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 || ULONG_MAX % 0xFFFF
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
306 /* glibc < 2.12 only provided _np API */
307 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
308 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
309 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
310 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
311 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
314 #endif /* MDB_USE_ROBUST */
316 #if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
317 #define MDB_ROBUST_SUPPORTED 1
321 #define MDB_USE_HASH 1
322 #define MDB_PIDLOCK 0
323 #define THREAD_RET DWORD
324 #define pthread_t HANDLE
325 #define pthread_mutex_t HANDLE
326 #define pthread_cond_t HANDLE
327 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
328 #define pthread_key_t DWORD
329 #define pthread_self() GetCurrentThreadId()
330 #define pthread_key_create(x,y) \
331 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
332 #define pthread_key_delete(x) TlsFree(x)
333 #define pthread_getspecific(x) TlsGetValue(x)
334 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
335 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
336 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
337 #define pthread_cond_signal(x) SetEvent(*x)
338 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
339 #define THREAD_CREATE(thr,start,arg) \
340 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
341 #define THREAD_FINISH(thr) \
342 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
343 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
344 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
345 #define mdb_mutex_consistent(mutex) 0
346 #define getpid() GetCurrentProcessId()
347 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
348 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
349 #define ErrCode() GetLastError()
350 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
351 #define close(fd) (CloseHandle(fd) ? 0 : -1)
352 #define munmap(ptr,len) UnmapViewOfFile(ptr)
353 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
354 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
356 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
359 #define THREAD_RET void *
360 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
361 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
363 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
364 #define MDB_PIDLOCK 1
366 #ifdef MDB_USE_POSIX_SEM
368 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
369 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
370 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
373 mdb_sem_wait(sem_t *sem)
376 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
380 #elif defined MDB_USE_SYSV_SEM
382 typedef struct mdb_mutex {
386 } mdb_mutex_t[1], *mdb_mutexref_t;
388 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
389 #define UNLOCK_MUTEX(mutex) do { \
390 struct sembuf sb = { 0, 1, SEM_UNDO }; \
391 sb.sem_num = (mutex)->semnum; \
392 *(mutex)->locked = 0; \
393 semop((mutex)->semid, &sb, 1); \
397 mdb_sem_wait(mdb_mutexref_t sem)
399 int rc, *locked = sem->locked;
400 struct sembuf sb = { 0, -1, SEM_UNDO };
401 sb.sem_num = sem->semnum;
403 if (!semop(sem->semid, &sb, 1)) {
404 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
408 } while ((rc = errno) == EINTR);
412 #define mdb_mutex_consistent(mutex) 0
414 #else /* MDB_USE_POSIX_MUTEX: */
415 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
416 * local variables keep it (mdb_mutexref_t).
418 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
419 * be the same, or an array[size 1] and a pointer.
422 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
424 /** Lock the reader or writer mutex.
425 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
427 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
428 /** Unlock the reader or writer mutex.
430 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
431 /** Mark mutex-protected data as repaired, after death of previous owner.
433 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
434 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
436 /** Get the error code for the last failed system function.
438 #define ErrCode() errno
440 /** An abstraction for a file handle.
441 * On POSIX systems file handles are small integers. On Windows
442 * they're opaque pointers.
446 /** A value for an invalid file handle.
447 * Mainly used to initialize file variables and signify that they are
450 #define INVALID_HANDLE_VALUE (-1)
452 /** Get the size of a memory page for the system.
453 * This is the basic size that the platform's memory manager uses, and is
454 * fundamental to the use of memory-mapped files.
456 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
459 #define Z MDB_FMT_Z /**< printf/scanf format modifier for size_t */
460 #define Y MDB_FMT_Y /**< printf/scanf format modifier for #mdb_size_t */
462 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
464 #elif defined(MDB_USE_SYSV_SEM)
465 #define MNAME_LEN (sizeof(int))
467 #define MNAME_LEN (sizeof(pthread_mutex_t))
470 #ifdef MDB_USE_SYSV_SEM
471 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
473 #define SYSV_SEM_FLAG 0
478 #ifdef MDB_ROBUST_SUPPORTED
479 /** Lock mutex, handle any error, set rc = result.
480 * Return 0 on success, nonzero (not rc) on error.
482 #define LOCK_MUTEX(rc, env, mutex) \
483 (((rc) = LOCK_MUTEX0(mutex)) && \
484 ((rc) = mdb_mutex_failed(env, mutex, rc)))
485 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
487 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
488 #define mdb_mutex_failed(env, mutex, rc) (rc)
492 /** A flag for opening a file and requesting synchronous data writes.
493 * This is only used when writing a meta page. It's not strictly needed;
494 * we could just do a normal write and then immediately perform a flush.
495 * But if this flag is available it saves us an extra system call.
497 * @note If O_DSYNC is undefined but exists in /usr/include,
498 * preferably set some compiler flag to get the definition.
502 # define MDB_DSYNC O_DSYNC
504 # define MDB_DSYNC O_SYNC
509 /** Function for flushing the data of a file. Define this to fsync
510 * if fdatasync() is not supported.
512 #ifndef MDB_FDATASYNC
513 # define MDB_FDATASYNC fdatasync
517 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
528 /** A page number in the database.
529 * Note that 64 bit page numbers are overkill, since pages themselves
530 * already represent 12-13 bits of addressable memory, and the OS will
531 * always limit applications to a maximum of 63 bits of address space.
533 * @note In the #MDB_node structure, we only store 48 bits of this value,
534 * which thus limits us to only 60 bits of addressable data.
536 typedef MDB_ID pgno_t;
538 /** A transaction ID.
539 * See struct MDB_txn.mt_txnid for details.
541 typedef MDB_ID txnid_t;
543 /** @defgroup debug Debug Macros
547 /** Enable debug output. Needs variable argument macros (a C99 feature).
548 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
549 * read from and written to the database (used for free space management).
555 static int mdb_debug;
556 static txnid_t mdb_debug_start;
558 /** Print a debug message with printf formatting.
559 * Requires double parenthesis around 2 or more args.
561 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
562 # define DPRINTF0(fmt, ...) \
563 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
565 # define DPRINTF(args) ((void) 0)
567 /** Print a debug string.
568 * The string is printed literally, with no format processing.
570 #define DPUTS(arg) DPRINTF(("%s", arg))
571 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
573 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
576 /** @brief The maximum size of a database page.
578 * It is 32k or 64k, since value-PAGEBASE must fit in
579 * #MDB_page.%mp_upper.
581 * LMDB will use database pages < OS pages if needed.
582 * That causes more I/O in write transactions: The OS must
583 * know (read) the whole page before writing a partial page.
585 * Note that we don't currently support Huge pages. On Linux,
586 * regular data files cannot use Huge pages, and in general
587 * Huge pages aren't actually pageable. We rely on the OS
588 * demand-pager to read our data and page it out when memory
589 * pressure from other processes is high. So until OSs have
590 * actual paging support for Huge pages, they're not viable.
592 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
594 /** The minimum number of keys required in a database page.
595 * Setting this to a larger value will place a smaller bound on the
596 * maximum size of a data item. Data items larger than this size will
597 * be pushed into overflow pages instead of being stored directly in
598 * the B-tree node. This value used to default to 4. With a page size
599 * of 4096 bytes that meant that any item larger than 1024 bytes would
600 * go into an overflow page. That also meant that on average 2-3KB of
601 * each overflow page was wasted space. The value cannot be lower than
602 * 2 because then there would no longer be a tree structure. With this
603 * value, items larger than 2KB will go into overflow pages, and on
604 * average only 1KB will be wasted.
606 #define MDB_MINKEYS 2
608 /** A stamp that identifies a file as an LMDB file.
609 * There's nothing special about this value other than that it is easily
610 * recognizable, and it will reflect any byte order mismatches.
612 #define MDB_MAGIC 0xBEEFC0DE
614 /** The version number for a database's datafile format. */
615 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
616 /** The version number for a database's lockfile format. */
617 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
619 /** @brief The max size of a key we can write, or 0 for computed max.
621 * This macro should normally be left alone or set to 0.
622 * Note that a database with big keys or dupsort data cannot be
623 * reliably modified by a liblmdb which uses a smaller max.
624 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
626 * Other values are allowed, for backwards compat. However:
627 * A value bigger than the computed max can break if you do not
628 * know what you are doing, and liblmdb <= 0.9.10 can break when
629 * modifying a DB with keys/dupsort data bigger than its max.
631 * Data items in an #MDB_DUPSORT database are also limited to
632 * this size, since they're actually keys of a sub-DB. Keys and
633 * #MDB_DUPSORT data items must fit on a node in a regular page.
635 #ifndef MDB_MAXKEYSIZE
636 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
639 /** The maximum size of a key we can write to the environment. */
641 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
643 #define ENV_MAXKEY(env) ((env)->me_maxkey)
646 /** @brief The maximum size of a data item.
648 * We only store a 32 bit value for node sizes.
650 #define MAXDATASIZE 0xffffffffUL
653 /** Key size which fits in a #DKBUF.
656 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
659 * This is used for printing a hex dump of a key's contents.
661 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
662 /** Display a key in hex.
664 * Invoke a function to display a key in hex.
666 #define DKEY(x) mdb_dkey(x, kbuf)
672 /** An invalid page number.
673 * Mainly used to denote an empty tree.
675 #define P_INVALID (~(pgno_t)0)
677 /** Test if the flags \b f are set in a flag word \b w. */
678 #define F_ISSET(w, f) (((w) & (f)) == (f))
680 /** Round \b n up to an even number. */
681 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
683 /** Used for offsets within a single page.
684 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
687 typedef uint16_t indx_t;
689 /** Default size of memory map.
690 * This is certainly too small for any actual applications. Apps should always set
691 * the size explicitly using #mdb_env_set_mapsize().
693 #define DEFAULT_MAPSIZE 1048576
695 /** @defgroup readers Reader Lock Table
696 * Readers don't acquire any locks for their data access. Instead, they
697 * simply record their transaction ID in the reader table. The reader
698 * mutex is needed just to find an empty slot in the reader table. The
699 * slot's address is saved in thread-specific data so that subsequent read
700 * transactions started by the same thread need no further locking to proceed.
702 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
704 * No reader table is used if the database is on a read-only filesystem, or
705 * if #MDB_NOLOCK is set.
707 * Since the database uses multi-version concurrency control, readers don't
708 * actually need any locking. This table is used to keep track of which
709 * readers are using data from which old transactions, so that we'll know
710 * when a particular old transaction is no longer in use. Old transactions
711 * that have discarded any data pages can then have those pages reclaimed
712 * for use by a later write transaction.
714 * The lock table is constructed such that reader slots are aligned with the
715 * processor's cache line size. Any slot is only ever used by one thread.
716 * This alignment guarantees that there will be no contention or cache
717 * thrashing as threads update their own slot info, and also eliminates
718 * any need for locking when accessing a slot.
720 * A writer thread will scan every slot in the table to determine the oldest
721 * outstanding reader transaction. Any freed pages older than this will be
722 * reclaimed by the writer. The writer doesn't use any locks when scanning
723 * this table. This means that there's no guarantee that the writer will
724 * see the most up-to-date reader info, but that's not required for correct
725 * operation - all we need is to know the upper bound on the oldest reader,
726 * we don't care at all about the newest reader. So the only consequence of
727 * reading stale information here is that old pages might hang around a
728 * while longer before being reclaimed. That's actually good anyway, because
729 * the longer we delay reclaiming old pages, the more likely it is that a
730 * string of contiguous pages can be found after coalescing old pages from
731 * many old transactions together.
734 /** Number of slots in the reader table.
735 * This value was chosen somewhat arbitrarily. 126 readers plus a
736 * couple mutexes fit exactly into 8KB on my development machine.
737 * Applications should set the table size using #mdb_env_set_maxreaders().
739 #define DEFAULT_READERS 126
741 /** The size of a CPU cache line in bytes. We want our lock structures
742 * aligned to this size to avoid false cache line sharing in the
744 * This value works for most CPUs. For Itanium this should be 128.
750 /** The information we store in a single slot of the reader table.
751 * In addition to a transaction ID, we also record the process and
752 * thread ID that owns a slot, so that we can detect stale information,
753 * e.g. threads or processes that went away without cleaning up.
754 * @note We currently don't check for stale records. We simply re-init
755 * the table when we know that we're the only process opening the
758 typedef struct MDB_rxbody {
759 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
760 * Multiple readers that start at the same time will probably have the
761 * same ID here. Again, it's not important to exclude them from
762 * anything; all we need to know is which version of the DB they
763 * started from so we can avoid overwriting any data used in that
764 * particular version.
766 volatile txnid_t mrb_txnid;
767 /** The process ID of the process owning this reader txn. */
768 volatile MDB_PID_T mrb_pid;
769 /** The thread ID of the thread owning this txn. */
770 volatile MDB_THR_T mrb_tid;
773 /** The actual reader record, with cacheline padding. */
774 typedef struct MDB_reader {
777 /** shorthand for mrb_txnid */
778 #define mr_txnid mru.mrx.mrb_txnid
779 #define mr_pid mru.mrx.mrb_pid
780 #define mr_tid mru.mrx.mrb_tid
781 /** cache line alignment */
782 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
786 /** The header for the reader table.
787 * The table resides in a memory-mapped file. (This is a different file
788 * than is used for the main database.)
790 * For POSIX the actual mutexes reside in the shared memory of this
791 * mapped file. On Windows, mutexes are named objects allocated by the
792 * kernel; we store the mutex names in this mapped file so that other
793 * processes can grab them. This same approach is also used on
794 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
795 * process-shared POSIX mutexes. For these cases where a named object
796 * is used, the object name is derived from a 64 bit FNV hash of the
797 * environment pathname. As such, naming collisions are extremely
798 * unlikely. If a collision occurs, the results are unpredictable.
800 typedef struct MDB_txbody {
801 /** Stamp identifying this as an LMDB file. It must be set
804 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
806 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
807 char mtb_rmname[MNAME_LEN];
808 #elif defined(MDB_USE_SYSV_SEM)
812 /** Mutex protecting access to this table.
813 * This is the reader table lock used with LOCK_MUTEX().
815 mdb_mutex_t mtb_rmutex;
817 /** The ID of the last transaction committed to the database.
818 * This is recorded here only for convenience; the value can always
819 * be determined by reading the main database meta pages.
821 volatile txnid_t mtb_txnid;
822 /** The number of slots that have been used in the reader table.
823 * This always records the maximum count, it is not decremented
824 * when readers release their slots.
826 volatile unsigned mtb_numreaders;
829 /** The actual reader table definition. */
830 typedef struct MDB_txninfo {
833 #define mti_magic mt1.mtb.mtb_magic
834 #define mti_format mt1.mtb.mtb_format
835 #define mti_rmutex mt1.mtb.mtb_rmutex
836 #define mti_rmname mt1.mtb.mtb_rmname
837 #define mti_txnid mt1.mtb.mtb_txnid
838 #define mti_numreaders mt1.mtb.mtb_numreaders
839 #ifdef MDB_USE_SYSV_SEM
840 #define mti_semid mt1.mtb.mtb_semid
841 #define mti_rlocked mt1.mtb.mtb_rlocked
843 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
846 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
847 char mt2_wmname[MNAME_LEN];
848 #define mti_wmname mt2.mt2_wmname
849 #elif defined MDB_USE_SYSV_SEM
851 #define mti_wlocked mt2.mt2_wlocked
853 mdb_mutex_t mt2_wmutex;
854 #define mti_wmutex mt2.mt2_wmutex
856 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
858 MDB_reader mti_readers[1];
861 /** Lockfile format signature: version, features and field layout */
862 #define MDB_LOCK_FORMAT \
864 ((MDB_LOCK_VERSION) \
865 /* Flags which describe functionality */ \
866 + (SYSV_SEM_FLAG << 18) \
867 + (((MDB_PIDLOCK) != 0) << 16)))
870 /** Common header for all page types.
871 * Overflow records occupy a number of contiguous pages with no
872 * headers on any page after the first.
874 typedef struct MDB_page {
875 #define mp_pgno mp_p.p_pgno
876 #define mp_next mp_p.p_next
878 pgno_t p_pgno; /**< page number */
879 struct MDB_page *p_next; /**< for in-memory list of freed pages */
882 /** @defgroup mdb_page Page Flags
884 * Flags for the page headers.
887 #define P_BRANCH 0x01 /**< branch page */
888 #define P_LEAF 0x02 /**< leaf page */
889 #define P_OVERFLOW 0x04 /**< overflow page */
890 #define P_META 0x08 /**< meta page */
891 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
892 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
893 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
894 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
895 #define P_KEEP 0x8000 /**< leave this page alone during spill */
897 uint16_t mp_flags; /**< @ref mdb_page */
898 #define mp_lower mp_pb.pb.pb_lower
899 #define mp_upper mp_pb.pb.pb_upper
900 #define mp_pages mp_pb.pb_pages
903 indx_t pb_lower; /**< lower bound of free space */
904 indx_t pb_upper; /**< upper bound of free space */
906 uint32_t pb_pages; /**< number of overflow pages */
908 indx_t mp_ptrs[1]; /**< dynamic size */
911 /** Size of the page header, excluding dynamic data at the end */
912 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
914 /** Address of first usable data byte in a page, after the header */
915 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
917 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
918 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
920 /** Number of nodes on a page */
921 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
923 /** The amount of space remaining in the page */
924 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
926 /** The percentage of space used in the page, in tenths of a percent. */
927 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
928 ((env)->me_psize - PAGEHDRSZ))
929 /** The minimum page fill factor, in tenths of a percent.
930 * Pages emptier than this are candidates for merging.
932 #define FILL_THRESHOLD 250
934 /** Test if a page is a leaf page */
935 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
936 /** Test if a page is a LEAF2 page */
937 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
938 /** Test if a page is a branch page */
939 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
940 /** Test if a page is an overflow page */
941 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
942 /** Test if a page is a sub page */
943 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
945 /** The number of overflow pages needed to store the given size. */
946 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
948 /** Link in #MDB_txn.%mt_loose_pgs list */
949 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
951 /** Header for a single key/data pair within a page.
952 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
953 * We guarantee 2-byte alignment for 'MDB_node's.
955 typedef struct MDB_node {
956 /** lo and hi are used for data size on leaf nodes and for
957 * child pgno on branch nodes. On 64 bit platforms, flags
958 * is also used for pgno. (Branch nodes have no flags).
959 * They are in host byte order in case that lets some
960 * accesses be optimized into a 32-bit word access.
962 #if BYTE_ORDER == LITTLE_ENDIAN
963 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
965 unsigned short mn_hi, mn_lo;
967 /** @defgroup mdb_node Node Flags
969 * Flags for node headers.
972 #define F_BIGDATA 0x01 /**< data put on overflow page */
973 #define F_SUBDATA 0x02 /**< data is a sub-database */
974 #define F_DUPDATA 0x04 /**< data has duplicates */
976 /** valid flags for #mdb_node_add() */
977 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
980 unsigned short mn_flags; /**< @ref mdb_node */
981 unsigned short mn_ksize; /**< key size */
982 char mn_data[1]; /**< key and data are appended here */
985 /** Size of the node header, excluding dynamic data at the end */
986 #define NODESIZE offsetof(MDB_node, mn_data)
988 /** Bit position of top word in page number, for shifting mn_flags */
989 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
991 /** Size of a node in a branch page with a given key.
992 * This is just the node header plus the key, there is no data.
994 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
996 /** Size of a node in a leaf page with a given key and data.
997 * This is node header plus key plus data size.
999 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
1001 /** Address of node \b i in page \b p */
1002 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
1004 /** Address of the key for the node */
1005 #define NODEKEY(node) (void *)((node)->mn_data)
1007 /** Address of the data for a node */
1008 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
1010 /** Get the page number pointed to by a branch node */
1011 #define NODEPGNO(node) \
1012 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
1013 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
1014 /** Set the page number in a branch node */
1015 #define SETPGNO(node,pgno) do { \
1016 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
1017 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
1019 /** Get the size of the data in a leaf node */
1020 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
1021 /** Set the size of the data for a leaf node */
1022 #define SETDSZ(node,size) do { \
1023 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
1024 /** The size of a key in a node */
1025 #define NODEKSZ(node) ((node)->mn_ksize)
1027 /** Copy a page number from src to dst */
1028 #ifdef MISALIGNED_OK
1029 #define COPY_PGNO(dst,src) dst = src
1031 #if SIZE_MAX > 4294967295UL
1032 #define COPY_PGNO(dst,src) do { \
1033 unsigned short *s, *d; \
1034 s = (unsigned short *)&(src); \
1035 d = (unsigned short *)&(dst); \
1042 #define COPY_PGNO(dst,src) do { \
1043 unsigned short *s, *d; \
1044 s = (unsigned short *)&(src); \
1045 d = (unsigned short *)&(dst); \
1051 /** The address of a key in a LEAF2 page.
1052 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
1053 * There are no node headers, keys are stored contiguously.
1055 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
1057 /** Set the \b node's key into \b keyptr, if requested. */
1058 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
1059 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
1061 /** Set the \b node's key into \b key. */
1062 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
1064 /** Information about a single database in the environment. */
1065 typedef struct MDB_db {
1066 uint32_t md_pad; /**< also ksize for LEAF2 pages */
1067 uint16_t md_flags; /**< @ref mdb_dbi_open */
1068 uint16_t md_depth; /**< depth of this tree */
1069 pgno_t md_branch_pages; /**< number of internal pages */
1070 pgno_t md_leaf_pages; /**< number of leaf pages */
1071 pgno_t md_overflow_pages; /**< number of overflow pages */
1072 mdb_size_t md_entries; /**< number of data items */
1073 pgno_t md_root; /**< the root page of this tree */
1076 /** mdb_dbi_open flags */
1077 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1078 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1079 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1080 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1082 /** Handle for the DB used to track free pages. */
1084 /** Handle for the default DB. */
1086 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1089 /** Number of meta pages - also hardcoded elsewhere */
1092 /** Meta page content.
1093 * A meta page is the start point for accessing a database snapshot.
1094 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1096 typedef struct MDB_meta {
1097 /** Stamp identifying this as an LMDB file. It must be set
1100 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1101 uint32_t mm_version;
1103 union { /* always zero since we don't support fixed mapping in MDB_VL32 */
1107 #define mm_address mm_un.mmun_address
1109 void *mm_address; /**< address for fixed mapping */
1111 pgno_t mm_mapsize; /**< size of mmap region */
1112 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1113 /** The size of pages used in this DB */
1114 #define mm_psize mm_dbs[FREE_DBI].md_pad
1115 /** Any persistent environment flags. @ref mdb_env */
1116 #define mm_flags mm_dbs[FREE_DBI].md_flags
1117 pgno_t mm_last_pg; /**< last used page in file */
1118 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1121 /** Buffer for a stack-allocated meta page.
1122 * The members define size and alignment, and silence type
1123 * aliasing warnings. They are not used directly; that could
1124 * mean incorrectly using several union members in parallel.
1126 typedef union MDB_metabuf {
1129 char mm_pad[PAGEHDRSZ];
1134 /** Auxiliary DB info.
1135 * The information here is mostly static/read-only. There is
1136 * only a single copy of this record in the environment.
1138 typedef struct MDB_dbx {
1139 MDB_val md_name; /**< name of the database */
1140 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1141 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1142 MDB_rel_func *md_rel; /**< user relocate function */
1143 void *md_relctx; /**< user-provided context for md_rel */
1146 /** A database transaction.
1147 * Every operation requires a transaction handle.
1150 MDB_txn *mt_parent; /**< parent of a nested txn */
1151 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1153 pgno_t mt_next_pgno; /**< next unallocated page */
1155 pgno_t mt_last_pgno; /**< last written page */
1157 /** The ID of this transaction. IDs are integers incrementing from 1.
1158 * Only committed write transactions increment the ID. If a transaction
1159 * aborts, the ID may be re-used by the next writer.
1162 MDB_env *mt_env; /**< the DB environment */
1163 /** The list of pages that became unused during this transaction.
1165 MDB_IDL mt_free_pgs;
1166 /** The list of loose pages that became unused and may be reused
1167 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1169 MDB_page *mt_loose_pgs;
1170 /* #Number of loose pages (#mt_loose_pgs) */
1172 /** The sorted list of dirty pages we temporarily wrote to disk
1173 * because the dirty list was full. page numbers in here are
1174 * shifted left by 1, deleted slots have the LSB set.
1176 MDB_IDL mt_spill_pgs;
1178 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1179 MDB_ID2L dirty_list;
1180 /** For read txns: This thread/txn's reader table slot, or NULL. */
1183 /** Array of records for each DB known in the environment. */
1185 /** Array of MDB_db records for each known DB */
1187 /** Array of sequence numbers for each DB handle */
1188 unsigned int *mt_dbiseqs;
1189 /** @defgroup mt_dbflag Transaction DB Flags
1193 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1194 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1195 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1196 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1197 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1199 /** In write txns, array of cursors for each DB */
1200 MDB_cursor **mt_cursors;
1201 /** Array of flags for each DB */
1202 unsigned char *mt_dbflags;
1204 /** List of read-only pages (actually chunks) */
1206 /** We map chunks of 16 pages. Even though Windows uses 4KB pages, all
1207 * mappings must begin on 64KB boundaries. So we round off all pgnos to
1208 * a chunk boundary. We do the same on Linux for symmetry, and also to
1209 * reduce the frequency of mmap/munmap calls.
1211 #define MDB_RPAGE_CHUNK 16
1212 #define MDB_TRPAGE_SIZE 4096 /**< size of #mt_rpages array of chunks */
1213 #define MDB_TRPAGE_MAX (MDB_TRPAGE_SIZE-1) /**< maximum chunk index */
1214 unsigned int mt_rpcheck; /**< threshold for reclaiming unref'd chunks */
1216 /** Number of DB records in use, or 0 when the txn is finished.
1217 * This number only ever increments until the txn finishes; we
1218 * don't decrement it when individual DB handles are closed.
1222 /** @defgroup mdb_txn Transaction Flags
1226 /** #mdb_txn_begin() flags */
1227 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1228 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1229 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1230 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1231 /* internal txn flags */
1232 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1233 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1234 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1235 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1236 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1237 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1238 /** most operations on the txn are currently illegal */
1239 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1241 unsigned int mt_flags; /**< @ref mdb_txn */
1242 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1243 * Includes ancestor txns' dirty pages not hidden by other txns'
1244 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1245 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1247 unsigned int mt_dirty_room;
1250 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1251 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1252 * raise this on a 64 bit machine.
1254 #define CURSOR_STACK 32
1258 /** Cursors are used for all DB operations.
1259 * A cursor holds a path of (page pointer, key index) from the DB
1260 * root to a position in the DB, plus other state. #MDB_DUPSORT
1261 * cursors include an xcursor to the current data item. Write txns
1262 * track their cursors and keep them up to date when data moves.
1263 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1264 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1267 /** Next cursor on this DB in this txn */
1268 MDB_cursor *mc_next;
1269 /** Backup of the original cursor if this cursor is a shadow */
1270 MDB_cursor *mc_backup;
1271 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1272 struct MDB_xcursor *mc_xcursor;
1273 /** The transaction that owns this cursor */
1275 /** The database handle this cursor operates on */
1277 /** The database record for this cursor */
1279 /** The database auxiliary record for this cursor */
1281 /** The @ref mt_dbflag for this database */
1282 unsigned char *mc_dbflag;
1283 unsigned short mc_snum; /**< number of pushed pages */
1284 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1285 /** @defgroup mdb_cursor Cursor Flags
1287 * Cursor state flags.
1290 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1291 #define C_EOF 0x02 /**< No more data */
1292 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1293 #define C_DEL 0x08 /**< last op was a cursor_del */
1294 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1295 #define C_WRITEMAP MDB_TXN_WRITEMAP /**< Copy of txn flag */
1296 /** Read-only cursor into the txn's original snapshot in the map.
1297 * Set for read-only txns, and in #mdb_page_alloc() for #FREE_DBI when
1298 * #MDB_DEVEL & 2. Only implements code which is necessary for this.
1300 #define C_ORIG_RDONLY MDB_TXN_RDONLY
1302 unsigned int mc_flags; /**< @ref mdb_cursor */
1303 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1304 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1306 MDB_page *mc_ovpg; /**< a referenced overflow page */
1310 /** Context for sorted-dup records.
