2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2016 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #if defined(MDB_VL32) || defined(__WIN64__)
39 #define _FILE_OFFSET_BITS 64
45 /* We use native NT APIs to setup the memory map, so that we can
46 * let the DB file grow incrementally instead of always preallocating
47 * the full size. These APIs are defined in <wdm.h> and <ntifs.h>
48 * but those headers are meant for driver-level development and
49 * conflict with the regular user-level headers, so we explicitly
50 * declare them here. Using these APIs also means we must link to
51 * ntdll.dll, which is not linked by default in user code.
54 NtCreateSection(OUT PHANDLE sh, IN ACCESS_MASK acc,
55 IN void * oa OPTIONAL,
56 IN PLARGE_INTEGER ms OPTIONAL,
57 IN ULONG pp, IN ULONG aa, IN HANDLE fh OPTIONAL);
59 typedef enum _SECTION_INHERIT {
65 NtMapViewOfSection(IN PHANDLE sh, IN HANDLE ph,
66 IN OUT PVOID *addr, IN ULONG_PTR zbits,
67 IN SIZE_T cs, IN OUT PLARGE_INTEGER off OPTIONAL,
68 IN OUT PSIZE_T vs, IN SECTION_INHERIT ih,
69 IN ULONG at, IN ULONG pp);
74 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
75 * as int64 which is wrong. MSVC doesn't define it at all, so just
79 #define MDB_THR_T DWORD
80 #include <sys/types.h>
83 # include <sys/param.h>
85 # define LITTLE_ENDIAN 1234
86 # define BIG_ENDIAN 4321
87 # define BYTE_ORDER LITTLE_ENDIAN
89 # define SSIZE_MAX INT_MAX
93 #include <sys/types.h>
95 #define MDB_PID_T pid_t
96 #define MDB_THR_T pthread_t
97 #include <sys/param.h>
100 #ifdef HAVE_SYS_FILE_H
101 #include <sys/file.h>
106 #if defined(__mips) && defined(__linux)
107 /* MIPS has cache coherency issues, requires explicit cache control */
108 #include <asm/cachectl.h>
109 extern int cacheflush(char *addr, int nbytes, int cache);
110 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
112 #define CACHEFLUSH(addr, bytes, cache)
115 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
116 /** fdatasync is broken on ext3/ext4fs on older kernels, see
117 * description in #mdb_env_open2 comments. You can safely
118 * define MDB_FDATASYNC_WORKS if this code will only be run
119 * on kernels 3.6 and newer.
121 #define BROKEN_FDATASYNC
127 #include <inttypes.h>
135 typedef SSIZE_T ssize_t;
140 #if defined(__sun) || defined(ANDROID)
141 /* Most platforms have posix_memalign, older may only have memalign */
142 #define HAVE_MEMALIGN 1
146 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
147 #include <netinet/in.h>
148 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
151 #if defined(__APPLE__) || defined (BSD)
152 # if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
153 # define MDB_USE_SYSV_SEM 1
155 # define MDB_FDATASYNC fsync
156 #elif defined(ANDROID)
157 # define MDB_FDATASYNC fsync
162 #ifdef MDB_USE_POSIX_SEM
163 # define MDB_USE_HASH 1
164 #include <semaphore.h>
165 #elif defined(MDB_USE_SYSV_SEM)
168 #ifdef _SEM_SEMUN_UNDEFINED
171 struct semid_ds *buf;
172 unsigned short *array;
174 #endif /* _SEM_SEMUN_UNDEFINED */
176 #define MDB_USE_POSIX_MUTEX 1
177 #endif /* MDB_USE_POSIX_SEM */
180 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
181 + defined(MDB_USE_POSIX_MUTEX) != 1
182 # error "Ambiguous shared-lock implementation"
186 #include <valgrind/memcheck.h>
187 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
188 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
189 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
190 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
191 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
193 #define VGMEMP_CREATE(h,r,z)
194 #define VGMEMP_ALLOC(h,a,s)
195 #define VGMEMP_FREE(h,a)
196 #define VGMEMP_DESTROY(h)
197 #define VGMEMP_DEFINED(a,s)
201 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
202 /* Solaris just defines one or the other */
203 # define LITTLE_ENDIAN 1234
204 # define BIG_ENDIAN 4321
205 # ifdef _LITTLE_ENDIAN
206 # define BYTE_ORDER LITTLE_ENDIAN
208 # define BYTE_ORDER BIG_ENDIAN
211 # define BYTE_ORDER __BYTE_ORDER
215 #ifndef LITTLE_ENDIAN
216 #define LITTLE_ENDIAN __LITTLE_ENDIAN
219 #define BIG_ENDIAN __BIG_ENDIAN
222 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
223 #define MISALIGNED_OK 1
229 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
230 # error "Unknown or unsupported endianness (BYTE_ORDER)"
231 #elif (-6 & 5) || CHAR_BIT!=8 || UINT_MAX!=0xffffffff || MDB_SIZE_MAX%UINT_MAX
232 # error "Two's complement, reasonably sized integer types, please"
236 /** Put infrequently used env functions in separate section */
238 # define ESECT __attribute__ ((section("__TEXT,text_env")))
240 # define ESECT __attribute__ ((section("text_env")))
247 #define CALL_CONV WINAPI
252 /** @defgroup internal LMDB Internals
255 /** @defgroup compat Compatibility Macros
256 * A bunch of macros to minimize the amount of platform-specific ifdefs
257 * needed throughout the rest of the code. When the features this library
258 * needs are similar enough to POSIX to be hidden in a one-or-two line
259 * replacement, this macro approach is used.
263 /** Features under development */
268 /** Wrapper around __func__, which is a C99 feature */
269 #if __STDC_VERSION__ >= 199901L
270 # define mdb_func_ __func__
271 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
272 # define mdb_func_ __FUNCTION__
274 /* If a debug message says <mdb_unknown>(), update the #if statements above */
275 # define mdb_func_ "<mdb_unknown>"
278 /* Internal error codes, not exposed outside liblmdb */
279 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
281 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
282 #elif defined MDB_USE_SYSV_SEM
283 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
284 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
285 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
289 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
291 /** Some platforms define the EOWNERDEAD error code
292 * even though they don't support Robust Mutexes.
293 * Compile with -DMDB_USE_ROBUST=0, or use some other
294 * mechanism like -DMDB_USE_SYSV_SEM instead of
295 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
296 * also Robust, but some systems don't support them
299 #ifndef MDB_USE_ROBUST
300 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
301 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
302 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
303 # define MDB_USE_ROBUST 0
305 # define MDB_USE_ROBUST 1
307 #endif /* !MDB_USE_ROBUST */
309 #if defined(MDB_USE_POSIX_MUTEX) && (MDB_USE_ROBUST)
310 /* glibc < 2.12 only provided _np API */
311 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
312 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
313 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
314 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
315 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
317 #endif /* MDB_USE_POSIX_MUTEX && MDB_USE_ROBUST */
319 #if defined(MDB_OWNERDEAD) && (MDB_USE_ROBUST)
320 #define MDB_ROBUST_SUPPORTED 1
324 #define MDB_USE_HASH 1
325 #define MDB_PIDLOCK 0
326 #define THREAD_RET DWORD
327 #define pthread_t HANDLE
328 #define pthread_mutex_t HANDLE
329 #define pthread_cond_t HANDLE
330 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
331 #define pthread_key_t DWORD
332 #define pthread_self() GetCurrentThreadId()
333 #define pthread_key_create(x,y) \
334 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
335 #define pthread_key_delete(x) TlsFree(x)
336 #define pthread_getspecific(x) TlsGetValue(x)
337 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
338 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
339 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
340 #define pthread_cond_signal(x) SetEvent(*x)
341 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
342 #define THREAD_CREATE(thr,start,arg) \
343 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
344 #define THREAD_FINISH(thr) \
345 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
346 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
347 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
348 #define mdb_mutex_consistent(mutex) 0
349 #define getpid() GetCurrentProcessId()
350 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
351 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
352 #define ErrCode() GetLastError()
353 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
354 #define close(fd) (CloseHandle(fd) ? 0 : -1)
355 #define munmap(ptr,len) UnmapViewOfFile(ptr)
356 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
357 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
359 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
362 #define THREAD_RET void *
363 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
364 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
366 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
367 #define MDB_PIDLOCK 1
369 #ifdef MDB_USE_POSIX_SEM
371 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
372 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
373 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
376 mdb_sem_wait(sem_t *sem)
379 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
383 #elif defined MDB_USE_SYSV_SEM
385 typedef struct mdb_mutex {
389 } mdb_mutex_t[1], *mdb_mutexref_t;
391 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
392 #define UNLOCK_MUTEX(mutex) do { \
393 struct sembuf sb = { 0, 1, SEM_UNDO }; \
394 sb.sem_num = (mutex)->semnum; \
395 *(mutex)->locked = 0; \
396 semop((mutex)->semid, &sb, 1); \
400 mdb_sem_wait(mdb_mutexref_t sem)
402 int rc, *locked = sem->locked;
403 struct sembuf sb = { 0, -1, SEM_UNDO };
404 sb.sem_num = sem->semnum;
406 if (!semop(sem->semid, &sb, 1)) {
407 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
411 } while ((rc = errno) == EINTR);
415 #define mdb_mutex_consistent(mutex) 0
417 #else /* MDB_USE_POSIX_MUTEX: */
418 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
419 * local variables keep it (mdb_mutexref_t).
421 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
422 * be the same, or an array[size 1] and a pointer.
425 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
427 /** Lock the reader or writer mutex.
428 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
430 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
431 /** Unlock the reader or writer mutex.
433 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
434 /** Mark mutex-protected data as repaired, after death of previous owner.
436 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
437 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
439 /** Get the error code for the last failed system function.
441 #define ErrCode() errno
443 /** An abstraction for a file handle.
444 * On POSIX systems file handles are small integers. On Windows
445 * they're opaque pointers.
449 /** A value for an invalid file handle.
450 * Mainly used to initialize file variables and signify that they are
453 #define INVALID_HANDLE_VALUE (-1)
455 /** Get the size of a memory page for the system.
456 * This is the basic size that the platform's memory manager uses, and is
457 * fundamental to the use of memory-mapped files.
459 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
462 #define Z MDB_FMT_Z /**< printf/scanf format modifier for size_t */
463 #define Yu MDB_PRIy(u) /**< printf format for #mdb_size_t */
464 #define Yd MDB_PRIy(d) /**< printf format for "signed #mdb_size_t" */
466 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
468 #elif defined(MDB_USE_SYSV_SEM)
469 #define MNAME_LEN (sizeof(int))
471 #define MNAME_LEN (sizeof(pthread_mutex_t))
474 #ifdef MDB_USE_SYSV_SEM
475 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
477 #define SYSV_SEM_FLAG 0
482 #ifdef MDB_ROBUST_SUPPORTED
483 /** Lock mutex, handle any error, set rc = result.
484 * Return 0 on success, nonzero (not rc) on error.
486 #define LOCK_MUTEX(rc, env, mutex) \
487 (((rc) = LOCK_MUTEX0(mutex)) && \
488 ((rc) = mdb_mutex_failed(env, mutex, rc)))
489 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
491 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
492 #define mdb_mutex_failed(env, mutex, rc) (rc)
496 /** A flag for opening a file and requesting synchronous data writes.
497 * This is only used when writing a meta page. It's not strictly needed;
498 * we could just do a normal write and then immediately perform a flush.
499 * But if this flag is available it saves us an extra system call.
501 * @note If O_DSYNC is undefined but exists in /usr/include,
502 * preferably set some compiler flag to get the definition.
506 # define MDB_DSYNC O_DSYNC
508 # define MDB_DSYNC O_SYNC
513 /** Function for flushing the data of a file. Define this to fsync
514 * if fdatasync() is not supported.
516 #ifndef MDB_FDATASYNC
517 # define MDB_FDATASYNC fdatasync
521 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
532 /** A page number in the database.
533 * Note that 64 bit page numbers are overkill, since pages themselves
534 * already represent 12-13 bits of addressable memory, and the OS will
535 * always limit applications to a maximum of 63 bits of address space.
537 * @note In the #MDB_node structure, we only store 48 bits of this value,
538 * which thus limits us to only 60 bits of addressable data.
540 typedef MDB_ID pgno_t;
542 /** A transaction ID.
543 * See struct MDB_txn.mt_txnid for details.
545 typedef MDB_ID txnid_t;
547 /** @defgroup debug Debug Macros
551 /** Enable debug output. Needs variable argument macros (a C99 feature).
552 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
553 * read from and written to the database (used for free space management).
559 static int mdb_debug;
560 static txnid_t mdb_debug_start;
562 /** Print a debug message with printf formatting.
563 * Requires double parenthesis around 2 or more args.
565 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
566 # define DPRINTF0(fmt, ...) \
567 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
569 # define DPRINTF(args) ((void) 0)
571 /** Print a debug string.
572 * The string is printed literally, with no format processing.
574 #define DPUTS(arg) DPRINTF(("%s", arg))
575 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
577 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
580 /** @brief The maximum size of a database page.
582 * It is 32k or 64k, since value-PAGEBASE must fit in
583 * #MDB_page.%mp_upper.
585 * LMDB will use database pages < OS pages if needed.
586 * That causes more I/O in write transactions: The OS must
587 * know (read) the whole page before writing a partial page.
589 * Note that we don't currently support Huge pages. On Linux,
590 * regular data files cannot use Huge pages, and in general
591 * Huge pages aren't actually pageable. We rely on the OS
592 * demand-pager to read our data and page it out when memory
593 * pressure from other processes is high. So until OSs have
594 * actual paging support for Huge pages, they're not viable.
596 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
598 /** The minimum number of keys required in a database page.
599 * Setting this to a larger value will place a smaller bound on the
600 * maximum size of a data item. Data items larger than this size will
601 * be pushed into overflow pages instead of being stored directly in
602 * the B-tree node. This value used to default to 4. With a page size
603 * of 4096 bytes that meant that any item larger than 1024 bytes would
604 * go into an overflow page. That also meant that on average 2-3KB of
605 * each overflow page was wasted space. The value cannot be lower than
606 * 2 because then there would no longer be a tree structure. With this
607 * value, items larger than 2KB will go into overflow pages, and on
608 * average only 1KB will be wasted.
610 #define MDB_MINKEYS 2
612 /** A stamp that identifies a file as an LMDB file.
613 * There's nothing special about this value other than that it is easily
614 * recognizable, and it will reflect any byte order mismatches.
616 #define MDB_MAGIC 0xBEEFC0DE
618 /** The version number for a database's datafile format. */
619 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
620 /** The version number for a database's lockfile format. */
621 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
623 /** @brief The max size of a key we can write, or 0 for computed max.
625 * This macro should normally be left alone or set to 0.
626 * Note that a database with big keys or dupsort data cannot be
627 * reliably modified by a liblmdb which uses a smaller max.
628 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
630 * Other values are allowed, for backwards compat. However:
631 * A value bigger than the computed max can break if you do not
632 * know what you are doing, and liblmdb <= 0.9.10 can break when
633 * modifying a DB with keys/dupsort data bigger than its max.
635 * Data items in an #MDB_DUPSORT database are also limited to
636 * this size, since they're actually keys of a sub-DB. Keys and
637 * #MDB_DUPSORT data items must fit on a node in a regular page.
639 #ifndef MDB_MAXKEYSIZE
640 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
643 /** The maximum size of a key we can write to the environment. */
645 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
647 #define ENV_MAXKEY(env) ((env)->me_maxkey)
650 /** @brief The maximum size of a data item.
652 * We only store a 32 bit value for node sizes.
654 #define MAXDATASIZE 0xffffffffUL
657 /** Key size which fits in a #DKBUF.
660 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
663 * This is used for printing a hex dump of a key's contents.
665 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
666 /** Display a key in hex.
668 * Invoke a function to display a key in hex.
670 #define DKEY(x) mdb_dkey(x, kbuf)
676 /** An invalid page number.
677 * Mainly used to denote an empty tree.
679 #define P_INVALID (~(pgno_t)0)
681 /** Test if the flags \b f are set in a flag word \b w. */
682 #define F_ISSET(w, f) (((w) & (f)) == (f))
684 /** Round \b n up to an even number. */
685 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
687 /** Used for offsets within a single page.
688 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
691 typedef uint16_t indx_t;
693 /** Default size of memory map.
694 * This is certainly too small for any actual applications. Apps should always set
695 * the size explicitly using #mdb_env_set_mapsize().
697 #define DEFAULT_MAPSIZE 1048576
699 /** @defgroup readers Reader Lock Table
700 * Readers don't acquire any locks for their data access. Instead, they
701 * simply record their transaction ID in the reader table. The reader
702 * mutex is needed just to find an empty slot in the reader table. The
703 * slot's address is saved in thread-specific data so that subsequent read
704 * transactions started by the same thread need no further locking to proceed.
706 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
708 * No reader table is used if the database is on a read-only filesystem, or
709 * if #MDB_NOLOCK is set.
711 * Since the database uses multi-version concurrency control, readers don't
712 * actually need any locking. This table is used to keep track of which
713 * readers are using data from which old transactions, so that we'll know
714 * when a particular old transaction is no longer in use. Old transactions
715 * that have discarded any data pages can then have those pages reclaimed
716 * for use by a later write transaction.
718 * The lock table is constructed such that reader slots are aligned with the
719 * processor's cache line size. Any slot is only ever used by one thread.
720 * This alignment guarantees that there will be no contention or cache
721 * thrashing as threads update their own slot info, and also eliminates
722 * any need for locking when accessing a slot.
724 * A writer thread will scan every slot in the table to determine the oldest
725 * outstanding reader transaction. Any freed pages older than this will be
726 * reclaimed by the writer. The writer doesn't use any locks when scanning
727 * this table. This means that there's no guarantee that the writer will
728 * see the most up-to-date reader info, but that's not required for correct
729 * operation - all we need is to know the upper bound on the oldest reader,
730 * we don't care at all about the newest reader. So the only consequence of
731 * reading stale information here is that old pages might hang around a
732 * while longer before being reclaimed. That's actually good anyway, because
733 * the longer we delay reclaiming old pages, the more likely it is that a
734 * string of contiguous pages can be found after coalescing old pages from
735 * many old transactions together.
738 /** Number of slots in the reader table.
739 * This value was chosen somewhat arbitrarily. 126 readers plus a
740 * couple mutexes fit exactly into 8KB on my development machine.
741 * Applications should set the table size using #mdb_env_set_maxreaders().
743 #define DEFAULT_READERS 126
745 /** The size of a CPU cache line in bytes. We want our lock structures
746 * aligned to this size to avoid false cache line sharing in the
748 * This value works for most CPUs. For Itanium this should be 128.
754 /** The information we store in a single slot of the reader table.
755 * In addition to a transaction ID, we also record the process and
756 * thread ID that owns a slot, so that we can detect stale information,
757 * e.g. threads or processes that went away without cleaning up.
758 * @note We currently don't check for stale records. We simply re-init
759 * the table when we know that we're the only process opening the
762 typedef struct MDB_rxbody {
763 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
764 * Multiple readers that start at the same time will probably have the
765 * same ID here. Again, it's not important to exclude them from
766 * anything; all we need to know is which version of the DB they
767 * started from so we can avoid overwriting any data used in that
768 * particular version.
770 volatile txnid_t mrb_txnid;
771 /** The process ID of the process owning this reader txn. */
772 volatile MDB_PID_T mrb_pid;
773 /** The thread ID of the thread owning this txn. */
774 volatile MDB_THR_T mrb_tid;
777 /** The actual reader record, with cacheline padding. */
778 typedef struct MDB_reader {
781 /** shorthand for mrb_txnid */
782 #define mr_txnid mru.mrx.mrb_txnid
783 #define mr_pid mru.mrx.mrb_pid
784 #define mr_tid mru.mrx.mrb_tid
785 /** cache line alignment */
786 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
790 /** The header for the reader table.
791 * The table resides in a memory-mapped file. (This is a different file
792 * than is used for the main database.)
794 * For POSIX the actual mutexes reside in the shared memory of this
795 * mapped file. On Windows, mutexes are named objects allocated by the
796 * kernel; we store the mutex names in this mapped file so that other
797 * processes can grab them. This same approach is also used on
798 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
799 * process-shared POSIX mutexes. For these cases where a named object
800 * is used, the object name is derived from a 64 bit FNV hash of the
801 * environment pathname. As such, naming collisions are extremely
802 * unlikely. If a collision occurs, the results are unpredictable.
804 typedef struct MDB_txbody {
805 /** Stamp identifying this as an LMDB file. It must be set
808 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
810 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
811 char mtb_rmname[MNAME_LEN];
812 #elif defined(MDB_USE_SYSV_SEM)
816 /** Mutex protecting access to this table.
817 * This is the reader table lock used with LOCK_MUTEX().
819 mdb_mutex_t mtb_rmutex;
821 /** The ID of the last transaction committed to the database.
822 * This is recorded here only for convenience; the value can always
823 * be determined by reading the main database meta pages.
825 volatile txnid_t mtb_txnid;
826 /** The number of slots that have been used in the reader table.
827 * This always records the maximum count, it is not decremented
828 * when readers release their slots.
830 volatile unsigned mtb_numreaders;
833 /** The actual reader table definition. */
834 typedef struct MDB_txninfo {
837 #define mti_magic mt1.mtb.mtb_magic
838 #define mti_format mt1.mtb.mtb_format
839 #define mti_rmutex mt1.mtb.mtb_rmutex
840 #define mti_rmname mt1.mtb.mtb_rmname
841 #define mti_txnid mt1.mtb.mtb_txnid
842 #define mti_numreaders mt1.mtb.mtb_numreaders
843 #ifdef MDB_USE_SYSV_SEM
844 #define mti_semid mt1.mtb.mtb_semid
845 #define mti_rlocked mt1.mtb.mtb_rlocked
847 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
850 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
851 char mt2_wmname[MNAME_LEN];
852 #define mti_wmname mt2.mt2_wmname
853 #elif defined MDB_USE_SYSV_SEM
855 #define mti_wlocked mt2.mt2_wlocked
857 mdb_mutex_t mt2_wmutex;
858 #define mti_wmutex mt2.mt2_wmutex
860 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
862 MDB_reader mti_readers[1];
865 /** Lockfile format signature: version, features and field layout */
866 #define MDB_LOCK_FORMAT \
868 ((MDB_LOCK_VERSION) \
869 /* Flags which describe functionality */ \
870 + (SYSV_SEM_FLAG << 18) \
871 + (((MDB_PIDLOCK) != 0) << 16)))
874 /** Common header for all page types. The page type depends on #mp_flags.
876 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with
877 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages
878 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header.
880 * #P_OVERFLOW records occupy one or more contiguous pages where only the
881 * first has a page header. They hold the real data of #F_BIGDATA nodes.
883 * #P_SUBP sub-pages are small leaf "pages" with duplicate data.
884 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page.
885 * (Duplicate data can also go in sub-databases, which use normal pages.)
887 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot.
889 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once
890 * in the snapshot: Either used by a database or listed in a freeDB record.
892 typedef struct MDB_page {
893 #define mp_pgno mp_p.p_pgno
894 #define mp_next mp_p.p_next
896 pgno_t p_pgno; /**< page number */
897 struct MDB_page *p_next; /**< for in-memory list of freed pages */
899 uint16_t mp_pad; /**< key size if this is a LEAF2 page */
900 /** @defgroup mdb_page Page Flags
902 * Flags for the page headers.
905 #define P_BRANCH 0x01 /**< branch page */
906 #define P_LEAF 0x02 /**< leaf page */
907 #define P_OVERFLOW 0x04 /**< overflow page */
908 #define P_META 0x08 /**< meta page */
909 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
910 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
911 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
912 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
913 #define P_KEEP 0x8000 /**< leave this page alone during spill */
915 uint16_t mp_flags; /**< @ref mdb_page */
916 #define mp_lower mp_pb.pb.pb_lower
917 #define mp_upper mp_pb.pb.pb_upper
918 #define mp_pages mp_pb.pb_pages
921 indx_t pb_lower; /**< lower bound of free space */
922 indx_t pb_upper; /**< upper bound of free space */
924 uint32_t pb_pages; /**< number of overflow pages */
926 indx_t mp_ptrs[1]; /**< dynamic size */
929 /** Size of the page header, excluding dynamic data at the end */
930 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
932 /** Address of first usable data byte in a page, after the header */
933 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
935 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
936 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
938 /** Number of nodes on a page */
939 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
941 /** The amount of space remaining in the page */
942 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
944 /** The percentage of space used in the page, in tenths of a percent. */
945 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
946 ((env)->me_psize - PAGEHDRSZ))
947 /** The minimum page fill factor, in tenths of a percent.
948 * Pages emptier than this are candidates for merging.
950 #define FILL_THRESHOLD 250
952 /** Test if a page is a leaf page */
953 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
954 /** Test if a page is a LEAF2 page */
955 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
956 /** Test if a page is a branch page */
957 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
958 /** Test if a page is an overflow page */
959 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
960 /** Test if a page is a sub page */
961 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
963 /** The number of overflow pages needed to store the given size. */
964 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
966 /** Link in #MDB_txn.%mt_loose_pgs list.
967 * Kept outside the page header, which is needed when reusing the page.
969 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
971 /** Header for a single key/data pair within a page.
972 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
973 * We guarantee 2-byte alignment for 'MDB_node's.
975 typedef struct MDB_node {
976 /** lo and hi are used for data size on leaf nodes and for
977 * child pgno on branch nodes. On 64 bit platforms, flags
978 * is also used for pgno. (Branch nodes have no flags).
979 * They are in host byte order in case that lets some
980 * accesses be optimized into a 32-bit word access.
982 #if BYTE_ORDER == LITTLE_ENDIAN
983 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
985 unsigned short mn_hi, mn_lo;
987 /** @defgroup mdb_node Node Flags
989 * Flags for node headers.
992 #define F_BIGDATA 0x01 /**< data put on overflow page */
993 #define F_SUBDATA 0x02 /**< data is a sub-database */
994 #define F_DUPDATA 0x04 /**< data has duplicates */
996 /** valid flags for #mdb_node_add() */
997 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
1000 unsigned short mn_flags; /**< @ref mdb_node */
1001 unsigned short mn_ksize; /**< key size */
1002 char mn_data[1]; /**< key and data are appended here */
1005 /** Size of the node header, excluding dynamic data at the end */
1006 #define NODESIZE offsetof(MDB_node, mn_data)
1008 /** Bit position of top word in page number, for shifting mn_flags */
1009 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
1011 /** Size of a node in a branch page with a given key.
1012 * This is just the node header plus the key, there is no data.
1014 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
1016 /** Size of a node in a leaf page with a given key and data.
1017 * This is node header plus key plus data size.
1019 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
1021 /** Address of node \b i in page \b p */
1022 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
1024 /** Address of the key for the node */
1025 #define NODEKEY(node) (void *)((node)->mn_data)
1027 /** Address of the data for a node */
1028 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
1030 /** Get the page number pointed to by a branch node */
1031 #define NODEPGNO(node) \
1032 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
1033 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
1034 /** Set the page number in a branch node */
1035 #define SETPGNO(node,pgno) do { \
1036 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
1037 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
1039 /** Get the size of the data in a leaf node */
1040 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
1041 /** Set the size of the data for a leaf node */
1042 #define SETDSZ(node,size) do { \
1043 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
1044 /** The size of a key in a node */
1045 #define NODEKSZ(node) ((node)->mn_ksize)
1047 /** Copy a page number from src to dst */
1048 #ifdef MISALIGNED_OK
1049 #define COPY_PGNO(dst,src) dst = src
1051 #if MDB_SIZE_MAX > 0xffffffffU
1052 #define COPY_PGNO(dst,src) do { \
1053 unsigned short *s, *d; \
1054 s = (unsigned short *)&(src); \
1055 d = (unsigned short *)&(dst); \
1062 #define COPY_PGNO(dst,src) do { \
1063 unsigned short *s, *d; \
1064 s = (unsigned short *)&(src); \
1065 d = (unsigned short *)&(dst); \
1071 /** The address of a key in a LEAF2 page.
1072 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
1073 * There are no node headers, keys are stored contiguously.
1075 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
1077 /** Set the \b node's key into \b keyptr, if requested. */
1078 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
1079 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
1081 /** Set the \b node's key into \b key. */
1082 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
1084 /** Information about a single database in the environment. */
1085 typedef struct MDB_db {
1086 uint32_t md_pad; /**< also ksize for LEAF2 pages */
1087 uint16_t md_flags; /**< @ref mdb_dbi_open */
1088 uint16_t md_depth; /**< depth of this tree */
1089 pgno_t md_branch_pages; /**< number of internal pages */
1090 pgno_t md_leaf_pages; /**< number of leaf pages */
1091 pgno_t md_overflow_pages; /**< number of overflow pages */
1092 mdb_size_t md_entries; /**< number of data items */
1093 pgno_t md_root; /**< the root page of this tree */
1096 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1097 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1098 /** #mdb_dbi_open() flags */
1099 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1100 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1102 /** Handle for the DB used to track free pages. */
1104 /** Handle for the default DB. */
1106 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1109 /** Number of meta pages - also hardcoded elsewhere */
1112 /** Meta page content.
1113 * A meta page is the start point for accessing a database snapshot.
1114 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1116 typedef struct MDB_meta {
1117 /** Stamp identifying this as an LMDB file. It must be set
1120 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1121 uint32_t mm_version;
1123 union { /* always zero since we don't support fixed mapping in MDB_VL32 */
1127 #define mm_address mm_un.mmun_address
1129 void *mm_address; /**< address for fixed mapping */
1131 pgno_t mm_mapsize; /**< size of mmap region */
1132 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1133 /** The size of pages used in this DB */
1134 #define mm_psize mm_dbs[FREE_DBI].md_pad
1135 /** Any persistent environment flags. @ref mdb_env */
1136 #define mm_flags mm_dbs[FREE_DBI].md_flags
1137 /** Last used page in the datafile.
1138 * Actually the file may be shorter if the freeDB lists the final pages.
1141 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1144 /** Buffer for a stack-allocated meta page.
1145 * The members define size and alignment, and silence type
1146 * aliasing warnings. They are not used directly; that could
1147 * mean incorrectly using several union members in parallel.
1149 typedef union MDB_metabuf {
1152 char mm_pad[PAGEHDRSZ];
1157 /** Auxiliary DB info.
1158 * The information here is mostly static/read-only. There is
1159 * only a single copy of this record in the environment.
1161 typedef struct MDB_dbx {
1162 MDB_val md_name; /**< name of the database */
1163 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1164 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1165 MDB_rel_func *md_rel; /**< user relocate function */
1166 void *md_relctx; /**< user-provided context for md_rel */
1169 /** A database transaction.
1170 * Every operation requires a transaction handle.
1173 MDB_txn *mt_parent; /**< parent of a nested txn */
1174 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1176 pgno_t mt_next_pgno; /**< next unallocated page */
1178 pgno_t mt_last_pgno; /**< last written page */
1180 /** The ID of this transaction. IDs are integers incrementing from 1.
1181 * Only committed write transactions increment the ID. If a transaction
1182 * aborts, the ID may be re-used by the next writer.
1185 MDB_env *mt_env; /**< the DB environment */
1186 /** The list of pages that became unused during this transaction.
1188 MDB_IDL mt_free_pgs;
1189 /** The list of loose pages that became unused and may be reused
1190 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1192 MDB_page *mt_loose_pgs;
1193 /** Number of loose pages (#mt_loose_pgs) */
1195 /** The sorted list of dirty pages we temporarily wrote to disk
1196 * because the dirty list was full. page numbers in here are
1197 * shifted left by 1, deleted slots have the LSB set.
