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 2011-2015 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.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
82 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
83 /** fdatasync is broken on ext3/ext4fs on older kernels, see
84 * description in #mdb_env_open2 comments. You can safely
85 * define MDB_FDATASYNC_WORKS if this code will only be run
86 * on kernels 3.6 and newer.
88 #define BROKEN_FDATASYNC
102 typedef SSIZE_T ssize_t;
107 #if defined(__sun) || defined(ANDROID)
108 /* Most platforms have posix_memalign, older may only have memalign */
109 #define HAVE_MEMALIGN 1
113 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
114 #include <netinet/in.h>
115 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
118 #if defined(__APPLE__) || defined (BSD)
119 # if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
120 # define MDB_USE_SYSV_SEM 1
122 # define MDB_FDATASYNC fsync
123 #elif defined(ANDROID)
124 # define MDB_FDATASYNC fsync
129 #ifdef MDB_USE_POSIX_SEM
130 # define MDB_USE_HASH 1
131 #include <semaphore.h>
132 #elif defined(MDB_USE_SYSV_SEM)
135 #ifdef _SEM_SEMUN_UNDEFINED
138 struct semid_ds *buf;
139 unsigned short *array;
141 #endif /* _SEM_SEMUN_UNDEFINED */
143 #define MDB_USE_POSIX_MUTEX 1
144 #endif /* MDB_USE_POSIX_SEM */
147 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
148 + defined(MDB_USE_POSIX_MUTEX) != 1
149 # error "Ambiguous shared-lock implementation"
153 #include <valgrind/memcheck.h>
154 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
155 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
156 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
157 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
158 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
160 #define VGMEMP_CREATE(h,r,z)
161 #define VGMEMP_ALLOC(h,a,s)
162 #define VGMEMP_FREE(h,a)
163 #define VGMEMP_DESTROY(h)
164 #define VGMEMP_DEFINED(a,s)
168 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
169 /* Solaris just defines one or the other */
170 # define LITTLE_ENDIAN 1234
171 # define BIG_ENDIAN 4321
172 # ifdef _LITTLE_ENDIAN
173 # define BYTE_ORDER LITTLE_ENDIAN
175 # define BYTE_ORDER BIG_ENDIAN
178 # define BYTE_ORDER __BYTE_ORDER
182 #ifndef LITTLE_ENDIAN
183 #define LITTLE_ENDIAN __LITTLE_ENDIAN
186 #define BIG_ENDIAN __BIG_ENDIAN
189 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
190 #define MISALIGNED_OK 1
196 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
197 # error "Unknown or unsupported endianness (BYTE_ORDER)"
198 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
199 # error "Two's complement, reasonably sized integer types, please"
203 /** Put infrequently used env functions in separate section */
205 # define ESECT __attribute__ ((section("__TEXT,text_env")))
207 # define ESECT __attribute__ ((section("text_env")))
214 #define CALL_CONV WINAPI
219 /** @defgroup internal LMDB Internals
222 /** @defgroup compat Compatibility Macros
223 * A bunch of macros to minimize the amount of platform-specific ifdefs
224 * needed throughout the rest of the code. When the features this library
225 * needs are similar enough to POSIX to be hidden in a one-or-two line
226 * replacement, this macro approach is used.
230 /** Features under development */
235 /** Wrapper around __func__, which is a C99 feature */
236 #if __STDC_VERSION__ >= 199901L
237 # define mdb_func_ __func__
238 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
239 # define mdb_func_ __FUNCTION__
241 /* If a debug message says <mdb_unknown>(), update the #if statements above */
242 # define mdb_func_ "<mdb_unknown>"
245 /* Internal error codes, not exposed outside liblmdb */
246 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
248 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
249 #elif defined MDB_USE_SYSV_SEM
250 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
251 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
252 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
256 /** Some platforms define the EOWNERDEAD error code
257 * even though they don't support Robust Mutexes.
258 * Compile with -DMDB_USE_ROBUST=0, or use some other
259 * mechanism like -DMDB_USE_SYSV_SEM instead of
260 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
261 * also Robust, but some systems don't support them
264 #ifndef MDB_USE_ROBUST
265 /* Android currently lacks Robust Mutex support */
266 #if defined(ANDROID) && defined(MDB_USE_POSIX_MUTEX) && !defined(MDB_USE_ROBUST)
267 #define MDB_USE_ROBUST 0
269 #define MDB_USE_ROBUST 1
271 #endif /* MDB_USE_ROBUST */
273 #if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
274 #define MDB_ROBUST_SUPPORTED 1
278 #define MDB_USE_HASH 1
279 #define MDB_PIDLOCK 0
280 #define THREAD_RET DWORD
281 #define pthread_t HANDLE
282 #define pthread_mutex_t HANDLE
283 #define pthread_cond_t HANDLE
284 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
285 #define pthread_key_t DWORD
286 #define pthread_self() GetCurrentThreadId()
287 #define pthread_key_create(x,y) \
288 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
289 #define pthread_key_delete(x) TlsFree(x)
290 #define pthread_getspecific(x) TlsGetValue(x)
291 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
292 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
293 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
294 #define pthread_cond_signal(x) SetEvent(*x)
295 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
296 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
297 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
298 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
299 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
300 #define mdb_mutex_consistent(mutex) 0
301 #define getpid() GetCurrentProcessId()
302 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
303 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
304 #define ErrCode() GetLastError()
305 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
306 #define close(fd) (CloseHandle(fd) ? 0 : -1)
307 #define munmap(ptr,len) UnmapViewOfFile(ptr)
308 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
309 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
311 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
315 #define THREAD_RET void *
316 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
317 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
318 #define Z "z" /**< printf format modifier for size_t */
320 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
321 #define MDB_PIDLOCK 1
323 #ifdef MDB_USE_POSIX_SEM
325 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
326 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
327 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
330 mdb_sem_wait(sem_t *sem)
333 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
337 #elif defined MDB_USE_SYSV_SEM
339 typedef struct mdb_mutex {
343 } mdb_mutex_t[1], *mdb_mutexref_t;
345 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
346 #define UNLOCK_MUTEX(mutex) do { \
347 struct sembuf sb = { 0, 1, SEM_UNDO }; \
348 sb.sem_num = (mutex)->semnum; \
349 *(mutex)->locked = 0; \
350 semop((mutex)->semid, &sb, 1); \
354 mdb_sem_wait(mdb_mutexref_t sem)
356 int rc, *locked = sem->locked;
357 struct sembuf sb = { 0, -1, SEM_UNDO };
358 sb.sem_num = sem->semnum;
360 if (!semop(sem->semid, &sb, 1)) {
361 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
365 } while ((rc = errno) == EINTR);
369 #define mdb_mutex_consistent(mutex) 0
371 #else /* MDB_USE_POSIX_MUTEX: */
372 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
373 * local variables keep it (mdb_mutexref_t).
375 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
376 * be the same, or an array[size 1] and a pointer.
379 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
381 /** Lock the reader or writer mutex.
382 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
384 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
385 /** Unlock the reader or writer mutex.
387 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
388 /** Mark mutex-protected data as repaired, after death of previous owner.
390 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
391 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
393 /** Get the error code for the last failed system function.
395 #define ErrCode() errno
397 /** An abstraction for a file handle.
398 * On POSIX systems file handles are small integers. On Windows
399 * they're opaque pointers.
403 /** A value for an invalid file handle.
404 * Mainly used to initialize file variables and signify that they are
407 #define INVALID_HANDLE_VALUE (-1)
409 /** Get the size of a memory page for the system.
410 * This is the basic size that the platform's memory manager uses, and is
411 * fundamental to the use of memory-mapped files.
413 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
416 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
418 #elif defined(MDB_USE_SYSV_SEM)
419 #define MNAME_LEN (sizeof(int))
421 #define MNAME_LEN (sizeof(pthread_mutex_t))
424 #ifdef MDB_USE_SYSV_SEM
425 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
427 #define SYSV_SEM_FLAG 0
432 #ifdef MDB_ROBUST_SUPPORTED
433 /** Lock mutex, handle any error, set rc = result.
434 * Return 0 on success, nonzero (not rc) on error.
436 #define LOCK_MUTEX(rc, env, mutex) \
437 (((rc) = LOCK_MUTEX0(mutex)) && \
438 ((rc) = mdb_mutex_failed(env, mutex, rc)))
439 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
441 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
442 #define mdb_mutex_failed(env, mutex, rc) (rc)
446 /** A flag for opening a file and requesting synchronous data writes.
447 * This is only used when writing a meta page. It's not strictly needed;
448 * we could just do a normal write and then immediately perform a flush.
449 * But if this flag is available it saves us an extra system call.
451 * @note If O_DSYNC is undefined but exists in /usr/include,
452 * preferably set some compiler flag to get the definition.
453 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
456 # define MDB_DSYNC O_DSYNC
460 /** Function for flushing the data of a file. Define this to fsync
461 * if fdatasync() is not supported.
463 #ifndef MDB_FDATASYNC
464 # define MDB_FDATASYNC fdatasync
468 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
479 /** A page number in the database.
480 * Note that 64 bit page numbers are overkill, since pages themselves
481 * already represent 12-13 bits of addressable memory, and the OS will
482 * always limit applications to a maximum of 63 bits of address space.
484 * @note In the #MDB_node structure, we only store 48 bits of this value,
485 * which thus limits us to only 60 bits of addressable data.
487 typedef MDB_ID pgno_t;
489 /** A transaction ID.
490 * See struct MDB_txn.mt_txnid for details.
492 typedef MDB_ID txnid_t;
494 /** @defgroup debug Debug Macros
498 /** Enable debug output. Needs variable argument macros (a C99 feature).
499 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
500 * read from and written to the database (used for free space management).
506 static int mdb_debug;
507 static txnid_t mdb_debug_start;
509 /** Print a debug message with printf formatting.
510 * Requires double parenthesis around 2 or more args.
512 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
513 # define DPRINTF0(fmt, ...) \
514 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
516 # define DPRINTF(args) ((void) 0)
518 /** Print a debug string.
519 * The string is printed literally, with no format processing.
521 #define DPUTS(arg) DPRINTF(("%s", arg))
522 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
524 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
527 /** @brief The maximum size of a database page.
529 * It is 32k or 64k, since value-PAGEBASE must fit in
530 * #MDB_page.%mp_upper.
532 * LMDB will use database pages < OS pages if needed.
533 * That causes more I/O in write transactions: The OS must
534 * know (read) the whole page before writing a partial page.
536 * Note that we don't currently support Huge pages. On Linux,
537 * regular data files cannot use Huge pages, and in general
538 * Huge pages aren't actually pageable. We rely on the OS
539 * demand-pager to read our data and page it out when memory
540 * pressure from other processes is high. So until OSs have
541 * actual paging support for Huge pages, they're not viable.
543 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
545 /** The minimum number of keys required in a database page.
546 * Setting this to a larger value will place a smaller bound on the
547 * maximum size of a data item. Data items larger than this size will
548 * be pushed into overflow pages instead of being stored directly in
549 * the B-tree node. This value used to default to 4. With a page size
550 * of 4096 bytes that meant that any item larger than 1024 bytes would
551 * go into an overflow page. That also meant that on average 2-3KB of
552 * each overflow page was wasted space. The value cannot be lower than
553 * 2 because then there would no longer be a tree structure. With this
554 * value, items larger than 2KB will go into overflow pages, and on
555 * average only 1KB will be wasted.
557 #define MDB_MINKEYS 2
559 /** A stamp that identifies a file as an LMDB file.
560 * There's nothing special about this value other than that it is easily
561 * recognizable, and it will reflect any byte order mismatches.
563 #define MDB_MAGIC 0xBEEFC0DE
565 /** The version number for a database's datafile format. */
566 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
567 /** The version number for a database's lockfile format. */
568 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
570 /** @brief The max size of a key we can write, or 0 for computed max.
572 * This macro should normally be left alone or set to 0.
573 * Note that a database with big keys or dupsort data cannot be
574 * reliably modified by a liblmdb which uses a smaller max.
575 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
577 * Other values are allowed, for backwards compat. However:
578 * A value bigger than the computed max can break if you do not
579 * know what you are doing, and liblmdb <= 0.9.10 can break when
580 * modifying a DB with keys/dupsort data bigger than its max.
582 * Data items in an #MDB_DUPSORT database are also limited to
583 * this size, since they're actually keys of a sub-DB. Keys and
584 * #MDB_DUPSORT data items must fit on a node in a regular page.
586 #ifndef MDB_MAXKEYSIZE
587 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
590 /** The maximum size of a key we can write to the environment. */
592 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
594 #define ENV_MAXKEY(env) ((env)->me_maxkey)
597 /** @brief The maximum size of a data item.
599 * We only store a 32 bit value for node sizes.
601 #define MAXDATASIZE 0xffffffffUL
604 /** Key size which fits in a #DKBUF.
607 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
610 * This is used for printing a hex dump of a key's contents.
612 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
613 /** Display a key in hex.
615 * Invoke a function to display a key in hex.
617 #define DKEY(x) mdb_dkey(x, kbuf)
623 /** An invalid page number.
624 * Mainly used to denote an empty tree.
626 #define P_INVALID (~(pgno_t)0)
628 /** Test if the flags \b f are set in a flag word \b w. */
629 #define F_ISSET(w, f) (((w) & (f)) == (f))
631 /** Round \b n up to an even number. */
632 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
634 /** Used for offsets within a single page.
635 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
638 typedef uint16_t indx_t;
640 /** Default size of memory map.
641 * This is certainly too small for any actual applications. Apps should always set
642 * the size explicitly using #mdb_env_set_mapsize().
644 #define DEFAULT_MAPSIZE 1048576
646 /** @defgroup readers Reader Lock Table
647 * Readers don't acquire any locks for their data access. Instead, they
648 * simply record their transaction ID in the reader table. The reader
649 * mutex is needed just to find an empty slot in the reader table. The
650 * slot's address is saved in thread-specific data so that subsequent read
651 * transactions started by the same thread need no further locking to proceed.
653 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
655 * No reader table is used if the database is on a read-only filesystem, or
656 * if #MDB_NOLOCK is set.
658 * Since the database uses multi-version concurrency control, readers don't
659 * actually need any locking. This table is used to keep track of which
660 * readers are using data from which old transactions, so that we'll know
661 * when a particular old transaction is no longer in use. Old transactions
662 * that have discarded any data pages can then have those pages reclaimed
663 * for use by a later write transaction.
665 * The lock table is constructed such that reader slots are aligned with the
666 * processor's cache line size. Any slot is only ever used by one thread.
667 * This alignment guarantees that there will be no contention or cache
668 * thrashing as threads update their own slot info, and also eliminates
669 * any need for locking when accessing a slot.
671 * A writer thread will scan every slot in the table to determine the oldest
672 * outstanding reader transaction. Any freed pages older than this will be
673 * reclaimed by the writer. The writer doesn't use any locks when scanning
674 * this table. This means that there's no guarantee that the writer will
675 * see the most up-to-date reader info, but that's not required for correct
676 * operation - all we need is to know the upper bound on the oldest reader,
677 * we don't care at all about the newest reader. So the only consequence of
678 * reading stale information here is that old pages might hang around a
679 * while longer before being reclaimed. That's actually good anyway, because
680 * the longer we delay reclaiming old pages, the more likely it is that a
681 * string of contiguous pages can be found after coalescing old pages from
682 * many old transactions together.
685 /** Number of slots in the reader table.
686 * This value was chosen somewhat arbitrarily. 126 readers plus a
687 * couple mutexes fit exactly into 8KB on my development machine.
688 * Applications should set the table size using #mdb_env_set_maxreaders().
690 #define DEFAULT_READERS 126
692 /** The size of a CPU cache line in bytes. We want our lock structures
693 * aligned to this size to avoid false cache line sharing in the
695 * This value works for most CPUs. For Itanium this should be 128.
701 /** The information we store in a single slot of the reader table.
702 * In addition to a transaction ID, we also record the process and
703 * thread ID that owns a slot, so that we can detect stale information,
704 * e.g. threads or processes that went away without cleaning up.
705 * @note We currently don't check for stale records. We simply re-init
706 * the table when we know that we're the only process opening the
709 typedef struct MDB_rxbody {
710 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
711 * Multiple readers that start at the same time will probably have the
712 * same ID here. Again, it's not important to exclude them from
713 * anything; all we need to know is which version of the DB they
714 * started from so we can avoid overwriting any data used in that
715 * particular version.
717 volatile txnid_t mrb_txnid;
718 /** The process ID of the process owning this reader txn. */
719 volatile MDB_PID_T mrb_pid;
720 /** The thread ID of the thread owning this txn. */
721 volatile MDB_THR_T mrb_tid;
724 /** The actual reader record, with cacheline padding. */
725 typedef struct MDB_reader {
728 /** shorthand for mrb_txnid */
729 #define mr_txnid mru.mrx.mrb_txnid
730 #define mr_pid mru.mrx.mrb_pid
731 #define mr_tid mru.mrx.mrb_tid
732 /** cache line alignment */
733 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
737 /** The header for the reader table.
738 * The table resides in a memory-mapped file. (This is a different file
739 * than is used for the main database.)
741 * For POSIX the actual mutexes reside in the shared memory of this
742 * mapped file. On Windows, mutexes are named objects allocated by the
743 * kernel; we store the mutex names in this mapped file so that other
744 * processes can grab them. This same approach is also used on
745 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
746 * process-shared POSIX mutexes. For these cases where a named object
747 * is used, the object name is derived from a 64 bit FNV hash of the
748 * environment pathname. As such, naming collisions are extremely
749 * unlikely. If a collision occurs, the results are unpredictable.
751 typedef struct MDB_txbody {
752 /** Stamp identifying this as an LMDB file. It must be set
755 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
757 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
758 char mtb_rmname[MNAME_LEN];
759 #elif defined(MDB_USE_SYSV_SEM)
763 /** Mutex protecting access to this table.
764 * This is the reader table lock used with LOCK_MUTEX().
766 mdb_mutex_t mtb_rmutex;
768 /** The ID of the last transaction committed to the database.
769 * This is recorded here only for convenience; the value can always
770 * be determined by reading the main database meta pages.
772 volatile txnid_t mtb_txnid;
773 /** The number of slots that have been used in the reader table.
774 * This always records the maximum count, it is not decremented
775 * when readers release their slots.
777 volatile unsigned mtb_numreaders;
780 /** The actual reader table definition. */
781 typedef struct MDB_txninfo {
784 #define mti_magic mt1.mtb.mtb_magic
785 #define mti_format mt1.mtb.mtb_format
786 #define mti_rmutex mt1.mtb.mtb_rmutex
787 #define mti_rmname mt1.mtb.mtb_rmname
788 #define mti_txnid mt1.mtb.mtb_txnid
789 #define mti_numreaders mt1.mtb.mtb_numreaders
790 #ifdef MDB_USE_SYSV_SEM
791 #define mti_semid mt1.mtb.mtb_semid
792 #define mti_rlocked mt1.mtb.mtb_rlocked
794 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
797 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
798 char mt2_wmname[MNAME_LEN];
799 #define mti_wmname mt2.mt2_wmname
800 #elif defined MDB_USE_SYSV_SEM
802 #define mti_wlocked mt2.mt2_wlocked
804 mdb_mutex_t mt2_wmutex;
805 #define mti_wmutex mt2.mt2_wmutex
807 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
809 MDB_reader mti_readers[1];
812 /** Lockfile format signature: version, features and field layout */
813 #define MDB_LOCK_FORMAT \
815 ((MDB_LOCK_VERSION) \
816 /* Flags which describe functionality */ \
817 + (SYSV_SEM_FLAG << 18) \
818 + (((MDB_PIDLOCK) != 0) << 16)))
821 /** Common header for all page types.
822 * Overflow records occupy a number of contiguous pages with no
823 * headers on any page after the first.
825 typedef struct MDB_page {
826 #define mp_pgno mp_p.p_pgno
827 #define mp_next mp_p.p_next
829 pgno_t p_pgno; /**< page number */
830 struct MDB_page *p_next; /**< for in-memory list of freed pages */
833 /** @defgroup mdb_page Page Flags
835 * Flags for the page headers.
838 #define P_BRANCH 0x01 /**< branch page */
839 #define P_LEAF 0x02 /**< leaf page */
840 #define P_OVERFLOW 0x04 /**< overflow page */
841 #define P_META 0x08 /**< meta page */
842 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
843 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
844 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
845 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
846 #define P_KEEP 0x8000 /**< leave this page alone during spill */
848 uint16_t mp_flags; /**< @ref mdb_page */
849 #define mp_lower mp_pb.pb.pb_lower
850 #define mp_upper mp_pb.pb.pb_upper
851 #define mp_pages mp_pb.pb_pages
854 indx_t pb_lower; /**< lower bound of free space */
855 indx_t pb_upper; /**< upper bound of free space */
857 uint32_t pb_pages; /**< number of overflow pages */
859 indx_t mp_ptrs[1]; /**< dynamic size */
862 /** Size of the page header, excluding dynamic data at the end */
863 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
865 /** Address of first usable data byte in a page, after the header */
866 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
868 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
869 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
871 /** Number of nodes on a page */
872 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
874 /** The amount of space remaining in the page */
875 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
877 /** The percentage of space used in the page, in tenths of a percent. */
878 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
879 ((env)->me_psize - PAGEHDRSZ))
880 /** The minimum page fill factor, in tenths of a percent.
881 * Pages emptier than this are candidates for merging.
883 #define FILL_THRESHOLD 250
885 /** Test if a page is a leaf page */
886 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
887 /** Test if a page is a LEAF2 page */
888 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
889 /** Test if a page is a branch page */
890 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
891 /** Test if a page is an overflow page */
892 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
893 /** Test if a page is a sub page */
894 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
896 /** The number of overflow pages needed to store the given size. */
897 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
899 /** Link in #MDB_txn.%mt_loose_pgs list */
900 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
902 /** Header for a single key/data pair within a page.
903 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
904 * We guarantee 2-byte alignment for 'MDB_node's.
906 typedef struct MDB_node {
907 /** lo and hi are used for data size on leaf nodes and for
908 * child pgno on branch nodes. On 64 bit platforms, flags
909 * is also used for pgno. (Branch nodes have no flags).
910 * They are in host byte order in case that lets some
911 * accesses be optimized into a 32-bit word access.
913 #if BYTE_ORDER == LITTLE_ENDIAN
914 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
916 unsigned short mn_hi, mn_lo;
918 /** @defgroup mdb_node Node Flags
920 * Flags for node headers.
923 #define F_BIGDATA 0x01 /**< data put on overflow page */
924 #define F_SUBDATA 0x02 /**< data is a sub-database */
925 #define F_DUPDATA 0x04 /**< data has duplicates */
927 /** valid flags for #mdb_node_add() */
928 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
931 unsigned short mn_flags; /**< @ref mdb_node */
932 unsigned short mn_ksize; /**< key size */
933 char mn_data[1]; /**< key and data are appended here */
936 /** Size of the node header, excluding dynamic data at the end */
937 #define NODESIZE offsetof(MDB_node, mn_data)
939 /** Bit position of top word in page number, for shifting mn_flags */
940 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
942 /** Size of a node in a branch page with a given key.
943 * This is just the node header plus the key, there is no data.
945 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
947 /** Size of a node in a leaf page with a given key and data.
948 * This is node header plus key plus data size.
950 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
952 /** Address of node \b i in page \b p */
953 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
955 /** Address of the key for the node */
956 #define NODEKEY(node) (void *)((node)->mn_data)
958 /** Address of the data for a node */
959 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
961 /** Get the page number pointed to by a branch node */
962 #define NODEPGNO(node) \
963 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
964 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
965 /** Set the page number in a branch node */
966 #define SETPGNO(node,pgno) do { \
967 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
968 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
970 /** Get the size of the data in a leaf node */
971 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
972 /** Set the size of the data for a leaf node */
973 #define SETDSZ(node,size) do { \
974 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
975 /** The size of a key in a node */
976 #define NODEKSZ(node) ((node)->mn_ksize)
978 /** Copy a page number from src to dst */
980 #define COPY_PGNO(dst,src) dst = src
982 #if SIZE_MAX > 4294967295UL
983 #define COPY_PGNO(dst,src) do { \
984 unsigned short *s, *d; \
985 s = (unsigned short *)&(src); \
986 d = (unsigned short *)&(dst); \
993 #define COPY_PGNO(dst,src) do { \
994 unsigned short *s, *d; \
995 s = (unsigned short *)&(src); \
996 d = (unsigned short *)&(dst); \
1002 /** The address of a key in a LEAF2 page.
1003 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
1004 * There are no node headers, keys are stored contiguously.
1006 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
1008 /** Set the \b node's key into \b keyptr, if requested. */
1009 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
1010 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
1012 /** Set the \b node's key into \b key. */
1013 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
1015 /** Information about a single database in the environment. */
1016 typedef struct MDB_db {
1017 uint32_t md_pad; /**< also ksize for LEAF2 pages */
1018 uint16_t md_flags; /**< @ref mdb_dbi_open */
1019 uint16_t md_depth; /**< depth of this tree */
1020 pgno_t md_branch_pages; /**< number of internal pages */
1021 pgno_t md_leaf_pages; /**< number of leaf pages */
1022 pgno_t md_overflow_pages; /**< number of overflow pages */
1023 size_t md_entries; /**< number of data items */
1024 pgno_t md_root; /**< the root page of this tree */
1027 /** mdb_dbi_open flags */
1028 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1029 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1030 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1031 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1033 /** Handle for the DB used to track free pages. */
1035 /** Handle for the default DB. */
1037 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1040 /** Number of meta pages - also hardcoded elsewhere */
1043 /** Meta page content.
1044 * A meta page is the start point for accessing a database snapshot.
1045 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1047 typedef struct MDB_meta {
1048 /** Stamp identifying this as an LMDB file. It must be set
1051 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1052 uint32_t mm_version;
1053 void *mm_address; /**< address for fixed mapping */
1054 size_t mm_mapsize; /**< size of mmap region */
1055 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1056 /** The size of pages used in this DB */
1057 #define mm_psize mm_dbs[FREE_DBI].md_pad
1058 /** Any persistent environment flags. @ref mdb_env */
1059 #define mm_flags mm_dbs[FREE_DBI].md_flags
1060 pgno_t mm_last_pg; /**< last used page in file */
1061 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1064 /** Buffer for a stack-allocated meta page.
1065 * The members define size and alignment, and silence type
1066 * aliasing warnings. They are not used directly; that could
1067 * mean incorrectly using several union members in parallel.
1069 typedef union MDB_metabuf {
1072 char mm_pad[PAGEHDRSZ];
1077 /** Auxiliary DB info.
1078 * The information here is mostly static/read-only. There is
1079 * only a single copy of this record in the environment.
1081 typedef struct MDB_dbx {
1082 MDB_val md_name; /**< name of the database */
1083 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1084 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1085 MDB_rel_func *md_rel; /**< user relocate function */
1086 void *md_relctx; /**< user-provided context for md_rel */
1089 /** A database transaction.