1311 * We could have gone to a fully recursive design, with arbitrarily
1312 * deep nesting of sub-databases. But for now we only handle these
1313 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1315 typedef struct MDB_xcursor {
1316 /** A sub-cursor for traversing the Dup DB */
1317 MDB_cursor mx_cursor;
1318 /** The database record for this Dup DB */
1320 /** The auxiliary DB record for this Dup DB */
1322 /** The @ref mt_dbflag for this Dup DB */
1323 unsigned char mx_dbflag;
1326 /** State of FreeDB old pages, stored in the MDB_env */
1327 typedef struct MDB_pgstate {
1328 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1329 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1332 /** The database environment. */
1334 HANDLE me_fd; /**< The main data file */
1335 HANDLE me_lfd; /**< The lock file */
1336 HANDLE me_mfd; /**< just for writing the meta pages */
1337 #if defined(MDB_VL32) && defined(_WIN32)
1338 HANDLE me_fmh; /**< File Mapping handle */
1340 /** Failed to update the meta page. Probably an I/O error. */
1341 #define MDB_FATAL_ERROR 0x80000000U
1342 /** Some fields are initialized. */
1343 #define MDB_ENV_ACTIVE 0x20000000U
1344 /** me_txkey is set */
1345 #define MDB_ENV_TXKEY 0x10000000U
1346 /** fdatasync is unreliable */
1347 #define MDB_FSYNCONLY 0x08000000U
1348 uint32_t me_flags; /**< @ref mdb_env */
1349 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1350 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1351 unsigned int me_maxreaders; /**< size of the reader table */
1352 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1353 volatile int me_close_readers;
1354 MDB_dbi me_numdbs; /**< number of DBs opened */
1355 MDB_dbi me_maxdbs; /**< size of the DB table */
1356 MDB_PID_T me_pid; /**< process ID of this env */
1357 char *me_path; /**< path to the DB files */
1358 char *me_map; /**< the memory map of the data file */
1359 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1360 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1361 void *me_pbuf; /**< scratch area for DUPSORT put() */
1362 MDB_txn *me_txn; /**< current write transaction */
1363 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1364 mdb_size_t me_mapsize; /**< size of the data memory map */
1365 off_t me_size; /**< current file size */
1366 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1367 MDB_dbx *me_dbxs; /**< array of static DB info */
1368 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1369 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1370 pthread_key_t me_txkey; /**< thread-key for readers */
1371 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1372 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1373 # define me_pglast me_pgstate.mf_pglast
1374 # define me_pghead me_pgstate.mf_pghead
1375 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1376 /** IDL of pages that became unused in a write txn */
1377 MDB_IDL me_free_pgs;
1378 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1379 MDB_ID2L me_dirty_list;
1380 /** Max number of freelist items that can fit in a single overflow page */
1382 /** Max size of a node on a page */
1383 unsigned int me_nodemax;
1384 #if !(MDB_MAXKEYSIZE)
1385 unsigned int me_maxkey; /**< max size of a key */
1387 int me_live_reader; /**< have liveness lock in reader table */
1389 int me_pidquery; /**< Used in OpenProcess */
1391 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1392 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1393 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1395 mdb_mutex_t me_rmutex;
1396 mdb_mutex_t me_wmutex;
1399 MDB_ID3L me_rpages; /**< like #mt_rpages, but global to env */
1400 pthread_mutex_t me_rpmutex; /**< control access to #me_rpages */
1401 #define MDB_ERPAGE_SIZE 16384
1402 #define MDB_ERPAGE_MAX (MDB_ERPAGE_SIZE-1)
1403 unsigned int me_rpcheck;
1405 void *me_userctx; /**< User-settable context */
1406 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1409 /** Nested transaction */
1410 typedef struct MDB_ntxn {
1411 MDB_txn mnt_txn; /**< the transaction */
1412 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1415 /** max number of pages to commit in one writev() call */
1416 #define MDB_COMMIT_PAGES 64
1417 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1418 #undef MDB_COMMIT_PAGES
1419 #define MDB_COMMIT_PAGES IOV_MAX
1422 /** max bytes to write in one call */
1423 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1425 /** Check \b txn and \b dbi arguments to a function */
1426 #define TXN_DBI_EXIST(txn, dbi, validity) \
1427 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1429 /** Check for misused \b dbi handles */
1430 #define TXN_DBI_CHANGED(txn, dbi) \
1431 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1433 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1434 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1435 static int mdb_page_touch(MDB_cursor *mc);
1437 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1438 "reset-tmp", "fail-begin", "fail-beginchild"}
1440 /* mdb_txn_end operation number, for logging */
1441 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1442 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1444 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1445 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1446 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1447 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1448 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1450 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1451 static int mdb_page_search_root(MDB_cursor *mc,
1452 MDB_val *key, int modify);
1453 #define MDB_PS_MODIFY 1
1454 #define MDB_PS_ROOTONLY 2
1455 #define MDB_PS_FIRST 4
1456 #define MDB_PS_LAST 8
1457 static int mdb_page_search(MDB_cursor *mc,
1458 MDB_val *key, int flags);
1459 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1461 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1462 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1463 pgno_t newpgno, unsigned int nflags);
1465 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1466 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1467 static int mdb_env_write_meta(MDB_txn *txn);
1468 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1469 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1471 static void mdb_env_close0(MDB_env *env, int excl);
1473 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1474 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1475 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1476 static void mdb_node_del(MDB_cursor *mc, int ksize);
1477 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1478 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1479 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1480 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1481 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1483 static int mdb_rebalance(MDB_cursor *mc);
1484 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1486 static void mdb_cursor_pop(MDB_cursor *mc);
1487 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1489 static int mdb_cursor_del0(MDB_cursor *mc);
1490 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1491 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1492 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1493 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1494 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1496 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1497 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1499 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1500 static void mdb_xcursor_init0(MDB_cursor *mc);
1501 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1502 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1504 static int mdb_drop0(MDB_cursor *mc, int subs);
1505 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1506 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1509 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1512 /** Compare two items pointing at size_t's of unknown alignment. */
1513 #ifdef MISALIGNED_OK
1514 # define mdb_cmp_clong mdb_cmp_long
1516 # define mdb_cmp_clong mdb_cmp_cint
1520 static SECURITY_DESCRIPTOR mdb_null_sd;
1521 static SECURITY_ATTRIBUTES mdb_all_sa;
1522 static int mdb_sec_inited;
1524 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize);
1527 /** Return the library version info. */
1529 mdb_version(int *major, int *minor, int *patch)
1531 if (major) *major = MDB_VERSION_MAJOR;
1532 if (minor) *minor = MDB_VERSION_MINOR;
1533 if (patch) *patch = MDB_VERSION_PATCH;
1534 return MDB_VERSION_STRING;
1537 /** Table of descriptions for LMDB @ref errors */
1538 static char *const mdb_errstr[] = {
1539 "MDB_KEYEXIST: Key/data pair already exists",
1540 "MDB_NOTFOUND: No matching key/data pair found",
1541 "MDB_PAGE_NOTFOUND: Requested page not found",
1542 "MDB_CORRUPTED: Located page was wrong type",
1543 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1544 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1545 "MDB_INVALID: File is not an LMDB file",
1546 "MDB_MAP_FULL: Environment mapsize limit reached",
1547 "MDB_DBS_FULL: Environment maxdbs limit reached",
1548 "MDB_READERS_FULL: Environment maxreaders limit reached",
1549 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1550 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1551 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1552 "MDB_PAGE_FULL: Internal error - page has no more space",
1553 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1554 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1555 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1556 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1557 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1558 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1559 "MDB_PROBLEM: Unexpected problem - txn should abort",
1563 mdb_strerror(int err)
1566 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1567 * This works as long as no function between the call to mdb_strerror
1568 * and the actual use of the message uses more than 4K of stack.
1570 #define MSGSIZE 1024
1571 #define PADSIZE 4096
1572 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1576 return ("Successful return: 0");
1578 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1579 i = err - MDB_KEYEXIST;
1580 return mdb_errstr[i];
1584 /* These are the C-runtime error codes we use. The comment indicates
1585 * their numeric value, and the Win32 error they would correspond to
1586 * if the error actually came from a Win32 API. A major mess, we should
1587 * have used LMDB-specific error codes for everything.
1590 case ENOENT: /* 2, FILE_NOT_FOUND */
1591 case EIO: /* 5, ACCESS_DENIED */
1592 case ENOMEM: /* 12, INVALID_ACCESS */
1593 case EACCES: /* 13, INVALID_DATA */
1594 case EBUSY: /* 16, CURRENT_DIRECTORY */
1595 case EINVAL: /* 22, BAD_COMMAND */
1596 case ENOSPC: /* 28, OUT_OF_PAPER */
1597 return strerror(err);
1602 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1603 FORMAT_MESSAGE_IGNORE_INSERTS,
1604 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1607 return strerror(err);
1611 /** assert(3) variant in cursor context */
1612 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1613 /** assert(3) variant in transaction context */
1614 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1615 /** assert(3) variant in environment context */
1616 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1619 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1620 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1623 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1624 const char *func, const char *file, int line)
1627 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1628 file, line, expr_txt, func);
1629 if (env->me_assert_func)
1630 env->me_assert_func(env, buf);
1631 fprintf(stderr, "%s\n", buf);
1635 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1639 /** Return the page number of \b mp which may be sub-page, for debug output */
1641 mdb_dbg_pgno(MDB_page *mp)
1644 COPY_PGNO(ret, mp->mp_pgno);
1648 /** Display a key in hexadecimal and return the address of the result.
1649 * @param[in] key the key to display
1650 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1651 * @return The key in hexadecimal form.
1654 mdb_dkey(MDB_val *key, char *buf)
1657 unsigned char *c = key->mv_data;
1663 if (key->mv_size > DKBUF_MAXKEYSIZE)
1664 return "MDB_MAXKEYSIZE";
1665 /* may want to make this a dynamic check: if the key is mostly
1666 * printable characters, print it as-is instead of converting to hex.
1670 for (i=0; i<key->mv_size; i++)
1671 ptr += sprintf(ptr, "%02x", *c++);
1673 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1679 mdb_leafnode_type(MDB_node *n)
1681 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1682 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1683 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1686 /** Display all the keys in the page. */
1688 mdb_page_list(MDB_page *mp)
1690 pgno_t pgno = mdb_dbg_pgno(mp);
1691 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1693 unsigned int i, nkeys, nsize, total = 0;
1697 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1698 case P_BRANCH: type = "Branch page"; break;
1699 case P_LEAF: type = "Leaf page"; break;
1700 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1701 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1702 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1704 fprintf(stderr, "Overflow page %"Y"u pages %u%s\n",
1705 pgno, mp->mp_pages, state);
1708 fprintf(stderr, "Meta-page %"Y"u txnid %"Y"u\n",
1709 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1712 fprintf(stderr, "Bad page %"Y"u flags 0x%u\n", pgno, mp->mp_flags);
1716 nkeys = NUMKEYS(mp);
1717 fprintf(stderr, "%s %"Y"u numkeys %d%s\n", type, pgno, nkeys, state);
1719 for (i=0; i<nkeys; i++) {
1720 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1721 key.mv_size = nsize = mp->mp_pad;
1722 key.mv_data = LEAF2KEY(mp, i, nsize);
1724 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1727 node = NODEPTR(mp, i);
1728 key.mv_size = node->mn_ksize;
1729 key.mv_data = node->mn_data;
1730 nsize = NODESIZE + key.mv_size;
1731 if (IS_BRANCH(mp)) {
1732 fprintf(stderr, "key %d: page %"Y"u, %s\n", i, NODEPGNO(node),
1736 if (F_ISSET(node->mn_flags, F_BIGDATA))
1737 nsize += sizeof(pgno_t);
1739 nsize += NODEDSZ(node);
1741 nsize += sizeof(indx_t);
1742 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1743 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1745 total = EVEN(total);
1747 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1748 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1752 mdb_cursor_chk(MDB_cursor *mc)
1758 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1759 for (i=0; i<mc->mc_top; i++) {
1761 node = NODEPTR(mp, mc->mc_ki[i]);
1762 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1765 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1767 if (mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
1768 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1769 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1770 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1778 /** Count all the pages in each DB and in the freelist
1779 * and make sure it matches the actual number of pages
1781 * All named DBs must be open for a correct count.
1783 static void mdb_audit(MDB_txn *txn)
1787 MDB_ID freecount, count;
1792 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1793 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1794 freecount += *(MDB_ID *)data.mv_data;
1795 mdb_tassert(txn, rc == MDB_NOTFOUND);
1798 for (i = 0; i<txn->mt_numdbs; i++) {
1800 if (!(txn->mt_dbflags[i] & DB_VALID))
1802 mdb_cursor_init(&mc, txn, i, &mx);
1803 if (txn->mt_dbs[i].md_root == P_INVALID)
1805 count += txn->mt_dbs[i].md_branch_pages +
1806 txn->mt_dbs[i].md_leaf_pages +
1807 txn->mt_dbs[i].md_overflow_pages;
1808 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1809 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1810 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1813 mp = mc.mc_pg[mc.mc_top];
1814 for (j=0; j<NUMKEYS(mp); j++) {
1815 MDB_node *leaf = NODEPTR(mp, j);
1816 if (leaf->mn_flags & F_SUBDATA) {
1818 memcpy(&db, NODEDATA(leaf), sizeof(db));
1819 count += db.md_branch_pages + db.md_leaf_pages +
1820 db.md_overflow_pages;
1824 mdb_tassert(txn, rc == MDB_NOTFOUND);
1827 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1828 fprintf(stderr, "audit: %"Y"u freecount: %"Y"u count: %"Y"u total: %"Y"u next_pgno: %"Y"u\n",
1829 txn->mt_txnid, freecount, count+NUM_METAS,
1830 freecount+count+NUM_METAS, txn->mt_next_pgno);
1836 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1838 return txn->mt_dbxs[dbi].md_cmp(a, b);
1842 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1844 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1845 #if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
1846 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(mdb_size_t))
1847 dcmp = mdb_cmp_clong;
1852 /** Allocate memory for a page.
1853 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1856 mdb_page_malloc(MDB_txn *txn, unsigned num)
1858 MDB_env *env = txn->mt_env;
1859 MDB_page *ret = env->me_dpages;
1860 size_t psize = env->me_psize, sz = psize, off;
1861 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1862 * For a single page alloc, we init everything after the page header.
1863 * For multi-page, we init the final page; if the caller needed that
1864 * many pages they will be filling in at least up to the last page.
1868 VGMEMP_ALLOC(env, ret, sz);
1869 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1870 env->me_dpages = ret->mp_next;
1873 psize -= off = PAGEHDRSZ;
1878 if ((ret = malloc(sz)) != NULL) {
1879 VGMEMP_ALLOC(env, ret, sz);
1880 if (!(env->me_flags & MDB_NOMEMINIT)) {
1881 memset((char *)ret + off, 0, psize);
1885 txn->mt_flags |= MDB_TXN_ERROR;
1889 /** Free a single page.
1890 * Saves single pages to a list, for future reuse.
1891 * (This is not used for multi-page overflow pages.)
1894 mdb_page_free(MDB_env *env, MDB_page *mp)
1896 mp->mp_next = env->me_dpages;
1897 VGMEMP_FREE(env, mp);
1898 env->me_dpages = mp;
1901 /** Free a dirty page */
1903 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1905 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1906 mdb_page_free(env, dp);
1908 /* large pages just get freed directly */
1909 VGMEMP_FREE(env, dp);
1914 /** Return all dirty pages to dpage list */
1916 mdb_dlist_free(MDB_txn *txn)
1918 MDB_env *env = txn->mt_env;
1919 MDB_ID2L dl = txn->mt_u.dirty_list;
1920 unsigned i, n = dl[0].mid;
1922 for (i = 1; i <= n; i++) {
1923 mdb_dpage_free(env, dl[i].mptr);
1930 mdb_page_unref(MDB_txn *txn, MDB_page *mp)
1933 MDB_ID3L tl = txn->mt_rpages;
1935 if (mp->mp_flags & (P_SUBP|P_DIRTY))
1937 rem = mp->mp_pgno & (MDB_RPAGE_CHUNK-1);
1938 pgno = mp->mp_pgno ^ rem;
1939 x = mdb_mid3l_search(tl, pgno);
1940 if (x != tl[0].mid && tl[x+1].mid == mp->mp_pgno)
1945 #define MDB_PAGE_UNREF(txn, mp) mdb_page_unref(txn, mp)
1948 mdb_cursor_unref(MDB_cursor *mc)
1951 if (!mc->mc_snum || !mc->mc_pg[0] || IS_SUBP(mc->mc_pg[0]))
1953 for (i=0; i<mc->mc_snum; i++)
1954 mdb_page_unref(mc->mc_txn, mc->mc_pg[i]);
1956 mdb_page_unref(mc->mc_txn, mc->mc_ovpg);
1959 mc->mc_snum = mc->mc_top = 0;
1960 mc->mc_pg[0] = NULL;
1961 mc->mc_flags &= ~C_INITIALIZED;
1964 #define MDB_PAGE_UNREF(txn, mp)
1965 #endif /* MDB_VL32 */
1967 /** Loosen or free a single page.
1968 * Saves single pages to a list for future reuse
1969 * in this same txn. It has been pulled from the freeDB
1970 * and already resides on the dirty list, but has been
1971 * deleted. Use these pages first before pulling again
1974 * If the page wasn't dirtied in this txn, just add it
1975 * to this txn's free list.
1978 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1981 pgno_t pgno = mp->mp_pgno;
1982 MDB_txn *txn = mc->mc_txn;
1984 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1985 if (txn->mt_parent) {
1986 MDB_ID2 *dl = txn->mt_u.dirty_list;
1987 /* If txn has a parent, make sure the page is in our
1991 unsigned x = mdb_mid2l_search(dl, pgno);
1992 if (x <= dl[0].mid && dl[x].mid == pgno) {
1993 if (mp != dl[x].mptr) { /* bad cursor? */
1994 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1995 txn->mt_flags |= MDB_TXN_ERROR;
2003 /* no parent txn, so it's just ours */
2008 DPRINTF(("loosen db %d page %"Y"u", DDBI(mc),
2010 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
2011 txn->mt_loose_pgs = mp;
2012 txn->mt_loose_count++;
2013 mp->mp_flags |= P_LOOSE;
2015 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
2023 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
2024 * @param[in] mc A cursor handle for the current operation.
2025 * @param[in] pflags Flags of the pages to update:
2026 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
2027 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
2028 * @return 0 on success, non-zero on failure.
2031 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
2033 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
2034 MDB_txn *txn = mc->mc_txn;
2035 MDB_cursor *m3, *m0 = mc;
2040 int rc = MDB_SUCCESS, level;
2042 /* Mark pages seen by cursors */
2043 if (mc->mc_flags & C_UNTRACK)
2044 mc = NULL; /* will find mc in mt_cursors */
2045 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
2046 for (; mc; mc=mc->mc_next) {
2047 if (!(mc->mc_flags & C_INITIALIZED))
2049 for (m3 = mc;; m3 = &mx->mx_cursor) {
2051 for (j=0; j<m3->mc_snum; j++) {
2053 if ((mp->mp_flags & Mask) == pflags)
2054 mp->mp_flags ^= P_KEEP;
2056 mx = m3->mc_xcursor;
2057 /* Proceed to mx if it is at a sub-database */
2058 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
2060 if (! (mp && (mp->mp_flags & P_LEAF)))
2062 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
2063 if (!(leaf->mn_flags & F_SUBDATA))
2072 /* Mark dirty root pages */
2073 for (i=0; i<txn->mt_numdbs; i++) {
2074 if (txn->mt_dbflags[i] & DB_DIRTY) {
2075 pgno_t pgno = txn->mt_dbs[i].md_root;
2076 if (pgno == P_INVALID)
2078 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
2080 if ((dp->mp_flags & Mask) == pflags && level <= 1)
2081 dp->mp_flags ^= P_KEEP;
2089 static int mdb_page_flush(MDB_txn *txn, int keep);
2091 /** Spill pages from the dirty list back to disk.
2092 * This is intended to prevent running into #MDB_TXN_FULL situations,
2093 * but note that they may still occur in a few cases:
2094 * 1) our estimate of the txn size could be too small. Currently this
2095 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
2096 * 2) child txns may run out of space if their parents dirtied a
2097 * lot of pages and never spilled them. TODO: we probably should do
2098 * a preemptive spill during #mdb_txn_begin() of a child txn, if
2099 * the parent's dirty_room is below a given threshold.
2101 * Otherwise, if not using nested txns, it is expected that apps will
2102 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
2103 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
2104 * If the txn never references them again, they can be left alone.
2105 * If the txn only reads them, they can be used without any fuss.
2106 * If the txn writes them again, they can be dirtied immediately without
2107 * going thru all of the work of #mdb_page_touch(). Such references are
2108 * handled by #mdb_page_unspill().
2110 * Also note, we never spill DB root pages, nor pages of active cursors,
2111 * because we'll need these back again soon anyway. And in nested txns,
2112 * we can't spill a page in a child txn if it was already spilled in a
2113 * parent txn. That would alter the parent txns' data even though
2114 * the child hasn't committed yet, and we'd have no way to undo it if
2115 * the child aborted.
2117 * @param[in] m0 cursor A cursor handle identifying the transaction and
2118 * database for which we are checking space.
2119 * @param[in] key For a put operation, the key being stored.
2120 * @param[in] data For a put operation, the data being stored.
2121 * @return 0 on success, non-zero on failure.
2124 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
2126 MDB_txn *txn = m0->mc_txn;
2128 MDB_ID2L dl = txn->mt_u.dirty_list;
2129 unsigned int i, j, need;
2132 if (m0->mc_flags & C_SUB)
2135 /* Estimate how much space this op will take */
2136 i = m0->mc_db->md_depth;
2137 /* Named DBs also dirty the main DB */
2138 if (m0->mc_dbi >= CORE_DBS)
2139 i += txn->mt_dbs[MAIN_DBI].md_depth;
2140 /* For puts, roughly factor in the key+data size */
2142 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2143 i += i; /* double it for good measure */
2146 if (txn->mt_dirty_room > i)
2149 if (!txn->mt_spill_pgs) {
2150 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2151 if (!txn->mt_spill_pgs)
2154 /* purge deleted slots */
2155 MDB_IDL sl = txn->mt_spill_pgs;
2156 unsigned int num = sl[0];
2158 for (i=1; i<=num; i++) {
2165 /* Preserve pages which may soon be dirtied again */
2166 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2169 /* Less aggressive spill - we originally spilled the entire dirty list,
2170 * with a few exceptions for cursor pages and DB root pages. But this
2171 * turns out to be a lot of wasted effort because in a large txn many
2172 * of those pages will need to be used again. So now we spill only 1/8th
2173 * of the dirty pages. Testing revealed this to be a good tradeoff,
2174 * better than 1/2, 1/4, or 1/10.
2176 if (need < MDB_IDL_UM_MAX / 8)
2177 need = MDB_IDL_UM_MAX / 8;
2179 /* Save the page IDs of all the pages we're flushing */
2180 /* flush from the tail forward, this saves a lot of shifting later on. */
2181 for (i=dl[0].mid; i && need; i--) {
2182 MDB_ID pn = dl[i].mid << 1;
2184 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2186 /* Can't spill twice, make sure it's not already in a parent's
2189 if (txn->mt_parent) {
2191 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2192 if (tx2->mt_spill_pgs) {
2193 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2194 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2195 dp->mp_flags |= P_KEEP;
2203 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2207 mdb_midl_sort(txn->mt_spill_pgs);
2209 /* Flush the spilled part of dirty list */
2210 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2213 /* Reset any dirty pages we kept that page_flush didn't see */
2214 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2217 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2221 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2223 mdb_find_oldest(MDB_txn *txn)
2226 txnid_t mr, oldest = txn->mt_txnid - 1;
2227 if (txn->mt_env->me_txns) {
2228 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2229 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2240 /** Add a page to the txn's dirty list */
2242 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2245 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2247 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2248 insert = mdb_mid2l_append;
2250 insert = mdb_mid2l_insert;
2252 mid.mid = mp->mp_pgno;
2254 rc = insert(txn->mt_u.dirty_list, &mid);
2255 mdb_tassert(txn, rc == 0);
2256 txn->mt_dirty_room--;
2259 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2260 * me_pghead and mt_next_pgno.
2262 * If there are free pages available from older transactions, they
2263 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2264 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2265 * and move me_pglast to say which records were consumed. Only this
2266 * function can create me_pghead and move me_pglast/mt_next_pgno.
2267 * When #MDB_DEVEL & 2, it is not affected by #mdb_freelist_save(): it
2268 * then uses the transaction's original snapshot of the freeDB.
2269 * @param[in] mc cursor A cursor handle identifying the transaction and
2270 * database for which we are allocating.
2271 * @param[in] num the number of pages to allocate.
2272 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2273 * will always be satisfied by a single contiguous chunk of memory.
2274 * @return 0 on success, non-zero on failure.
2277 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2279 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2280 /* Get at most <Max_retries> more freeDB records once me_pghead
2281 * has enough pages. If not enough, use new pages from the map.
2282 * If <Paranoid> and mc is updating the freeDB, only get new
2283 * records if me_pghead is empty. Then the freelist cannot play
2284 * catch-up with itself by growing while trying to save it.
2286 enum { Paranoid = 1, Max_retries = 500 };
2288 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2290 int rc, retry = num * 60;
2291 MDB_txn *txn = mc->mc_txn;
2292 MDB_env *env = txn->mt_env;
2293 pgno_t pgno, *mop = env->me_pghead;
2294 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2296 txnid_t oldest = 0, last;
2301 /* If there are any loose pages, just use them */
2302 if (num == 1 && txn->mt_loose_pgs) {
2303 np = txn->mt_loose_pgs;
2304 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2305 txn->mt_loose_count--;
2306 DPRINTF(("db %d use loose page %"Y"u", DDBI(mc),
2314 /* If our dirty list is already full, we can't do anything */
2315 if (txn->mt_dirty_room == 0) {
2320 for (op = MDB_FIRST;; op = MDB_NEXT) {
2325 /* Seek a big enough contiguous page range. Prefer
2326 * pages at the tail, just truncating the list.
2332 if (mop[i-n2] == pgno+n2)
2339 if (op == MDB_FIRST) { /* 1st iteration */
2340 /* Prepare to fetch more and coalesce */
2341 last = env->me_pglast;
2342 oldest = env->me_pgoldest;
2343 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2344 #if (MDB_DEVEL) & 2 /* "& 2" so MDB_DEVEL=1 won't hide bugs breaking freeDB */
2345 /* Use original snapshot. TODO: Should need less care in code
2346 * which modifies the database. Maybe we can delete some code?
2348 m2.mc_flags |= C_ORIG_RDONLY;
2349 m2.mc_db = &env->me_metas[(txn->mt_txnid-1) & 1]->mm_dbs[FREE_DBI];
2350 m2.mc_dbflag = (unsigned char *)""; /* probably unnecessary */
2354 key.mv_data = &last; /* will look up last+1 */
2355 key.mv_size = sizeof(last);
2357 if (Paranoid && mc->mc_dbi == FREE_DBI)
2360 if (Paranoid && retry < 0 && mop_len)
2364 /* Do not fetch more if the record will be too recent */
2365 if (oldest <= last) {
2367 oldest = mdb_find_oldest(txn);
2368 env->me_pgoldest = oldest;
2374 rc = mdb_cursor_get(&m2, &key, NULL, op);
2376 if (rc == MDB_NOTFOUND)
2380 last = *(txnid_t*)key.mv_data;
2381 if (oldest <= last) {
2383 oldest = mdb_find_oldest(txn);
2384 env->me_pgoldest = oldest;
2390 np = m2.mc_pg[m2.mc_top];
2391 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2392 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2395 idl = (MDB_ID *) data.mv_data;
2398 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2403 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2405 mop = env->me_pghead;
2407 env->me_pglast = last;
2409 DPRINTF(("IDL read txn %"Y"u root %"Y"u num %u",
2410 last, txn->mt_dbs[FREE_DBI].md_root, i));
2412 DPRINTF(("IDL %"Y"u", idl[j]));
2414 /* Merge in descending sorted order */
2415 mdb_midl_xmerge(mop, idl);
2419 /* Use new pages from the map when nothing suitable in the freeDB */
2421 pgno = txn->mt_next_pgno;
2422 if (pgno + num >= env->me_maxpg) {
2423 DPUTS("DB size maxed out");
2427 #if defined(_WIN32) && !defined(MDB_VL32)
2428 if (!(env->me_flags & MDB_RDONLY)) {
2430 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
2431 p = VirtualAlloc(p, env->me_psize * num, MEM_COMMIT,
2432 (env->me_flags & MDB_WRITEMAP) ? PAGE_READWRITE:
2435 DPUTS("VirtualAlloc failed");
2443 if (env->me_flags & MDB_WRITEMAP) {
2444 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2446 if (!(np = mdb_page_malloc(txn, num))) {
2452 mop[0] = mop_len -= num;
2453 /* Move any stragglers down */
2454 for (j = i-num; j < mop_len; )
2455 mop[++j] = mop[++i];
2457 txn->mt_next_pgno = pgno + num;
2460 mdb_page_dirty(txn, np);
2466 txn->mt_flags |= MDB_TXN_ERROR;
2470 /** Copy the used portions of a non-overflow page.
2471 * @param[in] dst page to copy into
2472 * @param[in] src page to copy from
2473 * @param[in] psize size of a page
2476 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2478 enum { Align = sizeof(pgno_t) };
2479 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2481 /* If page isn't full, just copy the used portion. Adjust
2482 * alignment so memcpy may copy words instead of bytes.
2484 if ((unused &= -Align) && !IS_LEAF2(src)) {
2485 upper = (upper + PAGEBASE) & -Align;
2486 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2487 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2490 memcpy(dst, src, psize - unused);
2494 /** Pull a page off the txn's spill list, if present.
2495 * If a page being referenced was spilled to disk in this txn, bring
2496 * it back and make it dirty/writable again.
2497 * @param[in] txn the transaction handle.
2498 * @param[in] mp the page being referenced. It must not be dirty.
2499 * @param[out] ret the writable page, if any. ret is unchanged if
2500 * mp wasn't spilled.
2503 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2505 MDB_env *env = txn->mt_env;
2508 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2510 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2511 if (!tx2->mt_spill_pgs)
2513 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2514 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2517 if (txn->mt_dirty_room == 0)
2518 return MDB_TXN_FULL;
2519 if (IS_OVERFLOW(mp))
2523 if (env->me_flags & MDB_WRITEMAP) {
2526 np = mdb_page_malloc(txn, num);
2530 memcpy(np, mp, num * env->me_psize);
2532 mdb_page_copy(np, mp, env->me_psize);
2535 /* If in current txn, this page is no longer spilled.
2536 * If it happens to be the last page, truncate the spill list.
2537 * Otherwise mark it as deleted by setting the LSB.