1199 MDB_IDL mt_spill_pgs;
1201 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1202 MDB_ID2L dirty_list;
1203 /** For read txns: This thread/txn's reader table slot, or NULL. */
1206 /** Array of records for each DB known in the environment. */
1208 /** Array of MDB_db records for each known DB */
1210 /** Array of sequence numbers for each DB handle */
1211 unsigned int *mt_dbiseqs;
1212 /** @defgroup mt_dbflag Transaction DB Flags
1216 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1217 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1218 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1219 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1220 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1222 /** In write txns, array of cursors for each DB */
1223 MDB_cursor **mt_cursors;
1224 /** Array of flags for each DB */
1225 unsigned char *mt_dbflags;
1227 /** List of read-only pages (actually chunks) */
1229 /** We map chunks of 16 pages. Even though Windows uses 4KB pages, all
1230 * mappings must begin on 64KB boundaries. So we round off all pgnos to
1231 * a chunk boundary. We do the same on Linux for symmetry, and also to
1232 * reduce the frequency of mmap/munmap calls.
1234 #define MDB_RPAGE_CHUNK 16
1235 #define MDB_TRPAGE_SIZE 4096 /**< size of #mt_rpages array of chunks */
1236 #define MDB_TRPAGE_MAX (MDB_TRPAGE_SIZE-1) /**< maximum chunk index */
1237 unsigned int mt_rpcheck; /**< threshold for reclaiming unref'd chunks */
1239 /** Number of DB records in use, or 0 when the txn is finished.
1240 * This number only ever increments until the txn finishes; we
1241 * don't decrement it when individual DB handles are closed.
1245 /** @defgroup mdb_txn Transaction Flags
1249 /** #mdb_txn_begin() flags */
1250 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1251 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1252 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1253 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1254 /* internal txn flags */
1255 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1256 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1257 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1258 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1259 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1260 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1261 /** most operations on the txn are currently illegal */
1262 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1264 unsigned int mt_flags; /**< @ref mdb_txn */
1265 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1266 * Includes ancestor txns' dirty pages not hidden by other txns'
1267 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1268 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1270 unsigned int mt_dirty_room;
1273 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1274 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1275 * raise this on a 64 bit machine.
1277 #define CURSOR_STACK 32
1281 /** Cursors are used for all DB operations.
1282 * A cursor holds a path of (page pointer, key index) from the DB
1283 * root to a position in the DB, plus other state. #MDB_DUPSORT
1284 * cursors include an xcursor to the current data item. Write txns
1285 * track their cursors and keep them up to date when data moves.
1286 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1287 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1290 /** Next cursor on this DB in this txn */
1291 MDB_cursor *mc_next;
1292 /** Backup of the original cursor if this cursor is a shadow */
1293 MDB_cursor *mc_backup;
1294 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1295 struct MDB_xcursor *mc_xcursor;
1296 /** The transaction that owns this cursor */
1298 /** The database handle this cursor operates on */
1300 /** The database record for this cursor */
1302 /** The database auxiliary record for this cursor */
1304 /** The @ref mt_dbflag for this database */
1305 unsigned char *mc_dbflag;
1306 unsigned short mc_snum; /**< number of pushed pages */
1307 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1308 /** @defgroup mdb_cursor Cursor Flags
1310 * Cursor state flags.
1313 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1314 #define C_EOF 0x02 /**< No more data */
1315 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1316 #define C_DEL 0x08 /**< last op was a cursor_del */
1317 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1318 #define C_WRITEMAP MDB_TXN_WRITEMAP /**< Copy of txn flag */
1319 /** Read-only cursor into the txn's original snapshot in the map.
1320 * Set for read-only txns, and in #mdb_page_alloc() for #FREE_DBI when
1321 * #MDB_DEVEL & 2. Only implements code which is necessary for this.
1323 #define C_ORIG_RDONLY MDB_TXN_RDONLY
1325 unsigned int mc_flags; /**< @ref mdb_cursor */
1326 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1327 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1329 MDB_page *mc_ovpg; /**< a referenced overflow page */
1330 # define MC_OVPG(mc) ((mc)->mc_ovpg)
1331 # define MC_SET_OVPG(mc, pg) ((mc)->mc_ovpg = (pg))
1333 # define MC_OVPG(mc) ((MDB_page *)0)
1334 # define MC_SET_OVPG(mc, pg) ((void)0)
1338 /** Context for sorted-dup records.
1339 * We could have gone to a fully recursive design, with arbitrarily
1340 * deep nesting of sub-databases. But for now we only handle these
1341 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1343 typedef struct MDB_xcursor {
1344 /** A sub-cursor for traversing the Dup DB */
1345 MDB_cursor mx_cursor;
1346 /** The database record for this Dup DB */
1348 /** The auxiliary DB record for this Dup DB */
1350 /** The @ref mt_dbflag for this Dup DB */
1351 unsigned char mx_dbflag;
1354 /** State of FreeDB old pages, stored in the MDB_env */
1355 typedef struct MDB_pgstate {
1356 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1357 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1360 /** The database environment. */
1362 HANDLE me_fd; /**< The main data file */
1363 HANDLE me_lfd; /**< The lock file */
1364 HANDLE me_mfd; /**< just for writing the meta pages */
1365 #if defined(MDB_VL32) && defined(_WIN32)
1366 HANDLE me_fmh; /**< File Mapping handle */
1368 /** Failed to update the meta page. Probably an I/O error. */
1369 #define MDB_FATAL_ERROR 0x80000000U
1370 /** Some fields are initialized. */
1371 #define MDB_ENV_ACTIVE 0x20000000U
1372 /** me_txkey is set */
1373 #define MDB_ENV_TXKEY 0x10000000U
1374 /** fdatasync is unreliable */
1375 #define MDB_FSYNCONLY 0x08000000U
1376 uint32_t me_flags; /**< @ref mdb_env */
1377 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1378 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1379 unsigned int me_maxreaders; /**< size of the reader table */
1380 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1381 volatile int me_close_readers;
1382 MDB_dbi me_numdbs; /**< number of DBs opened */
1383 MDB_dbi me_maxdbs; /**< size of the DB table */
1384 MDB_PID_T me_pid; /**< process ID of this env */
1385 char *me_path; /**< path to the DB files */
1386 char *me_map; /**< the memory map of the data file */
1387 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1388 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1389 void *me_pbuf; /**< scratch area for DUPSORT put() */
1390 MDB_txn *me_txn; /**< current write transaction */
1391 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1392 mdb_size_t me_mapsize; /**< size of the data memory map */
1393 off_t me_size; /**< current file size */
1394 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1395 MDB_dbx *me_dbxs; /**< array of static DB info */
1396 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1397 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1398 pthread_key_t me_txkey; /**< thread-key for readers */
1399 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1400 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1401 # define me_pglast me_pgstate.mf_pglast
1402 # define me_pghead me_pgstate.mf_pghead
1403 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1404 /** IDL of pages that became unused in a write txn */
1405 MDB_IDL me_free_pgs;
1406 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1407 MDB_ID2L me_dirty_list;
1408 /** Max number of freelist items that can fit in a single overflow page */
1410 /** Max size of a node on a page */
1411 unsigned int me_nodemax;
1412 #if !(MDB_MAXKEYSIZE)
1413 unsigned int me_maxkey; /**< max size of a key */
1415 int me_live_reader; /**< have liveness lock in reader table */
1417 int me_pidquery; /**< Used in OpenProcess */
1419 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1420 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1421 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1423 mdb_mutex_t me_rmutex;
1424 mdb_mutex_t me_wmutex;
1427 MDB_ID3L me_rpages; /**< like #mt_rpages, but global to env */
1428 pthread_mutex_t me_rpmutex; /**< control access to #me_rpages */
1429 #define MDB_ERPAGE_SIZE 16384
1430 #define MDB_ERPAGE_MAX (MDB_ERPAGE_SIZE-1)
1431 unsigned int me_rpcheck;
1433 void *me_userctx; /**< User-settable context */
1434 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1437 /** Nested transaction */
1438 typedef struct MDB_ntxn {
1439 MDB_txn mnt_txn; /**< the transaction */
1440 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1443 /** max number of pages to commit in one writev() call */
1444 #define MDB_COMMIT_PAGES 64
1445 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1446 #undef MDB_COMMIT_PAGES
1447 #define MDB_COMMIT_PAGES IOV_MAX
1450 /** max bytes to write in one call */
1451 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1453 /** Check \b txn and \b dbi arguments to a function */
1454 #define TXN_DBI_EXIST(txn, dbi, validity) \
1455 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1457 /** Check for misused \b dbi handles */
1458 #define TXN_DBI_CHANGED(txn, dbi) \
1459 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1461 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1462 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1463 static int mdb_page_touch(MDB_cursor *mc);
1465 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1466 "reset-tmp", "fail-begin", "fail-beginchild"}
1468 /* mdb_txn_end operation number, for logging */
1469 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1470 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1472 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1473 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1474 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1475 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1476 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1478 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1479 static int mdb_page_search_root(MDB_cursor *mc,
1480 MDB_val *key, int modify);
1481 #define MDB_PS_MODIFY 1
1482 #define MDB_PS_ROOTONLY 2
1483 #define MDB_PS_FIRST 4
1484 #define MDB_PS_LAST 8
1485 static int mdb_page_search(MDB_cursor *mc,
1486 MDB_val *key, int flags);
1487 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1489 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1490 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1491 pgno_t newpgno, unsigned int nflags);
1493 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1494 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1495 static int mdb_env_write_meta(MDB_txn *txn);
1496 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1497 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1499 static void mdb_env_close0(MDB_env *env, int excl);
1501 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1502 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1503 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1504 static void mdb_node_del(MDB_cursor *mc, int ksize);
1505 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1506 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1507 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1508 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1509 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1511 static int mdb_rebalance(MDB_cursor *mc);
1512 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1514 static void mdb_cursor_pop(MDB_cursor *mc);
1515 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1517 static int mdb_cursor_del0(MDB_cursor *mc);
1518 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1519 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1520 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1521 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1522 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1524 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1525 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1527 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1528 static void mdb_xcursor_init0(MDB_cursor *mc);
1529 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1530 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1532 static int mdb_drop0(MDB_cursor *mc, int subs);
1533 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1534 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1537 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1540 /** Compare two items pointing at '#mdb_size_t's of unknown alignment. */
1541 #ifdef MISALIGNED_OK
1542 # define mdb_cmp_clong mdb_cmp_long
1544 # define mdb_cmp_clong mdb_cmp_cint
1547 /** True if we need #mdb_cmp_clong() instead of \b cmp for #MDB_INTEGERDUP */
1548 #define NEED_CMP_CLONG(cmp, ksize) \
1549 (UINT_MAX < MDB_SIZE_MAX && \
1550 (cmp) == mdb_cmp_int && (ksize) == sizeof(mdb_size_t))
1553 static SECURITY_DESCRIPTOR mdb_null_sd;
1554 static SECURITY_ATTRIBUTES mdb_all_sa;
1555 static int mdb_sec_inited;
1557 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize);
1560 /** Return the library version info. */
1562 mdb_version(int *major, int *minor, int *patch)
1564 if (major) *major = MDB_VERSION_MAJOR;
1565 if (minor) *minor = MDB_VERSION_MINOR;
1566 if (patch) *patch = MDB_VERSION_PATCH;
1567 return MDB_VERSION_STRING;
1570 /** Table of descriptions for LMDB @ref errors */
1571 static char *const mdb_errstr[] = {
1572 "MDB_KEYEXIST: Key/data pair already exists",
1573 "MDB_NOTFOUND: No matching key/data pair found",
1574 "MDB_PAGE_NOTFOUND: Requested page not found",
1575 "MDB_CORRUPTED: Located page was wrong type",
1576 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1577 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1578 "MDB_INVALID: File is not an LMDB file",
1579 "MDB_MAP_FULL: Environment mapsize limit reached",
1580 "MDB_DBS_FULL: Environment maxdbs limit reached",
1581 "MDB_READERS_FULL: Environment maxreaders limit reached",
1582 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1583 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1584 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1585 "MDB_PAGE_FULL: Internal error - page has no more space",
1586 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1587 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1588 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1589 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1590 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1591 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1592 "MDB_PROBLEM: Unexpected problem - txn should abort",
1596 mdb_strerror(int err)
1599 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1600 * This works as long as no function between the call to mdb_strerror
1601 * and the actual use of the message uses more than 4K of stack.
1603 #define MSGSIZE 1024
1604 #define PADSIZE 4096
1605 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1609 return ("Successful return: 0");
1611 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1612 i = err - MDB_KEYEXIST;
1613 return mdb_errstr[i];
1617 /* These are the C-runtime error codes we use. The comment indicates
1618 * their numeric value, and the Win32 error they would correspond to
1619 * if the error actually came from a Win32 API. A major mess, we should
1620 * have used LMDB-specific error codes for everything.
1623 case ENOENT: /* 2, FILE_NOT_FOUND */
1624 case EIO: /* 5, ACCESS_DENIED */
1625 case ENOMEM: /* 12, INVALID_ACCESS */
1626 case EACCES: /* 13, INVALID_DATA */
1627 case EBUSY: /* 16, CURRENT_DIRECTORY */
1628 case EINVAL: /* 22, BAD_COMMAND */
1629 case ENOSPC: /* 28, OUT_OF_PAPER */
1630 return strerror(err);
1635 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1636 FORMAT_MESSAGE_IGNORE_INSERTS,
1637 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1640 return strerror(err);
1644 /** assert(3) variant in cursor context */
1645 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1646 /** assert(3) variant in transaction context */
1647 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1648 /** assert(3) variant in environment context */
1649 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1652 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1653 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1656 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1657 const char *func, const char *file, int line)
1660 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1661 file, line, expr_txt, func);
1662 if (env->me_assert_func)
1663 env->me_assert_func(env, buf);
1664 fprintf(stderr, "%s\n", buf);
1668 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1672 /** Return the page number of \b mp which may be sub-page, for debug output */
1674 mdb_dbg_pgno(MDB_page *mp)
1677 COPY_PGNO(ret, mp->mp_pgno);
1681 /** Display a key in hexadecimal and return the address of the result.
1682 * @param[in] key the key to display
1683 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1684 * @return The key in hexadecimal form.
1687 mdb_dkey(MDB_val *key, char *buf)
1690 unsigned char *c = key->mv_data;
1696 if (key->mv_size > DKBUF_MAXKEYSIZE)
1697 return "MDB_MAXKEYSIZE";
1698 /* may want to make this a dynamic check: if the key is mostly
1699 * printable characters, print it as-is instead of converting to hex.
1703 for (i=0; i<key->mv_size; i++)
1704 ptr += sprintf(ptr, "%02x", *c++);
1706 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1712 mdb_leafnode_type(MDB_node *n)
1714 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1715 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1716 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1719 /** Display all the keys in the page. */
1721 mdb_page_list(MDB_page *mp)
1723 pgno_t pgno = mdb_dbg_pgno(mp);
1724 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1726 unsigned int i, nkeys, nsize, total = 0;
1730 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1731 case P_BRANCH: type = "Branch page"; break;
1732 case P_LEAF: type = "Leaf page"; break;
1733 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1734 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1735 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1737 fprintf(stderr, "Overflow page %"Yu" pages %u%s\n",
1738 pgno, mp->mp_pages, state);
1741 fprintf(stderr, "Meta-page %"Yu" txnid %"Yu"\n",
1742 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1745 fprintf(stderr, "Bad page %"Yu" flags 0x%X\n", pgno, mp->mp_flags);
1749 nkeys = NUMKEYS(mp);
1750 fprintf(stderr, "%s %"Yu" numkeys %d%s\n", type, pgno, nkeys, state);
1752 for (i=0; i<nkeys; i++) {
1753 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1754 key.mv_size = nsize = mp->mp_pad;
1755 key.mv_data = LEAF2KEY(mp, i, nsize);
1757 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1760 node = NODEPTR(mp, i);
1761 key.mv_size = node->mn_ksize;
1762 key.mv_data = node->mn_data;
1763 nsize = NODESIZE + key.mv_size;
1764 if (IS_BRANCH(mp)) {
1765 fprintf(stderr, "key %d: page %"Yu", %s\n", i, NODEPGNO(node),
1769 if (F_ISSET(node->mn_flags, F_BIGDATA))
1770 nsize += sizeof(pgno_t);
1772 nsize += NODEDSZ(node);
1774 nsize += sizeof(indx_t);
1775 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1776 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1778 total = EVEN(total);
1780 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1781 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1785 mdb_cursor_chk(MDB_cursor *mc)
1791 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1792 for (i=0; i<mc->mc_top; i++) {
1794 node = NODEPTR(mp, mc->mc_ki[i]);
1795 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1798 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1800 if (mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
1801 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1802 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1803 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1811 /** Count all the pages in each DB and in the freelist
1812 * and make sure it matches the actual number of pages
1814 * All named DBs must be open for a correct count.
1816 static void mdb_audit(MDB_txn *txn)
1820 MDB_ID freecount, count;
1825 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1826 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1827 freecount += *(MDB_ID *)data.mv_data;
1828 mdb_tassert(txn, rc == MDB_NOTFOUND);
1831 for (i = 0; i<txn->mt_numdbs; i++) {
1833 if (!(txn->mt_dbflags[i] & DB_VALID))
1835 mdb_cursor_init(&mc, txn, i, &mx);
1836 if (txn->mt_dbs[i].md_root == P_INVALID)
1838 count += txn->mt_dbs[i].md_branch_pages +
1839 txn->mt_dbs[i].md_leaf_pages +
1840 txn->mt_dbs[i].md_overflow_pages;
1841 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1842 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1843 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1846 mp = mc.mc_pg[mc.mc_top];
1847 for (j=0; j<NUMKEYS(mp); j++) {
1848 MDB_node *leaf = NODEPTR(mp, j);
1849 if (leaf->mn_flags & F_SUBDATA) {
1851 memcpy(&db, NODEDATA(leaf), sizeof(db));
1852 count += db.md_branch_pages + db.md_leaf_pages +
1853 db.md_overflow_pages;
1857 mdb_tassert(txn, rc == MDB_NOTFOUND);
1860 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1861 fprintf(stderr, "audit: %"Yu" freecount: %"Yu" count: %"Yu" total: %"Yu" next_pgno: %"Yu"\n",
1862 txn->mt_txnid, freecount, count+NUM_METAS,
1863 freecount+count+NUM_METAS, txn->mt_next_pgno);
1869 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1871 return txn->mt_dbxs[dbi].md_cmp(a, b);
1875 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1877 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1878 if (NEED_CMP_CLONG(dcmp, a->mv_size))
1879 dcmp = mdb_cmp_clong;
1883 /** Allocate memory for a page.
1884 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1887 mdb_page_malloc(MDB_txn *txn, unsigned num)
1889 MDB_env *env = txn->mt_env;
1890 MDB_page *ret = env->me_dpages;
1891 size_t psize = env->me_psize, sz = psize, off;
1892 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1893 * For a single page alloc, we init everything after the page header.
1894 * For multi-page, we init the final page; if the caller needed that
1895 * many pages they will be filling in at least up to the last page.
1899 VGMEMP_ALLOC(env, ret, sz);
1900 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1901 env->me_dpages = ret->mp_next;
1904 psize -= off = PAGEHDRSZ;
1909 if ((ret = malloc(sz)) != NULL) {
1910 VGMEMP_ALLOC(env, ret, sz);
1911 if (!(env->me_flags & MDB_NOMEMINIT)) {
1912 memset((char *)ret + off, 0, psize);
1916 txn->mt_flags |= MDB_TXN_ERROR;
1920 /** Free a single page.
1921 * Saves single pages to a list, for future reuse.
1922 * (This is not used for multi-page overflow pages.)
1925 mdb_page_free(MDB_env *env, MDB_page *mp)
1927 mp->mp_next = env->me_dpages;
1928 VGMEMP_FREE(env, mp);
1929 env->me_dpages = mp;
1932 /** Free a dirty page */
1934 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1936 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1937 mdb_page_free(env, dp);
1939 /* large pages just get freed directly */
1940 VGMEMP_FREE(env, dp);
1945 /** Return all dirty pages to dpage list */
1947 mdb_dlist_free(MDB_txn *txn)
1949 MDB_env *env = txn->mt_env;
1950 MDB_ID2L dl = txn->mt_u.dirty_list;
1951 unsigned i, n = dl[0].mid;
1953 for (i = 1; i <= n; i++) {
1954 mdb_dpage_free(env, dl[i].mptr);
1961 mdb_page_unref(MDB_txn *txn, MDB_page *mp)
1964 MDB_ID3L tl = txn->mt_rpages;
1966 if (mp->mp_flags & (P_SUBP|P_DIRTY))
1968 rem = mp->mp_pgno & (MDB_RPAGE_CHUNK-1);
1969 pgno = mp->mp_pgno ^ rem;
1970 x = mdb_mid3l_search(tl, pgno);
1971 if (x != tl[0].mid && tl[x+1].mid == mp->mp_pgno)
1976 #define MDB_PAGE_UNREF(txn, mp) mdb_page_unref(txn, mp)
1979 mdb_cursor_unref(MDB_cursor *mc)
1982 if (!mc->mc_snum || !mc->mc_pg[0] || IS_SUBP(mc->mc_pg[0]))
1984 for (i=0; i<mc->mc_snum; i++)
1985 mdb_page_unref(mc->mc_txn, mc->mc_pg[i]);
1987 mdb_page_unref(mc->mc_txn, mc->mc_ovpg);
1990 mc->mc_snum = mc->mc_top = 0;
1991 mc->mc_pg[0] = NULL;
1992 mc->mc_flags &= ~C_INITIALIZED;
1994 #define MDB_CURSOR_UNREF(mc, force) \
1995 (((force) || ((mc)->mc_flags & C_INITIALIZED)) \
1996 ? mdb_cursor_unref(mc) \
2000 #define MDB_PAGE_UNREF(txn, mp)
2001 #define MDB_CURSOR_UNREF(mc, force) ((void)0)
2002 #endif /* MDB_VL32 */
2004 /** Loosen or free a single page.
2005 * Saves single pages to a list for future reuse
2006 * in this same txn. It has been pulled from the freeDB
2007 * and already resides on the dirty list, but has been
2008 * deleted. Use these pages first before pulling again
2011 * If the page wasn't dirtied in this txn, just add it
2012 * to this txn's free list.
2015 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
2018 pgno_t pgno = mp->mp_pgno;
2019 MDB_txn *txn = mc->mc_txn;
2021 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
2022 if (txn->mt_parent) {
2023 MDB_ID2 *dl = txn->mt_u.dirty_list;
2024 /* If txn has a parent, make sure the page is in our
2028 unsigned x = mdb_mid2l_search(dl, pgno);
2029 if (x <= dl[0].mid && dl[x].mid == pgno) {
2030 if (mp != dl[x].mptr) { /* bad cursor? */
2031 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2032 txn->mt_flags |= MDB_TXN_ERROR;
2040 /* no parent txn, so it's just ours */
2045 DPRINTF(("loosen db %d page %"Yu, DDBI(mc), mp->mp_pgno));
2046 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
2047 txn->mt_loose_pgs = mp;
2048 txn->mt_loose_count++;
2049 mp->mp_flags |= P_LOOSE;
2051 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
2059 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
2060 * @param[in] mc A cursor handle for the current operation.
2061 * @param[in] pflags Flags of the pages to update:
2062 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
2063 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
2064 * @return 0 on success, non-zero on failure.
2067 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
2069 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
2070 MDB_txn *txn = mc->mc_txn;
2071 MDB_cursor *m3, *m0 = mc;
2076 int rc = MDB_SUCCESS, level;
2078 /* Mark pages seen by cursors */
2079 if (mc->mc_flags & C_UNTRACK)
2080 mc = NULL; /* will find mc in mt_cursors */
2081 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
2082 for (; mc; mc=mc->mc_next) {
2083 if (!(mc->mc_flags & C_INITIALIZED))
2085 for (m3 = mc;; m3 = &mx->mx_cursor) {
2087 for (j=0; j<m3->mc_snum; j++) {
2089 if ((mp->mp_flags & Mask) == pflags)
2090 mp->mp_flags ^= P_KEEP;
2092 mx = m3->mc_xcursor;
2093 /* Proceed to mx if it is at a sub-database */
2094 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
2096 if (! (mp && (mp->mp_flags & P_LEAF)))
2098 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
2099 if (!(leaf->mn_flags & F_SUBDATA))
2108 /* Mark dirty root pages */
2109 for (i=0; i<txn->mt_numdbs; i++) {
2110 if (txn->mt_dbflags[i] & DB_DIRTY) {
2111 pgno_t pgno = txn->mt_dbs[i].md_root;
2112 if (pgno == P_INVALID)
2114 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
2116 if ((dp->mp_flags & Mask) == pflags && level <= 1)
2117 dp->mp_flags ^= P_KEEP;
2125 static int mdb_page_flush(MDB_txn *txn, int keep);
2127 /** Spill pages from the dirty list back to disk.
2128 * This is intended to prevent running into #MDB_TXN_FULL situations,
2129 * but note that they may still occur in a few cases:
2130 * 1) our estimate of the txn size could be too small. Currently this
2131 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
2132 * 2) child txns may run out of space if their parents dirtied a
2133 * lot of pages and never spilled them. TODO: we probably should do
2134 * a preemptive spill during #mdb_txn_begin() of a child txn, if
2135 * the parent's dirty_room is below a given threshold.
2137 * Otherwise, if not using nested txns, it is expected that apps will
2138 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
2139 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
2140 * If the txn never references them again, they can be left alone.
2141 * If the txn only reads them, they can be used without any fuss.
2142 * If the txn writes them again, they can be dirtied immediately without
2143 * going thru all of the work of #mdb_page_touch(). Such references are
2144 * handled by #mdb_page_unspill().
2146 * Also note, we never spill DB root pages, nor pages of active cursors,
2147 * because we'll need these back again soon anyway. And in nested txns,
2148 * we can't spill a page in a child txn if it was already spilled in a
2149 * parent txn. That would alter the parent txns' data even though
2150 * the child hasn't committed yet, and we'd have no way to undo it if
2151 * the child aborted.
2153 * @param[in] m0 cursor A cursor handle identifying the transaction and
2154 * database for which we are checking space.
2155 * @param[in] key For a put operation, the key being stored.
2156 * @param[in] data For a put operation, the data being stored.
2157 * @return 0 on success, non-zero on failure.
2160 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
2162 MDB_txn *txn = m0->mc_txn;
2164 MDB_ID2L dl = txn->mt_u.dirty_list;
2165 unsigned int i, j, need;
2168 if (m0->mc_flags & C_SUB)
2171 /* Estimate how much space this op will take */
2172 i = m0->mc_db->md_depth;
2173 /* Named DBs also dirty the main DB */
2174 if (m0->mc_dbi >= CORE_DBS)
2175 i += txn->mt_dbs[MAIN_DBI].md_depth;
2176 /* For puts, roughly factor in the key+data size */
2178 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2179 i += i; /* double it for good measure */
2182 if (txn->mt_dirty_room > i)
2185 if (!txn->mt_spill_pgs) {
2186 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2187 if (!txn->mt_spill_pgs)
2190 /* purge deleted slots */
2191 MDB_IDL sl = txn->mt_spill_pgs;
2192 unsigned int num = sl[0];
2194 for (i=1; i<=num; i++) {
2201 /* Preserve pages which may soon be dirtied again */
2202 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2205 /* Less aggressive spill - we originally spilled the entire dirty list,
2206 * with a few exceptions for cursor pages and DB root pages. But this
2207 * turns out to be a lot of wasted effort because in a large txn many
2208 * of those pages will need to be used again. So now we spill only 1/8th
2209 * of the dirty pages. Testing revealed this to be a good tradeoff,
2210 * better than 1/2, 1/4, or 1/10.
2212 if (need < MDB_IDL_UM_MAX / 8)
2213 need = MDB_IDL_UM_MAX / 8;
2215 /* Save the page IDs of all the pages we're flushing */
2216 /* flush from the tail forward, this saves a lot of shifting later on. */
2217 for (i=dl[0].mid; i && need; i--) {
2218 MDB_ID pn = dl[i].mid << 1;
2220 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2222 /* Can't spill twice, make sure it's not already in a parent's
2225 if (txn->mt_parent) {
2227 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2228 if (tx2->mt_spill_pgs) {
2229 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2230 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2231 dp->mp_flags |= P_KEEP;
2239 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2243 mdb_midl_sort(txn->mt_spill_pgs);
2245 /* Flush the spilled part of dirty list */
2246 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2249 /* Reset any dirty pages we kept that page_flush didn't see */
2250 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2253 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2257 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2259 mdb_find_oldest(MDB_txn *txn)
2262 txnid_t mr, oldest = txn->mt_txnid - 1;
2263 if (txn->mt_env->me_txns) {
2264 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2265 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2276 /** Add a page to the txn's dirty list */
2278 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2281 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2283 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2284 insert = mdb_mid2l_append;
2286 insert = mdb_mid2l_insert;
2288 mid.mid = mp->mp_pgno;
2290 rc = insert(txn->mt_u.dirty_list, &mid);
2291 mdb_tassert(txn, rc == 0);
2292 txn->mt_dirty_room--;
2295 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2296 * me_pghead and mt_next_pgno.
2298 * If there are free pages available from older transactions, they
2299 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2300 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2301 * and move me_pglast to say which records were consumed. Only this
2302 * function can create me_pghead and move me_pglast/mt_next_pgno.
2303 * When #MDB_DEVEL & 2, it is not affected by #mdb_freelist_save(): it
2304 * then uses the transaction's original snapshot of the freeDB.
2305 * @param[in] mc cursor A cursor handle identifying the transaction and
2306 * database for which we are allocating.
2307 * @param[in] num the number of pages to allocate.
2308 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2309 * will always be satisfied by a single contiguous chunk of memory.
2310 * @return 0 on success, non-zero on failure.
2313 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2315 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2316 /* Get at most <Max_retries> more freeDB records once me_pghead
2317 * has enough pages. If not enough, use new pages from the map.
2318 * If <Paranoid> and mc is updating the freeDB, only get new
2319 * records if me_pghead is empty. Then the freelist cannot play
2320 * catch-up with itself by growing while trying to save it.
2322 enum { Paranoid = 1, Max_retries = 500 };
2324 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2326 int rc, retry = num * 60;
2327 MDB_txn *txn = mc->mc_txn;
2328 MDB_env *env = txn->mt_env;
2329 pgno_t pgno, *mop = env->me_pghead;
2330 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2332 txnid_t oldest = 0, last;
2337 /* If there are any loose pages, just use them */
2338 if (num == 1 && txn->mt_loose_pgs) {
2339 np = txn->mt_loose_pgs;
2340 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2341 txn->mt_loose_count--;
2342 DPRINTF(("db %d use loose page %"Yu, DDBI(mc), np->mp_pgno));
2349 /* If our dirty list is already full, we can't do anything */
2350 if (txn->mt_dirty_room == 0) {
2355 for (op = MDB_FIRST;; op = MDB_NEXT) {
2360 /* Seek a big enough contiguous page range. Prefer
2361 * pages at the tail, just truncating the list.
2367 if (mop[i-n2] == pgno+n2)
2374 if (op == MDB_FIRST) { /* 1st iteration */
2375 /* Prepare to fetch more and coalesce */
2376 last = env->me_pglast;
2377 oldest = env->me_pgoldest;
2378 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2379 #if (MDB_DEVEL) & 2 /* "& 2" so MDB_DEVEL=1 won't hide bugs breaking freeDB */
2380 /* Use original snapshot. TODO: Should need less care in code
2381 * which modifies the database. Maybe we can delete some code?
