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")))
213 /** @defgroup internal LMDB Internals
216 /** @defgroup compat Compatibility Macros
217 * A bunch of macros to minimize the amount of platform-specific ifdefs
218 * needed throughout the rest of the code. When the features this library
219 * needs are similar enough to POSIX to be hidden in a one-or-two line
220 * replacement, this macro approach is used.
224 /** Features under development */
229 /** Wrapper around __func__, which is a C99 feature */
230 #if __STDC_VERSION__ >= 199901L
231 # define mdb_func_ __func__
232 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
233 # define mdb_func_ __FUNCTION__
235 /* If a debug message says <mdb_unknown>(), update the #if statements above */
236 # define mdb_func_ "<mdb_unknown>"
239 /* Internal error codes, not exposed outside liblmdb */
240 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
242 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
243 #elif defined MDB_USE_SYSV_SEM
244 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
245 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
246 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
250 #define MDB_ROBUST_SUPPORTED 1
254 #define MDB_USE_HASH 1
255 #define MDB_PIDLOCK 0
256 #define THREAD_RET DWORD
257 #define pthread_t HANDLE
258 #define pthread_mutex_t HANDLE
259 #define pthread_cond_t HANDLE
260 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
261 #define pthread_key_t DWORD
262 #define pthread_self() GetCurrentThreadId()
263 #define pthread_key_create(x,y) \
264 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
265 #define pthread_key_delete(x) TlsFree(x)
266 #define pthread_getspecific(x) TlsGetValue(x)
267 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
268 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
269 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
270 #define pthread_cond_signal(x) SetEvent(*x)
271 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
272 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
273 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
274 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
275 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
276 #define mdb_mutex_consistent(mutex) 0
277 #define getpid() GetCurrentProcessId()
278 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
279 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
280 #define ErrCode() GetLastError()
281 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
282 #define close(fd) (CloseHandle(fd) ? 0 : -1)
283 #define munmap(ptr,len) UnmapViewOfFile(ptr)
284 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
285 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
287 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
291 #define THREAD_RET void *
292 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
293 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
294 #define Z "z" /**< printf format modifier for size_t */
296 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
297 #define MDB_PIDLOCK 1
299 #ifdef MDB_USE_POSIX_SEM
301 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
302 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
303 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
306 mdb_sem_wait(sem_t *sem)
309 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
313 #elif defined MDB_USE_SYSV_SEM
315 typedef struct mdb_mutex {
319 } mdb_mutex_t[1], *mdb_mutexref_t;
321 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
322 #define UNLOCK_MUTEX(mutex) do { \
323 struct sembuf sb = { 0, 1, SEM_UNDO }; \
324 sb.sem_num = (mutex)->semnum; \
325 *(mutex)->locked = 0; \
326 semop((mutex)->semid, &sb, 1); \
330 mdb_sem_wait(mdb_mutexref_t sem)
332 int rc, *locked = sem->locked;
333 struct sembuf sb = { 0, -1, SEM_UNDO };
334 sb.sem_num = sem->semnum;
336 if (!semop(sem->semid, &sb, 1)) {
337 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
341 } while ((rc = errno) == EINTR);
345 #define mdb_mutex_consistent(mutex) 0
347 #else /* MDB_USE_POSIX_MUTEX: */
348 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
349 * local variables keep it (mdb_mutexref_t).
351 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
352 * be the same, or an array[size 1] and a pointer.
355 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
357 /** Lock the reader or writer mutex.
358 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
360 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
361 /** Unlock the reader or writer mutex.
363 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
364 /** Mark mutex-protected data as repaired, after death of previous owner.
366 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
367 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
369 /** Get the error code for the last failed system function.
371 #define ErrCode() errno
373 /** An abstraction for a file handle.
374 * On POSIX systems file handles are small integers. On Windows
375 * they're opaque pointers.
379 /** A value for an invalid file handle.
380 * Mainly used to initialize file variables and signify that they are
383 #define INVALID_HANDLE_VALUE (-1)
385 /** Get the size of a memory page for the system.