2539 if (x == txn->mt_spill_pgs[0])
2540 txn->mt_spill_pgs[0]--;
2542 txn->mt_spill_pgs[x] |= 1;
2543 } /* otherwise, if belonging to a parent txn, the
2544 * page remains spilled until child commits
2547 mdb_page_dirty(txn, np);
2548 np->mp_flags |= P_DIRTY;
2556 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2557 * @param[in] mc cursor pointing to the page to be touched
2558 * @return 0 on success, non-zero on failure.
2561 mdb_page_touch(MDB_cursor *mc)
2563 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2564 MDB_txn *txn = mc->mc_txn;
2565 MDB_cursor *m2, *m3;
2569 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2570 if (txn->mt_flags & MDB_TXN_SPILLS) {
2572 rc = mdb_page_unspill(txn, mp, &np);
2578 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2579 (rc = mdb_page_alloc(mc, 1, &np)))
2582 DPRINTF(("touched db %d page %"Y"u -> %"Y"u", DDBI(mc),
2583 mp->mp_pgno, pgno));
2584 mdb_cassert(mc, mp->mp_pgno != pgno);
2585 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2586 /* Update the parent page, if any, to point to the new page */
2588 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2589 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2590 SETPGNO(node, pgno);
2592 mc->mc_db->md_root = pgno;
2594 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2595 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2597 /* If txn has a parent, make sure the page is in our
2601 unsigned x = mdb_mid2l_search(dl, pgno);
2602 if (x <= dl[0].mid && dl[x].mid == pgno) {
2603 if (mp != dl[x].mptr) { /* bad cursor? */
2604 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2605 txn->mt_flags |= MDB_TXN_ERROR;
2611 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2613 np = mdb_page_malloc(txn, 1);
2618 rc = mdb_mid2l_insert(dl, &mid);
2619 mdb_cassert(mc, rc == 0);
2624 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2626 np->mp_flags |= P_DIRTY;
2629 /* Adjust cursors pointing to mp */
2630 mc->mc_pg[mc->mc_top] = np;
2631 m2 = txn->mt_cursors[mc->mc_dbi];
2632 if (mc->mc_flags & C_SUB) {
2633 for (; m2; m2=m2->mc_next) {
2634 m3 = &m2->mc_xcursor->mx_cursor;
2635 if (m3->mc_snum < mc->mc_snum) continue;
2636 if (m3->mc_pg[mc->mc_top] == mp)
2637 m3->mc_pg[mc->mc_top] = np;
2640 for (; m2; m2=m2->mc_next) {
2641 if (m2->mc_snum < mc->mc_snum) continue;
2642 if (m2 == mc) continue;
2643 if (m2->mc_pg[mc->mc_top] == mp) {
2644 m2->mc_pg[mc->mc_top] = np;
2645 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2647 (m2->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
2649 MDB_node *leaf = NODEPTR(np, m2->mc_ki[mc->mc_top]);
2650 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
2651 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2656 MDB_PAGE_UNREF(mc->mc_txn, mp);
2660 txn->mt_flags |= MDB_TXN_ERROR;
2665 mdb_env_sync0(MDB_env *env, int force, pgno_t numpgs)
2668 if (env->me_flags & MDB_RDONLY)
2670 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2671 if (env->me_flags & MDB_WRITEMAP) {
2672 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2673 ? MS_ASYNC : MS_SYNC;
2674 if (MDB_MSYNC(env->me_map, env->me_psize * numpgs, flags))
2677 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2681 #ifdef BROKEN_FDATASYNC
2682 if (env->me_flags & MDB_FSYNCONLY) {
2683 if (fsync(env->me_fd))
2687 if (MDB_FDATASYNC(env->me_fd))
2695 mdb_env_sync(MDB_env *env, int force)
2697 MDB_meta *m = mdb_env_pick_meta(env);
2698 return mdb_env_sync0(env, force, m->mm_last_pg+1);
2701 /** Back up parent txn's cursors, then grab the originals for tracking */
2703 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2705 MDB_cursor *mc, *bk;
2710 for (i = src->mt_numdbs; --i >= 0; ) {
2711 if ((mc = src->mt_cursors[i]) != NULL) {
2712 size = sizeof(MDB_cursor);
2714 size += sizeof(MDB_xcursor);
2715 for (; mc; mc = bk->mc_next) {
2721 mc->mc_db = &dst->mt_dbs[i];
2722 /* Kill pointers into src to reduce abuse: The
2723 * user may not use mc until dst ends. But we need a valid
2724 * txn pointer here for cursor fixups to keep working.
2727 mc->mc_dbflag = &dst->mt_dbflags[i];
2728 if ((mx = mc->mc_xcursor) != NULL) {
2729 *(MDB_xcursor *)(bk+1) = *mx;
2730 mx->mx_cursor.mc_txn = dst;
2732 mc->mc_next = dst->mt_cursors[i];
2733 dst->mt_cursors[i] = mc;
2740 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2741 * @param[in] txn the transaction handle.
2742 * @param[in] merge true to keep changes to parent cursors, false to revert.
2743 * @return 0 on success, non-zero on failure.
2746 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2748 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2752 for (i = txn->mt_numdbs; --i >= 0; ) {
2753 for (mc = cursors[i]; mc; mc = next) {
2755 if ((bk = mc->mc_backup) != NULL) {
2757 /* Commit changes to parent txn */
2758 mc->mc_next = bk->mc_next;
2759 mc->mc_backup = bk->mc_backup;
2760 mc->mc_txn = bk->mc_txn;
2761 mc->mc_db = bk->mc_db;
2762 mc->mc_dbflag = bk->mc_dbflag;
2763 if ((mx = mc->mc_xcursor) != NULL)
2764 mx->mx_cursor.mc_txn = bk->mc_txn;
2766 /* Abort nested txn */
2768 if ((mx = mc->mc_xcursor) != NULL)
2769 *mx = *(MDB_xcursor *)(bk+1);
2773 /* Only malloced cursors are permanently tracked. */
2780 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2786 Pidset = F_SETLK, Pidcheck = F_GETLK
2790 /** Set or check a pid lock. Set returns 0 on success.
2791 * Check returns 0 if the process is certainly dead, nonzero if it may
2792 * be alive (the lock exists or an error happened so we do not know).
2794 * On Windows Pidset is a no-op, we merely check for the existence
2795 * of the process with the given pid. On POSIX we use a single byte
2796 * lock on the lockfile, set at an offset equal to the pid.
2799 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2801 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2804 if (op == Pidcheck) {
2805 h = OpenProcess(env->me_pidquery, FALSE, pid);
2806 /* No documented "no such process" code, but other program use this: */
2808 return ErrCode() != ERROR_INVALID_PARAMETER;
2809 /* A process exists until all handles to it close. Has it exited? */
2810 ret = WaitForSingleObject(h, 0) != 0;
2817 struct flock lock_info;
2818 memset(&lock_info, 0, sizeof(lock_info));
2819 lock_info.l_type = F_WRLCK;
2820 lock_info.l_whence = SEEK_SET;
2821 lock_info.l_start = pid;
2822 lock_info.l_len = 1;
2823 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2824 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2826 } else if ((rc = ErrCode()) == EINTR) {
2834 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2835 * @param[in] txn the transaction handle to initialize
2836 * @return 0 on success, non-zero on failure.
2839 mdb_txn_renew0(MDB_txn *txn)
2841 MDB_env *env = txn->mt_env;
2842 MDB_txninfo *ti = env->me_txns;
2844 unsigned int i, nr, flags = txn->mt_flags;
2846 int rc, new_notls = 0;
2848 if ((flags &= MDB_TXN_RDONLY) != 0) {
2850 meta = mdb_env_pick_meta(env);
2851 txn->mt_txnid = meta->mm_txnid;
2852 txn->mt_u.reader = NULL;
2854 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2855 pthread_getspecific(env->me_txkey);
2857 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2858 return MDB_BAD_RSLOT;
2860 MDB_PID_T pid = env->me_pid;
2861 MDB_THR_T tid = pthread_self();
2862 mdb_mutexref_t rmutex = env->me_rmutex;
2864 if (!env->me_live_reader) {
2865 rc = mdb_reader_pid(env, Pidset, pid);
2868 env->me_live_reader = 1;
2871 if (LOCK_MUTEX(rc, env, rmutex))
2873 nr = ti->mti_numreaders;
2874 for (i=0; i<nr; i++)
2875 if (ti->mti_readers[i].mr_pid == 0)
2877 if (i == env->me_maxreaders) {
2878 UNLOCK_MUTEX(rmutex);
2879 return MDB_READERS_FULL;
2881 r = &ti->mti_readers[i];
2882 /* Claim the reader slot, carefully since other code
2883 * uses the reader table un-mutexed: First reset the
2884 * slot, next publish it in mti_numreaders. After
2885 * that, it is safe for mdb_env_close() to touch it.
2886 * When it will be closed, we can finally claim it.
2889 r->mr_txnid = (txnid_t)-1;
2892 ti->mti_numreaders = ++nr;
2893 env->me_close_readers = nr;
2895 UNLOCK_MUTEX(rmutex);
2897 new_notls = (env->me_flags & MDB_NOTLS);
2898 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2903 do /* LY: Retry on a race, ITS#7970. */
2904 r->mr_txnid = ti->mti_txnid;
2905 while(r->mr_txnid != ti->mti_txnid);
2906 txn->mt_txnid = r->mr_txnid;
2907 txn->mt_u.reader = r;
2908 meta = env->me_metas[txn->mt_txnid & 1];
2912 /* Not yet touching txn == env->me_txn0, it may be active */
2914 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2916 txn->mt_txnid = ti->mti_txnid;
2917 meta = env->me_metas[txn->mt_txnid & 1];
2919 meta = mdb_env_pick_meta(env);
2920 txn->mt_txnid = meta->mm_txnid;
2924 if (txn->mt_txnid == mdb_debug_start)
2927 txn->mt_child = NULL;
2928 txn->mt_loose_pgs = NULL;
2929 txn->mt_loose_count = 0;
2930 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2931 txn->mt_u.dirty_list = env->me_dirty_list;
2932 txn->mt_u.dirty_list[0].mid = 0;
2933 txn->mt_free_pgs = env->me_free_pgs;
2934 txn->mt_free_pgs[0] = 0;
2935 txn->mt_spill_pgs = NULL;
2937 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2940 /* Copy the DB info and flags */
2941 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2943 /* Moved to here to avoid a data race in read TXNs */
2944 txn->mt_next_pgno = meta->mm_last_pg+1;
2946 txn->mt_last_pgno = txn->mt_next_pgno - 1;
2949 txn->mt_flags = flags;
2952 txn->mt_numdbs = env->me_numdbs;
2953 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2954 x = env->me_dbflags[i];
2955 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2956 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2958 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2959 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2961 if (env->me_flags & MDB_FATAL_ERROR) {
2962 DPUTS("environment had fatal error, must shutdown!");
2964 } else if (env->me_maxpg < txn->mt_next_pgno) {
2965 rc = MDB_MAP_RESIZED;
2969 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2974 mdb_txn_renew(MDB_txn *txn)
2978 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2981 rc = mdb_txn_renew0(txn);
2982 if (rc == MDB_SUCCESS) {
2983 DPRINTF(("renew txn %"Y"u%c %p on mdbenv %p, root page %"Y"u",
2984 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2985 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2991 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2995 int rc, size, tsize;
2997 flags &= MDB_TXN_BEGIN_FLAGS;
2998 flags |= env->me_flags & MDB_WRITEMAP;
3000 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
3004 /* Nested transactions: Max 1 child, write txns only, no writemap */
3005 flags |= parent->mt_flags;
3006 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
3007 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
3009 /* Child txns save MDB_pgstate and use own copy of cursors */
3010 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
3011 size += tsize = sizeof(MDB_ntxn);
3012 } else if (flags & MDB_RDONLY) {
3013 size = env->me_maxdbs * (sizeof(MDB_db)+1);
3014 size += tsize = sizeof(MDB_txn);
3016 /* Reuse preallocated write txn. However, do not touch it until
3017 * mdb_txn_renew0() succeeds, since it currently may be active.
3022 if ((txn = calloc(1, size)) == NULL) {
3023 DPRINTF(("calloc: %s", strerror(errno)));
3028 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
3029 if (!txn->mt_rpages) {
3033 txn->mt_rpages[0].mid = 0;
3034 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
3037 txn->mt_dbxs = env->me_dbxs; /* static */
3038 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
3039 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
3040 txn->mt_flags = flags;
3045 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
3046 txn->mt_dbiseqs = parent->mt_dbiseqs;
3047 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
3048 if (!txn->mt_u.dirty_list ||
3049 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
3051 free(txn->mt_u.dirty_list);
3055 txn->mt_txnid = parent->mt_txnid;
3056 txn->mt_dirty_room = parent->mt_dirty_room;
3057 txn->mt_u.dirty_list[0].mid = 0;
3058 txn->mt_spill_pgs = NULL;
3059 txn->mt_next_pgno = parent->mt_next_pgno;
3060 parent->mt_flags |= MDB_TXN_HAS_CHILD;
3061 parent->mt_child = txn;
3062 txn->mt_parent = parent;
3063 txn->mt_numdbs = parent->mt_numdbs;
3065 txn->mt_rpages = parent->mt_rpages;
3067 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3068 /* Copy parent's mt_dbflags, but clear DB_NEW */
3069 for (i=0; i<txn->mt_numdbs; i++)
3070 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
3072 ntxn = (MDB_ntxn *)txn;
3073 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
3074 if (env->me_pghead) {
3075 size = MDB_IDL_SIZEOF(env->me_pghead);
3076 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
3078 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
3083 rc = mdb_cursor_shadow(parent, txn);
3085 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
3086 } else { /* MDB_RDONLY */
3087 txn->mt_dbiseqs = env->me_dbiseqs;
3089 rc = mdb_txn_renew0(txn);
3092 if (txn != env->me_txn0) {
3094 free(txn->mt_rpages);
3099 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
3101 DPRINTF(("begin txn %"Y"u%c %p on mdbenv %p, root page %"Y"u",
3102 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
3103 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
3110 mdb_txn_env(MDB_txn *txn)
3112 if(!txn) return NULL;
3117 mdb_txn_id(MDB_txn *txn)
3120 return txn->mt_txnid;
3123 /** Export or close DBI handles opened in this txn. */
3125 mdb_dbis_update(MDB_txn *txn, int keep)
3128 MDB_dbi n = txn->mt_numdbs;
3129 MDB_env *env = txn->mt_env;
3130 unsigned char *tdbflags = txn->mt_dbflags;
3132 for (i = n; --i >= CORE_DBS;) {
3133 if (tdbflags[i] & DB_NEW) {
3135 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
3137 char *ptr = env->me_dbxs[i].md_name.mv_data;
3139 env->me_dbxs[i].md_name.mv_data = NULL;
3140 env->me_dbxs[i].md_name.mv_size = 0;
3141 env->me_dbflags[i] = 0;
3142 env->me_dbiseqs[i]++;
3148 if (keep && env->me_numdbs < n)
3152 /** End a transaction, except successful commit of a nested transaction.
3153 * May be called twice for readonly txns: First reset it, then abort.
3154 * @param[in] txn the transaction handle to end
3155 * @param[in] mode why and how to end the transaction
3158 mdb_txn_end(MDB_txn *txn, unsigned mode)
3160 MDB_env *env = txn->mt_env;
3162 static const char *const names[] = MDB_END_NAMES;
3165 /* Export or close DBI handles opened in this txn */
3166 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
3168 DPRINTF(("%s txn %"Y"u%c %p on mdbenv %p, root page %"Y"u",
3169 names[mode & MDB_END_OPMASK],
3170 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3171 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
3173 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3174 if (txn->mt_u.reader) {
3175 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
3176 if (!(env->me_flags & MDB_NOTLS)) {
3177 txn->mt_u.reader = NULL; /* txn does not own reader */
3178 } else if (mode & MDB_END_SLOT) {
3179 txn->mt_u.reader->mr_pid = 0;
3180 txn->mt_u.reader = NULL;
3181 } /* else txn owns the slot until it does MDB_END_SLOT */
3183 txn->mt_numdbs = 0; /* prevent further DBI activity */
3184 txn->mt_flags |= MDB_TXN_FINISHED;
3186 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
3187 pgno_t *pghead = env->me_pghead;
3189 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
3190 mdb_cursors_close(txn, 0);
3191 if (!(env->me_flags & MDB_WRITEMAP)) {
3192 mdb_dlist_free(txn);
3196 txn->mt_flags = MDB_TXN_FINISHED;
3198 if (!txn->mt_parent) {
3199 mdb_midl_shrink(&txn->mt_free_pgs);
3200 env->me_free_pgs = txn->mt_free_pgs;
3202 env->me_pghead = NULL;
3206 mode = 0; /* txn == env->me_txn0, do not free() it */
3208 /* The writer mutex was locked in mdb_txn_begin. */
3210 UNLOCK_MUTEX(env->me_wmutex);
3212 txn->mt_parent->mt_child = NULL;
3213 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3214 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3215 mdb_midl_free(txn->mt_free_pgs);
3216 mdb_midl_free(txn->mt_spill_pgs);
3217 free(txn->mt_u.dirty_list);
3220 mdb_midl_free(pghead);
3223 if (!txn->mt_parent) {
3224 MDB_ID3L el = env->me_rpages, tl = txn->mt_rpages;
3225 unsigned i, x, n = tl[0].mid;
3226 pthread_mutex_lock(&env->me_rpmutex);
3227 for (i = 1; i <= n; i++) {
3228 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
3229 /* tmp overflow pages that we didn't share in env */
3230 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3232 x = mdb_mid3l_search(el, tl[i].mid);
3233 if (tl[i].mptr == el[x].mptr) {
3236 /* another tmp overflow page */
3237 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3241 pthread_mutex_unlock(&env->me_rpmutex);
3243 if (mode & MDB_END_FREE)
3247 if (mode & MDB_END_FREE)
3252 mdb_txn_reset(MDB_txn *txn)
3257 /* This call is only valid for read-only txns */
3258 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3261 mdb_txn_end(txn, MDB_END_RESET);
3265 mdb_txn_abort(MDB_txn *txn)
3271 mdb_txn_abort(txn->mt_child);
3273 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3276 /** Save the freelist as of this transaction to the freeDB.
3277 * This changes the freelist. Keep trying until it stabilizes.
3279 * When (MDB_DEVEL) & 2, the changes do not affect #mdb_page_alloc(),
3280 * it then uses the transaction's original snapshot of the freeDB.
3283 mdb_freelist_save(MDB_txn *txn)
3285 /* env->me_pghead[] can grow and shrink during this call.
3286 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3287 * Page numbers cannot disappear from txn->mt_free_pgs[].
3290 MDB_env *env = txn->mt_env;
3291 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3292 txnid_t pglast = 0, head_id = 0;
3293 pgno_t freecnt = 0, *free_pgs, *mop;
3294 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3296 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3298 if (env->me_pghead) {
3299 /* Make sure first page of freeDB is touched and on freelist */
3300 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3301 if (rc && rc != MDB_NOTFOUND)
3305 if (!env->me_pghead && txn->mt_loose_pgs) {
3306 /* Put loose page numbers in mt_free_pgs, since
3307 * we may be unable to return them to me_pghead.
3309 MDB_page *mp = txn->mt_loose_pgs;
3310 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3312 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3313 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3314 txn->mt_loose_pgs = NULL;
3315 txn->mt_loose_count = 0;
3318 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3319 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3320 ? SSIZE_MAX : maxfree_1pg;
3323 /* Come back here after each Put() in case freelist changed */
3328 /* If using records from freeDB which we have not yet
3329 * deleted, delete them and any we reserved for me_pghead.
3331 while (pglast < env->me_pglast) {
3332 rc = mdb_cursor_first(&mc, &key, NULL);
3335 pglast = head_id = *(txnid_t *)key.mv_data;
3336 total_room = head_room = 0;
3337 mdb_tassert(txn, pglast <= env->me_pglast);
3338 rc = mdb_cursor_del(&mc, 0);
3343 /* Save the IDL of pages freed by this txn, to a single record */
3344 if (freecnt < txn->mt_free_pgs[0]) {
3346 /* Make sure last page of freeDB is touched and on freelist */
3347 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3348 if (rc && rc != MDB_NOTFOUND)
3351 free_pgs = txn->mt_free_pgs;
3352 /* Write to last page of freeDB */
3353 key.mv_size = sizeof(txn->mt_txnid);
3354 key.mv_data = &txn->mt_txnid;
3356 freecnt = free_pgs[0];
3357 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3358 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3361 /* Retry if mt_free_pgs[] grew during the Put() */
3362 free_pgs = txn->mt_free_pgs;
3363 } while (freecnt < free_pgs[0]);
3364 mdb_midl_sort(free_pgs);
3365 memcpy(data.mv_data, free_pgs, data.mv_size);
3368 unsigned int i = free_pgs[0];
3369 DPRINTF(("IDL write txn %"Y"u root %"Y"u num %u",
3370 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3372 DPRINTF(("IDL %"Y"u", free_pgs[i]));
3378 mop = env->me_pghead;
3379 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3381 /* Reserve records for me_pghead[]. Split it if multi-page,
3382 * to avoid searching freeDB for a page range. Use keys in
3383 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3385 if (total_room >= mop_len) {
3386 if (total_room == mop_len || --more < 0)
3388 } else if (head_room >= maxfree_1pg && head_id > 1) {
3389 /* Keep current record (overflow page), add a new one */
3393 /* (Re)write {key = head_id, IDL length = head_room} */
3394 total_room -= head_room;
3395 head_room = mop_len - total_room;
3396 if (head_room > maxfree_1pg && head_id > 1) {
3397 /* Overflow multi-page for part of me_pghead */
3398 head_room /= head_id; /* amortize page sizes */
3399 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3400 } else if (head_room < 0) {
3401 /* Rare case, not bothering to delete this record */
3404 key.mv_size = sizeof(head_id);
3405 key.mv_data = &head_id;
3406 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3407 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3410 /* IDL is initially empty, zero out at least the length */
3411 pgs = (pgno_t *)data.mv_data;
3412 j = head_room > clean_limit ? head_room : 0;
3416 total_room += head_room;
3419 /* Return loose page numbers to me_pghead, though usually none are
3420 * left at this point. The pages themselves remain in dirty_list.
3422 if (txn->mt_loose_pgs) {
3423 MDB_page *mp = txn->mt_loose_pgs;
3424 unsigned count = txn->mt_loose_count;
3426 /* Room for loose pages + temp IDL with same */
3427 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3429 mop = env->me_pghead;
3430 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3431 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3432 loose[ ++count ] = mp->mp_pgno;
3434 mdb_midl_sort(loose);
3435 mdb_midl_xmerge(mop, loose);
3436 txn->mt_loose_pgs = NULL;
3437 txn->mt_loose_count = 0;
3441 /* Fill in the reserved me_pghead records */
3447 rc = mdb_cursor_first(&mc, &key, &data);
3448 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3449 txnid_t id = *(txnid_t *)key.mv_data;
3450 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3453 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3455 if (len > mop_len) {
3457 data.mv_size = (len + 1) * sizeof(MDB_ID);
3459 data.mv_data = mop -= len;
3462 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3464 if (rc || !(mop_len -= len))
3471 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3472 * @param[in] txn the transaction that's being committed
3473 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3474 * @return 0 on success, non-zero on failure.
3477 mdb_page_flush(MDB_txn *txn, int keep)
3479 MDB_env *env = txn->mt_env;
3480 MDB_ID2L dl = txn->mt_u.dirty_list;
3481 unsigned psize = env->me_psize, j;
3482 int i, pagecount = dl[0].mid, rc;
3486 MDB_page *dp = NULL;
3490 struct iovec iov[MDB_COMMIT_PAGES];
3491 ssize_t wsize = 0, wres;
3492 off_t wpos = 0, next_pos = 1; /* impossible pos, so pos != next_pos */
3498 if (env->me_flags & MDB_WRITEMAP) {
3499 /* Clear dirty flags */
3500 while (++i <= pagecount) {
3502 /* Don't flush this page yet */
3503 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3504 dp->mp_flags &= ~P_KEEP;
3508 dp->mp_flags &= ~P_DIRTY;
3513 /* Write the pages */
3515 if (++i <= pagecount) {
3517 /* Don't flush this page yet */
3518 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3519 dp->mp_flags &= ~P_KEEP;
3524 /* clear dirty flag */
3525 dp->mp_flags &= ~P_DIRTY;
3528 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3533 /* Windows actually supports scatter/gather I/O, but only on
3534 * unbuffered file handles. Since we're relying on the OS page
3535 * cache for all our data, that's self-defeating. So we just
3536 * write pages one at a time. We use the ov structure to set
3537 * the write offset, to at least save the overhead of a Seek
3540 DPRINTF(("committing page %"Z"u", pgno));
3541 memset(&ov, 0, sizeof(ov));
3542 ov.Offset = pos & 0xffffffff;
3543 ov.OffsetHigh = pos >> 16 >> 16;
3544 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3546 DPRINTF(("WriteFile: %d", rc));
3550 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3551 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3554 /* Write previous page(s) */
3555 #ifdef MDB_USE_PWRITEV
3556 wres = pwritev(env->me_fd, iov, n, wpos);
3559 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3562 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3566 DPRINTF(("lseek: %s", strerror(rc)));
3569 wres = writev(env->me_fd, iov, n);
3572 if (wres != wsize) {
3577 DPRINTF(("Write error: %s", strerror(rc)));
3579 rc = EIO; /* TODO: Use which error code? */
3580 DPUTS("short write, filesystem full?");
3591 DPRINTF(("committing page %"Y"u", pgno));
3592 next_pos = pos + size;
3593 iov[n].iov_len = size;
3594 iov[n].iov_base = (char *)dp;
3600 if (pgno > txn->mt_last_pgno)
3601 txn->mt_last_pgno = pgno;
3604 /* MIPS has cache coherency issues, this is a no-op everywhere else
3605 * Note: for any size >= on-chip cache size, entire on-chip cache is
3608 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3610 for (i = keep; ++i <= pagecount; ) {
3612 /* This is a page we skipped above */
3615 dl[j].mid = dp->mp_pgno;
3618 mdb_dpage_free(env, dp);
3623 txn->mt_dirty_room += i - j;
3629 mdb_txn_commit(MDB_txn *txn)
3632 unsigned int i, end_mode;
3638 /* mdb_txn_end() mode for a commit which writes nothing */
3639 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3641 if (txn->mt_child) {
3642 rc = mdb_txn_commit(txn->mt_child);
3649 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3653 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3654 DPUTS("txn has failed/finished, can't commit");
3656 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3661 if (txn->mt_parent) {
3662 MDB_txn *parent = txn->mt_parent;
3666 unsigned x, y, len, ps_len;
3668 /* Append our free list to parent's */
3669 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3672 mdb_midl_free(txn->mt_free_pgs);
3673 /* Failures after this must either undo the changes
3674 * to the parent or set MDB_TXN_ERROR in the parent.
3677 parent->mt_next_pgno = txn->mt_next_pgno;
3678 parent->mt_flags = txn->mt_flags;
3680 /* Merge our cursors into parent's and close them */
3681 mdb_cursors_close(txn, 1);
3683 /* Update parent's DB table. */
3684 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3685 parent->mt_numdbs = txn->mt_numdbs;
3686 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3687 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3688 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3689 /* preserve parent's DB_NEW status */
3690 x = parent->mt_dbflags[i] & DB_NEW;
3691 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3694 dst = parent->mt_u.dirty_list;
3695 src = txn->mt_u.dirty_list;
3696 /* Remove anything in our dirty list from parent's spill list */
3697 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3699 pspill[0] = (pgno_t)-1;
3700 /* Mark our dirty pages as deleted in parent spill list */
3701 for (i=0, len=src[0].mid; ++i <= len; ) {
3702 MDB_ID pn = src[i].mid << 1;
3703 while (pn > pspill[x])
3705 if (pn == pspill[x]) {
3710 /* Squash deleted pagenums if we deleted any */
3711 for (x=y; ++x <= ps_len; )
3712 if (!(pspill[x] & 1))
3713 pspill[++y] = pspill[x];
3717 /* Remove anything in our spill list from parent's dirty list */
3718 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3719 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3720 MDB_ID pn = txn->mt_spill_pgs[i];
3722 continue; /* deleted spillpg */
3724 y = mdb_mid2l_search(dst, pn);
3725 if (y <= dst[0].mid && dst[y].mid == pn) {
3727 while (y < dst[0].mid) {
3736 /* Find len = length of merging our dirty list with parent's */
3738 dst[0].mid = 0; /* simplify loops */
3739 if (parent->mt_parent) {
3740 len = x + src[0].mid;
3741 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3742 for (i = x; y && i; y--) {
3743 pgno_t yp = src[y].mid;
3744 while (yp < dst[i].mid)
3746 if (yp == dst[i].mid) {
3751 } else { /* Simplify the above for single-ancestor case */
3752 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3754 /* Merge our dirty list with parent's */
3756 for (i = len; y; dst[i--] = src[y--]) {
3757 pgno_t yp = src[y].mid;
3758 while (yp < dst[x].mid)
3759 dst[i--] = dst[x--];
3760 if (yp == dst[x].mid)
3761 free(dst[x--].mptr);
3763 mdb_tassert(txn, i == x);
3765 free(txn->mt_u.dirty_list);
3766 parent->mt_dirty_room = txn->mt_dirty_room;
3767 if (txn->mt_spill_pgs) {
3768 if (parent->mt_spill_pgs) {
3769 /* TODO: Prevent failure here, so parent does not fail */
3770 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3772 parent->mt_flags |= MDB_TXN_ERROR;
3773 mdb_midl_free(txn->mt_spill_pgs);
3774 mdb_midl_sort(parent->mt_spill_pgs);
3776 parent->mt_spill_pgs = txn->mt_spill_pgs;
3780 /* Append our loose page list to parent's */
3781 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3783 *lp = txn->mt_loose_pgs;
3784 parent->mt_loose_count += txn->mt_loose_count;
3786 parent->mt_child = NULL;
3787 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3792 if (txn != env->me_txn) {
3793 DPUTS("attempt to commit unknown transaction");
3798 mdb_cursors_close(txn, 0);
3800 if (!txn->mt_u.dirty_list[0].mid &&
3801 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3804 DPRINTF(("committing txn %"Y"u %p on mdbenv %p, root page %"Y"u",
3805 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3807 /* Update DB root pointers */
3808 if (txn->mt_numdbs > CORE_DBS) {
3812 data.mv_size = sizeof(MDB_db);
3814 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3815 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3816 if (txn->mt_dbflags[i] & DB_DIRTY) {
3817 if (TXN_DBI_CHANGED(txn, i)) {
3821 data.mv_data = &txn->mt_dbs[i];
3822 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3830 rc = mdb_freelist_save(txn);
3834 mdb_midl_free(env->me_pghead);
3835 env->me_pghead = NULL;
3836 mdb_midl_shrink(&txn->mt_free_pgs);
3842 if ((rc = mdb_page_flush(txn, 0)))
3844 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3845 (rc = mdb_env_sync0(env, 0, txn->mt_next_pgno)))
3847 if ((rc = mdb_env_write_meta(txn)))
3849 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3852 mdb_txn_end(txn, end_mode);
3860 /** Read the environment parameters of a DB environment before
3861 * mapping it into memory.