2383 m2.mc_flags |= C_ORIG_RDONLY;
2384 m2.mc_db = &env->me_metas[(txn->mt_txnid-1) & 1]->mm_dbs[FREE_DBI];
2385 m2.mc_dbflag = (unsigned char *)""; /* probably unnecessary */
2389 key.mv_data = &last; /* will look up last+1 */
2390 key.mv_size = sizeof(last);
2392 if (Paranoid && mc->mc_dbi == FREE_DBI)
2395 if (Paranoid && retry < 0 && mop_len)
2399 /* Do not fetch more if the record will be too recent */
2400 if (oldest <= last) {
2402 oldest = mdb_find_oldest(txn);
2403 env->me_pgoldest = oldest;
2409 rc = mdb_cursor_get(&m2, &key, NULL, op);
2411 if (rc == MDB_NOTFOUND)
2415 last = *(txnid_t*)key.mv_data;
2416 if (oldest <= last) {
2418 oldest = mdb_find_oldest(txn);
2419 env->me_pgoldest = oldest;
2425 np = m2.mc_pg[m2.mc_top];
2426 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2427 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2430 idl = (MDB_ID *) data.mv_data;
2433 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2438 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2440 mop = env->me_pghead;
2442 env->me_pglast = last;
2444 DPRINTF(("IDL read txn %"Yu" root %"Yu" num %u",
2445 last, txn->mt_dbs[FREE_DBI].md_root, i));
2447 DPRINTF(("IDL %"Yu, idl[j]));
2449 /* Merge in descending sorted order */
2450 mdb_midl_xmerge(mop, idl);
2454 /* Use new pages from the map when nothing suitable in the freeDB */
2456 pgno = txn->mt_next_pgno;
2457 if (pgno + num >= env->me_maxpg) {
2458 DPUTS("DB size maxed out");
2462 #if defined(_WIN32) && !defined(MDB_VL32)
2463 if (!(env->me_flags & MDB_RDONLY)) {
2465 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
2466 p = VirtualAlloc(p, env->me_psize * num, MEM_COMMIT,
2467 (env->me_flags & MDB_WRITEMAP) ? PAGE_READWRITE:
2470 DPUTS("VirtualAlloc failed");
2478 if (env->me_flags & MDB_WRITEMAP) {
2479 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2481 if (!(np = mdb_page_malloc(txn, num))) {
2487 mop[0] = mop_len -= num;
2488 /* Move any stragglers down */
2489 for (j = i-num; j < mop_len; )
2490 mop[++j] = mop[++i];
2492 txn->mt_next_pgno = pgno + num;
2495 mdb_page_dirty(txn, np);
2501 txn->mt_flags |= MDB_TXN_ERROR;
2505 /** Copy the used portions of a non-overflow page.
2506 * @param[in] dst page to copy into
2507 * @param[in] src page to copy from
2508 * @param[in] psize size of a page
2511 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2513 enum { Align = sizeof(pgno_t) };
2514 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2516 /* If page isn't full, just copy the used portion. Adjust
2517 * alignment so memcpy may copy words instead of bytes.
2519 if ((unused &= -Align) && !IS_LEAF2(src)) {
2520 upper = (upper + PAGEBASE) & -Align;
2521 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2522 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2525 memcpy(dst, src, psize - unused);
2529 /** Pull a page off the txn's spill list, if present.
2530 * If a page being referenced was spilled to disk in this txn, bring
2531 * it back and make it dirty/writable again.
2532 * @param[in] txn the transaction handle.
2533 * @param[in] mp the page being referenced. It must not be dirty.
2534 * @param[out] ret the writable page, if any. ret is unchanged if
2535 * mp wasn't spilled.
2538 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2540 MDB_env *env = txn->mt_env;
2543 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2545 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2546 if (!tx2->mt_spill_pgs)
2548 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2549 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2552 if (txn->mt_dirty_room == 0)
2553 return MDB_TXN_FULL;
2554 if (IS_OVERFLOW(mp))
2558 if (env->me_flags & MDB_WRITEMAP) {
2561 np = mdb_page_malloc(txn, num);
2565 memcpy(np, mp, num * env->me_psize);
2567 mdb_page_copy(np, mp, env->me_psize);
2570 /* If in current txn, this page is no longer spilled.
2571 * If it happens to be the last page, truncate the spill list.
2572 * Otherwise mark it as deleted by setting the LSB.
2574 if (x == txn->mt_spill_pgs[0])
2575 txn->mt_spill_pgs[0]--;
2577 txn->mt_spill_pgs[x] |= 1;
2578 } /* otherwise, if belonging to a parent txn, the
2579 * page remains spilled until child commits
2582 mdb_page_dirty(txn, np);
2583 np->mp_flags |= P_DIRTY;
2591 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2592 * @param[in] mc cursor pointing to the page to be touched
2593 * @return 0 on success, non-zero on failure.
2596 mdb_page_touch(MDB_cursor *mc)
2598 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2599 MDB_txn *txn = mc->mc_txn;
2600 MDB_cursor *m2, *m3;
2604 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2605 if (txn->mt_flags & MDB_TXN_SPILLS) {
2607 rc = mdb_page_unspill(txn, mp, &np);
2613 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2614 (rc = mdb_page_alloc(mc, 1, &np)))
2617 DPRINTF(("touched db %d page %"Yu" -> %"Yu, DDBI(mc),
2618 mp->mp_pgno, pgno));
2619 mdb_cassert(mc, mp->mp_pgno != pgno);
2620 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2621 /* Update the parent page, if any, to point to the new page */
2623 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2624 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2625 SETPGNO(node, pgno);
2627 mc->mc_db->md_root = pgno;
2629 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2630 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2632 /* If txn has a parent, make sure the page is in our
2636 unsigned x = mdb_mid2l_search(dl, pgno);
2637 if (x <= dl[0].mid && dl[x].mid == pgno) {
2638 if (mp != dl[x].mptr) { /* bad cursor? */
2639 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2640 txn->mt_flags |= MDB_TXN_ERROR;
2646 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2648 np = mdb_page_malloc(txn, 1);
2653 rc = mdb_mid2l_insert(dl, &mid);
2654 mdb_cassert(mc, rc == 0);
2659 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2661 np->mp_flags |= P_DIRTY;
2664 /* Adjust cursors pointing to mp */
2665 mc->mc_pg[mc->mc_top] = np;
2666 m2 = txn->mt_cursors[mc->mc_dbi];
2667 if (mc->mc_flags & C_SUB) {
2668 for (; m2; m2=m2->mc_next) {
2669 m3 = &m2->mc_xcursor->mx_cursor;
2670 if (m3->mc_snum < mc->mc_snum) continue;
2671 if (m3->mc_pg[mc->mc_top] == mp)
2672 m3->mc_pg[mc->mc_top] = np;
2675 for (; m2; m2=m2->mc_next) {
2676 if (m2->mc_snum < mc->mc_snum) continue;
2677 if (m2 == mc) continue;
2678 if (m2->mc_pg[mc->mc_top] == mp) {
2679 m2->mc_pg[mc->mc_top] = np;
2680 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2682 (m2->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
2684 MDB_node *leaf = NODEPTR(np, m2->mc_ki[mc->mc_top]);
2685 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
2686 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2691 MDB_PAGE_UNREF(mc->mc_txn, mp);
2695 txn->mt_flags |= MDB_TXN_ERROR;
2700 mdb_env_sync0(MDB_env *env, int force, pgno_t numpgs)
2703 if (env->me_flags & MDB_RDONLY)
2705 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2706 if (env->me_flags & MDB_WRITEMAP) {
2707 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2708 ? MS_ASYNC : MS_SYNC;
2709 if (MDB_MSYNC(env->me_map, env->me_psize * numpgs, flags))
2712 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2716 #ifdef BROKEN_FDATASYNC
2717 if (env->me_flags & MDB_FSYNCONLY) {
2718 if (fsync(env->me_fd))
2722 if (MDB_FDATASYNC(env->me_fd))
2730 mdb_env_sync(MDB_env *env, int force)
2732 MDB_meta *m = mdb_env_pick_meta(env);
2733 return mdb_env_sync0(env, force, m->mm_last_pg+1);
2736 /** Back up parent txn's cursors, then grab the originals for tracking */
2738 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2740 MDB_cursor *mc, *bk;
2745 for (i = src->mt_numdbs; --i >= 0; ) {
2746 if ((mc = src->mt_cursors[i]) != NULL) {
2747 size = sizeof(MDB_cursor);
2749 size += sizeof(MDB_xcursor);
2750 for (; mc; mc = bk->mc_next) {
2756 mc->mc_db = &dst->mt_dbs[i];
2757 /* Kill pointers into src to reduce abuse: The
2758 * user may not use mc until dst ends. But we need a valid
2759 * txn pointer here for cursor fixups to keep working.
2762 mc->mc_dbflag = &dst->mt_dbflags[i];
2763 if ((mx = mc->mc_xcursor) != NULL) {
2764 *(MDB_xcursor *)(bk+1) = *mx;
2765 mx->mx_cursor.mc_txn = dst;
2767 mc->mc_next = dst->mt_cursors[i];
2768 dst->mt_cursors[i] = mc;
2775 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2776 * @param[in] txn the transaction handle.
2777 * @param[in] merge true to keep changes to parent cursors, false to revert.
2778 * @return 0 on success, non-zero on failure.
2781 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2783 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2787 for (i = txn->mt_numdbs; --i >= 0; ) {
2788 for (mc = cursors[i]; mc; mc = next) {
2790 if ((bk = mc->mc_backup) != NULL) {
2792 /* Commit changes to parent txn */
2793 mc->mc_next = bk->mc_next;
2794 mc->mc_backup = bk->mc_backup;
2795 mc->mc_txn = bk->mc_txn;
2796 mc->mc_db = bk->mc_db;
2797 mc->mc_dbflag = bk->mc_dbflag;
2798 if ((mx = mc->mc_xcursor) != NULL)
2799 mx->mx_cursor.mc_txn = bk->mc_txn;
2801 /* Abort nested txn */
2803 if ((mx = mc->mc_xcursor) != NULL)
2804 *mx = *(MDB_xcursor *)(bk+1);
2808 /* Only malloced cursors are permanently tracked. */
2815 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2821 Pidset = F_SETLK, Pidcheck = F_GETLK
2825 /** Set or check a pid lock. Set returns 0 on success.
2826 * Check returns 0 if the process is certainly dead, nonzero if it may
2827 * be alive (the lock exists or an error happened so we do not know).
2829 * On Windows Pidset is a no-op, we merely check for the existence
2830 * of the process with the given pid. On POSIX we use a single byte
2831 * lock on the lockfile, set at an offset equal to the pid.
2834 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2836 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2839 if (op == Pidcheck) {
2840 h = OpenProcess(env->me_pidquery, FALSE, pid);
2841 /* No documented "no such process" code, but other program use this: */
2843 return ErrCode() != ERROR_INVALID_PARAMETER;
2844 /* A process exists until all handles to it close. Has it exited? */
2845 ret = WaitForSingleObject(h, 0) != 0;
2852 struct flock lock_info;
2853 memset(&lock_info, 0, sizeof(lock_info));
2854 lock_info.l_type = F_WRLCK;
2855 lock_info.l_whence = SEEK_SET;
2856 lock_info.l_start = pid;
2857 lock_info.l_len = 1;
2858 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2859 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2861 } else if ((rc = ErrCode()) == EINTR) {
2869 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2870 * @param[in] txn the transaction handle to initialize
2871 * @return 0 on success, non-zero on failure.
2874 mdb_txn_renew0(MDB_txn *txn)
2876 MDB_env *env = txn->mt_env;
2877 MDB_txninfo *ti = env->me_txns;
2879 unsigned int i, nr, flags = txn->mt_flags;
2881 int rc, new_notls = 0;
2883 if ((flags &= MDB_TXN_RDONLY) != 0) {
2885 meta = mdb_env_pick_meta(env);
2886 txn->mt_txnid = meta->mm_txnid;
2887 txn->mt_u.reader = NULL;
2889 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2890 pthread_getspecific(env->me_txkey);
2892 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2893 return MDB_BAD_RSLOT;
2895 MDB_PID_T pid = env->me_pid;
2896 MDB_THR_T tid = pthread_self();
2897 mdb_mutexref_t rmutex = env->me_rmutex;
2899 if (!env->me_live_reader) {
2900 rc = mdb_reader_pid(env, Pidset, pid);
2903 env->me_live_reader = 1;
2906 if (LOCK_MUTEX(rc, env, rmutex))
2908 nr = ti->mti_numreaders;
2909 for (i=0; i<nr; i++)
2910 if (ti->mti_readers[i].mr_pid == 0)
2912 if (i == env->me_maxreaders) {
2913 UNLOCK_MUTEX(rmutex);
2914 return MDB_READERS_FULL;
2916 r = &ti->mti_readers[i];
2917 /* Claim the reader slot, carefully since other code
2918 * uses the reader table un-mutexed: First reset the
2919 * slot, next publish it in mti_numreaders. After
2920 * that, it is safe for mdb_env_close() to touch it.
2921 * When it will be closed, we can finally claim it.
2924 r->mr_txnid = (txnid_t)-1;
2927 ti->mti_numreaders = ++nr;
2928 env->me_close_readers = nr;
2930 UNLOCK_MUTEX(rmutex);
2932 new_notls = (env->me_flags & MDB_NOTLS);
2933 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2938 do /* LY: Retry on a race, ITS#7970. */
2939 r->mr_txnid = ti->mti_txnid;
2940 while(r->mr_txnid != ti->mti_txnid);
2941 txn->mt_txnid = r->mr_txnid;
2942 txn->mt_u.reader = r;
2943 meta = env->me_metas[txn->mt_txnid & 1];
2947 /* Not yet touching txn == env->me_txn0, it may be active */
2949 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2951 txn->mt_txnid = ti->mti_txnid;
2952 meta = env->me_metas[txn->mt_txnid & 1];
2954 meta = mdb_env_pick_meta(env);
2955 txn->mt_txnid = meta->mm_txnid;
2959 if (txn->mt_txnid == mdb_debug_start)
2962 txn->mt_child = NULL;
2963 txn->mt_loose_pgs = NULL;
2964 txn->mt_loose_count = 0;
2965 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2966 txn->mt_u.dirty_list = env->me_dirty_list;
2967 txn->mt_u.dirty_list[0].mid = 0;
2968 txn->mt_free_pgs = env->me_free_pgs;
2969 txn->mt_free_pgs[0] = 0;
2970 txn->mt_spill_pgs = NULL;
2972 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2975 /* Copy the DB info and flags */
2976 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2978 /* Moved to here to avoid a data race in read TXNs */
2979 txn->mt_next_pgno = meta->mm_last_pg+1;
2981 txn->mt_last_pgno = txn->mt_next_pgno - 1;
2984 txn->mt_flags = flags;
2987 txn->mt_numdbs = env->me_numdbs;
2988 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2989 x = env->me_dbflags[i];
2990 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2991 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2993 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2994 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2996 if (env->me_flags & MDB_FATAL_ERROR) {
2997 DPUTS("environment had fatal error, must shutdown!");
2999 } else if (env->me_maxpg < txn->mt_next_pgno) {
3000 rc = MDB_MAP_RESIZED;
3004 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
3009 mdb_txn_renew(MDB_txn *txn)
3013 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
3016 rc = mdb_txn_renew0(txn);
3017 if (rc == MDB_SUCCESS) {
3018 DPRINTF(("renew txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3019 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3020 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
3026 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
3030 int rc, size, tsize;
3032 flags &= MDB_TXN_BEGIN_FLAGS;
3033 flags |= env->me_flags & MDB_WRITEMAP;
3035 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
3039 /* Nested transactions: Max 1 child, write txns only, no writemap */
3040 flags |= parent->mt_flags;
3041 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
3042 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
3044 /* Child txns save MDB_pgstate and use own copy of cursors */
3045 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
3046 size += tsize = sizeof(MDB_ntxn);
3047 } else if (flags & MDB_RDONLY) {
3048 size = env->me_maxdbs * (sizeof(MDB_db)+1);
3049 size += tsize = sizeof(MDB_txn);
3051 /* Reuse preallocated write txn. However, do not touch it until
3052 * mdb_txn_renew0() succeeds, since it currently may be active.
3057 if ((txn = calloc(1, size)) == NULL) {
3058 DPRINTF(("calloc: %s", strerror(errno)));
3063 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
3064 if (!txn->mt_rpages) {
3068 txn->mt_rpages[0].mid = 0;
3069 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
3072 txn->mt_dbxs = env->me_dbxs; /* static */
3073 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
3074 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
3075 txn->mt_flags = flags;
3080 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
3081 txn->mt_dbiseqs = parent->mt_dbiseqs;
3082 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
3083 if (!txn->mt_u.dirty_list ||
3084 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
3086 free(txn->mt_u.dirty_list);
3090 txn->mt_txnid = parent->mt_txnid;
3091 txn->mt_dirty_room = parent->mt_dirty_room;
3092 txn->mt_u.dirty_list[0].mid = 0;
3093 txn->mt_spill_pgs = NULL;
3094 txn->mt_next_pgno = parent->mt_next_pgno;
3095 parent->mt_flags |= MDB_TXN_HAS_CHILD;
3096 parent->mt_child = txn;
3097 txn->mt_parent = parent;
3098 txn->mt_numdbs = parent->mt_numdbs;
3100 txn->mt_rpages = parent->mt_rpages;
3102 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3103 /* Copy parent's mt_dbflags, but clear DB_NEW */
3104 for (i=0; i<txn->mt_numdbs; i++)
3105 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
3107 ntxn = (MDB_ntxn *)txn;
3108 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
3109 if (env->me_pghead) {
3110 size = MDB_IDL_SIZEOF(env->me_pghead);
3111 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
3113 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
3118 rc = mdb_cursor_shadow(parent, txn);
3120 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
3121 } else { /* MDB_RDONLY */
3122 txn->mt_dbiseqs = env->me_dbiseqs;
3124 rc = mdb_txn_renew0(txn);
3127 if (txn != env->me_txn0) {
3129 free(txn->mt_rpages);
3134 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
3136 DPRINTF(("begin txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3137 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
3138 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
3145 mdb_txn_env(MDB_txn *txn)
3147 if(!txn) return NULL;
3152 mdb_txn_id(MDB_txn *txn)
3155 return txn->mt_txnid;
3158 /** Export or close DBI handles opened in this txn. */
3160 mdb_dbis_update(MDB_txn *txn, int keep)
3163 MDB_dbi n = txn->mt_numdbs;
3164 MDB_env *env = txn->mt_env;
3165 unsigned char *tdbflags = txn->mt_dbflags;
3167 for (i = n; --i >= CORE_DBS;) {
3168 if (tdbflags[i] & DB_NEW) {
3170 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
3172 char *ptr = env->me_dbxs[i].md_name.mv_data;
3174 env->me_dbxs[i].md_name.mv_data = NULL;
3175 env->me_dbxs[i].md_name.mv_size = 0;
3176 env->me_dbflags[i] = 0;
3177 env->me_dbiseqs[i]++;
3183 if (keep && env->me_numdbs < n)
3187 /** End a transaction, except successful commit of a nested transaction.
3188 * May be called twice for readonly txns: First reset it, then abort.
3189 * @param[in] txn the transaction handle to end
3190 * @param[in] mode why and how to end the transaction
3193 mdb_txn_end(MDB_txn *txn, unsigned mode)
3195 MDB_env *env = txn->mt_env;
3197 static const char *const names[] = MDB_END_NAMES;
3200 /* Export or close DBI handles opened in this txn */
3201 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
3203 DPRINTF(("%s txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3204 names[mode & MDB_END_OPMASK],
3205 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3206 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
3208 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3209 if (txn->mt_u.reader) {
3210 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
3211 if (!(env->me_flags & MDB_NOTLS)) {
3212 txn->mt_u.reader = NULL; /* txn does not own reader */
3213 } else if (mode & MDB_END_SLOT) {
3214 txn->mt_u.reader->mr_pid = 0;
3215 txn->mt_u.reader = NULL;
3216 } /* else txn owns the slot until it does MDB_END_SLOT */
3218 txn->mt_numdbs = 0; /* prevent further DBI activity */
3219 txn->mt_flags |= MDB_TXN_FINISHED;
3221 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
3222 pgno_t *pghead = env->me_pghead;
3224 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
3225 mdb_cursors_close(txn, 0);
3226 if (!(env->me_flags & MDB_WRITEMAP)) {
3227 mdb_dlist_free(txn);
3231 txn->mt_flags = MDB_TXN_FINISHED;
3233 if (!txn->mt_parent) {
3234 mdb_midl_shrink(&txn->mt_free_pgs);
3235 env->me_free_pgs = txn->mt_free_pgs;
3237 env->me_pghead = NULL;
3241 mode = 0; /* txn == env->me_txn0, do not free() it */
3243 /* The writer mutex was locked in mdb_txn_begin. */
3245 UNLOCK_MUTEX(env->me_wmutex);
3247 txn->mt_parent->mt_child = NULL;
3248 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3249 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3250 mdb_midl_free(txn->mt_free_pgs);
3251 mdb_midl_free(txn->mt_spill_pgs);
3252 free(txn->mt_u.dirty_list);
3255 mdb_midl_free(pghead);
3258 if (!txn->mt_parent) {
3259 MDB_ID3L el = env->me_rpages, tl = txn->mt_rpages;
3260 unsigned i, x, n = tl[0].mid;
3261 pthread_mutex_lock(&env->me_rpmutex);
3262 for (i = 1; i <= n; i++) {
3263 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
3264 /* tmp overflow pages that we didn't share in env */
3265 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3267 x = mdb_mid3l_search(el, tl[i].mid);
3268 if (tl[i].mptr == el[x].mptr) {
3271 /* another tmp overflow page */
3272 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3276 pthread_mutex_unlock(&env->me_rpmutex);
3278 if (mode & MDB_END_FREE)
3282 if (mode & MDB_END_FREE)
3287 mdb_txn_reset(MDB_txn *txn)
3292 /* This call is only valid for read-only txns */
3293 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3296 mdb_txn_end(txn, MDB_END_RESET);
3300 mdb_txn_abort(MDB_txn *txn)
3306 mdb_txn_abort(txn->mt_child);
3308 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3311 /** Save the freelist as of this transaction to the freeDB.
3312 * This changes the freelist. Keep trying until it stabilizes.
3314 * When (MDB_DEVEL) & 2, the changes do not affect #mdb_page_alloc(),
3315 * it then uses the transaction's original snapshot of the freeDB.
3318 mdb_freelist_save(MDB_txn *txn)
3320 /* env->me_pghead[] can grow and shrink during this call.
3321 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3322 * Page numbers cannot disappear from txn->mt_free_pgs[].
3325 MDB_env *env = txn->mt_env;
3326 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3327 txnid_t pglast = 0, head_id = 0;
3328 pgno_t freecnt = 0, *free_pgs, *mop;
3329 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3331 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3333 if (env->me_pghead) {
3334 /* Make sure first page of freeDB is touched and on freelist */
3335 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3336 if (rc && rc != MDB_NOTFOUND)
3340 if (!env->me_pghead && txn->mt_loose_pgs) {
3341 /* Put loose page numbers in mt_free_pgs, since
3342 * we may be unable to return them to me_pghead.
3344 MDB_page *mp = txn->mt_loose_pgs;
3345 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3347 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3348 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3349 txn->mt_loose_pgs = NULL;
3350 txn->mt_loose_count = 0;
3353 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3354 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3355 ? SSIZE_MAX : maxfree_1pg;
3358 /* Come back here after each Put() in case freelist changed */
3363 /* If using records from freeDB which we have not yet
3364 * deleted, delete them and any we reserved for me_pghead.
3366 while (pglast < env->me_pglast) {
3367 rc = mdb_cursor_first(&mc, &key, NULL);
3370 pglast = head_id = *(txnid_t *)key.mv_data;
3371 total_room = head_room = 0;
3372 mdb_tassert(txn, pglast <= env->me_pglast);
3373 rc = mdb_cursor_del(&mc, 0);
3378 /* Save the IDL of pages freed by this txn, to a single record */
3379 if (freecnt < txn->mt_free_pgs[0]) {
3381 /* Make sure last page of freeDB is touched and on freelist */
3382 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3383 if (rc && rc != MDB_NOTFOUND)
3386 free_pgs = txn->mt_free_pgs;
3387 /* Write to last page of freeDB */
3388 key.mv_size = sizeof(txn->mt_txnid);
3389 key.mv_data = &txn->mt_txnid;
3391 freecnt = free_pgs[0];
3392 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3393 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3396 /* Retry if mt_free_pgs[] grew during the Put() */
3397 free_pgs = txn->mt_free_pgs;
3398 } while (freecnt < free_pgs[0]);
3399 mdb_midl_sort(free_pgs);
3400 memcpy(data.mv_data, free_pgs, data.mv_size);
3403 unsigned int i = free_pgs[0];
3404 DPRINTF(("IDL write txn %"Yu" root %"Yu" num %u",
3405 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3407 DPRINTF(("IDL %"Yu, free_pgs[i]));
3413 mop = env->me_pghead;
3414 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3416 /* Reserve records for me_pghead[]. Split it if multi-page,
3417 * to avoid searching freeDB for a page range. Use keys in
3418 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3420 if (total_room >= mop_len) {
3421 if (total_room == mop_len || --more < 0)
3423 } else if (head_room >= maxfree_1pg && head_id > 1) {
3424 /* Keep current record (overflow page), add a new one */
3428 /* (Re)write {key = head_id, IDL length = head_room} */
3429 total_room -= head_room;
3430 head_room = mop_len - total_room;
3431 if (head_room > maxfree_1pg && head_id > 1) {
3432 /* Overflow multi-page for part of me_pghead */
3433 head_room /= head_id; /* amortize page sizes */
3434 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3435 } else if (head_room < 0) {
3436 /* Rare case, not bothering to delete this record */
3439 key.mv_size = sizeof(head_id);
3440 key.mv_data = &head_id;
3441 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3442 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3445 /* IDL is initially empty, zero out at least the length */
3446 pgs = (pgno_t *)data.mv_data;
3447 j = head_room > clean_limit ? head_room : 0;
3451 total_room += head_room;
3454 /* Return loose page numbers to me_pghead, though usually none are
3455 * left at this point. The pages themselves remain in dirty_list.
3457 if (txn->mt_loose_pgs) {
3458 MDB_page *mp = txn->mt_loose_pgs;
3459 unsigned count = txn->mt_loose_count;
3461 /* Room for loose pages + temp IDL with same */
3462 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3464 mop = env->me_pghead;
3465 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3466 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3467 loose[ ++count ] = mp->mp_pgno;
3469 mdb_midl_sort(loose);
3470 mdb_midl_xmerge(mop, loose);
3471 txn->mt_loose_pgs = NULL;
3472 txn->mt_loose_count = 0;
3476 /* Fill in the reserved me_pghead records */
3482 rc = mdb_cursor_first(&mc, &key, &data);
3483 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3484 txnid_t id = *(txnid_t *)key.mv_data;
3485 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3488 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3490 if (len > mop_len) {
3492 data.mv_size = (len + 1) * sizeof(MDB_ID);
3494 data.mv_data = mop -= len;
3497 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3499 if (rc || !(mop_len -= len))
3506 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3507 * @param[in] txn the transaction that's being committed
3508 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3509 * @return 0 on success, non-zero on failure.
3512 mdb_page_flush(MDB_txn *txn, int keep)
3514 MDB_env *env = txn->mt_env;
3515 MDB_ID2L dl = txn->mt_u.dirty_list;
3516 unsigned psize = env->me_psize, j;
3517 int i, pagecount = dl[0].mid, rc;
3521 MDB_page *dp = NULL;
3525 struct iovec iov[MDB_COMMIT_PAGES];
3526 ssize_t wsize = 0, wres;
3527 off_t wpos = 0, next_pos = 1; /* impossible pos, so pos != next_pos */
3533 if (env->me_flags & MDB_WRITEMAP) {
3534 /* Clear dirty flags */
3535 while (++i <= pagecount) {
3537 /* Don't flush this page yet */
3538 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3539 dp->mp_flags &= ~P_KEEP;
3543 dp->mp_flags &= ~P_DIRTY;
3548 /* Write the pages */
3550 if (++i <= pagecount) {
3552 /* Don't flush this page yet */
3553 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3554 dp->mp_flags &= ~P_KEEP;
3559 /* clear dirty flag */
3560 dp->mp_flags &= ~P_DIRTY;
3563 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3568 /* Windows actually supports scatter/gather I/O, but only on
3569 * unbuffered file handles. Since we're relying on the OS page
3570 * cache for all our data, that's self-defeating. So we just
3571 * write pages one at a time. We use the ov structure to set
3572 * the write offset, to at least save the overhead of a Seek
3575 DPRINTF(("committing page %"Yu, pgno));
3576 memset(&ov, 0, sizeof(ov));
3577 ov.Offset = pos & 0xffffffff;
3578 ov.OffsetHigh = pos >> 16 >> 16;
3579 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3581 DPRINTF(("WriteFile: %d", rc));
3585 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3586 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3589 /* Write previous page(s) */
3590 #ifdef MDB_USE_PWRITEV
3591 wres = pwritev(env->me_fd, iov, n, wpos);
3594 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3597 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3601 DPRINTF(("lseek: %s", strerror(rc)));
3604 wres = writev(env->me_fd, iov, n);
3607 if (wres != wsize) {
3612 DPRINTF(("Write error: %s", strerror(rc)));
3614 rc = EIO; /* TODO: Use which error code? */
3615 DPUTS("short write, filesystem full?");
3626 DPRINTF(("committing page %"Yu, pgno));
3627 next_pos = pos + size;
3628 iov[n].iov_len = size;
3629 iov[n].iov_base = (char *)dp;
3635 if (pgno > txn->mt_last_pgno)
3636 txn->mt_last_pgno = pgno;
3639 /* MIPS has cache coherency issues, this is a no-op everywhere else
3640 * Note: for any size >= on-chip cache size, entire on-chip cache is
3643 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3645 for (i = keep; ++i <= pagecount; ) {
3647 /* This is a page we skipped above */
3650 dl[j].mid = dp->mp_pgno;
3653 mdb_dpage_free(env, dp);
3658 txn->mt_dirty_room += i - j;
3664 mdb_txn_commit(MDB_txn *txn)
3667 unsigned int i, end_mode;
3673 /* mdb_txn_end() mode for a commit which writes nothing */
3674 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3676 if (txn->mt_child) {
3677 rc = mdb_txn_commit(txn->mt_child);
3684 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3688 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3689 DPUTS("txn has failed/finished, can't commit");
3691 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3696 if (txn->mt_parent) {
3697 MDB_txn *parent = txn->mt_parent;
3701 unsigned x, y, len, ps_len;
3703 /* Append our free list to parent's */
3704 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3707 mdb_midl_free(txn->mt_free_pgs);
3708 /* Failures after this must either undo the changes
3709 * to the parent or set MDB_TXN_ERROR in the parent.