1090 * Every operation requires a transaction handle.
1093 MDB_txn *mt_parent; /**< parent of a nested txn */
1094 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1096 pgno_t mt_next_pgno; /**< next unallocated page */
1097 /** The ID of this transaction. IDs are integers incrementing from 1.
1098 * Only committed write transactions increment the ID. If a transaction
1099 * aborts, the ID may be re-used by the next writer.
1102 MDB_env *mt_env; /**< the DB environment */
1103 /** The list of pages that became unused during this transaction.
1105 MDB_IDL mt_free_pgs;
1106 /** The list of loose pages that became unused and may be reused
1107 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1109 MDB_page *mt_loose_pgs;
1110 /* #Number of loose pages (#mt_loose_pgs) */
1112 /** The sorted list of dirty pages we temporarily wrote to disk
1113 * because the dirty list was full. page numbers in here are
1114 * shifted left by 1, deleted slots have the LSB set.
1116 MDB_IDL mt_spill_pgs;
1118 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1119 MDB_ID2L dirty_list;
1120 /** For read txns: This thread/txn's reader table slot, or NULL. */
1123 /** Array of records for each DB known in the environment. */
1125 /** Array of MDB_db records for each known DB */
1127 /** Array of sequence numbers for each DB handle */
1128 unsigned int *mt_dbiseqs;
1129 /** @defgroup mt_dbflag Transaction DB Flags
1133 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1134 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1135 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1136 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1137 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1139 /** In write txns, array of cursors for each DB */
1140 MDB_cursor **mt_cursors;
1141 /** Array of flags for each DB */
1142 unsigned char *mt_dbflags;
1143 /** Number of DB records in use, or 0 when the txn is finished.
1144 * This number only ever increments until the txn finishes; we
1145 * don't decrement it when individual DB handles are closed.
1149 /** @defgroup mdb_txn Transaction Flags
1153 /** #mdb_txn_begin() flags */
1154 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1155 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1156 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1157 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1158 /* internal txn flags */
1159 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1160 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1161 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1162 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1163 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1164 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1165 /** most operations on the txn are currently illegal */
1166 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1168 unsigned int mt_flags; /**< @ref mdb_txn */
1169 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1170 * Includes ancestor txns' dirty pages not hidden by other txns'
1171 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1172 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1174 unsigned int mt_dirty_room;
1177 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1178 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1179 * raise this on a 64 bit machine.
1181 #define CURSOR_STACK 32
1185 /** Cursors are used for all DB operations.
1186 * A cursor holds a path of (page pointer, key index) from the DB
1187 * root to a position in the DB, plus other state. #MDB_DUPSORT
1188 * cursors include an xcursor to the current data item. Write txns
1189 * track their cursors and keep them up to date when data moves.
1190 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1191 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1194 /** Next cursor on this DB in this txn */
1195 MDB_cursor *mc_next;
1196 /** Backup of the original cursor if this cursor is a shadow */
1197 MDB_cursor *mc_backup;
1198 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1199 struct MDB_xcursor *mc_xcursor;
1200 /** The transaction that owns this cursor */
1202 /** The database handle this cursor operates on */
1204 /** The database record for this cursor */
1206 /** The database auxiliary record for this cursor */
1208 /** The @ref mt_dbflag for this database */
1209 unsigned char *mc_dbflag;
1210 unsigned short mc_snum; /**< number of pushed pages */
1211 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1212 /** @defgroup mdb_cursor Cursor Flags
1214 * Cursor state flags.
1217 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1218 #define C_EOF 0x02 /**< No more data */
1219 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1220 #define C_DEL 0x08 /**< last op was a cursor_del */
1221 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1222 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1224 unsigned int mc_flags; /**< @ref mdb_cursor */
1225 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1226 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1229 /** Context for sorted-dup records.
1230 * We could have gone to a fully recursive design, with arbitrarily
1231 * deep nesting of sub-databases. But for now we only handle these
1232 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1234 typedef struct MDB_xcursor {
1235 /** A sub-cursor for traversing the Dup DB */
1236 MDB_cursor mx_cursor;
1237 /** The database record for this Dup DB */
1239 /** The auxiliary DB record for this Dup DB */
1241 /** The @ref mt_dbflag for this Dup DB */
1242 unsigned char mx_dbflag;
1245 /** State of FreeDB old pages, stored in the MDB_env */
1246 typedef struct MDB_pgstate {
1247 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1248 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1251 /** The database environment. */
1253 HANDLE me_fd; /**< The main data file */
1254 HANDLE me_lfd; /**< The lock file */
1255 HANDLE me_mfd; /**< just for writing the meta pages */
1256 /** Failed to update the meta page. Probably an I/O error. */
1257 #define MDB_FATAL_ERROR 0x80000000U
1258 /** Some fields are initialized. */
1259 #define MDB_ENV_ACTIVE 0x20000000U
1260 /** me_txkey is set */
1261 #define MDB_ENV_TXKEY 0x10000000U
1262 /** fdatasync is unreliable */
1263 #define MDB_FSYNCONLY 0x08000000U
1264 uint32_t me_flags; /**< @ref mdb_env */
1265 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1266 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1267 unsigned int me_maxreaders; /**< size of the reader table */
1268 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1269 volatile int me_close_readers;
1270 MDB_dbi me_numdbs; /**< number of DBs opened */
1271 MDB_dbi me_maxdbs; /**< size of the DB table */
1272 MDB_PID_T me_pid; /**< process ID of this env */
1273 char *me_path; /**< path to the DB files */
1274 char *me_map; /**< the memory map of the data file */
1275 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1276 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1277 void *me_pbuf; /**< scratch area for DUPSORT put() */
1278 MDB_txn *me_txn; /**< current write transaction */
1279 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1280 size_t me_mapsize; /**< size of the data memory map */
1281 off_t me_size; /**< current file size */
1282 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1283 MDB_dbx *me_dbxs; /**< array of static DB info */
1284 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1285 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1286 pthread_key_t me_txkey; /**< thread-key for readers */
1287 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1288 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1289 # define me_pglast me_pgstate.mf_pglast
1290 # define me_pghead me_pgstate.mf_pghead
1291 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1292 /** IDL of pages that became unused in a write txn */
1293 MDB_IDL me_free_pgs;
1294 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1295 MDB_ID2L me_dirty_list;
1296 /** Max number of freelist items that can fit in a single overflow page */
1298 /** Max size of a node on a page */
1299 unsigned int me_nodemax;
1300 #if !(MDB_MAXKEYSIZE)
1301 unsigned int me_maxkey; /**< max size of a key */
1303 int me_live_reader; /**< have liveness lock in reader table */
1305 int me_pidquery; /**< Used in OpenProcess */
1307 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1308 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1309 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1311 mdb_mutex_t me_rmutex;
1312 mdb_mutex_t me_wmutex;
1314 void *me_userctx; /**< User-settable context */
1315 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1318 /** Nested transaction */
1319 typedef struct MDB_ntxn {
1320 MDB_txn mnt_txn; /**< the transaction */
1321 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1324 /** max number of pages to commit in one writev() call */
1325 #define MDB_COMMIT_PAGES 64
1326 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1327 #undef MDB_COMMIT_PAGES
1328 #define MDB_COMMIT_PAGES IOV_MAX
1331 /** max bytes to write in one call */
1332 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1334 /** Check \b txn and \b dbi arguments to a function */
1335 #define TXN_DBI_EXIST(txn, dbi, validity) \
1336 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1338 /** Check for misused \b dbi handles */
1339 #define TXN_DBI_CHANGED(txn, dbi) \
1340 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1342 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1343 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1344 static int mdb_page_touch(MDB_cursor *mc);
1346 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1347 "reset-tmp", "fail-begin", "fail-beginchild"}
1349 /* mdb_txn_end operation number, for logging */
1350 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1351 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1353 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1354 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1355 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1356 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1357 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1359 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1360 static int mdb_page_search_root(MDB_cursor *mc,
1361 MDB_val *key, int modify);
1362 #define MDB_PS_MODIFY 1
1363 #define MDB_PS_ROOTONLY 2
1364 #define MDB_PS_FIRST 4
1365 #define MDB_PS_LAST 8
1366 static int mdb_page_search(MDB_cursor *mc,
1367 MDB_val *key, int flags);
1368 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1370 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1371 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1372 pgno_t newpgno, unsigned int nflags);
1374 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1375 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1376 static int mdb_env_write_meta(MDB_txn *txn);
1377 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1378 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1380 static void mdb_env_close0(MDB_env *env, int excl);
1382 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1383 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1384 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1385 static void mdb_node_del(MDB_cursor *mc, int ksize);
1386 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1387 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1388 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1389 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1390 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1392 static int mdb_rebalance(MDB_cursor *mc);
1393 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1395 static void mdb_cursor_pop(MDB_cursor *mc);
1396 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1398 static int mdb_cursor_del0(MDB_cursor *mc);
1399 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1400 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1401 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1402 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1403 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1405 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1406 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1408 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1409 static void mdb_xcursor_init0(MDB_cursor *mc);
1410 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1411 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1413 static int mdb_drop0(MDB_cursor *mc, int subs);
1414 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1415 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1418 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1421 /** Compare two items pointing at size_t's of unknown alignment. */
1422 #ifdef MISALIGNED_OK
1423 # define mdb_cmp_clong mdb_cmp_long
1425 # define mdb_cmp_clong mdb_cmp_cint
1429 static SECURITY_DESCRIPTOR mdb_null_sd;
1430 static SECURITY_ATTRIBUTES mdb_all_sa;
1431 static int mdb_sec_inited;
1434 /** Return the library version info. */
1436 mdb_version(int *major, int *minor, int *patch)
1438 if (major) *major = MDB_VERSION_MAJOR;
1439 if (minor) *minor = MDB_VERSION_MINOR;
1440 if (patch) *patch = MDB_VERSION_PATCH;
1441 return MDB_VERSION_STRING;
1444 /** Table of descriptions for LMDB @ref errors */
1445 static char *const mdb_errstr[] = {
1446 "MDB_KEYEXIST: Key/data pair already exists",
1447 "MDB_NOTFOUND: No matching key/data pair found",
1448 "MDB_PAGE_NOTFOUND: Requested page not found",
1449 "MDB_CORRUPTED: Located page was wrong type",
1450 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1451 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1452 "MDB_INVALID: File is not an LMDB file",
1453 "MDB_MAP_FULL: Environment mapsize limit reached",
1454 "MDB_DBS_FULL: Environment maxdbs limit reached",
1455 "MDB_READERS_FULL: Environment maxreaders limit reached",
1456 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1457 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1458 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1459 "MDB_PAGE_FULL: Internal error - page has no more space",
1460 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1461 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1462 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1463 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1464 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1465 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1469 mdb_strerror(int err)
1472 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1473 * This works as long as no function between the call to mdb_strerror
1474 * and the actual use of the message uses more than 4K of stack.
1477 char buf[1024], *ptr = buf;
1481 return ("Successful return: 0");
1483 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1484 i = err - MDB_KEYEXIST;
1485 return mdb_errstr[i];
1489 /* These are the C-runtime error codes we use. The comment indicates
1490 * their numeric value, and the Win32 error they would correspond to
1491 * if the error actually came from a Win32 API. A major mess, we should
1492 * have used LMDB-specific error codes for everything.
1495 case ENOENT: /* 2, FILE_NOT_FOUND */
1496 case EIO: /* 5, ACCESS_DENIED */
1497 case ENOMEM: /* 12, INVALID_ACCESS */
1498 case EACCES: /* 13, INVALID_DATA */
1499 case EBUSY: /* 16, CURRENT_DIRECTORY */
1500 case EINVAL: /* 22, BAD_COMMAND */
1501 case ENOSPC: /* 28, OUT_OF_PAPER */
1502 return strerror(err);
1507 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1508 FORMAT_MESSAGE_IGNORE_INSERTS,
1509 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1512 return strerror(err);
1516 /** assert(3) variant in cursor context */
1517 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1518 /** assert(3) variant in transaction context */
1519 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1520 /** assert(3) variant in environment context */
1521 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1524 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1525 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1528 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1529 const char *func, const char *file, int line)
1532 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1533 file, line, expr_txt, func);
1534 if (env->me_assert_func)
1535 env->me_assert_func(env, buf);
1536 fprintf(stderr, "%s\n", buf);
1540 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1544 /** Return the page number of \b mp which may be sub-page, for debug output */
1546 mdb_dbg_pgno(MDB_page *mp)
1549 COPY_PGNO(ret, mp->mp_pgno);
1553 /** Display a key in hexadecimal and return the address of the result.
1554 * @param[in] key the key to display
1555 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1556 * @return The key in hexadecimal form.
1559 mdb_dkey(MDB_val *key, char *buf)
1562 unsigned char *c = key->mv_data;
1568 if (key->mv_size > DKBUF_MAXKEYSIZE)
1569 return "MDB_MAXKEYSIZE";
1570 /* may want to make this a dynamic check: if the key is mostly
1571 * printable characters, print it as-is instead of converting to hex.
1575 for (i=0; i<key->mv_size; i++)
1576 ptr += sprintf(ptr, "%02x", *c++);
1578 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1584 mdb_leafnode_type(MDB_node *n)
1586 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1587 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1588 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1591 /** Display all the keys in the page. */
1593 mdb_page_list(MDB_page *mp)
1595 pgno_t pgno = mdb_dbg_pgno(mp);
1596 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1598 unsigned int i, nkeys, nsize, total = 0;
1602 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1603 case P_BRANCH: type = "Branch page"; break;
1604 case P_LEAF: type = "Leaf page"; break;
1605 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1606 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1607 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1609 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1610 pgno, mp->mp_pages, state);
1613 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1614 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1617 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1621 nkeys = NUMKEYS(mp);
1622 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1624 for (i=0; i<nkeys; i++) {
1625 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1626 key.mv_size = nsize = mp->mp_pad;
1627 key.mv_data = LEAF2KEY(mp, i, nsize);
1629 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1632 node = NODEPTR(mp, i);
1633 key.mv_size = node->mn_ksize;
1634 key.mv_data = node->mn_data;
1635 nsize = NODESIZE + key.mv_size;
1636 if (IS_BRANCH(mp)) {
1637 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1641 if (F_ISSET(node->mn_flags, F_BIGDATA))
1642 nsize += sizeof(pgno_t);
1644 nsize += NODEDSZ(node);
1646 nsize += sizeof(indx_t);
1647 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1648 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1650 total = EVEN(total);
1652 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1653 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1657 mdb_cursor_chk(MDB_cursor *mc)
1663 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1664 for (i=0; i<mc->mc_top; i++) {
1666 node = NODEPTR(mp, mc->mc_ki[i]);
1667 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1670 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1676 /** Count all the pages in each DB and in the freelist
1677 * and make sure it matches the actual number of pages
1679 * All named DBs must be open for a correct count.
1681 static void mdb_audit(MDB_txn *txn)
1685 MDB_ID freecount, count;
1690 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1691 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1692 freecount += *(MDB_ID *)data.mv_data;
1693 mdb_tassert(txn, rc == MDB_NOTFOUND);
1696 for (i = 0; i<txn->mt_numdbs; i++) {
1698 if (!(txn->mt_dbflags[i] & DB_VALID))
1700 mdb_cursor_init(&mc, txn, i, &mx);
1701 if (txn->mt_dbs[i].md_root == P_INVALID)
1703 count += txn->mt_dbs[i].md_branch_pages +
1704 txn->mt_dbs[i].md_leaf_pages +
1705 txn->mt_dbs[i].md_overflow_pages;
1706 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1707 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1708 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1711 mp = mc.mc_pg[mc.mc_top];
1712 for (j=0; j<NUMKEYS(mp); j++) {
1713 MDB_node *leaf = NODEPTR(mp, j);
1714 if (leaf->mn_flags & F_SUBDATA) {
1716 memcpy(&db, NODEDATA(leaf), sizeof(db));
1717 count += db.md_branch_pages + db.md_leaf_pages +
1718 db.md_overflow_pages;
1722 mdb_tassert(txn, rc == MDB_NOTFOUND);
1725 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1726 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1727 txn->mt_txnid, freecount, count+NUM_METAS,
1728 freecount+count+NUM_METAS, txn->mt_next_pgno);
1734 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1736 return txn->mt_dbxs[dbi].md_cmp(a, b);
1740 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1742 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1743 #if UINT_MAX < SIZE_MAX
1744 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1745 dcmp = mdb_cmp_clong;
1750 /** Allocate memory for a page.
1751 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1754 mdb_page_malloc(MDB_txn *txn, unsigned num)
1756 MDB_env *env = txn->mt_env;
1757 MDB_page *ret = env->me_dpages;
1758 size_t psize = env->me_psize, sz = psize, off;
1759 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1760 * For a single page alloc, we init everything after the page header.
1761 * For multi-page, we init the final page; if the caller needed that
1762 * many pages they will be filling in at least up to the last page.
1766 VGMEMP_ALLOC(env, ret, sz);
1767 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1768 env->me_dpages = ret->mp_next;
1771 psize -= off = PAGEHDRSZ;
1776 if ((ret = malloc(sz)) != NULL) {
1777 VGMEMP_ALLOC(env, ret, sz);
1778 if (!(env->me_flags & MDB_NOMEMINIT)) {
1779 memset((char *)ret + off, 0, psize);
1783 txn->mt_flags |= MDB_TXN_ERROR;
1787 /** Free a single page.
1788 * Saves single pages to a list, for future reuse.
1789 * (This is not used for multi-page overflow pages.)
1792 mdb_page_free(MDB_env *env, MDB_page *mp)
1794 mp->mp_next = env->me_dpages;
1795 VGMEMP_FREE(env, mp);
1796 env->me_dpages = mp;
1799 /** Free a dirty page */
1801 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1803 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1804 mdb_page_free(env, dp);
1806 /* large pages just get freed directly */
1807 VGMEMP_FREE(env, dp);
1812 /** Return all dirty pages to dpage list */
1814 mdb_dlist_free(MDB_txn *txn)
1816 MDB_env *env = txn->mt_env;
1817 MDB_ID2L dl = txn->mt_u.dirty_list;
1818 unsigned i, n = dl[0].mid;
1820 for (i = 1; i <= n; i++) {
1821 mdb_dpage_free(env, dl[i].mptr);
1826 /** Loosen or free a single page.
1827 * Saves single pages to a list for future reuse
1828 * in this same txn. It has been pulled from the freeDB
1829 * and already resides on the dirty list, but has been
1830 * deleted. Use these pages first before pulling again
1833 * If the page wasn't dirtied in this txn, just add it
1834 * to this txn's free list.
1837 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1840 pgno_t pgno = mp->mp_pgno;
1841 MDB_txn *txn = mc->mc_txn;
1843 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1844 if (txn->mt_parent) {
1845 MDB_ID2 *dl = txn->mt_u.dirty_list;
1846 /* If txn has a parent, make sure the page is in our
1850 unsigned x = mdb_mid2l_search(dl, pgno);
1851 if (x <= dl[0].mid && dl[x].mid == pgno) {
1852 if (mp != dl[x].mptr) { /* bad cursor? */
1853 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1854 txn->mt_flags |= MDB_TXN_ERROR;
1855 return MDB_CORRUPTED;
1862 /* no parent txn, so it's just ours */
1867 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1869 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1870 txn->mt_loose_pgs = mp;
1871 txn->mt_loose_count++;
1872 mp->mp_flags |= P_LOOSE;
1874 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1882 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1883 * @param[in] mc A cursor handle for the current operation.
1884 * @param[in] pflags Flags of the pages to update:
1885 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1886 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1887 * @return 0 on success, non-zero on failure.
1890 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1892 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1893 MDB_txn *txn = mc->mc_txn;
1899 int rc = MDB_SUCCESS, level;
1901 /* Mark pages seen by cursors */
1902 if (mc->mc_flags & C_UNTRACK)
1903 mc = NULL; /* will find mc in mt_cursors */
1904 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1905 for (; mc; mc=mc->mc_next) {
1906 if (!(mc->mc_flags & C_INITIALIZED))
1908 for (m3 = mc;; m3 = &mx->mx_cursor) {
1910 for (j=0; j<m3->mc_snum; j++) {
1912 if ((mp->mp_flags & Mask) == pflags)
1913 mp->mp_flags ^= P_KEEP;
1915 mx = m3->mc_xcursor;
1916 /* Proceed to mx if it is at a sub-database */
1917 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1919 if (! (mp && (mp->mp_flags & P_LEAF)))
1921 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1922 if (!(leaf->mn_flags & F_SUBDATA))
1931 /* Mark dirty root pages */
1932 for (i=0; i<txn->mt_numdbs; i++) {
1933 if (txn->mt_dbflags[i] & DB_DIRTY) {
1934 pgno_t pgno = txn->mt_dbs[i].md_root;
1935 if (pgno == P_INVALID)
1937 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1939 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1940 dp->mp_flags ^= P_KEEP;
1948 static int mdb_page_flush(MDB_txn *txn, int keep);
1950 /** Spill pages from the dirty list back to disk.
1951 * This is intended to prevent running into #MDB_TXN_FULL situations,
1952 * but note that they may still occur in a few cases:
1953 * 1) our estimate of the txn size could be too small. Currently this
1954 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1955 * 2) child txns may run out of space if their parents dirtied a
1956 * lot of pages and never spilled them. TODO: we probably should do
1957 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1958 * the parent's dirty_room is below a given threshold.
1960 * Otherwise, if not using nested txns, it is expected that apps will
1961 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1962 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1963 * If the txn never references them again, they can be left alone.
1964 * If the txn only reads them, they can be used without any fuss.
1965 * If the txn writes them again, they can be dirtied immediately without
1966 * going thru all of the work of #mdb_page_touch(). Such references are
1967 * handled by #mdb_page_unspill().
1969 * Also note, we never spill DB root pages, nor pages of active cursors,
1970 * because we'll need these back again soon anyway. And in nested txns,
1971 * we can't spill a page in a child txn if it was already spilled in a
1972 * parent txn. That would alter the parent txns' data even though
1973 * the child hasn't committed yet, and we'd have no way to undo it if
1974 * the child aborted.
1976 * @param[in] m0 cursor A cursor handle identifying the transaction and
1977 * database for which we are checking space.
1978 * @param[in] key For a put operation, the key being stored.
1979 * @param[in] data For a put operation, the data being stored.
1980 * @return 0 on success, non-zero on failure.
1983 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1985 MDB_txn *txn = m0->mc_txn;
1987 MDB_ID2L dl = txn->mt_u.dirty_list;
1988 unsigned int i, j, need;
1991 if (m0->mc_flags & C_SUB)
1994 /* Estimate how much space this op will take */
1995 i = m0->mc_db->md_depth;
1996 /* Named DBs also dirty the main DB */
1997 if (m0->mc_dbi >= CORE_DBS)
1998 i += txn->mt_dbs[MAIN_DBI].md_depth;
1999 /* For puts, roughly factor in the key+data size */
2001 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2002 i += i; /* double it for good measure */
2005 if (txn->mt_dirty_room > i)
2008 if (!txn->mt_spill_pgs) {
2009 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2010 if (!txn->mt_spill_pgs)
2013 /* purge deleted slots */
2014 MDB_IDL sl = txn->mt_spill_pgs;
2015 unsigned int num = sl[0];
2017 for (i=1; i<=num; i++) {
2024 /* Preserve pages which may soon be dirtied again */
2025 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2028 /* Less aggressive spill - we originally spilled the entire dirty list,
2029 * with a few exceptions for cursor pages and DB root pages. But this
2030 * turns out to be a lot of wasted effort because in a large txn many
2031 * of those pages will need to be used again. So now we spill only 1/8th
2032 * of the dirty pages. Testing revealed this to be a good tradeoff,
2033 * better than 1/2, 1/4, or 1/10.
2035 if (need < MDB_IDL_UM_MAX / 8)
2036 need = MDB_IDL_UM_MAX / 8;
2038 /* Save the page IDs of all the pages we're flushing */
2039 /* flush from the tail forward, this saves a lot of shifting later on. */
2040 for (i=dl[0].mid; i && need; i--) {
2041 MDB_ID pn = dl[i].mid << 1;
2043 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2045 /* Can't spill twice, make sure it's not already in a parent's
2048 if (txn->mt_parent) {
2050 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2051 if (tx2->mt_spill_pgs) {
2052 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2053 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2054 dp->mp_flags |= P_KEEP;
2062 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2066 mdb_midl_sort(txn->mt_spill_pgs);
2068 /* Flush the spilled part of dirty list */
2069 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2072 /* Reset any dirty pages we kept that page_flush didn't see */
2073 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2076 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2080 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2082 mdb_find_oldest(MDB_txn *txn)
2085 txnid_t mr, oldest = txn->mt_txnid - 1;
2086 if (txn->mt_env->me_txns) {
2087 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2088 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2099 /** Add a page to the txn's dirty list */
2101 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2104 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2106 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2107 insert = mdb_mid2l_append;
2109 insert = mdb_mid2l_insert;
2111 mid.mid = mp->mp_pgno;
2113 rc = insert(txn->mt_u.dirty_list, &mid);
2114 mdb_tassert(txn, rc == 0);
2115 txn->mt_dirty_room--;
2118 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2119 * me_pghead and mt_next_pgno.
2121 * If there are free pages available from older transactions, they
2122 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2123 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2124 * and move me_pglast to say which records were consumed. Only this
2125 * function can create me_pghead and move me_pglast/mt_next_pgno.
2126 * @param[in] mc cursor A cursor handle identifying the transaction and
2127 * database for which we are allocating.
2128 * @param[in] num the number of pages to allocate.
2129 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2130 * will always be satisfied by a single contiguous chunk of memory.
2131 * @return 0 on success, non-zero on failure.
2134 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2136 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2137 /* Get at most <Max_retries> more freeDB records once me_pghead
2138 * has enough pages. If not enough, use new pages from the map.
2139 * If <Paranoid> and mc is updating the freeDB, only get new
2140 * records if me_pghead is empty. Then the freelist cannot play
2141 * catch-up with itself by growing while trying to save it.
2143 enum { Paranoid = 1, Max_retries = 500 };
2145 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2147 int rc, retry = num * 60;
2148 MDB_txn *txn = mc->mc_txn;
2149 MDB_env *env = txn->mt_env;
2150 pgno_t pgno, *mop = env->me_pghead;
2151 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2153 txnid_t oldest = 0, last;
2158 /* If there are any loose pages, just use them */
2159 if (num == 1 && txn->mt_loose_pgs) {
2160 np = txn->mt_loose_pgs;
2161 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2162 txn->mt_loose_count--;
2163 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2171 /* If our dirty list is already full, we can't do anything */
2172 if (txn->mt_dirty_room == 0) {
2177 for (op = MDB_FIRST;; op = MDB_NEXT) {
2182 /* Seek a big enough contiguous page range. Prefer
2183 * pages at the tail, just truncating the list.