3862 * @param[in] env the environment handle
3863 * @param[out] meta address of where to store the meta information
3864 * @return 0 on success, non-zero on failure.
3867 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3873 enum { Size = sizeof(pbuf) };
3875 /* We don't know the page size yet, so use a minimum value.
3876 * Read both meta pages so we can use the latest one.
3879 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3883 memset(&ov, 0, sizeof(ov));
3885 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3886 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3889 rc = pread(env->me_fd, &pbuf, Size, off);
3892 if (rc == 0 && off == 0)
3894 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3895 DPRINTF(("read: %s", mdb_strerror(rc)));
3899 p = (MDB_page *)&pbuf;
3901 if (!F_ISSET(p->mp_flags, P_META)) {
3902 DPRINTF(("page %"Y"u not a meta page", p->mp_pgno));
3907 if (m->mm_magic != MDB_MAGIC) {
3908 DPUTS("meta has invalid magic");
3912 if (m->mm_version != MDB_DATA_VERSION) {
3913 DPRINTF(("database is version %u, expected version %u",
3914 m->mm_version, MDB_DATA_VERSION));
3915 return MDB_VERSION_MISMATCH;
3918 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3924 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3926 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3928 meta->mm_magic = MDB_MAGIC;
3929 meta->mm_version = MDB_DATA_VERSION;
3930 meta->mm_mapsize = env->me_mapsize;
3931 meta->mm_psize = env->me_psize;
3932 meta->mm_last_pg = NUM_METAS-1;
3933 meta->mm_flags = env->me_flags & 0xffff;
3934 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3935 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3936 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3939 /** Write the environment parameters of a freshly created DB environment.
3940 * @param[in] env the environment handle
3941 * @param[in] meta the #MDB_meta to write
3942 * @return 0 on success, non-zero on failure.
3945 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3953 memset(&ov, 0, sizeof(ov));
3954 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3956 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3959 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3960 len = pwrite(fd, ptr, size, pos); \
3961 if (len == -1 && ErrCode() == EINTR) continue; \
3962 rc = (len >= 0); break; } while(1)
3965 DPUTS("writing new meta page");
3967 psize = env->me_psize;
3969 p = calloc(NUM_METAS, psize);
3973 p->mp_flags = P_META;
3974 *(MDB_meta *)METADATA(p) = *meta;
3976 q = (MDB_page *)((char *)p + psize);
3978 q->mp_flags = P_META;
3979 *(MDB_meta *)METADATA(q) = *meta;
3981 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3984 else if ((unsigned) len == psize * NUM_METAS)
3992 /** Update the environment info to commit a transaction.
3993 * @param[in] txn the transaction that's being committed
3994 * @return 0 on success, non-zero on failure.
3997 mdb_env_write_meta(MDB_txn *txn)
4000 MDB_meta meta, metab, *mp;
4004 int rc, len, toggle;
4013 toggle = txn->mt_txnid & 1;
4014 DPRINTF(("writing meta page %d for root page %"Y"u",
4015 toggle, txn->mt_dbs[MAIN_DBI].md_root));
4018 flags = txn->mt_flags | env->me_flags;
4019 mp = env->me_metas[toggle];
4020 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
4021 /* Persist any increases of mapsize config */
4022 if (mapsize < env->me_mapsize)
4023 mapsize = env->me_mapsize;
4025 if (flags & MDB_WRITEMAP) {
4026 mp->mm_mapsize = mapsize;
4027 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4028 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4029 mp->mm_last_pg = txn->mt_next_pgno - 1;
4030 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
4031 !(defined(__i386__) || defined(__x86_64__))
4032 /* LY: issue a memory barrier, if not x86. ITS#7969 */
4033 __sync_synchronize();
4035 mp->mm_txnid = txn->mt_txnid;
4036 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
4037 unsigned meta_size = env->me_psize;
4038 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
4039 ptr = (char *)mp - PAGEHDRSZ;
4040 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
4041 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
4045 if (MDB_MSYNC(ptr, meta_size, rc)) {
4052 metab.mm_txnid = mp->mm_txnid;
4053 metab.mm_last_pg = mp->mm_last_pg;
4055 meta.mm_mapsize = mapsize;
4056 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4057 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4058 meta.mm_last_pg = txn->mt_next_pgno - 1;
4059 meta.mm_txnid = txn->mt_txnid;
4061 off = offsetof(MDB_meta, mm_mapsize);
4062 ptr = (char *)&meta + off;
4063 len = sizeof(MDB_meta) - off;
4064 off += (char *)mp - env->me_map;
4066 /* Write to the SYNC fd */
4067 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
4070 memset(&ov, 0, sizeof(ov));
4072 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
4077 rc = pwrite(mfd, ptr, len, off);
4080 rc = rc < 0 ? ErrCode() : EIO;
4085 DPUTS("write failed, disk error?");
4086 /* On a failure, the pagecache still contains the new data.
4087 * Write some old data back, to prevent it from being used.
4088 * Use the non-SYNC fd; we know it will fail anyway.
4090 meta.mm_last_pg = metab.mm_last_pg;
4091 meta.mm_txnid = metab.mm_txnid;
4093 memset(&ov, 0, sizeof(ov));
4095 WriteFile(env->me_fd, ptr, len, NULL, &ov);
4097 r2 = pwrite(env->me_fd, ptr, len, off);
4098 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
4101 env->me_flags |= MDB_FATAL_ERROR;
4104 /* MIPS has cache coherency issues, this is a no-op everywhere else */
4105 CACHEFLUSH(env->me_map + off, len, DCACHE);
4107 /* Memory ordering issues are irrelevant; since the entire writer
4108 * is wrapped by wmutex, all of these changes will become visible
4109 * after the wmutex is unlocked. Since the DB is multi-version,
4110 * readers will get consistent data regardless of how fresh or
4111 * how stale their view of these values is.
4114 env->me_txns->mti_txnid = txn->mt_txnid;
4119 /** Check both meta pages to see which one is newer.
4120 * @param[in] env the environment handle
4121 * @return newest #MDB_meta.
4124 mdb_env_pick_meta(const MDB_env *env)
4126 MDB_meta *const *metas = env->me_metas;
4127 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
4131 mdb_env_create(MDB_env **env)
4135 e = calloc(1, sizeof(MDB_env));
4139 e->me_maxreaders = DEFAULT_READERS;
4140 e->me_maxdbs = e->me_numdbs = CORE_DBS;
4141 e->me_fd = INVALID_HANDLE_VALUE;
4142 e->me_lfd = INVALID_HANDLE_VALUE;
4143 e->me_mfd = INVALID_HANDLE_VALUE;
4144 #ifdef MDB_USE_POSIX_SEM
4145 e->me_rmutex = SEM_FAILED;
4146 e->me_wmutex = SEM_FAILED;
4147 #elif defined MDB_USE_SYSV_SEM
4148 e->me_rmutex->semid = -1;
4149 e->me_wmutex->semid = -1;
4151 e->me_pid = getpid();
4152 GET_PAGESIZE(e->me_os_psize);
4153 VGMEMP_CREATE(e,0,0);
4159 /** @brief Map a result from an NTAPI call to WIN32. */
4161 mdb_nt2win32(NTSTATUS st)
4166 GetOverlappedResult(NULL, &o, &br, FALSE);
4167 return GetLastError();
4172 mdb_env_map(MDB_env *env, void *addr)
4175 unsigned int flags = env->me_flags;
4178 int access = SECTION_MAP_READ;
4182 ULONG pageprot = PAGE_READONLY, secprot, alloctype;
4184 if (flags & MDB_WRITEMAP) {
4185 access |= SECTION_MAP_WRITE;
4186 pageprot = PAGE_READWRITE;
4188 if (flags & MDB_RDONLY) {
4189 secprot = PAGE_READONLY;
4193 secprot = PAGE_READWRITE;
4194 msize = env->me_mapsize;
4195 alloctype = MEM_RESERVE;
4198 rc = NtCreateSection(&mh, access, NULL, NULL, secprot, SEC_RESERVE, env->me_fd);
4200 return mdb_nt2win32(rc);
4203 msize = NUM_METAS * env->me_psize;
4205 rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
4212 return mdb_nt2win32(rc);
4217 env->me_map = mmap(addr, NUM_METAS * env->me_psize, PROT_READ, MAP_SHARED,
4219 if (env->me_map == MAP_FAILED) {
4224 int prot = PROT_READ;
4225 if (flags & MDB_WRITEMAP) {
4227 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
4230 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
4232 if (env->me_map == MAP_FAILED) {
4237 if (flags & MDB_NORDAHEAD) {
4238 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
4240 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
4242 #ifdef POSIX_MADV_RANDOM
4243 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
4244 #endif /* POSIX_MADV_RANDOM */
4245 #endif /* MADV_RANDOM */
4249 /* Can happen because the address argument to mmap() is just a
4250 * hint. mmap() can pick another, e.g. if the range is in use.
4251 * The MAP_FIXED flag would prevent that, but then mmap could
4252 * instead unmap existing pages to make room for the new map.
4254 if (addr && env->me_map != addr)
4255 return EBUSY; /* TODO: Make a new MDB_* error code? */
4258 p = (MDB_page *)env->me_map;
4259 env->me_metas[0] = METADATA(p);
4260 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4266 mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
4268 /* If env is already open, caller is responsible for making
4269 * sure there are no active txns.
4279 meta = mdb_env_pick_meta(env);
4281 size = meta->mm_mapsize;
4283 /* Silently round up to minimum if the size is too small */
4284 mdb_size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4289 /* For MDB_VL32 this bit is a noop since we dynamically remap
4290 * chunks of the DB anyway.
4292 munmap(env->me_map, env->me_mapsize);
4293 env->me_mapsize = size;
4294 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4295 rc = mdb_env_map(env, old);
4298 #endif /* !MDB_VL32 */
4300 env->me_mapsize = size;
4302 env->me_maxpg = env->me_mapsize / env->me_psize;
4307 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4311 env->me_maxdbs = dbs + CORE_DBS;
4316 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4318 if (env->me_map || readers < 1)
4320 env->me_maxreaders = readers;
4325 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4327 if (!env || !readers)
4329 *readers = env->me_maxreaders;
4334 mdb_fsize(HANDLE fd, mdb_size_t *size)
4337 LARGE_INTEGER fsize;
4339 if (!GetFileSizeEx(fd, &fsize))
4342 *size = fsize.QuadPart;
4354 #ifdef BROKEN_FDATASYNC
4355 #include <sys/utsname.h>
4356 #include <sys/vfs.h>
4359 /** Further setup required for opening an LMDB environment
4362 mdb_env_open2(MDB_env *env)
4364 unsigned int flags = env->me_flags;
4365 int i, newenv = 0, rc;
4369 /* See if we should use QueryLimited */
4371 if ((rc & 0xff) > 5)
4372 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4374 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4377 #ifdef BROKEN_FDATASYNC
4378 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4379 * https://lkml.org/lkml/2012/9/3/83
4380 * Kernels after 3.6-rc6 are known good.
4381 * https://lkml.org/lkml/2012/9/10/556
4382 * See if the DB is on ext3/ext4, then check for new enough kernel
4383 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4388 fstatfs(env->me_fd, &st);
4389 while (st.f_type == 0xEF53) {
4393 if (uts.release[0] < '3') {
4394 if (!strncmp(uts.release, "2.6.32.", 7)) {
4395 i = atoi(uts.release+7);
4397 break; /* 2.6.32.60 and newer is OK */
4398 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4399 i = atoi(uts.release+7);
4401 break; /* 2.6.34.15 and newer is OK */
4403 } else if (uts.release[0] == '3') {
4404 i = atoi(uts.release+2);
4406 break; /* 3.6 and newer is OK */
4408 i = atoi(uts.release+4);
4410 break; /* 3.5.4 and newer is OK */
4411 } else if (i == 2) {
4412 i = atoi(uts.release+4);
4414 break; /* 3.2.30 and newer is OK */
4416 } else { /* 4.x and newer is OK */
4419 env->me_flags |= MDB_FSYNCONLY;
4425 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4428 DPUTS("new mdbenv");
4430 env->me_psize = env->me_os_psize;
4431 if (env->me_psize > MAX_PAGESIZE)
4432 env->me_psize = MAX_PAGESIZE;
4433 memset(&meta, 0, sizeof(meta));
4434 mdb_env_init_meta0(env, &meta);
4435 meta.mm_mapsize = DEFAULT_MAPSIZE;
4437 env->me_psize = meta.mm_psize;
4440 /* Was a mapsize configured? */
4441 if (!env->me_mapsize) {
4442 env->me_mapsize = meta.mm_mapsize;
4445 /* Make sure mapsize >= committed data size. Even when using
4446 * mm_mapsize, which could be broken in old files (ITS#7789).
4448 mdb_size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4449 if (env->me_mapsize < minsize)
4450 env->me_mapsize = minsize;
4452 meta.mm_mapsize = env->me_mapsize;
4454 if (newenv && !(flags & MDB_FIXEDMAP)) {
4455 /* mdb_env_map() may grow the datafile. Write the metapages
4456 * first, so the file will be valid if initialization fails.
4457 * Except with FIXEDMAP, since we do not yet know mm_address.
4458 * We could fill in mm_address later, but then a different
4459 * program might end up doing that - one with a memory layout
4460 * and map address which does not suit the main program.
4462 rc = mdb_env_init_meta(env, &meta);
4468 /* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
4472 rc = WriteFile(env->me_fd, &dummy, 1, &len, NULL);
4480 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4485 if (flags & MDB_FIXEDMAP)
4486 meta.mm_address = env->me_map;
4487 i = mdb_env_init_meta(env, &meta);
4488 if (i != MDB_SUCCESS) {
4493 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4494 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4496 #if !(MDB_MAXKEYSIZE)
4497 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4499 env->me_maxpg = env->me_mapsize / env->me_psize;
4503 MDB_meta *meta = mdb_env_pick_meta(env);
4504 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4506 DPRINTF(("opened database version %u, pagesize %u",
4507 meta->mm_version, env->me_psize));
4508 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4509 DPRINTF(("depth: %u", db->md_depth));
4510 DPRINTF(("entries: %"Y"u", db->md_entries));
4511 DPRINTF(("branch pages: %"Y"u", db->md_branch_pages));
4512 DPRINTF(("leaf pages: %"Y"u", db->md_leaf_pages));
4513 DPRINTF(("overflow pages: %"Y"u", db->md_overflow_pages));
4514 DPRINTF(("root: %"Y"u", db->md_root));
4522 /** Release a reader thread's slot in the reader lock table.
4523 * This function is called automatically when a thread exits.
4524 * @param[in] ptr This points to the slot in the reader lock table.
4527 mdb_env_reader_dest(void *ptr)
4529 MDB_reader *reader = ptr;
4535 /** Junk for arranging thread-specific callbacks on Windows. This is
4536 * necessarily platform and compiler-specific. Windows supports up
4537 * to 1088 keys. Let's assume nobody opens more than 64 environments
4538 * in a single process, for now. They can override this if needed.
4540 #ifndef MAX_TLS_KEYS
4541 #define MAX_TLS_KEYS 64
4543 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4544 static int mdb_tls_nkeys;
4546 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4550 case DLL_PROCESS_ATTACH: break;
4551 case DLL_THREAD_ATTACH: break;
4552 case DLL_THREAD_DETACH:
4553 for (i=0; i<mdb_tls_nkeys; i++) {
4554 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4556 mdb_env_reader_dest(r);
4560 case DLL_PROCESS_DETACH: break;
4565 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4567 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4571 /* Force some symbol references.
4572 * _tls_used forces the linker to create the TLS directory if not already done
4573 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4575 #pragma comment(linker, "/INCLUDE:_tls_used")
4576 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4577 #pragma const_seg(".CRT$XLB")
4578 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4579 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4582 #pragma comment(linker, "/INCLUDE:__tls_used")
4583 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4584 #pragma data_seg(".CRT$XLB")
4585 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4587 #endif /* WIN 32/64 */
4588 #endif /* !__GNUC__ */
4591 /** Downgrade the exclusive lock on the region back to shared */
4593 mdb_env_share_locks(MDB_env *env, int *excl)
4596 MDB_meta *meta = mdb_env_pick_meta(env);
4598 env->me_txns->mti_txnid = meta->mm_txnid;
4603 /* First acquire a shared lock. The Unlock will
4604 * then release the existing exclusive lock.
4606 memset(&ov, 0, sizeof(ov));
4607 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4610 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4616 struct flock lock_info;
4617 /* The shared lock replaces the existing lock */
4618 memset((void *)&lock_info, 0, sizeof(lock_info));
4619 lock_info.l_type = F_RDLCK;
4620 lock_info.l_whence = SEEK_SET;
4621 lock_info.l_start = 0;
4622 lock_info.l_len = 1;
4623 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4624 (rc = ErrCode()) == EINTR) ;
4625 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4632 /** Try to get exclusive lock, otherwise shared.
4633 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4636 mdb_env_excl_lock(MDB_env *env, int *excl)
4640 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4644 memset(&ov, 0, sizeof(ov));
4645 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4652 struct flock lock_info;
4653 memset((void *)&lock_info, 0, sizeof(lock_info));
4654 lock_info.l_type = F_WRLCK;
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) ;
4663 # ifndef MDB_USE_POSIX_MUTEX
4664 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4667 lock_info.l_type = F_RDLCK;
4668 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4669 (rc = ErrCode()) == EINTR) ;
4679 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4681 * @(#) $Revision: 5.1 $
4682 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4683 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4685 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4689 * Please do not copyright this code. This code is in the public domain.
4691 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4692 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4693 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4694 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4695 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4696 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4697 * PERFORMANCE OF THIS SOFTWARE.
4700 * chongo <Landon Curt Noll> /\oo/\
4701 * http://www.isthe.com/chongo/
4703 * Share and Enjoy! :-)
4706 typedef unsigned long long mdb_hash_t;
4707 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4709 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4710 * @param[in] val value to hash
4711 * @param[in] hval initial value for hash
4712 * @return 64 bit hash
4714 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4715 * hval arg on the first call.
4718 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4720 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4721 unsigned char *end = s + val->mv_size;
4723 * FNV-1a hash each octet of the string
4726 /* xor the bottom with the current octet */
4727 hval ^= (mdb_hash_t)*s++;
4729 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4730 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4731 (hval << 7) + (hval << 8) + (hval << 40);
4733 /* return our new hash value */
4737 /** Hash the string and output the encoded hash.
4738 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4739 * very short name limits. We don't care about the encoding being reversible,
4740 * we just want to preserve as many bits of the input as possible in a
4741 * small printable string.
4742 * @param[in] str string to hash
4743 * @param[out] encbuf an array of 11 chars to hold the hash
4745 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4748 mdb_pack85(unsigned long l, char *out)
4752 for (i=0; i<5; i++) {
4753 *out++ = mdb_a85[l % 85];
4759 mdb_hash_enc(MDB_val *val, char *encbuf)
4761 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4763 mdb_pack85(h, encbuf);
4764 mdb_pack85(h>>32, encbuf+5);
4769 /** Open and/or initialize the lock region for the environment.
4770 * @param[in] env The LMDB environment.
4771 * @param[in] lpath The pathname of the file used for the lock region.
4772 * @param[in] mode The Unix permissions for the file, if we create it.
4773 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4774 * @return 0 on success, non-zero on failure.
4777 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4780 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4782 # define MDB_ERRCODE_ROFS EROFS
4783 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4784 # define MDB_CLOEXEC O_CLOEXEC
4787 # define MDB_CLOEXEC 0
4790 #ifdef MDB_USE_SYSV_SEM
4799 rc = utf8_to_utf16(lpath, -1, &wlpath, NULL);
4802 env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
4803 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4804 FILE_ATTRIBUTE_NORMAL, NULL);
4807 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4809 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4811 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4816 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4817 /* Lose record locks when exec*() */
4818 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4819 fcntl(env->me_lfd, F_SETFD, fdflags);
4822 if (!(env->me_flags & MDB_NOTLS)) {
4823 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4826 env->me_flags |= MDB_ENV_TXKEY;
4828 /* Windows TLS callbacks need help finding their TLS info. */
4829 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4833 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4837 /* Try to get exclusive lock. If we succeed, then
4838 * nobody is using the lock region and we should initialize it.
4840 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4843 size = GetFileSize(env->me_lfd, NULL);
4845 size = lseek(env->me_lfd, 0, SEEK_END);
4846 if (size == -1) goto fail_errno;
4848 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4849 if (size < rsize && *excl > 0) {
4851 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4852 || !SetEndOfFile(env->me_lfd))
4855 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4859 size = rsize - sizeof(MDB_txninfo);
4860 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4865 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4867 if (!mh) goto fail_errno;
4868 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4870 if (!env->me_txns) goto fail_errno;
4872 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4874 if (m == MAP_FAILED) goto fail_errno;
4880 BY_HANDLE_FILE_INFORMATION stbuf;
4889 if (!mdb_sec_inited) {
4890 InitializeSecurityDescriptor(&mdb_null_sd,
4891 SECURITY_DESCRIPTOR_REVISION);
4892 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4893 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4894 mdb_all_sa.bInheritHandle = FALSE;
4895 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4898 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4899 idbuf.volume = stbuf.dwVolumeSerialNumber;
4900 idbuf.nhigh = stbuf.nFileIndexHigh;
4901 idbuf.nlow = stbuf.nFileIndexLow;
4902 val.mv_data = &idbuf;
4903 val.mv_size = sizeof(idbuf);
4904 mdb_hash_enc(&val, encbuf);
4905 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4906 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4907 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4908 if (!env->me_rmutex) goto fail_errno;
4909 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4910 if (!env->me_wmutex) goto fail_errno;
4911 #elif defined(MDB_USE_POSIX_SEM)
4920 #if defined(__NetBSD__)
4921 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4923 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4924 idbuf.dev = stbuf.st_dev;
4925 idbuf.ino = stbuf.st_ino;
4926 val.mv_data = &idbuf;
4927 val.mv_size = sizeof(idbuf);
4928 mdb_hash_enc(&val, encbuf);
4929 #ifdef MDB_SHORT_SEMNAMES
4930 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4932 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4933 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4934 /* Clean up after a previous run, if needed: Try to
4935 * remove both semaphores before doing anything else.
4937 sem_unlink(env->me_txns->mti_rmname);
4938 sem_unlink(env->me_txns->mti_wmname);
4939 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4940 O_CREAT|O_EXCL, mode, 1);
4941 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4942 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4943 O_CREAT|O_EXCL, mode, 1);
4944 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4945 #elif defined(MDB_USE_SYSV_SEM)
4946 unsigned short vals[2] = {1, 1};
4947 key_t key = ftok(lpath, 'M');
4950 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4954 if (semctl(semid, 0, SETALL, semu) < 0)
4956 env->me_txns->mti_semid = semid;
4957 env->me_txns->mti_rlocked = 0;
4958 env->me_txns->mti_wlocked = 0;
4959 #else /* MDB_USE_POSIX_MUTEX: */
4960 pthread_mutexattr_t mattr;
4962 /* Solaris needs this before initing a robust mutex. Otherwise
4963 * it may skip the init and return EBUSY "seems someone already
4964 * inited" or EINVAL "it was inited differently".
4966 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
4967 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
4969 if ((rc = pthread_mutexattr_init(&mattr)) != 0)
4971 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
4972 #ifdef MDB_ROBUST_SUPPORTED
4973 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
4975 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
4976 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
4977 pthread_mutexattr_destroy(&mattr);
4980 #endif /* _WIN32 || ... */
4982 env->me_txns->mti_magic = MDB_MAGIC;
4983 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4984 env->me_txns->mti_txnid = 0;
4985 env->me_txns->mti_numreaders = 0;
4988 #ifdef MDB_USE_SYSV_SEM
4989 struct semid_ds buf;
4991 if (env->me_txns->mti_magic != MDB_MAGIC) {
4992 DPUTS("lock region has invalid magic");
4996 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4997 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4998 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4999 rc = MDB_VERSION_MISMATCH;
5003 if (rc && rc != EACCES && rc != EAGAIN) {
5007 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
5008 if (!env->me_rmutex) goto fail_errno;
5009 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
5010 if (!env->me_wmutex) goto fail_errno;
5011 #elif defined(MDB_USE_POSIX_SEM)
5012 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
5013 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
5014 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
5015 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
5016 #elif defined(MDB_USE_SYSV_SEM)
5017 semid = env->me_txns->mti_semid;
5019 /* check for read access */
5020 if (semctl(semid, 0, IPC_STAT, semu) < 0)
5022 /* check for write access */
5023 if (semctl(semid, 0, IPC_SET, semu) < 0)
5027 #ifdef MDB_USE_SYSV_SEM
5028 env->me_rmutex->semid = semid;
5029 env->me_wmutex->semid = semid;
5030 env->me_rmutex->semnum = 0;
5031 env->me_wmutex->semnum = 1;
5032 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
5033 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
5037 env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
5039 pthread_mutex_init(&env->me_rpmutex, NULL);
5051 /** The name of the lock file in the DB environment */
5052 #define LOCKNAME "/lock.mdb"
5053 /** The name of the data file in the DB environment */
5054 #define DATANAME "/data.mdb"
5055 /** The suffix of the lock file when no subdir is used */
5056 #define LOCKSUFF "-lock"
5057 /** Only a subset of the @ref mdb_env flags can be changed
5058 * at runtime. Changing other flags requires closing the
5059 * environment and re-opening it with the new flags.
5061 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
5062 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
5063 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
5065 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
5066 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
5070 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
5072 int oflags, rc, len, excl = -1;
5073 char *lpath, *dpath;
5078 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
5082 if (flags & MDB_WRITEMAP) {
5083 /* silently ignore WRITEMAP in 32 bit mode */
5084 flags ^= MDB_WRITEMAP;
5086 if (flags & MDB_FIXEDMAP) {
5087 /* cannot support FIXEDMAP */
5093 if (flags & MDB_NOSUBDIR) {
5094 rc = len + sizeof(LOCKSUFF) + len + 1;
5096 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
5101 if (flags & MDB_NOSUBDIR) {
5102 dpath = lpath + len + sizeof(LOCKSUFF);
5103 sprintf(lpath, "%s" LOCKSUFF, path);
5104 strcpy(dpath, path);
5106 dpath = lpath + len + sizeof(LOCKNAME);
5107 sprintf(lpath, "%s" LOCKNAME, path);
5108 sprintf(dpath, "%s" DATANAME, path);
5112 flags |= env->me_flags;
5113 if (flags & MDB_RDONLY) {
5114 /* silently ignore WRITEMAP when we're only getting read access */
5115 flags &= ~MDB_WRITEMAP;
5117 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
5118 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
5123 env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
5124 if (!env->me_rpages) {
5128 env->me_rpages[0].mid = 0;
5129 env->me_rpcheck = MDB_ERPAGE_SIZE/2;
5132 env->me_flags = flags |= MDB_ENV_ACTIVE;
5136 env->me_path = strdup(path);
5137 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
5138 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
5139 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
5140 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
5144 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
5146 /* For RDONLY, get lockfile after we know datafile exists */
5147 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
5148 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
5154 if (F_ISSET(flags, MDB_RDONLY)) {
5155 oflags = GENERIC_READ;
5156 len = OPEN_EXISTING;
5158 oflags = GENERIC_READ|GENERIC_WRITE;
5161 mode = FILE_ATTRIBUTE_NORMAL;
5162 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
5165 env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
5166 NULL, len, mode, NULL);
5169 if (F_ISSET(flags, MDB_RDONLY))
5172 oflags = O_RDWR | O_CREAT;
5174 env->me_fd = open(dpath, oflags, mode);
5176 if (env->me_fd == INVALID_HANDLE_VALUE) {
5181 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
5182 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
5187 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
5188 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
5189 env->me_mfd = env->me_fd;
5191 /* Synchronous fd for meta writes. Needed even with
5192 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
5195 len = OPEN_EXISTING;
5196 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
5199 env->me_mfd = CreateFileW(wpath, oflags,
5200 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
5201 mode | FILE_FLAG_WRITE_THROUGH, NULL);
5205 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
5207 if (env->me_mfd == INVALID_HANDLE_VALUE) {
5212 DPRINTF(("opened dbenv %p", (void *) env));
5214 rc = mdb_env_share_locks(env, &excl);
5218 if (!(flags & MDB_RDONLY)) {
5220 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
5221 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
5222 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
5223 (txn = calloc(1, size)))
5225 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
5226 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
5227 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
5228 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
5231 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
5232 if (!txn->mt_rpages) {
5237 txn->mt_rpages[0].mid = 0;
5238 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
5240 txn->mt_dbxs = env->me_dbxs;
5241 txn->mt_flags = MDB_TXN_FINISHED;
5251 mdb_env_close0(env, excl);
5257 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5259 mdb_env_close0(MDB_env *env, int excl)
5263 if (!(env->me_flags & MDB_ENV_ACTIVE))
5266 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5268 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5269 free(env->me_dbxs[i].md_name.mv_data);
5274 free(env->me_dbiseqs);
5275 free(env->me_dbflags);
5277 free(env->me_dirty_list);
5279 if (env->me_txn0 && env->me_txn0->mt_rpages)
5280 free(env->me_txn0->mt_rpages);
5282 for (x=1; x<=env->me_rpages[0].mid; x++)
5283 munmap(env->me_rpages[x].mptr, env->me_rpages[x].mcnt * env->me_psize);
5285 free(env->me_rpages);
5288 mdb_midl_free(env->me_free_pgs);
5290 if (env->me_flags & MDB_ENV_TXKEY) {
5291 pthread_key_delete(env->me_txkey);
5293 /* Delete our key from the global list */
5294 for (i=0; i<mdb_tls_nkeys; i++)
5295 if (mdb_tls_keys[i] == env->me_txkey) {
5296 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5305 munmap(env->me_map, NUM_METAS*env->me_psize);
5307 munmap(env->me_map, env->me_mapsize);
5310 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
5311 (void) close(env->me_mfd);
5312 if (env->me_fd != INVALID_HANDLE_VALUE)
5313 (void) close(env->me_fd);
5315 MDB_PID_T pid = env->me_pid;
5316 /* Clearing readers is done in this function because
5317 * me_txkey with its destructor must be disabled first.