3712 parent->mt_next_pgno = txn->mt_next_pgno;
3713 parent->mt_flags = txn->mt_flags;
3715 /* Merge our cursors into parent's and close them */
3716 mdb_cursors_close(txn, 1);
3718 /* Update parent's DB table. */
3719 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3720 parent->mt_numdbs = txn->mt_numdbs;
3721 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3722 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3723 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3724 /* preserve parent's DB_NEW status */
3725 x = parent->mt_dbflags[i] & DB_NEW;
3726 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3729 dst = parent->mt_u.dirty_list;
3730 src = txn->mt_u.dirty_list;
3731 /* Remove anything in our dirty list from parent's spill list */
3732 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3734 pspill[0] = (pgno_t)-1;
3735 /* Mark our dirty pages as deleted in parent spill list */
3736 for (i=0, len=src[0].mid; ++i <= len; ) {
3737 MDB_ID pn = src[i].mid << 1;
3738 while (pn > pspill[x])
3740 if (pn == pspill[x]) {
3745 /* Squash deleted pagenums if we deleted any */
3746 for (x=y; ++x <= ps_len; )
3747 if (!(pspill[x] & 1))
3748 pspill[++y] = pspill[x];
3752 /* Remove anything in our spill list from parent's dirty list */
3753 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3754 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3755 MDB_ID pn = txn->mt_spill_pgs[i];
3757 continue; /* deleted spillpg */
3759 y = mdb_mid2l_search(dst, pn);
3760 if (y <= dst[0].mid && dst[y].mid == pn) {
3762 while (y < dst[0].mid) {
3771 /* Find len = length of merging our dirty list with parent's */
3773 dst[0].mid = 0; /* simplify loops */
3774 if (parent->mt_parent) {
3775 len = x + src[0].mid;
3776 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3777 for (i = x; y && i; y--) {
3778 pgno_t yp = src[y].mid;
3779 while (yp < dst[i].mid)
3781 if (yp == dst[i].mid) {
3786 } else { /* Simplify the above for single-ancestor case */
3787 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3789 /* Merge our dirty list with parent's */
3791 for (i = len; y; dst[i--] = src[y--]) {
3792 pgno_t yp = src[y].mid;
3793 while (yp < dst[x].mid)
3794 dst[i--] = dst[x--];
3795 if (yp == dst[x].mid)
3796 free(dst[x--].mptr);
3798 mdb_tassert(txn, i == x);
3800 free(txn->mt_u.dirty_list);
3801 parent->mt_dirty_room = txn->mt_dirty_room;
3802 if (txn->mt_spill_pgs) {
3803 if (parent->mt_spill_pgs) {
3804 /* TODO: Prevent failure here, so parent does not fail */
3805 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3807 parent->mt_flags |= MDB_TXN_ERROR;
3808 mdb_midl_free(txn->mt_spill_pgs);
3809 mdb_midl_sort(parent->mt_spill_pgs);
3811 parent->mt_spill_pgs = txn->mt_spill_pgs;
3815 /* Append our loose page list to parent's */
3816 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3818 *lp = txn->mt_loose_pgs;
3819 parent->mt_loose_count += txn->mt_loose_count;
3821 parent->mt_child = NULL;
3822 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3827 if (txn != env->me_txn) {
3828 DPUTS("attempt to commit unknown transaction");
3833 mdb_cursors_close(txn, 0);
3835 if (!txn->mt_u.dirty_list[0].mid &&
3836 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3839 DPRINTF(("committing txn %"Yu" %p on mdbenv %p, root page %"Yu,
3840 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3842 /* Update DB root pointers */
3843 if (txn->mt_numdbs > CORE_DBS) {
3847 data.mv_size = sizeof(MDB_db);
3849 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3850 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3851 if (txn->mt_dbflags[i] & DB_DIRTY) {
3852 if (TXN_DBI_CHANGED(txn, i)) {
3856 data.mv_data = &txn->mt_dbs[i];
3857 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3865 rc = mdb_freelist_save(txn);
3869 mdb_midl_free(env->me_pghead);
3870 env->me_pghead = NULL;
3871 mdb_midl_shrink(&txn->mt_free_pgs);
3877 if ((rc = mdb_page_flush(txn, 0)))
3879 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3880 (rc = mdb_env_sync0(env, 0, txn->mt_next_pgno)))
3882 if ((rc = mdb_env_write_meta(txn)))
3884 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3887 mdb_txn_end(txn, end_mode);
3895 /** Read the environment parameters of a DB environment before
3896 * mapping it into memory.
3897 * @param[in] env the environment handle
3898 * @param[out] meta address of where to store the meta information
3899 * @return 0 on success, non-zero on failure.
3902 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3908 enum { Size = sizeof(pbuf) };
3910 /* We don't know the page size yet, so use a minimum value.
3911 * Read both meta pages so we can use the latest one.
3914 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3918 memset(&ov, 0, sizeof(ov));
3920 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3921 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3924 rc = pread(env->me_fd, &pbuf, Size, off);
3927 if (rc == 0 && off == 0)
3929 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3930 DPRINTF(("read: %s", mdb_strerror(rc)));
3934 p = (MDB_page *)&pbuf;
3936 if (!F_ISSET(p->mp_flags, P_META)) {
3937 DPRINTF(("page %"Yu" not a meta page", p->mp_pgno));
3942 if (m->mm_magic != MDB_MAGIC) {
3943 DPUTS("meta has invalid magic");
3947 if (m->mm_version != MDB_DATA_VERSION) {
3948 DPRINTF(("database is version %u, expected version %u",
3949 m->mm_version, MDB_DATA_VERSION));
3950 return MDB_VERSION_MISMATCH;
3953 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3959 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3961 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3963 meta->mm_magic = MDB_MAGIC;
3964 meta->mm_version = MDB_DATA_VERSION;
3965 meta->mm_mapsize = env->me_mapsize;
3966 meta->mm_psize = env->me_psize;
3967 meta->mm_last_pg = NUM_METAS-1;
3968 meta->mm_flags = env->me_flags & 0xffff;
3969 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3970 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3971 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3974 /** Write the environment parameters of a freshly created DB environment.
3975 * @param[in] env the environment handle
3976 * @param[in] meta the #MDB_meta to write
3977 * @return 0 on success, non-zero on failure.
3980 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3988 memset(&ov, 0, sizeof(ov));
3989 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3991 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3994 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3995 len = pwrite(fd, ptr, size, pos); \
3996 if (len == -1 && ErrCode() == EINTR) continue; \
3997 rc = (len >= 0); break; } while(1)
4000 DPUTS("writing new meta page");
4002 psize = env->me_psize;
4004 p = calloc(NUM_METAS, psize);
4008 p->mp_flags = P_META;
4009 *(MDB_meta *)METADATA(p) = *meta;
4011 q = (MDB_page *)((char *)p + psize);
4013 q->mp_flags = P_META;
4014 *(MDB_meta *)METADATA(q) = *meta;
4016 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
4019 else if ((unsigned) len == psize * NUM_METAS)
4027 /** Update the environment info to commit a transaction.
4028 * @param[in] txn the transaction that's being committed
4029 * @return 0 on success, non-zero on failure.
4032 mdb_env_write_meta(MDB_txn *txn)
4035 MDB_meta meta, metab, *mp;
4039 int rc, len, toggle;
4048 toggle = txn->mt_txnid & 1;
4049 DPRINTF(("writing meta page %d for root page %"Yu,
4050 toggle, txn->mt_dbs[MAIN_DBI].md_root));
4053 flags = txn->mt_flags | env->me_flags;
4054 mp = env->me_metas[toggle];
4055 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
4056 /* Persist any increases of mapsize config */
4057 if (mapsize < env->me_mapsize)
4058 mapsize = env->me_mapsize;
4060 if (flags & MDB_WRITEMAP) {
4061 mp->mm_mapsize = mapsize;
4062 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4063 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4064 mp->mm_last_pg = txn->mt_next_pgno - 1;
4065 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
4066 !(defined(__i386__) || defined(__x86_64__))
4067 /* LY: issue a memory barrier, if not x86. ITS#7969 */
4068 __sync_synchronize();
4070 mp->mm_txnid = txn->mt_txnid;
4071 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
4072 unsigned meta_size = env->me_psize;
4073 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
4074 ptr = (char *)mp - PAGEHDRSZ;
4075 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
4076 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
4080 if (MDB_MSYNC(ptr, meta_size, rc)) {
4087 metab.mm_txnid = mp->mm_txnid;
4088 metab.mm_last_pg = mp->mm_last_pg;
4090 meta.mm_mapsize = mapsize;
4091 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4092 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4093 meta.mm_last_pg = txn->mt_next_pgno - 1;
4094 meta.mm_txnid = txn->mt_txnid;
4096 off = offsetof(MDB_meta, mm_mapsize);
4097 ptr = (char *)&meta + off;
4098 len = sizeof(MDB_meta) - off;
4099 off += (char *)mp - env->me_map;
4101 /* Write to the SYNC fd */
4102 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
4105 memset(&ov, 0, sizeof(ov));
4107 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
4112 rc = pwrite(mfd, ptr, len, off);
4115 rc = rc < 0 ? ErrCode() : EIO;
4120 DPUTS("write failed, disk error?");
4121 /* On a failure, the pagecache still contains the new data.
4122 * Write some old data back, to prevent it from being used.
4123 * Use the non-SYNC fd; we know it will fail anyway.
4125 meta.mm_last_pg = metab.mm_last_pg;
4126 meta.mm_txnid = metab.mm_txnid;
4128 memset(&ov, 0, sizeof(ov));
4130 WriteFile(env->me_fd, ptr, len, NULL, &ov);
4132 r2 = pwrite(env->me_fd, ptr, len, off);
4133 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
4136 env->me_flags |= MDB_FATAL_ERROR;
4139 /* MIPS has cache coherency issues, this is a no-op everywhere else */
4140 CACHEFLUSH(env->me_map + off, len, DCACHE);
4142 /* Memory ordering issues are irrelevant; since the entire writer
4143 * is wrapped by wmutex, all of these changes will become visible
4144 * after the wmutex is unlocked. Since the DB is multi-version,
4145 * readers will get consistent data regardless of how fresh or
4146 * how stale their view of these values is.
4149 env->me_txns->mti_txnid = txn->mt_txnid;
4154 /** Check both meta pages to see which one is newer.
4155 * @param[in] env the environment handle
4156 * @return newest #MDB_meta.
4159 mdb_env_pick_meta(const MDB_env *env)
4161 MDB_meta *const *metas = env->me_metas;
4162 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
4166 mdb_env_create(MDB_env **env)
4170 e = calloc(1, sizeof(MDB_env));
4174 e->me_maxreaders = DEFAULT_READERS;
4175 e->me_maxdbs = e->me_numdbs = CORE_DBS;
4176 e->me_fd = INVALID_HANDLE_VALUE;
4177 e->me_lfd = INVALID_HANDLE_VALUE;
4178 e->me_mfd = INVALID_HANDLE_VALUE;
4179 #ifdef MDB_USE_POSIX_SEM
4180 e->me_rmutex = SEM_FAILED;
4181 e->me_wmutex = SEM_FAILED;
4182 #elif defined MDB_USE_SYSV_SEM
4183 e->me_rmutex->semid = -1;
4184 e->me_wmutex->semid = -1;
4186 e->me_pid = getpid();
4187 GET_PAGESIZE(e->me_os_psize);
4188 VGMEMP_CREATE(e,0,0);
4194 /** @brief Map a result from an NTAPI call to WIN32. */
4196 mdb_nt2win32(NTSTATUS st)
4201 GetOverlappedResult(NULL, &o, &br, FALSE);
4202 return GetLastError();
4207 mdb_env_map(MDB_env *env, void *addr)
4210 unsigned int flags = env->me_flags;
4213 int access = SECTION_MAP_READ;
4217 ULONG pageprot = PAGE_READONLY, secprot, alloctype;
4219 if (flags & MDB_WRITEMAP) {
4220 access |= SECTION_MAP_WRITE;
4221 pageprot = PAGE_READWRITE;
4223 if (flags & MDB_RDONLY) {
4224 secprot = PAGE_READONLY;
4228 secprot = PAGE_READWRITE;
4229 msize = env->me_mapsize;
4230 alloctype = MEM_RESERVE;
4233 rc = NtCreateSection(&mh, access, NULL, NULL, secprot, SEC_RESERVE, env->me_fd);
4235 return mdb_nt2win32(rc);
4238 msize = NUM_METAS * env->me_psize;
4240 rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
4247 return mdb_nt2win32(rc);
4252 env->me_map = mmap(addr, NUM_METAS * env->me_psize, PROT_READ, MAP_SHARED,
4254 if (env->me_map == MAP_FAILED) {
4259 int prot = PROT_READ;
4260 if (flags & MDB_WRITEMAP) {
4262 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
4265 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
4267 if (env->me_map == MAP_FAILED) {
4272 if (flags & MDB_NORDAHEAD) {
4273 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
4275 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
4277 #ifdef POSIX_MADV_RANDOM
4278 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
4279 #endif /* POSIX_MADV_RANDOM */
4280 #endif /* MADV_RANDOM */
4284 /* Can happen because the address argument to mmap() is just a
4285 * hint. mmap() can pick another, e.g. if the range is in use.
4286 * The MAP_FIXED flag would prevent that, but then mmap could
4287 * instead unmap existing pages to make room for the new map.
4289 if (addr && env->me_map != addr)
4290 return EBUSY; /* TODO: Make a new MDB_* error code? */
4293 p = (MDB_page *)env->me_map;
4294 env->me_metas[0] = METADATA(p);
4295 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4301 mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
4303 /* If env is already open, caller is responsible for making
4304 * sure there are no active txns.
4314 meta = mdb_env_pick_meta(env);
4316 size = meta->mm_mapsize;
4318 /* Silently round up to minimum if the size is too small */
4319 mdb_size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4324 /* For MDB_VL32 this bit is a noop since we dynamically remap
4325 * chunks of the DB anyway.
4327 munmap(env->me_map, env->me_mapsize);
4328 env->me_mapsize = size;
4329 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4330 rc = mdb_env_map(env, old);
4333 #endif /* !MDB_VL32 */
4335 env->me_mapsize = size;
4337 env->me_maxpg = env->me_mapsize / env->me_psize;
4342 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4346 env->me_maxdbs = dbs + CORE_DBS;
4351 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4353 if (env->me_map || readers < 1)
4355 env->me_maxreaders = readers;
4360 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4362 if (!env || !readers)
4364 *readers = env->me_maxreaders;
4369 mdb_fsize(HANDLE fd, mdb_size_t *size)
4372 LARGE_INTEGER fsize;
4374 if (!GetFileSizeEx(fd, &fsize))
4377 *size = fsize.QuadPart;
4389 #ifdef BROKEN_FDATASYNC
4390 #include <sys/utsname.h>
4391 #include <sys/vfs.h>
4394 /** Further setup required for opening an LMDB environment
4397 mdb_env_open2(MDB_env *env)
4399 unsigned int flags = env->me_flags;
4400 int i, newenv = 0, rc;
4404 /* See if we should use QueryLimited */
4406 if ((rc & 0xff) > 5)
4407 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4409 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4412 #ifdef BROKEN_FDATASYNC
4413 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4414 * https://lkml.org/lkml/2012/9/3/83
4415 * Kernels after 3.6-rc6 are known good.
4416 * https://lkml.org/lkml/2012/9/10/556
4417 * See if the DB is on ext3/ext4, then check for new enough kernel
4418 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4423 fstatfs(env->me_fd, &st);
4424 while (st.f_type == 0xEF53) {
4428 if (uts.release[0] < '3') {
4429 if (!strncmp(uts.release, "2.6.32.", 7)) {
4430 i = atoi(uts.release+7);
4432 break; /* 2.6.32.60 and newer is OK */
4433 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4434 i = atoi(uts.release+7);
4436 break; /* 2.6.34.15 and newer is OK */
4438 } else if (uts.release[0] == '3') {
4439 i = atoi(uts.release+2);
4441 break; /* 3.6 and newer is OK */
4443 i = atoi(uts.release+4);
4445 break; /* 3.5.4 and newer is OK */
4446 } else if (i == 2) {
4447 i = atoi(uts.release+4);
4449 break; /* 3.2.30 and newer is OK */
4451 } else { /* 4.x and newer is OK */
4454 env->me_flags |= MDB_FSYNCONLY;
4460 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4463 DPUTS("new mdbenv");
4465 env->me_psize = env->me_os_psize;
4466 if (env->me_psize > MAX_PAGESIZE)
4467 env->me_psize = MAX_PAGESIZE;
4468 memset(&meta, 0, sizeof(meta));
4469 mdb_env_init_meta0(env, &meta);
4470 meta.mm_mapsize = DEFAULT_MAPSIZE;
4472 env->me_psize = meta.mm_psize;
4475 /* Was a mapsize configured? */
4476 if (!env->me_mapsize) {
4477 env->me_mapsize = meta.mm_mapsize;
4480 /* Make sure mapsize >= committed data size. Even when using
4481 * mm_mapsize, which could be broken in old files (ITS#7789).
4483 mdb_size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4484 if (env->me_mapsize < minsize)
4485 env->me_mapsize = minsize;
4487 meta.mm_mapsize = env->me_mapsize;
4489 if (newenv && !(flags & MDB_FIXEDMAP)) {
4490 /* mdb_env_map() may grow the datafile. Write the metapages
4491 * first, so the file will be valid if initialization fails.
4492 * Except with FIXEDMAP, since we do not yet know mm_address.
4493 * We could fill in mm_address later, but then a different
4494 * program might end up doing that - one with a memory layout
4495 * and map address which does not suit the main program.
4497 rc = mdb_env_init_meta(env, &meta);
4503 /* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
4507 rc = WriteFile(env->me_fd, &dummy, 1, &len, NULL);
4515 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4520 if (flags & MDB_FIXEDMAP)
4521 meta.mm_address = env->me_map;
4522 i = mdb_env_init_meta(env, &meta);
4523 if (i != MDB_SUCCESS) {
4528 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4529 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4531 #if !(MDB_MAXKEYSIZE)
4532 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4534 env->me_maxpg = env->me_mapsize / env->me_psize;
4538 MDB_meta *meta = mdb_env_pick_meta(env);
4539 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4541 DPRINTF(("opened database version %u, pagesize %u",
4542 meta->mm_version, env->me_psize));
4543 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4544 DPRINTF(("depth: %u", db->md_depth));
4545 DPRINTF(("entries: %"Yu, db->md_entries));
4546 DPRINTF(("branch pages: %"Yu, db->md_branch_pages));
4547 DPRINTF(("leaf pages: %"Yu, db->md_leaf_pages));
4548 DPRINTF(("overflow pages: %"Yu, db->md_overflow_pages));
4549 DPRINTF(("root: %"Yu, db->md_root));
4557 /** Release a reader thread's slot in the reader lock table.
4558 * This function is called automatically when a thread exits.
4559 * @param[in] ptr This points to the slot in the reader lock table.
4562 mdb_env_reader_dest(void *ptr)
4564 MDB_reader *reader = ptr;
4570 /** Junk for arranging thread-specific callbacks on Windows. This is
4571 * necessarily platform and compiler-specific. Windows supports up
4572 * to 1088 keys. Let's assume nobody opens more than 64 environments
4573 * in a single process, for now. They can override this if needed.
4575 #ifndef MAX_TLS_KEYS
4576 #define MAX_TLS_KEYS 64
4578 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4579 static int mdb_tls_nkeys;
4581 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4585 case DLL_PROCESS_ATTACH: break;
4586 case DLL_THREAD_ATTACH: break;
4587 case DLL_THREAD_DETACH:
4588 for (i=0; i<mdb_tls_nkeys; i++) {
4589 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4591 mdb_env_reader_dest(r);
4595 case DLL_PROCESS_DETACH: break;
4600 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4602 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4606 /* Force some symbol references.
4607 * _tls_used forces the linker to create the TLS directory if not already done
4608 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4610 #pragma comment(linker, "/INCLUDE:_tls_used")
4611 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4612 #pragma const_seg(".CRT$XLB")
4613 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4614 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4617 #pragma comment(linker, "/INCLUDE:__tls_used")
4618 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4619 #pragma data_seg(".CRT$XLB")
4620 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4622 #endif /* WIN 32/64 */
4623 #endif /* !__GNUC__ */
4626 /** Downgrade the exclusive lock on the region back to shared */
4628 mdb_env_share_locks(MDB_env *env, int *excl)
4631 MDB_meta *meta = mdb_env_pick_meta(env);
4633 env->me_txns->mti_txnid = meta->mm_txnid;
4638 /* First acquire a shared lock. The Unlock will
4639 * then release the existing exclusive lock.
4641 memset(&ov, 0, sizeof(ov));
4642 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4645 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4651 struct flock lock_info;
4652 /* The shared lock replaces the existing lock */
4653 memset((void *)&lock_info, 0, sizeof(lock_info));
4654 lock_info.l_type = F_RDLCK;
4655 lock_info.l_whence = SEEK_SET;
4656 lock_info.l_start = 0;
4657 lock_info.l_len = 1;
4658 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4659 (rc = ErrCode()) == EINTR) ;
4660 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4667 /** Try to get exclusive lock, otherwise shared.
4668 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4671 mdb_env_excl_lock(MDB_env *env, int *excl)
4675 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4679 memset(&ov, 0, sizeof(ov));
4680 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4687 struct flock lock_info;
4688 memset((void *)&lock_info, 0, sizeof(lock_info));
4689 lock_info.l_type = F_WRLCK;
4690 lock_info.l_whence = SEEK_SET;
4691 lock_info.l_start = 0;
4692 lock_info.l_len = 1;
4693 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4694 (rc = ErrCode()) == EINTR) ;
4698 # ifndef MDB_USE_POSIX_MUTEX
4699 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4702 lock_info.l_type = F_RDLCK;
4703 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4704 (rc = ErrCode()) == EINTR) ;
4714 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4716 * @(#) $Revision: 5.1 $
4717 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4718 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4720 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4724 * Please do not copyright this code. This code is in the public domain.
4726 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4727 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4728 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4729 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4730 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4731 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4732 * PERFORMANCE OF THIS SOFTWARE.
4735 * chongo <Landon Curt Noll> /\oo/\
4736 * http://www.isthe.com/chongo/
4738 * Share and Enjoy! :-)
4741 typedef unsigned long long mdb_hash_t;
4742 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4744 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4745 * @param[in] val value to hash
4746 * @param[in] hval initial value for hash
4747 * @return 64 bit hash
4749 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4750 * hval arg on the first call.
4753 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4755 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4756 unsigned char *end = s + val->mv_size;
4758 * FNV-1a hash each octet of the string
4761 /* xor the bottom with the current octet */
4762 hval ^= (mdb_hash_t)*s++;
4764 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4765 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4766 (hval << 7) + (hval << 8) + (hval << 40);
4768 /* return our new hash value */
4772 /** Hash the string and output the encoded hash.
4773 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4774 * very short name limits. We don't care about the encoding being reversible,
4775 * we just want to preserve as many bits of the input as possible in a
4776 * small printable string.
4777 * @param[in] str string to hash
4778 * @param[out] encbuf an array of 11 chars to hold the hash
4780 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4783 mdb_pack85(unsigned long l, char *out)
4787 for (i=0; i<5; i++) {
4788 *out++ = mdb_a85[l % 85];
4794 mdb_hash_enc(MDB_val *val, char *encbuf)
4796 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4798 mdb_pack85(h, encbuf);
4799 mdb_pack85(h>>32, encbuf+5);
4804 /** Open and/or initialize the lock region for the environment.
4805 * @param[in] env The LMDB environment.
4806 * @param[in] lpath The pathname of the file used for the lock region.
4807 * @param[in] mode The Unix permissions for the file, if we create it.
4808 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4809 * @return 0 on success, non-zero on failure.
4812 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4815 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4817 # define MDB_ERRCODE_ROFS EROFS
4818 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4819 # define MDB_CLOEXEC O_CLOEXEC
4822 # define MDB_CLOEXEC 0
4825 #ifdef MDB_USE_SYSV_SEM
4834 rc = utf8_to_utf16(lpath, -1, &wlpath, NULL);
4837 env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
4838 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4839 FILE_ATTRIBUTE_NORMAL, NULL);
4842 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4844 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4846 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4851 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4852 /* Lose record locks when exec*() */
4853 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4854 fcntl(env->me_lfd, F_SETFD, fdflags);
4857 if (!(env->me_flags & MDB_NOTLS)) {
4858 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4861 env->me_flags |= MDB_ENV_TXKEY;
4863 /* Windows TLS callbacks need help finding their TLS info. */
4864 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4868 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4872 /* Try to get exclusive lock. If we succeed, then
4873 * nobody is using the lock region and we should initialize it.
4875 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4878 size = GetFileSize(env->me_lfd, NULL);
4880 size = lseek(env->me_lfd, 0, SEEK_END);
4881 if (size == -1) goto fail_errno;
4883 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4884 if (size < rsize && *excl > 0) {
4886 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4887 || !SetEndOfFile(env->me_lfd))
4890 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4894 size = rsize - sizeof(MDB_txninfo);
4895 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4900 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4902 if (!mh) goto fail_errno;
4903 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4905 if (!env->me_txns) goto fail_errno;
4907 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4909 if (m == MAP_FAILED) goto fail_errno;
4915 BY_HANDLE_FILE_INFORMATION stbuf;
4924 if (!mdb_sec_inited) {
4925 InitializeSecurityDescriptor(&mdb_null_sd,
4926 SECURITY_DESCRIPTOR_REVISION);
4927 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4928 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4929 mdb_all_sa.bInheritHandle = FALSE;
4930 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4933 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4934 idbuf.volume = stbuf.dwVolumeSerialNumber;
4935 idbuf.nhigh = stbuf.nFileIndexHigh;
4936 idbuf.nlow = stbuf.nFileIndexLow;
4937 val.mv_data = &idbuf;
4938 val.mv_size = sizeof(idbuf);
4939 mdb_hash_enc(&val, encbuf);
4940 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4941 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4942 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4943 if (!env->me_rmutex) goto fail_errno;
4944 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4945 if (!env->me_wmutex) goto fail_errno;
4946 #elif defined(MDB_USE_POSIX_SEM)
4955 #if defined(__NetBSD__)
4956 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4958 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4959 idbuf.dev = stbuf.st_dev;
4960 idbuf.ino = stbuf.st_ino;
4961 val.mv_data = &idbuf;
4962 val.mv_size = sizeof(idbuf);
4963 mdb_hash_enc(&val, encbuf);
4964 #ifdef MDB_SHORT_SEMNAMES
4965 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4967 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4968 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4969 /* Clean up after a previous run, if needed: Try to
4970 * remove both semaphores before doing anything else.
4972 sem_unlink(env->me_txns->mti_rmname);
4973 sem_unlink(env->me_txns->mti_wmname);
4974 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4975 O_CREAT|O_EXCL, mode, 1);
4976 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4977 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4978 O_CREAT|O_EXCL, mode, 1);
4979 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4980 #elif defined(MDB_USE_SYSV_SEM)
4981 unsigned short vals[2] = {1, 1};
4982 key_t key = ftok(lpath, 'M');
4985 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4989 if (semctl(semid, 0, SETALL, semu) < 0)
4991 env->me_txns->mti_semid = semid;
4992 env->me_txns->mti_rlocked = 0;
4993 env->me_txns->mti_wlocked = 0;
4994 #else /* MDB_USE_POSIX_MUTEX: */
4995 pthread_mutexattr_t mattr;
4997 /* Solaris needs this before initing a robust mutex. Otherwise
4998 * it may skip the init and return EBUSY "seems someone already
4999 * inited" or EINVAL "it was inited differently".
5001 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
5002 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
5004 if ((rc = pthread_mutexattr_init(&mattr)) != 0)
5006 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
5007 #ifdef MDB_ROBUST_SUPPORTED
5008 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
5010 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
5011 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
5012 pthread_mutexattr_destroy(&mattr);
5015 #endif /* _WIN32 || ... */
5017 env->me_txns->mti_magic = MDB_MAGIC;
5018 env->me_txns->mti_format = MDB_LOCK_FORMAT;
5019 env->me_txns->mti_txnid = 0;
5020 env->me_txns->mti_numreaders = 0;
5023 #ifdef MDB_USE_SYSV_SEM
5024 struct semid_ds buf;
5026 if (env->me_txns->mti_magic != MDB_MAGIC) {
5027 DPUTS("lock region has invalid magic");
5031 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
5032 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
5033 env->me_txns->mti_format, MDB_LOCK_FORMAT));
5034 rc = MDB_VERSION_MISMATCH;
5038 if (rc && rc != EACCES && rc != EAGAIN) {
5042 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
5043 if (!env->me_rmutex) goto fail_errno;
5044 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
5045 if (!env->me_wmutex) goto fail_errno;
5046 #elif defined(MDB_USE_POSIX_SEM)
5047 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
5048 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
5049 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
5050 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
5051 #elif defined(MDB_USE_SYSV_SEM)
5052 semid = env->me_txns->mti_semid;
5054 /* check for read access */
5055 if (semctl(semid, 0, IPC_STAT, semu) < 0)
5057 /* check for write access */
5058 if (semctl(semid, 0, IPC_SET, semu) < 0)
5062 #ifdef MDB_USE_SYSV_SEM
5063 env->me_rmutex->semid = semid;
5064 env->me_wmutex->semid = semid;
5065 env->me_rmutex->semnum = 0;
5066 env->me_wmutex->semnum = 1;
5067 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
5068 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
5072 env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
5074 pthread_mutex_init(&env->me_rpmutex, NULL);
5086 /** The name of the lock file in the DB environment */
5087 #define LOCKNAME "/lock.mdb"
5088 /** The name of the data file in the DB environment */
5089 #define DATANAME "/data.mdb"
5090 /** The suffix of the lock file when no subdir is used */
5091 #define LOCKSUFF "-lock"
5092 /** Only a subset of the @ref mdb_env flags can be changed
5093 * at runtime. Changing other flags requires closing the
5094 * environment and re-opening it with the new flags.
5096 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
5097 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
5098 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
5100 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
5101 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
5105 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
5107 int oflags, rc, len, excl = -1;
5108 char *lpath, *dpath;
5113 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
5117 if (flags & MDB_WRITEMAP) {
5118 /* silently ignore WRITEMAP in 32 bit mode */
5119 flags ^= MDB_WRITEMAP;
5121 if (flags & MDB_FIXEDMAP) {
5122 /* cannot support FIXEDMAP */
5128 if (flags & MDB_NOSUBDIR) {
5129 rc = len + sizeof(LOCKSUFF) + len + 1;
5131 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
5136 if (flags & MDB_NOSUBDIR) {
5137 dpath = lpath + len + sizeof(LOCKSUFF);
5138 sprintf(lpath, "%s" LOCKSUFF, path);
5139 strcpy(dpath, path);
5141 dpath = lpath + len + sizeof(LOCKNAME);
5142 sprintf(lpath, "%s" LOCKNAME, path);
5143 sprintf(dpath, "%s" DATANAME, path);
5147 flags |= env->me_flags;
5148 if (flags & MDB_RDONLY) {
5149 /* silently ignore WRITEMAP when we're only getting read access */
5150 flags &= ~MDB_WRITEMAP;
5152 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
5153 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
5158 env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
5159 if (!env->me_rpages) {
5163 env->me_rpages[0].mid = 0;
5164 env->me_rpcheck = MDB_ERPAGE_SIZE/2;
5167 env->me_flags = flags |= MDB_ENV_ACTIVE;
5171 env->me_path = strdup(path);
5172 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
5173 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
5174 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
5175 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
5179 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
5181 /* For RDONLY, get lockfile after we know datafile exists */
5182 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
5183 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
5189 if (F_ISSET(flags, MDB_RDONLY)) {
5190 oflags = GENERIC_READ;
5191 len = OPEN_EXISTING;
5193 oflags = GENERIC_READ|GENERIC_WRITE;
5196 mode = FILE_ATTRIBUTE_NORMAL;
5197 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
5200 env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
5201 NULL, len, mode, NULL);
5204 if (F_ISSET(flags, MDB_RDONLY))
5207 oflags = O_RDWR | O_CREAT;
5209 env->me_fd = open(dpath, oflags, mode);
5211 if (env->me_fd == INVALID_HANDLE_VALUE) {
5216 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
5217 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
5222 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
5223 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
5224 env->me_mfd = env->me_fd;
5226 /* Synchronous fd for meta writes. Needed even with
5227 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
5230 len = OPEN_EXISTING;
5231 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
5234 env->me_mfd = CreateFileW(wpath, oflags,
5235 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
5236 mode | FILE_FLAG_WRITE_THROUGH, NULL);
5240 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
5242 if (env->me_mfd == INVALID_HANDLE_VALUE) {
5247 DPRINTF(("opened dbenv %p", (void *) env));
5249 rc = mdb_env_share_locks(env, &excl);
5253 if (!(flags & MDB_RDONLY)) {
5255 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
5256 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
5257 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
5258 (txn = calloc(1, size)))
5260 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
5261 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
5262 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
5263 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
5266 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
5267 if (!txn->mt_rpages) {
5272 txn->mt_rpages[0].mid = 0;
5273 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
5275 txn->mt_dbxs = env->me_dbxs;
5276 txn->mt_flags = MDB_TXN_FINISHED;
5286 mdb_env_close0(env, excl);
5292 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5294 mdb_env_close0(MDB_env *env, int excl)
5298 if (!(env->me_flags & MDB_ENV_ACTIVE))
5301 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5303 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5304 free(env->me_dbxs[i].md_name.mv_data);
5309 free(env->me_dbiseqs);
5310 free(env->me_dbflags);
5312 free(env->me_dirty_list);
5314 if (env->me_txn0 && env->me_txn0->mt_rpages)
5315 free(env->me_txn0->mt_rpages);
5317 for (x=1; x<=env->me_rpages[0].mid; x++)
5318 munmap(env->me_rpages[x].mptr, env->me_rpages[x].mcnt * env->me_psize);
5320 free(env->me_rpages);
5323 mdb_midl_free(env->me_free_pgs);
5325 if (env->me_flags & MDB_ENV_TXKEY) {
5326 pthread_key_delete(env->me_txkey);
5328 /* Delete our key from the global list */
5329 for (i=0; i<mdb_tls_nkeys; i++)
5330 if (mdb_tls_keys[i] == env->me_txkey) {
5331 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5340 munmap(env->me_map, NUM_METAS*env->me_psize);
5342 munmap(env->me_map, env->me_mapsize);
5345 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
5346 (void) close(env->me_mfd);
5347 if (env->me_fd != INVALID_HANDLE_VALUE)
5348 (void) close(env->me_fd);
5350 MDB_PID_T pid = env->me_pid;
5351 /* Clearing readers is done in this function because
5352 * me_txkey with its destructor must be disabled first.