386 * This is the basic size that the platform's memory manager uses, and is
387 * fundamental to the use of memory-mapped files.
389 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
392 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
394 #elif defined(MDB_USE_SYSV_SEM)
395 #define MNAME_LEN (sizeof(int))
397 #define MNAME_LEN (sizeof(pthread_mutex_t))
400 #ifdef MDB_USE_SYSV_SEM
401 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
403 #define SYSV_SEM_FLAG 0
408 #ifdef MDB_ROBUST_SUPPORTED
409 /** Lock mutex, handle any error, set rc = result.
410 * Return 0 on success, nonzero (not rc) on error.
412 #define LOCK_MUTEX(rc, env, mutex) \
413 (((rc) = LOCK_MUTEX0(mutex)) && \
414 ((rc) = mdb_mutex_failed(env, mutex, rc)))
415 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
417 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
418 #define mdb_mutex_failed(env, mutex, rc) (rc)
422 /** A flag for opening a file and requesting synchronous data writes.
423 * This is only used when writing a meta page. It's not strictly needed;
424 * we could just do a normal write and then immediately perform a flush.
425 * But if this flag is available it saves us an extra system call.
427 * @note If O_DSYNC is undefined but exists in /usr/include,
428 * preferably set some compiler flag to get the definition.
429 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
432 # define MDB_DSYNC O_DSYNC
436 /** Function for flushing the data of a file. Define this to fsync
437 * if fdatasync() is not supported.
439 #ifndef MDB_FDATASYNC
440 # define MDB_FDATASYNC fdatasync
444 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
455 /** A page number in the database.
456 * Note that 64 bit page numbers are overkill, since pages themselves
457 * already represent 12-13 bits of addressable memory, and the OS will
458 * always limit applications to a maximum of 63 bits of address space.
460 * @note In the #MDB_node structure, we only store 48 bits of this value,
461 * which thus limits us to only 60 bits of addressable data.
463 typedef MDB_ID pgno_t;
465 /** A transaction ID.
466 * See struct MDB_txn.mt_txnid for details.
468 typedef MDB_ID txnid_t;
470 /** @defgroup debug Debug Macros
474 /** Enable debug output. Needs variable argument macros (a C99 feature).
475 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
476 * read from and written to the database (used for free space management).
482 static int mdb_debug;
483 static txnid_t mdb_debug_start;
485 /** Print a debug message with printf formatting.
486 * Requires double parenthesis around 2 or more args.
488 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
489 # define DPRINTF0(fmt, ...) \
490 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
492 # define DPRINTF(args) ((void) 0)
494 /** Print a debug string.
495 * The string is printed literally, with no format processing.
497 #define DPUTS(arg) DPRINTF(("%s", arg))
498 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
500 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
503 /** @brief The maximum size of a database page.
505 * It is 32k or 64k, since value-PAGEBASE must fit in
506 * #MDB_page.%mp_upper.
508 * LMDB will use database pages < OS pages if needed.
509 * That causes more I/O in write transactions: The OS must
510 * know (read) the whole page before writing a partial page.
512 * Note that we don't currently support Huge pages. On Linux,
513 * regular data files cannot use Huge pages, and in general
514 * Huge pages aren't actually pageable. We rely on the OS
515 * demand-pager to read our data and page it out when memory
516 * pressure from other processes is high. So until OSs have
517 * actual paging support for Huge pages, they're not viable.
519 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
521 /** The minimum number of keys required in a database page.
522 * Setting this to a larger value will place a smaller bound on the
523 * maximum size of a data item. Data items larger than this size will
524 * be pushed into overflow pages instead of being stored directly in
525 * the B-tree node. This value used to default to 4. With a page size
526 * of 4096 bytes that meant that any item larger than 1024 bytes would
527 * go into an overflow page. That also meant that on average 2-3KB of
528 * each overflow page was wasted space. The value cannot be lower than
529 * 2 because then there would no longer be a tree structure. With this
530 * value, items larger than 2KB will go into overflow pages, and on
531 * average only 1KB will be wasted.
533 #define MDB_MINKEYS 2
535 /** A stamp that identifies a file as an LMDB file.