2189 if (mop[i-n2] == pgno+n2)
2196 if (op == MDB_FIRST) { /* 1st iteration */
2197 /* Prepare to fetch more and coalesce */
2198 last = env->me_pglast;
2199 oldest = env->me_pgoldest;
2200 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2203 key.mv_data = &last; /* will look up last+1 */
2204 key.mv_size = sizeof(last);
2206 if (Paranoid && mc->mc_dbi == FREE_DBI)
2209 if (Paranoid && retry < 0 && mop_len)
2213 /* Do not fetch more if the record will be too recent */
2214 if (oldest <= last) {
2216 oldest = mdb_find_oldest(txn);
2217 env->me_pgoldest = oldest;
2223 rc = mdb_cursor_get(&m2, &key, NULL, op);
2225 if (rc == MDB_NOTFOUND)
2229 last = *(txnid_t*)key.mv_data;
2230 if (oldest <= last) {
2232 oldest = mdb_find_oldest(txn);
2233 env->me_pgoldest = oldest;
2239 np = m2.mc_pg[m2.mc_top];
2240 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2241 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2244 idl = (MDB_ID *) data.mv_data;
2247 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2252 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2254 mop = env->me_pghead;
2256 env->me_pglast = last;
2258 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2259 last, txn->mt_dbs[FREE_DBI].md_root, i));
2261 DPRINTF(("IDL %"Z"u", idl[j]));
2263 /* Merge in descending sorted order */
2264 mdb_midl_xmerge(mop, idl);
2268 /* Use new pages from the map when nothing suitable in the freeDB */
2270 pgno = txn->mt_next_pgno;
2271 if (pgno + num >= env->me_maxpg) {
2272 DPUTS("DB size maxed out");
2278 if (env->me_flags & MDB_WRITEMAP) {
2279 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2281 if (!(np = mdb_page_malloc(txn, num))) {
2287 mop[0] = mop_len -= num;
2288 /* Move any stragglers down */
2289 for (j = i-num; j < mop_len; )
2290 mop[++j] = mop[++i];
2292 txn->mt_next_pgno = pgno + num;
2295 mdb_page_dirty(txn, np);
2301 txn->mt_flags |= MDB_TXN_ERROR;
2305 /** Copy the used portions of a non-overflow page.
2306 * @param[in] dst page to copy into
2307 * @param[in] src page to copy from
2308 * @param[in] psize size of a page
2311 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2313 enum { Align = sizeof(pgno_t) };
2314 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2316 /* If page isn't full, just copy the used portion. Adjust
2317 * alignment so memcpy may copy words instead of bytes.
2319 if ((unused &= -Align) && !IS_LEAF2(src)) {
2320 upper = (upper + PAGEBASE) & -Align;
2321 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2322 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2325 memcpy(dst, src, psize - unused);
2329 /** Pull a page off the txn's spill list, if present.
2330 * If a page being referenced was spilled to disk in this txn, bring
2331 * it back and make it dirty/writable again.
2332 * @param[in] txn the transaction handle.
2333 * @param[in] mp the page being referenced. It must not be dirty.
2334 * @param[out] ret the writable page, if any. ret is unchanged if
2335 * mp wasn't spilled.
2338 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2340 MDB_env *env = txn->mt_env;
2343 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2345 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2346 if (!tx2->mt_spill_pgs)
2348 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2349 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2352 if (txn->mt_dirty_room == 0)
2353 return MDB_TXN_FULL;
2354 if (IS_OVERFLOW(mp))
2358 if (env->me_flags & MDB_WRITEMAP) {
2361 np = mdb_page_malloc(txn, num);
2365 memcpy(np, mp, num * env->me_psize);
2367 mdb_page_copy(np, mp, env->me_psize);
2370 /* If in current txn, this page is no longer spilled.
2371 * If it happens to be the last page, truncate the spill list.
2372 * Otherwise mark it as deleted by setting the LSB.
2374 if (x == txn->mt_spill_pgs[0])
2375 txn->mt_spill_pgs[0]--;
2377 txn->mt_spill_pgs[x] |= 1;
2378 } /* otherwise, if belonging to a parent txn, the
2379 * page remains spilled until child commits
2382 mdb_page_dirty(txn, np);
2383 np->mp_flags |= P_DIRTY;
2391 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2392 * @param[in] mc cursor pointing to the page to be touched
2393 * @return 0 on success, non-zero on failure.
2396 mdb_page_touch(MDB_cursor *mc)
2398 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2399 MDB_txn *txn = mc->mc_txn;
2400 MDB_cursor *m2, *m3;
2404 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2405 if (txn->mt_flags & MDB_TXN_SPILLS) {
2407 rc = mdb_page_unspill(txn, mp, &np);
2413 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2414 (rc = mdb_page_alloc(mc, 1, &np)))
2417 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2418 mp->mp_pgno, pgno));
2419 mdb_cassert(mc, mp->mp_pgno != pgno);
2420 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2421 /* Update the parent page, if any, to point to the new page */
2423 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2424 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2425 SETPGNO(node, pgno);
2427 mc->mc_db->md_root = pgno;
2429 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2430 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2432 /* If txn has a parent, make sure the page is in our
2436 unsigned x = mdb_mid2l_search(dl, pgno);
2437 if (x <= dl[0].mid && dl[x].mid == pgno) {
2438 if (mp != dl[x].mptr) { /* bad cursor? */
2439 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2440 txn->mt_flags |= MDB_TXN_ERROR;
2441 return MDB_CORRUPTED;
2446 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2448 np = mdb_page_malloc(txn, 1);
2453 rc = mdb_mid2l_insert(dl, &mid);
2454 mdb_cassert(mc, rc == 0);
2459 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2461 np->mp_flags |= P_DIRTY;
2464 /* Adjust cursors pointing to mp */
2465 mc->mc_pg[mc->mc_top] = np;
2466 m2 = txn->mt_cursors[mc->mc_dbi];
2467 if (mc->mc_flags & C_SUB) {
2468 for (; m2; m2=m2->mc_next) {
2469 m3 = &m2->mc_xcursor->mx_cursor;
2470 if (m3->mc_snum < mc->mc_snum) continue;
2471 if (m3->mc_pg[mc->mc_top] == mp)
2472 m3->mc_pg[mc->mc_top] = np;
2475 for (; m2; m2=m2->mc_next) {
2476 if (m2->mc_snum < mc->mc_snum) continue;
2477 if (m2->mc_pg[mc->mc_top] == mp) {
2478 m2->mc_pg[mc->mc_top] = np;
2479 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2481 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2483 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2484 if (!(leaf->mn_flags & F_SUBDATA))
2485 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2493 txn->mt_flags |= MDB_TXN_ERROR;
2498 mdb_env_sync(MDB_env *env, int force)
2501 if (env->me_flags & MDB_RDONLY)
2503 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2504 if (env->me_flags & MDB_WRITEMAP) {
2505 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2506 ? MS_ASYNC : MS_SYNC;
2507 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2510 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2514 #ifdef BROKEN_FDATASYNC
2515 if (env->me_flags & MDB_FSYNCONLY) {
2516 if (fsync(env->me_fd))
2520 if (MDB_FDATASYNC(env->me_fd))
2527 /** Back up parent txn's cursors, then grab the originals for tracking */
2529 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2531 MDB_cursor *mc, *bk;
2536 for (i = src->mt_numdbs; --i >= 0; ) {
2537 if ((mc = src->mt_cursors[i]) != NULL) {
2538 size = sizeof(MDB_cursor);
2540 size += sizeof(MDB_xcursor);
2541 for (; mc; mc = bk->mc_next) {
2547 mc->mc_db = &dst->mt_dbs[i];
2548 /* Kill pointers into src - and dst to reduce abuse: The
2549 * user may not use mc until dst ends. Otherwise we'd...
2551 mc->mc_txn = NULL; /* ...set this to dst */
2552 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2553 if ((mx = mc->mc_xcursor) != NULL) {
2554 *(MDB_xcursor *)(bk+1) = *mx;
2555 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2557 mc->mc_next = dst->mt_cursors[i];
2558 dst->mt_cursors[i] = mc;
2565 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2566 * @param[in] txn the transaction handle.
2567 * @param[in] merge true to keep changes to parent cursors, false to revert.
2568 * @return 0 on success, non-zero on failure.
2571 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2573 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2577 for (i = txn->mt_numdbs; --i >= 0; ) {
2578 for (mc = cursors[i]; mc; mc = next) {
2580 if ((bk = mc->mc_backup) != NULL) {
2582 /* Commit changes to parent txn */
2583 mc->mc_next = bk->mc_next;
2584 mc->mc_backup = bk->mc_backup;
2585 mc->mc_txn = bk->mc_txn;
2586 mc->mc_db = bk->mc_db;
2587 mc->mc_dbflag = bk->mc_dbflag;
2588 if ((mx = mc->mc_xcursor) != NULL)
2589 mx->mx_cursor.mc_txn = bk->mc_txn;
2591 /* Abort nested txn */
2593 if ((mx = mc->mc_xcursor) != NULL)
2594 *mx = *(MDB_xcursor *)(bk+1);
2598 /* Only malloced cursors are permanently tracked. */
2605 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2611 Pidset = F_SETLK, Pidcheck = F_GETLK
2615 /** Set or check a pid lock. Set returns 0 on success.
2616 * Check returns 0 if the process is certainly dead, nonzero if it may
2617 * be alive (the lock exists or an error happened so we do not know).
2619 * On Windows Pidset is a no-op, we merely check for the existence
2620 * of the process with the given pid. On POSIX we use a single byte
2621 * lock on the lockfile, set at an offset equal to the pid.
2624 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2626 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2629 if (op == Pidcheck) {
2630 h = OpenProcess(env->me_pidquery, FALSE, pid);
2631 /* No documented "no such process" code, but other program use this: */
2633 return ErrCode() != ERROR_INVALID_PARAMETER;
2634 /* A process exists until all handles to it close. Has it exited? */
2635 ret = WaitForSingleObject(h, 0) != 0;
2642 struct flock lock_info;
2643 memset(&lock_info, 0, sizeof(lock_info));
2644 lock_info.l_type = F_WRLCK;
2645 lock_info.l_whence = SEEK_SET;
2646 lock_info.l_start = pid;
2647 lock_info.l_len = 1;
2648 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2649 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2651 } else if ((rc = ErrCode()) == EINTR) {
2659 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2660 * @param[in] txn the transaction handle to initialize
2661 * @return 0 on success, non-zero on failure.
2664 mdb_txn_renew0(MDB_txn *txn)
2666 MDB_env *env = txn->mt_env;
2667 MDB_txninfo *ti = env->me_txns;
2669 unsigned int i, nr, flags = txn->mt_flags;
2671 int rc, new_notls = 0;
2673 if ((flags &= MDB_TXN_RDONLY) != 0) {
2675 meta = mdb_env_pick_meta(env);
2676 txn->mt_txnid = meta->mm_txnid;
2677 txn->mt_u.reader = NULL;
2679 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2680 pthread_getspecific(env->me_txkey);
2682 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2683 return MDB_BAD_RSLOT;
2685 MDB_PID_T pid = env->me_pid;
2686 MDB_THR_T tid = pthread_self();
2687 mdb_mutexref_t rmutex = env->me_rmutex;
2689 if (!env->me_live_reader) {
2690 rc = mdb_reader_pid(env, Pidset, pid);
2693 env->me_live_reader = 1;
2696 if (LOCK_MUTEX(rc, env, rmutex))
2698 nr = ti->mti_numreaders;
2699 for (i=0; i<nr; i++)
2700 if (ti->mti_readers[i].mr_pid == 0)
2702 if (i == env->me_maxreaders) {
2703 UNLOCK_MUTEX(rmutex);
2704 return MDB_READERS_FULL;
2706 r = &ti->mti_readers[i];
2707 /* Claim the reader slot, carefully since other code
2708 * uses the reader table un-mutexed: First reset the
2709 * slot, next publish it in mti_numreaders. After
2710 * that, it is safe for mdb_env_close() to touch it.
2711 * When it will be closed, we can finally claim it.
2714 r->mr_txnid = (txnid_t)-1;
2717 ti->mti_numreaders = ++nr;
2718 env->me_close_readers = nr;
2720 UNLOCK_MUTEX(rmutex);
2722 new_notls = (env->me_flags & MDB_NOTLS);
2723 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2728 do /* LY: Retry on a race, ITS#7970. */
2729 r->mr_txnid = ti->mti_txnid;
2730 while(r->mr_txnid != ti->mti_txnid);
2731 txn->mt_txnid = r->mr_txnid;
2732 txn->mt_u.reader = r;
2733 meta = env->me_metas[txn->mt_txnid & 1];
2737 /* Not yet touching txn == env->me_txn0, it may be active */
2739 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2741 txn->mt_txnid = ti->mti_txnid;
2742 meta = env->me_metas[txn->mt_txnid & 1];
2744 meta = mdb_env_pick_meta(env);
2745 txn->mt_txnid = meta->mm_txnid;
2749 if (txn->mt_txnid == mdb_debug_start)
2752 txn->mt_child = NULL;
2753 txn->mt_loose_pgs = NULL;
2754 txn->mt_loose_count = 0;
2755 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2756 txn->mt_u.dirty_list = env->me_dirty_list;
2757 txn->mt_u.dirty_list[0].mid = 0;
2758 txn->mt_free_pgs = env->me_free_pgs;
2759 txn->mt_free_pgs[0] = 0;
2760 txn->mt_spill_pgs = NULL;
2762 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2765 /* Copy the DB info and flags */
2766 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2768 /* Moved to here to avoid a data race in read TXNs */
2769 txn->mt_next_pgno = meta->mm_last_pg+1;
2771 txn->mt_flags = flags;
2774 txn->mt_numdbs = env->me_numdbs;
2775 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2776 x = env->me_dbflags[i];
2777 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2778 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2780 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2781 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2783 if (env->me_flags & MDB_FATAL_ERROR) {
2784 DPUTS("environment had fatal error, must shutdown!");
2786 } else if (env->me_maxpg < txn->mt_next_pgno) {
2787 rc = MDB_MAP_RESIZED;
2791 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2796 mdb_txn_renew(MDB_txn *txn)
2800 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2803 rc = mdb_txn_renew0(txn);
2804 if (rc == MDB_SUCCESS) {
2805 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2806 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2807 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2813 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2817 int rc, size, tsize;
2819 flags &= MDB_TXN_BEGIN_FLAGS;
2820 flags |= env->me_flags & MDB_WRITEMAP;
2822 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2826 /* Nested transactions: Max 1 child, write txns only, no writemap */
2827 flags |= parent->mt_flags;
2828 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2829 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2831 /* Child txns save MDB_pgstate and use own copy of cursors */
2832 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2833 size += tsize = sizeof(MDB_ntxn);
2834 } else if (flags & MDB_RDONLY) {
2835 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2836 size += tsize = sizeof(MDB_txn);
2838 /* Reuse preallocated write txn. However, do not touch it until
2839 * mdb_txn_renew0() succeeds, since it currently may be active.
2844 if ((txn = calloc(1, size)) == NULL) {
2845 DPRINTF(("calloc: %s", strerror(errno)));
2848 txn->mt_dbxs = env->me_dbxs; /* static */
2849 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2850 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2851 txn->mt_flags = flags;
2856 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2857 txn->mt_dbiseqs = parent->mt_dbiseqs;
2858 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2859 if (!txn->mt_u.dirty_list ||
2860 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2862 free(txn->mt_u.dirty_list);
2866 txn->mt_txnid = parent->mt_txnid;
2867 txn->mt_dirty_room = parent->mt_dirty_room;
2868 txn->mt_u.dirty_list[0].mid = 0;
2869 txn->mt_spill_pgs = NULL;
2870 txn->mt_next_pgno = parent->mt_next_pgno;
2871 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2872 parent->mt_child = txn;
2873 txn->mt_parent = parent;
2874 txn->mt_numdbs = parent->mt_numdbs;
2875 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2876 /* Copy parent's mt_dbflags, but clear DB_NEW */
2877 for (i=0; i<txn->mt_numdbs; i++)
2878 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2880 ntxn = (MDB_ntxn *)txn;
2881 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2882 if (env->me_pghead) {
2883 size = MDB_IDL_SIZEOF(env->me_pghead);
2884 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2886 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2891 rc = mdb_cursor_shadow(parent, txn);
2893 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2894 } else { /* MDB_RDONLY */
2895 txn->mt_dbiseqs = env->me_dbiseqs;
2897 rc = mdb_txn_renew0(txn);
2900 if (txn != env->me_txn0)
2903 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2905 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2906 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2907 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2914 mdb_txn_env(MDB_txn *txn)
2916 if(!txn) return NULL;
2921 mdb_txn_id(MDB_txn *txn)
2924 return txn->mt_txnid;
2927 /** Export or close DBI handles opened in this txn. */
2929 mdb_dbis_update(MDB_txn *txn, int keep)
2932 MDB_dbi n = txn->mt_numdbs;
2933 MDB_env *env = txn->mt_env;
2934 unsigned char *tdbflags = txn->mt_dbflags;
2936 for (i = n; --i >= CORE_DBS;) {
2937 if (tdbflags[i] & DB_NEW) {
2939 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2941 char *ptr = env->me_dbxs[i].md_name.mv_data;
2943 env->me_dbxs[i].md_name.mv_data = NULL;
2944 env->me_dbxs[i].md_name.mv_size = 0;
2945 env->me_dbflags[i] = 0;
2946 env->me_dbiseqs[i]++;
2952 if (keep && env->me_numdbs < n)
2956 /** End a transaction, except successful commit of a nested transaction.
2957 * May be called twice for readonly txns: First reset it, then abort.
2958 * @param[in] txn the transaction handle to end
2959 * @param[in] mode why and how to end the transaction
2962 mdb_txn_end(MDB_txn *txn, unsigned mode)
2964 MDB_env *env = txn->mt_env;
2966 static const char *const names[] = MDB_END_NAMES;
2969 /* Export or close DBI handles opened in this txn */
2970 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2972 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2973 names[mode & MDB_END_OPMASK],
2974 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2975 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2977 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2978 if (txn->mt_u.reader) {
2979 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2980 if (!(env->me_flags & MDB_NOTLS)) {
2981 txn->mt_u.reader = NULL; /* txn does not own reader */
2982 } else if (mode & MDB_END_SLOT) {
2983 txn->mt_u.reader->mr_pid = 0;
2984 txn->mt_u.reader = NULL;
2985 } /* else txn owns the slot until it does MDB_END_SLOT */
2987 txn->mt_numdbs = 0; /* prevent further DBI activity */
2988 txn->mt_flags |= MDB_TXN_FINISHED;
2990 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2991 pgno_t *pghead = env->me_pghead;
2993 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2994 mdb_cursors_close(txn, 0);
2995 if (!(env->me_flags & MDB_WRITEMAP)) {
2996 mdb_dlist_free(txn);
3000 txn->mt_flags = MDB_TXN_FINISHED;
3002 if (!txn->mt_parent) {
3003 mdb_midl_shrink(&txn->mt_free_pgs);
3004 env->me_free_pgs = txn->mt_free_pgs;
3006 env->me_pghead = NULL;
3010 mode = 0; /* txn == env->me_txn0, do not free() it */
3012 /* The writer mutex was locked in mdb_txn_begin. */
3014 UNLOCK_MUTEX(env->me_wmutex);
3016 txn->mt_parent->mt_child = NULL;
3017 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3018 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3019 mdb_midl_free(txn->mt_free_pgs);
3020 mdb_midl_free(txn->mt_spill_pgs);
3021 free(txn->mt_u.dirty_list);
3024 mdb_midl_free(pghead);
3027 if (mode & MDB_END_FREE)
3032 mdb_txn_reset(MDB_txn *txn)
3037 /* This call is only valid for read-only txns */
3038 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3041 mdb_txn_end(txn, MDB_END_RESET);
3045 mdb_txn_abort(MDB_txn *txn)
3051 mdb_txn_abort(txn->mt_child);
3053 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3056 /** Save the freelist as of this transaction to the freeDB.
3057 * This changes the freelist. Keep trying until it stabilizes.
3060 mdb_freelist_save(MDB_txn *txn)
3062 /* env->me_pghead[] can grow and shrink during this call.
3063 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3064 * Page numbers cannot disappear from txn->mt_free_pgs[].
3067 MDB_env *env = txn->mt_env;
3068 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3069 txnid_t pglast = 0, head_id = 0;
3070 pgno_t freecnt = 0, *free_pgs, *mop;
3071 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3073 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3075 if (env->me_pghead) {
3076 /* Make sure first page of freeDB is touched and on freelist */
3077 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3078 if (rc && rc != MDB_NOTFOUND)
3082 if (!env->me_pghead && txn->mt_loose_pgs) {
3083 /* Put loose page numbers in mt_free_pgs, since
3084 * we may be unable to return them to me_pghead.
3086 MDB_page *mp = txn->mt_loose_pgs;
3087 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3089 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3090 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3091 txn->mt_loose_pgs = NULL;
3092 txn->mt_loose_count = 0;
3095 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3096 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3097 ? SSIZE_MAX : maxfree_1pg;
3100 /* Come back here after each Put() in case freelist changed */
3105 /* If using records from freeDB which we have not yet
3106 * deleted, delete them and any we reserved for me_pghead.
3108 while (pglast < env->me_pglast) {
3109 rc = mdb_cursor_first(&mc, &key, NULL);
3112 pglast = head_id = *(txnid_t *)key.mv_data;
3113 total_room = head_room = 0;
3114 mdb_tassert(txn, pglast <= env->me_pglast);
3115 rc = mdb_cursor_del(&mc, 0);
3120 /* Save the IDL of pages freed by this txn, to a single record */
3121 if (freecnt < txn->mt_free_pgs[0]) {
3123 /* Make sure last page of freeDB is touched and on freelist */
3124 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3125 if (rc && rc != MDB_NOTFOUND)
3128 free_pgs = txn->mt_free_pgs;
3129 /* Write to last page of freeDB */
3130 key.mv_size = sizeof(txn->mt_txnid);
3131 key.mv_data = &txn->mt_txnid;
3133 freecnt = free_pgs[0];
3134 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3135 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3138 /* Retry if mt_free_pgs[] grew during the Put() */
3139 free_pgs = txn->mt_free_pgs;
3140 } while (freecnt < free_pgs[0]);
3141 mdb_midl_sort(free_pgs);
3142 memcpy(data.mv_data, free_pgs, data.mv_size);
3145 unsigned int i = free_pgs[0];
3146 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3147 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3149 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3155 mop = env->me_pghead;
3156 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3158 /* Reserve records for me_pghead[]. Split it if multi-page,
3159 * to avoid searching freeDB for a page range. Use keys in
3160 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3162 if (total_room >= mop_len) {
3163 if (total_room == mop_len || --more < 0)
3165 } else if (head_room >= maxfree_1pg && head_id > 1) {
3166 /* Keep current record (overflow page), add a new one */
3170 /* (Re)write {key = head_id, IDL length = head_room} */
3171 total_room -= head_room;
3172 head_room = mop_len - total_room;
3173 if (head_room > maxfree_1pg && head_id > 1) {
3174 /* Overflow multi-page for part of me_pghead */
3175 head_room /= head_id; /* amortize page sizes */
3176 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3177 } else if (head_room < 0) {
3178 /* Rare case, not bothering to delete this record */
3181 key.mv_size = sizeof(head_id);
3182 key.mv_data = &head_id;
3183 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3184 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3187 /* IDL is initially empty, zero out at least the length */
3188 pgs = (pgno_t *)data.mv_data;
3189 j = head_room > clean_limit ? head_room : 0;
3193 total_room += head_room;
3196 /* Return loose page numbers to me_pghead, though usually none are
3197 * left at this point. The pages themselves remain in dirty_list.
3199 if (txn->mt_loose_pgs) {
3200 MDB_page *mp = txn->mt_loose_pgs;
3201 unsigned count = txn->mt_loose_count;
3203 /* Room for loose pages + temp IDL with same */
3204 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3206 mop = env->me_pghead;
3207 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3208 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3209 loose[ ++count ] = mp->mp_pgno;
3211 mdb_midl_sort(loose);
3212 mdb_midl_xmerge(mop, loose);
3213 txn->mt_loose_pgs = NULL;
3214 txn->mt_loose_count = 0;
3218 /* Fill in the reserved me_pghead records */
3224 rc = mdb_cursor_first(&mc, &key, &data);
3225 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3226 txnid_t id = *(txnid_t *)key.mv_data;
3227 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3230 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3232 if (len > mop_len) {
3234 data.mv_size = (len + 1) * sizeof(MDB_ID);
3236 data.mv_data = mop -= len;
3239 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3241 if (rc || !(mop_len -= len))
3248 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3249 * @param[in] txn the transaction that's being committed
3250 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3251 * @return 0 on success, non-zero on failure.
3254 mdb_page_flush(MDB_txn *txn, int keep)
3256 MDB_env *env = txn->mt_env;
3257 MDB_ID2L dl = txn->mt_u.dirty_list;
3258 unsigned psize = env->me_psize, j;
3259 int i, pagecount = dl[0].mid, rc;
3260 size_t size = 0, pos = 0;
3262 MDB_page *dp = NULL;
3266 struct iovec iov[MDB_COMMIT_PAGES];
3267 ssize_t wpos = 0, wsize = 0, wres;
3268 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3274 if (env->me_flags & MDB_WRITEMAP) {
3275 /* Clear dirty flags */
3276 while (++i <= pagecount) {
3278 /* Don't flush this page yet */
3279 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3280 dp->mp_flags &= ~P_KEEP;
3284 dp->mp_flags &= ~P_DIRTY;
3289 /* Write the pages */
3291 if (++i <= pagecount) {
3293 /* Don't flush this page yet */
3294 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3295 dp->mp_flags &= ~P_KEEP;
3300 /* clear dirty flag */
3301 dp->mp_flags &= ~P_DIRTY;
3304 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3309 /* Windows actually supports scatter/gather I/O, but only on
3310 * unbuffered file handles. Since we're relying on the OS page
3311 * cache for all our data, that's self-defeating. So we just
3312 * write pages one at a time. We use the ov structure to set
3313 * the write offset, to at least save the overhead of a Seek
3316 DPRINTF(("committing page %"Z"u", pgno));
3317 memset(&ov, 0, sizeof(ov));
3318 ov.Offset = pos & 0xffffffff;
3319 ov.OffsetHigh = pos >> 16 >> 16;
3320 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3322 DPRINTF(("WriteFile: %d", rc));
3326 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3327 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3330 /* Write previous page(s) */
3331 #ifdef MDB_USE_PWRITEV
3332 wres = pwritev(env->me_fd, iov, n, wpos);
3335 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3338 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3342 DPRINTF(("lseek: %s", strerror(rc)));
3345 wres = writev(env->me_fd, iov, n);
3348 if (wres != wsize) {
3353 DPRINTF(("Write error: %s", strerror(rc)));
3355 rc = EIO; /* TODO: Use which error code? */
3356 DPUTS("short write, filesystem full?");
3367 DPRINTF(("committing page %"Z"u", pgno));
3368 next_pos = pos + size;
3369 iov[n].iov_len = size;
3370 iov[n].iov_base = (char *)dp;
3376 /* MIPS has cache coherency issues, this is a no-op everywhere else
3377 * Note: for any size >= on-chip cache size, entire on-chip cache is
3380 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3382 for (i = keep; ++i <= pagecount; ) {
3384 /* This is a page we skipped above */
3387 dl[j].mid = dp->mp_pgno;
3390 mdb_dpage_free(env, dp);
3395 txn->mt_dirty_room += i - j;
3401 mdb_txn_commit(MDB_txn *txn)
3404 unsigned int i, end_mode;
3410 /* mdb_txn_end() mode for a commit which writes nothing */
3411 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3413 if (txn->mt_child) {
3414 rc = mdb_txn_commit(txn->mt_child);
3421 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3425 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3426 DPUTS("txn has failed/finished, can't commit");
3428 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3433 if (txn->mt_parent) {
3434 MDB_txn *parent = txn->mt_parent;
3438 unsigned x, y, len, ps_len;
3440 /* Append our free list to parent's */
3441 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3444 mdb_midl_free(txn->mt_free_pgs);
3445 /* Failures after this must either undo the changes
3446 * to the parent or set MDB_TXN_ERROR in the parent.