5354 * We skip the the reader mutex, so we touch only
5355 * data owned by this process (me_close_readers and
5356 * our readers), and clear each reader atomically.
5358 for (i = env->me_close_readers; --i >= 0; )
5359 if (env->me_txns->mti_readers[i].mr_pid == pid)
5360 env->me_txns->mti_readers[i].mr_pid = 0;
5362 if (env->me_rmutex) {
5363 CloseHandle(env->me_rmutex);
5364 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5366 /* Windows automatically destroys the mutexes when
5367 * the last handle closes.
5369 #elif defined(MDB_USE_POSIX_SEM)
5370 if (env->me_rmutex != SEM_FAILED) {
5371 sem_close(env->me_rmutex);
5372 if (env->me_wmutex != SEM_FAILED)
5373 sem_close(env->me_wmutex);
5374 /* If we have the filelock: If we are the
5375 * only remaining user, clean up semaphores.
5378 mdb_env_excl_lock(env, &excl);
5380 sem_unlink(env->me_txns->mti_rmname);
5381 sem_unlink(env->me_txns->mti_wmname);
5384 #elif defined(MDB_USE_SYSV_SEM)
5385 if (env->me_rmutex->semid != -1) {
5386 /* If we have the filelock: If we are the
5387 * only remaining user, clean up semaphores.
5390 mdb_env_excl_lock(env, &excl);
5392 semctl(env->me_rmutex->semid, 0, IPC_RMID);
5395 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5397 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5400 /* Unlock the lockfile. Windows would have unlocked it
5401 * after closing anyway, but not necessarily at once.
5403 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5406 (void) close(env->me_lfd);
5410 if (env->me_fmh) CloseHandle(env->me_fmh);
5411 if (env->me_rpmutex) CloseHandle(env->me_rpmutex);
5413 pthread_mutex_destroy(&env->me_rpmutex);
5417 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5421 mdb_env_close(MDB_env *env)
5428 VGMEMP_DESTROY(env);
5429 while ((dp = env->me_dpages) != NULL) {
5430 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5431 env->me_dpages = dp->mp_next;
5435 mdb_env_close0(env, 0);
5439 /** Compare two items pointing at aligned #mdb_size_t's */
5441 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5443 return (*(mdb_size_t *)a->mv_data < *(mdb_size_t *)b->mv_data) ? -1 :
5444 *(mdb_size_t *)a->mv_data > *(mdb_size_t *)b->mv_data;
5447 /** Compare two items pointing at aligned unsigned int's.
5449 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5450 * but #mdb_cmp_clong() is called instead if the data type is #mdb_size_t.
5453 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5455 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5456 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5459 /** Compare two items pointing at unsigned ints of unknown alignment.
5460 * Nodes and keys are guaranteed to be 2-byte aligned.
5463 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5465 #if BYTE_ORDER == LITTLE_ENDIAN
5466 unsigned short *u, *c;
5469 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5470 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5473 } while(!x && u > (unsigned short *)a->mv_data);
5476 unsigned short *u, *c, *end;
5479 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5480 u = (unsigned short *)a->mv_data;
5481 c = (unsigned short *)b->mv_data;
5484 } while(!x && u < end);
5489 /** Compare two items lexically */
5491 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5498 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5504 diff = memcmp(a->mv_data, b->mv_data, len);
5505 return diff ? diff : len_diff<0 ? -1 : len_diff;
5508 /** Compare two items in reverse byte order */
5510 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5512 const unsigned char *p1, *p2, *p1_lim;
5516 p1_lim = (const unsigned char *)a->mv_data;
5517 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5518 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5520 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5526 while (p1 > p1_lim) {
5527 diff = *--p1 - *--p2;
5531 return len_diff<0 ? -1 : len_diff;
5534 /** Search for key within a page, using binary search.
5535 * Returns the smallest entry larger or equal to the key.
5536 * If exactp is non-null, stores whether the found entry was an exact match
5537 * in *exactp (1 or 0).
5538 * Updates the cursor index with the index of the found entry.
5539 * If no entry larger or equal to the key is found, returns NULL.
5542 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5544 unsigned int i = 0, nkeys;
5547 MDB_page *mp = mc->mc_pg[mc->mc_top];
5548 MDB_node *node = NULL;
5553 nkeys = NUMKEYS(mp);
5555 DPRINTF(("searching %u keys in %s %spage %"Yu,
5556 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5559 low = IS_LEAF(mp) ? 0 : 1;
5561 cmp = mc->mc_dbx->md_cmp;
5563 /* Branch pages have no data, so if using integer keys,
5564 * alignment is guaranteed. Use faster mdb_cmp_int.
5566 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5567 if (NODEPTR(mp, 1)->mn_ksize == sizeof(mdb_size_t))
5574 nodekey.mv_size = mc->mc_db->md_pad;
5575 node = NODEPTR(mp, 0); /* fake */
5576 while (low <= high) {
5577 i = (low + high) >> 1;
5578 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5579 rc = cmp(key, &nodekey);
5580 DPRINTF(("found leaf index %u [%s], rc = %i",
5581 i, DKEY(&nodekey), rc));
5590 while (low <= high) {
5591 i = (low + high) >> 1;
5593 node = NODEPTR(mp, i);
5594 nodekey.mv_size = NODEKSZ(node);
5595 nodekey.mv_data = NODEKEY(node);
5597 rc = cmp(key, &nodekey);
5600 DPRINTF(("found leaf index %u [%s], rc = %i",
5601 i, DKEY(&nodekey), rc));
5603 DPRINTF(("found branch index %u [%s -> %"Yu"], rc = %i",
5604 i, DKEY(&nodekey), NODEPGNO(node), rc));
5615 if (rc > 0) { /* Found entry is less than the key. */
5616 i++; /* Skip to get the smallest entry larger than key. */
5618 node = NODEPTR(mp, i);
5621 *exactp = (rc == 0 && nkeys > 0);
5622 /* store the key index */
5623 mc->mc_ki[mc->mc_top] = i;
5625 /* There is no entry larger or equal to the key. */
5628 /* nodeptr is fake for LEAF2 */
5634 mdb_cursor_adjust(MDB_cursor *mc, func)
5638 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5639 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5646 /** Pop a page off the top of the cursor's stack. */
5648 mdb_cursor_pop(MDB_cursor *mc)
5651 DPRINTF(("popping page %"Yu" off db %d cursor %p",
5652 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5658 mc->mc_flags &= ~C_INITIALIZED;
5663 /** Push a page onto the top of the cursor's stack. */
5665 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5667 DPRINTF(("pushing page %"Yu" on db %d cursor %p", mp->mp_pgno,
5668 DDBI(mc), (void *) mc));
5670 if (mc->mc_snum >= CURSOR_STACK) {
5671 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5672 return MDB_CURSOR_FULL;
5675 mc->mc_top = mc->mc_snum++;
5676 mc->mc_pg[mc->mc_top] = mp;
5677 mc->mc_ki[mc->mc_top] = 0;
5683 /** Map a read-only page.
5684 * There are two levels of tracking in use, a per-txn list and a per-env list.
5685 * ref'ing and unref'ing the per-txn list is faster since it requires no
5686 * locking. Pages are cached in the per-env list for global reuse, and a lock
5687 * is required. Pages are not immediately unmapped when their refcnt goes to
5688 * zero; they hang around in case they will be reused again soon.
5690 * When the per-txn list gets full, all pages with refcnt=0 are purged from the
5691 * list and their refcnts in the per-env list are decremented.
5693 * When the per-env list gets full, all pages with refcnt=0 are purged from the
5694 * list and their pages are unmapped.
5696 * @note "full" means the list has reached its respective rpcheck threshold.
5697 * This threshold slowly raises if no pages could be purged on a given check,
5698 * and returns to its original value when enough pages were purged.
5700 * If purging doesn't free any slots, filling the per-txn list will return
5701 * MDB_TXN_FULL, and filling the per-env list returns MDB_MAP_FULL.
5703 * Reference tracking in a txn is imperfect, pages can linger with non-zero
5704 * refcnt even without active references. It was deemed to be too invasive
5705 * to add unrefs in every required location. However, all pages are unref'd
5706 * at the end of the transaction. This guarantees that no stale references
5707 * linger in the per-env list.
5709 * Usually we map chunks of 16 pages at a time, but if an overflow page begins
5710 * at the tail of the chunk we extend the chunk to include the entire overflow
5711 * page. Unfortunately, pages can be turned into overflow pages after their
5712 * chunk was already mapped. In that case we must remap the chunk if the
5713 * overflow page is referenced. If the chunk's refcnt is 0 we can just remap
5714 * it, otherwise we temporarily map a new chunk just for the overflow page.
5716 * @note this chunk handling means we cannot guarantee that a data item
5717 * returned from the DB will stay alive for the duration of the transaction:
5718 * We unref pages as soon as a cursor moves away from the page
5719 * A subsequent op may cause a purge, which may unmap any unref'd chunks
5720 * The caller must copy the data if it must be used later in the same txn.
5722 * Also - our reference counting revolves around cursors, but overflow pages
5723 * aren't pointed to by a cursor's page stack. We have to remember them
5724 * explicitly, in the added mc_ovpg field. A single cursor can only hold a
5725 * reference to one overflow page at a time.
5727 * @param[in] txn the transaction for this access.
5728 * @param[in] pgno the page number for the page to retrieve.
5729 * @param[out] ret address of a pointer where the page's address will be stored.
5730 * @return 0 on success, non-zero on failure.
5733 mdb_rpage_get(MDB_txn *txn, pgno_t pg0, MDB_page **ret)
5735 MDB_env *env = txn->mt_env;
5737 MDB_ID3L tl = txn->mt_rpages;
5738 MDB_ID3L el = env->me_rpages;
5742 int rc, retries = 1;
5746 #define SET_OFF(off,val) off.QuadPart = val
5747 #define MAP(rc,env,addr,len,off) \
5749 rc = NtMapViewOfSection(env->me_fmh, GetCurrentProcess(), &addr, 0, \
5750 len, &off, &len, ViewUnmap, (env->me_flags & MDB_RDONLY) ? 0 : MEM_RESERVE, PAGE_READONLY); \
5751 if (rc) rc = mdb_nt2win32(rc)
5755 #define SET_OFF(off,val) off = val
5756 #define MAP(rc,env,addr,len,off) \
5757 addr = mmap(NULL, len, PROT_READ, MAP_SHARED, env->me_fd, off); \
5758 rc = (addr == MAP_FAILED) ? errno : 0
5761 /* remember the offset of the actual page number, so we can
5762 * return the correct pointer at the end.
5764 rem = pg0 & (MDB_RPAGE_CHUNK-1);
5768 x = mdb_mid3l_search(tl, pgno);
5769 if (x <= tl[0].mid && tl[x].mid == pgno) {
5770 if (x != tl[0].mid && tl[x+1].mid == pg0)
5772 /* check for overflow size */
5773 p = (MDB_page *)((char *)tl[x].mptr + rem * env->me_psize);
5774 if (IS_OVERFLOW(p) && p->mp_pages + rem > tl[x].mcnt) {
5775 id3.mcnt = p->mp_pages + rem;
5776 len = id3.mcnt * env->me_psize;
5777 SET_OFF(off, pgno * env->me_psize);
5778 MAP(rc, env, id3.mptr, len, off);
5781 /* check for local-only page */
5783 mdb_tassert(txn, tl[x].mid != pg0);
5784 /* hope there's room to insert this locally.
5785 * setting mid here tells later code to just insert
5786 * this id3 instead of searching for a match.
5791 /* ignore the mapping we got from env, use new one */
5792 tl[x].mptr = id3.mptr;
5793 tl[x].mcnt = id3.mcnt;
5794 /* if no active ref, see if we can replace in env */
5797 pthread_mutex_lock(&env->me_rpmutex);
5798 i = mdb_mid3l_search(el, tl[x].mid);
5799 if (el[i].mref == 1) {
5800 /* just us, replace it */
5801 munmap(el[i].mptr, el[i].mcnt * env->me_psize);
5802 el[i].mptr = tl[x].mptr;
5803 el[i].mcnt = tl[x].mcnt;
5805 /* there are others, remove ourself */
5808 pthread_mutex_unlock(&env->me_rpmutex);
5812 id3.mptr = tl[x].mptr;
5813 id3.mcnt = tl[x].mcnt;
5819 if (tl[0].mid >= MDB_TRPAGE_MAX - txn->mt_rpcheck) {
5821 /* purge unref'd pages from our list and unref in env */
5822 pthread_mutex_lock(&env->me_rpmutex);
5825 for (i=1; i<=tl[0].mid; i++) {
5828 /* tmp overflow pages don't go to env */
5829 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
5830 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
5833 x = mdb_mid3l_search(el, tl[i].mid);
5837 pthread_mutex_unlock(&env->me_rpmutex);
5839 /* we didn't find any unref'd chunks.
5840 * if we're out of room, fail.
5842 if (tl[0].mid >= MDB_TRPAGE_MAX)
5843 return MDB_TXN_FULL;
5844 /* otherwise, raise threshold for next time around
5847 txn->mt_rpcheck /= 2;
5849 /* we found some unused; consolidate the list */
5850 for (i=y+1; i<= tl[0].mid; i++)
5854 /* decrease the check threshold toward its original value */
5855 if (!txn->mt_rpcheck)
5856 txn->mt_rpcheck = 1;
5857 while (txn->mt_rpcheck < tl[0].mid && txn->mt_rpcheck < MDB_TRPAGE_SIZE/2)
5858 txn->mt_rpcheck *= 2;
5861 if (tl[0].mid < MDB_TRPAGE_SIZE) {
5865 /* don't map past last written page in read-only envs */
5866 if ((env->me_flags & MDB_RDONLY) && pgno + MDB_RPAGE_CHUNK-1 > txn->mt_last_pgno)
5867 id3.mcnt = txn->mt_last_pgno + 1 - pgno;
5869 id3.mcnt = MDB_RPAGE_CHUNK;
5870 len = id3.mcnt * env->me_psize;
5873 /* search for page in env */
5874 pthread_mutex_lock(&env->me_rpmutex);
5875 x = mdb_mid3l_search(el, pgno);
5876 if (x <= el[0].mid && el[x].mid == pgno) {
5877 id3.mptr = el[x].mptr;
5878 id3.mcnt = el[x].mcnt;
5879 /* check for overflow size */
5880 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5881 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
5882 id3.mcnt = p->mp_pages + rem;
5883 len = id3.mcnt * env->me_psize;
5884 SET_OFF(off, pgno * env->me_psize);
5885 MAP(rc, env, id3.mptr, len, off);
5889 munmap(el[x].mptr, env->me_psize * el[x].mcnt);
5890 el[x].mptr = id3.mptr;
5891 el[x].mcnt = id3.mcnt;
5894 pthread_mutex_unlock(&env->me_rpmutex);
5899 pthread_mutex_unlock(&env->me_rpmutex);
5902 if (el[0].mid >= MDB_ERPAGE_MAX - env->me_rpcheck) {
5903 /* purge unref'd pages */
5905 for (i=1; i<=el[0].mid; i++) {
5908 munmap(el[i].mptr, env->me_psize * el[i].mcnt);
5913 /* see if we can unref some local pages */
5918 if (el[0].mid >= MDB_ERPAGE_MAX) {
5919 pthread_mutex_unlock(&env->me_rpmutex);
5920 return MDB_MAP_FULL;
5922 env->me_rpcheck /= 2;
5924 for (i=y+1; i<= el[0].mid; i++)
5928 if (!env->me_rpcheck)
5929 env->me_rpcheck = 1;
5930 while (env->me_rpcheck < el[0].mid && env->me_rpcheck < MDB_ERPAGE_SIZE/2)
5931 env->me_rpcheck *= 2;
5934 SET_OFF(off, pgno * env->me_psize);
5935 MAP(rc, env, id3.mptr, len, off);
5938 pthread_mutex_unlock(&env->me_rpmutex);
5941 /* check for overflow size */
5942 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5943 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
5944 id3.mcnt = p->mp_pages + rem;
5945 munmap(id3.mptr, len);
5946 len = id3.mcnt * env->me_psize;
5947 MAP(rc, env, id3.mptr, len, off);
5951 mdb_mid3l_insert(el, &id3);
5952 pthread_mutex_unlock(&env->me_rpmutex);
5954 mdb_mid3l_insert(tl, &id3);
5956 return MDB_TXN_FULL;
5959 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5960 #if MDB_DEBUG /* we don't need this check any more */
5961 if (IS_OVERFLOW(p)) {
5962 mdb_tassert(txn, p->mp_pages + rem <= id3.mcnt);
5970 /** Find the address of the page corresponding to a given page number.
5971 * @param[in] mc the cursor accessing the page.
5972 * @param[in] pgno the page number for the page to retrieve.
5973 * @param[out] ret address of a pointer where the page's address will be stored.
5974 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5975 * @return 0 on success, non-zero on failure.
5978 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
5980 MDB_txn *txn = mc->mc_txn;
5984 if (! (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP))) {
5988 MDB_ID2L dl = tx2->mt_u.dirty_list;
5990 /* Spilled pages were dirtied in this txn and flushed
5991 * because the dirty list got full. Bring this page
5992 * back in from the map (but don't unspill it here,
5993 * leave that unless page_touch happens again).
5995 if (tx2->mt_spill_pgs) {
5996 MDB_ID pn = pgno << 1;
5997 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5998 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
6003 unsigned x = mdb_mid2l_search(dl, pgno);
6004 if (x <= dl[0].mid && dl[x].mid == pgno) {
6010 } while ((tx2 = tx2->mt_parent) != NULL);
6013 if (pgno >= txn->mt_next_pgno) {
6014 DPRINTF(("page %"Yu" not found", pgno));
6015 txn->mt_flags |= MDB_TXN_ERROR;
6016 return MDB_PAGE_NOTFOUND;
6024 int rc = mdb_rpage_get(txn, pgno, &p);
6028 MDB_env *env = txn->mt_env;
6029 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
6040 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
6041 * The cursor is at the root page, set up the rest of it.
6044 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
6046 MDB_page *mp = mc->mc_pg[mc->mc_top];
6050 while (IS_BRANCH(mp)) {
6054 DPRINTF(("branch page %"Yu" has %u keys", mp->mp_pgno, NUMKEYS(mp)));
6055 /* Don't assert on branch pages in the FreeDB. We can get here
6056 * while in the process of rebalancing a FreeDB branch page; we must
6057 * let that proceed. ITS#8336
6059 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
6060 DPRINTF(("found index 0 to page %"Yu, NODEPGNO(NODEPTR(mp, 0))));
6062 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
6064 if (flags & MDB_PS_LAST)
6065 i = NUMKEYS(mp) - 1;
6068 node = mdb_node_search(mc, key, &exact);
6070 i = NUMKEYS(mp) - 1;
6072 i = mc->mc_ki[mc->mc_top];
6074 mdb_cassert(mc, i > 0);
6078 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
6081 mdb_cassert(mc, i < NUMKEYS(mp));
6082 node = NODEPTR(mp, i);
6084 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6087 mc->mc_ki[mc->mc_top] = i;
6088 if ((rc = mdb_cursor_push(mc, mp)))
6091 if (flags & MDB_PS_MODIFY) {
6092 if ((rc = mdb_page_touch(mc)) != 0)
6094 mp = mc->mc_pg[mc->mc_top];
6099 DPRINTF(("internal error, index points to a %02X page!?",
6101 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6102 return MDB_CORRUPTED;
6105 DPRINTF(("found leaf page %"Yu" for key [%s]", mp->mp_pgno,
6106 key ? DKEY(key) : "null"));
6107 mc->mc_flags |= C_INITIALIZED;
6108 mc->mc_flags &= ~C_EOF;
6113 /** Search for the lowest key under the current branch page.
6114 * This just bypasses a NUMKEYS check in the current page
6115 * before calling mdb_page_search_root(), because the callers
6116 * are all in situations where the current page is known to
6120 mdb_page_search_lowest(MDB_cursor *mc)
6122 MDB_page *mp = mc->mc_pg[mc->mc_top];
6123 MDB_node *node = NODEPTR(mp, 0);
6126 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6129 mc->mc_ki[mc->mc_top] = 0;
6130 if ((rc = mdb_cursor_push(mc, mp)))
6132 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
6135 /** Search for the page a given key should be in.
6136 * Push it and its parent pages on the cursor stack.
6137 * @param[in,out] mc the cursor for this operation.
6138 * @param[in] key the key to search for, or NULL for first/last page.
6139 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
6140 * are touched (updated with new page numbers).
6141 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
6142 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
6143 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
6144 * @return 0 on success, non-zero on failure.
6147 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
6152 /* Make sure the txn is still viable, then find the root from
6153 * the txn's db table and set it as the root of the cursor's stack.
6155 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
6156 DPUTS("transaction may not be used now");
6159 /* Make sure we're using an up-to-date root */
6160 if (*mc->mc_dbflag & DB_STALE) {
6162 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6164 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
6165 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
6172 MDB_node *leaf = mdb_node_search(&mc2,
6173 &mc->mc_dbx->md_name, &exact);
6175 return MDB_NOTFOUND;
6176 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
6177 return MDB_INCOMPATIBLE; /* not a named DB */
6178 rc = mdb_node_read(&mc2, leaf, &data);
6181 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
6183 /* The txn may not know this DBI, or another process may
6184 * have dropped and recreated the DB with other flags.
6186 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
6187 return MDB_INCOMPATIBLE;
6188 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
6190 *mc->mc_dbflag &= ~DB_STALE;
6192 root = mc->mc_db->md_root;
6194 if (root == P_INVALID) { /* Tree is empty. */
6195 DPUTS("tree is empty");
6196 return MDB_NOTFOUND;
6200 mdb_cassert(mc, root > 1);
6201 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) {
6204 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[0]);
6206 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
6213 for (i=1; i<mc->mc_snum; i++)
6214 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[i]);
6220 DPRINTF(("db %d root page %"Yu" has flags 0x%X",
6221 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
6223 if (flags & MDB_PS_MODIFY) {
6224 if ((rc = mdb_page_touch(mc)))
6228 if (flags & MDB_PS_ROOTONLY)
6231 return mdb_page_search_root(mc, key, flags);
6235 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
6237 MDB_txn *txn = mc->mc_txn;
6238 pgno_t pg = mp->mp_pgno;
6239 unsigned x = 0, ovpages = mp->mp_pages;
6240 MDB_env *env = txn->mt_env;
6241 MDB_IDL sl = txn->mt_spill_pgs;
6242 MDB_ID pn = pg << 1;
6245 DPRINTF(("free ov page %"Yu" (%d)", pg, ovpages));
6246 /* If the page is dirty or on the spill list we just acquired it,
6247 * so we should give it back to our current free list, if any.
6248 * Otherwise put it onto the list of pages we freed in this txn.
6250 * Won't create me_pghead: me_pglast must be inited along with it.
6251 * Unsupported in nested txns: They would need to hide the page
6252 * range in ancestor txns' dirty and spilled lists.
6254 if (env->me_pghead &&
6256 ((mp->mp_flags & P_DIRTY) ||
6257 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
6261 MDB_ID2 *dl, ix, iy;
6262 rc = mdb_midl_need(&env->me_pghead, ovpages);
6265 if (!(mp->mp_flags & P_DIRTY)) {
6266 /* This page is no longer spilled */
6273 /* Remove from dirty list */
6274 dl = txn->mt_u.dirty_list;
6276 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
6282 mdb_cassert(mc, x > 1);
6284 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
6285 txn->mt_flags |= MDB_TXN_ERROR;
6289 txn->mt_dirty_room++;
6290 if (!(env->me_flags & MDB_WRITEMAP))
6291 mdb_dpage_free(env, mp);
6293 /* Insert in me_pghead */
6294 mop = env->me_pghead;
6295 j = mop[0] + ovpages;
6296 for (i = mop[0]; i && mop[i] < pg; i--)
6302 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
6306 mc->mc_db->md_overflow_pages -= ovpages;
6310 /** Return the data associated with a given node.
6311 * @param[in] mc The cursor for this operation.
6312 * @param[in] leaf The node being read.
6313 * @param[out] data Updated to point to the node's data.
6314 * @return 0 on success, non-zero on failure.
6317 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
6319 MDB_page *omp; /* overflow page */
6324 MDB_PAGE_UNREF(mc->mc_txn, MC_OVPG(mc));
6325 MC_SET_OVPG(mc, NULL);
6327 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6328 data->mv_size = NODEDSZ(leaf);
6329 data->mv_data = NODEDATA(leaf);
6333 /* Read overflow data.
6335 data->mv_size = NODEDSZ(leaf);
6336 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
6337 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
6338 DPRINTF(("read overflow page %"Yu" failed", pgno));
6341 data->mv_data = METADATA(omp);
6342 MC_SET_OVPG(mc, omp);
6348 mdb_get(MDB_txn *txn, MDB_dbi dbi,
6349 MDB_val *key, MDB_val *data)
6356 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
6358 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
6361 if (txn->mt_flags & MDB_TXN_BLOCKED)
6364 mdb_cursor_init(&mc, txn, dbi, &mx);
6365 rc = mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
6366 /* unref all the pages when MDB_VL32 - caller must copy the data
6367 * before doing anything else
6369 MDB_CURSOR_UNREF(&mc, 1);
6373 /** Find a sibling for a page.
6374 * Replaces the page at the top of the cursor's stack with the
6375 * specified sibling, if one exists.
6376 * @param[in] mc The cursor for this operation.
6377 * @param[in] move_right Non-zero if the right sibling is requested,
6378 * otherwise the left sibling.
6379 * @return 0 on success, non-zero on failure.
6382 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
6391 if (mc->mc_snum < 2) {
6392 return MDB_NOTFOUND; /* root has no siblings */
6396 op = mc->mc_pg[mc->mc_top];
6399 DPRINTF(("parent page is page %"Yu", index %u",
6400 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
6402 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6403 : (mc->mc_ki[mc->mc_top] == 0)) {
6404 DPRINTF(("no more keys left, moving to %s sibling",
6405 move_right ? "right" : "left"));
6406 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
6407 /* undo cursor_pop before returning */
6414 mc->mc_ki[mc->mc_top]++;
6416 mc->mc_ki[mc->mc_top]--;
6417 DPRINTF(("just moving to %s index key %u",
6418 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
6420 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
6422 MDB_PAGE_UNREF(mc->mc_txn, op);
6424 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6425 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
6426 /* mc will be inconsistent if caller does mc_snum++ as above */
6427 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
6431 mdb_cursor_push(mc, mp);
6433 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
6438 /** Move the cursor to the next data item. */
6440 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6446 if ((mc->mc_flags & C_EOF) ||
6447 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
6448 return MDB_NOTFOUND;
6450 if (!(mc->mc_flags & C_INITIALIZED))
6451 return mdb_cursor_first(mc, key, data);
6453 mp = mc->mc_pg[mc->mc_top];
6455 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6456 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6457 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6458 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
6459 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
6460 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
6461 if (rc == MDB_SUCCESS)
6462 MDB_GET_KEY(leaf, key);
6467 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6470 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6471 if (op == MDB_NEXT_DUP)
6472 return MDB_NOTFOUND;
6476 DPRINTF(("cursor_next: top page is %"Yu" in cursor %p",
6477 mdb_dbg_pgno(mp), (void *) mc));
6478 if (mc->mc_flags & C_DEL) {
6479 mc->mc_flags ^= C_DEL;
6483 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
6484 DPUTS("=====> move to next sibling page");
6485 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6486 mc->mc_flags |= C_EOF;
6489 mp = mc->mc_pg[mc->mc_top];
6490 DPRINTF(("next page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6492 mc->mc_ki[mc->mc_top]++;
6495 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6496 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6499 key->mv_size = mc->mc_db->md_pad;
6500 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6504 mdb_cassert(mc, IS_LEAF(mp));
6505 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6507 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6508 mdb_xcursor_init1(mc, leaf);
6511 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6514 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6515 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6516 if (rc != MDB_SUCCESS)
6521 MDB_GET_KEY(leaf, key);
6525 /** Move the cursor to the previous data item. */
6527 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6533 if (!(mc->mc_flags & C_INITIALIZED)) {
6534 rc = mdb_cursor_last(mc, key, data);
6537 mc->mc_ki[mc->mc_top]++;
6540 mp = mc->mc_pg[mc->mc_top];
6542 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6543 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6544 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6545 if (op == MDB_PREV || op == MDB_PREV_DUP) {
6546 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
6547 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
6548 if (rc == MDB_SUCCESS) {
6549 MDB_GET_KEY(leaf, key);
6550 mc->mc_flags &= ~C_EOF;
6556 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6559 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6560 if (op == MDB_PREV_DUP)
6561 return MDB_NOTFOUND;
6565 DPRINTF(("cursor_prev: top page is %"Yu" in cursor %p",
6566 mdb_dbg_pgno(mp), (void *) mc));
6568 mc->mc_flags &= ~(C_EOF|C_DEL);
6570 if (mc->mc_ki[mc->mc_top] == 0) {
6571 DPUTS("=====> move to prev sibling page");
6572 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
6575 mp = mc->mc_pg[mc->mc_top];
6576 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
6577 DPRINTF(("prev page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6579 mc->mc_ki[mc->mc_top]--;
6581 mc->mc_flags &= ~C_EOF;
6583 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6584 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6587 key->mv_size = mc->mc_db->md_pad;
6588 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6592 mdb_cassert(mc, IS_LEAF(mp));
6593 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6595 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6596 mdb_xcursor_init1(mc, leaf);
6599 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6602 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6603 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6604 if (rc != MDB_SUCCESS)
6609 MDB_GET_KEY(leaf, key);
6613 /** Set the cursor on a specific data item. */
6615 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6616 MDB_cursor_op op, int *exactp)
6620 MDB_node *leaf = NULL;
6623 if (key->mv_size == 0)
6624 return MDB_BAD_VALSIZE;
6626 if (mc->mc_xcursor) {
6627 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6628 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6631 /* See if we're already on the right page */
6632 if (mc->mc_flags & C_INITIALIZED) {
6635 mp = mc->mc_pg[mc->mc_top];
6637 mc->mc_ki[mc->mc_top] = 0;
6638 return MDB_NOTFOUND;
6640 if (mp->mp_flags & P_LEAF2) {
6641 nodekey.mv_size = mc->mc_db->md_pad;
6642 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
6644 leaf = NODEPTR(mp, 0);
6645 MDB_GET_KEY2(leaf, nodekey);
6647 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6649 /* Probably happens rarely, but first node on the page
6650 * was the one we wanted.
