while (st.f_type == 0xEF53) {
4732 if (uts.release[0] < '3') {
4733 if (!strncmp(uts.release, "2.6.32.", 7)) {
4734 i = atoi(uts.release+7);
4736 break; /* 2.6.32.60 and newer is OK */
4737 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4738 i = atoi(uts.release+7);
4740 break; /* 2.6.34.15 and newer is OK */
4742 } else if (uts.release[0] == '3') {
4743 i = atoi(uts.release+2);
4745 break; /* 3.6 and newer is OK */
4747 i = atoi(uts.release+4);
4749 break; /* 3.5.4 and newer is OK */
4750 } else if (i == 2) {
4751 i = atoi(uts.release+4);
4753 break; /* 3.2.30 and newer is OK */
4755 } else { /* 4.x and newer is OK */
4758 env->me_flags |= MDB_FSYNCONLY;
4764 if ((i = mdb_env_read_header(env, prev, &meta)) != 0) {
4767 DPUTS("new mdbenv");
4769 env->me_psize = env->me_os_psize;
4770 if (env->me_psize > MAX_PAGESIZE)
4771 env->me_psize = MAX_PAGESIZE;
4772 memset(&meta, 0, sizeof(meta));
4773 mdb_env_init_meta0(env, &meta);
4774 meta.mm_mapsize = DEFAULT_MAPSIZE;
4776 env->me_psize = meta.mm_psize;
4779 /* Was a mapsize configured? */
4780 if (!env->me_mapsize) {
4781 env->me_mapsize = meta.mm_mapsize;
4784 /* Make sure mapsize >= committed data size. Even when using
4785 * mm_mapsize, which could be broken in old files (ITS#7789).
4787 mdb_size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4788 if (env->me_mapsize < minsize)
4789 env->me_mapsize = minsize;
4791 meta.mm_mapsize = env->me_mapsize;
4793 if (newenv && !(flags & MDB_FIXEDMAP)) {
4794 /* mdb_env_map() may grow the datafile. Write the metapages
4795 * first, so the file will be valid if initialization fails.
4796 * Except with FIXEDMAP, since we do not yet know mm_address.
4797 * We could fill in mm_address later, but then a different
4798 * program might end up doing that - one with a memory layout
4799 * and map address which does not suit the main program.
4801 rc = mdb_env_init_meta(env, &meta);
4807 /* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
4811 rc = WriteFile(env->me_fd, &dummy, 1, &len, NULL);
4819 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4824 if (flags & MDB_FIXEDMAP)
4825 meta.mm_address = env->me_map;
4826 i = mdb_env_init_meta(env, &meta);
4827 if (i != MDB_SUCCESS) {
4832 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4833 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4835 #if !(MDB_MAXKEYSIZE)
4836 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4838 env->me_maxpg = env->me_mapsize / env->me_psize;
4842 MDB_meta *meta = mdb_env_pick_meta(env);
4843 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4845 DPRINTF(("opened database version %u, pagesize %u",
4846 meta->mm_version, env->me_psize));
4847 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4848 DPRINTF(("depth: %u", db->md_depth));
4849 DPRINTF(("entries: %"Yu, db->md_entries));
4850 DPRINTF(("branch pages: %"Yu, db->md_branch_pages));
4851 DPRINTF(("leaf pages: %"Yu, db->md_leaf_pages));
4852 DPRINTF(("overflow pages: %"Yu, db->md_overflow_pages));
4853 DPRINTF(("root: %"Yu, db->md_root));
4861 /** Release a reader thread's slot in the reader lock table.
4862 * This function is called automatically when a thread exits.
4863 * @param[in] ptr This points to the slot in the reader lock table.
4866 mdb_env_reader_dest(void *ptr)
4868 MDB_reader *reader = ptr;
4871 if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */
4873 /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */
4878 /** Junk for arranging thread-specific callbacks on Windows. This is
4879 * necessarily platform and compiler-specific. Windows supports up
4880 * to 1088 keys. Let's assume nobody opens more than 64 environments
4881 * in a single process, for now. They can override this if needed.
4883 #ifndef MAX_TLS_KEYS
4884 #define MAX_TLS_KEYS 64
4886 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4887 static int mdb_tls_nkeys;
4889 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4893 case DLL_PROCESS_ATTACH: break;
4894 case DLL_THREAD_ATTACH: break;
4895 case DLL_THREAD_DETACH:
4896 for (i=0; i<mdb_tls_nkeys; i++) {
4897 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4899 mdb_env_reader_dest(r);
4903 case DLL_PROCESS_DETACH: break;
4908 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4910 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4914 /* Force some symbol references.
4915 * _tls_used forces the linker to create the TLS directory if not already done
4916 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4918 #pragma comment(linker, "/INCLUDE:_tls_used")
4919 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4920 #pragma const_seg(".CRT$XLB")
4921 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4922 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4925 #pragma comment(linker, "/INCLUDE:__tls_used")
4926 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4927 #pragma data_seg(".CRT$XLB")
4928 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4930 #endif /* WIN 32/64 */
4931 #endif /* !__GNUC__ */
4934 /** Downgrade the exclusive lock on the region back to shared */
4936 mdb_env_share_locks(MDB_env *env, int *excl)
4939 MDB_meta *meta = mdb_env_pick_meta(env);
4941 env->me_txns->mti_txnid = meta->mm_txnid;
4946 /* First acquire a shared lock. The Unlock will
4947 * then release the existing exclusive lock.
4949 memset(&ov, 0, sizeof(ov));
4950 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4953 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4959 struct flock lock_info;
4960 /* The shared lock replaces the existing lock */
4961 memset((void *)&lock_info, 0, sizeof(lock_info));
4962 lock_info.l_type = F_RDLCK;
4963 lock_info.l_whence = SEEK_SET;
4964 lock_info.l_start = 0;
4965 lock_info.l_len = 1;
4966 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4967 (rc = ErrCode()) == EINTR) ;
4968 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4975 /** Try to get exclusive lock, otherwise shared.
4976 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4979 mdb_env_excl_lock(MDB_env *env, int *excl)
4983 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4987 memset(&ov, 0, sizeof(ov));
4988 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4995 struct flock lock_info;
4996 memset((void *)&lock_info, 0, sizeof(lock_info));
4997 lock_info.l_type = F_WRLCK;
4998 lock_info.l_whence = SEEK_SET;
4999 lock_info.l_start = 0;
5000 lock_info.l_len = 1;
5001 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
5002 (rc = ErrCode()) == EINTR) ;
5006 # ifndef MDB_USE_POSIX_MUTEX
5007 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
5010 lock_info.l_type = F_RDLCK;
5011 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
5012 (rc = ErrCode()) == EINTR) ;
5022 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
5024 * @(#) $Revision: 5.1 $
5025 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
5026 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
5028 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
5032 * Please do not copyright this code. This code is in the public domain.
5034 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
5035 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
5036 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
5037 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
5038 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
5039 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
5040 * PERFORMANCE OF THIS SOFTWARE.
5043 * chongo <Landon Curt Noll> /\oo/\
5044 * http://www.isthe.com/chongo/
5046 * Share and Enjoy! :-)
5049 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
5050 * @param[in] val value to hash
5051 * @param[in] len length of value
5052 * @return 64 bit hash
5055 mdb_hash(const void *val, size_t len)
5057 const unsigned char *s = (const unsigned char *) val, *end = s + len;
5058 mdb_hash_t hval = 0xcbf29ce484222325ULL;
5060 * FNV-1a hash each octet of the buffer
5063 hval = (hval ^ *s++) * 0x100000001b3ULL;
5065 /* return our new hash value */
5069 /** Hash the string and output the encoded hash.
5070 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
5071 * very short name limits. We don't care about the encoding being reversible,
5072 * we just want to preserve as many bits of the input as possible in a
5073 * small printable string.
5074 * @param[in] str string to hash
5075 * @param[out] encbuf an array of 11 chars to hold the hash
5077 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
5080 mdb_pack85(unsigned long long l, char *out)
5084 for (i=0; i<10 && l; i++) {
5085 *out++ = mdb_a85[l % 85];
5091 /** Init #MDB_env.me_mutexname[] except the char which #MUTEXNAME() will set.
5092 * Changes to this code must be reflected in #MDB_LOCK_FORMAT.
5095 mdb_env_mname_init(MDB_env *env)
5097 char *nm = env->me_mutexname;
5098 strcpy(nm, MUTEXNAME_PREFIX);
5099 mdb_pack85(env->me_txns->mti_mutexid, nm + sizeof(MUTEXNAME_PREFIX));
5102 /** Return env->me_mutexname after filling in ch ('r'/'w') for convenience */
5103 #define MUTEXNAME(env, ch) ( \
5104 (void) ((env)->me_mutexname[sizeof(MUTEXNAME_PREFIX)-1] = (ch)), \
5105 (env)->me_mutexname)
5109 /** Open and/or initialize the lock region for the environment.
5110 * @param[in] env The LMDB environment.
5111 * @param[in] fname Filename + scratch area, from #mdb_fname_init().
5112 * @param[in] mode The Unix permissions for the file, if we create it.
5113 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
5114 * @return 0 on success, non-zero on failure.
5117 mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl)
5120 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
5122 # define MDB_ERRCODE_ROFS EROFS
5124 #ifdef MDB_USE_SYSV_SEM
5131 rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd);
5133 /* Omit lockfile if read-only env on read-only filesystem */
5134 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
5140 if (!(env->me_flags & MDB_NOTLS)) {
5141 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
5144 env->me_flags |= MDB_ENV_TXKEY;
5146 /* Windows TLS callbacks need help finding their TLS info. */
5147 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
5151 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
5155 /* Try to get exclusive lock. If we succeed, then
5156 * nobody is using the lock region and we should initialize it.
5158 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
5161 size = GetFileSize(env->me_lfd, NULL);
5163 size = lseek(env->me_lfd, 0, SEEK_END);
5164 if (size == -1) goto fail_errno;
5166 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
5167 if (size < rsize && *excl > 0) {
5169 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
5170 || !SetEndOfFile(env->me_lfd))
5173 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
5177 size = rsize - sizeof(MDB_txninfo);
5178 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
5183 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
5185 if (!mh) goto fail_errno;
5186 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
5188 if (!env->me_txns) goto fail_errno;
5190 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
5192 if (m == MAP_FAILED) goto fail_errno;
5198 BY_HANDLE_FILE_INFORMATION stbuf;
5205 if (!mdb_sec_inited) {
5206 InitializeSecurityDescriptor(&mdb_null_sd,
5207 SECURITY_DESCRIPTOR_REVISION);
5208 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
5209 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
5210 mdb_all_sa.bInheritHandle = FALSE;
5211 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
5214 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
5215 idbuf.volume = stbuf.dwVolumeSerialNumber;
5216 idbuf.nhigh = stbuf.nFileIndexHigh;
5217 idbuf.nlow = stbuf.nFileIndexLow;
5218 env->me_txns->mti_mutexid = mdb_hash(&idbuf, sizeof(idbuf));
5219 mdb_env_mname_init(env);
5220 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, MUTEXNAME(env, 'r'));
5221 if (!env->me_rmutex) goto fail_errno;
5222 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, MUTEXNAME(env, 'w'));
5223 if (!env->me_wmutex) goto fail_errno;
5224 #elif defined(MDB_USE_POSIX_SEM)
5231 #if defined(__NetBSD__)
5232 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
5234 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
5235 memset(&idbuf, 0, sizeof(idbuf));
5236 idbuf.dev = stbuf.st_dev;
5237 idbuf.ino = stbuf.st_ino;
5238 env->me_txns->mti_mutexid = mdb_hash(&idbuf, sizeof(idbuf))
5239 #ifdef MDB_SHORT_SEMNAMES
5240 /* Max 9 base85-digits. We truncate here instead of in
5241 * mdb_env_mname_init() to keep the latter portable.
5243 % ((mdb_hash_t)85*85*85*85*85*85*85*85*85)
5246 mdb_env_mname_init(env);
5247 /* Clean up after a previous run, if needed: Try to
5248 * remove both semaphores before doing anything else.
5250 sem_unlink(MUTEXNAME(env, 'r'));
5251 sem_unlink(MUTEXNAME(env, 'w'));
5252 env->me_rmutex = sem_open(MUTEXNAME(env, 'r'), O_CREAT|O_EXCL, mode, 1);
5253 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
5254 env->me_wmutex = sem_open(MUTEXNAME(env, 'w'), O_CREAT|O_EXCL, mode, 1);
5255 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
5256 #elif defined(MDB_USE_SYSV_SEM)
5257 unsigned short vals[2] = {1, 1};
5258 key_t key = ftok(fname->mn_val, 'M'); /* fname is lockfile path now */
5261 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
5265 if (semctl(semid, 0, SETALL, semu) < 0)
5267 env->me_txns->mti_semid = semid;
5268 env->me_txns->mti_rlocked = 0;
5269 env->me_txns->mti_wlocked = 0;
5270 #else /* MDB_USE_POSIX_MUTEX: */
5271 pthread_mutexattr_t mattr;
5273 /* Solaris needs this before initing a robust mutex. Otherwise
5274 * it may skip the init and return EBUSY "seems someone already
5275 * inited" or EINVAL "it was inited differently".
5277 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
5278 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
5280 if ((rc = pthread_mutexattr_init(&mattr)) != 0)
5282 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
5283 #ifdef MDB_ROBUST_SUPPORTED
5284 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
5286 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
5287 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
5288 pthread_mutexattr_destroy(&mattr);
5291 #endif /* _WIN32 || ... */
5293 env->me_txns->mti_magic = MDB_MAGIC;
5294 env->me_txns->mti_format = MDB_LOCK_FORMAT;
5295 env->me_txns->mti_txnid = 0;
5296 env->me_txns->mti_numreaders = 0;
5299 #ifdef MDB_USE_SYSV_SEM
5300 struct semid_ds buf;
5302 if (env->me_txns->mti_magic != MDB_MAGIC) {
5303 DPUTS("lock region has invalid magic");
5307 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
5308 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
5309 env->me_txns->mti_format, MDB_LOCK_FORMAT));
5310 rc = MDB_VERSION_MISMATCH;
5314 if (rc && rc != EACCES && rc != EAGAIN) {
5318 mdb_env_mname_init(env);
5319 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, MUTEXNAME(env, 'r'));
5320 if (!env->me_rmutex) goto fail_errno;
5321 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, MUTEXNAME(env, 'w'));
5322 if (!env->me_wmutex) goto fail_errno;
5323 #elif defined(MDB_USE_POSIX_SEM)
5324 mdb_env_mname_init(env);
5325 env->me_rmutex = sem_open(MUTEXNAME(env, 'r'), 0);
5326 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
5327 env->me_wmutex = sem_open(MUTEXNAME(env, 'w'), 0);
5328 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
5329 #elif defined(MDB_USE_SYSV_SEM)
5330 semid = env->me_txns->mti_semid;
5332 /* check for read access */
5333 if (semctl(semid, 0, IPC_STAT, semu) < 0)
5335 /* check for write access */
5336 if (semctl(semid, 0, IPC_SET, semu) < 0)
5340 #ifdef MDB_USE_SYSV_SEM
5341 env->me_rmutex->semid = semid;
5342 env->me_wmutex->semid = semid;
5343 env->me_rmutex->semnum = 0;
5344 env->me_wmutex->semnum = 1;
5345 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
5346 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
5357 /** Only a subset of the @ref mdb_env flags can be changed
5358 * at runtime. Changing other flags requires closing the
5359 * environment and re-opening it with the new flags.
5361 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
5362 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
5363 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD|MDB_PREVMETA)
5365 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
5366 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
5370 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
5375 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
5379 if (flags & MDB_WRITEMAP) {
5380 /* silently ignore WRITEMAP in 32 bit mode */
5381 flags ^= MDB_WRITEMAP;
5383 if (flags & MDB_FIXEDMAP) {
5384 /* cannot support FIXEDMAP */
5388 flags |= env->me_flags;
5390 rc = mdb_fname_init(path, flags, &fname);
5396 env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
5397 if (!env->me_rpmutex) {
5402 rc = pthread_mutex_init(&env->me_rpmutex, NULL);
5407 flags |= MDB_ENV_ACTIVE; /* tell mdb_env_close0() to clean up */
5409 if (flags & MDB_RDONLY) {
5410 /* silently ignore WRITEMAP when we're only getting read access */
5411 flags &= ~MDB_WRITEMAP;
5413 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
5414 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
5418 env->me_flags = flags;
5424 env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
5425 if (!env->me_rpages) {
5429 env->me_rpages[0].mid = 0;
5430 env->me_rpcheck = MDB_ERPAGE_SIZE/2;
5434 env->me_path = strdup(path);
5435 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
5436 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
5437 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
5438 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
5442 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
5444 /* For RDONLY, get lockfile after we know datafile exists */
5445 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
5446 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
5451 rc = mdb_fopen(env, &fname,
5452 (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR,
5457 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
5458 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
5463 if ((rc = mdb_env_open2(env, flags & MDB_PREVMETA)) == MDB_SUCCESS) {
5464 if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) {
5465 /* Synchronous fd for meta writes. Needed even with
5466 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
5468 rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd);
5472 DPRINTF(("opened dbenv %p", (void *) env));
5474 rc = mdb_env_share_locks(env, &excl);
5478 if (!(flags & MDB_RDONLY)) {
5480 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
5481 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
5482 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
5483 (txn = calloc(1, size)))
5485 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
5486 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
5487 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
5488 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
5491 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
5492 if (!txn->mt_rpages) {
5497 txn->mt_rpages[0].mid = 0;
5498 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
5500 txn->mt_dbxs = env->me_dbxs;
5501 txn->mt_flags = MDB_TXN_FINISHED;
5511 mdb_env_close0(env, excl);
5513 mdb_fname_destroy(fname);
5517 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5519 mdb_env_close0(MDB_env *env, int excl)
5523 if (!(env->me_flags & MDB_ENV_ACTIVE))
5526 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5528 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5529 free(env->me_dbxs[i].md_name.mv_data);
5534 free(env->me_dbiseqs);
5535 free(env->me_dbflags);
5537 free(env->me_dirty_list);
5539 if (env->me_txn0 && env->me_txn0->mt_rpages)
5540 free(env->me_txn0->mt_rpages);
5541 if (env->me_rpages) {
5542 MDB_ID3L el = env->me_rpages;
5544 for (x=1; x<=el[0].mid; x++)
5545 munmap(el[x].mptr, el[x].mcnt * env->me_psize);
5550 mdb_midl_free(env->me_free_pgs);
5552 if (env->me_flags & MDB_ENV_TXKEY) {
5553 pthread_key_delete(env->me_txkey);
5555 /* Delete our key from the global list */
5556 for (i=0; i<mdb_tls_nkeys; i++)
5557 if (mdb_tls_keys[i] == env->me_txkey) {
5558 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5567 munmap(env->me_map, NUM_METAS*env->me_psize);
5569 munmap(env->me_map, env->me_mapsize);
5572 if (env->me_mfd != INVALID_HANDLE_VALUE)
5573 (void) close(env->me_mfd);
5574 if (env->me_fd != INVALID_HANDLE_VALUE)
5575 (void) close(env->me_fd);
5577 MDB_PID_T pid = getpid();
5578 /* Clearing readers is done in this function because
5579 * me_txkey with its destructor must be disabled first.