5319 * We skip the the reader mutex, so we touch only
5320 * data owned by this process (me_close_readers and
5321 * our readers), and clear each reader atomically.
5323 for (i = env->me_close_readers; --i >= 0; )
5324 if (env->me_txns->mti_readers[i].mr_pid == pid)
5325 env->me_txns->mti_readers[i].mr_pid = 0;
5327 if (env->me_rmutex) {
5328 CloseHandle(env->me_rmutex);
5329 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5331 /* Windows automatically destroys the mutexes when
5332 * the last handle closes.
5334 #elif defined(MDB_USE_POSIX_SEM)
5335 if (env->me_rmutex != SEM_FAILED) {
5336 sem_close(env->me_rmutex);
5337 if (env->me_wmutex != SEM_FAILED)
5338 sem_close(env->me_wmutex);
5339 /* If we have the filelock: If we are the
5340 * only remaining user, clean up semaphores.
5343 mdb_env_excl_lock(env, &excl);
5345 sem_unlink(env->me_txns->mti_rmname);
5346 sem_unlink(env->me_txns->mti_wmname);
5349 #elif defined(MDB_USE_SYSV_SEM)
5350 if (env->me_rmutex->semid != -1) {
5351 /* If we have the filelock: If we are the
5352 * only remaining user, clean up semaphores.
5355 mdb_env_excl_lock(env, &excl);
5357 semctl(env->me_rmutex->semid, 0, IPC_RMID);
5360 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5362 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5365 /* Unlock the lockfile. Windows would have unlocked it
5366 * after closing anyway, but not necessarily at once.
5368 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5371 (void) close(env->me_lfd);
5375 if (env->me_fmh) CloseHandle(env->me_fmh);
5376 if (env->me_rpmutex) CloseHandle(env->me_rpmutex);
5378 pthread_mutex_destroy(&env->me_rpmutex);
5382 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5386 mdb_env_close(MDB_env *env)
5393 VGMEMP_DESTROY(env);
5394 while ((dp = env->me_dpages) != NULL) {
5395 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5396 env->me_dpages = dp->mp_next;
5400 mdb_env_close0(env, 0);
5404 /** Compare two items pointing at aligned mdb_size_t's */
5406 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5408 return (*(mdb_size_t *)a->mv_data < *(mdb_size_t *)b->mv_data) ? -1 :
5409 *(mdb_size_t *)a->mv_data > *(mdb_size_t *)b->mv_data;
5412 /** Compare two items pointing at aligned unsigned int's.
5414 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5415 * but #mdb_cmp_clong() is called instead if the data type is mdb_size_t.
5418 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5420 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5421 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5424 /** Compare two items pointing at unsigned ints of unknown alignment.
5425 * Nodes and keys are guaranteed to be 2-byte aligned.
5428 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5430 #if BYTE_ORDER == LITTLE_ENDIAN
5431 unsigned short *u, *c;
5434 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5435 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5438 } while(!x && u > (unsigned short *)a->mv_data);
5441 unsigned short *u, *c, *end;
5444 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5445 u = (unsigned short *)a->mv_data;
5446 c = (unsigned short *)b->mv_data;
5449 } while(!x && u < end);
5454 /** Compare two items lexically */
5456 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5463 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5469 diff = memcmp(a->mv_data, b->mv_data, len);
5470 return diff ? diff : len_diff<0 ? -1 : len_diff;
5473 /** Compare two items in reverse byte order */
5475 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5477 const unsigned char *p1, *p2, *p1_lim;
5481 p1_lim = (const unsigned char *)a->mv_data;
5482 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5483 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5485 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5491 while (p1 > p1_lim) {
5492 diff = *--p1 - *--p2;
5496 return len_diff<0 ? -1 : len_diff;
5499 /** Search for key within a page, using binary search.
5500 * Returns the smallest entry larger or equal to the key.
5501 * If exactp is non-null, stores whether the found entry was an exact match
5502 * in *exactp (1 or 0).
5503 * Updates the cursor index with the index of the found entry.
5504 * If no entry larger or equal to the key is found, returns NULL.
5507 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5509 unsigned int i = 0, nkeys;
5512 MDB_page *mp = mc->mc_pg[mc->mc_top];
5513 MDB_node *node = NULL;
5518 nkeys = NUMKEYS(mp);
5520 DPRINTF(("searching %u keys in %s %spage %"Y"u",
5521 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5524 low = IS_LEAF(mp) ? 0 : 1;
5526 cmp = mc->mc_dbx->md_cmp;
5528 /* Branch pages have no data, so if using integer keys,
5529 * alignment is guaranteed. Use faster mdb_cmp_int.
5531 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5532 if (NODEPTR(mp, 1)->mn_ksize == sizeof(mdb_size_t))
5539 nodekey.mv_size = mc->mc_db->md_pad;
5540 node = NODEPTR(mp, 0); /* fake */
5541 while (low <= high) {
5542 i = (low + high) >> 1;
5543 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5544 rc = cmp(key, &nodekey);
5545 DPRINTF(("found leaf index %u [%s], rc = %i",
5546 i, DKEY(&nodekey), rc));
5555 while (low <= high) {
5556 i = (low + high) >> 1;
5558 node = NODEPTR(mp, i);
5559 nodekey.mv_size = NODEKSZ(node);
5560 nodekey.mv_data = NODEKEY(node);
5562 rc = cmp(key, &nodekey);
5565 DPRINTF(("found leaf index %u [%s], rc = %i",
5566 i, DKEY(&nodekey), rc));
5568 DPRINTF(("found branch index %u [%s -> %"Y"u], rc = %i",
5569 i, DKEY(&nodekey), NODEPGNO(node), rc));
5580 if (rc > 0) { /* Found entry is less than the key. */
5581 i++; /* Skip to get the smallest entry larger than key. */
5583 node = NODEPTR(mp, i);
5586 *exactp = (rc == 0 && nkeys > 0);
5587 /* store the key index */
5588 mc->mc_ki[mc->mc_top] = i;
5590 /* There is no entry larger or equal to the key. */
5593 /* nodeptr is fake for LEAF2 */
5599 mdb_cursor_adjust(MDB_cursor *mc, func)
5603 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5604 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5611 /** Pop a page off the top of the cursor's stack. */
5613 mdb_cursor_pop(MDB_cursor *mc)
5616 DPRINTF(("popping page %"Y"u off db %d cursor %p",
5617 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5623 mc->mc_flags &= ~C_INITIALIZED;
5628 /** Push a page onto the top of the cursor's stack. */
5630 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5632 DPRINTF(("pushing page %"Y"u on db %d cursor %p", mp->mp_pgno,
5633 DDBI(mc), (void *) mc));
5635 if (mc->mc_snum >= CURSOR_STACK) {
5636 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5637 return MDB_CURSOR_FULL;
5640 mc->mc_top = mc->mc_snum++;
5641 mc->mc_pg[mc->mc_top] = mp;
5642 mc->mc_ki[mc->mc_top] = 0;
5648 /** Map a read-only page.
5649 * There are two levels of tracking in use, a per-txn list and a per-env list.
5650 * ref'ing and unref'ing the per-txn list is faster since it requires no
5651 * locking. Pages are cached in the per-env list for global reuse, and a lock
5652 * is required. Pages are not immediately unmapped when their refcnt goes to
5653 * zero; they hang around in case they will be reused again soon.
5655 * When the per-txn list gets full, all pages with refcnt=0 are purged from the
5656 * list and their refcnts in the per-env list are decremented.
5658 * When the per-env list gets full, all pages with refcnt=0 are purged from the
5659 * list and their pages are unmapped.
5661 * @note "full" means the list has reached its respective rpcheck threshold.
5662 * This threshold slowly raises if no pages could be purged on a given check,
5663 * and returns to its original value when enough pages were purged.
5665 * If purging doesn't free any slots, filling the per-txn list will return
5666 * MDB_TXN_FULL, and filling the per-env list returns MDB_MAP_FULL.
5668 * Reference tracking in a txn is imperfect, pages can linger with non-zero
5669 * refcnt even without active references. It was deemed to be too invasive
5670 * to add unrefs in every required location. However, all pages are unref'd
5671 * at the end of the transaction. This guarantees that no stale references
5672 * linger in the per-env list.
5674 * Usually we map chunks of 16 pages at a time, but if an overflow page begins
5675 * at the tail of the chunk we extend the chunk to include the entire overflow
5676 * page. Unfortunately, pages can be turned into overflow pages after their
5677 * chunk was already mapped. In that case we must remap the chunk if the
5678 * overflow page is referenced. If the chunk's refcnt is 0 we can just remap
5679 * it, otherwise we temporarily map a new chunk just for the overflow page.
5681 * @note this chunk handling means we cannot guarantee that a data item
5682 * returned from the DB will stay alive for the duration of the transaction:
5683 * We unref pages as soon as a cursor moves away from the page
5684 * A subsequent op may cause a purge, which may unmap any unref'd chunks
5685 * The caller must copy the data if it must be used later in the same txn.
5687 * Also - our reference counting revolves around cursors, but overflow pages
5688 * aren't pointed to by a cursor's page stack. We have to remember them
5689 * explicitly, in the added mc_ovpg field. A single cursor can only hold a
5690 * reference to one overflow page at a time.
5692 * @param[in] txn the transaction for this access.
5693 * @param[in] pgno the page number for the page to retrieve.
5694 * @param[out] ret address of a pointer where the page's address will be stored.
5695 * @return 0 on success, non-zero on failure.
5698 mdb_rpage_get(MDB_txn *txn, pgno_t pg0, MDB_page **ret)
5700 MDB_env *env = txn->mt_env;
5702 MDB_ID3L tl = txn->mt_rpages;
5703 MDB_ID3L el = env->me_rpages;
5707 int rc, retries = 1;
5711 #define SET_OFF(off,val) off.QuadPart = val
5712 #define MAP(rc,env,addr,len,off) \
5714 rc = NtMapViewOfSection(env->me_fmh, GetCurrentProcess(), &addr, 0, \
5715 len, &off, &len, ViewUnmap, (env->me_flags & MDB_RDONLY) ? 0 : MEM_RESERVE, PAGE_READONLY); \
5716 if (rc) rc = mdb_nt2win32(rc)
5720 #define SET_OFF(off,val) off = val
5721 #define MAP(rc,env,addr,len,off) \
5722 addr = mmap(NULL, len, PROT_READ, MAP_SHARED, env->me_fd, off); \
5723 rc = (addr == MAP_FAILED) ? errno : 0
5726 /* remember the offset of the actual page number, so we can
5727 * return the correct pointer at the end.
5729 rem = pg0 & (MDB_RPAGE_CHUNK-1);
5733 x = mdb_mid3l_search(tl, pgno);
5734 if (x <= tl[0].mid && tl[x].mid == pgno) {
5735 if (x != tl[0].mid && tl[x+1].mid == pg0)
5737 /* check for overflow size */
5738 p = (MDB_page *)((char *)tl[x].mptr + rem * env->me_psize);
5739 if (IS_OVERFLOW(p) && p->mp_pages + rem > tl[x].mcnt) {
5740 id3.mcnt = p->mp_pages + rem;
5741 len = id3.mcnt * env->me_psize;
5742 SET_OFF(off, pgno * env->me_psize);
5743 MAP(rc, env, id3.mptr, len, off);
5746 /* check for local-only page */
5748 mdb_tassert(txn, tl[x].mid != pg0);
5749 /* hope there's room to insert this locally.
5750 * setting mid here tells later code to just insert
5751 * this id3 instead of searching for a match.
5756 /* ignore the mapping we got from env, use new one */
5757 tl[x].mptr = id3.mptr;
5758 tl[x].mcnt = id3.mcnt;
5759 /* if no active ref, see if we can replace in env */
5762 pthread_mutex_lock(&env->me_rpmutex);
5763 i = mdb_mid3l_search(el, tl[x].mid);
5764 if (el[i].mref == 1) {
5765 /* just us, replace it */
5766 munmap(el[i].mptr, el[i].mcnt * env->me_psize);
5767 el[i].mptr = tl[x].mptr;
5768 el[i].mcnt = tl[x].mcnt;
5770 /* there are others, remove ourself */
5773 pthread_mutex_unlock(&env->me_rpmutex);
5777 id3.mptr = tl[x].mptr;
5778 id3.mcnt = tl[x].mcnt;
5784 if (tl[0].mid >= MDB_TRPAGE_MAX - txn->mt_rpcheck) {
5786 /* purge unref'd pages from our list and unref in env */
5787 pthread_mutex_lock(&env->me_rpmutex);
5790 for (i=1; i<=tl[0].mid; i++) {
5793 /* tmp overflow pages don't go to env */
5794 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
5795 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
5798 x = mdb_mid3l_search(el, tl[i].mid);
5802 pthread_mutex_unlock(&env->me_rpmutex);
5804 /* we didn't find any unref'd chunks.
5805 * if we're out of room, fail.
5807 if (tl[0].mid >= MDB_TRPAGE_MAX)
5808 return MDB_TXN_FULL;
5809 /* otherwise, raise threshold for next time around
5812 txn->mt_rpcheck /= 2;
5814 /* we found some unused; consolidate the list */
5815 for (i=y+1; i<= tl[0].mid; i++)
5819 /* decrease the check threshold toward its original value */
5820 if (!txn->mt_rpcheck)
5821 txn->mt_rpcheck = 1;
5822 while (txn->mt_rpcheck < tl[0].mid && txn->mt_rpcheck < MDB_TRPAGE_SIZE/2)
5823 txn->mt_rpcheck *= 2;
5826 if (tl[0].mid < MDB_TRPAGE_SIZE) {
5830 /* don't map past last written page in read-only envs */
5831 if ((env->me_flags & MDB_RDONLY) && pgno + MDB_RPAGE_CHUNK-1 > txn->mt_last_pgno)
5832 id3.mcnt = txn->mt_last_pgno + 1 - pgno;
5834 id3.mcnt = MDB_RPAGE_CHUNK;
5835 len = id3.mcnt * env->me_psize;
5838 /* search for page in env */
5839 pthread_mutex_lock(&env->me_rpmutex);
5840 x = mdb_mid3l_search(el, pgno);
5841 if (x <= el[0].mid && el[x].mid == pgno) {
5842 id3.mptr = el[x].mptr;
5843 id3.mcnt = el[x].mcnt;
5844 /* check for overflow size */
5845 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5846 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
5847 id3.mcnt = p->mp_pages + rem;
5848 len = id3.mcnt * env->me_psize;
5849 SET_OFF(off, pgno * env->me_psize);
5850 MAP(rc, env, id3.mptr, len, off);
5854 munmap(el[x].mptr, env->me_psize * el[x].mcnt);
5855 el[x].mptr = id3.mptr;
5856 el[x].mcnt = id3.mcnt;
5859 pthread_mutex_unlock(&env->me_rpmutex);
5864 pthread_mutex_unlock(&env->me_rpmutex);
5867 if (el[0].mid >= MDB_ERPAGE_MAX - env->me_rpcheck) {
5868 /* purge unref'd pages */
5870 for (i=1; i<=el[0].mid; i++) {
5873 munmap(el[i].mptr, env->me_psize * el[i].mcnt);
5878 /* see if we can unref some local pages */
5883 if (el[0].mid >= MDB_ERPAGE_MAX) {
5884 pthread_mutex_unlock(&env->me_rpmutex);
5885 return MDB_MAP_FULL;
5887 env->me_rpcheck /= 2;
5889 for (i=y+1; i<= el[0].mid; i++)
5893 if (!env->me_rpcheck)
5894 env->me_rpcheck = 1;
5895 while (env->me_rpcheck < el[0].mid && env->me_rpcheck < MDB_ERPAGE_SIZE/2)
5896 env->me_rpcheck *= 2;
5899 SET_OFF(off, pgno * env->me_psize);
5900 MAP(rc, env, id3.mptr, len, off);
5903 pthread_mutex_unlock(&env->me_rpmutex);
5906 /* check for overflow size */
5907 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5908 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
5909 id3.mcnt = p->mp_pages + rem;
5910 munmap(id3.mptr, len);
5911 len = id3.mcnt * env->me_psize;
5912 MAP(rc, env, id3.mptr, len, off);
5916 mdb_mid3l_insert(el, &id3);
5917 pthread_mutex_unlock(&env->me_rpmutex);
5919 mdb_mid3l_insert(tl, &id3);
5921 return MDB_TXN_FULL;
5924 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5925 #if MDB_DEBUG /* we don't need this check any more */
5926 if (IS_OVERFLOW(p)) {
5927 mdb_tassert(txn, p->mp_pages + rem <= id3.mcnt);
5935 /** Find the address of the page corresponding to a given page number.
5936 * @param[in] mc the cursor accessing the page.
5937 * @param[in] pgno the page number for the page to retrieve.
5938 * @param[out] ret address of a pointer where the page's address will be stored.
5939 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5940 * @return 0 on success, non-zero on failure.
5943 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
5945 MDB_txn *txn = mc->mc_txn;
5947 MDB_env *env = txn->mt_env;
5952 if (! (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP))) {
5956 MDB_ID2L dl = tx2->mt_u.dirty_list;
5958 /* Spilled pages were dirtied in this txn and flushed
5959 * because the dirty list got full. Bring this page
5960 * back in from the map (but don't unspill it here,
5961 * leave that unless page_touch happens again).
5963 if (tx2->mt_spill_pgs) {
5964 MDB_ID pn = pgno << 1;
5965 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5966 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5968 int rc = mdb_rpage_get(txn, pgno, &p);
5972 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5978 unsigned x = mdb_mid2l_search(dl, pgno);
5979 if (x <= dl[0].mid && dl[x].mid == pgno) {
5985 } while ((tx2 = tx2->mt_parent) != NULL);
5988 if (pgno < txn->mt_next_pgno) {
5992 int rc = mdb_rpage_get(txn, pgno, &p);
5997 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
6000 DPRINTF(("page %"Y"u not found", pgno));
6001 txn->mt_flags |= MDB_TXN_ERROR;
6002 return MDB_PAGE_NOTFOUND;
6012 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
6013 * The cursor is at the root page, set up the rest of it.
6016 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
6018 MDB_page *mp = mc->mc_pg[mc->mc_top];
6022 while (IS_BRANCH(mp)) {
6026 DPRINTF(("branch page %"Y"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
6027 /* Don't assert on branch pages in the FreeDB. We can get here
6028 * while in the process of rebalancing a FreeDB branch page; we must
6029 * let that proceed. ITS#8336
6031 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
6032 DPRINTF(("found index 0 to page %"Y"u", NODEPGNO(NODEPTR(mp, 0))));
6034 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
6036 if (flags & MDB_PS_LAST)
6037 i = NUMKEYS(mp) - 1;
6040 node = mdb_node_search(mc, key, &exact);
6042 i = NUMKEYS(mp) - 1;
6044 i = mc->mc_ki[mc->mc_top];
6046 mdb_cassert(mc, i > 0);
6050 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
6053 mdb_cassert(mc, i < NUMKEYS(mp));
6054 node = NODEPTR(mp, i);
6056 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6059 mc->mc_ki[mc->mc_top] = i;
6060 if ((rc = mdb_cursor_push(mc, mp)))
6063 if (flags & MDB_PS_MODIFY) {
6064 if ((rc = mdb_page_touch(mc)) != 0)
6066 mp = mc->mc_pg[mc->mc_top];
6071 DPRINTF(("internal error, index points to a %02X page!?",
6073 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6074 return MDB_CORRUPTED;
6077 DPRINTF(("found leaf page %"Y"u for key [%s]", mp->mp_pgno,
6078 key ? DKEY(key) : "null"));
6079 mc->mc_flags |= C_INITIALIZED;
6080 mc->mc_flags &= ~C_EOF;
6085 /** Search for the lowest key under the current branch page.
6086 * This just bypasses a NUMKEYS check in the current page
6087 * before calling mdb_page_search_root(), because the callers
6088 * are all in situations where the current page is known to
6092 mdb_page_search_lowest(MDB_cursor *mc)
6094 MDB_page *mp = mc->mc_pg[mc->mc_top];
6095 MDB_node *node = NODEPTR(mp, 0);
6098 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6101 mc->mc_ki[mc->mc_top] = 0;
6102 if ((rc = mdb_cursor_push(mc, mp)))
6104 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
6107 /** Search for the page a given key should be in.
6108 * Push it and its parent pages on the cursor stack.
6109 * @param[in,out] mc the cursor for this operation.
6110 * @param[in] key the key to search for, or NULL for first/last page.
6111 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
6112 * are touched (updated with new page numbers).
6113 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
6114 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
6115 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
6116 * @return 0 on success, non-zero on failure.
6119 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
6124 /* Make sure the txn is still viable, then find the root from
6125 * the txn's db table and set it as the root of the cursor's stack.
6127 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
6128 DPUTS("transaction may not be used now");
6131 /* Make sure we're using an up-to-date root */
6132 if (*mc->mc_dbflag & DB_STALE) {
6134 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6136 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
6137 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
6144 MDB_node *leaf = mdb_node_search(&mc2,
6145 &mc->mc_dbx->md_name, &exact);
6147 return MDB_NOTFOUND;
6148 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
6149 return MDB_INCOMPATIBLE; /* not a named DB */
6150 rc = mdb_node_read(&mc2, leaf, &data);
6153 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
6155 /* The txn may not know this DBI, or another process may
6156 * have dropped and recreated the DB with other flags.
6158 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
6159 return MDB_INCOMPATIBLE;
6160 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
6162 *mc->mc_dbflag &= ~DB_STALE;
6164 root = mc->mc_db->md_root;
6166 if (root == P_INVALID) { /* Tree is empty. */
6167 DPUTS("tree is empty");
6168 return MDB_NOTFOUND;
6172 mdb_cassert(mc, root > 1);
6173 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) {
6176 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[0]);
6178 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
6185 for (i=1; i<mc->mc_snum; i++)
6186 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[i]);
6192 DPRINTF(("db %d root page %"Y"u has flags 0x%X",
6193 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
6195 if (flags & MDB_PS_MODIFY) {
6196 if ((rc = mdb_page_touch(mc)))
6200 if (flags & MDB_PS_ROOTONLY)
6203 return mdb_page_search_root(mc, key, flags);
6207 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
6209 MDB_txn *txn = mc->mc_txn;
6210 pgno_t pg = mp->mp_pgno;
6211 unsigned x = 0, ovpages = mp->mp_pages;
6212 MDB_env *env = txn->mt_env;
6213 MDB_IDL sl = txn->mt_spill_pgs;
6214 MDB_ID pn = pg << 1;
6217 DPRINTF(("free ov page %"Y"u (%d)", pg, ovpages));
6218 /* If the page is dirty or on the spill list we just acquired it,
6219 * so we should give it back to our current free list, if any.
6220 * Otherwise put it onto the list of pages we freed in this txn.
6222 * Won't create me_pghead: me_pglast must be inited along with it.
6223 * Unsupported in nested txns: They would need to hide the page
6224 * range in ancestor txns' dirty and spilled lists.
6226 if (env->me_pghead &&
6228 ((mp->mp_flags & P_DIRTY) ||
6229 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
6233 MDB_ID2 *dl, ix, iy;
6234 rc = mdb_midl_need(&env->me_pghead, ovpages);
6237 if (!(mp->mp_flags & P_DIRTY)) {
6238 /* This page is no longer spilled */
6245 /* Remove from dirty list */
6246 dl = txn->mt_u.dirty_list;
6248 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
6254 mdb_cassert(mc, x > 1);
6256 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
6257 txn->mt_flags |= MDB_TXN_ERROR;
6261 txn->mt_dirty_room++;
6262 if (!(env->me_flags & MDB_WRITEMAP))
6263 mdb_dpage_free(env, mp);
6265 /* Insert in me_pghead */
6266 mop = env->me_pghead;
6267 j = mop[0] + ovpages;
6268 for (i = mop[0]; i && mop[i] < pg; i--)
6274 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
6278 mc->mc_db->md_overflow_pages -= ovpages;
6282 /** Return the data associated with a given node.
6283 * @param[in] mc The cursor for this operation.
6284 * @param[in] leaf The node being read.
6285 * @param[out] data Updated to point to the node's data.
6286 * @return 0 on success, non-zero on failure.
6289 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
6291 MDB_page *omp; /* overflow page */
6297 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_ovpg);
6301 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6302 data->mv_size = NODEDSZ(leaf);
6303 data->mv_data = NODEDATA(leaf);
6307 /* Read overflow data.
6309 data->mv_size = NODEDSZ(leaf);
6310 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
6311 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
6312 DPRINTF(("read overflow page %"Y"u failed", pgno));
6315 data->mv_data = METADATA(omp);
6324 mdb_get(MDB_txn *txn, MDB_dbi dbi,
6325 MDB_val *key, MDB_val *data)
6332 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
6334 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
6337 if (txn->mt_flags & MDB_TXN_BLOCKED)
6340 mdb_cursor_init(&mc, txn, dbi, &mx);
6341 rc = mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
6344 /* unref all the pages - caller must copy the data
6345 * before doing anything else
6347 mdb_cursor_unref(&mc);
6353 /** Find a sibling for a page.
6354 * Replaces the page at the top of the cursor's stack with the
6355 * specified sibling, if one exists.
6356 * @param[in] mc The cursor for this operation.
6357 * @param[in] move_right Non-zero if the right sibling is requested,
6358 * otherwise the left sibling.
6359 * @return 0 on success, non-zero on failure.