6652 mc->mc_ki[mc->mc_top] = 0;
6659 unsigned int nkeys = NUMKEYS(mp);
6661 if (mp->mp_flags & P_LEAF2) {
6662 nodekey.mv_data = LEAF2KEY(mp,
6663 nkeys-1, nodekey.mv_size);
6665 leaf = NODEPTR(mp, nkeys-1);
6666 MDB_GET_KEY2(leaf, nodekey);
6668 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6670 /* last node was the one we wanted */
6671 mc->mc_ki[mc->mc_top] = nkeys-1;
6677 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6678 /* This is definitely the right page, skip search_page */
6679 if (mp->mp_flags & P_LEAF2) {
6680 nodekey.mv_data = LEAF2KEY(mp,
6681 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6683 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6684 MDB_GET_KEY2(leaf, nodekey);
6686 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6688 /* current node was the one we wanted */
6695 mc->mc_flags &= ~C_EOF;
6699 /* If any parents have right-sibs, search.
6700 * Otherwise, there's nothing further.
6702 for (i=0; i<mc->mc_top; i++)
6704 NUMKEYS(mc->mc_pg[i])-1)
6706 if (i == mc->mc_top) {
6707 /* There are no other pages */
6708 mc->mc_ki[mc->mc_top] = nkeys;
6709 return MDB_NOTFOUND;
6713 /* There are no other pages */
6714 mc->mc_ki[mc->mc_top] = 0;
6715 if (op == MDB_SET_RANGE && !exactp) {
6719 return MDB_NOTFOUND;
6725 rc = mdb_page_search(mc, key, 0);
6726 if (rc != MDB_SUCCESS)
6729 mp = mc->mc_pg[mc->mc_top];
6730 mdb_cassert(mc, IS_LEAF(mp));
6733 leaf = mdb_node_search(mc, key, exactp);
6734 if (exactp != NULL && !*exactp) {
6735 /* MDB_SET specified and not an exact match. */
6736 return MDB_NOTFOUND;
6740 DPUTS("===> inexact leaf not found, goto sibling");
6741 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6742 mc->mc_flags |= C_EOF;
6743 return rc; /* no entries matched */
6745 mp = mc->mc_pg[mc->mc_top];
6746 mdb_cassert(mc, IS_LEAF(mp));
6747 leaf = NODEPTR(mp, 0);
6751 mc->mc_flags |= C_INITIALIZED;
6752 mc->mc_flags &= ~C_EOF;
6755 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6756 key->mv_size = mc->mc_db->md_pad;
6757 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6762 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6763 mdb_xcursor_init1(mc, leaf);
6766 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6767 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6768 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6771 if (op == MDB_GET_BOTH) {
6777 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6778 if (rc != MDB_SUCCESS)
6781 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6784 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6786 dcmp = mc->mc_dbx->md_dcmp;
6787 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
6788 dcmp = mdb_cmp_clong;
6789 rc = dcmp(data, &olddata);
6791 if (op == MDB_GET_BOTH || rc > 0)
6792 return MDB_NOTFOUND;
6799 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6800 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6805 /* The key already matches in all other cases */
6806 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6807 MDB_GET_KEY(leaf, key);
6808 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6813 /** Move the cursor to the first item in the database. */
6815 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6820 if (mc->mc_xcursor) {
6821 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6822 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6825 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6826 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6827 if (rc != MDB_SUCCESS)
6830 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6832 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6833 mc->mc_flags |= C_INITIALIZED;
6834 mc->mc_flags &= ~C_EOF;
6836 mc->mc_ki[mc->mc_top] = 0;
6838 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6839 key->mv_size = mc->mc_db->md_pad;
6840 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6845 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6846 mdb_xcursor_init1(mc, leaf);
6847 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6851 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6855 MDB_GET_KEY(leaf, key);
6859 /** Move the cursor to the last item in the database. */
6861 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6866 if (mc->mc_xcursor) {
6867 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6868 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6871 if (!(mc->mc_flags & C_EOF)) {
6873 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6874 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6875 if (rc != MDB_SUCCESS)
6878 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6881 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6882 mc->mc_flags |= C_INITIALIZED|C_EOF;
6883 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6885 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6886 key->mv_size = mc->mc_db->md_pad;
6887 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6892 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6893 mdb_xcursor_init1(mc, leaf);
6894 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6898 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6903 MDB_GET_KEY(leaf, key);
6908 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6913 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6918 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6922 case MDB_GET_CURRENT:
6923 if (!(mc->mc_flags & C_INITIALIZED)) {
6926 MDB_page *mp = mc->mc_pg[mc->mc_top];
6927 int nkeys = NUMKEYS(mp);
6928 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6929 mc->mc_ki[mc->mc_top] = nkeys;
6935 key->mv_size = mc->mc_db->md_pad;
6936 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6938 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6939 MDB_GET_KEY(leaf, key);
6941 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6942 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6944 rc = mdb_node_read(mc, leaf, data);
6951 case MDB_GET_BOTH_RANGE:
6956 if (mc->mc_xcursor == NULL) {
6957 rc = MDB_INCOMPATIBLE;
6967 rc = mdb_cursor_set(mc, key, data, op,
6968 op == MDB_SET_RANGE ? NULL : &exact);
6971 case MDB_GET_MULTIPLE:
6972 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6976 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6977 rc = MDB_INCOMPATIBLE;
6981 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6982 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6985 case MDB_NEXT_MULTIPLE:
6990 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6991 rc = MDB_INCOMPATIBLE;
6994 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6995 if (rc == MDB_SUCCESS) {
6996 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6999 mx = &mc->mc_xcursor->mx_cursor;
7000 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
7002 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
7003 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
7009 case MDB_PREV_MULTIPLE:
7014 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7015 rc = MDB_INCOMPATIBLE;
7018 if (!(mc->mc_flags & C_INITIALIZED))
7019 rc = mdb_cursor_last(mc, key, data);
7022 if (rc == MDB_SUCCESS) {
7023 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
7024 if (mx->mc_flags & C_INITIALIZED) {
7025 rc = mdb_cursor_sibling(mx, 0);
7026 if (rc == MDB_SUCCESS)
7035 case MDB_NEXT_NODUP:
7036 rc = mdb_cursor_next(mc, key, data, op);
7040 case MDB_PREV_NODUP:
7041 rc = mdb_cursor_prev(mc, key, data, op);
7044 rc = mdb_cursor_first(mc, key, data);
7047 mfunc = mdb_cursor_first;
7049 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
7053 if (mc->mc_xcursor == NULL) {
7054 rc = MDB_INCOMPATIBLE;
7058 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7059 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7060 MDB_GET_KEY(leaf, key);
7061 rc = mdb_node_read(mc, leaf, data);
7065 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7069 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
7072 rc = mdb_cursor_last(mc, key, data);
7075 mfunc = mdb_cursor_last;
7078 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
7083 if (mc->mc_flags & C_DEL)
7084 mc->mc_flags ^= C_DEL;
7089 /** Touch all the pages in the cursor stack. Set mc_top.
7090 * Makes sure all the pages are writable, before attempting a write operation.
7091 * @param[in] mc The cursor to operate on.
7094 mdb_cursor_touch(MDB_cursor *mc)
7096 int rc = MDB_SUCCESS;
7098 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
7101 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
7103 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
7104 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
7107 *mc->mc_dbflag |= DB_DIRTY;
7112 rc = mdb_page_touch(mc);
7113 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
7114 mc->mc_top = mc->mc_snum-1;
7119 /** Do not spill pages to disk if txn is getting full, may fail instead */
7120 #define MDB_NOSPILL 0x8000
7123 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
7127 MDB_node *leaf = NULL;
7128 MDB_page *fp, *mp, *sub_root = NULL;
7130 MDB_val xdata, *rdata, dkey, olddata;
7132 int do_sub = 0, insert_key, insert_data;
7133 unsigned int mcount = 0, dcount = 0, nospill;
7136 unsigned int nflags;
7139 if (mc == NULL || key == NULL)
7142 env = mc->mc_txn->mt_env;
7144 /* Check this first so counter will always be zero on any
7147 if (flags & MDB_MULTIPLE) {
7148 dcount = data[1].mv_size;
7149 data[1].mv_size = 0;
7150 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
7151 return MDB_INCOMPATIBLE;
7154 nospill = flags & MDB_NOSPILL;
7155 flags &= ~MDB_NOSPILL;
7157 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7158 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7160 if (key->mv_size-1 >= ENV_MAXKEY(env))
7161 return MDB_BAD_VALSIZE;
7163 #if SIZE_MAX > MAXDATASIZE
7164 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
7165 return MDB_BAD_VALSIZE;
7167 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
7168 return MDB_BAD_VALSIZE;
7171 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
7172 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
7176 if (flags == MDB_CURRENT) {
7177 if (!(mc->mc_flags & C_INITIALIZED))
7180 } else if (mc->mc_db->md_root == P_INVALID) {
7181 /* new database, cursor has nothing to point to */
7184 mc->mc_flags &= ~C_INITIALIZED;
7189 if (flags & MDB_APPEND) {
7191 rc = mdb_cursor_last(mc, &k2, &d2);
7193 rc = mc->mc_dbx->md_cmp(key, &k2);
7196 mc->mc_ki[mc->mc_top]++;
7198 /* new key is <= last key */
7203 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
7205 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
7206 DPRINTF(("duplicate key [%s]", DKEY(key)));
7208 return MDB_KEYEXIST;
7210 if (rc && rc != MDB_NOTFOUND)
7214 if (mc->mc_flags & C_DEL)
7215 mc->mc_flags ^= C_DEL;
7217 /* Cursor is positioned, check for room in the dirty list */
7219 if (flags & MDB_MULTIPLE) {
7221 xdata.mv_size = data->mv_size * dcount;
7225 if ((rc2 = mdb_page_spill(mc, key, rdata)))
7229 if (rc == MDB_NO_ROOT) {
7231 /* new database, write a root leaf page */
7232 DPUTS("allocating new root leaf page");
7233 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
7236 mdb_cursor_push(mc, np);
7237 mc->mc_db->md_root = np->mp_pgno;
7238 mc->mc_db->md_depth++;
7239 *mc->mc_dbflag |= DB_DIRTY;
7240 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
7242 np->mp_flags |= P_LEAF2;
7243 mc->mc_flags |= C_INITIALIZED;
7245 /* make sure all cursor pages are writable */
7246 rc2 = mdb_cursor_touch(mc);
7251 insert_key = insert_data = rc;
7253 /* The key does not exist */
7254 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
7255 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
7256 LEAFSIZE(key, data) > env->me_nodemax)
7258 /* Too big for a node, insert in sub-DB. Set up an empty
7259 * "old sub-page" for prep_subDB to expand to a full page.
7261 fp_flags = P_LEAF|P_DIRTY;
7263 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
7264 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
7265 olddata.mv_size = PAGEHDRSZ;
7269 /* there's only a key anyway, so this is a no-op */
7270 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
7272 unsigned int ksize = mc->mc_db->md_pad;
7273 if (key->mv_size != ksize)
7274 return MDB_BAD_VALSIZE;
7275 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
7276 memcpy(ptr, key->mv_data, ksize);
7278 /* if overwriting slot 0 of leaf, need to
7279 * update branch key if there is a parent page
7281 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7282 unsigned short dtop = 1;
7284 /* slot 0 is always an empty key, find real slot */
7285 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7289 if (mc->mc_ki[mc->mc_top])
7290 rc2 = mdb_update_key(mc, key);
7301 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7302 olddata.mv_size = NODEDSZ(leaf);
7303 olddata.mv_data = NODEDATA(leaf);
7306 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
7307 /* Prepare (sub-)page/sub-DB to accept the new item,
7308 * if needed. fp: old sub-page or a header faking
7309 * it. mp: new (sub-)page. offset: growth in page
7310 * size. xdata: node data with new page or DB.
7312 unsigned i, offset = 0;
7313 mp = fp = xdata.mv_data = env->me_pbuf;
7314 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
7316 /* Was a single item before, must convert now */
7317 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7319 /* Just overwrite the current item */
7320 if (flags == MDB_CURRENT)
7322 dcmp = mc->mc_dbx->md_dcmp;
7323 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
7324 dcmp = mdb_cmp_clong;
7325 /* does data match? */
7326 if (!dcmp(data, &olddata)) {
7327 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
7328 return MDB_KEYEXIST;
7333 /* Back up original data item */
7334 dkey.mv_size = olddata.mv_size;
7335 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
7337 /* Make sub-page header for the dup items, with dummy body */
7338 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
7339 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
7340 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
7341 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7342 fp->mp_flags |= P_LEAF2;
7343 fp->mp_pad = data->mv_size;
7344 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
7346 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
7347 (dkey.mv_size & 1) + (data->mv_size & 1);
7349 fp->mp_upper = xdata.mv_size - PAGEBASE;
7350 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
7351 } else if (leaf->mn_flags & F_SUBDATA) {
7352 /* Data is on sub-DB, just store it */
7353 flags |= F_DUPDATA|F_SUBDATA;
7356 /* Data is on sub-page */
7357 fp = olddata.mv_data;
7360 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7361 offset = EVEN(NODESIZE + sizeof(indx_t) +
7365 offset = fp->mp_pad;
7366 if (SIZELEFT(fp) < offset) {
7367 offset *= 4; /* space for 4 more */
7370 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
7372 fp->mp_flags |= P_DIRTY;
7373 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
7374 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
7378 xdata.mv_size = olddata.mv_size + offset;
7381 fp_flags = fp->mp_flags;
7382 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
7383 /* Too big for a sub-page, convert to sub-DB */
7384 fp_flags &= ~P_SUBP;
7386 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7387 fp_flags |= P_LEAF2;
7388 dummy.md_pad = fp->mp_pad;
7389 dummy.md_flags = MDB_DUPFIXED;
7390 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7391 dummy.md_flags |= MDB_INTEGERKEY;
7397 dummy.md_branch_pages = 0;
7398 dummy.md_leaf_pages = 1;
7399 dummy.md_overflow_pages = 0;
7400 dummy.md_entries = NUMKEYS(fp);
7401 xdata.mv_size = sizeof(MDB_db);
7402 xdata.mv_data = &dummy;
7403 if ((rc = mdb_page_alloc(mc, 1, &mp)))
7405 offset = env->me_psize - olddata.mv_size;
7406 flags |= F_DUPDATA|F_SUBDATA;
7407 dummy.md_root = mp->mp_pgno;
7411 mp->mp_flags = fp_flags | P_DIRTY;
7412 mp->mp_pad = fp->mp_pad;
7413 mp->mp_lower = fp->mp_lower;
7414 mp->mp_upper = fp->mp_upper + offset;
7415 if (fp_flags & P_LEAF2) {
7416 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
7418 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
7419 olddata.mv_size - fp->mp_upper - PAGEBASE);
7420 for (i=0; i<NUMKEYS(fp); i++)
7421 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
7429 mdb_node_del(mc, 0);
7433 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
7434 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
7435 return MDB_INCOMPATIBLE;
7436 /* overflow page overwrites need special handling */
7437 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7440 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
7442 memcpy(&pg, olddata.mv_data, sizeof(pg));
7443 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
7445 ovpages = omp->mp_pages;
7447 /* Is the ov page large enough? */
7448 if (ovpages >= dpages) {
7449 if (!(omp->mp_flags & P_DIRTY) &&
7450 (level || (env->me_flags & MDB_WRITEMAP)))
7452 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
7455 level = 0; /* dirty in this txn or clean */
7458 if (omp->mp_flags & P_DIRTY) {
7459 /* yes, overwrite it. Note in this case we don't
7460 * bother to try shrinking the page if the new data
7461 * is smaller than the overflow threshold.
7464 /* It is writable only in a parent txn */
7465 size_t sz = (size_t) env->me_psize * ovpages, off;
7466 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
7472 /* Note - this page is already counted in parent's dirty_room */
7473 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
7474 mdb_cassert(mc, rc2 == 0);
7475 /* Currently we make the page look as with put() in the
7476 * parent txn, in case the user peeks at MDB_RESERVEd
7477 * or unused parts. Some users treat ovpages specially.
7479 if (!(flags & MDB_RESERVE)) {
7480 /* Skip the part where LMDB will put *data.
7481 * Copy end of page, adjusting alignment so
7482 * compiler may copy words instead of bytes.
7484 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
7485 memcpy((size_t *)((char *)np + off),
7486 (size_t *)((char *)omp + off), sz - off);
7489 memcpy(np, omp, sz); /* Copy beginning of page */
7492 SETDSZ(leaf, data->mv_size);
7493 if (F_ISSET(flags, MDB_RESERVE))
7494 data->mv_data = METADATA(omp);
7496 memcpy(METADATA(omp), data->mv_data, data->mv_size);
7500 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
7502 } else if (data->mv_size == olddata.mv_size) {
7503 /* same size, just replace it. Note that we could
7504 * also reuse this node if the new data is smaller,
7505 * but instead we opt to shrink the node in that case.
7507 if (F_ISSET(flags, MDB_RESERVE))
7508 data->mv_data = olddata.mv_data;
7509 else if (!(mc->mc_flags & C_SUB))
7510 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
7512 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
7517 mdb_node_del(mc, 0);
7523 nflags = flags & NODE_ADD_FLAGS;
7524 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
7525 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
7526 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
7527 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
7529 nflags |= MDB_SPLIT_REPLACE;
7530 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
7532 /* There is room already in this leaf page. */
7533 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
7535 /* Adjust other cursors pointing to mp */
7536 MDB_cursor *m2, *m3;
7537 MDB_dbi dbi = mc->mc_dbi;
7538 unsigned i = mc->mc_top;
7539 MDB_page *mp = mc->mc_pg[i];
7541 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7542 if (mc->mc_flags & C_SUB)
7543 m3 = &m2->mc_xcursor->mx_cursor;
7546 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
7547 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
7550 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7551 MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
7552 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7553 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7559 if (rc == MDB_SUCCESS) {
7560 /* Now store the actual data in the child DB. Note that we're
7561 * storing the user data in the keys field, so there are strict
7562 * size limits on dupdata. The actual data fields of the child
7563 * DB are all zero size.
7566 int xflags, new_dupdata;
7571 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7572 if (flags & MDB_CURRENT) {
7573 xflags = MDB_CURRENT|MDB_NOSPILL;
7575 mdb_xcursor_init1(mc, leaf);
7576 xflags = (flags & MDB_NODUPDATA) ?
7577 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
7580 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
7581 new_dupdata = (int)dkey.mv_size;
7582 /* converted, write the original data first */
7584 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
7587 /* we've done our job */
7590 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
7591 /* Adjust other cursors pointing to mp */
7593 MDB_xcursor *mx = mc->mc_xcursor;
7594 unsigned i = mc->mc_top;
7595 MDB_page *mp = mc->mc_pg[i];
7596 int nkeys = NUMKEYS(mp);
7598 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7599 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7600 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7601 if (m2->mc_pg[i] == mp) {
7602 if (m2->mc_ki[i] == mc->mc_ki[i]) {
7603 mdb_xcursor_init2(m2, mx, new_dupdata);
7604 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
7605 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
7606 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7607 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7612 ecount = mc->mc_xcursor->mx_db.md_entries;
7613 if (flags & MDB_APPENDDUP)
7614 xflags |= MDB_APPEND;
7615 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
7616 if (flags & F_SUBDATA) {
7617 void *db = NODEDATA(leaf);
7618 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7620 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
7622 /* Increment count unless we just replaced an existing item. */
7624 mc->mc_db->md_entries++;
7626 /* Invalidate txn if we created an empty sub-DB */
7629 /* If we succeeded and the key didn't exist before,
7630 * make sure the cursor is marked valid.
7632 mc->mc_flags |= C_INITIALIZED;
7634 if (flags & MDB_MULTIPLE) {
7637 /* let caller know how many succeeded, if any */
7638 data[1].mv_size = mcount;
7639 if (mcount < dcount) {
7640 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
7641 insert_key = insert_data = 0;
7648 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
7651 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7656 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7662 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7663 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7665 if (!(mc->mc_flags & C_INITIALIZED))
7668 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7669 return MDB_NOTFOUND;
7671 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7674 rc = mdb_cursor_touch(mc);
7678 mp = mc->mc_pg[mc->mc_top];
7681 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7683 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7684 if (flags & MDB_NODUPDATA) {
7685 /* mdb_cursor_del0() will subtract the final entry */
7686 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7687 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7689 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7690 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7692 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7695 /* If sub-DB still has entries, we're done */
7696 if (mc->mc_xcursor->mx_db.md_entries) {
7697 if (leaf->mn_flags & F_SUBDATA) {
7698 /* update subDB info */
7699 void *db = NODEDATA(leaf);
7700 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7703 /* shrink fake page */
7704 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7705 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7706 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7707 /* fix other sub-DB cursors pointed at fake pages on this page */
7708 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7709 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7710 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7711 if (m2->mc_pg[mc->mc_top] == mp) {
7712 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
7713 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7715 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7716 if (!(n2->mn_flags & F_SUBDATA))
7717 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7722 mc->mc_db->md_entries--;
7725 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7727 /* otherwise fall thru and delete the sub-DB */
7730 if (leaf->mn_flags & F_SUBDATA) {
7731 /* add all the child DB's pages to the free list */
7732 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7737 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7738 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7739 rc = MDB_INCOMPATIBLE;
7743 /* add overflow pages to free list */
7744 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7748 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7749 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7750 (rc = mdb_ovpage_free(mc, omp)))
7755 return mdb_cursor_del0(mc);
7758 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7762 /** Allocate and initialize new pages for a database.
7763 * @param[in] mc a cursor on the database being added to.
7764 * @param[in] flags flags defining what type of page is being allocated.
7765 * @param[in] num the number of pages to allocate. This is usually 1,
7766 * unless allocating overflow pages for a large record.
7767 * @param[out] mp Address of a page, or NULL on failure.
7768 * @return 0 on success, non-zero on failure.
7771 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7776 if ((rc = mdb_page_alloc(mc, num, &np)))
7778 DPRINTF(("allocated new mpage %"Yu", page size %u",
7779 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7780 np->mp_flags = flags | P_DIRTY;
7781 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7782 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7785 mc->mc_db->md_branch_pages++;
7786 else if (IS_LEAF(np))
7787 mc->mc_db->md_leaf_pages++;
7788 else if (IS_OVERFLOW(np)) {
7789 mc->mc_db->md_overflow_pages += num;
7797 /** Calculate the size of a leaf node.
7798 * The size depends on the environment's page size; if a data item
7799 * is too large it will be put onto an overflow page and the node
7800 * size will only include the key and not the data. Sizes are always
7801 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7802 * of the #MDB_node headers.
7803 * @param[in] env The environment handle.
7804 * @param[in] key The key for the node.
7805 * @param[in] data The data for the node.
7806 * @return The number of bytes needed to store the node.
7809 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7813 sz = LEAFSIZE(key, data);
7814 if (sz > env->me_nodemax) {
7815 /* put on overflow page */
7816 sz -= data->mv_size - sizeof(pgno_t);
7819 return EVEN(sz + sizeof(indx_t));
7822 /** Calculate the size of a branch node.
7823 * The size should depend on the environment's page size but since
7824 * we currently don't support spilling large keys onto overflow
7825 * pages, it's simply the size of the #MDB_node header plus the
7826 * size of the key. Sizes are always rounded up to an even number
7827 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7828 * @param[in] env The environment handle.
7829 * @param[in] key The key for the node.
7830 * @return The number of bytes needed to store the node.
7833 mdb_branch_size(MDB_env *env, MDB_val *key)
7838 if (sz > env->me_nodemax) {
7839 /* put on overflow page */
7840 /* not implemented */
7841 /* sz -= key->size - sizeof(pgno_t); */
7844 return sz + sizeof(indx_t);
7847 /** Add a node to the page pointed to by the cursor.
7848 * @param[in] mc The cursor for this operation.
7849 * @param[in] indx The index on the page where the new node should be added.
7850 * @param[in] key The key for the new node.
7851 * @param[in] data The data for the new node, if any.
7852 * @param[in] pgno The page number, if adding a branch node.
7853 * @param[in] flags Flags for the node.
7854 * @return 0 on success, non-zero on failure. Possible errors are:
7856 * <li>ENOMEM - failed to allocate overflow pages for the node.
7857 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7858 * should never happen since all callers already calculate the
7859 * page's free space before calling this function.
7863 mdb_node_add(MDB_cursor *mc, indx_t indx,
7864 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7867 size_t node_size = NODESIZE;
7871 MDB_page *mp = mc->mc_pg[mc->mc_top];
7872 MDB_page *ofp = NULL; /* overflow page */
7876 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7878 DPRINTF(("add to %s %spage %"Yu" index %i, data size %"Z"u key size %"Z"u [%s]",
7879 IS_LEAF(mp) ? "leaf" : "branch",
7880 IS_SUBP(mp) ? "sub-" : "",
7881 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7882 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7885 /* Move higher keys up one slot. */
7886 int ksize = mc->mc_db->md_pad, dif;
7887 char *ptr = LEAF2KEY(mp, indx, ksize);
7888 dif = NUMKEYS(mp) - indx;
7890 memmove(ptr+ksize, ptr, dif*ksize);
7891 /* insert new key */
7892 memcpy(ptr, key->mv_data, ksize);
7894 /* Just using these for counting */
7895 mp->mp_lower += sizeof(indx_t);
7896 mp->mp_upper -= ksize - sizeof(indx_t);
7900 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7902 node_size += key->mv_size;
7904 mdb_cassert(mc, key && data);
7905 if (F_ISSET(flags, F_BIGDATA)) {
7906 /* Data already on overflow page. */
7907 node_size += sizeof(pgno_t);
7908 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7909 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7911 /* Put data on overflow page. */
7912 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7913 data->mv_size, node_size+data->mv_size));
7914 node_size = EVEN(node_size + sizeof(pgno_t));
7915 if ((ssize_t)node_size > room)
7917 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7919 DPRINTF(("allocated overflow page %"Yu, ofp->mp_pgno));
7923 node_size += data->mv_size;
7926 node_size = EVEN(node_size);
7927 if ((ssize_t)node_size > room)
7931 /* Move higher pointers up one slot. */
7932 for (i = NUMKEYS(mp); i > indx; i--)
7933 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7935 /* Adjust free space offsets. */
7936 ofs = mp->mp_upper - node_size;
7937 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7938 mp->mp_ptrs[indx] = ofs;
7940 mp->mp_lower += sizeof(indx_t);
7942 /* Write the node data. */
7943 node = NODEPTR(mp, indx);
7944 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7945 node->mn_flags = flags;
7947 SETDSZ(node,data->mv_size);
7952 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7955 ndata = NODEDATA(node);
7957 if (F_ISSET(flags, F_BIGDATA))
7958 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7959 else if (F_ISSET(flags, MDB_RESERVE))
7960 data->mv_data = ndata;
7962 memcpy(ndata, data->mv_data, data->mv_size);
7964 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7965 ndata = METADATA(ofp);
7966 if (F_ISSET(flags, MDB_RESERVE))
7967 data->mv_data = ndata;
7969 memcpy(ndata, data->mv_data, data->mv_size);
7976 DPRINTF(("not enough room in page %"Yu", got %u ptrs",
7977 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7978 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7979 DPRINTF(("node size = %"Z"u", node_size));
7980 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7981 return MDB_PAGE_FULL;
7984 /** Delete the specified node from a page.
7985 * @param[in] mc Cursor pointing to the node to delete.
7986 * @param[in] ksize The size of a node. Only used if the page is
7987 * part of a #MDB_DUPFIXED database.
7990 mdb_node_del(MDB_cursor *mc, int ksize)
7992 MDB_page *mp = mc->mc_pg[mc->mc_top];
7993 indx_t indx = mc->mc_ki[mc->mc_top];
7995 indx_t i, j, numkeys, ptr;
7999 DPRINTF(("delete node %u on %s page %"Yu, indx,
8000 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
8001 numkeys = NUMKEYS(mp);
8002 mdb_cassert(mc, indx < numkeys);
8005 int x = numkeys - 1 - indx;
8006 base = LEAF2KEY(mp, indx, ksize);
8008 memmove(base, base + ksize, x * ksize);
8009 mp->mp_lower -= sizeof(indx_t);
8010 mp->mp_upper += ksize - sizeof(indx_t);
8014 node = NODEPTR(mp, indx);
8015 sz = NODESIZE + node->mn_ksize;
8017 if (F_ISSET(node->mn_flags, F_BIGDATA))
8018 sz += sizeof(pgno_t);
8020 sz += NODEDSZ(node);
8024 ptr = mp->mp_ptrs[indx];
8025 for (i = j = 0; i < numkeys; i++) {
8027 mp->mp_ptrs[j] = mp->mp_ptrs[i];
8028 if (mp->mp_ptrs[i] < ptr)
8029 mp->mp_ptrs[j] += sz;
8034 base = (char *)mp + mp->mp_upper + PAGEBASE;
8035 memmove(base + sz, base, ptr - mp->mp_upper);
8037 mp->mp_lower -= sizeof(indx_t);
8041 /** Compact the main page after deleting a node on a subpage.
8042 * @param[in] mp The main page to operate on.
8043 * @param[in] indx The index of the subpage on the main page.
8046 mdb_node_shrink(MDB_page *mp, indx_t indx)
8051 indx_t delta, nsize, len, ptr;
8054 node = NODEPTR(mp, indx);
8055 sp = (MDB_page *)NODEDATA(node);
8056 delta = SIZELEFT(sp);
8057 nsize = NODEDSZ(node) - delta;
8059 /* Prepare to shift upward, set len = length(subpage part to shift) */
8063 return; /* do not make the node uneven-sized */
8065 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
8066 for (i = NUMKEYS(sp); --i >= 0; )
8067 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
8070 sp->mp_upper = sp->mp_lower;
8071 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
8072 SETDSZ(node, nsize);
8074 /* Shift <lower nodes...initial part of subpage> upward */
8075 base = (char *)mp + mp->mp_upper + PAGEBASE;
8076 memmove(base + delta, base, (char *)sp + len - base);
8078 ptr = mp->mp_ptrs[indx];
8079 for (i = NUMKEYS(mp); --i >= 0; ) {
8080 if (mp->mp_ptrs[i] <= ptr)
8081 mp->mp_ptrs[i] += delta;
8083 mp->mp_upper += delta;
8086 /** Initial setup of a sorted-dups cursor.
8087 * Sorted duplicates are implemented as a sub-database for the given key.
8088 * The duplicate data items are actually keys of the sub-database.
8089 * Operations on the duplicate data items are performed using a sub-cursor
8090 * initialized when the sub-database is first accessed. This function does
8091 * the preliminary setup of the sub-cursor, filling in the fields that
8092 * depend only on the parent DB.