5581 * We skip the the reader mutex, so we touch only
5582 * data owned by this process (me_close_readers and
5583 * our readers), and clear each reader atomically.
5585 for (i = env->me_close_readers; --i >= 0; )
5586 if (env->me_txns->mti_readers[i].mr_pid == pid)
5587 env->me_txns->mti_readers[i].mr_pid = 0;
5589 if (env->me_rmutex) {
5590 CloseHandle(env->me_rmutex);
5591 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5593 /* Windows automatically destroys the mutexes when
5594 * the last handle closes.
5596 #elif defined(MDB_USE_POSIX_SEM)
5597 if (env->me_rmutex != SEM_FAILED) {
5598 sem_close(env->me_rmutex);
5599 if (env->me_wmutex != SEM_FAILED)
5600 sem_close(env->me_wmutex);
5601 /* If we have the filelock: If we are the
5602 * only remaining user, clean up semaphores.
5605 mdb_env_excl_lock(env, &excl);
5607 sem_unlink(MUTEXNAME(env, 'r'));
5608 sem_unlink(MUTEXNAME(env, 'w'));
5611 #elif defined(MDB_USE_SYSV_SEM)
5612 if (env->me_rmutex->semid != -1) {
5613 /* If we have the filelock: If we are the
5614 * only remaining user, clean up semaphores.
5617 mdb_env_excl_lock(env, &excl);
5619 semctl(env->me_rmutex->semid, 0, IPC_RMID);
5622 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5624 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5627 /* Unlock the lockfile. Windows would have unlocked it
5628 * after closing anyway, but not necessarily at once.
5630 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5633 (void) close(env->me_lfd);
5637 if (env->me_fmh) CloseHandle(env->me_fmh);
5638 if (env->me_rpmutex) CloseHandle(env->me_rpmutex);
5640 pthread_mutex_destroy(&env->me_rpmutex);
5644 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5648 mdb_env_close(MDB_env *env)
5655 VGMEMP_DESTROY(env);
5656 while ((dp = env->me_dpages) != NULL) {
5657 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5658 env->me_dpages = dp->mp_next;
5662 mdb_env_close0(env, 0);
5666 /** Compare two items pointing at aligned #mdb_size_t's */
5668 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5670 return (*(mdb_size_t *)a->mv_data < *(mdb_size_t *)b->mv_data) ? -1 :
5671 *(mdb_size_t *)a->mv_data > *(mdb_size_t *)b->mv_data;
5674 /** Compare two items pointing at aligned unsigned int's.
5676 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5677 * but #mdb_cmp_clong() is called instead if the data type is #mdb_size_t.
5680 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5682 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5683 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5686 /** Compare two items pointing at unsigned ints of unknown alignment.
5687 * Nodes and keys are guaranteed to be 2-byte aligned.
5690 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5692 #if BYTE_ORDER == LITTLE_ENDIAN
5693 unsigned short *u, *c;
5696 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5697 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5700 } while(!x && u > (unsigned short *)a->mv_data);
5703 unsigned short *u, *c, *end;
5706 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5707 u = (unsigned short *)a->mv_data;
5708 c = (unsigned short *)b->mv_data;
5711 } while(!x && u < end);
5716 /** Compare two items lexically */
5718 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5725 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5731 diff = memcmp(a->mv_data, b->mv_data, len);
5732 return diff ? diff : len_diff<0 ? -1 : len_diff;
5735 /** Compare two items in reverse byte order */
5737 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5739 const unsigned char *p1, *p2, *p1_lim;
5743 p1_lim = (const unsigned char *)a->mv_data;
5744 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5745 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5747 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5753 while (p1 > p1_lim) {
5754 diff = *--p1 - *--p2;
5758 return len_diff<0 ? -1 : len_diff;
5761 /** Search for key within a page, using binary search.
5762 * Returns the smallest entry larger or equal to the key.
5763 * If exactp is non-null, stores whether the found entry was an exact match
5764 * in *exactp (1 or 0).
5765 * Updates the cursor index with the index of the found entry.
5766 * If no entry larger or equal to the key is found, returns NULL.
5769 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5771 unsigned int i = 0, nkeys;
5774 MDB_page *mp = mc->mc_pg[mc->mc_top];
5775 MDB_node *node = NULL;
5780 nkeys = NUMKEYS(mp);
5782 DPRINTF(("searching %u keys in %s %spage %"Yu,
5783 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5786 low = IS_LEAF(mp) ? 0 : 1;
5788 cmp = mc->mc_dbx->md_cmp;
5790 /* Branch pages have no data, so if using integer keys,
5791 * alignment is guaranteed. Use faster mdb_cmp_int.
5793 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5794 if (NODEPTR(mp, 1)->mn_ksize == sizeof(mdb_size_t))
5801 nodekey.mv_size = mc->mc_db->md_pad;
5802 node = NODEPTR(mp, 0); /* fake */
5803 while (low <= high) {
5804 i = (low + high) >> 1;
5805 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5806 rc = cmp(key, &nodekey);
5807 DPRINTF(("found leaf index %u [%s], rc = %i",
5808 i, DKEY(&nodekey), rc));
5817 while (low <= high) {
5818 i = (low + high) >> 1;
5820 node = NODEPTR(mp, i);
5821 nodekey.mv_size = NODEKSZ(node);
5822 nodekey.mv_data = NODEKEY(node);
5824 rc = cmp(key, &nodekey);
5827 DPRINTF(("found leaf index %u [%s], rc = %i",
5828 i, DKEY(&nodekey), rc));
5830 DPRINTF(("found branch index %u [%s -> %"Yu"], rc = %i",
5831 i, DKEY(&nodekey), NODEPGNO(node), rc));
5842 if (rc > 0) { /* Found entry is less than the key. */
5843 i++; /* Skip to get the smallest entry larger than key. */
5845 node = NODEPTR(mp, i);
5848 *exactp = (rc == 0 && nkeys > 0);
5849 /* store the key index */
5850 mc->mc_ki[mc->mc_top] = i;
5852 /* There is no entry larger or equal to the key. */
5855 /* nodeptr is fake for LEAF2 */
5861 mdb_cursor_adjust(MDB_cursor *mc, func)
5865 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5866 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5873 /** Pop a page off the top of the cursor's stack. */
5875 mdb_cursor_pop(MDB_cursor *mc)
5878 DPRINTF(("popping page %"Yu" off db %d cursor %p",
5879 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5885 mc->mc_flags &= ~C_INITIALIZED;
5890 /** Push a page onto the top of the cursor's stack.
5891 * Set #MDB_TXN_ERROR on failure.
5894 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5896 DPRINTF(("pushing page %"Yu" on db %d cursor %p", mp->mp_pgno,
5897 DDBI(mc), (void *) mc));
5899 if (mc->mc_snum >= CURSOR_STACK) {
5900 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5901 return MDB_CURSOR_FULL;
5904 mc->mc_top = mc->mc_snum++;
5905 mc->mc_pg[mc->mc_top] = mp;
5906 mc->mc_ki[mc->mc_top] = 0;
5912 /** Map a read-only page.
5913 * There are two levels of tracking in use, a per-txn list and a per-env list.
5914 * ref'ing and unref'ing the per-txn list is faster since it requires no
5915 * locking. Pages are cached in the per-env list for global reuse, and a lock
5916 * is required. Pages are not immediately unmapped when their refcnt goes to
5917 * zero; they hang around in case they will be reused again soon.
5919 * When the per-txn list gets full, all pages with refcnt=0 are purged from the
5920 * list and their refcnts in the per-env list are decremented.
5922 * When the per-env list gets full, all pages with refcnt=0 are purged from the
5923 * list and their pages are unmapped.
5925 * @note "full" means the list has reached its respective rpcheck threshold.
5926 * This threshold slowly raises if no pages could be purged on a given check,
5927 * and returns to its original value when enough pages were purged.
5929 * If purging doesn't free any slots, filling the per-txn list will return
5930 * MDB_TXN_FULL, and filling the per-env list returns MDB_MAP_FULL.
5932 * Reference tracking in a txn is imperfect, pages can linger with non-zero
5933 * refcnt even without active references. It was deemed to be too invasive
5934 * to add unrefs in every required location. However, all pages are unref'd
5935 * at the end of the transaction. This guarantees that no stale references
5936 * linger in the per-env list.
5938 * Usually we map chunks of 16 pages at a time, but if an overflow page begins
5939 * at the tail of the chunk we extend the chunk to include the entire overflow
5940 * page. Unfortunately, pages can be turned into overflow pages after their
5941 * chunk was already mapped. In that case we must remap the chunk if the
5942 * overflow page is referenced. If the chunk's refcnt is 0 we can just remap
5943 * it, otherwise we temporarily map a new chunk just for the overflow page.
5945 * @note this chunk handling means we cannot guarantee that a data item
5946 * returned from the DB will stay alive for the duration of the transaction:
5947 * We unref pages as soon as a cursor moves away from the page
5948 * A subsequent op may cause a purge, which may unmap any unref'd chunks
5949 * The caller must copy the data if it must be used later in the same txn.
5951 * Also - our reference counting revolves around cursors, but overflow pages
5952 * aren't pointed to by a cursor's page stack. We have to remember them
5953 * explicitly, in the added mc_ovpg field. A single cursor can only hold a
5954 * reference to one overflow page at a time.
5956 * @param[in] txn the transaction for this access.
5957 * @param[in] pgno the page number for the page to retrieve.
5958 * @param[out] ret address of a pointer where the page's address will be stored.
5959 * @return 0 on success, non-zero on failure.
5962 mdb_rpage_get(MDB_txn *txn, pgno_t pg0, MDB_page **ret)
5964 MDB_env *env = txn->mt_env;
5966 MDB_ID3L tl = txn->mt_rpages;
5967 MDB_ID3L el = env->me_rpages;
5971 int rc, retries = 1;
5975 #define SET_OFF(off,val) off.QuadPart = val
5976 #define MAP(rc,env,addr,len,off) \
5978 rc = NtMapViewOfSection(env->me_fmh, GetCurrentProcess(), &addr, 0, \
5979 len, &off, &len, ViewUnmap, (env->me_flags & MDB_RDONLY) ? 0 : MEM_RESERVE, PAGE_READONLY); \
5980 if (rc) rc = mdb_nt2win32(rc)
5984 #define SET_OFF(off,val) off = val
5985 #define MAP(rc,env,addr,len,off) \
5986 addr = mmap(NULL, len, PROT_READ, MAP_SHARED, env->me_fd, off); \
5987 rc = (addr == MAP_FAILED) ? errno : 0
5990 /* remember the offset of the actual page number, so we can
5991 * return the correct pointer at the end.
5993 rem = pg0 & (MDB_RPAGE_CHUNK-1);
5997 x = mdb_mid3l_search(tl, pgno);
5998 if (x <= tl[0].mid && tl[x].mid == pgno) {
5999 if (x != tl[0].mid && tl[x+1].mid == pg0)
6001 /* check for overflow size */
6002 p = (MDB_page *)((char *)tl[x].mptr + rem * env->me_psize);
6003 if (IS_OVERFLOW(p) && p->mp_pages + rem > tl[x].mcnt) {
6004 id3.mcnt = p->mp_pages + rem;
6005 len = id3.mcnt * env->me_psize;
6006 SET_OFF(off, pgno * env->me_psize);
6007 MAP(rc, env, id3.mptr, len, off);
6010 /* check for local-only page */
6012 mdb_tassert(txn, tl[x].mid != pg0);
6013 /* hope there's room to insert this locally.
6014 * setting mid here tells later code to just insert
6015 * this id3 instead of searching for a match.
6020 /* ignore the mapping we got from env, use new one */
6021 tl[x].mptr = id3.mptr;
6022 tl[x].mcnt = id3.mcnt;
6023 /* if no active ref, see if we can replace in env */
6026 pthread_mutex_lock(&env->me_rpmutex);
6027 i = mdb_mid3l_search(el, tl[x].mid);
6028 if (el[i].mref == 1) {
6029 /* just us, replace it */
6030 munmap(el[i].mptr, el[i].mcnt * env->me_psize);
6031 el[i].mptr = tl[x].mptr;
6032 el[i].mcnt = tl[x].mcnt;
6034 /* there are others, remove ourself */
6037 pthread_mutex_unlock(&env->me_rpmutex);
6041 id3.mptr = tl[x].mptr;
6042 id3.mcnt = tl[x].mcnt;
6048 if (tl[0].mid >= MDB_TRPAGE_MAX - txn->mt_rpcheck) {
6050 /* purge unref'd pages from our list and unref in env */
6051 pthread_mutex_lock(&env->me_rpmutex);
6054 for (i=1; i<=tl[0].mid; i++) {
6057 /* tmp overflow pages don't go to env */
6058 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
6059 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
6062 x = mdb_mid3l_search(el, tl[i].mid);
6066 pthread_mutex_unlock(&env->me_rpmutex);
6068 /* we didn't find any unref'd chunks.
6069 * if we're out of room, fail.
6071 if (tl[0].mid >= MDB_TRPAGE_MAX)
6072 return MDB_TXN_FULL;
6073 /* otherwise, raise threshold for next time around
6076 txn->mt_rpcheck /= 2;
6078 /* we found some unused; consolidate the list */
6079 for (i=y+1; i<= tl[0].mid; i++)
6083 /* decrease the check threshold toward its original value */
6084 if (!txn->mt_rpcheck)
6085 txn->mt_rpcheck = 1;
6086 while (txn->mt_rpcheck < tl[0].mid && txn->mt_rpcheck < MDB_TRPAGE_SIZE/2)
6087 txn->mt_rpcheck *= 2;
6090 if (tl[0].mid < MDB_TRPAGE_SIZE) {
6094 /* don't map past last written page in read-only envs */
6095 if ((env->me_flags & MDB_RDONLY) && pgno + MDB_RPAGE_CHUNK-1 > txn->mt_last_pgno)
6096 id3.mcnt = txn->mt_last_pgno + 1 - pgno;
6098 id3.mcnt = MDB_RPAGE_CHUNK;
6099 len = id3.mcnt * env->me_psize;
6102 /* search for page in env */
6103 pthread_mutex_lock(&env->me_rpmutex);
6104 x = mdb_mid3l_search(el, pgno);
6105 if (x <= el[0].mid && el[x].mid == pgno) {
6106 id3.mptr = el[x].mptr;
6107 id3.mcnt = el[x].mcnt;
6108 /* check for overflow size */
6109 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
6110 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
6111 id3.mcnt = p->mp_pages + rem;
6112 len = id3.mcnt * env->me_psize;
6113 SET_OFF(off, pgno * env->me_psize);
6114 MAP(rc, env, id3.mptr, len, off);
6118 munmap(el[x].mptr, env->me_psize * el[x].mcnt);
6119 el[x].mptr = id3.mptr;
6120 el[x].mcnt = id3.mcnt;
6123 pthread_mutex_unlock(&env->me_rpmutex);
6128 pthread_mutex_unlock(&env->me_rpmutex);
6131 if (el[0].mid >= MDB_ERPAGE_MAX - env->me_rpcheck) {
6132 /* purge unref'd pages */
6134 for (i=1; i<=el[0].mid; i++) {
6137 munmap(el[i].mptr, env->me_psize * el[i].mcnt);
6142 /* see if we can unref some local pages */
6147 if (el[0].mid >= MDB_ERPAGE_MAX) {
6148 pthread_mutex_unlock(&env->me_rpmutex);
6149 return MDB_MAP_FULL;
6151 env->me_rpcheck /= 2;
6153 for (i=y+1; i<= el[0].mid; i++)
6157 if (!env->me_rpcheck)
6158 env->me_rpcheck = 1;
6159 while (env->me_rpcheck < el[0].mid && env->me_rpcheck < MDB_ERPAGE_SIZE/2)
6160 env->me_rpcheck *= 2;
6163 SET_OFF(off, pgno * env->me_psize);
6164 MAP(rc, env, id3.mptr, len, off);
6167 pthread_mutex_unlock(&env->me_rpmutex);
6170 /* check for overflow size */
6171 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
6172 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
6173 id3.mcnt = p->mp_pages + rem;
6174 munmap(id3.mptr, len);
6175 len = id3.mcnt * env->me_psize;
6176 MAP(rc, env, id3.mptr, len, off);
6180 mdb_mid3l_insert(el, &id3);
6181 pthread_mutex_unlock(&env->me_rpmutex);
6183 mdb_mid3l_insert(tl, &id3);
6185 return MDB_TXN_FULL;
6188 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
6189 #if MDB_DEBUG /* we don't need this check any more */
6190 if (IS_OVERFLOW(p)) {
6191 mdb_tassert(txn, p->mp_pages + rem <= id3.mcnt);
6199 /** Find the address of the page corresponding to a given page number.
6200 * Set #MDB_TXN_ERROR on failure.
6201 * @param[in] mc the cursor accessing the page.
6202 * @param[in] pgno the page number for the page to retrieve.
6203 * @param[out] ret address of a pointer where the page's address will be stored.
6204 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
6205 * @return 0 on success, non-zero on failure.
6208 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
6210 MDB_txn *txn = mc->mc_txn;
6214 if (! (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP))) {
6218 MDB_ID2L dl = tx2->mt_u.dirty_list;
6220 /* Spilled pages were dirtied in this txn and flushed
6221 * because the dirty list got full. Bring this page
6222 * back in from the map (but don't unspill it here,
6223 * leave that unless page_touch happens again).
6225 if (tx2->mt_spill_pgs) {
6226 MDB_ID pn = pgno << 1;
6227 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
6228 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
6233 unsigned x = mdb_mid2l_search(dl, pgno);
6234 if (x <= dl[0].mid && dl[x].mid == pgno) {
6240 } while ((tx2 = tx2->mt_parent) != NULL);
6243 if (pgno >= txn->mt_next_pgno) {
6244 DPRINTF(("page %"Yu" not found", pgno));
6245 txn->mt_flags |= MDB_TXN_ERROR;
6246 return MDB_PAGE_NOTFOUND;
6254 int rc = mdb_rpage_get(txn, pgno, &p);
6256 txn->mt_flags |= MDB_TXN_ERROR;
6260 MDB_env *env = txn->mt_env;
6261 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
6272 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
6273 * The cursor is at the root page, set up the rest of it.