3449 parent->mt_next_pgno = txn->mt_next_pgno;
3450 parent->mt_flags = txn->mt_flags;
3452 /* Merge our cursors into parent's and close them */
3453 mdb_cursors_close(txn, 1);
3455 /* Update parent's DB table. */
3456 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3457 parent->mt_numdbs = txn->mt_numdbs;
3458 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3459 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3460 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3461 /* preserve parent's DB_NEW status */
3462 x = parent->mt_dbflags[i] & DB_NEW;
3463 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3466 dst = parent->mt_u.dirty_list;
3467 src = txn->mt_u.dirty_list;
3468 /* Remove anything in our dirty list from parent's spill list */
3469 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3471 pspill[0] = (pgno_t)-1;
3472 /* Mark our dirty pages as deleted in parent spill list */
3473 for (i=0, len=src[0].mid; ++i <= len; ) {
3474 MDB_ID pn = src[i].mid << 1;
3475 while (pn > pspill[x])
3477 if (pn == pspill[x]) {
3482 /* Squash deleted pagenums if we deleted any */
3483 for (x=y; ++x <= ps_len; )
3484 if (!(pspill[x] & 1))
3485 pspill[++y] = pspill[x];
3489 /* Find len = length of merging our dirty list with parent's */
3491 dst[0].mid = 0; /* simplify loops */
3492 if (parent->mt_parent) {
3493 len = x + src[0].mid;
3494 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3495 for (i = x; y && i; y--) {
3496 pgno_t yp = src[y].mid;
3497 while (yp < dst[i].mid)
3499 if (yp == dst[i].mid) {
3504 } else { /* Simplify the above for single-ancestor case */
3505 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3507 /* Merge our dirty list with parent's */
3509 for (i = len; y; dst[i--] = src[y--]) {
3510 pgno_t yp = src[y].mid;
3511 while (yp < dst[x].mid)
3512 dst[i--] = dst[x--];
3513 if (yp == dst[x].mid)
3514 free(dst[x--].mptr);
3516 mdb_tassert(txn, i == x);
3518 free(txn->mt_u.dirty_list);
3519 parent->mt_dirty_room = txn->mt_dirty_room;
3520 if (txn->mt_spill_pgs) {
3521 if (parent->mt_spill_pgs) {
3522 /* TODO: Prevent failure here, so parent does not fail */
3523 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3525 parent->mt_flags |= MDB_TXN_ERROR;
3526 mdb_midl_free(txn->mt_spill_pgs);
3527 mdb_midl_sort(parent->mt_spill_pgs);
3529 parent->mt_spill_pgs = txn->mt_spill_pgs;
3533 /* Append our loose page list to parent's */
3534 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3536 *lp = txn->mt_loose_pgs;
3537 parent->mt_loose_count += txn->mt_loose_count;
3539 parent->mt_child = NULL;
3540 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3545 if (txn != env->me_txn) {
3546 DPUTS("attempt to commit unknown transaction");
3551 mdb_cursors_close(txn, 0);
3553 if (!txn->mt_u.dirty_list[0].mid &&
3554 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3557 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3558 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3560 /* Update DB root pointers */
3561 if (txn->mt_numdbs > CORE_DBS) {
3565 data.mv_size = sizeof(MDB_db);
3567 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3568 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3569 if (txn->mt_dbflags[i] & DB_DIRTY) {
3570 if (TXN_DBI_CHANGED(txn, i)) {
3574 data.mv_data = &txn->mt_dbs[i];
3575 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3583 rc = mdb_freelist_save(txn);
3587 mdb_midl_free(env->me_pghead);
3588 env->me_pghead = NULL;
3589 mdb_midl_shrink(&txn->mt_free_pgs);
3595 if ((rc = mdb_page_flush(txn, 0)))
3597 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3598 (rc = mdb_env_sync(env, 0)))
3600 if ((rc = mdb_env_write_meta(txn)))
3602 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3605 mdb_txn_end(txn, end_mode);
3613 /** Read the environment parameters of a DB environment before
3614 * mapping it into memory.
3615 * @param[in] env the environment handle
3616 * @param[out] meta address of where to store the meta information
3617 * @return 0 on success, non-zero on failure.
3620 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3626 enum { Size = sizeof(pbuf) };
3628 /* We don't know the page size yet, so use a minimum value.
3629 * Read both meta pages so we can use the latest one.
3632 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3636 memset(&ov, 0, sizeof(ov));
3638 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3639 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3642 rc = pread(env->me_fd, &pbuf, Size, off);
3645 if (rc == 0 && off == 0)
3647 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3648 DPRINTF(("read: %s", mdb_strerror(rc)));
3652 p = (MDB_page *)&pbuf;
3654 if (!F_ISSET(p->mp_flags, P_META)) {
3655 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3660 if (m->mm_magic != MDB_MAGIC) {
3661 DPUTS("meta has invalid magic");
3665 if (m->mm_version != MDB_DATA_VERSION) {
3666 DPRINTF(("database is version %u, expected version %u",
3667 m->mm_version, MDB_DATA_VERSION));
3668 return MDB_VERSION_MISMATCH;
3671 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3677 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3679 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3681 meta->mm_magic = MDB_MAGIC;
3682 meta->mm_version = MDB_DATA_VERSION;
3683 meta->mm_mapsize = env->me_mapsize;
3684 meta->mm_psize = env->me_psize;
3685 meta->mm_last_pg = NUM_METAS-1;
3686 meta->mm_flags = env->me_flags & 0xffff;
3687 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3688 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3689 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3692 /** Write the environment parameters of a freshly created DB environment.
3693 * @param[in] env the environment handle
3694 * @param[in] meta the #MDB_meta to write
3695 * @return 0 on success, non-zero on failure.
3698 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3706 memset(&ov, 0, sizeof(ov));
3707 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3709 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3712 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3713 len = pwrite(fd, ptr, size, pos); \
3714 if (len == -1 && ErrCode() == EINTR) continue; \
3715 rc = (len >= 0); break; } while(1)
3718 DPUTS("writing new meta page");
3720 psize = env->me_psize;
3722 p = calloc(NUM_METAS, psize);
3726 p->mp_flags = P_META;
3727 *(MDB_meta *)METADATA(p) = *meta;
3729 q = (MDB_page *)((char *)p + psize);
3731 q->mp_flags = P_META;
3732 *(MDB_meta *)METADATA(q) = *meta;
3734 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3737 else if ((unsigned) len == psize * NUM_METAS)
3745 /** Update the environment info to commit a transaction.
3746 * @param[in] txn the transaction that's being committed
3747 * @return 0 on success, non-zero on failure.
3750 mdb_env_write_meta(MDB_txn *txn)
3753 MDB_meta meta, metab, *mp;
3757 int rc, len, toggle;
3766 toggle = txn->mt_txnid & 1;
3767 DPRINTF(("writing meta page %d for root page %"Z"u",
3768 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3771 flags = txn->mt_flags | env->me_flags;
3772 mp = env->me_metas[toggle];
3773 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3774 /* Persist any increases of mapsize config */
3775 if (mapsize < env->me_mapsize)
3776 mapsize = env->me_mapsize;
3778 if (flags & MDB_WRITEMAP) {
3779 mp->mm_mapsize = mapsize;
3780 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3781 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3782 mp->mm_last_pg = txn->mt_next_pgno - 1;
3783 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3784 !(defined(__i386__) || defined(__x86_64__))
3785 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3786 __sync_synchronize();
3788 mp->mm_txnid = txn->mt_txnid;
3789 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3790 unsigned meta_size = env->me_psize;
3791 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3792 ptr = (char *)mp - PAGEHDRSZ;
3793 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3794 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3798 if (MDB_MSYNC(ptr, meta_size, rc)) {
3805 metab.mm_txnid = mp->mm_txnid;
3806 metab.mm_last_pg = mp->mm_last_pg;
3808 meta.mm_mapsize = mapsize;
3809 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3810 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3811 meta.mm_last_pg = txn->mt_next_pgno - 1;
3812 meta.mm_txnid = txn->mt_txnid;
3814 off = offsetof(MDB_meta, mm_mapsize);
3815 ptr = (char *)&meta + off;
3816 len = sizeof(MDB_meta) - off;
3817 off += (char *)mp - env->me_map;
3819 /* Write to the SYNC fd */
3820 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3823 memset(&ov, 0, sizeof(ov));
3825 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3830 rc = pwrite(mfd, ptr, len, off);
3833 rc = rc < 0 ? ErrCode() : EIO;
3838 DPUTS("write failed, disk error?");
3839 /* On a failure, the pagecache still contains the new data.
3840 * Write some old data back, to prevent it from being used.
3841 * Use the non-SYNC fd; we know it will fail anyway.
3843 meta.mm_last_pg = metab.mm_last_pg;
3844 meta.mm_txnid = metab.mm_txnid;
3846 memset(&ov, 0, sizeof(ov));
3848 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3850 r2 = pwrite(env->me_fd, ptr, len, off);
3851 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3854 env->me_flags |= MDB_FATAL_ERROR;
3857 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3858 CACHEFLUSH(env->me_map + off, len, DCACHE);
3860 /* Memory ordering issues are irrelevant; since the entire writer
3861 * is wrapped by wmutex, all of these changes will become visible
3862 * after the wmutex is unlocked. Since the DB is multi-version,
3863 * readers will get consistent data regardless of how fresh or
3864 * how stale their view of these values is.
3867 env->me_txns->mti_txnid = txn->mt_txnid;
3872 /** Check both meta pages to see which one is newer.
3873 * @param[in] env the environment handle
3874 * @return newest #MDB_meta.
3877 mdb_env_pick_meta(const MDB_env *env)
3879 MDB_meta *const *metas = env->me_metas;
3880 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3884 mdb_env_create(MDB_env **env)
3888 e = calloc(1, sizeof(MDB_env));
3892 e->me_maxreaders = DEFAULT_READERS;
3893 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3894 e->me_fd = INVALID_HANDLE_VALUE;
3895 e->me_lfd = INVALID_HANDLE_VALUE;
3896 e->me_mfd = INVALID_HANDLE_VALUE;
3897 #ifdef MDB_USE_POSIX_SEM
3898 e->me_rmutex = SEM_FAILED;
3899 e->me_wmutex = SEM_FAILED;
3900 #elif defined MDB_USE_SYSV_SEM
3901 e->me_rmutex->semid = -1;
3902 e->me_wmutex->semid = -1;
3904 e->me_pid = getpid();
3905 GET_PAGESIZE(e->me_os_psize);
3906 VGMEMP_CREATE(e,0,0);
3912 mdb_env_map(MDB_env *env, void *addr)
3915 unsigned int flags = env->me_flags;
3919 LONG sizelo, sizehi;
3922 if (flags & MDB_RDONLY) {
3923 /* Don't set explicit map size, use whatever exists */
3928 msize = env->me_mapsize;
3929 sizelo = msize & 0xffffffff;
3930 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3932 /* Windows won't create mappings for zero length files.
3933 * and won't map more than the file size.
3934 * Just set the maxsize right now.
3936 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3937 || !SetEndOfFile(env->me_fd)
3938 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3942 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3943 PAGE_READWRITE : PAGE_READONLY,
3944 sizehi, sizelo, NULL);
3947 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3948 FILE_MAP_WRITE : FILE_MAP_READ,
3950 rc = env->me_map ? 0 : ErrCode();
3955 int prot = PROT_READ;
3956 if (flags & MDB_WRITEMAP) {
3958 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3961 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3963 if (env->me_map == MAP_FAILED) {
3968 if (flags & MDB_NORDAHEAD) {
3969 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3971 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3973 #ifdef POSIX_MADV_RANDOM
3974 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3975 #endif /* POSIX_MADV_RANDOM */
3976 #endif /* MADV_RANDOM */
3980 /* Can happen because the address argument to mmap() is just a
3981 * hint. mmap() can pick another, e.g. if the range is in use.
3982 * The MAP_FIXED flag would prevent that, but then mmap could
3983 * instead unmap existing pages to make room for the new map.
3985 if (addr && env->me_map != addr)
3986 return EBUSY; /* TODO: Make a new MDB_* error code? */
3988 p = (MDB_page *)env->me_map;
3989 env->me_metas[0] = METADATA(p);
3990 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3996 mdb_env_set_mapsize(MDB_env *env, size_t size)
3998 /* If env is already open, caller is responsible for making
3999 * sure there are no active txns.
4007 meta = mdb_env_pick_meta(env);
4009 size = meta->mm_mapsize;
4011 /* Silently round up to minimum if the size is too small */
4012 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4016 munmap(env->me_map, env->me_mapsize);
4017 env->me_mapsize = size;
4018 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4019 rc = mdb_env_map(env, old);
4023 env->me_mapsize = size;
4025 env->me_maxpg = env->me_mapsize / env->me_psize;
4030 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4034 env->me_maxdbs = dbs + CORE_DBS;
4039 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4041 if (env->me_map || readers < 1)
4043 env->me_maxreaders = readers;
4048 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4050 if (!env || !readers)
4052 *readers = env->me_maxreaders;
4057 mdb_fsize(HANDLE fd, size_t *size)
4060 LARGE_INTEGER fsize;
4062 if (!GetFileSizeEx(fd, &fsize))
4065 *size = fsize.QuadPart;
4077 #ifdef BROKEN_FDATASYNC
4078 #include <sys/utsname.h>
4079 #include <sys/vfs.h>
4082 /** Further setup required for opening an LMDB environment
4085 mdb_env_open2(MDB_env *env)
4087 unsigned int flags = env->me_flags;
4088 int i, newenv = 0, rc;
4092 /* See if we should use QueryLimited */
4094 if ((rc & 0xff) > 5)
4095 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4097 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4100 #ifdef BROKEN_FDATASYNC
4101 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4102 * https://lkml.org/lkml/2012/9/3/83
4103 * Kernels after 3.6-rc6 are known good.
4104 * https://lkml.org/lkml/2012/9/10/556
4105 * See if the DB is on ext3/ext4, then check for new enough kernel
4106 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4111 fstatfs(env->me_fd, &st);
4112 while (st.f_type == 0xEF53) {
4116 if (uts.release[0] < '3') {
4117 if (!strncmp(uts.release, "2.6.32.", 7)) {
4118 i = atoi(uts.release+7);
4120 break; /* 2.6.32.60 and newer is OK */
4121 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4122 i = atoi(uts.release+7);
4124 break; /* 2.6.34.15 and newer is OK */
4126 } else if (uts.release[0] == '3') {
4127 i = atoi(uts.release+2);
4129 break; /* 3.6 and newer is OK */
4131 i = atoi(uts.release+4);
4133 break; /* 3.5.4 and newer is OK */
4134 } else if (i == 2) {
4135 i = atoi(uts.release+4);
4137 break; /* 3.2.30 and newer is OK */
4139 } else { /* 4.x and newer is OK */
4142 env->me_flags |= MDB_FSYNCONLY;
4148 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4151 DPUTS("new mdbenv");
4153 env->me_psize = env->me_os_psize;
4154 if (env->me_psize > MAX_PAGESIZE)
4155 env->me_psize = MAX_PAGESIZE;
4156 memset(&meta, 0, sizeof(meta));
4157 mdb_env_init_meta0(env, &meta);
4158 meta.mm_mapsize = DEFAULT_MAPSIZE;
4160 env->me_psize = meta.mm_psize;
4163 /* Was a mapsize configured? */
4164 if (!env->me_mapsize) {
4165 env->me_mapsize = meta.mm_mapsize;
4168 /* Make sure mapsize >= committed data size. Even when using
4169 * mm_mapsize, which could be broken in old files (ITS#7789).
4171 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4172 if (env->me_mapsize < minsize)
4173 env->me_mapsize = minsize;
4175 meta.mm_mapsize = env->me_mapsize;
4177 if (newenv && !(flags & MDB_FIXEDMAP)) {
4178 /* mdb_env_map() may grow the datafile. Write the metapages
4179 * first, so the file will be valid if initialization fails.
4180 * Except with FIXEDMAP, since we do not yet know mm_address.
4181 * We could fill in mm_address later, but then a different
4182 * program might end up doing that - one with a memory layout
4183 * and map address which does not suit the main program.
4185 rc = mdb_env_init_meta(env, &meta);
4191 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4196 if (flags & MDB_FIXEDMAP)
4197 meta.mm_address = env->me_map;
4198 i = mdb_env_init_meta(env, &meta);
4199 if (i != MDB_SUCCESS) {
4204 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4205 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4207 #if !(MDB_MAXKEYSIZE)
4208 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4210 env->me_maxpg = env->me_mapsize / env->me_psize;
4214 MDB_meta *meta = mdb_env_pick_meta(env);
4215 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4217 DPRINTF(("opened database version %u, pagesize %u",
4218 meta->mm_version, env->me_psize));
4219 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4220 DPRINTF(("depth: %u", db->md_depth));
4221 DPRINTF(("entries: %"Z"u", db->md_entries));
4222 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4223 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4224 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4225 DPRINTF(("root: %"Z"u", db->md_root));
4233 /** Release a reader thread's slot in the reader lock table.
4234 * This function is called automatically when a thread exits.
4235 * @param[in] ptr This points to the slot in the reader lock table.
4238 mdb_env_reader_dest(void *ptr)
4240 MDB_reader *reader = ptr;
4246 /** Junk for arranging thread-specific callbacks on Windows. This is
4247 * necessarily platform and compiler-specific. Windows supports up
4248 * to 1088 keys. Let's assume nobody opens more than 64 environments
4249 * in a single process, for now. They can override this if needed.
4251 #ifndef MAX_TLS_KEYS
4252 #define MAX_TLS_KEYS 64
4254 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4255 static int mdb_tls_nkeys;
4257 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4261 case DLL_PROCESS_ATTACH: break;
4262 case DLL_THREAD_ATTACH: break;
4263 case DLL_THREAD_DETACH:
4264 for (i=0; i<mdb_tls_nkeys; i++) {
4265 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4267 mdb_env_reader_dest(r);
4271 case DLL_PROCESS_DETACH: break;
4276 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4278 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4282 /* Force some symbol references.
4283 * _tls_used forces the linker to create the TLS directory if not already done
4284 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4286 #pragma comment(linker, "/INCLUDE:_tls_used")
4287 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4288 #pragma const_seg(".CRT$XLB")
4289 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4290 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4293 #pragma comment(linker, "/INCLUDE:__tls_used")
4294 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4295 #pragma data_seg(".CRT$XLB")
4296 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4298 #endif /* WIN 32/64 */
4299 #endif /* !__GNUC__ */
4302 /** Downgrade the exclusive lock on the region back to shared */
4304 mdb_env_share_locks(MDB_env *env, int *excl)
4307 MDB_meta *meta = mdb_env_pick_meta(env);
4309 env->me_txns->mti_txnid = meta->mm_txnid;
4314 /* First acquire a shared lock. The Unlock will
4315 * then release the existing exclusive lock.
4317 memset(&ov, 0, sizeof(ov));
4318 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4321 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4327 struct flock lock_info;
4328 /* The shared lock replaces the existing lock */
4329 memset((void *)&lock_info, 0, sizeof(lock_info));
4330 lock_info.l_type = F_RDLCK;
4331 lock_info.l_whence = SEEK_SET;
4332 lock_info.l_start = 0;
4333 lock_info.l_len = 1;
4334 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4335 (rc = ErrCode()) == EINTR) ;
4336 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4343 /** Try to get exclusive lock, otherwise shared.
4344 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4347 mdb_env_excl_lock(MDB_env *env, int *excl)
4351 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4355 memset(&ov, 0, sizeof(ov));
4356 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4363 struct flock lock_info;
4364 memset((void *)&lock_info, 0, sizeof(lock_info));
4365 lock_info.l_type = F_WRLCK;
4366 lock_info.l_whence = SEEK_SET;
4367 lock_info.l_start = 0;
4368 lock_info.l_len = 1;
4369 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4370 (rc = ErrCode()) == EINTR) ;
4374 # ifndef MDB_USE_POSIX_MUTEX
4375 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4378 lock_info.l_type = F_RDLCK;
4379 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4380 (rc = ErrCode()) == EINTR) ;
4390 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4392 * @(#) $Revision: 5.1 $
4393 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4394 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4396 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4400 * Please do not copyright this code. This code is in the public domain.
4402 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4403 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4404 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4405 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4406 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4407 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4408 * PERFORMANCE OF THIS SOFTWARE.
4411 * chongo <Landon Curt Noll> /\oo/\
4412 * http://www.isthe.com/chongo/
4414 * Share and Enjoy! :-)
4417 typedef unsigned long long mdb_hash_t;
4418 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4420 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4421 * @param[in] val value to hash
4422 * @param[in] hval initial value for hash
4423 * @return 64 bit hash
4425 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4426 * hval arg on the first call.
4429 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4431 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4432 unsigned char *end = s + val->mv_size;
4434 * FNV-1a hash each octet of the string
4437 /* xor the bottom with the current octet */
4438 hval ^= (mdb_hash_t)*s++;
4440 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4441 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4442 (hval << 7) + (hval << 8) + (hval << 40);
4444 /* return our new hash value */
4448 /** Hash the string and output the encoded hash.
4449 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4450 * very short name limits. We don't care about the encoding being reversible,
4451 * we just want to preserve as many bits of the input as possible in a
4452 * small printable string.
4453 * @param[in] str string to hash
4454 * @param[out] encbuf an array of 11 chars to hold the hash
4456 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4459 mdb_pack85(unsigned long l, char *out)
4463 for (i=0; i<5; i++) {
4464 *out++ = mdb_a85[l % 85];
4470 mdb_hash_enc(MDB_val *val, char *encbuf)
4472 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4474 mdb_pack85(h, encbuf);
4475 mdb_pack85(h>>32, encbuf+5);
4480 /** Open and/or initialize the lock region for the environment.
4481 * @param[in] env The LMDB environment.
4482 * @param[in] lpath The pathname of the file used for the lock region.
4483 * @param[in] mode The Unix permissions for the file, if we create it.
4484 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4485 * @return 0 on success, non-zero on failure.
4488 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4491 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4493 # define MDB_ERRCODE_ROFS EROFS
4494 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4495 # define MDB_CLOEXEC O_CLOEXEC
4498 # define MDB_CLOEXEC 0
4501 #ifdef MDB_USE_SYSV_SEM
4509 env->me_lfd = CreateFileA(lpath, GENERIC_READ|GENERIC_WRITE,
4510 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4511 FILE_ATTRIBUTE_NORMAL, NULL);
4513 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4515 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4517 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4522 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4523 /* Lose record locks when exec*() */
4524 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4525 fcntl(env->me_lfd, F_SETFD, fdflags);
4528 if (!(env->me_flags & MDB_NOTLS)) {
4529 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4532 env->me_flags |= MDB_ENV_TXKEY;
4534 /* Windows TLS callbacks need help finding their TLS info. */
4535 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4539 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4543 /* Try to get exclusive lock. If we succeed, then
4544 * nobody is using the lock region and we should initialize it.
4546 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4549 size = GetFileSize(env->me_lfd, NULL);
4551 size = lseek(env->me_lfd, 0, SEEK_END);
4552 if (size == -1) goto fail_errno;
4554 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4555 if (size < rsize && *excl > 0) {
4557 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4558 || !SetEndOfFile(env->me_lfd))
4561 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4565 size = rsize - sizeof(MDB_txninfo);
4566 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4571 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4573 if (!mh) goto fail_errno;
4574 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4576 if (!env->me_txns) goto fail_errno;
4578 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4580 if (m == MAP_FAILED) goto fail_errno;
4586 BY_HANDLE_FILE_INFORMATION stbuf;
4595 if (!mdb_sec_inited) {
4596 InitializeSecurityDescriptor(&mdb_null_sd,
4597 SECURITY_DESCRIPTOR_REVISION);
4598 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4599 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4600 mdb_all_sa.bInheritHandle = FALSE;
4601 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4604 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4605 idbuf.volume = stbuf.dwVolumeSerialNumber;
4606 idbuf.nhigh = stbuf.nFileIndexHigh;
4607 idbuf.nlow = stbuf.nFileIndexLow;
4608 val.mv_data = &idbuf;
4609 val.mv_size = sizeof(idbuf);
4610 mdb_hash_enc(&val, encbuf);
4611 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4612 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4613 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4614 if (!env->me_rmutex) goto fail_errno;
4615 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4616 if (!env->me_wmutex) goto fail_errno;
4617 #elif defined(MDB_USE_POSIX_SEM)
4626 #if defined(__NetBSD__)
4627 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4629 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4630 idbuf.dev = stbuf.st_dev;
4631 idbuf.ino = stbuf.st_ino;
4632 val.mv_data = &idbuf;
4633 val.mv_size = sizeof(idbuf);
4634 mdb_hash_enc(&val, encbuf);
4635 #ifdef MDB_SHORT_SEMNAMES
4636 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4638 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4639 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4640 /* Clean up after a previous run, if needed: Try to
4641 * remove both semaphores before doing anything else.