8093 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8096 mdb_xcursor_init0(MDB_cursor *mc)
8098 MDB_xcursor *mx = mc->mc_xcursor;
8100 mx->mx_cursor.mc_xcursor = NULL;
8101 mx->mx_cursor.mc_txn = mc->mc_txn;
8102 mx->mx_cursor.mc_db = &mx->mx_db;
8103 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
8104 mx->mx_cursor.mc_dbi = mc->mc_dbi;
8105 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
8106 mx->mx_cursor.mc_snum = 0;
8107 mx->mx_cursor.mc_top = 0;
8108 MC_SET_OVPG(&mx->mx_cursor, NULL);
8109 mx->mx_cursor.mc_flags = C_SUB | (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP));
8110 mx->mx_dbx.md_name.mv_size = 0;
8111 mx->mx_dbx.md_name.mv_data = NULL;
8112 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
8113 mx->mx_dbx.md_dcmp = NULL;
8114 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
8117 /** Final setup of a sorted-dups cursor.
8118 * Sets up the fields that depend on the data from the main cursor.
8119 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8120 * @param[in] node The data containing the #MDB_db record for the
8121 * sorted-dup database.
8124 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
8126 MDB_xcursor *mx = mc->mc_xcursor;
8128 mx->mx_cursor.mc_flags &= C_SUB|C_ORIG_RDONLY|C_WRITEMAP;
8129 if (node->mn_flags & F_SUBDATA) {
8130 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
8131 mx->mx_cursor.mc_pg[0] = 0;
8132 mx->mx_cursor.mc_snum = 0;
8133 mx->mx_cursor.mc_top = 0;
8135 MDB_page *fp = NODEDATA(node);
8136 mx->mx_db.md_pad = 0;
8137 mx->mx_db.md_flags = 0;
8138 mx->mx_db.md_depth = 1;
8139 mx->mx_db.md_branch_pages = 0;
8140 mx->mx_db.md_leaf_pages = 1;
8141 mx->mx_db.md_overflow_pages = 0;
8142 mx->mx_db.md_entries = NUMKEYS(fp);
8143 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
8144 mx->mx_cursor.mc_snum = 1;
8145 mx->mx_cursor.mc_top = 0;
8146 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8147 mx->mx_cursor.mc_pg[0] = fp;
8148 mx->mx_cursor.mc_ki[0] = 0;
8149 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
8150 mx->mx_db.md_flags = MDB_DUPFIXED;
8151 mx->mx_db.md_pad = fp->mp_pad;
8152 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
8153 mx->mx_db.md_flags |= MDB_INTEGERKEY;
8156 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8157 mx->mx_db.md_root));
8158 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8159 if (NEED_CMP_CLONG(mx->mx_dbx.md_cmp, mx->mx_db.md_pad))
8160 mx->mx_dbx.md_cmp = mdb_cmp_clong;
8164 /** Fixup a sorted-dups cursor due to underlying update.
8165 * Sets up some fields that depend on the data from the main cursor.
8166 * Almost the same as init1, but skips initialization steps if the
8167 * xcursor had already been used.
8168 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
8169 * @param[in] src_mx The xcursor of an up-to-date cursor.
8170 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
8173 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
8175 MDB_xcursor *mx = mc->mc_xcursor;
8178 mx->mx_cursor.mc_snum = 1;
8179 mx->mx_cursor.mc_top = 0;
8180 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8181 mx->mx_cursor.mc_ki[0] = 0;
8182 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8183 #if UINT_MAX < MDB_SIZE_MAX /* matches mdb_xcursor_init1:NEED_CMP_CLONG() */
8184 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
8186 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
8189 mx->mx_db = src_mx->mx_db;
8190 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
8191 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8192 mx->mx_db.md_root));
8195 /** Initialize a cursor for a given transaction and database. */
8197 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
8200 mc->mc_backup = NULL;
8203 mc->mc_db = &txn->mt_dbs[dbi];
8204 mc->mc_dbx = &txn->mt_dbxs[dbi];
8205 mc->mc_dbflag = &txn->mt_dbflags[dbi];
8210 MC_SET_OVPG(mc, NULL);
8211 mc->mc_flags = txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
8212 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
8213 mdb_tassert(txn, mx != NULL);
8214 mc->mc_xcursor = mx;
8215 mdb_xcursor_init0(mc);
8217 mc->mc_xcursor = NULL;
8219 if (*mc->mc_dbflag & DB_STALE) {
8220 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
8225 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
8228 size_t size = sizeof(MDB_cursor);
8230 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
8233 if (txn->mt_flags & MDB_TXN_BLOCKED)
8236 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8239 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
8240 size += sizeof(MDB_xcursor);
8242 if ((mc = malloc(size)) != NULL) {
8243 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
8244 if (txn->mt_cursors) {
8245 mc->mc_next = txn->mt_cursors[dbi];
8246 txn->mt_cursors[dbi] = mc;
8247 mc->mc_flags |= C_UNTRACK;
8259 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
8261 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
8264 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
8267 if (txn->mt_flags & MDB_TXN_BLOCKED)
8270 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
8274 /* Return the count of duplicate data items for the current key */
8276 mdb_cursor_count(MDB_cursor *mc, mdb_size_t *countp)
8280 if (mc == NULL || countp == NULL)
8283 if (mc->mc_xcursor == NULL)
8284 return MDB_INCOMPATIBLE;
8286 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
8289 if (!(mc->mc_flags & C_INITIALIZED))
8292 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
8293 return MDB_NOTFOUND;
8295 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8296 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
8299 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
8302 *countp = mc->mc_xcursor->mx_db.md_entries;
8308 mdb_cursor_close(MDB_cursor *mc)
8311 MDB_CURSOR_UNREF(mc, 0);
8313 if (mc && !mc->mc_backup) {
8314 /* remove from txn, if tracked */
8315 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
8316 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
8317 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
8319 *prev = mc->mc_next;
8326 mdb_cursor_txn(MDB_cursor *mc)
8328 if (!mc) return NULL;
8333 mdb_cursor_dbi(MDB_cursor *mc)
8338 /** Replace the key for a branch node with a new key.
8339 * @param[in] mc Cursor pointing to the node to operate on.
8340 * @param[in] key The new key to use.
8341 * @return 0 on success, non-zero on failure.
8344 mdb_update_key(MDB_cursor *mc, MDB_val *key)
8350 int delta, ksize, oksize;
8351 indx_t ptr, i, numkeys, indx;
8354 indx = mc->mc_ki[mc->mc_top];
8355 mp = mc->mc_pg[mc->mc_top];
8356 node = NODEPTR(mp, indx);
8357 ptr = mp->mp_ptrs[indx];
8361 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
8362 k2.mv_data = NODEKEY(node);
8363 k2.mv_size = node->mn_ksize;
8364 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Yu,
8366 mdb_dkey(&k2, kbuf2),
8372 /* Sizes must be 2-byte aligned. */
8373 ksize = EVEN(key->mv_size);
8374 oksize = EVEN(node->mn_ksize);
8375 delta = ksize - oksize;
8377 /* Shift node contents if EVEN(key length) changed. */
8379 if (delta > 0 && SIZELEFT(mp) < delta) {
8381 /* not enough space left, do a delete and split */
8382 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
8383 pgno = NODEPGNO(node);
8384 mdb_node_del(mc, 0);
8385 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
8388 numkeys = NUMKEYS(mp);
8389 for (i = 0; i < numkeys; i++) {
8390 if (mp->mp_ptrs[i] <= ptr)
8391 mp->mp_ptrs[i] -= delta;
8394 base = (char *)mp + mp->mp_upper + PAGEBASE;
8395 len = ptr - mp->mp_upper + NODESIZE;
8396 memmove(base - delta, base, len);
8397 mp->mp_upper -= delta;
8399 node = NODEPTR(mp, indx);
8402 /* But even if no shift was needed, update ksize */
8403 if (node->mn_ksize != key->mv_size)
8404 node->mn_ksize = key->mv_size;
8407 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8413 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
8415 /** Perform \b act while tracking temporary cursor \b mn */
8416 #define WITH_CURSOR_TRACKING(mn, act) do { \
8417 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
8418 if ((mn).mc_flags & C_SUB) { \
8419 dummy.mc_flags = C_INITIALIZED; \
8420 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
8425 tracked->mc_next = *tp; \
8428 *tp = tracked->mc_next; \
8431 /** Move a node from csrc to cdst.
8434 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
8441 unsigned short flags;
8445 /* Mark src and dst as dirty. */
8446 if ((rc = mdb_page_touch(csrc)) ||
8447 (rc = mdb_page_touch(cdst)))
8450 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8451 key.mv_size = csrc->mc_db->md_pad;
8452 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
8454 data.mv_data = NULL;
8458 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
8459 mdb_cassert(csrc, !((size_t)srcnode & 1));
8460 srcpg = NODEPGNO(srcnode);
8461 flags = srcnode->mn_flags;
8462 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8463 unsigned int snum = csrc->mc_snum;
8465 /* must find the lowest key below src */
8466 rc = mdb_page_search_lowest(csrc);
8469 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8470 key.mv_size = csrc->mc_db->md_pad;
8471 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8473 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8474 key.mv_size = NODEKSZ(s2);
8475 key.mv_data = NODEKEY(s2);
8477 csrc->mc_snum = snum--;
8478 csrc->mc_top = snum;
8480 key.mv_size = NODEKSZ(srcnode);
8481 key.mv_data = NODEKEY(srcnode);
8483 data.mv_size = NODEDSZ(srcnode);
8484 data.mv_data = NODEDATA(srcnode);
8486 mn.mc_xcursor = NULL;
8487 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
8488 unsigned int snum = cdst->mc_snum;
8491 /* must find the lowest key below dst */
8492 mdb_cursor_copy(cdst, &mn);
8493 rc = mdb_page_search_lowest(&mn);
8496 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8497 bkey.mv_size = mn.mc_db->md_pad;
8498 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
8500 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8501 bkey.mv_size = NODEKSZ(s2);
8502 bkey.mv_data = NODEKEY(s2);
8504 mn.mc_snum = snum--;
8507 rc = mdb_update_key(&mn, &bkey);
8512 DPRINTF(("moving %s node %u [%s] on page %"Yu" to node %u on page %"Yu,
8513 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
8514 csrc->mc_ki[csrc->mc_top],
8516 csrc->mc_pg[csrc->mc_top]->mp_pgno,
8517 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
8519 /* Add the node to the destination page.
8521 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
8522 if (rc != MDB_SUCCESS)
8525 /* Delete the node from the source page.
8527 mdb_node_del(csrc, key.mv_size);
8530 /* Adjust other cursors pointing to mp */
8531 MDB_cursor *m2, *m3;
8532 MDB_dbi dbi = csrc->mc_dbi;
8533 MDB_page *mpd, *mps;
8535 mps = csrc->mc_pg[csrc->mc_top];
8536 /* If we're adding on the left, bump others up */
8538 mpd = cdst->mc_pg[csrc->mc_top];
8539 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8540 if (csrc->mc_flags & C_SUB)
8541 m3 = &m2->mc_xcursor->mx_cursor;
8544 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8547 m3->mc_pg[csrc->mc_top] == mpd &&
8548 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
8549 m3->mc_ki[csrc->mc_top]++;
8552 m3->mc_pg[csrc->mc_top] == mps &&
8553 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
8554 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8555 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8556 m3->mc_ki[csrc->mc_top-1]++;
8558 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8560 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8561 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8562 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8566 /* Adding on the right, bump others down */
8568 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8569 if (csrc->mc_flags & C_SUB)
8570 m3 = &m2->mc_xcursor->mx_cursor;
8573 if (m3 == csrc) continue;
8574 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8576 if (m3->mc_pg[csrc->mc_top] == mps) {
8577 if (!m3->mc_ki[csrc->mc_top]) {
8578 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8579 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8580 m3->mc_ki[csrc->mc_top-1]--;
8582 m3->mc_ki[csrc->mc_top]--;
8584 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8586 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8587 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8588 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8595 /* Update the parent separators.
8597 if (csrc->mc_ki[csrc->mc_top] == 0) {
8598 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
8599 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8600 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8602 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8603 key.mv_size = NODEKSZ(srcnode);
8604 key.mv_data = NODEKEY(srcnode);
8606 DPRINTF(("update separator for source page %"Yu" to [%s]",
8607 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
8608 mdb_cursor_copy(csrc, &mn);
8611 /* We want mdb_rebalance to find mn when doing fixups */
8612 WITH_CURSOR_TRACKING(mn,
8613 rc = mdb_update_key(&mn, &key));
8617 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8619 indx_t ix = csrc->mc_ki[csrc->mc_top];
8620 nullkey.mv_size = 0;
8621 csrc->mc_ki[csrc->mc_top] = 0;
8622 rc = mdb_update_key(csrc, &nullkey);
8623 csrc->mc_ki[csrc->mc_top] = ix;
8624 mdb_cassert(csrc, rc == MDB_SUCCESS);
8628 if (cdst->mc_ki[cdst->mc_top] == 0) {
8629 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
8630 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8631 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
8633 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
8634 key.mv_size = NODEKSZ(srcnode);
8635 key.mv_data = NODEKEY(srcnode);
8637 DPRINTF(("update separator for destination page %"Yu" to [%s]",
8638 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
8639 mdb_cursor_copy(cdst, &mn);
8642 /* We want mdb_rebalance to find mn when doing fixups */
8643 WITH_CURSOR_TRACKING(mn,
8644 rc = mdb_update_key(&mn, &key));
8648 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
8650 indx_t ix = cdst->mc_ki[cdst->mc_top];
8651 nullkey.mv_size = 0;
8652 cdst->mc_ki[cdst->mc_top] = 0;
8653 rc = mdb_update_key(cdst, &nullkey);
8654 cdst->mc_ki[cdst->mc_top] = ix;
8655 mdb_cassert(cdst, rc == MDB_SUCCESS);
8662 /** Merge one page into another.
8663 * The nodes from the page pointed to by \b csrc will
8664 * be copied to the page pointed to by \b cdst and then
8665 * the \b csrc page will be freed.
8666 * @param[in] csrc Cursor pointing to the source page.
8667 * @param[in] cdst Cursor pointing to the destination page.
8668 * @return 0 on success, non-zero on failure.
8671 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8673 MDB_page *psrc, *pdst;
8680 psrc = csrc->mc_pg[csrc->mc_top];
8681 pdst = cdst->mc_pg[cdst->mc_top];
8683 DPRINTF(("merging page %"Yu" into %"Yu, psrc->mp_pgno, pdst->mp_pgno));
8685 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8686 mdb_cassert(csrc, cdst->mc_snum > 1);
8688 /* Mark dst as dirty. */
8689 if ((rc = mdb_page_touch(cdst)))
8692 /* get dst page again now that we've touched it. */
8693 pdst = cdst->mc_pg[cdst->mc_top];
8695 /* Move all nodes from src to dst.
8697 j = nkeys = NUMKEYS(pdst);
8698 if (IS_LEAF2(psrc)) {
8699 key.mv_size = csrc->mc_db->md_pad;
8700 key.mv_data = METADATA(psrc);
8701 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8702 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8703 if (rc != MDB_SUCCESS)
8705 key.mv_data = (char *)key.mv_data + key.mv_size;
8708 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8709 srcnode = NODEPTR(psrc, i);
8710 if (i == 0 && IS_BRANCH(psrc)) {
8713 mdb_cursor_copy(csrc, &mn);
8714 mn.mc_xcursor = NULL;
8715 /* must find the lowest key below src */
8716 rc = mdb_page_search_lowest(&mn);
8719 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8720 key.mv_size = mn.mc_db->md_pad;
8721 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8723 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8724 key.mv_size = NODEKSZ(s2);
8725 key.mv_data = NODEKEY(s2);
8728 key.mv_size = srcnode->mn_ksize;
8729 key.mv_data = NODEKEY(srcnode);
8732 data.mv_size = NODEDSZ(srcnode);
8733 data.mv_data = NODEDATA(srcnode);
8734 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8735 if (rc != MDB_SUCCESS)
8740 DPRINTF(("dst page %"Yu" now has %u keys (%.1f%% filled)",
8741 pdst->mp_pgno, NUMKEYS(pdst),
8742 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8744 /* Unlink the src page from parent and add to free list.
8747 mdb_node_del(csrc, 0);
8748 if (csrc->mc_ki[csrc->mc_top] == 0) {
8750 rc = mdb_update_key(csrc, &key);
8758 psrc = csrc->mc_pg[csrc->mc_top];
8759 /* If not operating on FreeDB, allow this page to be reused
8760 * in this txn. Otherwise just add to free list.
8762 rc = mdb_page_loose(csrc, psrc);
8766 csrc->mc_db->md_leaf_pages--;
8768 csrc->mc_db->md_branch_pages--;
8770 /* Adjust other cursors pointing to mp */
8771 MDB_cursor *m2, *m3;
8772 MDB_dbi dbi = csrc->mc_dbi;
8773 unsigned int top = csrc->mc_top;
8775 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8776 if (csrc->mc_flags & C_SUB)
8777 m3 = &m2->mc_xcursor->mx_cursor;
8780 if (m3 == csrc) continue;
8781 if (m3->mc_snum < csrc->mc_snum) continue;
8782 if (m3->mc_pg[top] == psrc) {
8783 m3->mc_pg[top] = pdst;
8784 m3->mc_ki[top] += nkeys;
8785 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8786 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8787 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8790 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8792 MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
8793 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8794 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8799 unsigned int snum = cdst->mc_snum;
8800 uint16_t depth = cdst->mc_db->md_depth;
8801 mdb_cursor_pop(cdst);
8802 rc = mdb_rebalance(cdst);
8803 /* Did the tree height change? */
8804 if (depth != cdst->mc_db->md_depth)
8805 snum += cdst->mc_db->md_depth - depth;
8806 cdst->mc_snum = snum;
8807 cdst->mc_top = snum-1;
8812 /** Copy the contents of a cursor.
8813 * @param[in] csrc The cursor to copy from.
8814 * @param[out] cdst The cursor to copy to.
8817 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8821 cdst->mc_txn = csrc->mc_txn;
8822 cdst->mc_dbi = csrc->mc_dbi;
8823 cdst->mc_db = csrc->mc_db;
8824 cdst->mc_dbx = csrc->mc_dbx;
8825 cdst->mc_snum = csrc->mc_snum;
8826 cdst->mc_top = csrc->mc_top;
8827 cdst->mc_flags = csrc->mc_flags;
8828 MC_SET_OVPG(cdst, MC_OVPG(csrc));
8830 for (i=0; i<csrc->mc_snum; i++) {
8831 cdst->mc_pg[i] = csrc->mc_pg[i];
8832 cdst->mc_ki[i] = csrc->mc_ki[i];
8836 /** Rebalance the tree after a delete operation.
8837 * @param[in] mc Cursor pointing to the page where rebalancing
8839 * @return 0 on success, non-zero on failure.
8842 mdb_rebalance(MDB_cursor *mc)
8846 unsigned int ptop, minkeys, thresh;
8850 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8855 thresh = FILL_THRESHOLD;
8857 DPRINTF(("rebalancing %s page %"Yu" (has %u keys, %.1f%% full)",
8858 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8859 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8860 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8862 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8863 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8864 DPRINTF(("no need to rebalance page %"Yu", above fill threshold",
8865 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8869 if (mc->mc_snum < 2) {
8870 MDB_page *mp = mc->mc_pg[0];
8872 DPUTS("Can't rebalance a subpage, ignoring");
8875 if (NUMKEYS(mp) == 0) {
8876 DPUTS("tree is completely empty");
8877 mc->mc_db->md_root = P_INVALID;
8878 mc->mc_db->md_depth = 0;
8879 mc->mc_db->md_leaf_pages = 0;
8880 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8883 /* Adjust cursors pointing to mp */
8886 mc->mc_flags &= ~C_INITIALIZED;
8888 MDB_cursor *m2, *m3;
8889 MDB_dbi dbi = mc->mc_dbi;
8891 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8892 if (mc->mc_flags & C_SUB)
8893 m3 = &m2->mc_xcursor->mx_cursor;
8896 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8898 if (m3->mc_pg[0] == mp) {
8901 m3->mc_flags &= ~C_INITIALIZED;
8905 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8907 DPUTS("collapsing root page!");
8908 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8911 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8912 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
8915 mc->mc_db->md_depth--;
8916 mc->mc_db->md_branch_pages--;
8917 mc->mc_ki[0] = mc->mc_ki[1];
8918 for (i = 1; i<mc->mc_db->md_depth; i++) {
8919 mc->mc_pg[i] = mc->mc_pg[i+1];
8920 mc->mc_ki[i] = mc->mc_ki[i+1];
8923 /* Adjust other cursors pointing to mp */
8924 MDB_cursor *m2, *m3;
8925 MDB_dbi dbi = mc->mc_dbi;
8927 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8928 if (mc->mc_flags & C_SUB)
8929 m3 = &m2->mc_xcursor->mx_cursor;
8932 if (m3 == mc) continue;
8933 if (!(m3->mc_flags & C_INITIALIZED))
8935 if (m3->mc_pg[0] == mp) {
8936 for (i=0; i<mc->mc_db->md_depth; i++) {
8937 m3->mc_pg[i] = m3->mc_pg[i+1];
8938 m3->mc_ki[i] = m3->mc_ki[i+1];
8946 DPUTS("root page doesn't need rebalancing");
8950 /* The parent (branch page) must have at least 2 pointers,
8951 * otherwise the tree is invalid.
8953 ptop = mc->mc_top-1;
8954 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8956 /* Leaf page fill factor is below the threshold.
8957 * Try to move keys from left or right neighbor, or
8958 * merge with a neighbor page.
8963 mdb_cursor_copy(mc, &mn);
8964 mn.mc_xcursor = NULL;
8966 oldki = mc->mc_ki[mc->mc_top];
8967 if (mc->mc_ki[ptop] == 0) {
8968 /* We're the leftmost leaf in our parent.
8970 DPUTS("reading right neighbor");
8972 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8973 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8976 mn.mc_ki[mn.mc_top] = 0;
8977 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8980 /* There is at least one neighbor to the left.
8982 DPUTS("reading left neighbor");
8984 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8985 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8988 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8989 mc->mc_ki[mc->mc_top] = 0;
8993 DPRINTF(("found neighbor page %"Yu" (%u keys, %.1f%% full)",
8994 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8995 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8997 /* If the neighbor page is above threshold and has enough keys,
8998 * move one key from it. Otherwise we should try to merge them.
8999 * (A branch page must never have less than 2 keys.)
9001 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
9002 rc = mdb_node_move(&mn, mc, fromleft);
9004 /* if we inserted on left, bump position up */
9009 rc = mdb_page_merge(&mn, mc);
9011 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
9012 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
9013 /* We want mdb_rebalance to find mn when doing fixups */
9014 WITH_CURSOR_TRACKING(mn,
9015 rc = mdb_page_merge(mc, &mn));
9016 mdb_cursor_copy(&mn, mc);
9018 mc->mc_flags &= ~C_EOF;
9020 mc->mc_ki[mc->mc_top] = oldki;
9024 /** Complete a delete operation started by #mdb_cursor_del(). */
9026 mdb_cursor_del0(MDB_cursor *mc)
9032 MDB_cursor *m2, *m3;
9033 MDB_dbi dbi = mc->mc_dbi;
9035 ki = mc->mc_ki[mc->mc_top];
9036 mp = mc->mc_pg[mc->mc_top];
9037 mdb_node_del(mc, mc->mc_db->md_pad);
9038 mc->mc_db->md_entries--;
9040 /* Adjust other cursors pointing to mp */
9041 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9042 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9043 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9045 if (m3 == mc || m3->mc_snum < mc->mc_snum)
9047 if (m3->mc_pg[mc->mc_top] == mp) {
9048 if (m3->mc_ki[mc->mc_top] == ki) {
9049 m3->mc_flags |= C_DEL;
9050 } else if (m3->mc_ki[mc->mc_top] > ki) {
9051 m3->mc_ki[mc->mc_top]--;
9053 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
9054 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9055 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
9056 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9061 rc = mdb_rebalance(mc);
9063 if (rc == MDB_SUCCESS) {
9064 /* DB is totally empty now, just bail out.
9065 * Other cursors adjustments were already done
9066 * by mdb_rebalance and aren't needed here.
9071 mp = mc->mc_pg[mc->mc_top];
9072 nkeys = NUMKEYS(mp);
9074 /* Adjust other cursors pointing to mp */
9075 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
9076 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9077 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9079 if (m3->mc_snum < mc->mc_snum)
9081 if (m3->mc_pg[mc->mc_top] == mp) {
9082 /* if m3 points past last node in page, find next sibling */
9083 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
9084 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9085 rc = mdb_cursor_sibling(m3, 1);
9086 if (rc == MDB_NOTFOUND) {
9087 m3->mc_flags |= C_EOF;
9092 if (mc->mc_db->md_flags & MDB_DUPSORT) {
9093 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
9094 if (node->mn_flags & F_DUPDATA) {
9095 mdb_xcursor_init1(m3, node);
9096 m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL;
9102 mc->mc_flags |= C_DEL;
9106 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9111 mdb_del(MDB_txn *txn, MDB_dbi dbi,
9112 MDB_val *key, MDB_val *data)
9114 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9117 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9118 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9120 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
9121 /* must ignore any data */
9125 return mdb_del0(txn, dbi, key, data, 0);
9129 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
9130 MDB_val *key, MDB_val *data, unsigned flags)
9135 MDB_val rdata, *xdata;
9139 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
9141 mdb_cursor_init(&mc, txn, dbi, &mx);
9150 flags |= MDB_NODUPDATA;
9152 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
9154 /* let mdb_page_split know about this cursor if needed:
9155 * delete will trigger a rebalance; if it needs to move
9156 * a node from one page to another, it will have to
9157 * update the parent's separator key(s). If the new sepkey
9158 * is larger than the current one, the parent page may
9159 * run out of space, triggering a split. We need this
9160 * cursor to be consistent until the end of the rebalance.
9162 mc.mc_flags |= C_UNTRACK;
9163 mc.mc_next = txn->mt_cursors[dbi];
9164 txn->mt_cursors[dbi] = &mc;
9165 rc = mdb_cursor_del(&mc, flags);
9166 txn->mt_cursors[dbi] = mc.mc_next;
9171 /** Split a page and insert a new node.
9172 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
9173 * The cursor will be updated to point to the actual page and index where
9174 * the node got inserted after the split.
9175 * @param[in] newkey The key for the newly inserted node.
9176 * @param[in] newdata The data for the newly inserted node.
9177 * @param[in] newpgno The page number, if the new node is a branch node.
9178 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
9179 * @return 0 on success, non-zero on failure.
9182 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
9183 unsigned int nflags)
9186 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
9189 int i, j, split_indx, nkeys, pmax;
9190 MDB_env *env = mc->mc_txn->mt_env;
9192 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
9193 MDB_page *copy = NULL;
9194 MDB_page *mp, *rp, *pp;
9199 mp = mc->mc_pg[mc->mc_top];
9200 newindx = mc->mc_ki[mc->mc_top];
9201 nkeys = NUMKEYS(mp);
9203 DPRINTF(("-----> splitting %s page %"Yu" and adding [%s] at index %i/%i",
9204 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
9205 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
9207 /* Create a right sibling. */
9208 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
9210 rp->mp_pad = mp->mp_pad;
9211 DPRINTF(("new right sibling: page %"Yu, rp->mp_pgno));
9213 /* Usually when splitting the root page, the cursor
9214 * height is 1. But when called from mdb_update_key,
9215 * the cursor height may be greater because it walks
9216 * up the stack while finding the branch slot to update.
9218 if (mc->mc_top < 1) {
9219 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
9221 /* shift current top to make room for new parent */
9222 for (i=mc->mc_snum; i>0; i--) {
9223 mc->mc_pg[i] = mc->mc_pg[i-1];
9224 mc->mc_ki[i] = mc->mc_ki[i-1];
9228 mc->mc_db->md_root = pp->mp_pgno;
9229 DPRINTF(("root split! new root = %"Yu, pp->mp_pgno));
9230 new_root = mc->mc_db->md_depth++;
9232 /* Add left (implicit) pointer. */
9233 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
9234 /* undo the pre-push */
9235 mc->mc_pg[0] = mc->mc_pg[1];
9236 mc->mc_ki[0] = mc->mc_ki[1];
9237 mc->mc_db->md_root = mp->mp_pgno;
9238 mc->mc_db->md_depth--;
9245 ptop = mc->mc_top-1;
9246 DPRINTF(("parent branch page is %"Yu, mc->mc_pg[ptop]->mp_pgno));
9249 mdb_cursor_copy(mc, &mn);
9250 mn.mc_xcursor = NULL;
9251 mn.mc_pg[mn.mc_top] = rp;
9252 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
9254 if (nflags & MDB_APPEND) {
9255 mn.mc_ki[mn.mc_top] = 0;
9257 split_indx = newindx;
9261 split_indx = (nkeys+1) / 2;
9266 unsigned int lsize, rsize, ksize;
9267 /* Move half of the keys to the right sibling */
9268 x = mc->mc_ki[mc->mc_top] - split_indx;
9269 ksize = mc->mc_db->md_pad;
9270 split = LEAF2KEY(mp, split_indx, ksize);
9271 rsize = (nkeys - split_indx) * ksize;
9272 lsize = (nkeys - split_indx) * sizeof(indx_t);
9273 mp->mp_lower -= lsize;
9274 rp->mp_lower += lsize;
9275 mp->mp_upper += rsize - lsize;
9276 rp->mp_upper -= rsize - lsize;
9277 sepkey.mv_size = ksize;
9278 if (newindx == split_indx) {
9279 sepkey.mv_data = newkey->mv_data;
9281 sepkey.mv_data = split;
9284 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
9285 memcpy(rp->mp_ptrs, split, rsize);
9286 sepkey.mv_data = rp->mp_ptrs;
9287 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
9288 memcpy(ins, newkey->mv_data, ksize);
9289 mp->mp_lower += sizeof(indx_t);
9290 mp->mp_upper -= ksize - sizeof(indx_t);
9293 memcpy(rp->mp_ptrs, split, x * ksize);
9294 ins = LEAF2KEY(rp, x, ksize);
9295 memcpy(ins, newkey->mv_data, ksize);
9296 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
9297 rp->mp_lower += sizeof(indx_t);
9298 rp->mp_upper -= ksize - sizeof(indx_t);
9299 mc->mc_ki[mc->mc_top] = x;
9302 int psize, nsize, k;
9303 /* Maximum free space in an empty page */
9304 pmax = env->me_psize - PAGEHDRSZ;
9306 nsize = mdb_leaf_size(env, newkey, newdata);
9308 nsize = mdb_branch_size(env, newkey);
9309 nsize = EVEN(nsize);
9311 /* grab a page to hold a temporary copy */
9312 copy = mdb_page_malloc(mc->mc_txn, 1);
9317 copy->mp_pgno = mp->mp_pgno;
9318 copy->mp_flags = mp->mp_flags;
9319 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
9320 copy->mp_upper = env->me_psize - PAGEBASE;
9322 /* prepare to insert */
9323 for (i=0, j=0; i<nkeys; i++) {
9325 copy->mp_ptrs[j++] = 0;
9327 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
9330 /* When items are relatively large the split point needs
9331 * to be checked, because being off-by-one will make the
9332 * difference between success or failure in mdb_node_add.
9334 * It's also relevant if a page happens to be laid out
9335 * such that one half of its nodes are all "small" and
9336 * the other half of its nodes are "large." If the new
9337 * item is also "large" and falls on the half with
9338 * "large" nodes, it also may not fit.
9340 * As a final tweak, if the new item goes on the last
9341 * spot on the page (and thus, onto the new page), bias
9342 * the split so the new page is emptier than the old page.
9343 * This yields better packing during sequential inserts.