6276 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
6278 MDB_page *mp = mc->mc_pg[mc->mc_top];
6282 while (IS_BRANCH(mp)) {
6286 DPRINTF(("branch page %"Yu" has %u keys", mp->mp_pgno, NUMKEYS(mp)));
6287 /* Don't assert on branch pages in the FreeDB. We can get here
6288 * while in the process of rebalancing a FreeDB branch page; we must
6289 * let that proceed. ITS#8336
6291 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
6292 DPRINTF(("found index 0 to page %"Yu, NODEPGNO(NODEPTR(mp, 0))));
6294 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
6296 if (flags & MDB_PS_LAST) {
6297 i = NUMKEYS(mp) - 1;
6298 /* if already init'd, see if we're already in right place */
6299 if (mc->mc_flags & C_INITIALIZED) {
6300 if (mc->mc_ki[mc->mc_top] == i) {
6301 mc->mc_top = mc->mc_snum++;
6302 mp = mc->mc_pg[mc->mc_top];
6309 node = mdb_node_search(mc, key, &exact);
6311 i = NUMKEYS(mp) - 1;
6313 i = mc->mc_ki[mc->mc_top];
6315 mdb_cassert(mc, i > 0);
6319 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
6322 mdb_cassert(mc, i < NUMKEYS(mp));
6323 node = NODEPTR(mp, i);
6325 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6328 mc->mc_ki[mc->mc_top] = i;
6329 if ((rc = mdb_cursor_push(mc, mp)))
6333 if (flags & MDB_PS_MODIFY) {
6334 if ((rc = mdb_page_touch(mc)) != 0)
6336 mp = mc->mc_pg[mc->mc_top];
6341 DPRINTF(("internal error, index points to a %02X page!?",
6343 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6344 return MDB_CORRUPTED;
6347 DPRINTF(("found leaf page %"Yu" for key [%s]", mp->mp_pgno,
6348 key ? DKEY(key) : "null"));
6349 mc->mc_flags |= C_INITIALIZED;
6350 mc->mc_flags &= ~C_EOF;
6355 /** Search for the lowest key under the current branch page.
6356 * This just bypasses a NUMKEYS check in the current page
6357 * before calling mdb_page_search_root(), because the callers
6358 * are all in situations where the current page is known to
6362 mdb_page_search_lowest(MDB_cursor *mc)
6364 MDB_page *mp = mc->mc_pg[mc->mc_top];
6365 MDB_node *node = NODEPTR(mp, 0);
6368 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6371 mc->mc_ki[mc->mc_top] = 0;
6372 if ((rc = mdb_cursor_push(mc, mp)))
6374 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
6377 /** Search for the page a given key should be in.
6378 * Push it and its parent pages on the cursor stack.
6379 * @param[in,out] mc the cursor for this operation.
6380 * @param[in] key the key to search for, or NULL for first/last page.
6381 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
6382 * are touched (updated with new page numbers).
6383 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
6384 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
6385 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
6386 * @return 0 on success, non-zero on failure.
6389 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
6394 /* Make sure the txn is still viable, then find the root from
6395 * the txn's db table and set it as the root of the cursor's stack.
6397 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
6398 DPUTS("transaction may not be used now");
6401 /* Make sure we're using an up-to-date root */
6402 if (*mc->mc_dbflag & DB_STALE) {
6404 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6406 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
6407 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
6414 MDB_node *leaf = mdb_node_search(&mc2,
6415 &mc->mc_dbx->md_name, &exact);
6417 return MDB_NOTFOUND;
6418 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
6419 return MDB_INCOMPATIBLE; /* not a named DB */
6420 rc = mdb_node_read(&mc2, leaf, &data);
6423 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
6425 /* The txn may not know this DBI, or another process may
6426 * have dropped and recreated the DB with other flags.
6428 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
6429 return MDB_INCOMPATIBLE;
6430 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
6432 *mc->mc_dbflag &= ~DB_STALE;
6434 root = mc->mc_db->md_root;
6436 if (root == P_INVALID) { /* Tree is empty. */
6437 DPUTS("tree is empty");
6438 return MDB_NOTFOUND;
6442 mdb_cassert(mc, root > 1);
6443 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) {
6446 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[0]);
6448 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
6455 for (i=1; i<mc->mc_snum; i++)
6456 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[i]);
6462 DPRINTF(("db %d root page %"Yu" has flags 0x%X",
6463 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
6465 if (flags & MDB_PS_MODIFY) {
6466 if ((rc = mdb_page_touch(mc)))
6470 if (flags & MDB_PS_ROOTONLY)
6473 return mdb_page_search_root(mc, key, flags);
6477 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
6479 MDB_txn *txn = mc->mc_txn;
6480 pgno_t pg = mp->mp_pgno;
6481 unsigned x = 0, ovpages = mp->mp_pages;
6482 MDB_env *env = txn->mt_env;
6483 MDB_IDL sl = txn->mt_spill_pgs;
6484 MDB_ID pn = pg << 1;
6487 DPRINTF(("free ov page %"Yu" (%d)", pg, ovpages));
6488 /* If the page is dirty or on the spill list we just acquired it,
6489 * so we should give it back to our current free list, if any.
6490 * Otherwise put it onto the list of pages we freed in this txn.
6492 * Won't create me_pghead: me_pglast must be inited along with it.
6493 * Unsupported in nested txns: They would need to hide the page
6494 * range in ancestor txns' dirty and spilled lists.
6496 if (env->me_pghead &&
6498 ((mp->mp_flags & P_DIRTY) ||
6499 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
6503 MDB_ID2 *dl, ix, iy;
6504 rc = mdb_midl_need(&env->me_pghead, ovpages);
6507 if (!(mp->mp_flags & P_DIRTY)) {
6508 /* This page is no longer spilled */
6515 /* Remove from dirty list */
6516 dl = txn->mt_u.dirty_list;
6518 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
6524 mdb_cassert(mc, x > 1);
6526 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
6527 txn->mt_flags |= MDB_TXN_ERROR;
6531 txn->mt_dirty_room++;
6532 if (!(env->me_flags & MDB_WRITEMAP))
6533 mdb_dpage_free(env, mp);
6535 /* Insert in me_pghead */
6536 mop = env->me_pghead;
6537 j = mop[0] + ovpages;
6538 for (i = mop[0]; i && mop[i] < pg; i--)
6544 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
6549 if (mc->mc_ovpg == mp)
6552 mc->mc_db->md_overflow_pages -= ovpages;
6556 /** Return the data associated with a given node.
6557 * @param[in] mc The cursor for this operation.
6558 * @param[in] leaf The node being read.
6559 * @param[out] data Updated to point to the node's data.
6560 * @return 0 on success, non-zero on failure.
6563 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
6565 MDB_page *omp; /* overflow page */
6570 MDB_PAGE_UNREF(mc->mc_txn, MC_OVPG(mc));
6571 MC_SET_OVPG(mc, NULL);
6573 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6574 data->mv_size = NODEDSZ(leaf);
6575 data->mv_data = NODEDATA(leaf);
6579 /* Read overflow data.
6581 data->mv_size = NODEDSZ(leaf);
6582 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
6583 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
6584 DPRINTF(("read overflow page %"Yu" failed", pgno));
6587 data->mv_data = METADATA(omp);
6588 MC_SET_OVPG(mc, omp);
6594 mdb_get(MDB_txn *txn, MDB_dbi dbi,
6595 MDB_val *key, MDB_val *data)
6602 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
6604 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
6607 if (txn->mt_flags & MDB_TXN_BLOCKED)
6610 mdb_cursor_init(&mc, txn, dbi, &mx);
6611 rc = mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
6612 /* unref all the pages when MDB_VL32 - caller must copy the data
6613 * before doing anything else
6615 MDB_CURSOR_UNREF(&mc, 1);
6619 /** Find a sibling for a page.
6620 * Replaces the page at the top of the cursor's stack with the
6621 * specified sibling, if one exists.
6622 * @param[in] mc The cursor for this operation.
6623 * @param[in] move_right Non-zero if the right sibling is requested,
6624 * otherwise the left sibling.
6625 * @return 0 on success, non-zero on failure.
6628 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
6637 if (mc->mc_snum < 2) {
6638 return MDB_NOTFOUND; /* root has no siblings */
6642 op = mc->mc_pg[mc->mc_top];
6645 DPRINTF(("parent page is page %"Yu", index %u",
6646 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
6648 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6649 : (mc->mc_ki[mc->mc_top] == 0)) {
6650 DPRINTF(("no more keys left, moving to %s sibling",
6651 move_right ? "right" : "left"));
6652 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
6653 /* undo cursor_pop before returning */
6660 mc->mc_ki[mc->mc_top]++;
6662 mc->mc_ki[mc->mc_top]--;
6663 DPRINTF(("just moving to %s index key %u",
6664 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
6666 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
6668 MDB_PAGE_UNREF(mc->mc_txn, op);
6670 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6671 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
6672 /* mc will be inconsistent if caller does mc_snum++ as above */
6673 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
6677 mdb_cursor_push(mc, mp);
6679 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
6684 /** Move the cursor to the next data item. */
6686 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6692 if ((mc->mc_flags & C_DEL && op == MDB_NEXT_DUP))
6693 return MDB_NOTFOUND;
6695 if (!(mc->mc_flags & C_INITIALIZED))
6696 return mdb_cursor_first(mc, key, data);
6698 mp = mc->mc_pg[mc->mc_top];
6700 if (mc->mc_flags & C_EOF) {
6701 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mp)-1)
6702 return MDB_NOTFOUND;
6703 mc->mc_flags ^= C_EOF;
6706 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6707 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6708 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6709 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
6710 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
6711 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
6712 if (rc == MDB_SUCCESS)
6713 MDB_GET_KEY(leaf, key);
6718 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6721 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6722 if (op == MDB_NEXT_DUP)
6723 return MDB_NOTFOUND;
6727 DPRINTF(("cursor_next: top page is %"Yu" in cursor %p",
6728 mdb_dbg_pgno(mp), (void *) mc));
6729 if (mc->mc_flags & C_DEL) {
6730 mc->mc_flags ^= C_DEL;
6734 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
6735 DPUTS("=====> move to next sibling page");
6736 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6737 mc->mc_flags |= C_EOF;
6740 mp = mc->mc_pg[mc->mc_top];
6741 DPRINTF(("next page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6743 mc->mc_ki[mc->mc_top]++;
6746 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6747 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6750 key->mv_size = mc->mc_db->md_pad;
6751 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6755 mdb_cassert(mc, IS_LEAF(mp));
6756 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6758 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6759 mdb_xcursor_init1(mc, leaf);
6762 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6765 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6766 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6767 if (rc != MDB_SUCCESS)
6772 MDB_GET_KEY(leaf, key);
6776 /** Move the cursor to the previous data item. */
6778 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6784 if (!(mc->mc_flags & C_INITIALIZED)) {
6785 rc = mdb_cursor_last(mc, key, data);
6788 mc->mc_ki[mc->mc_top]++;
6791 mp = mc->mc_pg[mc->mc_top];
6793 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6794 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6795 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6796 if (op == MDB_PREV || op == MDB_PREV_DUP) {
6797 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
6798 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
6799 if (rc == MDB_SUCCESS) {
6800 MDB_GET_KEY(leaf, key);
6801 mc->mc_flags &= ~C_EOF;
6807 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6810 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6811 if (op == MDB_PREV_DUP)
6812 return MDB_NOTFOUND;
6816 DPRINTF(("cursor_prev: top page is %"Yu" in cursor %p",
6817 mdb_dbg_pgno(mp), (void *) mc));
6819 mc->mc_flags &= ~(C_EOF|C_DEL);
6821 if (mc->mc_ki[mc->mc_top] == 0) {
6822 DPUTS("=====> move to prev sibling page");
6823 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
6826 mp = mc->mc_pg[mc->mc_top];
6827 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
6828 DPRINTF(("prev page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6830 mc->mc_ki[mc->mc_top]--;
6832 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6833 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6836 key->mv_size = mc->mc_db->md_pad;
6837 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6841 mdb_cassert(mc, IS_LEAF(mp));
6842 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6844 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6845 mdb_xcursor_init1(mc, leaf);
6848 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6851 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6852 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6853 if (rc != MDB_SUCCESS)
6858 MDB_GET_KEY(leaf, key);
6862 /** Set the cursor on a specific data item. */
6864 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6865 MDB_cursor_op op, int *exactp)
6869 MDB_node *leaf = NULL;
6872 if (key->mv_size == 0)
6873 return MDB_BAD_VALSIZE;
6875 if (mc->mc_xcursor) {
6876 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6877 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6880 /* See if we're already on the right page */
6881 if (mc->mc_flags & C_INITIALIZED) {
6884 mp = mc->mc_pg[mc->mc_top];
6886 mc->mc_ki[mc->mc_top] = 0;
6887 return MDB_NOTFOUND;
6889 if (mp->mp_flags & P_LEAF2) {
6890 nodekey.mv_size = mc->mc_db->md_pad;
6891 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
6893 leaf = NODEPTR(mp, 0);
6894 MDB_GET_KEY2(leaf, nodekey);
6896 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6898 /* Probably happens rarely, but first node on the page
6899 * was the one we wanted.
6901 mc->mc_ki[mc->mc_top] = 0;
6908 unsigned int nkeys = NUMKEYS(mp);
6910 if (mp->mp_flags & P_LEAF2) {
6911 nodekey.mv_data = LEAF2KEY(mp,
6912 nkeys-1, nodekey.mv_size);
6914 leaf = NODEPTR(mp, nkeys-1);
6915 MDB_GET_KEY2(leaf, nodekey);
6917 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6919 /* last node was the one we wanted */
6920 mc->mc_ki[mc->mc_top] = nkeys-1;
6926 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6927 /* This is definitely the right page, skip search_page */
6928 if (mp->mp_flags & P_LEAF2) {
6929 nodekey.mv_data = LEAF2KEY(mp,
6930 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6932 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6933 MDB_GET_KEY2(leaf, nodekey);
6935 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6937 /* current node was the one we wanted */
6944 mc->mc_flags &= ~C_EOF;
6948 /* If any parents have right-sibs, search.
6949 * Otherwise, there's nothing further.
6951 for (i=0; i<mc->mc_top; i++)
6953 NUMKEYS(mc->mc_pg[i])-1)
6955 if (i == mc->mc_top) {
6956 /* There are no other pages */
6957 mc->mc_ki[mc->mc_top] = nkeys;
6958 return MDB_NOTFOUND;
6962 /* There are no other pages */
6963 mc->mc_ki[mc->mc_top] = 0;
6964 if (op == MDB_SET_RANGE && !exactp) {
6968 return MDB_NOTFOUND;
6974 rc = mdb_page_search(mc, key, 0);
6975 if (rc != MDB_SUCCESS)
6978 mp = mc->mc_pg[mc->mc_top];
6979 mdb_cassert(mc, IS_LEAF(mp));
6982 leaf = mdb_node_search(mc, key, exactp);
6983 if (exactp != NULL && !*exactp) {
6984 /* MDB_SET specified and not an exact match. */
6985 return MDB_NOTFOUND;
6989 DPUTS("===> inexact leaf not found, goto sibling");
6990 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6991 mc->mc_flags |= C_EOF;
6992 return rc; /* no entries matched */
6994 mp = mc->mc_pg[mc->mc_top];
6995 mdb_cassert(mc, IS_LEAF(mp));
6996 leaf = NODEPTR(mp, 0);
7000 mc->mc_flags |= C_INITIALIZED;
7001 mc->mc_flags &= ~C_EOF;
7004 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
7005 key->mv_size = mc->mc_db->md_pad;
7006 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
7011 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7012 mdb_xcursor_init1(mc, leaf);
7015 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7016 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
7017 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
7020 if (op == MDB_GET_BOTH) {
7026 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
7027 if (rc != MDB_SUCCESS)
7030 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
7033 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
7035 dcmp = mc->mc_dbx->md_dcmp;
7036 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
7037 dcmp = mdb_cmp_clong;
7038 rc = dcmp(data, &olddata);
7040 if (op == MDB_GET_BOTH || rc > 0)
7041 return MDB_NOTFOUND;
7048 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
7049 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
7054 /* The key already matches in all other cases */
7055 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
7056 MDB_GET_KEY(leaf, key);
7057 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
7062 /** Move the cursor to the first item in the database. */
7064 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
7069 if (mc->mc_xcursor) {
7070 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
7071 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
7074 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
7075 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
7076 if (rc != MDB_SUCCESS)
7079 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
7081 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
7082 mc->mc_flags |= C_INITIALIZED;
7083 mc->mc_flags &= ~C_EOF;
7085 mc->mc_ki[mc->mc_top] = 0;
7087 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
7088 key->mv_size = mc->mc_db->md_pad;
7089 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
7094 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7095 mdb_xcursor_init1(mc, leaf);
7096 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
7100 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
7104 MDB_GET_KEY(leaf, key);
7108 /** Move the cursor to the last item in the database. */
7110 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
7115 if (mc->mc_xcursor) {
7116 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
7117 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
7120 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
7121 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
7122 if (rc != MDB_SUCCESS)
7125 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
7127 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
7128 mc->mc_flags |= C_INITIALIZED|C_EOF;
7129 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7131 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
7132 key->mv_size = mc->mc_db->md_pad;
7133 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
7138 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7139 mdb_xcursor_init1(mc, leaf);
7140 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
7144 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
7149 MDB_GET_KEY(leaf, key);
7154 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
7159 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
7164 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7168 case MDB_GET_CURRENT:
7169 if (!(mc->mc_flags & C_INITIALIZED)) {
7172 MDB_page *mp = mc->mc_pg[mc->mc_top];
7173 int nkeys = NUMKEYS(mp);
7174 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
7175 mc->mc_ki[mc->mc_top] = nkeys;
7181 key->mv_size = mc->mc_db->md_pad;
7182 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
7184 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7185 MDB_GET_KEY(leaf, key);
7187 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7188 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
7190 rc = mdb_node_read(mc, leaf, data);
7197 case MDB_GET_BOTH_RANGE:
7202 if (mc->mc_xcursor == NULL) {
7203 rc = MDB_INCOMPATIBLE;
7213 rc = mdb_cursor_set(mc, key, data, op,
7214 op == MDB_SET_RANGE ? NULL : &exact);
7217 case MDB_GET_MULTIPLE:
7218 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
7222 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7223 rc = MDB_INCOMPATIBLE;
7227 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
7228 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
7231 case MDB_NEXT_MULTIPLE:
7236 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7237 rc = MDB_INCOMPATIBLE;
7240 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
7241 if (rc == MDB_SUCCESS) {
7242 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
7245 mx = &mc->mc_xcursor->mx_cursor;
7246 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
7248 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
7249 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
7255 case MDB_PREV_MULTIPLE:
7260 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7261 rc = MDB_INCOMPATIBLE;
7264 if (!(mc->mc_flags & C_INITIALIZED))
7265 rc = mdb_cursor_last(mc, key, data);
7268 if (rc == MDB_SUCCESS) {
7269 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
7270 if (mx->mc_flags & C_INITIALIZED) {
7271 rc = mdb_cursor_sibling(mx, 0);
7272 if (rc == MDB_SUCCESS)
7281 case MDB_NEXT_NODUP:
7282 rc = mdb_cursor_next(mc, key, data, op);
7286 case MDB_PREV_NODUP:
7287 rc = mdb_cursor_prev(mc, key, data, op);
7290 rc = mdb_cursor_first(mc, key, data);
7293 mfunc = mdb_cursor_first;
7295 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
7299 if (mc->mc_xcursor == NULL) {
7300 rc = MDB_INCOMPATIBLE;
7303 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top])) {
7304 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7309 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7310 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7311 MDB_GET_KEY(leaf, key);
7312 rc = mdb_node_read(mc, leaf, data);
7316 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7320 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
7323 rc = mdb_cursor_last(mc, key, data);
7326 mfunc = mdb_cursor_last;
7329 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
7334 if (mc->mc_flags & C_DEL)
7335 mc->mc_flags ^= C_DEL;
7340 /** Touch all the pages in the cursor stack. Set mc_top.