4643 sem_unlink(env->me_txns->mti_rmname);
4644 sem_unlink(env->me_txns->mti_wmname);
4645 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4646 O_CREAT|O_EXCL, mode, 1);
4647 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4648 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4649 O_CREAT|O_EXCL, mode, 1);
4650 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4651 #elif defined(MDB_USE_SYSV_SEM)
4652 unsigned short vals[2] = {1, 1};
4653 key_t key = ftok(lpath, 'M');
4656 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4660 if (semctl(semid, 0, SETALL, semu) < 0)
4662 env->me_txns->mti_semid = semid;
4663 #else /* MDB_USE_POSIX_MUTEX: */
4664 pthread_mutexattr_t mattr;
4666 if ((rc = pthread_mutexattr_init(&mattr))
4667 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4668 #ifdef MDB_ROBUST_SUPPORTED
4669 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4671 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4672 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4674 pthread_mutexattr_destroy(&mattr);
4675 #endif /* _WIN32 || ... */
4677 env->me_txns->mti_magic = MDB_MAGIC;
4678 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4679 env->me_txns->mti_txnid = 0;
4680 env->me_txns->mti_numreaders = 0;
4683 #ifdef MDB_USE_SYSV_SEM
4684 struct semid_ds buf;
4686 if (env->me_txns->mti_magic != MDB_MAGIC) {
4687 DPUTS("lock region has invalid magic");
4691 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4692 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4693 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4694 rc = MDB_VERSION_MISMATCH;
4698 if (rc && rc != EACCES && rc != EAGAIN) {
4702 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4703 if (!env->me_rmutex) goto fail_errno;
4704 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4705 if (!env->me_wmutex) goto fail_errno;
4706 #elif defined(MDB_USE_POSIX_SEM)
4707 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4708 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4709 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4710 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4711 #elif defined(MDB_USE_SYSV_SEM)
4712 semid = env->me_txns->mti_semid;
4714 /* check for read access */
4715 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4717 /* check for write access */
4718 if (semctl(semid, 0, IPC_SET, semu) < 0)
4722 #ifdef MDB_USE_SYSV_SEM
4723 env->me_rmutex->semid = semid;
4724 env->me_wmutex->semid = semid;
4725 env->me_rmutex->semnum = 0;
4726 env->me_wmutex->semnum = 1;
4727 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4728 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4739 /** The name of the lock file in the DB environment */
4740 #define LOCKNAME "/lock.mdb"
4741 /** The name of the data file in the DB environment */
4742 #define DATANAME "/data.mdb"
4743 /** The suffix of the lock file when no subdir is used */
4744 #define LOCKSUFF "-lock"
4745 /** Only a subset of the @ref mdb_env flags can be changed
4746 * at runtime. Changing other flags requires closing the
4747 * environment and re-opening it with the new flags.
4749 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4750 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4751 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4753 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4754 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4758 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4760 int oflags, rc, len, excl = -1;
4761 char *lpath, *dpath;
4763 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4767 if (flags & MDB_NOSUBDIR) {
4768 rc = len + sizeof(LOCKSUFF) + len + 1;
4770 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4775 if (flags & MDB_NOSUBDIR) {
4776 dpath = lpath + len + sizeof(LOCKSUFF);
4777 sprintf(lpath, "%s" LOCKSUFF, path);
4778 strcpy(dpath, path);
4780 dpath = lpath + len + sizeof(LOCKNAME);
4781 sprintf(lpath, "%s" LOCKNAME, path);
4782 sprintf(dpath, "%s" DATANAME, path);
4786 flags |= env->me_flags;
4787 if (flags & MDB_RDONLY) {
4788 /* silently ignore WRITEMAP when we're only getting read access */
4789 flags &= ~MDB_WRITEMAP;
4791 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4792 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4795 env->me_flags = flags |= MDB_ENV_ACTIVE;
4799 env->me_path = strdup(path);
4800 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4801 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4802 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4803 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4807 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4809 /* For RDONLY, get lockfile after we know datafile exists */
4810 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4811 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4817 if (F_ISSET(flags, MDB_RDONLY)) {
4818 oflags = GENERIC_READ;
4819 len = OPEN_EXISTING;
4821 oflags = GENERIC_READ|GENERIC_WRITE;
4824 mode = FILE_ATTRIBUTE_NORMAL;
4825 env->me_fd = CreateFileA(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4826 NULL, len, mode, NULL);
4828 if (F_ISSET(flags, MDB_RDONLY))
4831 oflags = O_RDWR | O_CREAT;
4833 env->me_fd = open(dpath, oflags, mode);
4835 if (env->me_fd == INVALID_HANDLE_VALUE) {
4840 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4841 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4846 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4847 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4848 env->me_mfd = env->me_fd;
4850 /* Synchronous fd for meta writes. Needed even with
4851 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4854 len = OPEN_EXISTING;
4855 env->me_mfd = CreateFileA(dpath, oflags,
4856 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4857 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4860 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4862 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4867 DPRINTF(("opened dbenv %p", (void *) env));
4869 rc = mdb_env_share_locks(env, &excl);
4873 if (!(flags & MDB_RDONLY)) {
4875 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4876 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4877 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4878 (txn = calloc(1, size)))
4880 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4881 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4882 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4883 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4885 txn->mt_dbxs = env->me_dbxs;
4886 txn->mt_flags = MDB_TXN_FINISHED;
4896 mdb_env_close0(env, excl);
4902 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4904 mdb_env_close0(MDB_env *env, int excl)
4908 if (!(env->me_flags & MDB_ENV_ACTIVE))
4911 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4913 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4914 free(env->me_dbxs[i].md_name.mv_data);
4919 free(env->me_dbiseqs);
4920 free(env->me_dbflags);
4922 free(env->me_dirty_list);
4924 mdb_midl_free(env->me_free_pgs);
4926 if (env->me_flags & MDB_ENV_TXKEY) {
4927 pthread_key_delete(env->me_txkey);
4929 /* Delete our key from the global list */
4930 for (i=0; i<mdb_tls_nkeys; i++)
4931 if (mdb_tls_keys[i] == env->me_txkey) {
4932 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4940 munmap(env->me_map, env->me_mapsize);
4942 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4943 (void) close(env->me_mfd);
4944 if (env->me_fd != INVALID_HANDLE_VALUE)
4945 (void) close(env->me_fd);
4947 MDB_PID_T pid = env->me_pid;
4948 /* Clearing readers is done in this function because
4949 * me_txkey with its destructor must be disabled first.
4951 * We skip the the reader mutex, so we touch only
4952 * data owned by this process (me_close_readers and
4953 * our readers), and clear each reader atomically.
4955 for (i = env->me_close_readers; --i >= 0; )
4956 if (env->me_txns->mti_readers[i].mr_pid == pid)
4957 env->me_txns->mti_readers[i].mr_pid = 0;
4959 if (env->me_rmutex) {
4960 CloseHandle(env->me_rmutex);
4961 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4963 /* Windows automatically destroys the mutexes when
4964 * the last handle closes.
4966 #elif defined(MDB_USE_POSIX_SEM)
4967 if (env->me_rmutex != SEM_FAILED) {
4968 sem_close(env->me_rmutex);
4969 if (env->me_wmutex != SEM_FAILED)
4970 sem_close(env->me_wmutex);
4971 /* If we have the filelock: If we are the
4972 * only remaining user, clean up semaphores.
4975 mdb_env_excl_lock(env, &excl);
4977 sem_unlink(env->me_txns->mti_rmname);
4978 sem_unlink(env->me_txns->mti_wmname);
4981 #elif defined(MDB_USE_SYSV_SEM)
4982 if (env->me_rmutex->semid != -1) {
4983 /* If we have the filelock: If we are the
4984 * only remaining user, clean up semaphores.
4987 mdb_env_excl_lock(env, &excl);
4989 semctl(env->me_rmutex->semid, 0, IPC_RMID);
4992 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4994 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4997 /* Unlock the lockfile. Windows would have unlocked it
4998 * after closing anyway, but not necessarily at once.
5000 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5003 (void) close(env->me_lfd);
5006 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5010 mdb_env_close(MDB_env *env)
5017 VGMEMP_DESTROY(env);
5018 while ((dp = env->me_dpages) != NULL) {
5019 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5020 env->me_dpages = dp->mp_next;
5024 mdb_env_close0(env, 0);
5028 /** Compare two items pointing at aligned size_t's */
5030 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5032 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5033 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5036 /** Compare two items pointing at aligned unsigned int's.
5038 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5039 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5042 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5044 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5045 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5048 /** Compare two items pointing at unsigned ints of unknown alignment.
5049 * Nodes and keys are guaranteed to be 2-byte aligned.
5052 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5054 #if BYTE_ORDER == LITTLE_ENDIAN
5055 unsigned short *u, *c;
5058 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5059 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5062 } while(!x && u > (unsigned short *)a->mv_data);
5065 unsigned short *u, *c, *end;
5068 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5069 u = (unsigned short *)a->mv_data;
5070 c = (unsigned short *)b->mv_data;
5073 } while(!x && u < end);
5078 /** Compare two items lexically */
5080 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5087 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5093 diff = memcmp(a->mv_data, b->mv_data, len);
5094 return diff ? diff : len_diff<0 ? -1 : len_diff;
5097 /** Compare two items in reverse byte order */
5099 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5101 const unsigned char *p1, *p2, *p1_lim;
5105 p1_lim = (const unsigned char *)a->mv_data;
5106 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5107 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5109 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5115 while (p1 > p1_lim) {
5116 diff = *--p1 - *--p2;
5120 return len_diff<0 ? -1 : len_diff;
5123 /** Search for key within a page, using binary search.
5124 * Returns the smallest entry larger or equal to the key.
5125 * If exactp is non-null, stores whether the found entry was an exact match
5126 * in *exactp (1 or 0).
5127 * Updates the cursor index with the index of the found entry.
5128 * If no entry larger or equal to the key is found, returns NULL.
5131 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5133 unsigned int i = 0, nkeys;
5136 MDB_page *mp = mc->mc_pg[mc->mc_top];
5137 MDB_node *node = NULL;
5142 nkeys = NUMKEYS(mp);
5144 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5145 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5148 low = IS_LEAF(mp) ? 0 : 1;
5150 cmp = mc->mc_dbx->md_cmp;
5152 /* Branch pages have no data, so if using integer keys,
5153 * alignment is guaranteed. Use faster mdb_cmp_int.
5155 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5156 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5163 nodekey.mv_size = mc->mc_db->md_pad;
5164 node = NODEPTR(mp, 0); /* fake */
5165 while (low <= high) {
5166 i = (low + high) >> 1;
5167 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5168 rc = cmp(key, &nodekey);
5169 DPRINTF(("found leaf index %u [%s], rc = %i",
5170 i, DKEY(&nodekey), rc));
5179 while (low <= high) {
5180 i = (low + high) >> 1;
5182 node = NODEPTR(mp, i);
5183 nodekey.mv_size = NODEKSZ(node);
5184 nodekey.mv_data = NODEKEY(node);
5186 rc = cmp(key, &nodekey);
5189 DPRINTF(("found leaf index %u [%s], rc = %i",
5190 i, DKEY(&nodekey), rc));
5192 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5193 i, DKEY(&nodekey), NODEPGNO(node), rc));
5204 if (rc > 0) { /* Found entry is less than the key. */
5205 i++; /* Skip to get the smallest entry larger than key. */
5207 node = NODEPTR(mp, i);
5210 *exactp = (rc == 0 && nkeys > 0);
5211 /* store the key index */
5212 mc->mc_ki[mc->mc_top] = i;
5214 /* There is no entry larger or equal to the key. */
5217 /* nodeptr is fake for LEAF2 */
5223 mdb_cursor_adjust(MDB_cursor *mc, func)
5227 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5228 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5235 /** Pop a page off the top of the cursor's stack. */
5237 mdb_cursor_pop(MDB_cursor *mc)
5240 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5241 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5249 /** Push a page onto the top of the cursor's stack. */
5251 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5253 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5254 DDBI(mc), (void *) mc));
5256 if (mc->mc_snum >= CURSOR_STACK) {
5257 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5258 return MDB_CURSOR_FULL;
5261 mc->mc_top = mc->mc_snum++;
5262 mc->mc_pg[mc->mc_top] = mp;
5263 mc->mc_ki[mc->mc_top] = 0;
5268 /** Find the address of the page corresponding to a given page number.
5269 * @param[in] txn the transaction for this access.
5270 * @param[in] pgno the page number for the page to retrieve.
5271 * @param[out] ret address of a pointer where the page's address will be stored.
5272 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5273 * @return 0 on success, non-zero on failure.
5276 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5278 MDB_env *env = txn->mt_env;
5282 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5286 MDB_ID2L dl = tx2->mt_u.dirty_list;
5288 /* Spilled pages were dirtied in this txn and flushed
5289 * because the dirty list got full. Bring this page
5290 * back in from the map (but don't unspill it here,
5291 * leave that unless page_touch happens again).
5293 if (tx2->mt_spill_pgs) {
5294 MDB_ID pn = pgno << 1;
5295 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5296 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5297 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5302 unsigned x = mdb_mid2l_search(dl, pgno);
5303 if (x <= dl[0].mid && dl[x].mid == pgno) {
5309 } while ((tx2 = tx2->mt_parent) != NULL);
5312 if (pgno < txn->mt_next_pgno) {
5314 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5316 DPRINTF(("page %"Z"u not found", pgno));
5317 txn->mt_flags |= MDB_TXN_ERROR;
5318 return MDB_PAGE_NOTFOUND;
5328 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5329 * The cursor is at the root page, set up the rest of it.
5332 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5334 MDB_page *mp = mc->mc_pg[mc->mc_top];
5338 while (IS_BRANCH(mp)) {
5342 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5343 mdb_cassert(mc, NUMKEYS(mp) > 1);
5344 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5346 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5348 if (flags & MDB_PS_LAST)
5349 i = NUMKEYS(mp) - 1;
5352 node = mdb_node_search(mc, key, &exact);
5354 i = NUMKEYS(mp) - 1;
5356 i = mc->mc_ki[mc->mc_top];
5358 mdb_cassert(mc, i > 0);
5362 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5365 mdb_cassert(mc, i < NUMKEYS(mp));
5366 node = NODEPTR(mp, i);
5368 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5371 mc->mc_ki[mc->mc_top] = i;
5372 if ((rc = mdb_cursor_push(mc, mp)))
5375 if (flags & MDB_PS_MODIFY) {
5376 if ((rc = mdb_page_touch(mc)) != 0)
5378 mp = mc->mc_pg[mc->mc_top];
5383 DPRINTF(("internal error, index points to a %02X page!?",
5385 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5386 return MDB_CORRUPTED;
5389 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5390 key ? DKEY(key) : "null"));
5391 mc->mc_flags |= C_INITIALIZED;
5392 mc->mc_flags &= ~C_EOF;
5397 /** Search for the lowest key under the current branch page.
5398 * This just bypasses a NUMKEYS check in the current page
5399 * before calling mdb_page_search_root(), because the callers
5400 * are all in situations where the current page is known to
5404 mdb_page_search_lowest(MDB_cursor *mc)
5406 MDB_page *mp = mc->mc_pg[mc->mc_top];
5407 MDB_node *node = NODEPTR(mp, 0);
5410 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5413 mc->mc_ki[mc->mc_top] = 0;
5414 if ((rc = mdb_cursor_push(mc, mp)))
5416 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5419 /** Search for the page a given key should be in.
5420 * Push it and its parent pages on the cursor stack.
5421 * @param[in,out] mc the cursor for this operation.
5422 * @param[in] key the key to search for, or NULL for first/last page.
5423 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5424 * are touched (updated with new page numbers).
5425 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5426 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5427 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5428 * @return 0 on success, non-zero on failure.
5431 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5436 /* Make sure the txn is still viable, then find the root from
5437 * the txn's db table and set it as the root of the cursor's stack.
5439 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5440 DPUTS("transaction may not be used now");
5443 /* Make sure we're using an up-to-date root */
5444 if (*mc->mc_dbflag & DB_STALE) {
5446 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5448 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5449 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5456 MDB_node *leaf = mdb_node_search(&mc2,
5457 &mc->mc_dbx->md_name, &exact);
5459 return MDB_NOTFOUND;
5460 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5461 return MDB_INCOMPATIBLE; /* not a named DB */
5462 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5465 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5467 /* The txn may not know this DBI, or another process may
5468 * have dropped and recreated the DB with other flags.
5470 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5471 return MDB_INCOMPATIBLE;
5472 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5474 *mc->mc_dbflag &= ~DB_STALE;
5476 root = mc->mc_db->md_root;
5478 if (root == P_INVALID) { /* Tree is empty. */
5479 DPUTS("tree is empty");
5480 return MDB_NOTFOUND;
5484 mdb_cassert(mc, root > 1);
5485 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5486 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5492 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5493 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5495 if (flags & MDB_PS_MODIFY) {
5496 if ((rc = mdb_page_touch(mc)))
5500 if (flags & MDB_PS_ROOTONLY)
5503 return mdb_page_search_root(mc, key, flags);
5507 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5509 MDB_txn *txn = mc->mc_txn;
5510 pgno_t pg = mp->mp_pgno;
5511 unsigned x = 0, ovpages = mp->mp_pages;
5512 MDB_env *env = txn->mt_env;
5513 MDB_IDL sl = txn->mt_spill_pgs;
5514 MDB_ID pn = pg << 1;
5517 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5518 /* If the page is dirty or on the spill list we just acquired it,
5519 * so we should give it back to our current free list, if any.
5520 * Otherwise put it onto the list of pages we freed in this txn.
5522 * Won't create me_pghead: me_pglast must be inited along with it.
5523 * Unsupported in nested txns: They would need to hide the page
5524 * range in ancestor txns' dirty and spilled lists.
5526 if (env->me_pghead &&
5528 ((mp->mp_flags & P_DIRTY) ||
5529 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5533 MDB_ID2 *dl, ix, iy;
5534 rc = mdb_midl_need(&env->me_pghead, ovpages);
5537 if (!(mp->mp_flags & P_DIRTY)) {
5538 /* This page is no longer spilled */
5545 /* Remove from dirty list */
5546 dl = txn->mt_u.dirty_list;
5548 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5554 mdb_cassert(mc, x > 1);
5556 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5557 txn->mt_flags |= MDB_TXN_ERROR;
5558 return MDB_CORRUPTED;
5561 if (!(env->me_flags & MDB_WRITEMAP))
5562 mdb_dpage_free(env, mp);
5564 /* Insert in me_pghead */
5565 mop = env->me_pghead;
5566 j = mop[0] + ovpages;
5567 for (i = mop[0]; i && mop[i] < pg; i--)
5573 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5577 mc->mc_db->md_overflow_pages -= ovpages;
5581 /** Return the data associated with a given node.
5582 * @param[in] txn The transaction for this operation.
5583 * @param[in] leaf The node being read.
5584 * @param[out] data Updated to point to the node's data.
5585 * @return 0 on success, non-zero on failure.
5588 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5590 MDB_page *omp; /* overflow page */
5594 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5595 data->mv_size = NODEDSZ(leaf);
5596 data->mv_data = NODEDATA(leaf);
5600 /* Read overflow data.
5602 data->mv_size = NODEDSZ(leaf);
5603 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5604 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5605 DPRINTF(("read overflow page %"Z"u failed", pgno));
5608 data->mv_data = METADATA(omp);
5614 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5615 MDB_val *key, MDB_val *data)
5622 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5624 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5627 if (txn->mt_flags & MDB_TXN_BLOCKED)
5630 mdb_cursor_init(&mc, txn, dbi, &mx);
5631 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5634 /** Find a sibling for a page.
5635 * Replaces the page at the top of the cursor's stack with the
5636 * specified sibling, if one exists.
5637 * @param[in] mc The cursor for this operation.
5638 * @param[in] move_right Non-zero if the right sibling is requested,
5639 * otherwise the left sibling.
5640 * @return 0 on success, non-zero on failure.
5643 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5649 if (mc->mc_snum < 2) {
5650 return MDB_NOTFOUND; /* root has no siblings */
5654 DPRINTF(("parent page is page %"Z"u, index %u",
5655 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5657 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5658 : (mc->mc_ki[mc->mc_top] == 0)) {
5659 DPRINTF(("no more keys left, moving to %s sibling",
5660 move_right ? "right" : "left"));
5661 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5662 /* undo cursor_pop before returning */
5669 mc->mc_ki[mc->mc_top]++;
5671 mc->mc_ki[mc->mc_top]--;
5672 DPRINTF(("just moving to %s index key %u",
5673 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5675 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5677 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5678 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5679 /* mc will be inconsistent if caller does mc_snum++ as above */
5680 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5684 mdb_cursor_push(mc, mp);
5686 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5691 /** Move the cursor to the next data item. */
5693 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5699 if (mc->mc_flags & C_EOF) {
5700 return MDB_NOTFOUND;
5703 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5705 mp = mc->mc_pg[mc->mc_top];
5707 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5708 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5709 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5710 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5711 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5712 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5713 if (rc == MDB_SUCCESS)
5714 MDB_GET_KEY(leaf, key);
5719 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5720 if (op == MDB_NEXT_DUP)
5721 return MDB_NOTFOUND;
5725 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5726 mdb_dbg_pgno(mp), (void *) mc));
5727 if (mc->mc_flags & C_DEL)
5730 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5731 DPUTS("=====> move to next sibling page");
5732 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5733 mc->mc_flags |= C_EOF;
5736 mp = mc->mc_pg[mc->mc_top];
5737 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5739 mc->mc_ki[mc->mc_top]++;
5742 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5743 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5746 key->mv_size = mc->mc_db->md_pad;
5747 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5751 mdb_cassert(mc, IS_LEAF(mp));
5752 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5754 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5755 mdb_xcursor_init1(mc, leaf);
5758 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5761 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5762 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5763 if (rc != MDB_SUCCESS)
5768 MDB_GET_KEY(leaf, key);
5772 /** Move the cursor to the previous data item. */
5774 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5780 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5782 mp = mc->mc_pg[mc->mc_top];
5784 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5785 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5786 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5787 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5788 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5789 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5790 if (rc == MDB_SUCCESS) {
5791 MDB_GET_KEY(leaf, key);
5792 mc->mc_flags &= ~C_EOF;
5798 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5799 if (op == MDB_PREV_DUP)
5800 return MDB_NOTFOUND;
5804 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5805 mdb_dbg_pgno(mp), (void *) mc));
5807 if (mc->mc_ki[mc->mc_top] == 0) {
5808 DPUTS("=====> move to prev sibling page");
5809 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5812 mp = mc->mc_pg[mc->mc_top];
5813 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5814 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5816 mc->mc_ki[mc->mc_top]--;
5818 mc->mc_flags &= ~C_EOF;
5820 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5821 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5824 key->mv_size = mc->mc_db->md_pad;
5825 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5829 mdb_cassert(mc, IS_LEAF(mp));
5830 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5832 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5833 mdb_xcursor_init1(mc, leaf);
5836 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5839 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5840 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5841 if (rc != MDB_SUCCESS)
5846 MDB_GET_KEY(leaf, key);
5850 /** Set the cursor on a specific data item. */
5852 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5853 MDB_cursor_op op, int *exactp)
5857 MDB_node *leaf = NULL;
5860 if (key->mv_size == 0)
5861 return MDB_BAD_VALSIZE;
5864 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5866 /* See if we're already on the right page */
5867 if (mc->mc_flags & C_INITIALIZED) {
5870 mp = mc->mc_pg[mc->mc_top];
5872 mc->mc_ki[mc->mc_top] = 0;
5873 return MDB_NOTFOUND;
5875 if (mp->mp_flags & P_LEAF2) {
5876 nodekey.mv_size = mc->mc_db->md_pad;
5877 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5879 leaf = NODEPTR(mp, 0);
5880 MDB_GET_KEY2(leaf, nodekey);
5882 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5884 /* Probably happens rarely, but first node on the page
5885 * was the one we wanted.
5887 mc->mc_ki[mc->mc_top] = 0;
5894 unsigned int nkeys = NUMKEYS(mp);
5896 if (mp->mp_flags & P_LEAF2) {
5897 nodekey.mv_data = LEAF2KEY(mp,
5898 nkeys-1, nodekey.mv_size);
5900 leaf = NODEPTR(mp, nkeys-1);
5901 MDB_GET_KEY2(leaf, nodekey);
5903 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5905 /* last node was the one we wanted */
5906 mc->mc_ki[mc->mc_top] = nkeys-1;
5912 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5913 /* This is definitely the right page, skip search_page */
5914 if (mp->mp_flags & P_LEAF2) {
5915 nodekey.mv_data = LEAF2KEY(mp,
5916 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5918 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5919 MDB_GET_KEY2(leaf, nodekey);
5921 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5923 /* current node was the one we wanted */
5933 /* If any parents have right-sibs, search.
5934 * Otherwise, there's nothing further.
5936 for (i=0; i<mc->mc_top; i++)
5938 NUMKEYS(mc->mc_pg[i])-1)
5940 if (i == mc->mc_top) {
5941 /* There are no other pages */
5942 mc->mc_ki[mc->mc_top] = nkeys;
5943 return MDB_NOTFOUND;
5947 /* There are no other pages */
5948 mc->mc_ki[mc->mc_top] = 0;
5949 if (op == MDB_SET_RANGE && !exactp) {
5953 return MDB_NOTFOUND;
5957 rc = mdb_page_search(mc, key, 0);
5958 if (rc != MDB_SUCCESS)
5961 mp = mc->mc_pg[mc->mc_top];
5962 mdb_cassert(mc, IS_LEAF(mp));
5965 leaf = mdb_node_search(mc, key, exactp);
5966 if (exactp != NULL && !*exactp) {
5967 /* MDB_SET specified and not an exact match. */
5968 return MDB_NOTFOUND;
5972 DPUTS("===> inexact leaf not found, goto sibling");
5973 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5974 mc->mc_flags |= C_EOF;
5975 return rc; /* no entries matched */
5977 mp = mc->mc_pg[mc->mc_top];
5978 mdb_cassert(mc, IS_LEAF(mp));
5979 leaf = NODEPTR(mp, 0);
5983 mc->mc_flags |= C_INITIALIZED;
5984 mc->mc_flags &= ~C_EOF;
5987 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5988 key->mv_size = mc->mc_db->md_pad;
5989 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5994 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5995 mdb_xcursor_init1(mc, leaf);
5998 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5999 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6000 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6003 if (op == MDB_GET_BOTH) {
6009 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6010 if (rc != MDB_SUCCESS)
6013 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6016 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
6018 dcmp = mc->mc_dbx->md_dcmp;
6019 #if UINT_MAX < SIZE_MAX
6020 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6021 dcmp = mdb_cmp_clong;
6023 rc = dcmp(data, &olddata);
6025 if (op == MDB_GET_BOTH || rc > 0)
6026 return MDB_NOTFOUND;
6033 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6034 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6039 /* The key already matches in all other cases */
6040 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6041 MDB_GET_KEY(leaf, key);
6042 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6047 /** Move the cursor to the first item in the database. */
6049 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6055 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6057 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6058 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6059 if (rc != MDB_SUCCESS)
6062 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6064 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6065 mc->mc_flags |= C_INITIALIZED;
6066 mc->mc_flags &= ~C_EOF;
6068 mc->mc_ki[mc->mc_top] = 0;
6070 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6071 key->mv_size = mc->mc_db->md_pad;
6072 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6077 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6078 mdb_xcursor_init1(mc, leaf);
6079 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6083 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6087 MDB_GET_KEY(leaf, key);
6091 /** Move the cursor to the last item in the database. */
6093 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6099 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6101 if (!(mc->mc_flags & C_EOF)) {
6103 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6104 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6105 if (rc != MDB_SUCCESS)
6108 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6111 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6112 mc->mc_flags |= C_INITIALIZED|C_EOF;
6113 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6115 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6116 key->mv_size = mc->mc_db->md_pad;
6117 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6122 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6123 mdb_xcursor_init1(mc, leaf);
6124 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6128 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6133 MDB_GET_KEY(leaf, key);
6138 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6143 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6148 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6152 case MDB_GET_CURRENT:
6153 if (!(mc->mc_flags & C_INITIALIZED)) {
6156 MDB_page *mp = mc->mc_pg[mc->mc_top];
6157 int nkeys = NUMKEYS(mp);
6158 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6159 mc->mc_ki[mc->mc_top] = nkeys;
6165 key->mv_size = mc->mc_db->md_pad;
6166 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6168 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6169 MDB_GET_KEY(leaf, key);
6171 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6172 if (mc->mc_flags & C_DEL)
6173 mdb_xcursor_init1(mc, leaf);
6174 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6176 rc = mdb_node_read(mc->mc_txn, leaf, data);
6183 case MDB_GET_BOTH_RANGE:
6188 if (mc->mc_xcursor == NULL) {
6189 rc = MDB_INCOMPATIBLE;
6199 rc = mdb_cursor_set(mc, key, data, op,
6200 op == MDB_SET_RANGE ? NULL : &exact);
6203 case MDB_GET_MULTIPLE:
6204 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6208 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6209 rc = MDB_INCOMPATIBLE;
6213 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6214 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6217 case MDB_NEXT_MULTIPLE:
6222 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6223 rc = MDB_INCOMPATIBLE;
6226 if (!(mc->mc_flags & C_INITIALIZED))
6227 rc = mdb_cursor_first(mc, key, data);
6229 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6230 if (rc == MDB_SUCCESS) {
6231 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6234 mx = &mc->mc_xcursor->mx_cursor;
6235 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6237 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6238 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6246 case MDB_NEXT_NODUP:
6247 if (!(mc->mc_flags & C_INITIALIZED))
6248 rc = mdb_cursor_first(mc, key, data);
6250 rc = mdb_cursor_next(mc, key, data, op);
6254 case MDB_PREV_NODUP:
6255 if (!(mc->mc_flags & C_INITIALIZED)) {
6256 rc = mdb_cursor_last(mc, key, data);
6259 mc->mc_flags |= C_INITIALIZED;
6260 mc->mc_ki[mc->mc_top]++;
6262 rc = mdb_cursor_prev(mc, key, data, op);
6265 rc = mdb_cursor_first(mc, key, data);
6268 mfunc = mdb_cursor_first;
6270 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6274 if (mc->mc_xcursor == NULL) {
6275 rc = MDB_INCOMPATIBLE;
6279 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6280 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6281 MDB_GET_KEY(leaf, key);
6282 rc = mdb_node_read(mc->mc_txn, leaf, data);
6286 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6290 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6293 rc = mdb_cursor_last(mc, key, data);
6296 mfunc = mdb_cursor_last;
6299 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6304 if (mc->mc_flags & C_DEL)
6305 mc->mc_flags ^= C_DEL;
6310 /** Touch all the pages in the cursor stack. Set mc_top.