536 * There's nothing special about this value other than that it is easily
537 * recognizable, and it will reflect any byte order mismatches.
539 #define MDB_MAGIC 0xBEEFC0DE
541 /** The version number for a database's datafile format. */
542 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
543 /** The version number for a database's lockfile format. */
544 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
546 /** @brief The max size of a key we can write, or 0 for computed max.
548 * This macro should normally be left alone or set to 0.
549 * Note that a database with big keys or dupsort data cannot be
550 * reliably modified by a liblmdb which uses a smaller max.
551 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
553 * Other values are allowed, for backwards compat. However:
554 * A value bigger than the computed max can break if you do not
555 * know what you are doing, and liblmdb <= 0.9.10 can break when
556 * modifying a DB with keys/dupsort data bigger than its max.
558 * Data items in an #MDB_DUPSORT database are also limited to
559 * this size, since they're actually keys of a sub-DB. Keys and
560 * #MDB_DUPSORT data items must fit on a node in a regular page.
562 #ifndef MDB_MAXKEYSIZE
563 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
566 /** The maximum size of a key we can write to the environment. */
568 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
570 #define ENV_MAXKEY(env) ((env)->me_maxkey)
573 /** @brief The maximum size of a data item.
575 * We only store a 32 bit value for node sizes.
577 #define MAXDATASIZE 0xffffffffUL
580 /** Key size which fits in a #DKBUF.
583 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
586 * This is used for printing a hex dump of a key's contents.
588 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
589 /** Display a key in hex.
591 * Invoke a function to display a key in hex.
593 #define DKEY(x) mdb_dkey(x, kbuf)
599 /** An invalid page number.
600 * Mainly used to denote an empty tree.
602 #define P_INVALID (~(pgno_t)0)
604 /** Test if the flags \b f are set in a flag word \b w. */
605 #define F_ISSET(w, f) (((w) & (f)) == (f))
607 /** Round \b n up to an even number. */
608 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
610 /** Used for offsets within a single page.
611 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
614 typedef uint16_t indx_t;
616 /** Default size of memory map.
617 * This is certainly too small for any actual applications. Apps should always set
618 * the size explicitly using #mdb_env_set_mapsize().
620 #define DEFAULT_MAPSIZE 1048576
622 /** @defgroup readers Reader Lock Table
623 * Readers don't acquire any locks for their data access. Instead, they
624 * simply record their transaction ID in the reader table. The reader
625 * mutex is needed just to find an empty slot in the reader table. The
626 * slot's address is saved in thread-specific data so that subsequent read
627 * transactions started by the same thread need no further locking to proceed.
629 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
631 * No reader table is used if the database is on a read-only filesystem, or
632 * if #MDB_NOLOCK is set.
634 * Since the database uses multi-version concurrency control, readers don't
635 * actually need any locking. This table is used to keep track of which
636 * readers are using data from which old transactions, so that we'll know
637 * when a particular old transaction is no longer in use. Old transactions
638 * that have discarded any data pages can then have those pages reclaimed
639 * for use by a later write transaction.
641 * The lock table is constructed such that reader slots are aligned with the
642 * processor's cache line size. Any slot is only ever used by one thread.
643 * This alignment guarantees that there will be no contention or cache
644 * thrashing as threads update their own slot info, and also eliminates
645 * any need for locking when accessing a slot.
647 * A writer thread will scan every slot in the table to determine the oldest
648 * outstanding reader transaction. Any freed pages older than this will be
649 * reclaimed by the writer. The writer doesn't use any locks when scanning
650 * this table. This means that there's no guarantee that the writer will
651 * see the most up-to-date reader info, but that's not required for correct
652 * operation - all we need is to know the upper bound on the oldest reader,
653 * we don't care at all about the newest reader. So the only consequence of
654 * reading stale information here is that old pages might hang around a
655 * while longer before being reclaimed. That's actually good anyway, because
656 * the longer we delay reclaiming old pages, the more likely it is that a
657 * string of contiguous pages can be found after coalescing old pages from
658 * many old transactions together.