7341 * Makes sure all the pages are writable, before attempting a write operation.
7342 * @param[in] mc The cursor to operate on.
7345 mdb_cursor_touch(MDB_cursor *mc)
7347 int rc = MDB_SUCCESS;
7349 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & (DB_DIRTY|DB_DUPDATA))) {
7350 /* Touch DB record of named DB */
7353 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
7355 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
7356 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
7359 *mc->mc_dbflag |= DB_DIRTY;
7364 rc = mdb_page_touch(mc);
7365 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
7366 mc->mc_top = mc->mc_snum-1;
7371 /** Do not spill pages to disk if txn is getting full, may fail instead */
7372 #define MDB_NOSPILL 0x8000
7375 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
7379 MDB_node *leaf = NULL;
7380 MDB_page *fp, *mp, *sub_root = NULL;
7382 MDB_val xdata, *rdata, dkey, olddata;
7384 int do_sub = 0, insert_key, insert_data;
7385 unsigned int mcount = 0, dcount = 0, nospill;
7388 unsigned int nflags;
7391 if (mc == NULL || key == NULL)
7394 env = mc->mc_txn->mt_env;
7396 /* Check this first so counter will always be zero on any
7399 if (flags & MDB_MULTIPLE) {
7400 dcount = data[1].mv_size;
7401 data[1].mv_size = 0;
7402 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
7403 return MDB_INCOMPATIBLE;
7406 nospill = flags & MDB_NOSPILL;
7407 flags &= ~MDB_NOSPILL;
7409 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7410 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7412 if (key->mv_size-1 >= ENV_MAXKEY(env))
7413 return MDB_BAD_VALSIZE;
7415 #if SIZE_MAX > MAXDATASIZE
7416 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
7417 return MDB_BAD_VALSIZE;
7419 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
7420 return MDB_BAD_VALSIZE;
7423 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
7424 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
7428 if (flags == MDB_CURRENT) {
7429 if (!(mc->mc_flags & C_INITIALIZED))
7432 } else if (mc->mc_db->md_root == P_INVALID) {
7433 /* new database, cursor has nothing to point to */
7436 mc->mc_flags &= ~C_INITIALIZED;
7441 if (flags & MDB_APPEND) {
7443 rc = mdb_cursor_last(mc, &k2, &d2);
7445 rc = mc->mc_dbx->md_cmp(key, &k2);
7448 mc->mc_ki[mc->mc_top]++;
7450 /* new key is <= last key */
7455 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
7457 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
7458 DPRINTF(("duplicate key [%s]", DKEY(key)));
7460 return MDB_KEYEXIST;
7462 if (rc && rc != MDB_NOTFOUND)
7466 if (mc->mc_flags & C_DEL)
7467 mc->mc_flags ^= C_DEL;
7469 /* Cursor is positioned, check for room in the dirty list */
7471 if (flags & MDB_MULTIPLE) {
7473 xdata.mv_size = data->mv_size * dcount;
7477 if ((rc2 = mdb_page_spill(mc, key, rdata)))
7481 if (rc == MDB_NO_ROOT) {
7483 /* new database, write a root leaf page */
7484 DPUTS("allocating new root leaf page");
7485 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
7488 mdb_cursor_push(mc, np);
7489 mc->mc_db->md_root = np->mp_pgno;
7490 mc->mc_db->md_depth++;
7491 *mc->mc_dbflag |= DB_DIRTY;
7492 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
7494 np->mp_flags |= P_LEAF2;
7495 mc->mc_flags |= C_INITIALIZED;
7497 /* make sure all cursor pages are writable */
7498 rc2 = mdb_cursor_touch(mc);
7503 insert_key = insert_data = rc;
7505 /* The key does not exist */
7506 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
7507 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
7508 LEAFSIZE(key, data) > env->me_nodemax)
7510 /* Too big for a node, insert in sub-DB. Set up an empty
7511 * "old sub-page" for prep_subDB to expand to a full page.
7513 fp_flags = P_LEAF|P_DIRTY;
7515 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
7516 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
7517 olddata.mv_size = PAGEHDRSZ;
7521 /* there's only a key anyway, so this is a no-op */
7522 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
7524 unsigned int ksize = mc->mc_db->md_pad;
7525 if (key->mv_size != ksize)
7526 return MDB_BAD_VALSIZE;
7527 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
7528 memcpy(ptr, key->mv_data, ksize);
7530 /* if overwriting slot 0 of leaf, need to
7531 * update branch key if there is a parent page
7533 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7534 unsigned short dtop = 1;
7536 /* slot 0 is always an empty key, find real slot */
7537 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7541 if (mc->mc_ki[mc->mc_top])
7542 rc2 = mdb_update_key(mc, key);
7553 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7554 olddata.mv_size = NODEDSZ(leaf);
7555 olddata.mv_data = NODEDATA(leaf);
7558 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
7559 /* Prepare (sub-)page/sub-DB to accept the new item,
7560 * if needed. fp: old sub-page or a header faking
7561 * it. mp: new (sub-)page. offset: growth in page
7562 * size. xdata: node data with new page or DB.
7564 unsigned i, offset = 0;
7565 mp = fp = xdata.mv_data = env->me_pbuf;
7566 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
7568 /* Was a single item before, must convert now */
7569 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7571 /* Just overwrite the current item */
7572 if (flags == MDB_CURRENT)
7574 dcmp = mc->mc_dbx->md_dcmp;
7575 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
7576 dcmp = mdb_cmp_clong;
7577 /* does data match? */
7578 if (!dcmp(data, &olddata)) {
7579 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
7580 return MDB_KEYEXIST;
7585 /* Back up original data item */
7586 dkey.mv_size = olddata.mv_size;
7587 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
7589 /* Make sub-page header for the dup items, with dummy body */
7590 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
7591 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
7592 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
7593 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7594 fp->mp_flags |= P_LEAF2;
7595 fp->mp_pad = data->mv_size;
7596 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
7598 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
7599 (dkey.mv_size & 1) + (data->mv_size & 1);
7601 fp->mp_upper = xdata.mv_size - PAGEBASE;
7602 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
7603 } else if (leaf->mn_flags & F_SUBDATA) {
7604 /* Data is on sub-DB, just store it */
7605 flags |= F_DUPDATA|F_SUBDATA;
7608 /* Data is on sub-page */
7609 fp = olddata.mv_data;
7612 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7613 offset = EVEN(NODESIZE + sizeof(indx_t) +
7617 offset = fp->mp_pad;
7618 if (SIZELEFT(fp) < offset) {
7619 offset *= 4; /* space for 4 more */
7622 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
7624 fp->mp_flags |= P_DIRTY;
7625 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
7626 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
7630 xdata.mv_size = olddata.mv_size + offset;
7633 fp_flags = fp->mp_flags;
7634 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
7635 /* Too big for a sub-page, convert to sub-DB */
7636 fp_flags &= ~P_SUBP;
7638 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7639 fp_flags |= P_LEAF2;
7640 dummy.md_pad = fp->mp_pad;
7641 dummy.md_flags = MDB_DUPFIXED;
7642 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7643 dummy.md_flags |= MDB_INTEGERKEY;
7649 dummy.md_branch_pages = 0;
7650 dummy.md_leaf_pages = 1;
7651 dummy.md_overflow_pages = 0;
7652 dummy.md_entries = NUMKEYS(fp);
7653 xdata.mv_size = sizeof(MDB_db);
7654 xdata.mv_data = &dummy;
7655 if ((rc = mdb_page_alloc(mc, 1, &mp)))
7657 offset = env->me_psize - olddata.mv_size;
7658 flags |= F_DUPDATA|F_SUBDATA;
7659 dummy.md_root = mp->mp_pgno;
7663 mp->mp_flags = fp_flags | P_DIRTY;
7664 mp->mp_pad = fp->mp_pad;
7665 mp->mp_lower = fp->mp_lower;
7666 mp->mp_upper = fp->mp_upper + offset;
7667 if (fp_flags & P_LEAF2) {
7668 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
7670 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
7671 olddata.mv_size - fp->mp_upper - PAGEBASE);
7672 memcpy((char *)(&mp->mp_ptrs), (char *)(&fp->mp_ptrs), NUMKEYS(fp) * sizeof(mp->mp_ptrs[0]));
7673 for (i=0; i<NUMKEYS(fp); i++)
7674 mp->mp_ptrs[i] += offset;
7682 mdb_node_del(mc, 0);
7686 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
7687 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
7688 return MDB_INCOMPATIBLE;
7689 /* overflow page overwrites need special handling */
7690 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7693 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
7695 memcpy(&pg, olddata.mv_data, sizeof(pg));
7696 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
7698 ovpages = omp->mp_pages;
7700 /* Is the ov page large enough? */
7701 if (ovpages >= dpages) {
7702 if (!(omp->mp_flags & P_DIRTY) &&
7703 (level || (env->me_flags & MDB_WRITEMAP)))
7705 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
7708 level = 0; /* dirty in this txn or clean */
7711 if (omp->mp_flags & P_DIRTY) {
7712 /* yes, overwrite it. Note in this case we don't
7713 * bother to try shrinking the page if the new data
7714 * is smaller than the overflow threshold.
7717 /* It is writable only in a parent txn */
7718 size_t sz = (size_t) env->me_psize * ovpages, off;
7719 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
7725 /* Note - this page is already counted in parent's dirty_room */
7726 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
7727 mdb_cassert(mc, rc2 == 0);
7728 /* Currently we make the page look as with put() in the
7729 * parent txn, in case the user peeks at MDB_RESERVEd
7730 * or unused parts. Some users treat ovpages specially.
7732 if (!(flags & MDB_RESERVE)) {
7733 /* Skip the part where LMDB will put *data.
7734 * Copy end of page, adjusting alignment so
7735 * compiler may copy words instead of bytes.
7737 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
7738 memcpy((size_t *)((char *)np + off),
7739 (size_t *)((char *)omp + off), sz - off);
7742 memcpy(np, omp, sz); /* Copy beginning of page */
7745 SETDSZ(leaf, data->mv_size);
7746 if (F_ISSET(flags, MDB_RESERVE))
7747 data->mv_data = METADATA(omp);
7749 memcpy(METADATA(omp), data->mv_data, data->mv_size);
7753 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
7755 } else if (data->mv_size == olddata.mv_size) {
7756 /* same size, just replace it. Note that we could
7757 * also reuse this node if the new data is smaller,
7758 * but instead we opt to shrink the node in that case.
7760 if (F_ISSET(flags, MDB_RESERVE))
7761 data->mv_data = olddata.mv_data;
7762 else if (!(mc->mc_flags & C_SUB))
7763 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
7765 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
7770 mdb_node_del(mc, 0);
7776 nflags = flags & NODE_ADD_FLAGS;
7777 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
7778 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
7779 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
7780 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
7782 nflags |= MDB_SPLIT_REPLACE;
7783 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
7785 /* There is room already in this leaf page. */
7786 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
7788 /* Adjust other cursors pointing to mp */
7789 MDB_cursor *m2, *m3;
7790 MDB_dbi dbi = mc->mc_dbi;
7791 unsigned i = mc->mc_top;
7792 MDB_page *mp = mc->mc_pg[i];
7794 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7795 if (mc->mc_flags & C_SUB)
7796 m3 = &m2->mc_xcursor->mx_cursor;
7799 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
7800 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
7803 XCURSOR_REFRESH(m3, i, mp);
7808 if (rc == MDB_SUCCESS) {
7809 /* Now store the actual data in the child DB. Note that we're
7810 * storing the user data in the keys field, so there are strict
7811 * size limits on dupdata. The actual data fields of the child
7812 * DB are all zero size.
7815 int xflags, new_dupdata;
7820 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7821 if (flags & MDB_CURRENT) {
7822 xflags = MDB_CURRENT|MDB_NOSPILL;
7824 mdb_xcursor_init1(mc, leaf);
7825 xflags = (flags & MDB_NODUPDATA) ?
7826 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
7829 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
7830 new_dupdata = (int)dkey.mv_size;
7831 /* converted, write the original data first */
7833 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
7836 /* we've done our job */
7839 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
7840 /* Adjust other cursors pointing to mp */
7842 MDB_xcursor *mx = mc->mc_xcursor;
7843 unsigned i = mc->mc_top;
7844 MDB_page *mp = mc->mc_pg[i];
7846 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7847 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7848 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7849 if (m2->mc_pg[i] == mp) {
7850 if (m2->mc_ki[i] == mc->mc_ki[i]) {
7851 mdb_xcursor_init2(m2, mx, new_dupdata);
7852 } else if (!insert_key) {
7853 XCURSOR_REFRESH(m2, i, mp);
7858 ecount = mc->mc_xcursor->mx_db.md_entries;
7859 if (flags & MDB_APPENDDUP)
7860 xflags |= MDB_APPEND;
7861 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
7862 if (flags & F_SUBDATA) {
7863 void *db = NODEDATA(leaf);
7864 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7866 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
7868 /* Increment count unless we just replaced an existing item. */
7870 mc->mc_db->md_entries++;
7872 /* Invalidate txn if we created an empty sub-DB */
7875 /* If we succeeded and the key didn't exist before,
7876 * make sure the cursor is marked valid.
7878 mc->mc_flags |= C_INITIALIZED;
7880 if (flags & MDB_MULTIPLE) {
7883 /* let caller know how many succeeded, if any */
7884 data[1].mv_size = mcount;
7885 if (mcount < dcount) {
7886 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
7887 insert_key = insert_data = 0;
7894 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
7897 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7902 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7908 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7909 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7911 if (!(mc->mc_flags & C_INITIALIZED))
7914 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7915 return MDB_NOTFOUND;
7917 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7920 rc = mdb_cursor_touch(mc);
7924 mp = mc->mc_pg[mc->mc_top];
7927 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7929 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7930 if (flags & MDB_NODUPDATA) {
7931 /* mdb_cursor_del0() will subtract the final entry */
7932 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7933 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7935 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7936 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7938 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7941 /* If sub-DB still has entries, we're done */
7942 if (mc->mc_xcursor->mx_db.md_entries) {
7943 if (leaf->mn_flags & F_SUBDATA) {
7944 /* update subDB info */
7945 void *db = NODEDATA(leaf);
7946 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7949 /* shrink fake page */
7950 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7951 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7952 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7953 /* fix other sub-DB cursors pointed at fake pages on this page */
7954 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7955 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7956 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7957 if (m2->mc_pg[mc->mc_top] == mp) {
7958 XCURSOR_REFRESH(m2, mc->mc_top, mp);
7962 mc->mc_db->md_entries--;
7965 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7967 /* otherwise fall thru and delete the sub-DB */
7970 if (leaf->mn_flags & F_SUBDATA) {
7971 /* add all the child DB's pages to the free list */
7972 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7977 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7978 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7979 rc = MDB_INCOMPATIBLE;
7983 /* add overflow pages to free list */
7984 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7988 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7989 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7990 (rc = mdb_ovpage_free(mc, omp)))
7995 return mdb_cursor_del0(mc);
7998 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8002 /** Allocate and initialize new pages for a database.
8003 * Set #MDB_TXN_ERROR on failure.
8004 * @param[in] mc a cursor on the database being added to.
8005 * @param[in] flags flags defining what type of page is being allocated.
8006 * @param[in] num the number of pages to allocate. This is usually 1,
8007 * unless allocating overflow pages for a large record.
8008 * @param[out] mp Address of a page, or NULL on failure.
8009 * @return 0 on success, non-zero on failure.
8012 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
8017 if ((rc = mdb_page_alloc(mc, num, &np)))
8019 DPRINTF(("allocated new mpage %"Yu", page size %u",
8020 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
8021 np->mp_flags = flags | P_DIRTY;
8022 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
8023 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
8026 mc->mc_db->md_branch_pages++;
8027 else if (IS_LEAF(np))
8028 mc->mc_db->md_leaf_pages++;
8029 else if (IS_OVERFLOW(np)) {
8030 mc->mc_db->md_overflow_pages += num;
8038 /** Calculate the size of a leaf node.
8039 * The size depends on the environment's page size; if a data item
8040 * is too large it will be put onto an overflow page and the node
8041 * size will only include the key and not the data. Sizes are always
8042 * rounded up to an even number of bytes, to guarantee 2-byte alignment
8043 * of the #MDB_node headers.
8044 * @param[in] env The environment handle.
8045 * @param[in] key The key for the node.
8046 * @param[in] data The data for the node.
8047 * @return The number of bytes needed to store the node.
8050 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
8054 sz = LEAFSIZE(key, data);
8055 if (sz > env->me_nodemax) {
8056 /* put on overflow page */
8057 sz -= data->mv_size - sizeof(pgno_t);
8060 return EVEN(sz + sizeof(indx_t));
8063 /** Calculate the size of a branch node.