6311 * Makes sure all the pages are writable, before attempting a write operation.
6312 * @param[in] mc The cursor to operate on.
6315 mdb_cursor_touch(MDB_cursor *mc)
6317 int rc = MDB_SUCCESS;
6319 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6322 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6324 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6325 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6328 *mc->mc_dbflag |= DB_DIRTY;
6333 rc = mdb_page_touch(mc);
6334 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6335 mc->mc_top = mc->mc_snum-1;
6340 /** Do not spill pages to disk if txn is getting full, may fail instead */
6341 #define MDB_NOSPILL 0x8000
6344 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6348 MDB_node *leaf = NULL;
6349 MDB_page *fp, *mp, *sub_root = NULL;
6351 MDB_val xdata, *rdata, dkey, olddata;
6353 int do_sub = 0, insert_key, insert_data;
6354 unsigned int mcount = 0, dcount = 0, nospill;
6357 unsigned int nflags;
6360 if (mc == NULL || key == NULL)
6363 env = mc->mc_txn->mt_env;
6365 /* Check this first so counter will always be zero on any
6368 if (flags & MDB_MULTIPLE) {
6369 dcount = data[1].mv_size;
6370 data[1].mv_size = 0;
6371 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6372 return MDB_INCOMPATIBLE;
6375 nospill = flags & MDB_NOSPILL;
6376 flags &= ~MDB_NOSPILL;
6378 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6379 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6381 if (key->mv_size-1 >= ENV_MAXKEY(env))
6382 return MDB_BAD_VALSIZE;
6384 #if SIZE_MAX > MAXDATASIZE
6385 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6386 return MDB_BAD_VALSIZE;
6388 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6389 return MDB_BAD_VALSIZE;
6392 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6393 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6397 if (flags == MDB_CURRENT) {
6398 if (!(mc->mc_flags & C_INITIALIZED))
6401 } else if (mc->mc_db->md_root == P_INVALID) {
6402 /* new database, cursor has nothing to point to */
6405 mc->mc_flags &= ~C_INITIALIZED;
6410 if (flags & MDB_APPEND) {
6412 rc = mdb_cursor_last(mc, &k2, &d2);
6414 rc = mc->mc_dbx->md_cmp(key, &k2);
6417 mc->mc_ki[mc->mc_top]++;
6419 /* new key is <= last key */
6424 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6426 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6427 DPRINTF(("duplicate key [%s]", DKEY(key)));
6429 return MDB_KEYEXIST;
6431 if (rc && rc != MDB_NOTFOUND)
6435 if (mc->mc_flags & C_DEL)
6436 mc->mc_flags ^= C_DEL;
6438 /* Cursor is positioned, check for room in the dirty list */
6440 if (flags & MDB_MULTIPLE) {
6442 xdata.mv_size = data->mv_size * dcount;
6446 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6450 if (rc == MDB_NO_ROOT) {
6452 /* new database, write a root leaf page */
6453 DPUTS("allocating new root leaf page");
6454 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6457 mdb_cursor_push(mc, np);
6458 mc->mc_db->md_root = np->mp_pgno;
6459 mc->mc_db->md_depth++;
6460 *mc->mc_dbflag |= DB_DIRTY;
6461 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6463 np->mp_flags |= P_LEAF2;
6464 mc->mc_flags |= C_INITIALIZED;
6466 /* make sure all cursor pages are writable */
6467 rc2 = mdb_cursor_touch(mc);
6472 insert_key = insert_data = rc;
6474 /* The key does not exist */
6475 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6476 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6477 LEAFSIZE(key, data) > env->me_nodemax)
6479 /* Too big for a node, insert in sub-DB. Set up an empty
6480 * "old sub-page" for prep_subDB to expand to a full page.
6482 fp_flags = P_LEAF|P_DIRTY;
6484 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6485 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6486 olddata.mv_size = PAGEHDRSZ;
6490 /* there's only a key anyway, so this is a no-op */
6491 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6493 unsigned int ksize = mc->mc_db->md_pad;
6494 if (key->mv_size != ksize)
6495 return MDB_BAD_VALSIZE;
6496 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6497 memcpy(ptr, key->mv_data, ksize);
6499 /* if overwriting slot 0 of leaf, need to
6500 * update branch key if there is a parent page
6502 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6503 unsigned short top = mc->mc_top;
6505 /* slot 0 is always an empty key, find real slot */
6506 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6508 if (mc->mc_ki[mc->mc_top])
6509 rc2 = mdb_update_key(mc, key);
6520 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6521 olddata.mv_size = NODEDSZ(leaf);
6522 olddata.mv_data = NODEDATA(leaf);
6525 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6526 /* Prepare (sub-)page/sub-DB to accept the new item,
6527 * if needed. fp: old sub-page or a header faking
6528 * it. mp: new (sub-)page. offset: growth in page
6529 * size. xdata: node data with new page or DB.
6531 unsigned i, offset = 0;
6532 mp = fp = xdata.mv_data = env->me_pbuf;
6533 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6535 /* Was a single item before, must convert now */
6536 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6538 /* Just overwrite the current item */
6539 if (flags == MDB_CURRENT)
6541 dcmp = mc->mc_dbx->md_dcmp;
6542 #if UINT_MAX < SIZE_MAX
6543 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6544 dcmp = mdb_cmp_clong;
6546 /* does data match? */
6547 if (!dcmp(data, &olddata)) {
6548 if (flags & MDB_NODUPDATA)
6549 return MDB_KEYEXIST;
6554 /* Back up original data item */
6555 dkey.mv_size = olddata.mv_size;
6556 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6558 /* Make sub-page header for the dup items, with dummy body */
6559 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6560 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6561 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6562 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6563 fp->mp_flags |= P_LEAF2;
6564 fp->mp_pad = data->mv_size;
6565 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6567 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6568 (dkey.mv_size & 1) + (data->mv_size & 1);
6570 fp->mp_upper = xdata.mv_size - PAGEBASE;
6571 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6572 } else if (leaf->mn_flags & F_SUBDATA) {
6573 /* Data is on sub-DB, just store it */
6574 flags |= F_DUPDATA|F_SUBDATA;
6577 /* Data is on sub-page */
6578 fp = olddata.mv_data;
6581 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6582 offset = EVEN(NODESIZE + sizeof(indx_t) +
6586 offset = fp->mp_pad;
6587 if (SIZELEFT(fp) < offset) {
6588 offset *= 4; /* space for 4 more */
6591 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6593 fp->mp_flags |= P_DIRTY;
6594 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6595 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6599 xdata.mv_size = olddata.mv_size + offset;
6602 fp_flags = fp->mp_flags;
6603 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6604 /* Too big for a sub-page, convert to sub-DB */
6605 fp_flags &= ~P_SUBP;
6607 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6608 fp_flags |= P_LEAF2;
6609 dummy.md_pad = fp->mp_pad;
6610 dummy.md_flags = MDB_DUPFIXED;
6611 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6612 dummy.md_flags |= MDB_INTEGERKEY;
6618 dummy.md_branch_pages = 0;
6619 dummy.md_leaf_pages = 1;
6620 dummy.md_overflow_pages = 0;
6621 dummy.md_entries = NUMKEYS(fp);
6622 xdata.mv_size = sizeof(MDB_db);
6623 xdata.mv_data = &dummy;
6624 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6626 offset = env->me_psize - olddata.mv_size;
6627 flags |= F_DUPDATA|F_SUBDATA;
6628 dummy.md_root = mp->mp_pgno;
6632 mp->mp_flags = fp_flags | P_DIRTY;
6633 mp->mp_pad = fp->mp_pad;
6634 mp->mp_lower = fp->mp_lower;
6635 mp->mp_upper = fp->mp_upper + offset;
6636 if (fp_flags & P_LEAF2) {
6637 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6639 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6640 olddata.mv_size - fp->mp_upper - PAGEBASE);
6641 for (i=0; i<NUMKEYS(fp); i++)
6642 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6650 mdb_node_del(mc, 0);
6654 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6655 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6656 return MDB_INCOMPATIBLE;
6657 /* overflow page overwrites need special handling */
6658 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6661 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6663 memcpy(&pg, olddata.mv_data, sizeof(pg));
6664 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6666 ovpages = omp->mp_pages;
6668 /* Is the ov page large enough? */
6669 if (ovpages >= dpages) {
6670 if (!(omp->mp_flags & P_DIRTY) &&
6671 (level || (env->me_flags & MDB_WRITEMAP)))
6673 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6676 level = 0; /* dirty in this txn or clean */
6679 if (omp->mp_flags & P_DIRTY) {
6680 /* yes, overwrite it. Note in this case we don't
6681 * bother to try shrinking the page if the new data
6682 * is smaller than the overflow threshold.
6685 /* It is writable only in a parent txn */
6686 size_t sz = (size_t) env->me_psize * ovpages, off;
6687 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6693 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6694 mdb_cassert(mc, rc2 == 0);
6695 if (!(flags & MDB_RESERVE)) {
6696 /* Copy end of page, adjusting alignment so
6697 * compiler may copy words instead of bytes.
6699 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6700 memcpy((size_t *)((char *)np + off),
6701 (size_t *)((char *)omp + off), sz - off);
6704 memcpy(np, omp, sz); /* Copy beginning of page */
6707 SETDSZ(leaf, data->mv_size);
6708 if (F_ISSET(flags, MDB_RESERVE))
6709 data->mv_data = METADATA(omp);
6711 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6715 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6717 } else if (data->mv_size == olddata.mv_size) {
6718 /* same size, just replace it. Note that we could
6719 * also reuse this node if the new data is smaller,
6720 * but instead we opt to shrink the node in that case.
6722 if (F_ISSET(flags, MDB_RESERVE))
6723 data->mv_data = olddata.mv_data;
6724 else if (!(mc->mc_flags & C_SUB))
6725 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6727 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6732 mdb_node_del(mc, 0);
6738 nflags = flags & NODE_ADD_FLAGS;
6739 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6740 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6741 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6742 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6744 nflags |= MDB_SPLIT_REPLACE;
6745 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6747 /* There is room already in this leaf page. */
6748 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6749 if (rc == 0 && insert_key) {
6750 /* Adjust other cursors pointing to mp */
6751 MDB_cursor *m2, *m3;
6752 MDB_dbi dbi = mc->mc_dbi;
6753 unsigned i = mc->mc_top;
6754 MDB_page *mp = mc->mc_pg[i];
6756 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6757 if (mc->mc_flags & C_SUB)
6758 m3 = &m2->mc_xcursor->mx_cursor;
6761 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6762 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6769 if (rc == MDB_SUCCESS) {
6770 /* Now store the actual data in the child DB. Note that we're
6771 * storing the user data in the keys field, so there are strict
6772 * size limits on dupdata. The actual data fields of the child
6773 * DB are all zero size.
6776 int xflags, new_dupdata;
6781 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6782 if (flags & MDB_CURRENT) {
6783 xflags = MDB_CURRENT|MDB_NOSPILL;
6785 mdb_xcursor_init1(mc, leaf);
6786 xflags = (flags & MDB_NODUPDATA) ?
6787 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6790 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6791 new_dupdata = (int)dkey.mv_size;
6792 /* converted, write the original data first */
6794 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6797 /* we've done our job */
6800 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6801 /* Adjust other cursors pointing to mp */
6803 MDB_xcursor *mx = mc->mc_xcursor;
6804 unsigned i = mc->mc_top;
6805 MDB_page *mp = mc->mc_pg[i];
6807 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6808 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6809 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6810 if (m2->mc_pg[i] == mp) {
6811 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6812 mdb_xcursor_init2(m2, mx, new_dupdata);
6813 } else if (!insert_key) {
6814 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6815 if (!(n2->mn_flags & F_SUBDATA))
6816 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6821 ecount = mc->mc_xcursor->mx_db.md_entries;
6822 if (flags & MDB_APPENDDUP)
6823 xflags |= MDB_APPEND;
6824 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6825 if (flags & F_SUBDATA) {
6826 void *db = NODEDATA(leaf);
6827 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6829 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6831 /* Increment count unless we just replaced an existing item. */
6833 mc->mc_db->md_entries++;
6835 /* Invalidate txn if we created an empty sub-DB */
6838 /* If we succeeded and the key didn't exist before,
6839 * make sure the cursor is marked valid.
6841 mc->mc_flags |= C_INITIALIZED;
6843 if (flags & MDB_MULTIPLE) {
6846 /* let caller know how many succeeded, if any */
6847 data[1].mv_size = mcount;
6848 if (mcount < dcount) {
6849 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6850 insert_key = insert_data = 0;
6857 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6860 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6865 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6871 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6872 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6874 if (!(mc->mc_flags & C_INITIALIZED))
6877 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6878 return MDB_NOTFOUND;
6880 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6883 rc = mdb_cursor_touch(mc);
6887 mp = mc->mc_pg[mc->mc_top];
6890 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6892 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6893 if (flags & MDB_NODUPDATA) {
6894 /* mdb_cursor_del0() will subtract the final entry */
6895 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6897 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6898 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6900 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6903 /* If sub-DB still has entries, we're done */
6904 if (mc->mc_xcursor->mx_db.md_entries) {
6905 if (leaf->mn_flags & F_SUBDATA) {
6906 /* update subDB info */
6907 void *db = NODEDATA(leaf);
6908 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6911 /* shrink fake page */
6912 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6913 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6914 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6915 /* fix other sub-DB cursors pointed at fake pages on this page */
6916 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6917 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6918 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6919 if (m2->mc_pg[mc->mc_top] == mp) {
6920 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6921 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6923 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6924 if (!(n2->mn_flags & F_SUBDATA))
6925 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6930 mc->mc_db->md_entries--;
6931 mc->mc_flags |= C_DEL;
6934 /* otherwise fall thru and delete the sub-DB */
6937 if (leaf->mn_flags & F_SUBDATA) {
6938 /* add all the child DB's pages to the free list */
6939 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6944 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6945 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6946 rc = MDB_INCOMPATIBLE;
6950 /* add overflow pages to free list */
6951 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6955 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6956 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6957 (rc = mdb_ovpage_free(mc, omp)))
6962 return mdb_cursor_del0(mc);
6965 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6969 /** Allocate and initialize new pages for a database.
6970 * @param[in] mc a cursor on the database being added to.
6971 * @param[in] flags flags defining what type of page is being allocated.
6972 * @param[in] num the number of pages to allocate. This is usually 1,
6973 * unless allocating overflow pages for a large record.
6974 * @param[out] mp Address of a page, or NULL on failure.
6975 * @return 0 on success, non-zero on failure.
6978 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6983 if ((rc = mdb_page_alloc(mc, num, &np)))
6985 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6986 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6987 np->mp_flags = flags | P_DIRTY;
6988 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6989 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6992 mc->mc_db->md_branch_pages++;
6993 else if (IS_LEAF(np))
6994 mc->mc_db->md_leaf_pages++;
6995 else if (IS_OVERFLOW(np)) {
6996 mc->mc_db->md_overflow_pages += num;
7004 /** Calculate the size of a leaf node.
7005 * The size depends on the environment's page size; if a data item
7006 * is too large it will be put onto an overflow page and the node
7007 * size will only include the key and not the data. Sizes are always
7008 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7009 * of the #MDB_node headers.
7010 * @param[in] env The environment handle.
7011 * @param[in] key The key for the node.
7012 * @param[in] data The data for the node.
7013 * @return The number of bytes needed to store the node.
7016 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7020 sz = LEAFSIZE(key, data);
7021 if (sz > env->me_nodemax) {
7022 /* put on overflow page */
7023 sz -= data->mv_size - sizeof(pgno_t);
7026 return EVEN(sz + sizeof(indx_t));
7029 /** Calculate the size of a branch node.
7030 * The size should depend on the environment's page size but since
7031 * we currently don't support spilling large keys onto overflow
7032 * pages, it's simply the size of the #MDB_node header plus the
7033 * size of the key. Sizes are always rounded up to an even number
7034 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7035 * @param[in] env The environment handle.
7036 * @param[in] key The key for the node.
7037 * @return The number of bytes needed to store the node.
7040 mdb_branch_size(MDB_env *env, MDB_val *key)
7045 if (sz > env->me_nodemax) {
7046 /* put on overflow page */
7047 /* not implemented */
7048 /* sz -= key->size - sizeof(pgno_t); */
7051 return sz + sizeof(indx_t);
7054 /** Add a node to the page pointed to by the cursor.
7055 * @param[in] mc The cursor for this operation.
7056 * @param[in] indx The index on the page where the new node should be added.
7057 * @param[in] key The key for the new node.
7058 * @param[in] data The data for the new node, if any.
7059 * @param[in] pgno The page number, if adding a branch node.
7060 * @param[in] flags Flags for the node.
7061 * @return 0 on success, non-zero on failure. Possible errors are:
7063 * <li>ENOMEM - failed to allocate overflow pages for the node.
7064 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7065 * should never happen since all callers already calculate the
7066 * page's free space before calling this function.
7070 mdb_node_add(MDB_cursor *mc, indx_t indx,
7071 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7074 size_t node_size = NODESIZE;
7078 MDB_page *mp = mc->mc_pg[mc->mc_top];
7079 MDB_page *ofp = NULL; /* overflow page */
7083 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7085 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7086 IS_LEAF(mp) ? "leaf" : "branch",
7087 IS_SUBP(mp) ? "sub-" : "",
7088 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7089 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7092 /* Move higher keys up one slot. */
7093 int ksize = mc->mc_db->md_pad, dif;
7094 char *ptr = LEAF2KEY(mp, indx, ksize);
7095 dif = NUMKEYS(mp) - indx;
7097 memmove(ptr+ksize, ptr, dif*ksize);
7098 /* insert new key */
7099 memcpy(ptr, key->mv_data, ksize);
7101 /* Just using these for counting */
7102 mp->mp_lower += sizeof(indx_t);
7103 mp->mp_upper -= ksize - sizeof(indx_t);
7107 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7109 node_size += key->mv_size;
7111 mdb_cassert(mc, key && data);
7112 if (F_ISSET(flags, F_BIGDATA)) {
7113 /* Data already on overflow page. */
7114 node_size += sizeof(pgno_t);
7115 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7116 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7118 /* Put data on overflow page. */
7119 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7120 data->mv_size, node_size+data->mv_size));
7121 node_size = EVEN(node_size + sizeof(pgno_t));
7122 if ((ssize_t)node_size > room)
7124 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7126 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7130 node_size += data->mv_size;
7133 node_size = EVEN(node_size);
7134 if ((ssize_t)node_size > room)
7138 /* Move higher pointers up one slot. */
7139 for (i = NUMKEYS(mp); i > indx; i--)
7140 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7142 /* Adjust free space offsets. */
7143 ofs = mp->mp_upper - node_size;
7144 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7145 mp->mp_ptrs[indx] = ofs;
7147 mp->mp_lower += sizeof(indx_t);
7149 /* Write the node data. */
7150 node = NODEPTR(mp, indx);
7151 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7152 node->mn_flags = flags;
7154 SETDSZ(node,data->mv_size);
7159 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7162 ndata = NODEDATA(node);
7164 if (F_ISSET(flags, F_BIGDATA))
7165 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7166 else if (F_ISSET(flags, MDB_RESERVE))
7167 data->mv_data = ndata;
7169 memcpy(ndata, data->mv_data, data->mv_size);
7171 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7172 ndata = METADATA(ofp);
7173 if (F_ISSET(flags, MDB_RESERVE))
7174 data->mv_data = ndata;
7176 memcpy(ndata, data->mv_data, data->mv_size);
7183 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7184 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7185 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7186 DPRINTF(("node size = %"Z"u", node_size));
7187 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7188 return MDB_PAGE_FULL;
7191 /** Delete the specified node from a page.
7192 * @param[in] mc Cursor pointing to the node to delete.
7193 * @param[in] ksize The size of a node. Only used if the page is
7194 * part of a #MDB_DUPFIXED database.
7197 mdb_node_del(MDB_cursor *mc, int ksize)
7199 MDB_page *mp = mc->mc_pg[mc->mc_top];
7200 indx_t indx = mc->mc_ki[mc->mc_top];
7202 indx_t i, j, numkeys, ptr;
7206 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7207 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7208 numkeys = NUMKEYS(mp);
7209 mdb_cassert(mc, indx < numkeys);
7212 int x = numkeys - 1 - indx;
7213 base = LEAF2KEY(mp, indx, ksize);
7215 memmove(base, base + ksize, x * ksize);
7216 mp->mp_lower -= sizeof(indx_t);
7217 mp->mp_upper += ksize - sizeof(indx_t);
7221 node = NODEPTR(mp, indx);
7222 sz = NODESIZE + node->mn_ksize;
7224 if (F_ISSET(node->mn_flags, F_BIGDATA))
7225 sz += sizeof(pgno_t);
7227 sz += NODEDSZ(node);
7231 ptr = mp->mp_ptrs[indx];
7232 for (i = j = 0; i < numkeys; i++) {
7234 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7235 if (mp->mp_ptrs[i] < ptr)
7236 mp->mp_ptrs[j] += sz;
7241 base = (char *)mp + mp->mp_upper + PAGEBASE;
7242 memmove(base + sz, base, ptr - mp->mp_upper);
7244 mp->mp_lower -= sizeof(indx_t);
7248 /** Compact the main page after deleting a node on a subpage.
7249 * @param[in] mp The main page to operate on.
7250 * @param[in] indx The index of the subpage on the main page.
7253 mdb_node_shrink(MDB_page *mp, indx_t indx)
7258 indx_t delta, nsize, len, ptr;
7261 node = NODEPTR(mp, indx);
7262 sp = (MDB_page *)NODEDATA(node);
7263 delta = SIZELEFT(sp);
7264 nsize = NODEDSZ(node) - delta;
7266 /* Prepare to shift upward, set len = length(subpage part to shift) */
7270 return; /* do not make the node uneven-sized */
7272 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7273 for (i = NUMKEYS(sp); --i >= 0; )
7274 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7277 sp->mp_upper = sp->mp_lower;
7278 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7279 SETDSZ(node, nsize);
7281 /* Shift <lower nodes...initial part of subpage> upward */
7282 base = (char *)mp + mp->mp_upper + PAGEBASE;
7283 memmove(base + delta, base, (char *)sp + len - base);
7285 ptr = mp->mp_ptrs[indx];
7286 for (i = NUMKEYS(mp); --i >= 0; ) {
7287 if (mp->mp_ptrs[i] <= ptr)
7288 mp->mp_ptrs[i] += delta;
7290 mp->mp_upper += delta;
7293 /** Initial setup of a sorted-dups cursor.
7294 * Sorted duplicates are implemented as a sub-database for the given key.
7295 * The duplicate data items are actually keys of the sub-database.
7296 * Operations on the duplicate data items are performed using a sub-cursor
7297 * initialized when the sub-database is first accessed. This function does
7298 * the preliminary setup of the sub-cursor, filling in the fields that
7299 * depend only on the parent DB.
7300 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7303 mdb_xcursor_init0(MDB_cursor *mc)
7305 MDB_xcursor *mx = mc->mc_xcursor;
7307 mx->mx_cursor.mc_xcursor = NULL;
7308 mx->mx_cursor.mc_txn = mc->mc_txn;
7309 mx->mx_cursor.mc_db = &mx->mx_db;
7310 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7311 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7312 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7313 mx->mx_cursor.mc_snum = 0;
7314 mx->mx_cursor.mc_top = 0;
7315 mx->mx_cursor.mc_flags = C_SUB;
7316 mx->mx_dbx.md_name.mv_size = 0;
7317 mx->mx_dbx.md_name.mv_data = NULL;
7318 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7319 mx->mx_dbx.md_dcmp = NULL;
7320 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7323 /** Final setup of a sorted-dups cursor.
7324 * Sets up the fields that depend on the data from the main cursor.
7325 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7326 * @param[in] node The data containing the #MDB_db record for the
7327 * sorted-dup database.