6362 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
6371 if (mc->mc_snum < 2) {
6372 return MDB_NOTFOUND; /* root has no siblings */
6376 op = mc->mc_pg[mc->mc_top];
6379 DPRINTF(("parent page is page %"Y"u, index %u",
6380 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
6382 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6383 : (mc->mc_ki[mc->mc_top] == 0)) {
6384 DPRINTF(("no more keys left, moving to %s sibling",
6385 move_right ? "right" : "left"));
6386 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
6387 /* undo cursor_pop before returning */
6394 mc->mc_ki[mc->mc_top]++;
6396 mc->mc_ki[mc->mc_top]--;
6397 DPRINTF(("just moving to %s index key %u",
6398 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
6400 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
6402 MDB_PAGE_UNREF(mc->mc_txn, op);
6404 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6405 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
6406 /* mc will be inconsistent if caller does mc_snum++ as above */
6407 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
6411 mdb_cursor_push(mc, mp);
6413 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
6418 /** Move the cursor to the next data item. */
6420 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6426 if ((mc->mc_flags & C_EOF) ||
6427 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
6428 return MDB_NOTFOUND;
6430 if (!(mc->mc_flags & C_INITIALIZED))
6431 return mdb_cursor_first(mc, key, data);
6433 mp = mc->mc_pg[mc->mc_top];
6435 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6436 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6437 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6438 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
6439 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
6440 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
6441 if (rc == MDB_SUCCESS)
6442 MDB_GET_KEY(leaf, key);
6448 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6449 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
6454 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6455 if (op == MDB_NEXT_DUP)
6456 return MDB_NOTFOUND;
6460 DPRINTF(("cursor_next: top page is %"Y"u in cursor %p",
6461 mdb_dbg_pgno(mp), (void *) mc));
6462 if (mc->mc_flags & C_DEL) {
6463 mc->mc_flags ^= C_DEL;
6467 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
6468 DPUTS("=====> move to next sibling page");
6469 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6470 mc->mc_flags |= C_EOF;
6473 mp = mc->mc_pg[mc->mc_top];
6474 DPRINTF(("next page is %"Y"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6476 mc->mc_ki[mc->mc_top]++;
6479 DPRINTF(("==> cursor points to page %"Y"u with %u keys, key index %u",
6480 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6483 key->mv_size = mc->mc_db->md_pad;
6484 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6488 mdb_cassert(mc, IS_LEAF(mp));
6489 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6491 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6492 mdb_xcursor_init1(mc, leaf);
6495 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6498 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6499 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6500 if (rc != MDB_SUCCESS)
6505 MDB_GET_KEY(leaf, key);
6509 /** Move the cursor to the previous data item. */
6511 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6517 if (!(mc->mc_flags & C_INITIALIZED)) {
6518 rc = mdb_cursor_last(mc, key, data);
6521 mc->mc_ki[mc->mc_top]++;
6524 mp = mc->mc_pg[mc->mc_top];
6526 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6527 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6528 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6529 if (op == MDB_PREV || op == MDB_PREV_DUP) {
6530 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
6531 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
6532 if (rc == MDB_SUCCESS) {
6533 MDB_GET_KEY(leaf, key);
6534 mc->mc_flags &= ~C_EOF;
6541 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6542 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
6547 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6548 if (op == MDB_PREV_DUP)
6549 return MDB_NOTFOUND;
6553 DPRINTF(("cursor_prev: top page is %"Y"u in cursor %p",
6554 mdb_dbg_pgno(mp), (void *) mc));
6556 mc->mc_flags &= ~(C_EOF|C_DEL);
6558 if (mc->mc_ki[mc->mc_top] == 0) {
6559 DPUTS("=====> move to prev sibling page");
6560 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
6563 mp = mc->mc_pg[mc->mc_top];
6564 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
6565 DPRINTF(("prev page is %"Y"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6567 mc->mc_ki[mc->mc_top]--;
6569 mc->mc_flags &= ~C_EOF;
6571 DPRINTF(("==> cursor points to page %"Y"u with %u keys, key index %u",
6572 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6575 key->mv_size = mc->mc_db->md_pad;
6576 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6580 mdb_cassert(mc, IS_LEAF(mp));
6581 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6583 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6584 mdb_xcursor_init1(mc, leaf);
6587 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6590 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6591 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6592 if (rc != MDB_SUCCESS)
6597 MDB_GET_KEY(leaf, key);
6601 /** Set the cursor on a specific data item. */
6603 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6604 MDB_cursor_op op, int *exactp)
6608 MDB_node *leaf = NULL;
6611 if (key->mv_size == 0)
6612 return MDB_BAD_VALSIZE;
6615 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6617 /* See if we're already on the right page */
6618 if (mc->mc_flags & C_INITIALIZED) {
6621 mp = mc->mc_pg[mc->mc_top];
6623 mc->mc_ki[mc->mc_top] = 0;
6624 return MDB_NOTFOUND;
6626 if (mp->mp_flags & P_LEAF2) {
6627 nodekey.mv_size = mc->mc_db->md_pad;
6628 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
6630 leaf = NODEPTR(mp, 0);
6631 MDB_GET_KEY2(leaf, nodekey);
6633 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6635 /* Probably happens rarely, but first node on the page
6636 * was the one we wanted.
6638 mc->mc_ki[mc->mc_top] = 0;
6645 unsigned int nkeys = NUMKEYS(mp);
6647 if (mp->mp_flags & P_LEAF2) {
6648 nodekey.mv_data = LEAF2KEY(mp,
6649 nkeys-1, nodekey.mv_size);
6651 leaf = NODEPTR(mp, nkeys-1);
6652 MDB_GET_KEY2(leaf, nodekey);
6654 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6656 /* last node was the one we wanted */
6657 mc->mc_ki[mc->mc_top] = nkeys-1;
6663 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6664 /* This is definitely the right page, skip search_page */
6665 if (mp->mp_flags & P_LEAF2) {
6666 nodekey.mv_data = LEAF2KEY(mp,
6667 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6669 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6670 MDB_GET_KEY2(leaf, nodekey);
6672 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6674 /* current node was the one we wanted */
6684 /* If any parents have right-sibs, search.
6685 * Otherwise, there's nothing further.
6687 for (i=0; i<mc->mc_top; i++)
6689 NUMKEYS(mc->mc_pg[i])-1)
6691 if (i == mc->mc_top) {
6692 /* There are no other pages */
6693 mc->mc_ki[mc->mc_top] = nkeys;
6694 return MDB_NOTFOUND;
6698 /* There are no other pages */
6699 mc->mc_ki[mc->mc_top] = 0;
6700 if (op == MDB_SET_RANGE && !exactp) {
6704 return MDB_NOTFOUND;
6710 rc = mdb_page_search(mc, key, 0);
6711 if (rc != MDB_SUCCESS)
6714 mp = mc->mc_pg[mc->mc_top];
6715 mdb_cassert(mc, IS_LEAF(mp));
6718 leaf = mdb_node_search(mc, key, exactp);
6719 if (exactp != NULL && !*exactp) {
6720 /* MDB_SET specified and not an exact match. */
6721 return MDB_NOTFOUND;
6725 DPUTS("===> inexact leaf not found, goto sibling");
6726 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6727 mc->mc_flags |= C_EOF;
6728 return rc; /* no entries matched */
6730 mp = mc->mc_pg[mc->mc_top];
6731 mdb_cassert(mc, IS_LEAF(mp));
6732 leaf = NODEPTR(mp, 0);
6736 mc->mc_flags |= C_INITIALIZED;
6737 mc->mc_flags &= ~C_EOF;
6740 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6741 key->mv_size = mc->mc_db->md_pad;
6742 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6748 if (mc->mc_xcursor && mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6749 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
6752 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6753 mdb_xcursor_init1(mc, leaf);
6756 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6757 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6758 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6761 if (op == MDB_GET_BOTH) {
6767 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6768 if (rc != MDB_SUCCESS)
6771 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6774 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6776 dcmp = mc->mc_dbx->md_dcmp;
6777 #if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
6778 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(mdb_size_t))
6779 dcmp = mdb_cmp_clong;
6781 rc = dcmp(data, &olddata);
6783 if (op == MDB_GET_BOTH || rc > 0)
6784 return MDB_NOTFOUND;
6791 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6792 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6797 /* The key already matches in all other cases */
6798 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6799 MDB_GET_KEY(leaf, key);
6800 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6805 /** Move the cursor to the first item in the database. */
6807 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6812 if (mc->mc_xcursor) {
6814 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6815 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
6818 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6821 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6822 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6823 if (rc != MDB_SUCCESS)
6826 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6828 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6829 mc->mc_flags |= C_INITIALIZED;
6830 mc->mc_flags &= ~C_EOF;
6832 mc->mc_ki[mc->mc_top] = 0;
6834 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6835 key->mv_size = mc->mc_db->md_pad;
6836 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6841 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6842 mdb_xcursor_init1(mc, leaf);
6843 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6847 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6851 MDB_GET_KEY(leaf, key);
6855 /** Move the cursor to the last item in the database. */
6857 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6862 if (mc->mc_xcursor) {
6864 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6865 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
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 UINT_MAX < SIZE_MAX || defined(MDB_VL32)
7324 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(mdb_size_t))
7325 dcmp = mdb_cmp_clong;
7327 /* does data match? */
7328 if (!dcmp(data, &olddata)) {
7329 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
7330 return MDB_KEYEXIST;
7335 /* Back up original data item */
7336 dkey.mv_size = olddata.mv_size;
7337 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
7339 /* Make sub-page header for the dup items, with dummy body */
7340 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
7341 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
7342 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
7343 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7344 fp->mp_flags |= P_LEAF2;
7345 fp->mp_pad = data->mv_size;
7346 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
7348 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
7349 (dkey.mv_size & 1) + (data->mv_size & 1);
7351 fp->mp_upper = xdata.mv_size - PAGEBASE;
7352 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
7353 } else if (leaf->mn_flags & F_SUBDATA) {
7354 /* Data is on sub-DB, just store it */
7355 flags |= F_DUPDATA|F_SUBDATA;
7358 /* Data is on sub-page */
7359 fp = olddata.mv_data;
7362 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7363 offset = EVEN(NODESIZE + sizeof(indx_t) +
7367 offset = fp->mp_pad;
7368 if (SIZELEFT(fp) < offset) {
7369 offset *= 4; /* space for 4 more */
7372 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
7374 fp->mp_flags |= P_DIRTY;
7375 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
7376 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
7380 xdata.mv_size = olddata.mv_size + offset;
7383 fp_flags = fp->mp_flags;
7384 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
7385 /* Too big for a sub-page, convert to sub-DB */
7386 fp_flags &= ~P_SUBP;
7388 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7389 fp_flags |= P_LEAF2;
7390 dummy.md_pad = fp->mp_pad;
7391 dummy.md_flags = MDB_DUPFIXED;
7392 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7393 dummy.md_flags |= MDB_INTEGERKEY;
7399 dummy.md_branch_pages = 0;
7400 dummy.md_leaf_pages = 1;
7401 dummy.md_overflow_pages = 0;
7402 dummy.md_entries = NUMKEYS(fp);
7403 xdata.mv_size = sizeof(MDB_db);
7404 xdata.mv_data = &dummy;
7405 if ((rc = mdb_page_alloc(mc, 1, &mp)))
7407 offset = env->me_psize - olddata.mv_size;
7408 flags |= F_DUPDATA|F_SUBDATA;
7409 dummy.md_root = mp->mp_pgno;
7413 mp->mp_flags = fp_flags | P_DIRTY;
7414 mp->mp_pad = fp->mp_pad;
7415 mp->mp_lower = fp->mp_lower;
7416 mp->mp_upper = fp->mp_upper + offset;
7417 if (fp_flags & P_LEAF2) {
7418 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
7420 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
7421 olddata.mv_size - fp->mp_upper - PAGEBASE);
7422 for (i=0; i<NUMKEYS(fp); i++)
7423 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
7431 mdb_node_del(mc, 0);
7435 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
7436 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
7437 return MDB_INCOMPATIBLE;
7438 /* overflow page overwrites need special handling */
7439 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7442 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
7444 memcpy(&pg, olddata.mv_data, sizeof(pg));
7445 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
7447 ovpages = omp->mp_pages;
7449 /* Is the ov page large enough? */
7450 if (ovpages >= dpages) {
7451 if (!(omp->mp_flags & P_DIRTY) &&
7452 (level || (env->me_flags & MDB_WRITEMAP)))
7454 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
7457 level = 0; /* dirty in this txn or clean */
7460 if (omp->mp_flags & P_DIRTY) {
7461 /* yes, overwrite it. Note in this case we don't
7462 * bother to try shrinking the page if the new data
7463 * is smaller than the overflow threshold.
7466 /* It is writable only in a parent txn */
7467 size_t sz = (size_t) env->me_psize * ovpages, off;
7468 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
7474 /* Note - this page is already counted in parent's dirty_room */
7475 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
7476 mdb_cassert(mc, rc2 == 0);
7477 /* Currently we make the page look as with put() in the
7478 * parent txn, in case the user peeks at MDB_RESERVEd
7479 * or unused parts. Some users treat ovpages specially.
7481 if (!(flags & MDB_RESERVE)) {
7482 /* Skip the part where LMDB will put *data.
7483 * Copy end of page, adjusting alignment so
7484 * compiler may copy words instead of bytes.
7486 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
7487 memcpy((size_t *)((char *)np + off),
7488 (size_t *)((char *)omp + off), sz - off);
7491 memcpy(np, omp, sz); /* Copy beginning of page */
7494 SETDSZ(leaf, data->mv_size);
7495 if (F_ISSET(flags, MDB_RESERVE))
7496 data->mv_data = METADATA(omp);
7498 memcpy(METADATA(omp), data->mv_data, data->mv_size);
7502 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
7504 } else if (data->mv_size == olddata.mv_size) {
7505 /* same size, just replace it. Note that we could
7506 * also reuse this node if the new data is smaller,
7507 * but instead we opt to shrink the node in that case.
7509 if (F_ISSET(flags, MDB_RESERVE))
7510 data->mv_data = olddata.mv_data;
7511 else if (!(mc->mc_flags & C_SUB))
7512 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
7514 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
7519 mdb_node_del(mc, 0);
7525 nflags = flags & NODE_ADD_FLAGS;
7526 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
7527 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
7528 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
7529 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
7531 nflags |= MDB_SPLIT_REPLACE;
7532 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
7534 /* There is room already in this leaf page. */
7535 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
7537 /* Adjust other cursors pointing to mp */
7538 MDB_cursor *m2, *m3;
7539 MDB_dbi dbi = mc->mc_dbi;
7540 unsigned i = mc->mc_top;
7541 MDB_page *mp = mc->mc_pg[i];
7543 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7544 if (mc->mc_flags & C_SUB)
7545 m3 = &m2->mc_xcursor->mx_cursor;
7548 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
7549 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
7552 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7553 MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
7554 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7555 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7561 if (rc == MDB_SUCCESS) {
7562 /* Now store the actual data in the child DB. Note that we're
7563 * storing the user data in the keys field, so there are strict
7564 * size limits on dupdata. The actual data fields of the child
7565 * DB are all zero size.
7568 int xflags, new_dupdata;
7573 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7574 if (flags & MDB_CURRENT) {
7575 xflags = MDB_CURRENT|MDB_NOSPILL;
7577 mdb_xcursor_init1(mc, leaf);
7578 xflags = (flags & MDB_NODUPDATA) ?
7579 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
7582 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
7583 new_dupdata = (int)dkey.mv_size;
7584 /* converted, write the original data first */
7586 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
7589 /* we've done our job */
7592 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
7593 /* Adjust other cursors pointing to mp */
7595 MDB_xcursor *mx = mc->mc_xcursor;
7596 unsigned i = mc->mc_top;
7597 MDB_page *mp = mc->mc_pg[i];
7598 int nkeys = NUMKEYS(mp);
7600 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7601 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7602 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7603 if (m2->mc_pg[i] == mp) {
7604 if (m2->mc_ki[i] == mc->mc_ki[i]) {
7605 mdb_xcursor_init2(m2, mx, new_dupdata);
7606 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
7607 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
7608 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7609 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7614 ecount = mc->mc_xcursor->mx_db.md_entries;
7615 if (flags & MDB_APPENDDUP)
7616 xflags |= MDB_APPEND;
7617 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
7618 if (flags & F_SUBDATA) {
7619 void *db = NODEDATA(leaf);
7620 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7622 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
7624 /* Increment count unless we just replaced an existing item. */
7626 mc->mc_db->md_entries++;
7628 /* Invalidate txn if we created an empty sub-DB */
7631 /* If we succeeded and the key didn't exist before,
7632 * make sure the cursor is marked valid.
7634 mc->mc_flags |= C_INITIALIZED;
7636 if (flags & MDB_MULTIPLE) {
7639 /* let caller know how many succeeded, if any */
7640 data[1].mv_size = mcount;
7641 if (mcount < dcount) {
7642 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
7643 insert_key = insert_data = 0;
7650 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
7653 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7658 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7664 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7665 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7667 if (!(mc->mc_flags & C_INITIALIZED))
7670 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7671 return MDB_NOTFOUND;
7673 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7676 rc = mdb_cursor_touch(mc);
7680 mp = mc->mc_pg[mc->mc_top];
7683 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7685 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7686 if (flags & MDB_NODUPDATA) {
7687 /* mdb_cursor_del0() will subtract the final entry */
7688 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7689 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7691 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7692 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7694 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7697 /* If sub-DB still has entries, we're done */
7698 if (mc->mc_xcursor->mx_db.md_entries) {
7699 if (leaf->mn_flags & F_SUBDATA) {
7700 /* update subDB info */
7701 void *db = NODEDATA(leaf);
7702 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7705 /* shrink fake page */
7706 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7707 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7708 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7709 /* fix other sub-DB cursors pointed at fake pages on this page */
7710 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7711 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7712 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7713 if (m2->mc_pg[mc->mc_top] == mp) {
7714 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
7715 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7717 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7718 if (!(n2->mn_flags & F_SUBDATA))
7719 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7724 mc->mc_db->md_entries--;
7727 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7729 /* otherwise fall thru and delete the sub-DB */
7732 if (leaf->mn_flags & F_SUBDATA) {
7733 /* add all the child DB's pages to the free list */
7734 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7739 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7740 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7741 rc = MDB_INCOMPATIBLE;
7745 /* add overflow pages to free list */
7746 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7750 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7751 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7752 (rc = mdb_ovpage_free(mc, omp)))
7757 return mdb_cursor_del0(mc);
7760 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7764 /** Allocate and initialize new pages for a database.
7765 * @param[in] mc a cursor on the database being added to.
7766 * @param[in] flags flags defining what type of page is being allocated.
7767 * @param[in] num the number of pages to allocate. This is usually 1,
7768 * unless allocating overflow pages for a large record.
7769 * @param[out] mp Address of a page, or NULL on failure.
7770 * @return 0 on success, non-zero on failure.
7773 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7778 if ((rc = mdb_page_alloc(mc, num, &np)))
7780 DPRINTF(("allocated new mpage %"Y"u, page size %u",
7781 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7782 np->mp_flags = flags | P_DIRTY;
7783 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7784 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7787 mc->mc_db->md_branch_pages++;
7788 else if (IS_LEAF(np))
7789 mc->mc_db->md_leaf_pages++;
7790 else if (IS_OVERFLOW(np)) {
7791 mc->mc_db->md_overflow_pages += num;
7799 /** Calculate the size of a leaf node.
7800 * The size depends on the environment's page size; if a data item
7801 * is too large it will be put onto an overflow page and the node
7802 * size will only include the key and not the data. Sizes are always
7803 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7804 * of the #MDB_node headers.
7805 * @param[in] env The environment handle.
7806 * @param[in] key The key for the node.
7807 * @param[in] data The data for the node.
7808 * @return The number of bytes needed to store the node.
7811 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7815 sz = LEAFSIZE(key, data);
7816 if (sz > env->me_nodemax) {
7817 /* put on overflow page */
7818 sz -= data->mv_size - sizeof(pgno_t);
7821 return EVEN(sz + sizeof(indx_t));
7824 /** Calculate the size of a branch node.
7825 * The size should depend on the environment's page size but since
7826 * we currently don't support spilling large keys onto overflow
7827 * pages, it's simply the size of the #MDB_node header plus the
7828 * size of the key. Sizes are always rounded up to an even number
7829 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7830 * @param[in] env The environment handle.
7831 * @param[in] key The key for the node.
7832 * @return The number of bytes needed to store the node.
7835 mdb_branch_size(MDB_env *env, MDB_val *key)
7840 if (sz > env->me_nodemax) {
7841 /* put on overflow page */
7842 /* not implemented */
7843 /* sz -= key->size - sizeof(pgno_t); */
7846 return sz + sizeof(indx_t);
7849 /** Add a node to the page pointed to by the cursor.
7850 * @param[in] mc The cursor for this operation.
7851 * @param[in] indx The index on the page where the new node should be added.
7852 * @param[in] key The key for the new node.
7853 * @param[in] data The data for the new node, if any.
7854 * @param[in] pgno The page number, if adding a branch node.
7855 * @param[in] flags Flags for the node.
7856 * @return 0 on success, non-zero on failure. Possible errors are:
7858 * <li>ENOMEM - failed to allocate overflow pages for the node.
7859 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7860 * should never happen since all callers already calculate the
7861 * page's free space before calling this function.
7865 mdb_node_add(MDB_cursor *mc, indx_t indx,
7866 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7869 size_t node_size = NODESIZE;
7873 MDB_page *mp = mc->mc_pg[mc->mc_top];
7874 MDB_page *ofp = NULL; /* overflow page */
7878 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7880 DPRINTF(("add to %s %spage %"Y"u index %i, data size %"Z"u key size %"Z"u [%s]",
7881 IS_LEAF(mp) ? "leaf" : "branch",
7882 IS_SUBP(mp) ? "sub-" : "",
7883 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7884 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7887 /* Move higher keys up one slot. */
7888 int ksize = mc->mc_db->md_pad, dif;
7889 char *ptr = LEAF2KEY(mp, indx, ksize);
7890 dif = NUMKEYS(mp) - indx;
7892 memmove(ptr+ksize, ptr, dif*ksize);
7893 /* insert new key */
7894 memcpy(ptr, key->mv_data, ksize);
7896 /* Just using these for counting */
7897 mp->mp_lower += sizeof(indx_t);
7898 mp->mp_upper -= ksize - sizeof(indx_t);
7902 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7904 node_size += key->mv_size;
7906 mdb_cassert(mc, key && data);
7907 if (F_ISSET(flags, F_BIGDATA)) {
7908 /* Data already on overflow page. */
7909 node_size += sizeof(pgno_t);
7910 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7911 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7913 /* Put data on overflow page. */
7914 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7915 data->mv_size, node_size+data->mv_size));
7916 node_size = EVEN(node_size + sizeof(pgno_t));
7917 if ((ssize_t)node_size > room)
7919 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7921 DPRINTF(("allocated overflow page %"Y"u", ofp->mp_pgno));
7925 node_size += data->mv_size;
7928 node_size = EVEN(node_size);
7929 if ((ssize_t)node_size > room)
7933 /* Move higher pointers up one slot. */
7934 for (i = NUMKEYS(mp); i > indx; i--)
7935 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7937 /* Adjust free space offsets. */
7938 ofs = mp->mp_upper - node_size;
7939 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7940 mp->mp_ptrs[indx] = ofs;
7942 mp->mp_lower += sizeof(indx_t);
7944 /* Write the node data. */
7945 node = NODEPTR(mp, indx);
7946 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7947 node->mn_flags = flags;
7949 SETDSZ(node,data->mv_size);
7954 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7957 ndata = NODEDATA(node);
7959 if (F_ISSET(flags, F_BIGDATA))
7960 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7961 else if (F_ISSET(flags, MDB_RESERVE))
7962 data->mv_data = ndata;
7964 memcpy(ndata, data->mv_data, data->mv_size);
7966 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7967 ndata = METADATA(ofp);
7968 if (F_ISSET(flags, MDB_RESERVE))
7969 data->mv_data = ndata;
7971 memcpy(ndata, data->mv_data, data->mv_size);
7978 DPRINTF(("not enough room in page %"Y"u, got %u ptrs",
7979 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7980 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7981 DPRINTF(("node size = %"Z"u", node_size));
7982 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7983 return MDB_PAGE_FULL;
7986 /** Delete the specified node from a page.
7987 * @param[in] mc Cursor pointing to the node to delete.
7988 * @param[in] ksize The size of a node. Only used if the page is
7989 * part of a #MDB_DUPFIXED database.
7992 mdb_node_del(MDB_cursor *mc, int ksize)
7994 MDB_page *mp = mc->mc_pg[mc->mc_top];
7995 indx_t indx = mc->mc_ki[mc->mc_top];
7997 indx_t i, j, numkeys, ptr;
8001 DPRINTF(("delete node %u on %s page %"Y"u", indx,
8002 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
8003 numkeys = NUMKEYS(mp);
8004 mdb_cassert(mc, indx < numkeys);
8007 int x = numkeys - 1 - indx;
8008 base = LEAF2KEY(mp, indx, ksize);
8010 memmove(base, base + ksize, x * ksize);
8011 mp->mp_lower -= sizeof(indx_t);
8012 mp->mp_upper += ksize - sizeof(indx_t);
8016 node = NODEPTR(mp, indx);
8017 sz = NODESIZE + node->mn_ksize;
8019 if (F_ISSET(node->mn_flags, F_BIGDATA))
8020 sz += sizeof(pgno_t);
8022 sz += NODEDSZ(node);
8026 ptr = mp->mp_ptrs[indx];
8027 for (i = j = 0; i < numkeys; i++) {
8029 mp->mp_ptrs[j] = mp->mp_ptrs[i];
8030 if (mp->mp_ptrs[i] < ptr)
8031 mp->mp_ptrs[j] += sz;
8036 base = (char *)mp + mp->mp_upper + PAGEBASE;
8037 memmove(base + sz, base, ptr - mp->mp_upper);
8039 mp->mp_lower -= sizeof(indx_t);
8043 /** Compact the main page after deleting a node on a subpage.
8044 * @param[in] mp The main page to operate on.
8045 * @param[in] indx The index of the subpage on the main page.
8048 mdb_node_shrink(MDB_page *mp, indx_t indx)
8053 indx_t delta, nsize, len, ptr;
8056 node = NODEPTR(mp, indx);
8057 sp = (MDB_page *)NODEDATA(node);
8058 delta = SIZELEFT(sp);
8059 nsize = NODEDSZ(node) - delta;
8061 /* Prepare to shift upward, set len = length(subpage part to shift) */
8065 return; /* do not make the node uneven-sized */
8067 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
8068 for (i = NUMKEYS(sp); --i >= 0; )
8069 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
8072 sp->mp_upper = sp->mp_lower;
8073 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
8074 SETDSZ(node, nsize);
8076 /* Shift <lower nodes...initial part of subpage> upward */
8077 base = (char *)mp + mp->mp_upper + PAGEBASE;
8078 memmove(base + delta, base, (char *)sp + len - base);
8080 ptr = mp->mp_ptrs[indx];
8081 for (i = NUMKEYS(mp); --i >= 0; ) {
8082 if (mp->mp_ptrs[i] <= ptr)
8083 mp->mp_ptrs[i] += delta;
8085 mp->mp_upper += delta;
8088 /** Initial setup of a sorted-dups cursor.
8089 * Sorted duplicates are implemented as a sub-database for the given key.
8090 * The duplicate data items are actually keys of the sub-database.
8091 * Operations on the duplicate data items are performed using a sub-cursor
8092 * initialized when the sub-database is first accessed. This function does
8093 * the preliminary setup of the sub-cursor, filling in the fields that
8094 * depend only on the parent DB.
8095 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8098 mdb_xcursor_init0(MDB_cursor *mc)
8100 MDB_xcursor *mx = mc->mc_xcursor;
8102 mx->mx_cursor.mc_xcursor = NULL;
8103 mx->mx_cursor.mc_txn = mc->mc_txn;
8104 mx->mx_cursor.mc_db = &mx->mx_db;
8105 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
8106 mx->mx_cursor.mc_dbi = mc->mc_dbi;
8107 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
8108 mx->mx_cursor.mc_snum = 0;
8109 mx->mx_cursor.mc_top = 0;
8111 mx->mx_cursor.mc_ovpg = 0;
8113 mx->mx_cursor.mc_flags = C_SUB | (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP));
8114 mx->mx_dbx.md_name.mv_size = 0;
8115 mx->mx_dbx.md_name.mv_data = NULL;
8116 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
8117 mx->mx_dbx.md_dcmp = NULL;
8118 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
8121 /** Final setup of a sorted-dups cursor.
8122 * Sets up the fields that depend on the data from the main cursor.
8123 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8124 * @param[in] node The data containing the #MDB_db record for the
8125 * sorted-dup database.
8128 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
8130 MDB_xcursor *mx = mc->mc_xcursor;
8132 mx->mx_cursor.mc_flags &= C_SUB|C_ORIG_RDONLY|C_WRITEMAP;
8133 if (node->mn_flags & F_SUBDATA) {
8134 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
8135 mx->mx_cursor.mc_pg[0] = 0;
8136 mx->mx_cursor.mc_snum = 0;
8137 mx->mx_cursor.mc_top = 0;
8139 MDB_page *fp = NODEDATA(node);
8140 mx->mx_db.md_pad = 0;
8141 mx->mx_db.md_flags = 0;
8142 mx->mx_db.md_depth = 1;
8143 mx->mx_db.md_branch_pages = 0;
8144 mx->mx_db.md_leaf_pages = 1;
8145 mx->mx_db.md_overflow_pages = 0;
8146 mx->mx_db.md_entries = NUMKEYS(fp);
8147 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
8148 mx->mx_cursor.mc_snum = 1;
8149 mx->mx_cursor.mc_top = 0;
8150 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8151 mx->mx_cursor.mc_pg[0] = fp;
8152 mx->mx_cursor.mc_ki[0] = 0;
8153 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
8154 mx->mx_db.md_flags = MDB_DUPFIXED;
8155 mx->mx_db.md_pad = fp->mp_pad;
8156 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
8157 mx->mx_db.md_flags |= MDB_INTEGERKEY;
8160 DPRINTF(("Sub-db -%u root page %"Y"u", mx->mx_cursor.mc_dbi,
8161 mx->mx_db.md_root));
8162 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8163 #if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
8164 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(mdb_size_t))
8165 mx->mx_dbx.md_cmp = mdb_cmp_clong;
8170 /** Fixup a sorted-dups cursor due to underlying update.
8171 * Sets up some fields that depend on the data from the main cursor.
8172 * Almost the same as init1, but skips initialization steps if the
8173 * xcursor had already been used.
8174 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
8175 * @param[in] src_mx The xcursor of an up-to-date cursor.