8064 * The size should depend on the environment's page size but since
8065 * we currently don't support spilling large keys onto overflow
8066 * pages, it's simply the size of the #MDB_node header plus the
8067 * size of the key. Sizes are always rounded up to an even number
8068 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
8069 * @param[in] env The environment handle.
8070 * @param[in] key The key for the node.
8071 * @return The number of bytes needed to store the node.
8074 mdb_branch_size(MDB_env *env, MDB_val *key)
8079 if (sz > env->me_nodemax) {
8080 /* put on overflow page */
8081 /* not implemented */
8082 /* sz -= key->size - sizeof(pgno_t); */
8085 return sz + sizeof(indx_t);
8088 /** Add a node to the page pointed to by the cursor.
8089 * Set #MDB_TXN_ERROR on failure.
8090 * @param[in] mc The cursor for this operation.
8091 * @param[in] indx The index on the page where the new node should be added.
8092 * @param[in] key The key for the new node.
8093 * @param[in] data The data for the new node, if any.
8094 * @param[in] pgno The page number, if adding a branch node.
8095 * @param[in] flags Flags for the node.
8096 * @return 0 on success, non-zero on failure. Possible errors are:
8098 * <li>ENOMEM - failed to allocate overflow pages for the node.
8099 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
8100 * should never happen since all callers already calculate the
8101 * page's free space before calling this function.
8105 mdb_node_add(MDB_cursor *mc, indx_t indx,
8106 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
8109 size_t node_size = NODESIZE;
8113 MDB_page *mp = mc->mc_pg[mc->mc_top];
8114 MDB_page *ofp = NULL; /* overflow page */
8118 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
8120 DPRINTF(("add to %s %spage %"Yu" index %i, data size %"Z"u key size %"Z"u [%s]",
8121 IS_LEAF(mp) ? "leaf" : "branch",
8122 IS_SUBP(mp) ? "sub-" : "",
8123 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
8124 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
8127 /* Move higher keys up one slot. */
8128 int ksize = mc->mc_db->md_pad, dif;
8129 char *ptr = LEAF2KEY(mp, indx, ksize);
8130 dif = NUMKEYS(mp) - indx;
8132 memmove(ptr+ksize, ptr, dif*ksize);
8133 /* insert new key */
8134 memcpy(ptr, key->mv_data, ksize);
8136 /* Just using these for counting */
8137 mp->mp_lower += sizeof(indx_t);
8138 mp->mp_upper -= ksize - sizeof(indx_t);
8142 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
8144 node_size += key->mv_size;
8146 mdb_cassert(mc, key && data);
8147 if (F_ISSET(flags, F_BIGDATA)) {
8148 /* Data already on overflow page. */
8149 node_size += sizeof(pgno_t);
8150 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
8151 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
8153 /* Put data on overflow page. */
8154 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
8155 data->mv_size, node_size+data->mv_size));
8156 node_size = EVEN(node_size + sizeof(pgno_t));
8157 if ((ssize_t)node_size > room)
8159 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
8161 DPRINTF(("allocated overflow page %"Yu, ofp->mp_pgno));
8165 node_size += data->mv_size;
8168 node_size = EVEN(node_size);
8169 if ((ssize_t)node_size > room)
8173 /* Move higher pointers up one slot. */
8174 for (i = NUMKEYS(mp); i > indx; i--)
8175 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
8177 /* Adjust free space offsets. */
8178 ofs = mp->mp_upper - node_size;
8179 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
8180 mp->mp_ptrs[indx] = ofs;
8182 mp->mp_lower += sizeof(indx_t);
8184 /* Write the node data. */
8185 node = NODEPTR(mp, indx);
8186 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
8187 node->mn_flags = flags;
8189 SETDSZ(node,data->mv_size);
8194 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8197 ndata = NODEDATA(node);
8199 if (F_ISSET(flags, F_BIGDATA))
8200 memcpy(ndata, data->mv_data, sizeof(pgno_t));
8201 else if (F_ISSET(flags, MDB_RESERVE))
8202 data->mv_data = ndata;
8204 memcpy(ndata, data->mv_data, data->mv_size);
8206 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
8207 ndata = METADATA(ofp);
8208 if (F_ISSET(flags, MDB_RESERVE))
8209 data->mv_data = ndata;
8211 memcpy(ndata, data->mv_data, data->mv_size);
8218 DPRINTF(("not enough room in page %"Yu", got %u ptrs",
8219 mdb_dbg_pgno(mp), NUMKEYS(mp)));
8220 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
8221 DPRINTF(("node size = %"Z"u", node_size));
8222 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8223 return MDB_PAGE_FULL;
8226 /** Delete the specified node from a page.
8227 * @param[in] mc Cursor pointing to the node to delete.
8228 * @param[in] ksize The size of a node. Only used if the page is
8229 * part of a #MDB_DUPFIXED database.
8232 mdb_node_del(MDB_cursor *mc, int ksize)
8234 MDB_page *mp = mc->mc_pg[mc->mc_top];
8235 indx_t indx = mc->mc_ki[mc->mc_top];
8237 indx_t i, j, numkeys, ptr;
8241 DPRINTF(("delete node %u on %s page %"Yu, indx,
8242 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
8243 numkeys = NUMKEYS(mp);
8244 mdb_cassert(mc, indx < numkeys);
8247 int x = numkeys - 1 - indx;
8248 base = LEAF2KEY(mp, indx, ksize);
8250 memmove(base, base + ksize, x * ksize);
8251 mp->mp_lower -= sizeof(indx_t);
8252 mp->mp_upper += ksize - sizeof(indx_t);
8256 node = NODEPTR(mp, indx);
8257 sz = NODESIZE + node->mn_ksize;
8259 if (F_ISSET(node->mn_flags, F_BIGDATA))
8260 sz += sizeof(pgno_t);
8262 sz += NODEDSZ(node);
8266 ptr = mp->mp_ptrs[indx];
8267 for (i = j = 0; i < numkeys; i++) {
8269 mp->mp_ptrs[j] = mp->mp_ptrs[i];
8270 if (mp->mp_ptrs[i] < ptr)
8271 mp->mp_ptrs[j] += sz;
8276 base = (char *)mp + mp->mp_upper + PAGEBASE;
8277 memmove(base + sz, base, ptr - mp->mp_upper);
8279 mp->mp_lower -= sizeof(indx_t);
8283 /** Compact the main page after deleting a node on a subpage.
8284 * @param[in] mp The main page to operate on.
8285 * @param[in] indx The index of the subpage on the main page.
8288 mdb_node_shrink(MDB_page *mp, indx_t indx)
8293 indx_t delta, nsize, len, ptr;
8296 node = NODEPTR(mp, indx);
8297 sp = (MDB_page *)NODEDATA(node);
8298 delta = SIZELEFT(sp);
8299 nsize = NODEDSZ(node) - delta;
8301 /* Prepare to shift upward, set len = length(subpage part to shift) */
8305 return; /* do not make the node uneven-sized */
8307 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
8308 for (i = NUMKEYS(sp); --i >= 0; )
8309 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
8312 sp->mp_upper = sp->mp_lower;
8313 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
8314 SETDSZ(node, nsize);
8316 /* Shift <lower nodes...initial part of subpage> upward */
8317 base = (char *)mp + mp->mp_upper + PAGEBASE;
8318 memmove(base + delta, base, (char *)sp + len - base);
8320 ptr = mp->mp_ptrs[indx];
8321 for (i = NUMKEYS(mp); --i >= 0; ) {
8322 if (mp->mp_ptrs[i] <= ptr)
8323 mp->mp_ptrs[i] += delta;
8325 mp->mp_upper += delta;
8328 /** Initial setup of a sorted-dups cursor.
8329 * Sorted duplicates are implemented as a sub-database for the given key.
8330 * The duplicate data items are actually keys of the sub-database.
8331 * Operations on the duplicate data items are performed using a sub-cursor
8332 * initialized when the sub-database is first accessed. This function does
8333 * the preliminary setup of the sub-cursor, filling in the fields that
8334 * depend only on the parent DB.
8335 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8338 mdb_xcursor_init0(MDB_cursor *mc)
8340 MDB_xcursor *mx = mc->mc_xcursor;
8342 mx->mx_cursor.mc_xcursor = NULL;
8343 mx->mx_cursor.mc_txn = mc->mc_txn;
8344 mx->mx_cursor.mc_db = &mx->mx_db;
8345 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
8346 mx->mx_cursor.mc_dbi = mc->mc_dbi;
8347 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
8348 mx->mx_cursor.mc_snum = 0;
8349 mx->mx_cursor.mc_top = 0;
8350 MC_SET_OVPG(&mx->mx_cursor, NULL);
8351 mx->mx_cursor.mc_flags = C_SUB | (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP));
8352 mx->mx_dbx.md_name.mv_size = 0;
8353 mx->mx_dbx.md_name.mv_data = NULL;
8354 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
8355 mx->mx_dbx.md_dcmp = NULL;
8356 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
8359 /** Final setup of a sorted-dups cursor.
8360 * Sets up the fields that depend on the data from the main cursor.
8361 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8362 * @param[in] node The data containing the #MDB_db record for the
8363 * sorted-dup database.
8366 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
8368 MDB_xcursor *mx = mc->mc_xcursor;
8370 mx->mx_cursor.mc_flags &= C_SUB|C_ORIG_RDONLY|C_WRITEMAP;
8371 if (node->mn_flags & F_SUBDATA) {
8372 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
8373 mx->mx_cursor.mc_pg[0] = 0;
8374 mx->mx_cursor.mc_snum = 0;
8375 mx->mx_cursor.mc_top = 0;
8377 MDB_page *fp = NODEDATA(node);
8378 mx->mx_db.md_pad = 0;
8379 mx->mx_db.md_flags = 0;
8380 mx->mx_db.md_depth = 1;
8381 mx->mx_db.md_branch_pages = 0;
8382 mx->mx_db.md_leaf_pages = 1;
8383 mx->mx_db.md_overflow_pages = 0;
8384 mx->mx_db.md_entries = NUMKEYS(fp);
8385 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
8386 mx->mx_cursor.mc_snum = 1;
8387 mx->mx_cursor.mc_top = 0;
8388 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8389 mx->mx_cursor.mc_pg[0] = fp;
8390 mx->mx_cursor.mc_ki[0] = 0;
8391 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
8392 mx->mx_db.md_flags = MDB_DUPFIXED;
8393 mx->mx_db.md_pad = fp->mp_pad;
8394 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
8395 mx->mx_db.md_flags |= MDB_INTEGERKEY;
8398 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8399 mx->mx_db.md_root));
8400 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
8401 if (NEED_CMP_CLONG(mx->mx_dbx.md_cmp, mx->mx_db.md_pad))
8402 mx->mx_dbx.md_cmp = mdb_cmp_clong;
8406 /** Fixup a sorted-dups cursor due to underlying update.
8407 * Sets up some fields that depend on the data from the main cursor.
8408 * Almost the same as init1, but skips initialization steps if the
8409 * xcursor had already been used.
8410 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
8411 * @param[in] src_mx The xcursor of an up-to-date cursor.
8412 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
8415 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
8417 MDB_xcursor *mx = mc->mc_xcursor;
8420 mx->mx_cursor.mc_snum = 1;
8421 mx->mx_cursor.mc_top = 0;
8422 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8423 mx->mx_cursor.mc_ki[0] = 0;
8424 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
8425 #if UINT_MAX < MDB_SIZE_MAX /* matches mdb_xcursor_init1:NEED_CMP_CLONG() */
8426 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
8428 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
8431 mx->mx_db = src_mx->mx_db;
8432 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
8433 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8434 mx->mx_db.md_root));
8437 /** Initialize a cursor for a given transaction and database. */
8439 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
8442 mc->mc_backup = NULL;
8445 mc->mc_db = &txn->mt_dbs[dbi];
8446 mc->mc_dbx = &txn->mt_dbxs[dbi];
8447 mc->mc_dbflag = &txn->mt_dbflags[dbi];
8452 MC_SET_OVPG(mc, NULL);
8453 mc->mc_flags = txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
8454 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
8455 mdb_tassert(txn, mx != NULL);
8456 mc->mc_xcursor = mx;
8457 mdb_xcursor_init0(mc);
8459 mc->mc_xcursor = NULL;
8461 if (*mc->mc_dbflag & DB_STALE) {
8462 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
8467 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
8470 size_t size = sizeof(MDB_cursor);
8472 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
8475 if (txn->mt_flags & MDB_TXN_BLOCKED)
8478 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8481 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
8482 size += sizeof(MDB_xcursor);
8484 if ((mc = malloc(size)) != NULL) {
8485 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
8486 if (txn->mt_cursors) {
8487 mc->mc_next = txn->mt_cursors[dbi];
8488 txn->mt_cursors[dbi] = mc;
8489 mc->mc_flags |= C_UNTRACK;
8501 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
8503 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
8506 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
8509 if (txn->mt_flags & MDB_TXN_BLOCKED)
8512 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
8516 /* Return the count of duplicate data items for the current key */
8518 mdb_cursor_count(MDB_cursor *mc, mdb_size_t *countp)
8522 if (mc == NULL || countp == NULL)
8525 if (mc->mc_xcursor == NULL)
8526 return MDB_INCOMPATIBLE;
8528 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
8531 if (!(mc->mc_flags & C_INITIALIZED))
8535 return MDB_NOTFOUND;
8537 if (mc->mc_flags & C_EOF) {
8538 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
8539 return MDB_NOTFOUND;
8540 mc->mc_flags ^= C_EOF;
8543 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8544 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
8547 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
8550 *countp = mc->mc_xcursor->mx_db.md_entries;
8556 mdb_cursor_close(MDB_cursor *mc)
8559 MDB_CURSOR_UNREF(mc, 0);
8561 if (mc && !mc->mc_backup) {
8562 /* Remove from txn, if tracked.
8563 * A read-only txn (!C_UNTRACK) may have been freed already,
8564 * so do not peek inside it. Only write txns track cursors.
8566 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
8567 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
8568 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
8570 *prev = mc->mc_next;
8577 mdb_cursor_txn(MDB_cursor *mc)
8579 if (!mc) return NULL;
8584 mdb_cursor_dbi(MDB_cursor *mc)
8589 /** Replace the key for a branch node with a new key.
8590 * Set #MDB_TXN_ERROR on failure.
8591 * @param[in] mc Cursor pointing to the node to operate on.
8592 * @param[in] key The new key to use.
8593 * @return 0 on success, non-zero on failure.
8596 mdb_update_key(MDB_cursor *mc, MDB_val *key)
8602 int delta, ksize, oksize;
8603 indx_t ptr, i, numkeys, indx;
8606 indx = mc->mc_ki[mc->mc_top];
8607 mp = mc->mc_pg[mc->mc_top];
8608 node = NODEPTR(mp, indx);
8609 ptr = mp->mp_ptrs[indx];
8613 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
8614 k2.mv_data = NODEKEY(node);
8615 k2.mv_size = node->mn_ksize;
8616 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Yu,
8618 mdb_dkey(&k2, kbuf2),
8624 /* Sizes must be 2-byte aligned. */
8625 ksize = EVEN(key->mv_size);
8626 oksize = EVEN(node->mn_ksize);
8627 delta = ksize - oksize;
8629 /* Shift node contents if EVEN(key length) changed. */
8631 if (delta > 0 && SIZELEFT(mp) < delta) {
8633 /* not enough space left, do a delete and split */
8634 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
8635 pgno = NODEPGNO(node);
8636 mdb_node_del(mc, 0);
8637 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
8640 numkeys = NUMKEYS(mp);
8641 for (i = 0; i < numkeys; i++) {
8642 if (mp->mp_ptrs[i] <= ptr)
8643 mp->mp_ptrs[i] -= delta;
8646 base = (char *)mp + mp->mp_upper + PAGEBASE;
8647 len = ptr - mp->mp_upper + NODESIZE;
8648 memmove(base - delta, base, len);
8649 mp->mp_upper -= delta;
8651 node = NODEPTR(mp, indx);
8654 /* But even if no shift was needed, update ksize */
8655 if (node->mn_ksize != key->mv_size)
8656 node->mn_ksize = key->mv_size;
8659 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8665 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
8667 /** Perform \b act while tracking temporary cursor \b mn */
8668 #define WITH_CURSOR_TRACKING(mn, act) do { \
8669 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
8670 if ((mn).mc_flags & C_SUB) { \
8671 dummy.mc_flags = C_INITIALIZED; \
8672 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
8677 tracked->mc_next = *tp; \
8680 *tp = tracked->mc_next; \
8683 /** Move a node from csrc to cdst.
8686 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
8693 unsigned short flags;
8697 /* Mark src and dst as dirty. */
8698 if ((rc = mdb_page_touch(csrc)) ||
8699 (rc = mdb_page_touch(cdst)))
8702 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8703 key.mv_size = csrc->mc_db->md_pad;
8704 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
8706 data.mv_data = NULL;
8710 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
8711 mdb_cassert(csrc, !((size_t)srcnode & 1));
8712 srcpg = NODEPGNO(srcnode);
8713 flags = srcnode->mn_flags;
8714 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8715 unsigned int snum = csrc->mc_snum;
8717 /* must find the lowest key below src */
8718 rc = mdb_page_search_lowest(csrc);
8721 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8722 key.mv_size = csrc->mc_db->md_pad;
8723 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8725 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8726 key.mv_size = NODEKSZ(s2);
8727 key.mv_data = NODEKEY(s2);
8729 csrc->mc_snum = snum--;
8730 csrc->mc_top = snum;
8732 key.mv_size = NODEKSZ(srcnode);
8733 key.mv_data = NODEKEY(srcnode);
8735 data.mv_size = NODEDSZ(srcnode);
8736 data.mv_data = NODEDATA(srcnode);
8738 mn.mc_xcursor = NULL;
8739 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
8740 unsigned int snum = cdst->mc_snum;
8743 /* must find the lowest key below dst */
8744 mdb_cursor_copy(cdst, &mn);
8745 rc = mdb_page_search_lowest(&mn);
8748 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8749 bkey.mv_size = mn.mc_db->md_pad;
8750 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
8752 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8753 bkey.mv_size = NODEKSZ(s2);
8754 bkey.mv_data = NODEKEY(s2);
8756 mn.mc_snum = snum--;
8759 rc = mdb_update_key(&mn, &bkey);
8764 DPRINTF(("moving %s node %u [%s] on page %"Yu" to node %u on page %"Yu,
8765 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
8766 csrc->mc_ki[csrc->mc_top],
8768 csrc->mc_pg[csrc->mc_top]->mp_pgno,
8769 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
8771 /* Add the node to the destination page.