7330 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7332 MDB_xcursor *mx = mc->mc_xcursor;
7334 if (node->mn_flags & F_SUBDATA) {
7335 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7336 mx->mx_cursor.mc_pg[0] = 0;
7337 mx->mx_cursor.mc_snum = 0;
7338 mx->mx_cursor.mc_top = 0;
7339 mx->mx_cursor.mc_flags = C_SUB;
7341 MDB_page *fp = NODEDATA(node);
7342 mx->mx_db.md_pad = 0;
7343 mx->mx_db.md_flags = 0;
7344 mx->mx_db.md_depth = 1;
7345 mx->mx_db.md_branch_pages = 0;
7346 mx->mx_db.md_leaf_pages = 1;
7347 mx->mx_db.md_overflow_pages = 0;
7348 mx->mx_db.md_entries = NUMKEYS(fp);
7349 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7350 mx->mx_cursor.mc_snum = 1;
7351 mx->mx_cursor.mc_top = 0;
7352 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7353 mx->mx_cursor.mc_pg[0] = fp;
7354 mx->mx_cursor.mc_ki[0] = 0;
7355 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7356 mx->mx_db.md_flags = MDB_DUPFIXED;
7357 mx->mx_db.md_pad = fp->mp_pad;
7358 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7359 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7362 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7363 mx->mx_db.md_root));
7364 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7365 #if UINT_MAX < SIZE_MAX
7366 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7367 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7372 /** Fixup a sorted-dups cursor due to underlying update.
7373 * Sets up some fields that depend on the data from the main cursor.
7374 * Almost the same as init1, but skips initialization steps if the
7375 * xcursor had already been used.
7376 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7377 * @param[in] src_mx The xcursor of an up-to-date cursor.
7378 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7381 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7383 MDB_xcursor *mx = mc->mc_xcursor;
7386 mx->mx_cursor.mc_snum = 1;
7387 mx->mx_cursor.mc_top = 0;
7388 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7389 mx->mx_cursor.mc_ki[0] = 0;
7390 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7391 #if UINT_MAX < SIZE_MAX
7392 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7394 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7397 mx->mx_db = src_mx->mx_db;
7398 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7399 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7400 mx->mx_db.md_root));
7403 /** Initialize a cursor for a given transaction and database. */
7405 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7408 mc->mc_backup = NULL;
7411 mc->mc_db = &txn->mt_dbs[dbi];
7412 mc->mc_dbx = &txn->mt_dbxs[dbi];
7413 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7419 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7420 mdb_tassert(txn, mx != NULL);
7421 mc->mc_xcursor = mx;
7422 mdb_xcursor_init0(mc);
7424 mc->mc_xcursor = NULL;
7426 if (*mc->mc_dbflag & DB_STALE) {
7427 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7432 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7435 size_t size = sizeof(MDB_cursor);
7437 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7440 if (txn->mt_flags & MDB_TXN_BLOCKED)
7443 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7446 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7447 size += sizeof(MDB_xcursor);
7449 if ((mc = malloc(size)) != NULL) {
7450 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7451 if (txn->mt_cursors) {
7452 mc->mc_next = txn->mt_cursors[dbi];
7453 txn->mt_cursors[dbi] = mc;
7454 mc->mc_flags |= C_UNTRACK;
7466 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7468 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7471 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7474 if (txn->mt_flags & MDB_TXN_BLOCKED)
7477 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7481 /* Return the count of duplicate data items for the current key */
7483 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7487 if (mc == NULL || countp == NULL)
7490 if (mc->mc_xcursor == NULL)
7491 return MDB_INCOMPATIBLE;
7493 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7496 if (!(mc->mc_flags & C_INITIALIZED))
7499 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7500 return MDB_NOTFOUND;
7502 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7503 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7506 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7509 *countp = mc->mc_xcursor->mx_db.md_entries;
7515 mdb_cursor_close(MDB_cursor *mc)
7517 if (mc && !mc->mc_backup) {
7518 /* remove from txn, if tracked */
7519 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7520 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7521 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7523 *prev = mc->mc_next;
7530 mdb_cursor_txn(MDB_cursor *mc)
7532 if (!mc) return NULL;
7537 mdb_cursor_dbi(MDB_cursor *mc)
7542 /** Replace the key for a branch node with a new key.
7543 * @param[in] mc Cursor pointing to the node to operate on.
7544 * @param[in] key The new key to use.
7545 * @return 0 on success, non-zero on failure.
7548 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7554 int delta, ksize, oksize;
7555 indx_t ptr, i, numkeys, indx;
7558 indx = mc->mc_ki[mc->mc_top];
7559 mp = mc->mc_pg[mc->mc_top];
7560 node = NODEPTR(mp, indx);
7561 ptr = mp->mp_ptrs[indx];
7565 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7566 k2.mv_data = NODEKEY(node);
7567 k2.mv_size = node->mn_ksize;
7568 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7570 mdb_dkey(&k2, kbuf2),
7576 /* Sizes must be 2-byte aligned. */
7577 ksize = EVEN(key->mv_size);
7578 oksize = EVEN(node->mn_ksize);
7579 delta = ksize - oksize;
7581 /* Shift node contents if EVEN(key length) changed. */
7583 if (delta > 0 && SIZELEFT(mp) < delta) {
7585 /* not enough space left, do a delete and split */
7586 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7587 pgno = NODEPGNO(node);
7588 mdb_node_del(mc, 0);
7589 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7592 numkeys = NUMKEYS(mp);
7593 for (i = 0; i < numkeys; i++) {
7594 if (mp->mp_ptrs[i] <= ptr)
7595 mp->mp_ptrs[i] -= delta;
7598 base = (char *)mp + mp->mp_upper + PAGEBASE;
7599 len = ptr - mp->mp_upper + NODESIZE;
7600 memmove(base - delta, base, len);
7601 mp->mp_upper -= delta;
7603 node = NODEPTR(mp, indx);
7606 /* But even if no shift was needed, update ksize */
7607 if (node->mn_ksize != key->mv_size)
7608 node->mn_ksize = key->mv_size;
7611 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7617 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7619 /** Move a node from csrc to cdst.
7622 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7629 unsigned short flags;
7633 /* Mark src and dst as dirty. */
7634 if ((rc = mdb_page_touch(csrc)) ||
7635 (rc = mdb_page_touch(cdst)))
7638 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7639 key.mv_size = csrc->mc_db->md_pad;
7640 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7642 data.mv_data = NULL;
7646 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7647 mdb_cassert(csrc, !((size_t)srcnode & 1));
7648 srcpg = NODEPGNO(srcnode);
7649 flags = srcnode->mn_flags;
7650 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7651 unsigned int snum = csrc->mc_snum;
7653 /* must find the lowest key below src */
7654 rc = mdb_page_search_lowest(csrc);
7657 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7658 key.mv_size = csrc->mc_db->md_pad;
7659 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7661 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7662 key.mv_size = NODEKSZ(s2);
7663 key.mv_data = NODEKEY(s2);
7665 csrc->mc_snum = snum--;
7666 csrc->mc_top = snum;
7668 key.mv_size = NODEKSZ(srcnode);
7669 key.mv_data = NODEKEY(srcnode);
7671 data.mv_size = NODEDSZ(srcnode);
7672 data.mv_data = NODEDATA(srcnode);
7674 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7675 unsigned int snum = cdst->mc_snum;
7678 /* must find the lowest key below dst */
7679 mdb_cursor_copy(cdst, &mn);
7680 rc = mdb_page_search_lowest(&mn);
7683 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7684 bkey.mv_size = mn.mc_db->md_pad;
7685 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7687 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7688 bkey.mv_size = NODEKSZ(s2);
7689 bkey.mv_data = NODEKEY(s2);
7691 mn.mc_snum = snum--;
7694 rc = mdb_update_key(&mn, &bkey);
7699 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7700 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7701 csrc->mc_ki[csrc->mc_top],
7703 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7704 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7706 /* Add the node to the destination page.
7708 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7709 if (rc != MDB_SUCCESS)
7712 /* Delete the node from the source page.
7714 mdb_node_del(csrc, key.mv_size);
7717 /* Adjust other cursors pointing to mp */
7718 MDB_cursor *m2, *m3;
7719 MDB_dbi dbi = csrc->mc_dbi;
7720 MDB_page *mpd, *mps;
7722 mps = csrc->mc_pg[csrc->mc_top];
7723 /* If we're adding on the left, bump others up */
7724 if (!cdst->mc_ki[csrc->mc_top]) {
7725 mpd = cdst->mc_pg[csrc->mc_top];
7726 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7727 if (csrc->mc_flags & C_SUB)
7728 m3 = &m2->mc_xcursor->mx_cursor;
7732 m3->mc_pg[csrc->mc_top] == mpd &&
7733 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7734 m3->mc_ki[csrc->mc_top]++;
7737 m3->mc_pg[csrc->mc_top] == mps &&
7738 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7739 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7740 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7741 m3->mc_ki[csrc->mc_top-1]++;
7745 /* Adding on the right, bump others down */
7747 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7748 if (csrc->mc_flags & C_SUB)
7749 m3 = &m2->mc_xcursor->mx_cursor;
7752 if (m3 == csrc) continue;
7753 if (m3->mc_pg[csrc->mc_top] == mps) {
7754 if (!m3->mc_ki[csrc->mc_top]) {
7755 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7756 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7757 m3->mc_ki[csrc->mc_top-1]--;
7759 m3->mc_ki[csrc->mc_top]--;
7766 /* Update the parent separators.
7768 if (csrc->mc_ki[csrc->mc_top] == 0) {
7769 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7770 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7771 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7773 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7774 key.mv_size = NODEKSZ(srcnode);
7775 key.mv_data = NODEKEY(srcnode);
7777 DPRINTF(("update separator for source page %"Z"u to [%s]",
7778 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7779 mdb_cursor_copy(csrc, &mn);
7782 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7785 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7787 indx_t ix = csrc->mc_ki[csrc->mc_top];
7788 nullkey.mv_size = 0;
7789 csrc->mc_ki[csrc->mc_top] = 0;
7790 rc = mdb_update_key(csrc, &nullkey);
7791 csrc->mc_ki[csrc->mc_top] = ix;
7792 mdb_cassert(csrc, rc == MDB_SUCCESS);
7796 if (cdst->mc_ki[cdst->mc_top] == 0) {
7797 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7798 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7799 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7801 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7802 key.mv_size = NODEKSZ(srcnode);
7803 key.mv_data = NODEKEY(srcnode);
7805 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7806 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7807 mdb_cursor_copy(cdst, &mn);
7810 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7813 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7815 indx_t ix = cdst->mc_ki[cdst->mc_top];
7816 nullkey.mv_size = 0;
7817 cdst->mc_ki[cdst->mc_top] = 0;
7818 rc = mdb_update_key(cdst, &nullkey);
7819 cdst->mc_ki[cdst->mc_top] = ix;
7820 mdb_cassert(cdst, rc == MDB_SUCCESS);
7827 /** Merge one page into another.
7828 * The nodes from the page pointed to by \b csrc will
7829 * be copied to the page pointed to by \b cdst and then
7830 * the \b csrc page will be freed.
7831 * @param[in] csrc Cursor pointing to the source page.
7832 * @param[in] cdst Cursor pointing to the destination page.
7833 * @return 0 on success, non-zero on failure.
7836 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7838 MDB_page *psrc, *pdst;
7845 psrc = csrc->mc_pg[csrc->mc_top];
7846 pdst = cdst->mc_pg[cdst->mc_top];
7848 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7850 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7851 mdb_cassert(csrc, cdst->mc_snum > 1);
7853 /* Mark dst as dirty. */
7854 if ((rc = mdb_page_touch(cdst)))
7857 /* get dst page again now that we've touched it. */
7858 pdst = cdst->mc_pg[cdst->mc_top];
7860 /* Move all nodes from src to dst.
7862 j = nkeys = NUMKEYS(pdst);
7863 if (IS_LEAF2(psrc)) {
7864 key.mv_size = csrc->mc_db->md_pad;
7865 key.mv_data = METADATA(psrc);
7866 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7867 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7868 if (rc != MDB_SUCCESS)
7870 key.mv_data = (char *)key.mv_data + key.mv_size;
7873 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7874 srcnode = NODEPTR(psrc, i);
7875 if (i == 0 && IS_BRANCH(psrc)) {
7878 mdb_cursor_copy(csrc, &mn);
7879 /* must find the lowest key below src */
7880 rc = mdb_page_search_lowest(&mn);
7883 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7884 key.mv_size = mn.mc_db->md_pad;
7885 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7887 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7888 key.mv_size = NODEKSZ(s2);
7889 key.mv_data = NODEKEY(s2);
7892 key.mv_size = srcnode->mn_ksize;
7893 key.mv_data = NODEKEY(srcnode);
7896 data.mv_size = NODEDSZ(srcnode);
7897 data.mv_data = NODEDATA(srcnode);
7898 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7899 if (rc != MDB_SUCCESS)
7904 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7905 pdst->mp_pgno, NUMKEYS(pdst),
7906 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7908 /* Unlink the src page from parent and add to free list.
7911 mdb_node_del(csrc, 0);
7912 if (csrc->mc_ki[csrc->mc_top] == 0) {
7914 rc = mdb_update_key(csrc, &key);
7922 psrc = csrc->mc_pg[csrc->mc_top];
7923 /* If not operating on FreeDB, allow this page to be reused
7924 * in this txn. Otherwise just add to free list.
7926 rc = mdb_page_loose(csrc, psrc);
7930 csrc->mc_db->md_leaf_pages--;
7932 csrc->mc_db->md_branch_pages--;
7934 /* Adjust other cursors pointing to mp */
7935 MDB_cursor *m2, *m3;
7936 MDB_dbi dbi = csrc->mc_dbi;
7937 unsigned int top = csrc->mc_top;
7939 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7940 if (csrc->mc_flags & C_SUB)
7941 m3 = &m2->mc_xcursor->mx_cursor;
7944 if (m3 == csrc) continue;
7945 if (m3->mc_snum < csrc->mc_snum) continue;
7946 if (m3->mc_pg[top] == psrc) {
7947 m3->mc_pg[top] = pdst;
7948 m3->mc_ki[top] += nkeys;
7949 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
7954 unsigned int snum = cdst->mc_snum;
7955 uint16_t depth = cdst->mc_db->md_depth;
7956 mdb_cursor_pop(cdst);
7957 rc = mdb_rebalance(cdst);
7958 /* Did the tree height change? */
7959 if (depth != cdst->mc_db->md_depth)
7960 snum += cdst->mc_db->md_depth - depth;
7961 cdst->mc_snum = snum;
7962 cdst->mc_top = snum-1;
7967 /** Copy the contents of a cursor.
7968 * @param[in] csrc The cursor to copy from.
7969 * @param[out] cdst The cursor to copy to.
7972 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7976 cdst->mc_txn = csrc->mc_txn;
7977 cdst->mc_dbi = csrc->mc_dbi;
7978 cdst->mc_db = csrc->mc_db;
7979 cdst->mc_dbx = csrc->mc_dbx;
7980 cdst->mc_snum = csrc->mc_snum;
7981 cdst->mc_top = csrc->mc_top;
7982 cdst->mc_flags = csrc->mc_flags;
7984 for (i=0; i<csrc->mc_snum; i++) {
7985 cdst->mc_pg[i] = csrc->mc_pg[i];
7986 cdst->mc_ki[i] = csrc->mc_ki[i];
7990 /** Rebalance the tree after a delete operation.
7991 * @param[in] mc Cursor pointing to the page where rebalancing
7993 * @return 0 on success, non-zero on failure.
7996 mdb_rebalance(MDB_cursor *mc)
8000 unsigned int ptop, minkeys, thresh;
8004 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8009 thresh = FILL_THRESHOLD;
8011 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
8012 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8013 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8014 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8016 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8017 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8018 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
8019 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8023 if (mc->mc_snum < 2) {
8024 MDB_page *mp = mc->mc_pg[0];
8026 DPUTS("Can't rebalance a subpage, ignoring");
8029 if (NUMKEYS(mp) == 0) {
8030 DPUTS("tree is completely empty");
8031 mc->mc_db->md_root = P_INVALID;
8032 mc->mc_db->md_depth = 0;
8033 mc->mc_db->md_leaf_pages = 0;
8034 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8037 /* Adjust cursors pointing to mp */
8040 mc->mc_flags &= ~C_INITIALIZED;
8042 MDB_cursor *m2, *m3;
8043 MDB_dbi dbi = mc->mc_dbi;
8045 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8046 if (mc->mc_flags & C_SUB)
8047 m3 = &m2->mc_xcursor->mx_cursor;
8050 if (m3->mc_snum < mc->mc_snum) continue;
8051 if (m3->mc_pg[0] == mp) {
8054 m3->mc_flags &= ~C_INITIALIZED;
8058 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8060 DPUTS("collapsing root page!");
8061 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8064 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8065 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
8068 mc->mc_db->md_depth--;
8069 mc->mc_db->md_branch_pages--;
8070 mc->mc_ki[0] = mc->mc_ki[1];
8071 for (i = 1; i<mc->mc_db->md_depth; i++) {
8072 mc->mc_pg[i] = mc->mc_pg[i+1];
8073 mc->mc_ki[i] = mc->mc_ki[i+1];
8076 /* Adjust other cursors pointing to mp */
8077 MDB_cursor *m2, *m3;
8078 MDB_dbi dbi = mc->mc_dbi;
8080 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8081 if (mc->mc_flags & C_SUB)
8082 m3 = &m2->mc_xcursor->mx_cursor;
8085 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
8086 if (m3->mc_pg[0] == mp) {
8087 for (i=0; i<m3->mc_snum; i++) {
8088 m3->mc_pg[i] = m3->mc_pg[i+1];
8089 m3->mc_ki[i] = m3->mc_ki[i+1];
8097 DPUTS("root page doesn't need rebalancing");
8101 /* The parent (branch page) must have at least 2 pointers,
8102 * otherwise the tree is invalid.
8104 ptop = mc->mc_top-1;
8105 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8107 /* Leaf page fill factor is below the threshold.
8108 * Try to move keys from left or right neighbor, or
8109 * merge with a neighbor page.
8114 mdb_cursor_copy(mc, &mn);
8115 mn.mc_xcursor = NULL;
8117 oldki = mc->mc_ki[mc->mc_top];
8118 if (mc->mc_ki[ptop] == 0) {
8119 /* We're the leftmost leaf in our parent.
8121 DPUTS("reading right neighbor");
8123 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8124 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8127 mn.mc_ki[mn.mc_top] = 0;
8128 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8130 /* There is at least one neighbor to the left.
8132 DPUTS("reading left neighbor");
8134 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8135 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8138 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8139 mc->mc_ki[mc->mc_top] = 0;
8142 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8143 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8144 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8146 /* If the neighbor page is above threshold and has enough keys,
8147 * move one key from it. Otherwise we should try to merge them.
8148 * (A branch page must never have less than 2 keys.)
8150 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8151 rc = mdb_node_move(&mn, mc);
8152 if (!mc->mc_ki[mc->mc_top]) {
8153 /* if we inserted on left, bump position up */
8157 if (mc->mc_ki[ptop] == 0) {
8158 rc = mdb_page_merge(&mn, mc);
8161 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8162 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8163 /* We want mdb_rebalance to find mn when doing fixups */
8164 if (mc->mc_flags & C_SUB) {
8165 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8166 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8167 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8169 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8170 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8172 rc = mdb_page_merge(mc, &mn);
8173 if (mc->mc_flags & C_SUB)
8174 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8176 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8177 mdb_cursor_copy(&mn, mc);
8179 mc->mc_flags &= ~C_EOF;
8181 mc->mc_ki[mc->mc_top] = oldki;
8185 /** Complete a delete operation started by #mdb_cursor_del(). */
8187 mdb_cursor_del0(MDB_cursor *mc)
8193 MDB_cursor *m2, *m3;
8194 MDB_dbi dbi = mc->mc_dbi;
8196 ki = mc->mc_ki[mc->mc_top];
8197 mp = mc->mc_pg[mc->mc_top];
8198 mdb_node_del(mc, mc->mc_db->md_pad);
8199 mc->mc_db->md_entries--;
8201 /* Adjust other cursors pointing to mp */
8202 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8203 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8204 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8206 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8208 if (m3->mc_pg[mc->mc_top] == mp) {
8209 if (m3->mc_ki[mc->mc_top] >= ki) {
8210 m3->mc_flags |= C_DEL;
8211 if (m3->mc_ki[mc->mc_top] > ki)
8212 m3->mc_ki[mc->mc_top]--;
8213 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8214 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8219 rc = mdb_rebalance(mc);
8221 if (rc == MDB_SUCCESS) {
8222 /* DB is totally empty now, just bail out.
8223 * Other cursors adjustments were already done
8224 * by mdb_rebalance and aren't needed here.
8229 mp = mc->mc_pg[mc->mc_top];
8230 nkeys = NUMKEYS(mp);
8232 /* Adjust other cursors pointing to mp */
8233 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8234 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8235 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8237 if (m3->mc_snum < mc->mc_snum)
8239 if (m3->mc_pg[mc->mc_top] == mp) {
8240 /* if m3 points past last node in page, find next sibling */
8241 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8242 rc = mdb_cursor_sibling(m3, 1);
8243 if (rc == MDB_NOTFOUND) {
8244 m3->mc_flags |= C_EOF;
8250 mc->mc_flags |= C_DEL;
8254 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8259 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8260 MDB_val *key, MDB_val *data)
8262 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8265 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8266 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8268 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8269 /* must ignore any data */
8273 return mdb_del0(txn, dbi, key, data, 0);
8277 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8278 MDB_val *key, MDB_val *data, unsigned flags)
8283 MDB_val rdata, *xdata;
8287 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8289 mdb_cursor_init(&mc, txn, dbi, &mx);
8298 flags |= MDB_NODUPDATA;
8300 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8302 /* let mdb_page_split know about this cursor if needed:
8303 * delete will trigger a rebalance; if it needs to move
8304 * a node from one page to another, it will have to
8305 * update the parent's separator key(s). If the new sepkey
8306 * is larger than the current one, the parent page may
8307 * run out of space, triggering a split. We need this
8308 * cursor to be consistent until the end of the rebalance.
8310 mc.mc_flags |= C_UNTRACK;
8311 mc.mc_next = txn->mt_cursors[dbi];
8312 txn->mt_cursors[dbi] = &mc;
8313 rc = mdb_cursor_del(&mc, flags);
8314 txn->mt_cursors[dbi] = mc.mc_next;
8319 /** Split a page and insert a new node.
8320 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8321 * The cursor will be updated to point to the actual page and index where
8322 * the node got inserted after the split.
8323 * @param[in] newkey The key for the newly inserted node.
8324 * @param[in] newdata The data for the newly inserted node.
8325 * @param[in] newpgno The page number, if the new node is a branch node.
8326 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8327 * @return 0 on success, non-zero on failure.
8330 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8331 unsigned int nflags)
8334 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8337 int i, j, split_indx, nkeys, pmax;
8338 MDB_env *env = mc->mc_txn->mt_env;
8340 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8341 MDB_page *copy = NULL;
8342 MDB_page *mp, *rp, *pp;
8347 mp = mc->mc_pg[mc->mc_top];
8348 newindx = mc->mc_ki[mc->mc_top];
8349 nkeys = NUMKEYS(mp);
8351 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8352 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8353 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8355 /* Create a right sibling. */
8356 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8358 rp->mp_pad = mp->mp_pad;
8359 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8361 if (mc->mc_snum < 2) {
8362 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8364 /* shift current top to make room for new parent */
8365 mc->mc_pg[1] = mc->mc_pg[0];
8366 mc->mc_ki[1] = mc->mc_ki[0];
8369 mc->mc_db->md_root = pp->mp_pgno;
8370 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8371 new_root = mc->mc_db->md_depth++;
8373 /* Add left (implicit) pointer. */
8374 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8375 /* undo the pre-push */
8376 mc->mc_pg[0] = mc->mc_pg[1];
8377 mc->mc_ki[0] = mc->mc_ki[1];
8378 mc->mc_db->md_root = mp->mp_pgno;
8379 mc->mc_db->md_depth--;
8386 ptop = mc->mc_top-1;
8387 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8390 mc->mc_flags |= C_SPLITTING;
8391 mdb_cursor_copy(mc, &mn);
8392 mn.mc_pg[mn.mc_top] = rp;
8393 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8395 if (nflags & MDB_APPEND) {
8396 mn.mc_ki[mn.mc_top] = 0;
8398 split_indx = newindx;
8402 split_indx = (nkeys+1) / 2;
8407 unsigned int lsize, rsize, ksize;
8408 /* Move half of the keys to the right sibling */
8409 x = mc->mc_ki[mc->mc_top] - split_indx;
8410 ksize = mc->mc_db->md_pad;
8411 split = LEAF2KEY(mp, split_indx, ksize);
8412 rsize = (nkeys - split_indx) * ksize;
8413 lsize = (nkeys - split_indx) * sizeof(indx_t);
8414 mp->mp_lower -= lsize;
8415 rp->mp_lower += lsize;
8416 mp->mp_upper += rsize - lsize;
8417 rp->mp_upper -= rsize - lsize;
8418 sepkey.mv_size = ksize;
8419 if (newindx == split_indx) {
8420 sepkey.mv_data = newkey->mv_data;
8422 sepkey.mv_data = split;
8425 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8426 memcpy(rp->mp_ptrs, split, rsize);
8427 sepkey.mv_data = rp->mp_ptrs;
8428 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8429 memcpy(ins, newkey->mv_data, ksize);
8430 mp->mp_lower += sizeof(indx_t);
8431 mp->mp_upper -= ksize - sizeof(indx_t);
8434 memcpy(rp->mp_ptrs, split, x * ksize);
8435 ins = LEAF2KEY(rp, x, ksize);
8436 memcpy(ins, newkey->mv_data, ksize);
8437 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8438 rp->mp_lower += sizeof(indx_t);
8439 rp->mp_upper -= ksize - sizeof(indx_t);
8440 mc->mc_ki[mc->mc_top] = x;
8441 mc->mc_pg[mc->mc_top] = rp;
8445 int psize, nsize, k;
8446 /* Maximum free space in an empty page */
8447 pmax = env->me_psize - PAGEHDRSZ;
8449 nsize = mdb_leaf_size(env, newkey, newdata);
8451 nsize = mdb_branch_size(env, newkey);
8452 nsize = EVEN(nsize);
8454 /* grab a page to hold a temporary copy */
8455 copy = mdb_page_malloc(mc->mc_txn, 1);
8460 copy->mp_pgno = mp->mp_pgno;
8461 copy->mp_flags = mp->mp_flags;
8462 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8463 copy->mp_upper = env->me_psize - PAGEBASE;
8465 /* prepare to insert */
8466 for (i=0, j=0; i<nkeys; i++) {
8468 copy->mp_ptrs[j++] = 0;
8470 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8473 /* When items are relatively large the split point needs
8474 * to be checked, because being off-by-one will make the
8475 * difference between success or failure in mdb_node_add.
8477 * It's also relevant if a page happens to be laid out
8478 * such that one half of its nodes are all "small" and
8479 * the other half of its nodes are "large." If the new
8480 * item is also "large" and falls on the half with
8481 * "large" nodes, it also may not fit.