9345 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
9346 /* Find split point */
9348 if (newindx <= split_indx || newindx >= nkeys) {
9350 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
9355 for (; i!=k; i+=j) {
9360 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9361 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
9363 if (F_ISSET(node->mn_flags, F_BIGDATA))
9364 psize += sizeof(pgno_t);
9366 psize += NODEDSZ(node);
9368 psize = EVEN(psize);
9370 if (psize > pmax || i == k-j) {
9371 split_indx = i + (j<0);
9376 if (split_indx == newindx) {
9377 sepkey.mv_size = newkey->mv_size;
9378 sepkey.mv_data = newkey->mv_data;
9380 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
9381 sepkey.mv_size = node->mn_ksize;
9382 sepkey.mv_data = NODEKEY(node);
9387 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
9389 /* Copy separator key to the parent.
9391 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
9392 int snum = mc->mc_snum;
9396 /* We want other splits to find mn when doing fixups */
9397 WITH_CURSOR_TRACKING(mn,
9398 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
9403 if (mc->mc_snum > snum) {
9406 /* Right page might now have changed parent.
9407 * Check if left page also changed parent.
9409 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9410 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9411 for (i=0; i<ptop; i++) {
9412 mc->mc_pg[i] = mn.mc_pg[i];
9413 mc->mc_ki[i] = mn.mc_ki[i];
9415 mc->mc_pg[ptop] = mn.mc_pg[ptop];
9416 if (mn.mc_ki[ptop]) {
9417 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
9419 /* find right page's left sibling */
9420 mc->mc_ki[ptop] = mn.mc_ki[ptop];
9421 mdb_cursor_sibling(mc, 0);
9426 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
9429 if (rc != MDB_SUCCESS) {
9432 if (nflags & MDB_APPEND) {
9433 mc->mc_pg[mc->mc_top] = rp;
9434 mc->mc_ki[mc->mc_top] = 0;
9435 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
9438 for (i=0; i<mc->mc_top; i++)
9439 mc->mc_ki[i] = mn.mc_ki[i];
9440 } else if (!IS_LEAF2(mp)) {
9442 mc->mc_pg[mc->mc_top] = rp;
9447 rkey.mv_data = newkey->mv_data;
9448 rkey.mv_size = newkey->mv_size;
9454 /* Update index for the new key. */
9455 mc->mc_ki[mc->mc_top] = j;
9457 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9458 rkey.mv_data = NODEKEY(node);
9459 rkey.mv_size = node->mn_ksize;
9461 xdata.mv_data = NODEDATA(node);
9462 xdata.mv_size = NODEDSZ(node);
9465 pgno = NODEPGNO(node);
9466 flags = node->mn_flags;
9469 if (!IS_LEAF(mp) && j == 0) {
9470 /* First branch index doesn't need key data. */
9474 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
9480 mc->mc_pg[mc->mc_top] = copy;
9485 } while (i != split_indx);
9487 nkeys = NUMKEYS(copy);
9488 for (i=0; i<nkeys; i++)
9489 mp->mp_ptrs[i] = copy->mp_ptrs[i];
9490 mp->mp_lower = copy->mp_lower;
9491 mp->mp_upper = copy->mp_upper;
9492 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
9493 env->me_psize - copy->mp_upper - PAGEBASE);
9495 /* reset back to original page */
9496 if (newindx < split_indx) {
9497 mc->mc_pg[mc->mc_top] = mp;
9499 mc->mc_pg[mc->mc_top] = rp;
9501 /* Make sure mc_ki is still valid.
9503 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9504 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9505 for (i=0; i<=ptop; i++) {
9506 mc->mc_pg[i] = mn.mc_pg[i];
9507 mc->mc_ki[i] = mn.mc_ki[i];
9511 if (nflags & MDB_RESERVE) {
9512 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
9513 if (!(node->mn_flags & F_BIGDATA))
9514 newdata->mv_data = NODEDATA(node);
9517 if (newindx >= split_indx) {
9518 mc->mc_pg[mc->mc_top] = rp;
9520 /* Make sure mc_ki is still valid.
9522 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9523 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9524 for (i=0; i<=ptop; i++) {
9525 mc->mc_pg[i] = mn.mc_pg[i];
9526 mc->mc_ki[i] = mn.mc_ki[i];
9533 /* Adjust other cursors pointing to mp */
9534 MDB_cursor *m2, *m3;
9535 MDB_dbi dbi = mc->mc_dbi;
9536 nkeys = NUMKEYS(mp);
9538 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9539 if (mc->mc_flags & C_SUB)
9540 m3 = &m2->mc_xcursor->mx_cursor;
9545 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9549 /* sub cursors may be on different DB */
9550 if (m3->mc_pg[0] != mp)
9553 for (k=new_root; k>=0; k--) {
9554 m3->mc_ki[k+1] = m3->mc_ki[k];
9555 m3->mc_pg[k+1] = m3->mc_pg[k];
9557 if (m3->mc_ki[0] >= nkeys) {
9562 m3->mc_pg[0] = mc->mc_pg[0];
9566 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
9567 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
9568 m3->mc_ki[mc->mc_top]++;
9569 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9570 m3->mc_pg[mc->mc_top] = rp;
9571 m3->mc_ki[mc->mc_top] -= nkeys;
9572 for (i=0; i<mc->mc_top; i++) {
9573 m3->mc_ki[i] = mn.mc_ki[i];
9574 m3->mc_pg[i] = mn.mc_pg[i];
9577 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
9578 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
9581 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
9583 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9584 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
9585 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9589 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
9592 if (copy) /* tmp page */
9593 mdb_page_free(env, copy);
9595 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9600 mdb_put(MDB_txn *txn, MDB_dbi dbi,
9601 MDB_val *key, MDB_val *data, unsigned int flags)
9607 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9610 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
9613 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9614 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9616 mdb_cursor_init(&mc, txn, dbi, &mx);
9617 mc.mc_next = txn->mt_cursors[dbi];
9618 txn->mt_cursors[dbi] = &mc;
9619 rc = mdb_cursor_put(&mc, key, data, flags);
9620 txn->mt_cursors[dbi] = mc.mc_next;
9625 #define MDB_WBUF (1024*1024)
9627 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
9629 /** State needed for a double-buffering compacting copy. */
9630 typedef struct mdb_copy {
9631 pthread_mutex_t mc_mutex;
9632 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
9639 pgno_t mc_next_pgno;
9641 int mc_toggle; /**< Buffer number in provider */
9642 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
9643 volatile int mc_error; /**< Error code, never cleared if set */
9646 /** Dedicated writer thread for compacting copy. */
9647 static THREAD_RET ESECT CALL_CONV
9648 mdb_env_copythr(void *arg)
9652 int toggle = 0, wsize, rc;
9655 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9658 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9661 pthread_mutex_lock(&my->mc_mutex);
9664 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9665 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
9667 wsize = my->mc_wlen[toggle];
9668 ptr = my->mc_wbuf[toggle];
9671 while (wsize > 0 && !my->mc_error) {
9672 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9676 } else if (len > 0) {
9689 /* If there's an overflow page tail, write it too */
9690 if (my->mc_olen[toggle]) {
9691 wsize = my->mc_olen[toggle];
9692 ptr = my->mc_over[toggle];
9693 my->mc_olen[toggle] = 0;
9696 my->mc_wlen[toggle] = 0;
9698 /* Return the empty buffer to provider */
9700 pthread_cond_signal(&my->mc_cond);
9702 pthread_mutex_unlock(&my->mc_mutex);
9703 return (THREAD_RET)0;
9707 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9709 * @param[in] my control structure.
9710 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9713 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9715 pthread_mutex_lock(&my->mc_mutex);
9716 my->mc_new += adjust;
9717 pthread_cond_signal(&my->mc_cond);
9718 while (my->mc_new & 2) /* both buffers in use */
9719 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9720 pthread_mutex_unlock(&my->mc_mutex);
9722 my->mc_toggle ^= (adjust & 1);
9723 /* Both threads reset mc_wlen, to be safe from threading errors */
9724 my->mc_wlen[my->mc_toggle] = 0;
9725 return my->mc_error;
9728 /** Depth-first tree traversal for compacting copy. */
9730 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9732 MDB_cursor mc = {0};
9734 MDB_page *mo, *mp, *leaf;
9739 /* Empty DB, nothing to do */
9740 if (*pg == P_INVALID)
9744 mc.mc_txn = my->mc_txn;
9745 mc.mc_flags = my->mc_txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
9747 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9750 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9754 /* Make cursor pages writable */
9755 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9759 for (i=0; i<mc.mc_top; i++) {
9760 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9761 mc.mc_pg[i] = (MDB_page *)ptr;
9762 ptr += my->mc_env->me_psize;
9765 /* This is writable space for a leaf page. Usually not needed. */
9766 leaf = (MDB_page *)ptr;
9768 toggle = my->mc_toggle;
9769 while (mc.mc_snum > 0) {
9771 mp = mc.mc_pg[mc.mc_top];
9775 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9776 for (i=0; i<n; i++) {
9777 ni = NODEPTR(mp, i);
9778 if (ni->mn_flags & F_BIGDATA) {
9782 /* Need writable leaf */
9784 mc.mc_pg[mc.mc_top] = leaf;
9785 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9787 ni = NODEPTR(mp, i);
9790 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9791 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
9792 rc = mdb_page_get(&mc, pg, &omp, NULL);
9795 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9796 rc = mdb_env_cthr_toggle(my, 1);
9799 toggle = my->mc_toggle;
9801 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9802 memcpy(mo, omp, my->mc_env->me_psize);
9803 mo->mp_pgno = my->mc_next_pgno;
9804 my->mc_next_pgno += omp->mp_pages;
9805 my->mc_wlen[toggle] += my->mc_env->me_psize;
9806 if (omp->mp_pages > 1) {
9807 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9808 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9809 rc = mdb_env_cthr_toggle(my, 1);
9812 toggle = my->mc_toggle;
9814 } else if (ni->mn_flags & F_SUBDATA) {
9817 /* Need writable leaf */
9819 mc.mc_pg[mc.mc_top] = leaf;
9820 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9822 ni = NODEPTR(mp, i);
9825 memcpy(&db, NODEDATA(ni), sizeof(db));
9826 my->mc_toggle = toggle;
9827 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9830 toggle = my->mc_toggle;
9831 memcpy(NODEDATA(ni), &db, sizeof(db));
9836 mc.mc_ki[mc.mc_top]++;
9837 if (mc.mc_ki[mc.mc_top] < n) {
9840 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9842 rc = mdb_page_get(&mc, pg, &mp, NULL);
9847 mc.mc_ki[mc.mc_top] = 0;
9848 if (IS_BRANCH(mp)) {
9849 /* Whenever we advance to a sibling branch page,
9850 * we must proceed all the way down to its first leaf.
9852 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9855 mc.mc_pg[mc.mc_top] = mp;
9859 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9860 rc = mdb_env_cthr_toggle(my, 1);
9863 toggle = my->mc_toggle;
9865 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9866 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9867 mo->mp_pgno = my->mc_next_pgno++;
9868 my->mc_wlen[toggle] += my->mc_env->me_psize;
9870 /* Update parent if there is one */
9871 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9872 SETPGNO(ni, mo->mp_pgno);
9873 mdb_cursor_pop(&mc);
9875 /* Otherwise we're done */
9885 /** Copy environment with compaction. */
9887 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9892 MDB_txn *txn = NULL;
9894 pgno_t root, new_root;
9895 int rc = MDB_SUCCESS;
9898 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
9899 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
9903 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9904 if (my.mc_wbuf[0] == NULL) {
9905 /* _aligned_malloc() sets errno, but we use Windows error codes */
9906 rc = ERROR_NOT_ENOUGH_MEMORY;
9910 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
9912 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
9914 #ifdef HAVE_MEMALIGN
9915 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9916 if (my.mc_wbuf[0] == NULL) {
9923 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
9929 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9930 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9931 my.mc_next_pgno = NUM_METAS;
9934 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
9938 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9942 mp = (MDB_page *)my.mc_wbuf[0];
9943 memset(mp, 0, NUM_METAS * env->me_psize);
9945 mp->mp_flags = P_META;
9946 mm = (MDB_meta *)METADATA(mp);
9947 mdb_env_init_meta0(env, mm);
9948 mm->mm_address = env->me_metas[0]->mm_address;
9950 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9952 mp->mp_flags = P_META;
9953 *(MDB_meta *)METADATA(mp) = *mm;
9954 mm = (MDB_meta *)METADATA(mp);
9956 /* Set metapage 1 with current main DB */
9957 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
9958 if (root != P_INVALID) {
9959 /* Count free pages + freeDB pages. Subtract from last_pg
9960 * to find the new last_pg, which also becomes the new root.
9962 MDB_ID freecount = 0;
9965 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9966 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9967 freecount += *(MDB_ID *)data.mv_data;
9968 if (rc != MDB_NOTFOUND)
9970 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9971 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9972 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9974 new_root = txn->mt_next_pgno - 1 - freecount;
9975 mm->mm_last_pg = new_root;
9976 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9977 mm->mm_dbs[MAIN_DBI].md_root = new_root;
9979 /* When the DB is empty, handle it specially to
9980 * fix any breakage like page leaks from ITS#8174.
9982 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
9984 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
9985 mm->mm_txnid = 1; /* use metapage 1 */
9988 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9990 rc = mdb_env_cwalk(&my, &root, 0);
9991 if (rc == MDB_SUCCESS && root != new_root) {
9992 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
9998 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
9999 rc = THREAD_FINISH(thr);
10000 mdb_txn_abort(txn);
10004 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
10005 if (my.mc_cond) CloseHandle(my.mc_cond);
10006 if (my.mc_mutex) CloseHandle(my.mc_mutex);
10008 free(my.mc_wbuf[0]);
10009 pthread_cond_destroy(&my.mc_cond);
10011 pthread_mutex_destroy(&my.mc_mutex);
10013 return rc ? rc : my.mc_error;
10016 /** Copy environment as-is. */
10018 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
10020 MDB_txn *txn = NULL;
10021 mdb_mutexref_t wmutex = NULL;
10023 mdb_size_t wsize, w3;
10027 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
10031 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
10034 /* Do the lock/unlock of the reader mutex before starting the
10035 * write txn. Otherwise other read txns could block writers.
10037 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
10041 if (env->me_txns) {
10042 /* We must start the actual read txn after blocking writers */
10043 mdb_txn_end(txn, MDB_END_RESET_TMP);
10045 /* Temporarily block writers until we snapshot the meta pages */
10046 wmutex = env->me_wmutex;
10047 if (LOCK_MUTEX(rc, env, wmutex))
10050 rc = mdb_txn_renew0(txn);
10052 UNLOCK_MUTEX(wmutex);
10057 wsize = env->me_psize * NUM_METAS;
10061 DO_WRITE(rc, fd, ptr, w2, len);
10065 } else if (len > 0) {
10071 /* Non-blocking or async handles are not supported */
10077 UNLOCK_MUTEX(wmutex);
10082 w3 = txn->mt_next_pgno * env->me_psize;
10084 mdb_size_t fsize = 0;
10085 if ((rc = mdb_fsize(env->me_fd, &fsize)))
10090 wsize = w3 - wsize;
10091 while (wsize > 0) {
10092 if (wsize > MAX_WRITE)
10096 DO_WRITE(rc, fd, ptr, w2, len);
10100 } else if (len > 0) {
10112 mdb_txn_abort(txn);
10117 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
10119 if (flags & MDB_CP_COMPACT)
10120 return mdb_env_copyfd1(env, fd);
10122 return mdb_env_copyfd0(env, fd);
10126 mdb_env_copyfd(MDB_env *env, HANDLE fd)
10128 return mdb_env_copyfd2(env, fd, 0);
10132 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
10136 HANDLE newfd = INVALID_HANDLE_VALUE;
10141 if (env->me_flags & MDB_NOSUBDIR) {
10142 lpath = (char *)path;
10144 len = strlen(path);
10145 len += sizeof(DATANAME);
10146 lpath = malloc(len);
10149 sprintf(lpath, "%s" DATANAME, path);
10152 /* The destination path must exist, but the destination file must not.
10153 * We don't want the OS to cache the writes, since the source data is
10154 * already in the OS cache.
10157 rc = utf8_to_utf16(lpath, -1, &wpath, NULL);
10160 newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
10161 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
10164 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
10166 if (newfd == INVALID_HANDLE_VALUE) {
10171 if (env->me_psize >= env->me_os_psize) {
10173 /* Set O_DIRECT if the file system supports it */
10174 if ((rc = fcntl(newfd, F_GETFL)) != -1)
10175 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
10177 #ifdef F_NOCACHE /* __APPLE__ */
10178 rc = fcntl(newfd, F_NOCACHE, 1);
10186 rc = mdb_env_copyfd2(env, newfd, flags);
10189 if (!(env->me_flags & MDB_NOSUBDIR))
10191 if (newfd != INVALID_HANDLE_VALUE)
10192 if (close(newfd) < 0 && rc == MDB_SUCCESS)
10199 mdb_env_copy(MDB_env *env, const char *path)
10201 return mdb_env_copy2(env, path, 0);
10205 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
10207 if (flag & ~CHANGEABLE)
10210 env->me_flags |= flag;
10212 env->me_flags &= ~flag;
10213 return MDB_SUCCESS;
10217 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
10222 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
10223 return MDB_SUCCESS;
10227 mdb_env_set_userctx(MDB_env *env, void *ctx)
10231 env->me_userctx = ctx;
10232 return MDB_SUCCESS;
10236 mdb_env_get_userctx(MDB_env *env)
10238 return env ? env->me_userctx : NULL;
10242 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
10247 env->me_assert_func = func;
10249 return MDB_SUCCESS;
10253 mdb_env_get_path(MDB_env *env, const char **arg)
10258 *arg = env->me_path;
10259 return MDB_SUCCESS;
10263 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
10269 return MDB_SUCCESS;
10272 /** Common code for #mdb_stat() and #mdb_env_stat().
10273 * @param[in] env the environment to operate in.
10274 * @param[in] db the #MDB_db record containing the stats to return.
10275 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
10276 * @return 0, this function always succeeds.
10279 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
10281 arg->ms_psize = env->me_psize;
10282 arg->ms_depth = db->md_depth;
10283 arg->ms_branch_pages = db->md_branch_pages;
10284 arg->ms_leaf_pages = db->md_leaf_pages;
10285 arg->ms_overflow_pages = db->md_overflow_pages;
10286 arg->ms_entries = db->md_entries;
10288 return MDB_SUCCESS;
10292 mdb_env_stat(MDB_env *env, MDB_stat *arg)
10296 if (env == NULL || arg == NULL)
10299 meta = mdb_env_pick_meta(env);
10301 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
10305 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
10309 if (env == NULL || arg == NULL)
10312 meta = mdb_env_pick_meta(env);
10313 arg->me_mapaddr = meta->mm_address;
10314 arg->me_last_pgno = meta->mm_last_pg;
10315 arg->me_last_txnid = meta->mm_txnid;
10317 arg->me_mapsize = env->me_mapsize;
10318 arg->me_maxreaders = env->me_maxreaders;
10319 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
10320 return MDB_SUCCESS;
10323 /** Set the default comparison functions for a database.
10324 * Called immediately after a database is opened to set the defaults.
10325 * The user can then override them with #mdb_set_compare() or
10326 * #mdb_set_dupsort().
10327 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
10328 * @param[in] dbi A database handle returned by #mdb_dbi_open()
10331 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
10333 uint16_t f = txn->mt_dbs[dbi].md_flags;
10335 txn->mt_dbxs[dbi].md_cmp =
10336 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
10337 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
10339 txn->mt_dbxs[dbi].md_dcmp =
10340 !(f & MDB_DUPSORT) ? 0 :
10341 ((f & MDB_INTEGERDUP)
10342 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
10343 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
10346 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
10352 int rc, dbflag, exact;
10353 unsigned int unused = 0, seq;
10357 if (flags & ~VALID_FLAGS)
10359 if (txn->mt_flags & MDB_TXN_BLOCKED)
10360 return MDB_BAD_TXN;
10365 if (flags & PERSISTENT_FLAGS) {
10366 uint16_t f2 = flags & PERSISTENT_FLAGS;
10367 /* make sure flag changes get committed */
10368 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
10369 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
10370 txn->mt_flags |= MDB_TXN_DIRTY;
10373 mdb_default_cmp(txn, MAIN_DBI);
10374 return MDB_SUCCESS;
10377 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
10378 mdb_default_cmp(txn, MAIN_DBI);
10381 /* Is the DB already open? */
10382 len = strlen(name);
10383 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
10384 if (!txn->mt_dbxs[i].md_name.mv_size) {
10385 /* Remember this free slot */
10386 if (!unused) unused = i;
10389 if (len == txn->mt_dbxs[i].md_name.mv_size &&
10390 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
10392 return MDB_SUCCESS;
10396 /* If no free slot and max hit, fail */
10397 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
10398 return MDB_DBS_FULL;
10400 /* Cannot mix named databases with some mainDB flags */
10401 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
10402 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
10404 /* Find the DB info */
10405 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
10408 key.mv_data = (void *)name;
10409 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
10410 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
10411 if (rc == MDB_SUCCESS) {
10412 /* make sure this is actually a DB */
10413 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
10414 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
10415 return MDB_INCOMPATIBLE;
10416 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
10420 /* Done here so we cannot fail after creating a new DB */
10421 if ((namedup = strdup(name)) == NULL)
10425 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
10426 data.mv_size = sizeof(MDB_db);
10427 data.mv_data = &dummy;
10428 memset(&dummy, 0, sizeof(dummy));
10429 dummy.md_root = P_INVALID;
10430 dummy.md_flags = flags & PERSISTENT_FLAGS;
10431 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
10432 dbflag |= DB_DIRTY;
10438 /* Got info, register DBI in this txn */
10439 unsigned int slot = unused ? unused : txn->mt_numdbs;
10440 txn->mt_dbxs[slot].md_name.mv_data = namedup;
10441 txn->mt_dbxs[slot].md_name.mv_size = len;
10442 txn->mt_dbxs[slot].md_rel = NULL;
10443 txn->mt_dbflags[slot] = dbflag;
10444 /* txn-> and env-> are the same in read txns, use
10445 * tmp variable to avoid undefined assignment
10447 seq = ++txn->mt_env->me_dbiseqs[slot];
10448 txn->mt_dbiseqs[slot] = seq;
10450 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
10452 mdb_default_cmp(txn, slot);
10462 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
10464 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
10467 if (txn->mt_flags & MDB_TXN_BLOCKED)
10468 return MDB_BAD_TXN;
10470 if (txn->mt_dbflags[dbi] & DB_STALE) {
10473 /* Stale, must read the DB's root. cursor_init does it for us. */
10474 mdb_cursor_init(&mc, txn, dbi, &mx);
10476 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
10479 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
10482 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
10484 ptr = env->me_dbxs[dbi].md_name.mv_data;
10485 /* If there was no name, this was already closed */
10487 env->me_dbxs[dbi].md_name.mv_data = NULL;
10488 env->me_dbxs[dbi].md_name.mv_size = 0;
10489 env->me_dbflags[dbi] = 0;
10490 env->me_dbiseqs[dbi]++;
10495 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
10497 /* We could return the flags for the FREE_DBI too but what's the point? */
10498 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10500 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
10501 return MDB_SUCCESS;
10504 /** Add all the DB's pages to the free list.
10505 * @param[in] mc Cursor on the DB to free.
10506 * @param[in] subs non-Zero to check for sub-DBs in this DB.
10507 * @return 0 on success, non-zero on failure.
10510 mdb_drop0(MDB_cursor *mc, int subs)
10514 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
10515 if (rc == MDB_SUCCESS) {
10516 MDB_txn *txn = mc->mc_txn;
10521 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
10522 * This also avoids any P_LEAF2 pages, which have no nodes.
10523 * Also if the DB doesn't have sub-DBs and has no overflow
10524 * pages, omit scanning leaves.
10526 if ((mc->mc_flags & C_SUB) ||
10527 (!subs && !mc->mc_db->md_overflow_pages))
10528 mdb_cursor_pop(mc);
10530 mdb_cursor_copy(mc, &mx);
10532 /* bump refcount for mx's pages */
10533 for (i=0; i<mc->mc_snum; i++)
10534 mdb_page_get(&mx, mc->mc_pg[i]->mp_pgno, &mx.mc_pg[i], NULL);
10536 while (mc->mc_snum > 0) {
10537 MDB_page *mp = mc->mc_pg[mc->mc_top];
10538 unsigned n = NUMKEYS(mp);
10540 for (i=0; i<n; i++) {
10541 ni = NODEPTR(mp, i);
10542 if (ni->mn_flags & F_BIGDATA) {
10545 memcpy(&pg, NODEDATA(ni), sizeof(pg));
10546 rc = mdb_page_get(mc, pg, &omp, NULL);
10549 mdb_cassert(mc, IS_OVERFLOW(omp));
10550 rc = mdb_midl_append_range(&txn->mt_free_pgs,
10551 pg, omp->mp_pages);
10554 mc->mc_db->md_overflow_pages -= omp->mp_pages;
10555 if (!mc->mc_db->md_overflow_pages && !subs)
10557 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
10558 mdb_xcursor_init1(mc, ni);
10559 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
10564 if (!subs && !mc->mc_db->md_overflow_pages)
10567 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
10569 for (i=0; i<n; i++) {
10571 ni = NODEPTR(mp, i);
10574 mdb_midl_xappend(txn->mt_free_pgs, pg);
10579 mc->mc_ki[mc->mc_top] = i;
10580 rc = mdb_cursor_sibling(mc, 1);
10582 if (rc != MDB_NOTFOUND)
10584 /* no more siblings, go back to beginning
10585 * of previous level.
10588 mdb_cursor_pop(mc);
10590 for (i=1; i<mc->mc_snum; i++) {
10592 mc->mc_pg[i] = mx.mc_pg[i];
10597 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
10600 txn->mt_flags |= MDB_TXN_ERROR;
10601 /* drop refcount for mx's pages */
10602 MDB_CURSOR_UNREF(&mx, 0);
10603 } else if (rc == MDB_NOTFOUND) {
10606 mc->mc_flags &= ~C_INITIALIZED;
10610 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
10612 MDB_cursor *mc, *m2;
10615 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10618 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
10621 if (TXN_DBI_CHANGED(txn, dbi))
10622 return MDB_BAD_DBI;
10624 rc = mdb_cursor_open(txn, dbi, &mc);
10628 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
10629 /* Invalidate the dropped DB's cursors */
10630 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
10631 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
10635 /* Can't delete the main DB */
10636 if (del && dbi >= CORE_DBS) {
10637 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
10639 txn->mt_dbflags[dbi] = DB_STALE;
10640 mdb_dbi_close(txn->mt_env, dbi);
10642 txn->mt_flags |= MDB_TXN_ERROR;
10645 /* reset the DB record, mark it dirty */
10646 txn->mt_dbflags[dbi] |= DB_DIRTY;
10647 txn->mt_dbs[dbi].md_depth = 0;
10648 txn->mt_dbs[dbi].md_branch_pages = 0;
10649 txn->mt_dbs[dbi].md_leaf_pages = 0;
10650 txn->mt_dbs[dbi].md_overflow_pages = 0;
10651 txn->mt_dbs[dbi].md_entries = 0;
10652 txn->mt_dbs[dbi].md_root = P_INVALID;
10654 txn->mt_flags |= MDB_TXN_DIRTY;
10657 mdb_cursor_close(mc);
10661 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10663 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10666 txn->mt_dbxs[dbi].md_cmp = cmp;
10667 return MDB_SUCCESS;
10670 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10672 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10675 txn->mt_dbxs[dbi].md_dcmp = cmp;
10676 return MDB_SUCCESS;
10679 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
10681 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10684 txn->mt_dbxs[dbi].md_rel = rel;
10685 return MDB_SUCCESS;
10688 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
10690 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10693 txn->mt_dbxs[dbi].md_relctx = ctx;
10694 return MDB_SUCCESS;
10698 mdb_env_get_maxkeysize(MDB_env *env)
10700 return ENV_MAXKEY(env);
10704 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
10706 unsigned int i, rdrs;
10709 int rc = 0, first = 1;
10713 if (!env->me_txns) {
10714 return func("(no reader locks)\n", ctx);
10716 rdrs = env->me_txns->mti_numreaders;
10717 mr = env->me_txns->mti_readers;
10718 for (i=0; i<rdrs; i++) {
10719 if (mr[i].mr_pid) {
10720 txnid_t txnid = mr[i].mr_txnid;
10721 sprintf(buf, txnid == (txnid_t)-1 ?
10722 "%10d %"Z"x -\n" : "%10d %"Z"x %"Yu"\n",
10723 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10726 rc = func(" pid thread txnid\n", ctx);
10730 rc = func(buf, ctx);
10736 rc = func("(no active readers)\n", ctx);
10741 /** Insert pid into list if not already present.
10742 * return -1 if already present.
10745 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10747 /* binary search of pid in list */
10749 unsigned cursor = 1;
10751 unsigned n = ids[0];
10754 unsigned pivot = n >> 1;
10755 cursor = base + pivot + 1;
10756 val = pid - ids[cursor];
10761 } else if ( val > 0 ) {
10766 /* found, so it's a duplicate */
10775 for (n = ids[0]; n > cursor; n--)
10782 mdb_reader_check(MDB_env *env, int *dead)
10788 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10791 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
10793 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10795 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10796 unsigned int i, j, rdrs;
10798 MDB_PID_T *pids, pid;
10799 int rc = MDB_SUCCESS, count = 0;
10801 rdrs = env->me_txns->mti_numreaders;
10802 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10806 mr = env->me_txns->mti_readers;
10807 for (i=0; i<rdrs; i++) {
10808 pid = mr[i].mr_pid;
10809 if (pid && pid != env->me_pid) {
10810 if (mdb_pid_insert(pids, pid) == 0) {
10811 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10812 /* Stale reader found */
10815 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10816 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10818 rdrs = 0; /* the above checked all readers */
10820 /* Recheck, a new process may have reused pid */
10821 if (mdb_reader_pid(env, Pidcheck, pid))
10825 for (; j<rdrs; j++)
10826 if (mr[j].mr_pid == pid) {
10827 DPRINTF(("clear stale reader pid %u txn %"Yd,
10828 (unsigned) pid, mr[j].mr_txnid));
10833 UNLOCK_MUTEX(rmutex);
10844 #ifdef MDB_ROBUST_SUPPORTED
10845 /** Handle #LOCK_MUTEX0() failure.
10846 * Try to repair the lock file if the mutex owner died.
10847 * @param[in] env the environment handle
10848 * @param[in] mutex LOCK_MUTEX0() mutex
10849 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10850 * @return 0 on success with the mutex locked, or an error code on failure.
10853 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10858 if (rc == MDB_OWNERDEAD) {
10859 /* We own the mutex. Clean up after dead previous owner. */
10861 rlocked = (mutex == env->me_rmutex);
10863 /* Keep mti_txnid updated, otherwise next writer can
10864 * overwrite data which latest meta page refers to.
10866 meta = mdb_env_pick_meta(env);
10867 env->me_txns->mti_txnid = meta->mm_txnid;
10868 /* env is hosed if the dead thread was ours */
10870 env->me_flags |= MDB_FATAL_ERROR;
10871 env->me_txn = NULL;
10875 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10876 (rc ? "this process' env is hosed" : "recovering")));
10877 rc2 = mdb_reader_check0(env, rlocked, NULL);
10879 rc2 = mdb_mutex_consistent(mutex);
10880 if (rc || (rc = rc2)) {
10881 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10882 UNLOCK_MUTEX(mutex);
10888 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10893 #endif /* MDB_ROBUST_SUPPORTED */
10896 #if defined(_WIN32)
10897 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
10901 need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
10902 if (need == 0xFFFD)
10906 result = malloc(sizeof(wchar_t) * need);
10909 MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
10915 #endif /* defined(_WIN32) */