661 /** Number of slots in the reader table.
662 * This value was chosen somewhat arbitrarily. 126 readers plus a
663 * couple mutexes fit exactly into 8KB on my development machine.
664 * Applications should set the table size using #mdb_env_set_maxreaders().
666 #define DEFAULT_READERS 126
668 /** The size of a CPU cache line in bytes. We want our lock structures
669 * aligned to this size to avoid false cache line sharing in the
671 * This value works for most CPUs. For Itanium this should be 128.
677 /** The information we store in a single slot of the reader table.
678 * In addition to a transaction ID, we also record the process and
679 * thread ID that owns a slot, so that we can detect stale information,
680 * e.g. threads or processes that went away without cleaning up.
681 * @note We currently don't check for stale records. We simply re-init
682 * the table when we know that we're the only process opening the
685 typedef struct MDB_rxbody {
686 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
687 * Multiple readers that start at the same time will probably have the
688 * same ID here. Again, it's not important to exclude them from
689 * anything; all we need to know is which version of the DB they
690 * started from so we can avoid overwriting any data used in that
691 * particular version.
693 volatile txnid_t mrb_txnid;
694 /** The process ID of the process owning this reader txn. */
695 volatile MDB_PID_T mrb_pid;
696 /** The thread ID of the thread owning this txn. */
697 volatile MDB_THR_T mrb_tid;
700 /** The actual reader record, with cacheline padding. */
701 typedef struct MDB_reader {
704 /** shorthand for mrb_txnid */
705 #define mr_txnid mru.mrx.mrb_txnid
706 #define mr_pid mru.mrx.mrb_pid
707 #define mr_tid mru.mrx.mrb_tid
708 /** cache line alignment */
709 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
713 /** The header for the reader table.
714 * The table resides in a memory-mapped file. (This is a different file
715 * than is used for the main database.)
717 * For POSIX the actual mutexes reside in the shared memory of this
718 * mapped file. On Windows, mutexes are named objects allocated by the
719 * kernel; we store the mutex names in this mapped file so that other
720 * processes can grab them. This same approach is also used on
721 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
722 * process-shared POSIX mutexes. For these cases where a named object
723 * is used, the object name is derived from a 64 bit FNV hash of the
724 * environment pathname. As such, naming collisions are extremely
725 * unlikely. If a collision occurs, the results are unpredictable.
727 typedef struct MDB_txbody {
728 /** Stamp identifying this as an LMDB file. It must be set
731 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
733 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
734 char mtb_rmname[MNAME_LEN];
735 #elif defined(MDB_USE_SYSV_SEM)
739 /** Mutex protecting access to this table.
740 * This is the reader table lock used with LOCK_MUTEX().
742 mdb_mutex_t mtb_rmutex;
744 /** The ID of the last transaction committed to the database.
745 * This is recorded here only for convenience; the value can always
746 * be determined by reading the main database meta pages.
748 volatile txnid_t mtb_txnid;
749 /** The number of slots that have been used in the reader table.
750 * This always records the maximum count, it is not decremented
751 * when readers release their slots.
753 volatile unsigned mtb_numreaders;
756 /** The actual reader table definition. */
757 typedef struct MDB_txninfo {
760 #define mti_magic mt1.mtb.mtb_magic
761 #define mti_format mt1.mtb.mtb_format
762 #define mti_rmutex mt1.mtb.mtb_rmutex
763 #define mti_rmname mt1.mtb.mtb_rmname
764 #define mti_txnid mt1.mtb.mtb_txnid
765 #define mti_numreaders mt1.mtb.mtb_numreaders
766 #ifdef MDB_USE_SYSV_SEM
767 #define mti_semid mt1.mtb.mtb_semid
768 #define mti_rlocked mt1.mtb.mtb_rlocked
770 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
773 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
774 char mt2_wmname[MNAME_LEN];
775 #define mti_wmname mt2.mt2_wmname
776 #elif defined MDB_USE_SYSV_SEM
778 #define mti_wlocked mt2.mt2_wlocked
780 mdb_mutex_t mt2_wmutex;
781 #define mti_wmutex mt2.mt2_wmutex
783 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
785 MDB_reader mti_readers[1];
788 /** Lockfile format signature: version, features and field layout */
789 #define MDB_LOCK_FORMAT \
791 ((MDB_LOCK_VERSION) \
792 /* Flags which describe functionality */ \
793 + (SYSV_SEM_FLAG << 18) \
794 + (((MDB_PIDLOCK) != 0) << 16)))
797 /** Common header for all page types.