8773 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
8774 if (rc != MDB_SUCCESS)
8777 /* Delete the node from the source page.
8779 mdb_node_del(csrc, key.mv_size);
8782 /* Adjust other cursors pointing to mp */
8783 MDB_cursor *m2, *m3;
8784 MDB_dbi dbi = csrc->mc_dbi;
8785 MDB_page *mpd, *mps;
8787 mps = csrc->mc_pg[csrc->mc_top];
8788 /* If we're adding on the left, bump others up */
8790 mpd = cdst->mc_pg[csrc->mc_top];
8791 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8792 if (csrc->mc_flags & C_SUB)
8793 m3 = &m2->mc_xcursor->mx_cursor;
8796 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8799 m3->mc_pg[csrc->mc_top] == mpd &&
8800 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
8801 m3->mc_ki[csrc->mc_top]++;
8804 m3->mc_pg[csrc->mc_top] == mps &&
8805 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
8806 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8807 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8808 m3->mc_ki[csrc->mc_top-1]++;
8811 XCURSOR_REFRESH(m3, csrc->mc_top, m3->mc_pg[csrc->mc_top]);
8814 /* Adding on the right, bump others down */
8816 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8817 if (csrc->mc_flags & C_SUB)
8818 m3 = &m2->mc_xcursor->mx_cursor;
8821 if (m3 == csrc) continue;
8822 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8824 if (m3->mc_pg[csrc->mc_top] == mps) {
8825 if (!m3->mc_ki[csrc->mc_top]) {
8826 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8827 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8828 m3->mc_ki[csrc->mc_top-1]--;
8830 m3->mc_ki[csrc->mc_top]--;
8833 XCURSOR_REFRESH(m3, csrc->mc_top, m3->mc_pg[csrc->mc_top]);
8839 /* Update the parent separators.
8841 if (csrc->mc_ki[csrc->mc_top] == 0) {
8842 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
8843 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8844 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8846 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8847 key.mv_size = NODEKSZ(srcnode);
8848 key.mv_data = NODEKEY(srcnode);
8850 DPRINTF(("update separator for source page %"Yu" to [%s]",
8851 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
8852 mdb_cursor_copy(csrc, &mn);
8855 /* We want mdb_rebalance to find mn when doing fixups */
8856 WITH_CURSOR_TRACKING(mn,
8857 rc = mdb_update_key(&mn, &key));
8861 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8863 indx_t ix = csrc->mc_ki[csrc->mc_top];
8864 nullkey.mv_size = 0;
8865 csrc->mc_ki[csrc->mc_top] = 0;
8866 rc = mdb_update_key(csrc, &nullkey);
8867 csrc->mc_ki[csrc->mc_top] = ix;
8868 mdb_cassert(csrc, rc == MDB_SUCCESS);
8872 if (cdst->mc_ki[cdst->mc_top] == 0) {
8873 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
8874 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8875 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
8877 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
8878 key.mv_size = NODEKSZ(srcnode);
8879 key.mv_data = NODEKEY(srcnode);
8881 DPRINTF(("update separator for destination page %"Yu" to [%s]",
8882 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
8883 mdb_cursor_copy(cdst, &mn);
8886 /* We want mdb_rebalance to find mn when doing fixups */
8887 WITH_CURSOR_TRACKING(mn,
8888 rc = mdb_update_key(&mn, &key));
8892 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
8894 indx_t ix = cdst->mc_ki[cdst->mc_top];
8895 nullkey.mv_size = 0;
8896 cdst->mc_ki[cdst->mc_top] = 0;
8897 rc = mdb_update_key(cdst, &nullkey);
8898 cdst->mc_ki[cdst->mc_top] = ix;
8899 mdb_cassert(cdst, rc == MDB_SUCCESS);
8906 /** Merge one page into another.
8907 * The nodes from the page pointed to by \b csrc will
8908 * be copied to the page pointed to by \b cdst and then
8909 * the \b csrc page will be freed.
8910 * @param[in] csrc Cursor pointing to the source page.
8911 * @param[in] cdst Cursor pointing to the destination page.
8912 * @return 0 on success, non-zero on failure.
8915 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8917 MDB_page *psrc, *pdst;
8924 psrc = csrc->mc_pg[csrc->mc_top];
8925 pdst = cdst->mc_pg[cdst->mc_top];
8927 DPRINTF(("merging page %"Yu" into %"Yu, psrc->mp_pgno, pdst->mp_pgno));
8929 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8930 mdb_cassert(csrc, cdst->mc_snum > 1);
8932 /* Mark dst as dirty. */
8933 if ((rc = mdb_page_touch(cdst)))
8936 /* get dst page again now that we've touched it. */
8937 pdst = cdst->mc_pg[cdst->mc_top];
8939 /* Move all nodes from src to dst.
8941 j = nkeys = NUMKEYS(pdst);
8942 if (IS_LEAF2(psrc)) {
8943 key.mv_size = csrc->mc_db->md_pad;
8944 key.mv_data = METADATA(psrc);
8945 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8946 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8947 if (rc != MDB_SUCCESS)
8949 key.mv_data = (char *)key.mv_data + key.mv_size;
8952 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8953 srcnode = NODEPTR(psrc, i);
8954 if (i == 0 && IS_BRANCH(psrc)) {
8957 mdb_cursor_copy(csrc, &mn);
8958 mn.mc_xcursor = NULL;
8959 /* must find the lowest key below src */
8960 rc = mdb_page_search_lowest(&mn);
8963 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8964 key.mv_size = mn.mc_db->md_pad;
8965 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8967 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8968 key.mv_size = NODEKSZ(s2);
8969 key.mv_data = NODEKEY(s2);
8972 key.mv_size = srcnode->mn_ksize;
8973 key.mv_data = NODEKEY(srcnode);
8976 data.mv_size = NODEDSZ(srcnode);
8977 data.mv_data = NODEDATA(srcnode);
8978 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8979 if (rc != MDB_SUCCESS)
8984 DPRINTF(("dst page %"Yu" now has %u keys (%.1f%% filled)",
8985 pdst->mp_pgno, NUMKEYS(pdst),
8986 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8988 /* Unlink the src page from parent and add to free list.
8991 mdb_node_del(csrc, 0);
8992 if (csrc->mc_ki[csrc->mc_top] == 0) {
8994 rc = mdb_update_key(csrc, &key);
9002 psrc = csrc->mc_pg[csrc->mc_top];
9003 /* If not operating on FreeDB, allow this page to be reused
9004 * in this txn. Otherwise just add to free list.
9006 rc = mdb_page_loose(csrc, psrc);
9010 csrc->mc_db->md_leaf_pages--;
9012 csrc->mc_db->md_branch_pages--;
9014 /* Adjust other cursors pointing to mp */
9015 MDB_cursor *m2, *m3;
9016 MDB_dbi dbi = csrc->mc_dbi;
9017 unsigned int top = csrc->mc_top;
9019 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9020 if (csrc->mc_flags & C_SUB)
9021 m3 = &m2->mc_xcursor->mx_cursor;
9024 if (m3 == csrc) continue;
9025 if (m3->mc_snum < csrc->mc_snum) continue;
9026 if (m3->mc_pg[top] == psrc) {
9027 m3->mc_pg[top] = pdst;
9028 m3->mc_ki[top] += nkeys;
9029 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
9030 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
9031 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
9035 XCURSOR_REFRESH(m3, top, m3->mc_pg[top]);
9039 unsigned int snum = cdst->mc_snum;
9040 uint16_t depth = cdst->mc_db->md_depth;
9041 mdb_cursor_pop(cdst);
9042 rc = mdb_rebalance(cdst);
9043 /* Did the tree height change? */
9044 if (depth != cdst->mc_db->md_depth)
9045 snum += cdst->mc_db->md_depth - depth;
9046 cdst->mc_snum = snum;
9047 cdst->mc_top = snum-1;
9052 /** Copy the contents of a cursor.
9053 * @param[in] csrc The cursor to copy from.
9054 * @param[out] cdst The cursor to copy to.
9057 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
9061 cdst->mc_txn = csrc->mc_txn;
9062 cdst->mc_dbi = csrc->mc_dbi;
9063 cdst->mc_db = csrc->mc_db;
9064 cdst->mc_dbx = csrc->mc_dbx;
9065 cdst->mc_snum = csrc->mc_snum;
9066 cdst->mc_top = csrc->mc_top;
9067 cdst->mc_flags = csrc->mc_flags;
9068 MC_SET_OVPG(cdst, MC_OVPG(csrc));
9070 for (i=0; i<csrc->mc_snum; i++) {
9071 cdst->mc_pg[i] = csrc->mc_pg[i];
9072 cdst->mc_ki[i] = csrc->mc_ki[i];
9076 /** Rebalance the tree after a delete operation.
9077 * @param[in] mc Cursor pointing to the page where rebalancing
9079 * @return 0 on success, non-zero on failure.
9082 mdb_rebalance(MDB_cursor *mc)
9086 unsigned int ptop, minkeys, thresh;
9090 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
9095 thresh = FILL_THRESHOLD;
9097 DPRINTF(("rebalancing %s page %"Yu" (has %u keys, %.1f%% full)",
9098 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
9099 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
9100 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
9102 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
9103 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
9104 DPRINTF(("no need to rebalance page %"Yu", above fill threshold",
9105 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
9109 if (mc->mc_snum < 2) {
9110 MDB_page *mp = mc->mc_pg[0];
9112 DPUTS("Can't rebalance a subpage, ignoring");
9115 if (NUMKEYS(mp) == 0) {
9116 DPUTS("tree is completely empty");
9117 mc->mc_db->md_root = P_INVALID;
9118 mc->mc_db->md_depth = 0;
9119 mc->mc_db->md_leaf_pages = 0;
9120 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
9123 /* Adjust cursors pointing to mp */
9126 mc->mc_flags &= ~C_INITIALIZED;
9128 MDB_cursor *m2, *m3;
9129 MDB_dbi dbi = mc->mc_dbi;
9131 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9132 if (mc->mc_flags & C_SUB)
9133 m3 = &m2->mc_xcursor->mx_cursor;
9136 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
9138 if (m3->mc_pg[0] == mp) {
9141 m3->mc_flags &= ~C_INITIALIZED;
9145 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
9147 DPUTS("collapsing root page!");
9148 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
9151 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
9152 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
9155 mc->mc_db->md_depth--;
9156 mc->mc_db->md_branch_pages--;
9157 mc->mc_ki[0] = mc->mc_ki[1];
9158 for (i = 1; i<mc->mc_db->md_depth; i++) {
9159 mc->mc_pg[i] = mc->mc_pg[i+1];
9160 mc->mc_ki[i] = mc->mc_ki[i+1];
9163 /* Adjust other cursors pointing to mp */
9164 MDB_cursor *m2, *m3;
9165 MDB_dbi dbi = mc->mc_dbi;
9167 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9168 if (mc->mc_flags & C_SUB)
9169 m3 = &m2->mc_xcursor->mx_cursor;
9172 if (m3 == mc) continue;
9173 if (!(m3->mc_flags & C_INITIALIZED))
9175 if (m3->mc_pg[0] == mp) {
9176 for (i=0; i<mc->mc_db->md_depth; i++) {
9177 m3->mc_pg[i] = m3->mc_pg[i+1];
9178 m3->mc_ki[i] = m3->mc_ki[i+1];
9186 DPUTS("root page doesn't need rebalancing");
9190 /* The parent (branch page) must have at least 2 pointers,
9191 * otherwise the tree is invalid.
9193 ptop = mc->mc_top-1;
9194 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
9196 /* Leaf page fill factor is below the threshold.
9197 * Try to move keys from left or right neighbor, or
9198 * merge with a neighbor page.
9203 mdb_cursor_copy(mc, &mn);
9204 mn.mc_xcursor = NULL;
9206 oldki = mc->mc_ki[mc->mc_top];
9207 if (mc->mc_ki[ptop] == 0) {
9208 /* We're the leftmost leaf in our parent.
9210 DPUTS("reading right neighbor");
9212 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
9213 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
9216 mn.mc_ki[mn.mc_top] = 0;
9217 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
9220 /* There is at least one neighbor to the left.
9222 DPUTS("reading left neighbor");
9224 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
9225 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
9228 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
9229 mc->mc_ki[mc->mc_top] = 0;
9233 DPRINTF(("found neighbor page %"Yu" (%u keys, %.1f%% full)",
9234 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
9235 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
9237 /* If the neighbor page is above threshold and has enough keys,
9238 * move one key from it. Otherwise we should try to merge them.
9239 * (A branch page must never have less than 2 keys.)
9241 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
9242 rc = mdb_node_move(&mn, mc, fromleft);
9244 /* if we inserted on left, bump position up */
9249 rc = mdb_page_merge(&mn, mc);
9251 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
9252 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
9253 /* We want mdb_rebalance to find mn when doing fixups */
9254 WITH_CURSOR_TRACKING(mn,
9255 rc = mdb_page_merge(mc, &mn));
9256 mdb_cursor_copy(&mn, mc);
9258 mc->mc_flags &= ~C_EOF;
9260 mc->mc_ki[mc->mc_top] = oldki;
9264 /** Complete a delete operation started by #mdb_cursor_del(). */
9266 mdb_cursor_del0(MDB_cursor *mc)
9272 MDB_cursor *m2, *m3;
9273 MDB_dbi dbi = mc->mc_dbi;
9275 ki = mc->mc_ki[mc->mc_top];
9276 mp = mc->mc_pg[mc->mc_top];
9277 mdb_node_del(mc, mc->mc_db->md_pad);
9278 mc->mc_db->md_entries--;
9280 /* Adjust other cursors pointing to mp */
9281 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9282 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9283 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9285 if (m3 == mc || m3->mc_snum < mc->mc_snum)
9287 if (m3->mc_pg[mc->mc_top] == mp) {
9288 if (m3->mc_ki[mc->mc_top] == ki) {
9289 m3->mc_flags |= C_DEL;
9290 if (mc->mc_db->md_flags & MDB_DUPSORT) {
9291 /* Sub-cursor referred into dataset which is gone */
9292 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
9295 } else if (m3->mc_ki[mc->mc_top] > ki) {
9296 m3->mc_ki[mc->mc_top]--;
9298 XCURSOR_REFRESH(m3, mc->mc_top, mp);
9302 rc = mdb_rebalance(mc);
9304 if (rc == MDB_SUCCESS) {
9305 /* DB is totally empty now, just bail out.
9306 * Other cursors adjustments were already done
9307 * by mdb_rebalance and aren't needed here.
9312 mp = mc->mc_pg[mc->mc_top];
9313 nkeys = NUMKEYS(mp);
9315 /* Adjust other cursors pointing to mp */
9316 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
9317 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9318 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9320 if (m3->mc_snum < mc->mc_snum)
9322 if (m3->mc_pg[mc->mc_top] == mp) {
9323 /* if m3 points past last node in page, find next sibling */
9324 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
9325 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9326 rc = mdb_cursor_sibling(m3, 1);
9327 if (rc == MDB_NOTFOUND) {
9328 m3->mc_flags |= C_EOF;
9333 if (mc->mc_db->md_flags & MDB_DUPSORT) {
9334 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
9335 /* If this node has dupdata, it may need to be reinited
9336 * because its data has moved.
9337 * If the xcursor was not initd it must be reinited.
9338 * Else if node points to a subDB, nothing is needed.
9339 * Else (xcursor was initd, not a subDB) needs mc_pg[0] reset.
9341 if (node->mn_flags & F_DUPDATA) {
9342 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
9343 if (!(node->mn_flags & F_SUBDATA))
9344 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9346 mdb_xcursor_init1(m3, node);
9347 m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL;
9354 mc->mc_flags |= C_DEL;
9358 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9363 mdb_del(MDB_txn *txn, MDB_dbi dbi,
9364 MDB_val *key, MDB_val *data)
9366 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9369 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9370 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9372 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
9373 /* must ignore any data */
9377 return mdb_del0(txn, dbi, key, data, 0);
9381 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
9382 MDB_val *key, MDB_val *data, unsigned flags)
9387 MDB_val rdata, *xdata;
9391 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
9393 mdb_cursor_init(&mc, txn, dbi, &mx);
9402 flags |= MDB_NODUPDATA;
9404 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
9406 /* let mdb_page_split know about this cursor if needed:
9407 * delete will trigger a rebalance; if it needs to move
9408 * a node from one page to another, it will have to
9409 * update the parent's separator key(s). If the new sepkey
9410 * is larger than the current one, the parent page may
9411 * run out of space, triggering a split. We need this
9412 * cursor to be consistent until the end of the rebalance.
9414 mc.mc_next = txn->mt_cursors[dbi];
9415 txn->mt_cursors[dbi] = &mc;
9416 rc = mdb_cursor_del(&mc, flags);
9417 txn->mt_cursors[dbi] = mc.mc_next;
9422 /** Split a page and insert a new node.
9423 * Set #MDB_TXN_ERROR on failure.
9424 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
9425 * The cursor will be updated to point to the actual page and index where
9426 * the node got inserted after the split.
9427 * @param[in] newkey The key for the newly inserted node.
9428 * @param[in] newdata The data for the newly inserted node.
9429 * @param[in] newpgno The page number, if the new node is a branch node.
9430 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
9431 * @return 0 on success, non-zero on failure.
9434 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
9435 unsigned int nflags)
9438 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
9441 int i, j, split_indx, nkeys, pmax;
9442 MDB_env *env = mc->mc_txn->mt_env;
9444 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
9445 MDB_page *copy = NULL;
9446 MDB_page *mp, *rp, *pp;
9451 mp = mc->mc_pg[mc->mc_top];
9452 newindx = mc->mc_ki[mc->mc_top];
9453 nkeys = NUMKEYS(mp);
9455 DPRINTF(("-----> splitting %s page %"Yu" and adding [%s] at index %i/%i",
9456 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
9457 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
9459 /* Create a right sibling. */
9460 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
9462 rp->mp_pad = mp->mp_pad;
9463 DPRINTF(("new right sibling: page %"Yu, rp->mp_pgno));
9465 /* Usually when splitting the root page, the cursor
9466 * height is 1. But when called from mdb_update_key,
9467 * the cursor height may be greater because it walks
9468 * up the stack while finding the branch slot to update.