8176 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
8179 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
8181 MDB_xcursor *mx = mc->mc_xcursor;
8184 mx->mx_cursor.mc_snum = 1;
8185 mx->mx_cursor.mc_top = 0;
8186 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8187 mx->mx_cursor.mc_ki[0] = 0;
8188 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8189 #if UINT_MAX < SIZE_MAX
8190 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
8192 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
8195 mx->mx_db = src_mx->mx_db;
8196 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
8197 DPRINTF(("Sub-db -%u root page %"Y"u", mx->mx_cursor.mc_dbi,
8198 mx->mx_db.md_root));
8201 /** Initialize a cursor for a given transaction and database. */
8203 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
8206 mc->mc_backup = NULL;
8209 mc->mc_db = &txn->mt_dbs[dbi];
8210 mc->mc_dbx = &txn->mt_dbxs[dbi];
8211 mc->mc_dbflag = &txn->mt_dbflags[dbi];
8219 mc->mc_flags = txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
8220 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
8221 mdb_tassert(txn, mx != NULL);
8222 mc->mc_xcursor = mx;
8223 mdb_xcursor_init0(mc);
8225 mc->mc_xcursor = NULL;
8227 if (*mc->mc_dbflag & DB_STALE) {
8228 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
8233 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
8236 size_t size = sizeof(MDB_cursor);
8238 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
8241 if (txn->mt_flags & MDB_TXN_BLOCKED)
8244 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8247 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
8248 size += sizeof(MDB_xcursor);
8250 if ((mc = malloc(size)) != NULL) {
8251 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
8252 if (txn->mt_cursors) {
8253 mc->mc_next = txn->mt_cursors[dbi];
8254 txn->mt_cursors[dbi] = mc;
8255 mc->mc_flags |= C_UNTRACK;
8267 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
8269 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
8272 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
8275 if (txn->mt_flags & MDB_TXN_BLOCKED)
8278 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
8282 /* Return the count of duplicate data items for the current key */
8284 mdb_cursor_count(MDB_cursor *mc, mdb_size_t *countp)
8288 if (mc == NULL || countp == NULL)
8291 if (mc->mc_xcursor == NULL)
8292 return MDB_INCOMPATIBLE;
8294 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
8297 if (!(mc->mc_flags & C_INITIALIZED))
8300 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
8301 return MDB_NOTFOUND;
8303 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8304 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
8307 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
8310 *countp = mc->mc_xcursor->mx_db.md_entries;
8316 mdb_cursor_close(MDB_cursor *mc)
8318 if (mc && !mc->mc_backup) {
8319 /* remove from txn, if tracked */
8320 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
8321 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
8322 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
8324 *prev = mc->mc_next;
8331 mdb_cursor_txn(MDB_cursor *mc)
8333 if (!mc) return NULL;
8338 mdb_cursor_dbi(MDB_cursor *mc)
8343 /** Replace the key for a branch node with a new key.
8344 * @param[in] mc Cursor pointing to the node to operate on.
8345 * @param[in] key The new key to use.
8346 * @return 0 on success, non-zero on failure.
8349 mdb_update_key(MDB_cursor *mc, MDB_val *key)
8355 int delta, ksize, oksize;
8356 indx_t ptr, i, numkeys, indx;
8359 indx = mc->mc_ki[mc->mc_top];
8360 mp = mc->mc_pg[mc->mc_top];
8361 node = NODEPTR(mp, indx);
8362 ptr = mp->mp_ptrs[indx];
8366 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
8367 k2.mv_data = NODEKEY(node);
8368 k2.mv_size = node->mn_ksize;
8369 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Y"u",
8371 mdb_dkey(&k2, kbuf2),
8377 /* Sizes must be 2-byte aligned. */
8378 ksize = EVEN(key->mv_size);
8379 oksize = EVEN(node->mn_ksize);
8380 delta = ksize - oksize;
8382 /* Shift node contents if EVEN(key length) changed. */
8384 if (delta > 0 && SIZELEFT(mp) < delta) {
8386 /* not enough space left, do a delete and split */
8387 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
8388 pgno = NODEPGNO(node);
8389 mdb_node_del(mc, 0);
8390 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
8393 numkeys = NUMKEYS(mp);
8394 for (i = 0; i < numkeys; i++) {
8395 if (mp->mp_ptrs[i] <= ptr)
8396 mp->mp_ptrs[i] -= delta;
8399 base = (char *)mp + mp->mp_upper + PAGEBASE;
8400 len = ptr - mp->mp_upper + NODESIZE;
8401 memmove(base - delta, base, len);
8402 mp->mp_upper -= delta;
8404 node = NODEPTR(mp, indx);
8407 /* But even if no shift was needed, update ksize */
8408 if (node->mn_ksize != key->mv_size)
8409 node->mn_ksize = key->mv_size;
8412 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8418 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
8420 /** Perform \b act while tracking temporary cursor \b mn */
8421 #define WITH_CURSOR_TRACKING(mn, act) do { \
8422 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
8423 if ((mn).mc_flags & C_SUB) { \
8424 dummy.mc_flags = C_INITIALIZED; \
8425 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
8430 tracked->mc_next = *tp; \
8433 *tp = tracked->mc_next; \
8436 /** Move a node from csrc to cdst.
8439 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
8446 unsigned short flags;
8450 /* Mark src and dst as dirty. */
8451 if ((rc = mdb_page_touch(csrc)) ||
8452 (rc = mdb_page_touch(cdst)))
8455 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8456 key.mv_size = csrc->mc_db->md_pad;
8457 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
8459 data.mv_data = NULL;
8463 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
8464 mdb_cassert(csrc, !((size_t)srcnode & 1));
8465 srcpg = NODEPGNO(srcnode);
8466 flags = srcnode->mn_flags;
8467 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8468 unsigned int snum = csrc->mc_snum;
8470 /* must find the lowest key below src */
8471 rc = mdb_page_search_lowest(csrc);
8474 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8475 key.mv_size = csrc->mc_db->md_pad;
8476 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8478 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8479 key.mv_size = NODEKSZ(s2);
8480 key.mv_data = NODEKEY(s2);
8482 csrc->mc_snum = snum--;
8483 csrc->mc_top = snum;
8485 key.mv_size = NODEKSZ(srcnode);
8486 key.mv_data = NODEKEY(srcnode);
8488 data.mv_size = NODEDSZ(srcnode);
8489 data.mv_data = NODEDATA(srcnode);
8491 mn.mc_xcursor = NULL;
8492 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
8493 unsigned int snum = cdst->mc_snum;
8496 /* must find the lowest key below dst */
8497 mdb_cursor_copy(cdst, &mn);
8498 rc = mdb_page_search_lowest(&mn);
8501 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8502 bkey.mv_size = mn.mc_db->md_pad;
8503 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
8505 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8506 bkey.mv_size = NODEKSZ(s2);
8507 bkey.mv_data = NODEKEY(s2);
8509 mn.mc_snum = snum--;
8512 rc = mdb_update_key(&mn, &bkey);
8517 DPRINTF(("moving %s node %u [%s] on page %"Y"u to node %u on page %"Y"u",
8518 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
8519 csrc->mc_ki[csrc->mc_top],
8521 csrc->mc_pg[csrc->mc_top]->mp_pgno,
8522 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
8524 /* Add the node to the destination page.
8526 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
8527 if (rc != MDB_SUCCESS)
8530 /* Delete the node from the source page.
8532 mdb_node_del(csrc, key.mv_size);
8535 /* Adjust other cursors pointing to mp */
8536 MDB_cursor *m2, *m3;
8537 MDB_dbi dbi = csrc->mc_dbi;
8538 MDB_page *mpd, *mps;
8540 mps = csrc->mc_pg[csrc->mc_top];
8541 /* If we're adding on the left, bump others up */
8543 mpd = cdst->mc_pg[csrc->mc_top];
8544 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8545 if (csrc->mc_flags & C_SUB)
8546 m3 = &m2->mc_xcursor->mx_cursor;
8549 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8552 m3->mc_pg[csrc->mc_top] == mpd &&
8553 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
8554 m3->mc_ki[csrc->mc_top]++;
8557 m3->mc_pg[csrc->mc_top] == mps &&
8558 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
8559 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8560 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8561 m3->mc_ki[csrc->mc_top-1]++;
8563 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8565 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8566 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8567 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8571 /* Adding on the right, bump others down */
8573 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8574 if (csrc->mc_flags & C_SUB)
8575 m3 = &m2->mc_xcursor->mx_cursor;
8578 if (m3 == csrc) continue;
8579 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8581 if (m3->mc_pg[csrc->mc_top] == mps) {
8582 if (!m3->mc_ki[csrc->mc_top]) {
8583 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8584 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8585 m3->mc_ki[csrc->mc_top-1]--;
8587 m3->mc_ki[csrc->mc_top]--;
8589 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8591 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8592 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8593 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8600 /* Update the parent separators.
8602 if (csrc->mc_ki[csrc->mc_top] == 0) {
8603 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
8604 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8605 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8607 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8608 key.mv_size = NODEKSZ(srcnode);
8609 key.mv_data = NODEKEY(srcnode);
8611 DPRINTF(("update separator for source page %"Y"u to [%s]",
8612 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
8613 mdb_cursor_copy(csrc, &mn);
8616 /* We want mdb_rebalance to find mn when doing fixups */
8617 WITH_CURSOR_TRACKING(mn,
8618 rc = mdb_update_key(&mn, &key));
8622 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8624 indx_t ix = csrc->mc_ki[csrc->mc_top];
8625 nullkey.mv_size = 0;
8626 csrc->mc_ki[csrc->mc_top] = 0;
8627 rc = mdb_update_key(csrc, &nullkey);
8628 csrc->mc_ki[csrc->mc_top] = ix;
8629 mdb_cassert(csrc, rc == MDB_SUCCESS);
8633 if (cdst->mc_ki[cdst->mc_top] == 0) {
8634 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
8635 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8636 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
8638 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
8639 key.mv_size = NODEKSZ(srcnode);
8640 key.mv_data = NODEKEY(srcnode);
8642 DPRINTF(("update separator for destination page %"Y"u to [%s]",
8643 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
8644 mdb_cursor_copy(cdst, &mn);
8647 /* We want mdb_rebalance to find mn when doing fixups */
8648 WITH_CURSOR_TRACKING(mn,
8649 rc = mdb_update_key(&mn, &key));
8653 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
8655 indx_t ix = cdst->mc_ki[cdst->mc_top];
8656 nullkey.mv_size = 0;
8657 cdst->mc_ki[cdst->mc_top] = 0;
8658 rc = mdb_update_key(cdst, &nullkey);
8659 cdst->mc_ki[cdst->mc_top] = ix;
8660 mdb_cassert(cdst, rc == MDB_SUCCESS);
8667 /** Merge one page into another.
8668 * The nodes from the page pointed to by \b csrc will
8669 * be copied to the page pointed to by \b cdst and then
8670 * the \b csrc page will be freed.
8671 * @param[in] csrc Cursor pointing to the source page.
8672 * @param[in] cdst Cursor pointing to the destination page.
8673 * @return 0 on success, non-zero on failure.
8676 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8678 MDB_page *psrc, *pdst;
8685 psrc = csrc->mc_pg[csrc->mc_top];
8686 pdst = cdst->mc_pg[cdst->mc_top];
8688 DPRINTF(("merging page %"Y"u into %"Y"u", psrc->mp_pgno, pdst->mp_pgno));
8690 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8691 mdb_cassert(csrc, cdst->mc_snum > 1);
8693 /* Mark dst as dirty. */
8694 if ((rc = mdb_page_touch(cdst)))
8697 /* get dst page again now that we've touched it. */
8698 pdst = cdst->mc_pg[cdst->mc_top];
8700 /* Move all nodes from src to dst.
8702 j = nkeys = NUMKEYS(pdst);
8703 if (IS_LEAF2(psrc)) {
8704 key.mv_size = csrc->mc_db->md_pad;
8705 key.mv_data = METADATA(psrc);
8706 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8707 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8708 if (rc != MDB_SUCCESS)
8710 key.mv_data = (char *)key.mv_data + key.mv_size;
8713 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8714 srcnode = NODEPTR(psrc, i);
8715 if (i == 0 && IS_BRANCH(psrc)) {
8718 mdb_cursor_copy(csrc, &mn);
8719 mn.mc_xcursor = NULL;
8720 /* must find the lowest key below src */
8721 rc = mdb_page_search_lowest(&mn);
8724 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8725 key.mv_size = mn.mc_db->md_pad;
8726 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8728 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8729 key.mv_size = NODEKSZ(s2);
8730 key.mv_data = NODEKEY(s2);
8733 key.mv_size = srcnode->mn_ksize;
8734 key.mv_data = NODEKEY(srcnode);
8737 data.mv_size = NODEDSZ(srcnode);
8738 data.mv_data = NODEDATA(srcnode);
8739 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8740 if (rc != MDB_SUCCESS)
8745 DPRINTF(("dst page %"Y"u now has %u keys (%.1f%% filled)",
8746 pdst->mp_pgno, NUMKEYS(pdst),
8747 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8749 /* Unlink the src page from parent and add to free list.
8752 mdb_node_del(csrc, 0);
8753 if (csrc->mc_ki[csrc->mc_top] == 0) {
8755 rc = mdb_update_key(csrc, &key);
8763 psrc = csrc->mc_pg[csrc->mc_top];
8764 /* If not operating on FreeDB, allow this page to be reused
8765 * in this txn. Otherwise just add to free list.
8767 rc = mdb_page_loose(csrc, psrc);
8771 csrc->mc_db->md_leaf_pages--;
8773 csrc->mc_db->md_branch_pages--;
8775 /* Adjust other cursors pointing to mp */
8776 MDB_cursor *m2, *m3;
8777 MDB_dbi dbi = csrc->mc_dbi;
8778 unsigned int top = csrc->mc_top;
8780 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8781 if (csrc->mc_flags & C_SUB)
8782 m3 = &m2->mc_xcursor->mx_cursor;
8785 if (m3 == csrc) continue;
8786 if (m3->mc_snum < csrc->mc_snum) continue;
8787 if (m3->mc_pg[top] == psrc) {
8788 m3->mc_pg[top] = pdst;
8789 m3->mc_ki[top] += nkeys;
8790 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8791 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8792 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8795 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8797 MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
8798 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8799 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8804 unsigned int snum = cdst->mc_snum;
8805 uint16_t depth = cdst->mc_db->md_depth;
8806 mdb_cursor_pop(cdst);
8807 rc = mdb_rebalance(cdst);
8808 /* Did the tree height change? */
8809 if (depth != cdst->mc_db->md_depth)
8810 snum += cdst->mc_db->md_depth - depth;
8811 cdst->mc_snum = snum;
8812 cdst->mc_top = snum-1;
8817 /** Copy the contents of a cursor.
8818 * @param[in] csrc The cursor to copy from.
8819 * @param[out] cdst The cursor to copy to.
8822 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8826 cdst->mc_txn = csrc->mc_txn;
8827 cdst->mc_dbi = csrc->mc_dbi;
8828 cdst->mc_db = csrc->mc_db;
8829 cdst->mc_dbx = csrc->mc_dbx;
8830 cdst->mc_snum = csrc->mc_snum;
8831 cdst->mc_top = csrc->mc_top;
8832 cdst->mc_flags = csrc->mc_flags;
8834 cdst->mc_ovpg = csrc->mc_ovpg;
8837 for (i=0; i<csrc->mc_snum; i++) {
8838 cdst->mc_pg[i] = csrc->mc_pg[i];
8839 cdst->mc_ki[i] = csrc->mc_ki[i];
8843 /** Rebalance the tree after a delete operation.
8844 * @param[in] mc Cursor pointing to the page where rebalancing
8846 * @return 0 on success, non-zero on failure.
8849 mdb_rebalance(MDB_cursor *mc)
8853 unsigned int ptop, minkeys, thresh;
8857 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8862 thresh = FILL_THRESHOLD;
8864 DPRINTF(("rebalancing %s page %"Y"u (has %u keys, %.1f%% full)",
8865 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8866 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8867 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8869 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8870 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8871 DPRINTF(("no need to rebalance page %"Y"u, above fill threshold",
8872 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8876 if (mc->mc_snum < 2) {
8877 MDB_page *mp = mc->mc_pg[0];
8879 DPUTS("Can't rebalance a subpage, ignoring");
8882 if (NUMKEYS(mp) == 0) {
8883 DPUTS("tree is completely empty");
8884 mc->mc_db->md_root = P_INVALID;
8885 mc->mc_db->md_depth = 0;
8886 mc->mc_db->md_leaf_pages = 0;
8887 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8890 /* Adjust cursors pointing to mp */
8893 mc->mc_flags &= ~C_INITIALIZED;
8895 MDB_cursor *m2, *m3;
8896 MDB_dbi dbi = mc->mc_dbi;
8898 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8899 if (mc->mc_flags & C_SUB)
8900 m3 = &m2->mc_xcursor->mx_cursor;
8903 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8905 if (m3->mc_pg[0] == mp) {
8908 m3->mc_flags &= ~C_INITIALIZED;
8912 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8914 DPUTS("collapsing root page!");
8915 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8918 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8919 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
8922 mc->mc_db->md_depth--;
8923 mc->mc_db->md_branch_pages--;
8924 mc->mc_ki[0] = mc->mc_ki[1];
8925 for (i = 1; i<mc->mc_db->md_depth; i++) {
8926 mc->mc_pg[i] = mc->mc_pg[i+1];
8927 mc->mc_ki[i] = mc->mc_ki[i+1];
8930 /* Adjust other cursors pointing to mp */
8931 MDB_cursor *m2, *m3;
8932 MDB_dbi dbi = mc->mc_dbi;
8934 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8935 if (mc->mc_flags & C_SUB)
8936 m3 = &m2->mc_xcursor->mx_cursor;
8939 if (m3 == mc) continue;
8940 if (!(m3->mc_flags & C_INITIALIZED))
8942 if (m3->mc_pg[0] == mp) {
8943 for (i=0; i<mc->mc_db->md_depth; i++) {
8944 m3->mc_pg[i] = m3->mc_pg[i+1];
8945 m3->mc_ki[i] = m3->mc_ki[i+1];
8953 DPUTS("root page doesn't need rebalancing");
8957 /* The parent (branch page) must have at least 2 pointers,
8958 * otherwise the tree is invalid.
8960 ptop = mc->mc_top-1;
8961 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8963 /* Leaf page fill factor is below the threshold.
8964 * Try to move keys from left or right neighbor, or
8965 * merge with a neighbor page.
8970 mdb_cursor_copy(mc, &mn);
8971 mn.mc_xcursor = NULL;
8973 oldki = mc->mc_ki[mc->mc_top];
8974 if (mc->mc_ki[ptop] == 0) {
8975 /* We're the leftmost leaf in our parent.
8977 DPUTS("reading right neighbor");
8979 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8980 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8983 mn.mc_ki[mn.mc_top] = 0;
8984 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8987 /* There is at least one neighbor to the left.
8989 DPUTS("reading left neighbor");
8991 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8992 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8995 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8996 mc->mc_ki[mc->mc_top] = 0;
9000 DPRINTF(("found neighbor page %"Y"u (%u keys, %.1f%% full)",
9001 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
9002 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
9004 /* If the neighbor page is above threshold and has enough keys,
9005 * move one key from it. Otherwise we should try to merge them.
9006 * (A branch page must never have less than 2 keys.)
9008 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
9009 rc = mdb_node_move(&mn, mc, fromleft);
9011 /* if we inserted on left, bump position up */
9016 rc = mdb_page_merge(&mn, mc);
9018 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
9019 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
9020 /* We want mdb_rebalance to find mn when doing fixups */
9021 WITH_CURSOR_TRACKING(mn,
9022 rc = mdb_page_merge(mc, &mn));
9023 mdb_cursor_copy(&mn, mc);
9025 mc->mc_flags &= ~C_EOF;
9027 mc->mc_ki[mc->mc_top] = oldki;
9031 /** Complete a delete operation started by #mdb_cursor_del(). */
9033 mdb_cursor_del0(MDB_cursor *mc)
9039 MDB_cursor *m2, *m3;
9040 MDB_dbi dbi = mc->mc_dbi;
9042 ki = mc->mc_ki[mc->mc_top];
9043 mp = mc->mc_pg[mc->mc_top];
9044 mdb_node_del(mc, mc->mc_db->md_pad);
9045 mc->mc_db->md_entries--;
9047 /* Adjust other cursors pointing to mp */
9048 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9049 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9050 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9052 if (m3 == mc || m3->mc_snum < mc->mc_snum)
9054 if (m3->mc_pg[mc->mc_top] == mp) {
9055 if (m3->mc_ki[mc->mc_top] == ki) {
9056 m3->mc_flags |= C_DEL;
9057 } else if (m3->mc_ki[mc->mc_top] > ki) {
9058 m3->mc_ki[mc->mc_top]--;
9060 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
9061 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9062 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
9063 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9068 rc = mdb_rebalance(mc);
9070 if (rc == MDB_SUCCESS) {
9071 /* DB is totally empty now, just bail out.
9072 * Other cursors adjustments were already done
9073 * by mdb_rebalance and aren't needed here.
9078 mp = mc->mc_pg[mc->mc_top];
9079 nkeys = NUMKEYS(mp);
9081 /* Adjust other cursors pointing to mp */
9082 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
9083 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9084 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9086 if (m3->mc_snum < mc->mc_snum)
9088 if (m3->mc_pg[mc->mc_top] == mp) {
9089 /* if m3 points past last node in page, find next sibling */
9090 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
9091 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9092 rc = mdb_cursor_sibling(m3, 1);
9093 if (rc == MDB_NOTFOUND) {
9094 m3->mc_flags |= C_EOF;
9099 if (mc->mc_db->md_flags & MDB_DUPSORT) {
9100 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
9101 if (node->mn_flags & F_DUPDATA) {
9102 mdb_xcursor_init1(m3, node);
9103 m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL;
9109 mc->mc_flags |= C_DEL;
9113 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9118 mdb_del(MDB_txn *txn, MDB_dbi dbi,
9119 MDB_val *key, MDB_val *data)
9121 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9124 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9125 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9127 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
9128 /* must ignore any data */
9132 return mdb_del0(txn, dbi, key, data, 0);
9136 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
9137 MDB_val *key, MDB_val *data, unsigned flags)
9142 MDB_val rdata, *xdata;
9146 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
9148 mdb_cursor_init(&mc, txn, dbi, &mx);
9157 flags |= MDB_NODUPDATA;
9159 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
9161 /* let mdb_page_split know about this cursor if needed:
9162 * delete will trigger a rebalance; if it needs to move
9163 * a node from one page to another, it will have to
9164 * update the parent's separator key(s). If the new sepkey
9165 * is larger than the current one, the parent page may
9166 * run out of space, triggering a split. We need this
9167 * cursor to be consistent until the end of the rebalance.
9169 mc.mc_flags |= C_UNTRACK;
9170 mc.mc_next = txn->mt_cursors[dbi];
9171 txn->mt_cursors[dbi] = &mc;
9172 rc = mdb_cursor_del(&mc, flags);
9173 txn->mt_cursors[dbi] = mc.mc_next;
9178 /** Split a page and insert a new node.
9179 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
9180 * The cursor will be updated to point to the actual page and index where
9181 * the node got inserted after the split.
9182 * @param[in] newkey The key for the newly inserted node.
9183 * @param[in] newdata The data for the newly inserted node.
9184 * @param[in] newpgno The page number, if the new node is a branch node.
9185 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
9186 * @return 0 on success, non-zero on failure.
9189 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
9190 unsigned int nflags)
9193 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
9196 int i, j, split_indx, nkeys, pmax;
9197 MDB_env *env = mc->mc_txn->mt_env;
9199 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
9200 MDB_page *copy = NULL;
9201 MDB_page *mp, *rp, *pp;
9206 mp = mc->mc_pg[mc->mc_top];
9207 newindx = mc->mc_ki[mc->mc_top];
9208 nkeys = NUMKEYS(mp);
9210 DPRINTF(("-----> splitting %s page %"Y"u and adding [%s] at index %i/%i",
9211 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
9212 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
9214 /* Create a right sibling. */
9215 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
9217 rp->mp_pad = mp->mp_pad;
9218 DPRINTF(("new right sibling: page %"Y"u", rp->mp_pgno));
9220 /* Usually when splitting the root page, the cursor
9221 * height is 1. But when called from mdb_update_key,
9222 * the cursor height may be greater because it walks
9223 * up the stack while finding the branch slot to update.
9225 if (mc->mc_top < 1) {
9226 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
9228 /* shift current top to make room for new parent */
9229 for (i=mc->mc_snum; i>0; i--) {
9230 mc->mc_pg[i] = mc->mc_pg[i-1];
9231 mc->mc_ki[i] = mc->mc_ki[i-1];
9235 mc->mc_db->md_root = pp->mp_pgno;
9236 DPRINTF(("root split! new root = %"Y"u", pp->mp_pgno));
9237 new_root = mc->mc_db->md_depth++;
9239 /* Add left (implicit) pointer. */
9240 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
9241 /* undo the pre-push */
9242 mc->mc_pg[0] = mc->mc_pg[1];
9243 mc->mc_ki[0] = mc->mc_ki[1];
9244 mc->mc_db->md_root = mp->mp_pgno;
9245 mc->mc_db->md_depth--;
9252 ptop = mc->mc_top-1;
9253 DPRINTF(("parent branch page is %"Y"u", mc->mc_pg[ptop]->mp_pgno));
9256 mdb_cursor_copy(mc, &mn);
9257 mn.mc_xcursor = NULL;
9258 mn.mc_pg[mn.mc_top] = rp;
9259 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
9261 if (nflags & MDB_APPEND) {
9262 mn.mc_ki[mn.mc_top] = 0;
9264 split_indx = newindx;
9268 split_indx = (nkeys+1) / 2;
9273 unsigned int lsize, rsize, ksize;
9274 /* Move half of the keys to the right sibling */
9275 x = mc->mc_ki[mc->mc_top] - split_indx;
9276 ksize = mc->mc_db->md_pad;
9277 split = LEAF2KEY(mp, split_indx, ksize);
9278 rsize = (nkeys - split_indx) * ksize;
9279 lsize = (nkeys - split_indx) * sizeof(indx_t);
9280 mp->mp_lower -= lsize;
9281 rp->mp_lower += lsize;
9282 mp->mp_upper += rsize - lsize;
9283 rp->mp_upper -= rsize - lsize;
9284 sepkey.mv_size = ksize;
9285 if (newindx == split_indx) {
9286 sepkey.mv_data = newkey->mv_data;
9288 sepkey.mv_data = split;
9291 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
9292 memcpy(rp->mp_ptrs, split, rsize);
9293 sepkey.mv_data = rp->mp_ptrs;
9294 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
9295 memcpy(ins, newkey->mv_data, ksize);
9296 mp->mp_lower += sizeof(indx_t);
9297 mp->mp_upper -= ksize - sizeof(indx_t);
9300 memcpy(rp->mp_ptrs, split, x * ksize);
9301 ins = LEAF2KEY(rp, x, ksize);
9302 memcpy(ins, newkey->mv_data, ksize);
9303 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
9304 rp->mp_lower += sizeof(indx_t);
9305 rp->mp_upper -= ksize - sizeof(indx_t);
9306 mc->mc_ki[mc->mc_top] = x;
9309 int psize, nsize, k;
9310 /* Maximum free space in an empty page */
9311 pmax = env->me_psize - PAGEHDRSZ;
9313 nsize = mdb_leaf_size(env, newkey, newdata);
9315 nsize = mdb_branch_size(env, newkey);
9316 nsize = EVEN(nsize);
9318 /* grab a page to hold a temporary copy */
9319 copy = mdb_page_malloc(mc->mc_txn, 1);
9324 copy->mp_pgno = mp->mp_pgno;
9325 copy->mp_flags = mp->mp_flags;
9326 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
9327 copy->mp_upper = env->me_psize - PAGEBASE;
9329 /* prepare to insert */
9330 for (i=0, j=0; i<nkeys; i++) {
9332 copy->mp_ptrs[j++] = 0;
9334 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
9337 /* When items are relatively large the split point needs
9338 * to be checked, because being off-by-one will make the
9339 * difference between success or failure in mdb_node_add.
9341 * It's also relevant if a page happens to be laid out
9342 * such that one half of its nodes are all "small" and
9343 * the other half of its nodes are "large." If the new
9344 * item is also "large" and falls on the half with
9345 * "large" nodes, it also may not fit.
9347 * As a final tweak, if the new item goes on the last
9348 * spot on the page (and thus, onto the new page), bias
9349 * the split so the new page is emptier than the old page.
9350 * This yields better packing during sequential inserts.
9352 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
9353 /* Find split point */
9355 if (newindx <= split_indx || newindx >= nkeys) {
9357 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
9362 for (; i!=k; i+=j) {
9367 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9368 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
9370 if (F_ISSET(node->mn_flags, F_BIGDATA))
9371 psize += sizeof(pgno_t);
9373 psize += NODEDSZ(node);
9375 psize = EVEN(psize);
9377 if (psize > pmax || i == k-j) {
9378 split_indx = i + (j<0);
9383 if (split_indx == newindx) {
9384 sepkey.mv_size = newkey->mv_size;
9385 sepkey.mv_data = newkey->mv_data;
9387 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
9388 sepkey.mv_size = node->mn_ksize;
9389 sepkey.mv_data = NODEKEY(node);
9394 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
9396 /* Copy separator key to the parent.
9398 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
9399 int snum = mc->mc_snum;
9403 /* We want other splits to find mn when doing fixups */
9404 WITH_CURSOR_TRACKING(mn,
9405 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
9410 if (mc->mc_snum > snum) {
9413 /* Right page might now have changed parent.
9414 * Check if left page also changed parent.