8483 * As a final tweak, if the new item goes on the last
8484 * spot on the page (and thus, onto the new page), bias
8485 * the split so the new page is emptier than the old page.
8486 * This yields better packing during sequential inserts.
8488 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8489 /* Find split point */
8491 if (newindx <= split_indx || newindx >= nkeys) {
8493 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8498 for (; i!=k; i+=j) {
8503 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8504 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8506 if (F_ISSET(node->mn_flags, F_BIGDATA))
8507 psize += sizeof(pgno_t);
8509 psize += NODEDSZ(node);
8511 psize = EVEN(psize);
8513 if (psize > pmax || i == k-j) {
8514 split_indx = i + (j<0);
8519 if (split_indx == newindx) {
8520 sepkey.mv_size = newkey->mv_size;
8521 sepkey.mv_data = newkey->mv_data;
8523 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8524 sepkey.mv_size = node->mn_ksize;
8525 sepkey.mv_data = NODEKEY(node);
8530 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8532 /* Copy separator key to the parent.
8534 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8538 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8543 if (mn.mc_snum == mc->mc_snum) {
8544 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8545 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8546 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8547 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8552 /* Right page might now have changed parent.
8553 * Check if left page also changed parent.
8555 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8556 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8557 for (i=0; i<ptop; i++) {
8558 mc->mc_pg[i] = mn.mc_pg[i];
8559 mc->mc_ki[i] = mn.mc_ki[i];
8561 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8562 if (mn.mc_ki[ptop]) {
8563 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8565 /* find right page's left sibling */
8566 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8567 mdb_cursor_sibling(mc, 0);
8572 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8575 mc->mc_flags ^= C_SPLITTING;
8576 if (rc != MDB_SUCCESS) {
8579 if (nflags & MDB_APPEND) {
8580 mc->mc_pg[mc->mc_top] = rp;
8581 mc->mc_ki[mc->mc_top] = 0;
8582 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8585 for (i=0; i<mc->mc_top; i++)
8586 mc->mc_ki[i] = mn.mc_ki[i];
8587 } else if (!IS_LEAF2(mp)) {
8589 mc->mc_pg[mc->mc_top] = rp;
8594 rkey.mv_data = newkey->mv_data;
8595 rkey.mv_size = newkey->mv_size;
8601 /* Update index for the new key. */
8602 mc->mc_ki[mc->mc_top] = j;
8604 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8605 rkey.mv_data = NODEKEY(node);
8606 rkey.mv_size = node->mn_ksize;
8608 xdata.mv_data = NODEDATA(node);
8609 xdata.mv_size = NODEDSZ(node);
8612 pgno = NODEPGNO(node);
8613 flags = node->mn_flags;
8616 if (!IS_LEAF(mp) && j == 0) {
8617 /* First branch index doesn't need key data. */
8621 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8627 mc->mc_pg[mc->mc_top] = copy;
8632 } while (i != split_indx);
8634 nkeys = NUMKEYS(copy);
8635 for (i=0; i<nkeys; i++)
8636 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8637 mp->mp_lower = copy->mp_lower;
8638 mp->mp_upper = copy->mp_upper;
8639 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8640 env->me_psize - copy->mp_upper - PAGEBASE);
8642 /* reset back to original page */
8643 if (newindx < split_indx) {
8644 mc->mc_pg[mc->mc_top] = mp;
8645 if (nflags & MDB_RESERVE) {
8646 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8647 if (!(node->mn_flags & F_BIGDATA))
8648 newdata->mv_data = NODEDATA(node);
8651 mc->mc_pg[mc->mc_top] = rp;
8653 /* Make sure mc_ki is still valid.
8655 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8656 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8657 for (i=0; i<=ptop; i++) {
8658 mc->mc_pg[i] = mn.mc_pg[i];
8659 mc->mc_ki[i] = mn.mc_ki[i];
8666 /* Adjust other cursors pointing to mp */
8667 MDB_cursor *m2, *m3;
8668 MDB_dbi dbi = mc->mc_dbi;
8669 int fixup = NUMKEYS(mp);
8671 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8672 if (mc->mc_flags & C_SUB)
8673 m3 = &m2->mc_xcursor->mx_cursor;
8678 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8680 if (m3->mc_flags & C_SPLITTING)
8685 for (k=new_root; k>=0; k--) {
8686 m3->mc_ki[k+1] = m3->mc_ki[k];
8687 m3->mc_pg[k+1] = m3->mc_pg[k];
8689 if (m3->mc_ki[0] >= split_indx) {
8694 m3->mc_pg[0] = mc->mc_pg[0];
8698 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8699 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8700 m3->mc_ki[mc->mc_top]++;
8701 if (m3->mc_ki[mc->mc_top] >= fixup) {
8702 m3->mc_pg[mc->mc_top] = rp;
8703 m3->mc_ki[mc->mc_top] -= fixup;
8704 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8706 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8707 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8712 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8715 if (copy) /* tmp page */
8716 mdb_page_free(env, copy);
8718 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8723 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8724 MDB_val *key, MDB_val *data, unsigned int flags)
8729 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8732 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8735 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8736 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8738 mdb_cursor_init(&mc, txn, dbi, &mx);
8739 return mdb_cursor_put(&mc, key, data, flags);
8743 #define MDB_WBUF (1024*1024)
8746 /** State needed for a compacting copy. */
8747 typedef struct mdb_copy {
8748 pthread_mutex_t mc_mutex;
8749 pthread_cond_t mc_cond;
8756 pgno_t mc_next_pgno;
8759 volatile int mc_new;
8764 /** Dedicated writer thread for compacting copy. */
8765 static THREAD_RET ESECT CALL_CONV
8766 mdb_env_copythr(void *arg)
8770 int toggle = 0, wsize, rc;
8773 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8776 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8779 pthread_mutex_lock(&my->mc_mutex);
8781 pthread_cond_signal(&my->mc_cond);
8784 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8785 if (my->mc_new < 0) {
8790 wsize = my->mc_wlen[toggle];
8791 ptr = my->mc_wbuf[toggle];
8794 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8798 } else if (len > 0) {
8812 /* If there's an overflow page tail, write it too */
8813 if (my->mc_olen[toggle]) {
8814 wsize = my->mc_olen[toggle];
8815 ptr = my->mc_over[toggle];
8816 my->mc_olen[toggle] = 0;
8819 my->mc_wlen[toggle] = 0;
8821 pthread_cond_signal(&my->mc_cond);
8823 pthread_cond_signal(&my->mc_cond);
8824 pthread_mutex_unlock(&my->mc_mutex);
8825 return (THREAD_RET)0;
8829 /** Tell the writer thread there's a buffer ready to write */
8831 mdb_env_cthr_toggle(mdb_copy *my, int st)
8833 int toggle = my->mc_toggle ^ 1;
8834 pthread_mutex_lock(&my->mc_mutex);
8835 if (my->mc_status) {
8836 pthread_mutex_unlock(&my->mc_mutex);
8837 return my->mc_status;
8839 while (my->mc_new == 1)
8840 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8842 my->mc_toggle = toggle;
8843 pthread_cond_signal(&my->mc_cond);
8844 pthread_mutex_unlock(&my->mc_mutex);
8848 /** Depth-first tree traversal for compacting copy. */
8850 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8853 MDB_txn *txn = my->mc_txn;
8855 MDB_page *mo, *mp, *leaf;
8860 /* Empty DB, nothing to do */
8861 if (*pg == P_INVALID)
8868 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8871 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8875 /* Make cursor pages writable */
8876 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8880 for (i=0; i<mc.mc_top; i++) {
8881 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8882 mc.mc_pg[i] = (MDB_page *)ptr;
8883 ptr += my->mc_env->me_psize;
8886 /* This is writable space for a leaf page. Usually not needed. */
8887 leaf = (MDB_page *)ptr;
8889 toggle = my->mc_toggle;
8890 while (mc.mc_snum > 0) {
8892 mp = mc.mc_pg[mc.mc_top];
8896 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8897 for (i=0; i<n; i++) {
8898 ni = NODEPTR(mp, i);
8899 if (ni->mn_flags & F_BIGDATA) {
8903 /* Need writable leaf */
8905 mc.mc_pg[mc.mc_top] = leaf;
8906 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8908 ni = NODEPTR(mp, i);
8911 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8912 rc = mdb_page_get(txn, pg, &omp, NULL);
8915 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8916 rc = mdb_env_cthr_toggle(my, 1);
8919 toggle = my->mc_toggle;
8921 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8922 memcpy(mo, omp, my->mc_env->me_psize);
8923 mo->mp_pgno = my->mc_next_pgno;
8924 my->mc_next_pgno += omp->mp_pages;
8925 my->mc_wlen[toggle] += my->mc_env->me_psize;
8926 if (omp->mp_pages > 1) {
8927 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8928 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8929 rc = mdb_env_cthr_toggle(my, 1);
8932 toggle = my->mc_toggle;
8934 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8935 } else if (ni->mn_flags & F_SUBDATA) {
8938 /* Need writable leaf */
8940 mc.mc_pg[mc.mc_top] = leaf;
8941 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8943 ni = NODEPTR(mp, i);
8946 memcpy(&db, NODEDATA(ni), sizeof(db));
8947 my->mc_toggle = toggle;
8948 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8951 toggle = my->mc_toggle;
8952 memcpy(NODEDATA(ni), &db, sizeof(db));
8957 mc.mc_ki[mc.mc_top]++;
8958 if (mc.mc_ki[mc.mc_top] < n) {
8961 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8963 rc = mdb_page_get(txn, pg, &mp, NULL);
8968 mc.mc_ki[mc.mc_top] = 0;
8969 if (IS_BRANCH(mp)) {
8970 /* Whenever we advance to a sibling branch page,
8971 * we must proceed all the way down to its first leaf.
8973 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8976 mc.mc_pg[mc.mc_top] = mp;
8980 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8981 rc = mdb_env_cthr_toggle(my, 1);
8984 toggle = my->mc_toggle;
8986 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8987 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8988 mo->mp_pgno = my->mc_next_pgno++;
8989 my->mc_wlen[toggle] += my->mc_env->me_psize;
8991 /* Update parent if there is one */
8992 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8993 SETPGNO(ni, mo->mp_pgno);
8994 mdb_cursor_pop(&mc);
8996 /* Otherwise we're done */
9006 /** Copy environment with compaction. */
9008 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9013 MDB_txn *txn = NULL;
9018 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
9019 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
9020 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9021 if (my.mc_wbuf[0] == NULL)
9024 pthread_mutex_init(&my.mc_mutex, NULL);
9025 pthread_cond_init(&my.mc_cond, NULL);
9026 #ifdef HAVE_MEMALIGN
9027 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9028 if (my.mc_wbuf[0] == NULL)
9031 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
9036 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9037 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9042 my.mc_next_pgno = NUM_METAS;
9048 THREAD_CREATE(thr, mdb_env_copythr, &my);
9050 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9054 mp = (MDB_page *)my.mc_wbuf[0];
9055 memset(mp, 0, NUM_METAS * env->me_psize);
9057 mp->mp_flags = P_META;
9058 mm = (MDB_meta *)METADATA(mp);
9059 mdb_env_init_meta0(env, mm);
9060 mm->mm_address = env->me_metas[0]->mm_address;
9062 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9064 mp->mp_flags = P_META;
9065 *(MDB_meta *)METADATA(mp) = *mm;
9066 mm = (MDB_meta *)METADATA(mp);
9068 /* Count the number of free pages, subtract from lastpg to find
9069 * number of active pages
9072 MDB_ID freecount = 0;
9075 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9076 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9077 freecount += *(MDB_ID *)data.mv_data;
9078 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9079 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9080 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9082 /* Set metapage 1 */
9083 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
9084 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9085 if (mm->mm_last_pg > NUM_METAS-1) {
9086 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
9089 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
9092 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9094 pthread_mutex_lock(&my.mc_mutex);
9096 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9097 pthread_mutex_unlock(&my.mc_mutex);
9098 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
9099 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
9100 rc = mdb_env_cthr_toggle(&my, 1);
9101 mdb_env_cthr_toggle(&my, -1);
9102 pthread_mutex_lock(&my.mc_mutex);
9104 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9105 pthread_mutex_unlock(&my.mc_mutex);
9110 CloseHandle(my.mc_cond);
9111 CloseHandle(my.mc_mutex);
9112 _aligned_free(my.mc_wbuf[0]);
9114 pthread_cond_destroy(&my.mc_cond);
9115 pthread_mutex_destroy(&my.mc_mutex);
9116 free(my.mc_wbuf[0]);
9121 /** Copy environment as-is. */
9123 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9125 MDB_txn *txn = NULL;
9126 mdb_mutexref_t wmutex = NULL;
9132 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9136 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9139 /* Do the lock/unlock of the reader mutex before starting the
9140 * write txn. Otherwise other read txns could block writers.
9142 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9147 /* We must start the actual read txn after blocking writers */
9148 mdb_txn_end(txn, MDB_END_RESET_TMP);
9150 /* Temporarily block writers until we snapshot the meta pages */
9151 wmutex = env->me_wmutex;
9152 if (LOCK_MUTEX(rc, env, wmutex))
9155 rc = mdb_txn_renew0(txn);
9157 UNLOCK_MUTEX(wmutex);
9162 wsize = env->me_psize * NUM_METAS;
9166 DO_WRITE(rc, fd, ptr, w2, len);
9170 } else if (len > 0) {
9176 /* Non-blocking or async handles are not supported */
9182 UNLOCK_MUTEX(wmutex);
9187 w2 = txn->mt_next_pgno * env->me_psize;
9190 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9197 if (wsize > MAX_WRITE)
9201 DO_WRITE(rc, fd, ptr, w2, len);
9205 } else if (len > 0) {
9222 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9224 if (flags & MDB_CP_COMPACT)
9225 return mdb_env_copyfd1(env, fd);
9227 return mdb_env_copyfd0(env, fd);
9231 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9233 return mdb_env_copyfd2(env, fd, 0);
9237 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9241 HANDLE newfd = INVALID_HANDLE_VALUE;
9243 if (env->me_flags & MDB_NOSUBDIR) {
9244 lpath = (char *)path;
9247 len += sizeof(DATANAME);
9248 lpath = malloc(len);
9251 sprintf(lpath, "%s" DATANAME, path);
9254 /* The destination path must exist, but the destination file must not.
9255 * We don't want the OS to cache the writes, since the source data is
9256 * already in the OS cache.
9259 newfd = CreateFileA(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9260 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9262 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9264 if (newfd == INVALID_HANDLE_VALUE) {
9269 if (env->me_psize >= env->me_os_psize) {
9271 /* Set O_DIRECT if the file system supports it */
9272 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9273 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9275 #ifdef F_NOCACHE /* __APPLE__ */
9276 rc = fcntl(newfd, F_NOCACHE, 1);
9284 rc = mdb_env_copyfd2(env, newfd, flags);
9287 if (!(env->me_flags & MDB_NOSUBDIR))
9289 if (newfd != INVALID_HANDLE_VALUE)
9290 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9297 mdb_env_copy(MDB_env *env, const char *path)
9299 return mdb_env_copy2(env, path, 0);
9303 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9305 if (flag & ~CHANGEABLE)
9308 env->me_flags |= flag;
9310 env->me_flags &= ~flag;
9315 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9320 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9325 mdb_env_set_userctx(MDB_env *env, void *ctx)
9329 env->me_userctx = ctx;
9334 mdb_env_get_userctx(MDB_env *env)
9336 return env ? env->me_userctx : NULL;
9340 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9345 env->me_assert_func = func;
9351 mdb_env_get_path(MDB_env *env, const char **arg)
9356 *arg = env->me_path;
9361 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9370 /** Common code for #mdb_stat() and #mdb_env_stat().
9371 * @param[in] env the environment to operate in.
9372 * @param[in] db the #MDB_db record containing the stats to return.
9373 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9374 * @return 0, this function always succeeds.
9377 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9379 arg->ms_psize = env->me_psize;
9380 arg->ms_depth = db->md_depth;
9381 arg->ms_branch_pages = db->md_branch_pages;
9382 arg->ms_leaf_pages = db->md_leaf_pages;
9383 arg->ms_overflow_pages = db->md_overflow_pages;
9384 arg->ms_entries = db->md_entries;
9390 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9394 if (env == NULL || arg == NULL)
9397 meta = mdb_env_pick_meta(env);
9399 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9403 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9407 if (env == NULL || arg == NULL)
9410 meta = mdb_env_pick_meta(env);
9411 arg->me_mapaddr = meta->mm_address;
9412 arg->me_last_pgno = meta->mm_last_pg;
9413 arg->me_last_txnid = meta->mm_txnid;
9415 arg->me_mapsize = env->me_mapsize;
9416 arg->me_maxreaders = env->me_maxreaders;
9417 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9421 /** Set the default comparison functions for a database.
9422 * Called immediately after a database is opened to set the defaults.
9423 * The user can then override them with #mdb_set_compare() or
9424 * #mdb_set_dupsort().
9425 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9426 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9429 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9431 uint16_t f = txn->mt_dbs[dbi].md_flags;
9433 txn->mt_dbxs[dbi].md_cmp =
9434 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9435 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9437 txn->mt_dbxs[dbi].md_dcmp =
9438 !(f & MDB_DUPSORT) ? 0 :
9439 ((f & MDB_INTEGERDUP)
9440 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9441 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9444 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9450 int rc, dbflag, exact;
9451 unsigned int unused = 0, seq;
9454 if (flags & ~VALID_FLAGS)
9456 if (txn->mt_flags & MDB_TXN_BLOCKED)
9462 if (flags & PERSISTENT_FLAGS) {
9463 uint16_t f2 = flags & PERSISTENT_FLAGS;
9464 /* make sure flag changes get committed */
9465 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9466 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9467 txn->mt_flags |= MDB_TXN_DIRTY;
9470 mdb_default_cmp(txn, MAIN_DBI);
9474 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9475 mdb_default_cmp(txn, MAIN_DBI);
9478 /* Is the DB already open? */
9480 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9481 if (!txn->mt_dbxs[i].md_name.mv_size) {
9482 /* Remember this free slot */
9483 if (!unused) unused = i;
9486 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9487 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9493 /* If no free slot and max hit, fail */
9494 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9495 return MDB_DBS_FULL;
9497 /* Cannot mix named databases with some mainDB flags */
9498 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9499 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9501 /* Find the DB info */
9502 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9505 key.mv_data = (void *)name;
9506 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9507 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9508 if (rc == MDB_SUCCESS) {
9509 /* make sure this is actually a DB */
9510 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9511 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9512 return MDB_INCOMPATIBLE;
9513 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9514 /* Create if requested */
9515 data.mv_size = sizeof(MDB_db);
9516 data.mv_data = &dummy;
9517 memset(&dummy, 0, sizeof(dummy));
9518 dummy.md_root = P_INVALID;
9519 dummy.md_flags = flags & PERSISTENT_FLAGS;
9520 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9524 /* OK, got info, add to table */
9525 if (rc == MDB_SUCCESS) {
9526 unsigned int slot = unused ? unused : txn->mt_numdbs;
9527 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9528 txn->mt_dbxs[slot].md_name.mv_size = len;
9529 txn->mt_dbxs[slot].md_rel = NULL;
9530 txn->mt_dbflags[slot] = dbflag;
9531 /* txn-> and env-> are the same in read txns, use
9532 * tmp variable to avoid undefined assignment
9534 seq = ++txn->mt_env->me_dbiseqs[slot];
9535 txn->mt_dbiseqs[slot] = seq;
9537 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9539 mdb_default_cmp(txn, slot);
9549 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9551 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9554 if (txn->mt_flags & MDB_TXN_BLOCKED)
9557 if (txn->mt_dbflags[dbi] & DB_STALE) {
9560 /* Stale, must read the DB's root. cursor_init does it for us. */
9561 mdb_cursor_init(&mc, txn, dbi, &mx);
9563 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9566 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9569 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9571 ptr = env->me_dbxs[dbi].md_name.mv_data;
9572 /* If there was no name, this was already closed */
9574 env->me_dbxs[dbi].md_name.mv_data = NULL;
9575 env->me_dbxs[dbi].md_name.mv_size = 0;
9576 env->me_dbflags[dbi] = 0;
9577 env->me_dbiseqs[dbi]++;
9582 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9584 /* We could return the flags for the FREE_DBI too but what's the point? */
9585 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9587 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9591 /** Add all the DB's pages to the free list.
9592 * @param[in] mc Cursor on the DB to free.
9593 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9594 * @return 0 on success, non-zero on failure.
9597 mdb_drop0(MDB_cursor *mc, int subs)
9601 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9602 if (rc == MDB_SUCCESS) {
9603 MDB_txn *txn = mc->mc_txn;
9608 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9609 * This also avoids any P_LEAF2 pages, which have no nodes.
9611 if (mc->mc_flags & C_SUB)
9614 mdb_cursor_copy(mc, &mx);
9615 while (mc->mc_snum > 0) {
9616 MDB_page *mp = mc->mc_pg[mc->mc_top];
9617 unsigned n = NUMKEYS(mp);
9619 for (i=0; i<n; i++) {
9620 ni = NODEPTR(mp, i);
9621 if (ni->mn_flags & F_BIGDATA) {
9624 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9625 rc = mdb_page_get(txn, pg, &omp, NULL);
9628 mdb_cassert(mc, IS_OVERFLOW(omp));
9629 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9633 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9634 mdb_xcursor_init1(mc, ni);
9635 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9641 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9643 for (i=0; i<n; i++) {
9645 ni = NODEPTR(mp, i);
9648 mdb_midl_xappend(txn->mt_free_pgs, pg);
9653 mc->mc_ki[mc->mc_top] = i;
9654 rc = mdb_cursor_sibling(mc, 1);
9656 if (rc != MDB_NOTFOUND)
9658 /* no more siblings, go back to beginning
9659 * of previous level.
9663 for (i=1; i<mc->mc_snum; i++) {
9665 mc->mc_pg[i] = mx.mc_pg[i];
9670 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9673 txn->mt_flags |= MDB_TXN_ERROR;
9674 } else if (rc == MDB_NOTFOUND) {
9680 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9682 MDB_cursor *mc, *m2;
9685 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9688 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9691 if (TXN_DBI_CHANGED(txn, dbi))
9694 rc = mdb_cursor_open(txn, dbi, &mc);
9698 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9699 /* Invalidate the dropped DB's cursors */
9700 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9701 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9705 /* Can't delete the main DB */
9706 if (del && dbi >= CORE_DBS) {
9707 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9709 txn->mt_dbflags[dbi] = DB_STALE;
9710 mdb_dbi_close(txn->mt_env, dbi);
9712 txn->mt_flags |= MDB_TXN_ERROR;
9715 /* reset the DB record, mark it dirty */
9716 txn->mt_dbflags[dbi] |= DB_DIRTY;
9717 txn->mt_dbs[dbi].md_depth = 0;
9718 txn->mt_dbs[dbi].md_branch_pages = 0;
9719 txn->mt_dbs[dbi].md_leaf_pages = 0;
9720 txn->mt_dbs[dbi].md_overflow_pages = 0;
9721 txn->mt_dbs[dbi].md_entries = 0;
9722 txn->mt_dbs[dbi].md_root = P_INVALID;
9724 txn->mt_flags |= MDB_TXN_DIRTY;
9727 mdb_cursor_close(mc);
9731 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9733 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9736 txn->mt_dbxs[dbi].md_cmp = cmp;
9740 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9742 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9745 txn->mt_dbxs[dbi].md_dcmp = cmp;
9749 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9751 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9754 txn->mt_dbxs[dbi].md_rel = rel;
9758 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9760 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9763 txn->mt_dbxs[dbi].md_relctx = ctx;
9768 mdb_env_get_maxkeysize(MDB_env *env)
9770 return ENV_MAXKEY(env);
9774 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9776 unsigned int i, rdrs;
9779 int rc = 0, first = 1;
9783 if (!env->me_txns) {
9784 return func("(no reader locks)\n", ctx);
9786 rdrs = env->me_txns->mti_numreaders;
9787 mr = env->me_txns->mti_readers;
9788 for (i=0; i<rdrs; i++) {
9790 txnid_t txnid = mr[i].mr_txnid;
9791 sprintf(buf, txnid == (txnid_t)-1 ?
9792 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9793 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9796 rc = func(" pid thread txnid\n", ctx);
9800 rc = func(buf, ctx);
9806 rc = func("(no active readers)\n", ctx);
9811 /** Insert pid into list if not already present.
9812 * return -1 if already present.
9815 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9817 /* binary search of pid in list */
9819 unsigned cursor = 1;
9821 unsigned n = ids[0];
9824 unsigned pivot = n >> 1;
9825 cursor = base + pivot + 1;
9826 val = pid - ids[cursor];
9831 } else if ( val > 0 ) {
9836 /* found, so it's a duplicate */
9845 for (n = ids[0]; n > cursor; n--)
9852 mdb_reader_check(MDB_env *env, int *dead)
9858 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9861 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9863 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9865 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9866 unsigned int i, j, rdrs;
9868 MDB_PID_T *pids, pid;
9869 int rc = MDB_SUCCESS, count = 0;
9871 rdrs = env->me_txns->mti_numreaders;
9872 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9876 mr = env->me_txns->mti_readers;
9877 for (i=0; i<rdrs; i++) {
9879 if (pid && pid != env->me_pid) {
9880 if (mdb_pid_insert(pids, pid) == 0) {
9881 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9882 /* Stale reader found */
9885 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9886 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9888 rdrs = 0; /* the above checked all readers */
9890 /* Recheck, a new process may have reused pid */
9891 if (mdb_reader_pid(env, Pidcheck, pid))
9896 if (mr[j].mr_pid == pid) {
9897 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9898 (unsigned) pid, mr[j].mr_txnid));
9903 UNLOCK_MUTEX(rmutex);
9914 #ifdef MDB_ROBUST_SUPPORTED
9915 /** Handle #LOCK_MUTEX0() failure.
9916 * Try to repair the lock file if the mutex owner died.
9917 * @param[in] env the environment handle
9918 * @param[in] mutex LOCK_MUTEX0() mutex
9919 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9920 * @return 0 on success with the mutex locked, or an error code on failure.
9923 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9928 if (rc == MDB_OWNERDEAD) {
9929 /* We own the mutex. Clean up after dead previous owner. */
9931 rlocked = (mutex == env->me_rmutex);
9933 /* Keep mti_txnid updated, otherwise next writer can
9934 * overwrite data which latest meta page refers to.
9936 meta = mdb_env_pick_meta(env);
9937 env->me_txns->mti_txnid = meta->mm_txnid;
9938 /* env is hosed if the dead thread was ours */
9940 env->me_flags |= MDB_FATAL_ERROR;
9945 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9946 (rc ? "this process' env is hosed" : "recovering")));
9947 rc2 = mdb_reader_check0(env, rlocked, NULL);
9949 rc2 = mdb_mutex_consistent(mutex);
9950 if (rc || (rc = rc2)) {
9951 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9952 UNLOCK_MUTEX(mutex);
9958 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9963 #endif /* MDB_ROBUST_SUPPORTED */