9470 if (mc->mc_top < 1) {
9471 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
9473 /* shift current top to make room for new parent */
9474 for (i=mc->mc_snum; i>0; i--) {
9475 mc->mc_pg[i] = mc->mc_pg[i-1];
9476 mc->mc_ki[i] = mc->mc_ki[i-1];
9480 mc->mc_db->md_root = pp->mp_pgno;
9481 DPRINTF(("root split! new root = %"Yu, pp->mp_pgno));
9482 new_root = mc->mc_db->md_depth++;
9484 /* Add left (implicit) pointer. */
9485 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
9486 /* undo the pre-push */
9487 mc->mc_pg[0] = mc->mc_pg[1];
9488 mc->mc_ki[0] = mc->mc_ki[1];
9489 mc->mc_db->md_root = mp->mp_pgno;
9490 mc->mc_db->md_depth--;
9497 ptop = mc->mc_top-1;
9498 DPRINTF(("parent branch page is %"Yu, mc->mc_pg[ptop]->mp_pgno));
9501 mdb_cursor_copy(mc, &mn);
9502 mn.mc_xcursor = NULL;
9503 mn.mc_pg[mn.mc_top] = rp;
9504 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
9506 if (nflags & MDB_APPEND) {
9507 mn.mc_ki[mn.mc_top] = 0;
9509 split_indx = newindx;
9513 split_indx = (nkeys+1) / 2;
9518 unsigned int lsize, rsize, ksize;
9519 /* Move half of the keys to the right sibling */
9520 x = mc->mc_ki[mc->mc_top] - split_indx;
9521 ksize = mc->mc_db->md_pad;
9522 split = LEAF2KEY(mp, split_indx, ksize);
9523 rsize = (nkeys - split_indx) * ksize;
9524 lsize = (nkeys - split_indx) * sizeof(indx_t);
9525 mp->mp_lower -= lsize;
9526 rp->mp_lower += lsize;
9527 mp->mp_upper += rsize - lsize;
9528 rp->mp_upper -= rsize - lsize;
9529 sepkey.mv_size = ksize;
9530 if (newindx == split_indx) {
9531 sepkey.mv_data = newkey->mv_data;
9533 sepkey.mv_data = split;
9536 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
9537 memcpy(rp->mp_ptrs, split, rsize);
9538 sepkey.mv_data = rp->mp_ptrs;
9539 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
9540 memcpy(ins, newkey->mv_data, ksize);
9541 mp->mp_lower += sizeof(indx_t);
9542 mp->mp_upper -= ksize - sizeof(indx_t);
9545 memcpy(rp->mp_ptrs, split, x * ksize);
9546 ins = LEAF2KEY(rp, x, ksize);
9547 memcpy(ins, newkey->mv_data, ksize);
9548 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
9549 rp->mp_lower += sizeof(indx_t);
9550 rp->mp_upper -= ksize - sizeof(indx_t);
9551 mc->mc_ki[mc->mc_top] = x;
9554 int psize, nsize, k;
9555 /* Maximum free space in an empty page */
9556 pmax = env->me_psize - PAGEHDRSZ;
9558 nsize = mdb_leaf_size(env, newkey, newdata);
9560 nsize = mdb_branch_size(env, newkey);
9561 nsize = EVEN(nsize);
9563 /* grab a page to hold a temporary copy */
9564 copy = mdb_page_malloc(mc->mc_txn, 1);
9569 copy->mp_pgno = mp->mp_pgno;
9570 copy->mp_flags = mp->mp_flags;
9571 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
9572 copy->mp_upper = env->me_psize - PAGEBASE;
9574 /* prepare to insert */
9575 for (i=0, j=0; i<nkeys; i++) {
9577 copy->mp_ptrs[j++] = 0;
9579 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
9582 /* When items are relatively large the split point needs
9583 * to be checked, because being off-by-one will make the
9584 * difference between success or failure in mdb_node_add.
9586 * It's also relevant if a page happens to be laid out
9587 * such that one half of its nodes are all "small" and
9588 * the other half of its nodes are "large." If the new
9589 * item is also "large" and falls on the half with
9590 * "large" nodes, it also may not fit.
9592 * As a final tweak, if the new item goes on the last
9593 * spot on the page (and thus, onto the new page), bias
9594 * the split so the new page is emptier than the old page.
9595 * This yields better packing during sequential inserts.
9597 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
9598 /* Find split point */
9600 if (newindx <= split_indx || newindx >= nkeys) {
9602 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
9607 for (; i!=k; i+=j) {
9612 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9613 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
9615 if (F_ISSET(node->mn_flags, F_BIGDATA))
9616 psize += sizeof(pgno_t);
9618 psize += NODEDSZ(node);
9620 psize = EVEN(psize);
9622 if (psize > pmax || i == k-j) {
9623 split_indx = i + (j<0);
9628 if (split_indx == newindx) {
9629 sepkey.mv_size = newkey->mv_size;
9630 sepkey.mv_data = newkey->mv_data;
9632 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
9633 sepkey.mv_size = node->mn_ksize;
9634 sepkey.mv_data = NODEKEY(node);
9639 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
9641 /* Copy separator key to the parent.
9643 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
9644 int snum = mc->mc_snum;
9648 /* We want other splits to find mn when doing fixups */
9649 WITH_CURSOR_TRACKING(mn,
9650 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
9655 if (mc->mc_snum > snum) {
9658 /* Right page might now have changed parent.
9659 * Check if left page also changed parent.
9661 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9662 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9663 for (i=0; i<ptop; i++) {
9664 mc->mc_pg[i] = mn.mc_pg[i];
9665 mc->mc_ki[i] = mn.mc_ki[i];
9667 mc->mc_pg[ptop] = mn.mc_pg[ptop];
9668 if (mn.mc_ki[ptop]) {
9669 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
9671 /* find right page's left sibling */
9672 mc->mc_ki[ptop] = mn.mc_ki[ptop];
9673 rc = mdb_cursor_sibling(mc, 0);
9678 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
9681 if (rc != MDB_SUCCESS) {
9682 if (rc == MDB_NOTFOUND) /* improper mdb_cursor_sibling() result */
9686 if (nflags & MDB_APPEND) {
9687 mc->mc_pg[mc->mc_top] = rp;
9688 mc->mc_ki[mc->mc_top] = 0;
9689 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
9692 for (i=0; i<mc->mc_top; i++)
9693 mc->mc_ki[i] = mn.mc_ki[i];
9694 } else if (!IS_LEAF2(mp)) {
9696 mc->mc_pg[mc->mc_top] = rp;
9701 rkey.mv_data = newkey->mv_data;
9702 rkey.mv_size = newkey->mv_size;
9708 /* Update index for the new key. */
9709 mc->mc_ki[mc->mc_top] = j;
9711 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9712 rkey.mv_data = NODEKEY(node);
9713 rkey.mv_size = node->mn_ksize;
9715 xdata.mv_data = NODEDATA(node);
9716 xdata.mv_size = NODEDSZ(node);
9719 pgno = NODEPGNO(node);
9720 flags = node->mn_flags;
9723 if (!IS_LEAF(mp) && j == 0) {
9724 /* First branch index doesn't need key data. */
9728 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
9734 mc->mc_pg[mc->mc_top] = copy;
9739 } while (i != split_indx);
9741 nkeys = NUMKEYS(copy);
9742 for (i=0; i<nkeys; i++)
9743 mp->mp_ptrs[i] = copy->mp_ptrs[i];
9744 mp->mp_lower = copy->mp_lower;
9745 mp->mp_upper = copy->mp_upper;
9746 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
9747 env->me_psize - copy->mp_upper - PAGEBASE);
9749 /* reset back to original page */
9750 if (newindx < split_indx) {
9751 mc->mc_pg[mc->mc_top] = mp;
9753 mc->mc_pg[mc->mc_top] = rp;
9755 /* Make sure mc_ki is still valid.
9757 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9758 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9759 for (i=0; i<=ptop; i++) {
9760 mc->mc_pg[i] = mn.mc_pg[i];
9761 mc->mc_ki[i] = mn.mc_ki[i];
9765 if (nflags & MDB_RESERVE) {
9766 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
9767 if (!(node->mn_flags & F_BIGDATA))
9768 newdata->mv_data = NODEDATA(node);
9771 if (newindx >= split_indx) {
9772 mc->mc_pg[mc->mc_top] = rp;
9774 /* Make sure mc_ki is still valid.
9776 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9777 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9778 for (i=0; i<=ptop; i++) {
9779 mc->mc_pg[i] = mn.mc_pg[i];
9780 mc->mc_ki[i] = mn.mc_ki[i];
9787 /* Adjust other cursors pointing to mp */
9788 MDB_cursor *m2, *m3;
9789 MDB_dbi dbi = mc->mc_dbi;
9790 nkeys = NUMKEYS(mp);
9792 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9793 if (mc->mc_flags & C_SUB)
9794 m3 = &m2->mc_xcursor->mx_cursor;
9799 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9803 /* sub cursors may be on different DB */
9804 if (m3->mc_pg[0] != mp)
9807 for (k=new_root; k>=0; k--) {
9808 m3->mc_ki[k+1] = m3->mc_ki[k];
9809 m3->mc_pg[k+1] = m3->mc_pg[k];
9811 if (m3->mc_ki[0] >= nkeys) {
9816 m3->mc_pg[0] = mc->mc_pg[0];
9820 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
9821 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
9822 m3->mc_ki[mc->mc_top]++;
9823 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9824 m3->mc_pg[mc->mc_top] = rp;
9825 m3->mc_ki[mc->mc_top] -= nkeys;
9826 for (i=0; i<mc->mc_top; i++) {
9827 m3->mc_ki[i] = mn.mc_ki[i];
9828 m3->mc_pg[i] = mn.mc_pg[i];
9831 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
9832 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
9836 XCURSOR_REFRESH(m3, mc->mc_top, m3->mc_pg[mc->mc_top]);
9839 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
9842 if (copy) /* tmp page */
9843 mdb_page_free(env, copy);
9845 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9850 mdb_put(MDB_txn *txn, MDB_dbi dbi,
9851 MDB_val *key, MDB_val *data, unsigned int flags)
9857 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9860 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
9863 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9864 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9866 mdb_cursor_init(&mc, txn, dbi, &mx);
9867 mc.mc_next = txn->mt_cursors[dbi];
9868 txn->mt_cursors[dbi] = &mc;
9869 rc = mdb_cursor_put(&mc, key, data, flags);
9870 txn->mt_cursors[dbi] = mc.mc_next;
9875 #define MDB_WBUF (1024*1024)
9877 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
9879 /** State needed for a double-buffering compacting copy. */
9880 typedef struct mdb_copy {
9883 pthread_mutex_t mc_mutex;
9884 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
9889 pgno_t mc_next_pgno;
9891 int mc_toggle; /**< Buffer number in provider */
9892 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
9893 /** Error code. Never cleared if set. Both threads can set nonzero
9894 * to fail the copy. Not mutex-protected, LMDB expects atomic int.
9896 volatile int mc_error;
9899 /** Dedicated writer thread for compacting copy. */
9900 static THREAD_RET ESECT CALL_CONV
9901 mdb_env_copythr(void *arg)
9905 int toggle = 0, wsize, rc;
9908 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9911 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9915 sigaddset(&set, SIGPIPE);
9916 if ((rc = pthread_sigmask(SIG_BLOCK, &set, NULL)) != 0)
9921 pthread_mutex_lock(&my->mc_mutex);
9924 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9925 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
9927 wsize = my->mc_wlen[toggle];
9928 ptr = my->mc_wbuf[toggle];
9931 while (wsize > 0 && !my->mc_error) {
9932 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9935 #if defined(SIGPIPE) && !defined(_WIN32)
9937 /* Collect the pending SIGPIPE, otherwise at least OS X
9938 * gives it to the process on thread-exit (ITS#8504).
9941 sigwait(&set, &tmp);
9945 } else if (len > 0) {
9958 /* If there's an overflow page tail, write it too */
9959 if (my->mc_olen[toggle]) {
9960 wsize = my->mc_olen[toggle];
9961 ptr = my->mc_over[toggle];
9962 my->mc_olen[toggle] = 0;
9965 my->mc_wlen[toggle] = 0;
9967 /* Return the empty buffer to provider */
9969 pthread_cond_signal(&my->mc_cond);
9971 pthread_mutex_unlock(&my->mc_mutex);
9972 return (THREAD_RET)0;
9976 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9978 * @param[in] my control structure.
9979 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9982 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9984 pthread_mutex_lock(&my->mc_mutex);
9985 my->mc_new += adjust;
9986 pthread_cond_signal(&my->mc_cond);
9987 while (my->mc_new & 2) /* both buffers in use */
9988 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9989 pthread_mutex_unlock(&my->mc_mutex);
9991 my->mc_toggle ^= (adjust & 1);
9992 /* Both threads reset mc_wlen, to be safe from threading errors */
9993 my->mc_wlen[my->mc_toggle] = 0;
9994 return my->mc_error;
9997 /** Depth-first tree traversal for compacting copy.
9998 * @param[in] my control structure.
9999 * @param[in,out] pg database root.
10000 * @param[in] flags includes #F_DUPDATA if it is a sorted-duplicate sub-DB.
10003 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
10005 MDB_cursor mc = {0};
10007 MDB_page *mo, *mp, *leaf;
10012 /* Empty DB, nothing to do */
10013 if (*pg == P_INVALID)
10014 return MDB_SUCCESS;
10017 mc.mc_txn = my->mc_txn;
10018 mc.mc_flags = my->mc_txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
10020 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
10023 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
10027 /* Make cursor pages writable */
10028 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
10032 for (i=0; i<mc.mc_top; i++) {
10033 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
10034 mc.mc_pg[i] = (MDB_page *)ptr;
10035 ptr += my->mc_env->me_psize;
10038 /* This is writable space for a leaf page. Usually not needed. */
10039 leaf = (MDB_page *)ptr;
10041 toggle = my->mc_toggle;
10042 while (mc.mc_snum > 0) {
10044 mp = mc.mc_pg[mc.mc_top];
10048 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
10049 for (i=0; i<n; i++) {
10050 ni = NODEPTR(mp, i);
10051 if (ni->mn_flags & F_BIGDATA) {
10055 /* Need writable leaf */
10057 mc.mc_pg[mc.mc_top] = leaf;
10058 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
10060 ni = NODEPTR(mp, i);
10063 memcpy(&pg, NODEDATA(ni), sizeof(pg));
10064 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
10065 rc = mdb_page_get(&mc, pg, &omp, NULL);
10068 if (my->mc_wlen[toggle] >= MDB_WBUF) {
10069 rc = mdb_env_cthr_toggle(my, 1);
10072 toggle = my->mc_toggle;
10074 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
10075 memcpy(mo, omp, my->mc_env->me_psize);
10076 mo->mp_pgno = my->mc_next_pgno;
10077 my->mc_next_pgno += omp->mp_pages;
10078 my->mc_wlen[toggle] += my->mc_env->me_psize;
10079 if (omp->mp_pages > 1) {
10080 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
10081 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
10082 rc = mdb_env_cthr_toggle(my, 1);
10085 toggle = my->mc_toggle;
10087 } else if (ni->mn_flags & F_SUBDATA) {
10090 /* Need writable leaf */
10092 mc.mc_pg[mc.mc_top] = leaf;
10093 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
10095 ni = NODEPTR(mp, i);
10098 memcpy(&db, NODEDATA(ni), sizeof(db));
10099 my->mc_toggle = toggle;
10100 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
10103 toggle = my->mc_toggle;
10104 memcpy(NODEDATA(ni), &db, sizeof(db));
10109 mc.mc_ki[mc.mc_top]++;
10110 if (mc.mc_ki[mc.mc_top] < n) {
10113 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
10115 rc = mdb_page_get(&mc, pg, &mp, NULL);
10120 mc.mc_ki[mc.mc_top] = 0;
10121 if (IS_BRANCH(mp)) {
10122 /* Whenever we advance to a sibling branch page,
10123 * we must proceed all the way down to its first leaf.
10125 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
10128 mc.mc_pg[mc.mc_top] = mp;
10132 if (my->mc_wlen[toggle] >= MDB_WBUF) {
10133 rc = mdb_env_cthr_toggle(my, 1);
10136 toggle = my->mc_toggle;
10138 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
10139 mdb_page_copy(mo, mp, my->mc_env->me_psize);
10140 mo->mp_pgno = my->mc_next_pgno++;
10141 my->mc_wlen[toggle] += my->mc_env->me_psize;
10143 /* Update parent if there is one */
10144 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
10145 SETPGNO(ni, mo->mp_pgno);
10146 mdb_cursor_pop(&mc);
10148 /* Otherwise we're done */
10158 /** Copy environment with compaction. */
10160 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
10165 MDB_txn *txn = NULL;
10167 pgno_t root, new_root;
10168 int rc = MDB_SUCCESS;
10171 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
10172 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
10176 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
10177 if (my.mc_wbuf[0] == NULL) {
10178 /* _aligned_malloc() sets errno, but we use Windows error codes */
10179 rc = ERROR_NOT_ENOUGH_MEMORY;
10183 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
10185 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
10187 #ifdef HAVE_MEMALIGN
10188 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
10189 if (my.mc_wbuf[0] == NULL) {
10196 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
10202 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
10203 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
10204 my.mc_next_pgno = NUM_METAS;
10207 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
10211 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
10215 mp = (MDB_page *)my.mc_wbuf[0];
10216 memset(mp, 0, NUM_METAS * env->me_psize);
10218 mp->mp_flags = P_META;
10219 mm = (MDB_meta *)METADATA(mp);
10220 mdb_env_init_meta0(env, mm);
10221 mm->mm_address = env->me_metas[0]->mm_address;
10223 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
10225 mp->mp_flags = P_META;
10226 *(MDB_meta *)METADATA(mp) = *mm;
10227 mm = (MDB_meta *)METADATA(mp);
10229 /* Set metapage 1 with current main DB */
10230 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
10231 if (root != P_INVALID) {
10232 /* Count free pages + freeDB pages. Subtract from last_pg
10233 * to find the new last_pg, which also becomes the new root.
10235 MDB_ID freecount = 0;
10238 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
10239 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
10240 freecount += *(MDB_ID *)data.mv_data;
10241 if (rc != MDB_NOTFOUND)
10243 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
10244 txn->mt_dbs[FREE_DBI].md_leaf_pages +
10245 txn->mt_dbs[FREE_DBI].md_overflow_pages;
10247 new_root = txn->mt_next_pgno - 1 - freecount;
10248 mm->mm_last_pg = new_root;
10249 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
10250 mm->mm_dbs[MAIN_DBI].md_root = new_root;
10252 /* When the DB is empty, handle it specially to
10253 * fix any breakage like page leaks from ITS#8174.