9416 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9417 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9418 for (i=0; i<ptop; i++) {
9419 mc->mc_pg[i] = mn.mc_pg[i];
9420 mc->mc_ki[i] = mn.mc_ki[i];
9422 mc->mc_pg[ptop] = mn.mc_pg[ptop];
9423 if (mn.mc_ki[ptop]) {
9424 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
9426 /* find right page's left sibling */
9427 mc->mc_ki[ptop] = mn.mc_ki[ptop];
9428 mdb_cursor_sibling(mc, 0);
9433 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
9436 if (rc != MDB_SUCCESS) {
9439 if (nflags & MDB_APPEND) {
9440 mc->mc_pg[mc->mc_top] = rp;
9441 mc->mc_ki[mc->mc_top] = 0;
9442 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
9445 for (i=0; i<mc->mc_top; i++)
9446 mc->mc_ki[i] = mn.mc_ki[i];
9447 } else if (!IS_LEAF2(mp)) {
9449 mc->mc_pg[mc->mc_top] = rp;
9454 rkey.mv_data = newkey->mv_data;
9455 rkey.mv_size = newkey->mv_size;
9461 /* Update index for the new key. */
9462 mc->mc_ki[mc->mc_top] = j;
9464 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9465 rkey.mv_data = NODEKEY(node);
9466 rkey.mv_size = node->mn_ksize;
9468 xdata.mv_data = NODEDATA(node);
9469 xdata.mv_size = NODEDSZ(node);
9472 pgno = NODEPGNO(node);
9473 flags = node->mn_flags;
9476 if (!IS_LEAF(mp) && j == 0) {
9477 /* First branch index doesn't need key data. */
9481 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
9487 mc->mc_pg[mc->mc_top] = copy;
9492 } while (i != split_indx);
9494 nkeys = NUMKEYS(copy);
9495 for (i=0; i<nkeys; i++)
9496 mp->mp_ptrs[i] = copy->mp_ptrs[i];
9497 mp->mp_lower = copy->mp_lower;
9498 mp->mp_upper = copy->mp_upper;
9499 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
9500 env->me_psize - copy->mp_upper - PAGEBASE);
9502 /* reset back to original page */
9503 if (newindx < split_indx) {
9504 mc->mc_pg[mc->mc_top] = mp;
9506 mc->mc_pg[mc->mc_top] = rp;
9508 /* Make sure mc_ki is still valid.
9510 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9511 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9512 for (i=0; i<=ptop; i++) {
9513 mc->mc_pg[i] = mn.mc_pg[i];
9514 mc->mc_ki[i] = mn.mc_ki[i];
9518 if (nflags & MDB_RESERVE) {
9519 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
9520 if (!(node->mn_flags & F_BIGDATA))
9521 newdata->mv_data = NODEDATA(node);
9524 if (newindx >= split_indx) {
9525 mc->mc_pg[mc->mc_top] = rp;
9527 /* Make sure mc_ki is still valid.
9529 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9530 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9531 for (i=0; i<=ptop; i++) {
9532 mc->mc_pg[i] = mn.mc_pg[i];
9533 mc->mc_ki[i] = mn.mc_ki[i];
9540 /* Adjust other cursors pointing to mp */
9541 MDB_cursor *m2, *m3;
9542 MDB_dbi dbi = mc->mc_dbi;
9543 nkeys = NUMKEYS(mp);
9545 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9546 if (mc->mc_flags & C_SUB)
9547 m3 = &m2->mc_xcursor->mx_cursor;
9552 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9556 /* sub cursors may be on different DB */
9557 if (m3->mc_pg[0] != mp)
9560 for (k=new_root; k>=0; k--) {
9561 m3->mc_ki[k+1] = m3->mc_ki[k];
9562 m3->mc_pg[k+1] = m3->mc_pg[k];
9564 if (m3->mc_ki[0] >= nkeys) {
9569 m3->mc_pg[0] = mc->mc_pg[0];
9573 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
9574 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
9575 m3->mc_ki[mc->mc_top]++;
9576 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9577 m3->mc_pg[mc->mc_top] = rp;
9578 m3->mc_ki[mc->mc_top] -= nkeys;
9579 for (i=0; i<mc->mc_top; i++) {
9580 m3->mc_ki[i] = mn.mc_ki[i];
9581 m3->mc_pg[i] = mn.mc_pg[i];
9584 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
9585 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
9588 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
9590 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9591 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
9592 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9596 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
9599 if (copy) /* tmp page */
9600 mdb_page_free(env, copy);
9602 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9607 mdb_put(MDB_txn *txn, MDB_dbi dbi,
9608 MDB_val *key, MDB_val *data, unsigned int flags)
9614 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9617 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
9620 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9621 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9623 mdb_cursor_init(&mc, txn, dbi, &mx);
9624 mc.mc_next = txn->mt_cursors[dbi];
9625 txn->mt_cursors[dbi] = &mc;
9626 rc = mdb_cursor_put(&mc, key, data, flags);
9627 txn->mt_cursors[dbi] = mc.mc_next;
9632 #define MDB_WBUF (1024*1024)
9634 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
9636 /** State needed for a double-buffering compacting copy. */
9637 typedef struct mdb_copy {
9638 pthread_mutex_t mc_mutex;
9639 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
9646 pgno_t mc_next_pgno;
9648 int mc_toggle; /**< Buffer number in provider */
9649 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
9650 volatile int mc_error; /**< Error code, never cleared if set */
9653 /** Dedicated writer thread for compacting copy. */
9654 static THREAD_RET ESECT CALL_CONV
9655 mdb_env_copythr(void *arg)
9659 int toggle = 0, wsize, rc;
9662 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9665 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9668 pthread_mutex_lock(&my->mc_mutex);
9671 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9672 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
9674 wsize = my->mc_wlen[toggle];
9675 ptr = my->mc_wbuf[toggle];
9678 while (wsize > 0 && !my->mc_error) {
9679 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9683 } else if (len > 0) {
9696 /* If there's an overflow page tail, write it too */
9697 if (my->mc_olen[toggle]) {
9698 wsize = my->mc_olen[toggle];
9699 ptr = my->mc_over[toggle];
9700 my->mc_olen[toggle] = 0;
9703 my->mc_wlen[toggle] = 0;
9705 /* Return the empty buffer to provider */
9707 pthread_cond_signal(&my->mc_cond);
9709 pthread_mutex_unlock(&my->mc_mutex);
9710 return (THREAD_RET)0;
9714 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9716 * @param[in] my control structure.
9717 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9720 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9722 pthread_mutex_lock(&my->mc_mutex);
9723 my->mc_new += adjust;
9724 pthread_cond_signal(&my->mc_cond);
9725 while (my->mc_new & 2) /* both buffers in use */
9726 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9727 pthread_mutex_unlock(&my->mc_mutex);
9729 my->mc_toggle ^= (adjust & 1);
9730 /* Both threads reset mc_wlen, to be safe from threading errors */
9731 my->mc_wlen[my->mc_toggle] = 0;
9732 return my->mc_error;
9735 /** Depth-first tree traversal for compacting copy. */
9737 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9739 MDB_cursor mc = {0};
9741 MDB_page *mo, *mp, *leaf;
9746 /* Empty DB, nothing to do */
9747 if (*pg == P_INVALID)
9751 mc.mc_txn = my->mc_txn;
9752 mc.mc_flags = my->mc_txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
9754 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9757 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9761 /* Make cursor pages writable */
9762 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9766 for (i=0; i<mc.mc_top; i++) {
9767 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9768 mc.mc_pg[i] = (MDB_page *)ptr;
9769 ptr += my->mc_env->me_psize;
9772 /* This is writable space for a leaf page. Usually not needed. */
9773 leaf = (MDB_page *)ptr;
9775 toggle = my->mc_toggle;
9776 while (mc.mc_snum > 0) {
9778 mp = mc.mc_pg[mc.mc_top];
9782 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9783 for (i=0; i<n; i++) {
9784 ni = NODEPTR(mp, i);
9785 if (ni->mn_flags & F_BIGDATA) {
9789 /* Need writable leaf */
9791 mc.mc_pg[mc.mc_top] = leaf;
9792 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9794 ni = NODEPTR(mp, i);
9797 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9798 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
9799 rc = mdb_page_get(&mc, pg, &omp, NULL);
9802 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9803 rc = mdb_env_cthr_toggle(my, 1);
9806 toggle = my->mc_toggle;
9808 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9809 memcpy(mo, omp, my->mc_env->me_psize);
9810 mo->mp_pgno = my->mc_next_pgno;
9811 my->mc_next_pgno += omp->mp_pages;
9812 my->mc_wlen[toggle] += my->mc_env->me_psize;
9813 if (omp->mp_pages > 1) {
9814 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9815 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9816 rc = mdb_env_cthr_toggle(my, 1);
9819 toggle = my->mc_toggle;
9821 } else if (ni->mn_flags & F_SUBDATA) {
9824 /* Need writable leaf */
9826 mc.mc_pg[mc.mc_top] = leaf;
9827 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9829 ni = NODEPTR(mp, i);
9832 memcpy(&db, NODEDATA(ni), sizeof(db));
9833 my->mc_toggle = toggle;
9834 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9837 toggle = my->mc_toggle;
9838 memcpy(NODEDATA(ni), &db, sizeof(db));
9843 mc.mc_ki[mc.mc_top]++;
9844 if (mc.mc_ki[mc.mc_top] < n) {
9847 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9849 rc = mdb_page_get(&mc, pg, &mp, NULL);
9854 mc.mc_ki[mc.mc_top] = 0;
9855 if (IS_BRANCH(mp)) {
9856 /* Whenever we advance to a sibling branch page,
9857 * we must proceed all the way down to its first leaf.
9859 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9862 mc.mc_pg[mc.mc_top] = mp;
9866 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9867 rc = mdb_env_cthr_toggle(my, 1);
9870 toggle = my->mc_toggle;
9872 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9873 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9874 mo->mp_pgno = my->mc_next_pgno++;
9875 my->mc_wlen[toggle] += my->mc_env->me_psize;
9877 /* Update parent if there is one */
9878 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9879 SETPGNO(ni, mo->mp_pgno);
9880 mdb_cursor_pop(&mc);
9882 /* Otherwise we're done */
9892 /** Copy environment with compaction. */
9894 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9899 MDB_txn *txn = NULL;
9901 pgno_t root, new_root;
9902 int rc = MDB_SUCCESS;
9905 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
9906 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
9910 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9911 if (my.mc_wbuf[0] == NULL) {
9912 /* _aligned_malloc() sets errno, but we use Windows error codes */
9913 rc = ERROR_NOT_ENOUGH_MEMORY;
9917 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
9919 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
9921 #ifdef HAVE_MEMALIGN
9922 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9923 if (my.mc_wbuf[0] == NULL) {
9930 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
9936 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9937 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9938 my.mc_next_pgno = NUM_METAS;
9941 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
9945 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9949 mp = (MDB_page *)my.mc_wbuf[0];
9950 memset(mp, 0, NUM_METAS * env->me_psize);
9952 mp->mp_flags = P_META;
9953 mm = (MDB_meta *)METADATA(mp);
9954 mdb_env_init_meta0(env, mm);
9955 mm->mm_address = env->me_metas[0]->mm_address;
9957 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9959 mp->mp_flags = P_META;
9960 *(MDB_meta *)METADATA(mp) = *mm;
9961 mm = (MDB_meta *)METADATA(mp);
9963 /* Set metapage 1 with current main DB */
9964 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
9965 if (root != P_INVALID) {
9966 /* Count free pages + freeDB pages. Subtract from last_pg
9967 * to find the new last_pg, which also becomes the new root.
9969 MDB_ID freecount = 0;
9972 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9973 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9974 freecount += *(MDB_ID *)data.mv_data;
9975 if (rc != MDB_NOTFOUND)
9977 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9978 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9979 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9981 new_root = txn->mt_next_pgno - 1 - freecount;
9982 mm->mm_last_pg = new_root;
9983 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9984 mm->mm_dbs[MAIN_DBI].md_root = new_root;
9986 /* When the DB is empty, handle it specially to
9987 * fix any breakage like page leaks from ITS#8174.
9989 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
9991 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
9992 mm->mm_txnid = 1; /* use metapage 1 */
9995 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9997 rc = mdb_env_cwalk(&my, &root, 0);
9998 if (rc == MDB_SUCCESS && root != new_root) {
9999 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
10005 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
10006 rc = THREAD_FINISH(thr);
10007 mdb_txn_abort(txn);
10011 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
10012 if (my.mc_cond) CloseHandle(my.mc_cond);
10013 if (my.mc_mutex) CloseHandle(my.mc_mutex);
10015 free(my.mc_wbuf[0]);
10016 pthread_cond_destroy(&my.mc_cond);
10018 pthread_mutex_destroy(&my.mc_mutex);
10020 return rc ? rc : my.mc_error;
10023 /** Copy environment as-is. */
10025 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
10027 MDB_txn *txn = NULL;
10028 mdb_mutexref_t wmutex = NULL;
10030 mdb_size_t wsize, w3;
10034 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
10038 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
10041 /* Do the lock/unlock of the reader mutex before starting the
10042 * write txn. Otherwise other read txns could block writers.
10044 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
10048 if (env->me_txns) {
10049 /* We must start the actual read txn after blocking writers */
10050 mdb_txn_end(txn, MDB_END_RESET_TMP);
10052 /* Temporarily block writers until we snapshot the meta pages */
10053 wmutex = env->me_wmutex;
10054 if (LOCK_MUTEX(rc, env, wmutex))
10057 rc = mdb_txn_renew0(txn);
10059 UNLOCK_MUTEX(wmutex);
10064 wsize = env->me_psize * NUM_METAS;
10068 DO_WRITE(rc, fd, ptr, w2, len);
10072 } else if (len > 0) {
10078 /* Non-blocking or async handles are not supported */
10084 UNLOCK_MUTEX(wmutex);
10089 w3 = txn->mt_next_pgno * env->me_psize;
10091 mdb_size_t fsize = 0;
10092 if ((rc = mdb_fsize(env->me_fd, &fsize)))
10097 wsize = w3 - wsize;
10098 while (wsize > 0) {
10099 if (wsize > MAX_WRITE)
10103 DO_WRITE(rc, fd, ptr, w2, len);
10107 } else if (len > 0) {
10119 mdb_txn_abort(txn);
10124 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
10126 if (flags & MDB_CP_COMPACT)
10127 return mdb_env_copyfd1(env, fd);
10129 return mdb_env_copyfd0(env, fd);
10133 mdb_env_copyfd(MDB_env *env, HANDLE fd)
10135 return mdb_env_copyfd2(env, fd, 0);
10139 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
10143 HANDLE newfd = INVALID_HANDLE_VALUE;
10148 if (env->me_flags & MDB_NOSUBDIR) {
10149 lpath = (char *)path;
10151 len = strlen(path);
10152 len += sizeof(DATANAME);
10153 lpath = malloc(len);
10156 sprintf(lpath, "%s" DATANAME, path);
10159 /* The destination path must exist, but the destination file must not.
10160 * We don't want the OS to cache the writes, since the source data is
10161 * already in the OS cache.
10164 rc = utf8_to_utf16(lpath, -1, &wpath, NULL);
10167 newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
10168 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
10171 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
10173 if (newfd == INVALID_HANDLE_VALUE) {
10178 if (env->me_psize >= env->me_os_psize) {
10180 /* Set O_DIRECT if the file system supports it */
10181 if ((rc = fcntl(newfd, F_GETFL)) != -1)
10182 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
10184 #ifdef F_NOCACHE /* __APPLE__ */
10185 rc = fcntl(newfd, F_NOCACHE, 1);
10193 rc = mdb_env_copyfd2(env, newfd, flags);
10196 if (!(env->me_flags & MDB_NOSUBDIR))
10198 if (newfd != INVALID_HANDLE_VALUE)
10199 if (close(newfd) < 0 && rc == MDB_SUCCESS)
10206 mdb_env_copy(MDB_env *env, const char *path)
10208 return mdb_env_copy2(env, path, 0);
10212 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
10214 if (flag & ~CHANGEABLE)
10217 env->me_flags |= flag;
10219 env->me_flags &= ~flag;
10220 return MDB_SUCCESS;
10224 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
10229 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
10230 return MDB_SUCCESS;
10234 mdb_env_set_userctx(MDB_env *env, void *ctx)
10238 env->me_userctx = ctx;
10239 return MDB_SUCCESS;
10243 mdb_env_get_userctx(MDB_env *env)
10245 return env ? env->me_userctx : NULL;
10249 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
10254 env->me_assert_func = func;
10256 return MDB_SUCCESS;
10260 mdb_env_get_path(MDB_env *env, const char **arg)
10265 *arg = env->me_path;
10266 return MDB_SUCCESS;
10270 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
10276 return MDB_SUCCESS;
10279 /** Common code for #mdb_stat() and #mdb_env_stat().
10280 * @param[in] env the environment to operate in.
10281 * @param[in] db the #MDB_db record containing the stats to return.
10282 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
10283 * @return 0, this function always succeeds.
10286 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
10288 arg->ms_psize = env->me_psize;
10289 arg->ms_depth = db->md_depth;
10290 arg->ms_branch_pages = db->md_branch_pages;
10291 arg->ms_leaf_pages = db->md_leaf_pages;
10292 arg->ms_overflow_pages = db->md_overflow_pages;
10293 arg->ms_entries = db->md_entries;
10295 return MDB_SUCCESS;
10299 mdb_env_stat(MDB_env *env, MDB_stat *arg)
10303 if (env == NULL || arg == NULL)
10306 meta = mdb_env_pick_meta(env);
10308 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
10312 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
10316 if (env == NULL || arg == NULL)
10319 meta = mdb_env_pick_meta(env);
10320 arg->me_mapaddr = meta->mm_address;
10321 arg->me_last_pgno = meta->mm_last_pg;
10322 arg->me_last_txnid = meta->mm_txnid;
10324 arg->me_mapsize = env->me_mapsize;
10325 arg->me_maxreaders = env->me_maxreaders;
10326 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
10327 return MDB_SUCCESS;
10330 /** Set the default comparison functions for a database.
10331 * Called immediately after a database is opened to set the defaults.
10332 * The user can then override them with #mdb_set_compare() or
10333 * #mdb_set_dupsort().
10334 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
10335 * @param[in] dbi A database handle returned by #mdb_dbi_open()
10338 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
10340 uint16_t f = txn->mt_dbs[dbi].md_flags;
10342 txn->mt_dbxs[dbi].md_cmp =
10343 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
10344 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
10346 txn->mt_dbxs[dbi].md_dcmp =
10347 !(f & MDB_DUPSORT) ? 0 :
10348 ((f & MDB_INTEGERDUP)
10349 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
10350 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
10353 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
10359 int rc, dbflag, exact;
10360 unsigned int unused = 0, seq;
10364 if (flags & ~VALID_FLAGS)
10366 if (txn->mt_flags & MDB_TXN_BLOCKED)
10367 return MDB_BAD_TXN;
10372 if (flags & PERSISTENT_FLAGS) {
10373 uint16_t f2 = flags & PERSISTENT_FLAGS;
10374 /* make sure flag changes get committed */
10375 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
10376 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
10377 txn->mt_flags |= MDB_TXN_DIRTY;
10380 mdb_default_cmp(txn, MAIN_DBI);
10381 return MDB_SUCCESS;
10384 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
10385 mdb_default_cmp(txn, MAIN_DBI);
10388 /* Is the DB already open? */
10389 len = strlen(name);
10390 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
10391 if (!txn->mt_dbxs[i].md_name.mv_size) {
10392 /* Remember this free slot */
10393 if (!unused) unused = i;
10396 if (len == txn->mt_dbxs[i].md_name.mv_size &&
10397 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
10399 return MDB_SUCCESS;
10403 /* If no free slot and max hit, fail */
10404 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
10405 return MDB_DBS_FULL;
10407 /* Cannot mix named databases with some mainDB flags */
10408 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
10409 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
10411 /* Find the DB info */
10412 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
10415 key.mv_data = (void *)name;
10416 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
10417 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
10418 if (rc == MDB_SUCCESS) {
10419 /* make sure this is actually a DB */
10420 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
10421 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
10422 return MDB_INCOMPATIBLE;
10423 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
10427 /* Done here so we cannot fail after creating a new DB */
10428 if ((namedup = strdup(name)) == NULL)
10432 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
10433 data.mv_size = sizeof(MDB_db);
10434 data.mv_data = &dummy;
10435 memset(&dummy, 0, sizeof(dummy));
10436 dummy.md_root = P_INVALID;
10437 dummy.md_flags = flags & PERSISTENT_FLAGS;
10438 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
10439 dbflag |= DB_DIRTY;
10445 /* Got info, register DBI in this txn */
10446 unsigned int slot = unused ? unused : txn->mt_numdbs;
10447 txn->mt_dbxs[slot].md_name.mv_data = namedup;
10448 txn->mt_dbxs[slot].md_name.mv_size = len;
10449 txn->mt_dbxs[slot].md_rel = NULL;
10450 txn->mt_dbflags[slot] = dbflag;
10451 /* txn-> and env-> are the same in read txns, use
10452 * tmp variable to avoid undefined assignment
10454 seq = ++txn->mt_env->me_dbiseqs[slot];
10455 txn->mt_dbiseqs[slot] = seq;
10457 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
10459 mdb_default_cmp(txn, slot);
10469 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
10471 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
10474 if (txn->mt_flags & MDB_TXN_BLOCKED)
10475 return MDB_BAD_TXN;
10477 if (txn->mt_dbflags[dbi] & DB_STALE) {
10480 /* Stale, must read the DB's root. cursor_init does it for us. */
10481 mdb_cursor_init(&mc, txn, dbi, &mx);
10483 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
10486 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
10489 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
10491 ptr = env->me_dbxs[dbi].md_name.mv_data;
10492 /* If there was no name, this was already closed */
10494 env->me_dbxs[dbi].md_name.mv_data = NULL;
10495 env->me_dbxs[dbi].md_name.mv_size = 0;
10496 env->me_dbflags[dbi] = 0;
10497 env->me_dbiseqs[dbi]++;
10502 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
10504 /* We could return the flags for the FREE_DBI too but what's the point? */
10505 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10507 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
10508 return MDB_SUCCESS;
10511 /** Add all the DB's pages to the free list.
10512 * @param[in] mc Cursor on the DB to free.
10513 * @param[in] subs non-Zero to check for sub-DBs in this DB.
10514 * @return 0 on success, non-zero on failure.
10517 mdb_drop0(MDB_cursor *mc, int subs)
10521 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
10522 if (rc == MDB_SUCCESS) {
10523 MDB_txn *txn = mc->mc_txn;
10528 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
10529 * This also avoids any P_LEAF2 pages, which have no nodes.
10530 * Also if the DB doesn't have sub-DBs and has no overflow
10531 * pages, omit scanning leaves.
10533 if ((mc->mc_flags & C_SUB) ||
10534 (!subs && !mc->mc_db->md_overflow_pages))
10535 mdb_cursor_pop(mc);
10537 mdb_cursor_copy(mc, &mx);
10539 /* bump refcount for mx's pages */
10540 for (i=0; i<mc->mc_snum; i++)
10541 mdb_page_get(&mx, mc->mc_pg[i]->mp_pgno, &mx.mc_pg[i], NULL);
10543 while (mc->mc_snum > 0) {
10544 MDB_page *mp = mc->mc_pg[mc->mc_top];
10545 unsigned n = NUMKEYS(mp);
10547 for (i=0; i<n; i++) {
10548 ni = NODEPTR(mp, i);
10549 if (ni->mn_flags & F_BIGDATA) {
10552 memcpy(&pg, NODEDATA(ni), sizeof(pg));
10553 rc = mdb_page_get(mc, pg, &omp, NULL);
10556 mdb_cassert(mc, IS_OVERFLOW(omp));
10557 rc = mdb_midl_append_range(&txn->mt_free_pgs,
10558 pg, omp->mp_pages);
10561 mc->mc_db->md_overflow_pages -= omp->mp_pages;
10562 if (!mc->mc_db->md_overflow_pages && !subs)
10564 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
10565 mdb_xcursor_init1(mc, ni);
10566 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
10571 if (!subs && !mc->mc_db->md_overflow_pages)
10574 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
10576 for (i=0; i<n; i++) {
10578 ni = NODEPTR(mp, i);
10581 mdb_midl_xappend(txn->mt_free_pgs, pg);
10586 mc->mc_ki[mc->mc_top] = i;
10587 rc = mdb_cursor_sibling(mc, 1);
10589 if (rc != MDB_NOTFOUND)
10591 /* no more siblings, go back to beginning
10592 * of previous level.
10595 mdb_cursor_pop(mc);
10597 for (i=1; i<mc->mc_snum; i++) {
10599 mc->mc_pg[i] = mx.mc_pg[i];
10604 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
10607 txn->mt_flags |= MDB_TXN_ERROR;
10609 /* drop refcount for mx's pages */
10610 mdb_cursor_unref(&mx);
10612 } else if (rc == MDB_NOTFOUND) {
10615 mc->mc_flags &= ~C_INITIALIZED;
10619 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
10621 MDB_cursor *mc, *m2;
10624 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10627 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
10630 if (TXN_DBI_CHANGED(txn, dbi))
10631 return MDB_BAD_DBI;
10633 rc = mdb_cursor_open(txn, dbi, &mc);
10637 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
10638 /* Invalidate the dropped DB's cursors */
10639 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
10640 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
10644 /* Can't delete the main DB */
10645 if (del && dbi >= CORE_DBS) {
10646 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
10648 txn->mt_dbflags[dbi] = DB_STALE;
10649 mdb_dbi_close(txn->mt_env, dbi);
10651 txn->mt_flags |= MDB_TXN_ERROR;
10654 /* reset the DB record, mark it dirty */
10655 txn->mt_dbflags[dbi] |= DB_DIRTY;
10656 txn->mt_dbs[dbi].md_depth = 0;
10657 txn->mt_dbs[dbi].md_branch_pages = 0;
10658 txn->mt_dbs[dbi].md_leaf_pages = 0;
10659 txn->mt_dbs[dbi].md_overflow_pages = 0;
10660 txn->mt_dbs[dbi].md_entries = 0;
10661 txn->mt_dbs[dbi].md_root = P_INVALID;
10663 txn->mt_flags |= MDB_TXN_DIRTY;
10666 mdb_cursor_close(mc);
10670 int mdb_set_compare(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_cmp = cmp;
10676 return MDB_SUCCESS;
10679 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10681 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10684 txn->mt_dbxs[dbi].md_dcmp = cmp;
10685 return MDB_SUCCESS;
10688 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
10690 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10693 txn->mt_dbxs[dbi].md_rel = rel;
10694 return MDB_SUCCESS;
10697 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
10699 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10702 txn->mt_dbxs[dbi].md_relctx = ctx;
10703 return MDB_SUCCESS;
10707 mdb_env_get_maxkeysize(MDB_env *env)
10709 return ENV_MAXKEY(env);
10713 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
10715 unsigned int i, rdrs;
10718 int rc = 0, first = 1;
10722 if (!env->me_txns) {
10723 return func("(no reader locks)\n", ctx);
10725 rdrs = env->me_txns->mti_numreaders;
10726 mr = env->me_txns->mti_readers;
10727 for (i=0; i<rdrs; i++) {
10728 if (mr[i].mr_pid) {
10729 txnid_t txnid = mr[i].mr_txnid;
10730 sprintf(buf, txnid == (txnid_t)-1 ?
10731 "%10d %"Z"x -\n" : "%10d %"Z"x %"Y"u\n",
10732 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10735 rc = func(" pid thread txnid\n", ctx);
10739 rc = func(buf, ctx);
10745 rc = func("(no active readers)\n", ctx);
10750 /** Insert pid into list if not already present.
10751 * return -1 if already present.
10754 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10756 /* binary search of pid in list */
10758 unsigned cursor = 1;
10760 unsigned n = ids[0];
10763 unsigned pivot = n >> 1;
10764 cursor = base + pivot + 1;
10765 val = pid - ids[cursor];
10770 } else if ( val > 0 ) {
10775 /* found, so it's a duplicate */
10784 for (n = ids[0]; n > cursor; n--)
10791 mdb_reader_check(MDB_env *env, int *dead)
10797 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10800 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
10802 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10804 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10805 unsigned int i, j, rdrs;
10807 MDB_PID_T *pids, pid;
10808 int rc = MDB_SUCCESS, count = 0;
10810 rdrs = env->me_txns->mti_numreaders;
10811 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10815 mr = env->me_txns->mti_readers;
10816 for (i=0; i<rdrs; i++) {
10817 pid = mr[i].mr_pid;
10818 if (pid && pid != env->me_pid) {
10819 if (mdb_pid_insert(pids, pid) == 0) {
10820 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10821 /* Stale reader found */
10824 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10825 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10827 rdrs = 0; /* the above checked all readers */
10829 /* Recheck, a new process may have reused pid */
10830 if (mdb_reader_pid(env, Pidcheck, pid))
10834 for (; j<rdrs; j++)
10835 if (mr[j].mr_pid == pid) {
10836 DPRINTF(("clear stale reader pid %u txn %"Y"d",
10837 (unsigned) pid, mr[j].mr_txnid));
10842 UNLOCK_MUTEX(rmutex);
10853 #ifdef MDB_ROBUST_SUPPORTED
10854 /** Handle #LOCK_MUTEX0() failure.
10855 * Try to repair the lock file if the mutex owner died.
10856 * @param[in] env the environment handle
10857 * @param[in] mutex LOCK_MUTEX0() mutex
10858 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10859 * @return 0 on success with the mutex locked, or an error code on failure.
10862 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10867 if (rc == MDB_OWNERDEAD) {
10868 /* We own the mutex. Clean up after dead previous owner. */
10870 rlocked = (mutex == env->me_rmutex);
10872 /* Keep mti_txnid updated, otherwise next writer can
10873 * overwrite data which latest meta page refers to.
10875 meta = mdb_env_pick_meta(env);
10876 env->me_txns->mti_txnid = meta->mm_txnid;
10877 /* env is hosed if the dead thread was ours */
10879 env->me_flags |= MDB_FATAL_ERROR;
10880 env->me_txn = NULL;
10884 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10885 (rc ? "this process' env is hosed" : "recovering")));
10886 rc2 = mdb_reader_check0(env, rlocked, NULL);
10888 rc2 = mdb_mutex_consistent(mutex);
10889 if (rc || (rc = rc2)) {
10890 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10891 UNLOCK_MUTEX(mutex);
10897 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10902 #endif /* MDB_ROBUST_SUPPORTED */
10905 #if defined(_WIN32)
10906 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
10910 need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
10911 if (need == 0xFFFD)
10915 result = malloc(sizeof(wchar_t) * need);
10918 MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
10924 #endif /* defined(_WIN32) */