10255 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
10257 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
10258 mm->mm_txnid = 1; /* use metapage 1 */
10261 my.mc_wlen[0] = env->me_psize * NUM_METAS;
10263 rc = mdb_env_cwalk(&my, &root, 0);
10264 if (rc == MDB_SUCCESS && root != new_root) {
10265 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
10271 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
10272 rc = THREAD_FINISH(thr);
10273 mdb_txn_abort(txn);
10277 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
10278 if (my.mc_cond) CloseHandle(my.mc_cond);
10279 if (my.mc_mutex) CloseHandle(my.mc_mutex);
10281 free(my.mc_wbuf[0]);
10282 pthread_cond_destroy(&my.mc_cond);
10284 pthread_mutex_destroy(&my.mc_mutex);
10286 return rc ? rc : my.mc_error;
10289 /** Copy environment as-is. */
10291 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
10293 MDB_txn *txn = NULL;
10294 mdb_mutexref_t wmutex = NULL;
10296 mdb_size_t wsize, w3;
10300 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
10304 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
10307 /* Do the lock/unlock of the reader mutex before starting the
10308 * write txn. Otherwise other read txns could block writers.
10310 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
10314 if (env->me_txns) {
10315 /* We must start the actual read txn after blocking writers */
10316 mdb_txn_end(txn, MDB_END_RESET_TMP);
10318 /* Temporarily block writers until we snapshot the meta pages */
10319 wmutex = env->me_wmutex;
10320 if (LOCK_MUTEX(rc, env, wmutex))
10323 rc = mdb_txn_renew0(txn);
10325 UNLOCK_MUTEX(wmutex);
10330 wsize = env->me_psize * NUM_METAS;
10334 DO_WRITE(rc, fd, ptr, w2, len);
10338 } else if (len > 0) {
10344 /* Non-blocking or async handles are not supported */
10350 UNLOCK_MUTEX(wmutex);
10355 w3 = txn->mt_next_pgno * env->me_psize;
10357 mdb_size_t fsize = 0;
10358 if ((rc = mdb_fsize(env->me_fd, &fsize)))
10363 wsize = w3 - wsize;
10364 while (wsize > 0) {
10365 if (wsize > MAX_WRITE)
10369 DO_WRITE(rc, fd, ptr, w2, len);
10373 } else if (len > 0) {
10385 mdb_txn_abort(txn);
10390 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
10392 if (flags & MDB_CP_COMPACT)
10393 return mdb_env_copyfd1(env, fd);
10395 return mdb_env_copyfd0(env, fd);
10399 mdb_env_copyfd(MDB_env *env, HANDLE fd)
10401 return mdb_env_copyfd2(env, fd, 0);
10405 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
10409 HANDLE newfd = INVALID_HANDLE_VALUE;
10411 rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname);
10412 if (rc == MDB_SUCCESS) {
10413 rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd);
10414 mdb_fname_destroy(fname);
10416 if (rc == MDB_SUCCESS) {
10417 rc = mdb_env_copyfd2(env, newfd, flags);
10418 if (close(newfd) < 0 && rc == MDB_SUCCESS)
10425 mdb_env_copy(MDB_env *env, const char *path)
10427 return mdb_env_copy2(env, path, 0);
10431 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
10433 if (flag & ~CHANGEABLE)
10436 env->me_flags |= flag;
10438 env->me_flags &= ~flag;
10439 return MDB_SUCCESS;
10443 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
10448 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
10449 return MDB_SUCCESS;
10453 mdb_env_set_userctx(MDB_env *env, void *ctx)
10457 env->me_userctx = ctx;
10458 return MDB_SUCCESS;
10462 mdb_env_get_userctx(MDB_env *env)
10464 return env ? env->me_userctx : NULL;
10468 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
10473 env->me_assert_func = func;
10475 return MDB_SUCCESS;
10479 mdb_env_get_path(MDB_env *env, const char **arg)
10484 *arg = env->me_path;
10485 return MDB_SUCCESS;
10489 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
10495 return MDB_SUCCESS;
10498 /** Common code for #mdb_stat() and #mdb_env_stat().
10499 * @param[in] env the environment to operate in.
10500 * @param[in] db the #MDB_db record containing the stats to return.
10501 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
10502 * @return 0, this function always succeeds.
10505 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
10507 arg->ms_psize = env->me_psize;
10508 arg->ms_depth = db->md_depth;
10509 arg->ms_branch_pages = db->md_branch_pages;
10510 arg->ms_leaf_pages = db->md_leaf_pages;
10511 arg->ms_overflow_pages = db->md_overflow_pages;
10512 arg->ms_entries = db->md_entries;
10514 return MDB_SUCCESS;
10518 mdb_env_stat(MDB_env *env, MDB_stat *arg)
10522 if (env == NULL || arg == NULL)
10525 meta = mdb_env_pick_meta(env);
10527 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
10531 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
10535 if (env == NULL || arg == NULL)
10538 meta = mdb_env_pick_meta(env);
10539 arg->me_mapaddr = meta->mm_address;
10540 arg->me_last_pgno = meta->mm_last_pg;
10541 arg->me_last_txnid = meta->mm_txnid;
10543 arg->me_mapsize = env->me_mapsize;
10544 arg->me_maxreaders = env->me_maxreaders;
10545 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
10546 return MDB_SUCCESS;
10549 /** Set the default comparison functions for a database.
10550 * Called immediately after a database is opened to set the defaults.
10551 * The user can then override them with #mdb_set_compare() or
10552 * #mdb_set_dupsort().
10553 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
10554 * @param[in] dbi A database handle returned by #mdb_dbi_open()
10557 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
10559 uint16_t f = txn->mt_dbs[dbi].md_flags;
10561 txn->mt_dbxs[dbi].md_cmp =
10562 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
10563 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
10565 txn->mt_dbxs[dbi].md_dcmp =
10566 !(f & MDB_DUPSORT) ? 0 :
10567 ((f & MDB_INTEGERDUP)
10568 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
10569 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
10572 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
10578 int rc, dbflag, exact;
10579 unsigned int unused = 0, seq;
10583 if (flags & ~VALID_FLAGS)
10585 if (txn->mt_flags & MDB_TXN_BLOCKED)
10586 return MDB_BAD_TXN;
10591 if (flags & PERSISTENT_FLAGS) {
10592 uint16_t f2 = flags & PERSISTENT_FLAGS;
10593 /* make sure flag changes get committed */
10594 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
10595 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
10596 txn->mt_flags |= MDB_TXN_DIRTY;
10599 mdb_default_cmp(txn, MAIN_DBI);
10600 return MDB_SUCCESS;
10603 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
10604 mdb_default_cmp(txn, MAIN_DBI);
10607 /* Is the DB already open? */
10608 len = strlen(name);
10609 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
10610 if (!txn->mt_dbxs[i].md_name.mv_size) {
10611 /* Remember this free slot */
10612 if (!unused) unused = i;
10615 if (len == txn->mt_dbxs[i].md_name.mv_size &&
10616 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
10618 return MDB_SUCCESS;
10622 /* If no free slot and max hit, fail */
10623 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
10624 return MDB_DBS_FULL;
10626 /* Cannot mix named databases with some mainDB flags */
10627 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
10628 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
10630 /* Find the DB info */
10631 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
10634 key.mv_data = (void *)name;
10635 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
10636 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
10637 if (rc == MDB_SUCCESS) {
10638 /* make sure this is actually a DB */
10639 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
10640 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
10641 return MDB_INCOMPATIBLE;
10643 if (rc != MDB_NOTFOUND || !(flags & MDB_CREATE))
10645 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
10649 /* Done here so we cannot fail after creating a new DB */
10650 if ((namedup = strdup(name)) == NULL)
10654 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
10655 data.mv_size = sizeof(MDB_db);
10656 data.mv_data = &dummy;
10657 memset(&dummy, 0, sizeof(dummy));
10658 dummy.md_root = P_INVALID;
10659 dummy.md_flags = flags & PERSISTENT_FLAGS;
10660 WITH_CURSOR_TRACKING(mc,
10661 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA));
10662 dbflag |= DB_DIRTY;
10668 /* Got info, register DBI in this txn */
10669 unsigned int slot = unused ? unused : txn->mt_numdbs;
10670 txn->mt_dbxs[slot].md_name.mv_data = namedup;
10671 txn->mt_dbxs[slot].md_name.mv_size = len;
10672 txn->mt_dbxs[slot].md_rel = NULL;
10673 txn->mt_dbflags[slot] = dbflag;
10674 /* txn-> and env-> are the same in read txns, use
10675 * tmp variable to avoid undefined assignment
10677 seq = ++txn->mt_env->me_dbiseqs[slot];
10678 txn->mt_dbiseqs[slot] = seq;
10680 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
10682 mdb_default_cmp(txn, slot);
10692 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
10694 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
10697 if (txn->mt_flags & MDB_TXN_BLOCKED)
10698 return MDB_BAD_TXN;
10700 if (txn->mt_dbflags[dbi] & DB_STALE) {
10703 /* Stale, must read the DB's root. cursor_init does it for us. */
10704 mdb_cursor_init(&mc, txn, dbi, &mx);
10706 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
10709 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
10712 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
10714 ptr = env->me_dbxs[dbi].md_name.mv_data;
10715 /* If there was no name, this was already closed */
10717 env->me_dbxs[dbi].md_name.mv_data = NULL;
10718 env->me_dbxs[dbi].md_name.mv_size = 0;
10719 env->me_dbflags[dbi] = 0;
10720 env->me_dbiseqs[dbi]++;
10725 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
10727 /* We could return the flags for the FREE_DBI too but what's the point? */
10728 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10730 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
10731 return MDB_SUCCESS;
10734 /** Add all the DB's pages to the free list.
10735 * @param[in] mc Cursor on the DB to free.
10736 * @param[in] subs non-Zero to check for sub-DBs in this DB.
10737 * @return 0 on success, non-zero on failure.
10740 mdb_drop0(MDB_cursor *mc, int subs)
10744 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
10745 if (rc == MDB_SUCCESS) {
10746 MDB_txn *txn = mc->mc_txn;
10751 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
10752 * This also avoids any P_LEAF2 pages, which have no nodes.
10753 * Also if the DB doesn't have sub-DBs and has no overflow
10754 * pages, omit scanning leaves.
10756 if ((mc->mc_flags & C_SUB) ||
10757 (!subs && !mc->mc_db->md_overflow_pages))
10758 mdb_cursor_pop(mc);
10760 mdb_cursor_copy(mc, &mx);
10762 /* bump refcount for mx's pages */
10763 for (i=0; i<mc->mc_snum; i++)
10764 mdb_page_get(&mx, mc->mc_pg[i]->mp_pgno, &mx.mc_pg[i], NULL);
10766 while (mc->mc_snum > 0) {
10767 MDB_page *mp = mc->mc_pg[mc->mc_top];
10768 unsigned n = NUMKEYS(mp);
10770 for (i=0; i<n; i++) {
10771 ni = NODEPTR(mp, i);
10772 if (ni->mn_flags & F_BIGDATA) {
10775 memcpy(&pg, NODEDATA(ni), sizeof(pg));
10776 rc = mdb_page_get(mc, pg, &omp, NULL);
10779 mdb_cassert(mc, IS_OVERFLOW(omp));
10780 rc = mdb_midl_append_range(&txn->mt_free_pgs,
10781 pg, omp->mp_pages);
10784 mc->mc_db->md_overflow_pages -= omp->mp_pages;
10785 if (!mc->mc_db->md_overflow_pages && !subs)
10787 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
10788 mdb_xcursor_init1(mc, ni);
10789 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
10794 if (!subs && !mc->mc_db->md_overflow_pages)
10797 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
10799 for (i=0; i<n; i++) {
10801 ni = NODEPTR(mp, i);
10804 mdb_midl_xappend(txn->mt_free_pgs, pg);
10809 mc->mc_ki[mc->mc_top] = i;
10810 rc = mdb_cursor_sibling(mc, 1);
10812 if (rc != MDB_NOTFOUND)
10814 /* no more siblings, go back to beginning
10815 * of previous level.
10818 mdb_cursor_pop(mc);
10820 for (i=1; i<mc->mc_snum; i++) {
10822 mc->mc_pg[i] = mx.mc_pg[i];
10827 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
10830 txn->mt_flags |= MDB_TXN_ERROR;
10831 /* drop refcount for mx's pages */
10832 MDB_CURSOR_UNREF(&mx, 0);
10833 } else if (rc == MDB_NOTFOUND) {
10836 mc->mc_flags &= ~C_INITIALIZED;
10840 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
10842 MDB_cursor *mc, *m2;
10845 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10848 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
10851 if (TXN_DBI_CHANGED(txn, dbi))
10852 return MDB_BAD_DBI;
10854 rc = mdb_cursor_open(txn, dbi, &mc);
10858 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
10859 /* Invalidate the dropped DB's cursors */
10860 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
10861 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
10865 /* Can't delete the main DB */
10866 if (del && dbi >= CORE_DBS) {
10867 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
10869 txn->mt_dbflags[dbi] = DB_STALE;
10870 mdb_dbi_close(txn->mt_env, dbi);
10872 txn->mt_flags |= MDB_TXN_ERROR;
10875 /* reset the DB record, mark it dirty */
10876 txn->mt_dbflags[dbi] |= DB_DIRTY;
10877 txn->mt_dbs[dbi].md_depth = 0;
10878 txn->mt_dbs[dbi].md_branch_pages = 0;
10879 txn->mt_dbs[dbi].md_leaf_pages = 0;
10880 txn->mt_dbs[dbi].md_overflow_pages = 0;
10881 txn->mt_dbs[dbi].md_entries = 0;
10882 txn->mt_dbs[dbi].md_root = P_INVALID;
10884 txn->mt_flags |= MDB_TXN_DIRTY;
10887 mdb_cursor_close(mc);
10891 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10893 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10896 txn->mt_dbxs[dbi].md_cmp = cmp;
10897 return MDB_SUCCESS;
10900 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10902 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10905 txn->mt_dbxs[dbi].md_dcmp = cmp;
10906 return MDB_SUCCESS;
10909 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
10911 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10914 txn->mt_dbxs[dbi].md_rel = rel;
10915 return MDB_SUCCESS;
10918 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
10920 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10923 txn->mt_dbxs[dbi].md_relctx = ctx;
10924 return MDB_SUCCESS;
10928 mdb_env_get_maxkeysize(MDB_env *env)
10930 return ENV_MAXKEY(env);
10934 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
10936 unsigned int i, rdrs;
10939 int rc = 0, first = 1;
10943 if (!env->me_txns) {
10944 return func("(no reader locks)\n", ctx);
10946 rdrs = env->me_txns->mti_numreaders;
10947 mr = env->me_txns->mti_readers;
10948 for (i=0; i<rdrs; i++) {
10949 if (mr[i].mr_pid) {
10950 txnid_t txnid = mr[i].mr_txnid;
10951 sprintf(buf, txnid == (txnid_t)-1 ?
10952 "%10d %"Z"x -\n" : "%10d %"Z"x %"Yu"\n",
10953 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10956 rc = func(" pid thread txnid\n", ctx);
10960 rc = func(buf, ctx);
10966 rc = func("(no active readers)\n", ctx);
10971 /** Insert pid into list if not already present.
10972 * return -1 if already present.
10975 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10977 /* binary search of pid in list */
10979 unsigned cursor = 1;
10981 unsigned n = ids[0];
10984 unsigned pivot = n >> 1;
10985 cursor = base + pivot + 1;
10986 val = pid - ids[cursor];
10991 } else if ( val > 0 ) {
10996 /* found, so it's a duplicate */
11005 for (n = ids[0]; n > cursor; n--)
11012 mdb_reader_check(MDB_env *env, int *dead)
11018 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
11021 /** As #mdb_reader_check(). \b rlocked is set if caller locked #me_rmutex. */
11023 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
11025 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
11026 unsigned int i, j, rdrs;
11028 MDB_PID_T *pids, pid;
11029 int rc = MDB_SUCCESS, count = 0;
11031 rdrs = env->me_txns->mti_numreaders;
11032 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
11036 mr = env->me_txns->mti_readers;
11037 for (i=0; i<rdrs; i++) {
11038 pid = mr[i].mr_pid;
11039 if (pid && pid != env->me_pid) {
11040 if (mdb_pid_insert(pids, pid) == 0) {
11041 if (!mdb_reader_pid(env, Pidcheck, pid)) {
11042 /* Stale reader found */
11045 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
11046 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
11048 rdrs = 0; /* the above checked all readers */
11050 /* Recheck, a new process may have reused pid */
11051 if (mdb_reader_pid(env, Pidcheck, pid))
11055 for (; j<rdrs; j++)
11056 if (mr[j].mr_pid == pid) {
11057 DPRINTF(("clear stale reader pid %u txn %"Yd,
11058 (unsigned) pid, mr[j].mr_txnid));
11063 UNLOCK_MUTEX(rmutex);
11074 #ifdef MDB_ROBUST_SUPPORTED
11075 /** Handle #LOCK_MUTEX0() failure.
11076 * Try to repair the lock file if the mutex owner died.
11077 * @param[in] env the environment handle
11078 * @param[in] mutex LOCK_MUTEX0() mutex
11079 * @param[in] rc LOCK_MUTEX0() error (nonzero)
11080 * @return 0 on success with the mutex locked, or an error code on failure.
11083 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
11088 if (rc == MDB_OWNERDEAD) {
11089 /* We own the mutex. Clean up after dead previous owner. */
11091 rlocked = (mutex == env->me_rmutex);
11093 /* Keep mti_txnid updated, otherwise next writer can
11094 * overwrite data which latest meta page refers to.
11096 meta = mdb_env_pick_meta(env);
11097 env->me_txns->mti_txnid = meta->mm_txnid;
11098 /* env is hosed if the dead thread was ours */
11100 env->me_flags |= MDB_FATAL_ERROR;
11101 env->me_txn = NULL;
11105 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
11106 (rc ? "this process' env is hosed" : "recovering")));
11107 rc2 = mdb_reader_check0(env, rlocked, NULL);
11109 rc2 = mdb_mutex_consistent(mutex);
11110 if (rc || (rc = rc2)) {
11111 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
11112 UNLOCK_MUTEX(mutex);
11118 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
11123 #endif /* MDB_ROBUST_SUPPORTED */
11125 #if defined(_WIN32)
11126 /** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */
11128 utf8_to_utf16(const char *src, MDB_name *dst, int xtra)
11131 wchar_t *result = NULL;
11132 for (;;) { /* malloc result, then fill it in */
11133 need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need);
11140 result = malloc(sizeof(wchar_t) * (need + xtra));
11145 dst->mn_alloced = 1;
11146 dst->mn_len = need - 1;
11147 dst->mn_val = result;
11148 return MDB_SUCCESS;
11151 #endif /* defined(_WIN32) */