798 * Overflow records occupy a number of contiguous pages with no
799 * headers on any page after the first.
801 typedef struct MDB_page {
802 #define mp_pgno mp_p.p_pgno
803 #define mp_next mp_p.p_next
805 pgno_t p_pgno; /**< page number */
806 struct MDB_page *p_next; /**< for in-memory list of freed pages */
809 /** @defgroup mdb_page Page Flags
811 * Flags for the page headers.
814 #define P_BRANCH 0x01 /**< branch page */
815 #define P_LEAF 0x02 /**< leaf page */
816 #define P_OVERFLOW 0x04 /**< overflow page */
817 #define P_META 0x08 /**< meta page */
818 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
819 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
820 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
821 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
822 #define P_KEEP 0x8000 /**< leave this page alone during spill */
824 uint16_t mp_flags; /**< @ref mdb_page */
825 #define mp_lower mp_pb.pb.pb_lower
826 #define mp_upper mp_pb.pb.pb_upper
827 #define mp_pages mp_pb.pb_pages
830 indx_t pb_lower; /**< lower bound of free space */
831 indx_t pb_upper; /**< upper bound of free space */
833 uint32_t pb_pages; /**< number of overflow pages */
835 indx_t mp_ptrs[1]; /**< dynamic size */
838 /** Size of the page header, excluding dynamic data at the end */
839 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
841 /** Address of first usable data byte in a page, after the header */
842 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
844 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
845 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
847 /** Number of nodes on a page */
848 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
850 /** The amount of space remaining in the page */
851 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
853 /** The percentage of space used in the page, in tenths of a percent. */
854 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
855 ((env)->me_psize - PAGEHDRSZ))
856 /** The minimum page fill factor, in tenths of a percent.
857 * Pages emptier than this are candidates for merging.
859 #define FILL_THRESHOLD 250
861 /** Test if a page is a leaf page */
862 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
863 /** Test if a page is a LEAF2 page */
864 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
865 /** Test if a page is a branch page */
866 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
867 /** Test if a page is an overflow page */
868 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
869 /** Test if a page is a sub page */
870 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
872 /** The number of overflow pages needed to store the given size. */
873 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
875 /** Link in #MDB_txn.%mt_loose_pgs list */
876 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
878 /** Header for a single key/data pair within a page.
879 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
880 * We guarantee 2-byte alignment for 'MDB_node's.
882 typedef struct MDB_node {
883 /** lo and hi are used for data size on leaf nodes and for
884 * child pgno on branch nodes. On 64 bit platforms, flags
885 * is also used for pgno. (Branch nodes have no flags).
886 * They are in host byte order in case that lets some
887 * accesses be optimized into a 32-bit word access.
889 #if BYTE_ORDER == LITTLE_ENDIAN
890 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
892 unsigned short mn_hi, mn_lo;
894 /** @defgroup mdb_node Node Flags
896 * Flags for node headers.
899 #define F_BIGDATA 0x01 /**< data put on overflow page */
900 #define F_SUBDATA 0x02 /**< data is a sub-database */
901 #define F_DUPDATA 0x04 /**< data has duplicates */
903 /** valid flags for #mdb_node_add() */
904 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
907 unsigned short mn_flags; /**< @ref mdb_node */
908 unsigned short mn_ksize; /**< key size */
909 char mn_data[1]; /**< key and data are appended here */
912 /** Size of the node header, excluding dynamic data at the end */
913 #define NODESIZE offsetof(MDB_node, mn_data)
915 /** Bit position of top word in page number, for shifting mn_flags */
916 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
918 /** Size of a node in a branch page with a given key.
919 * This is just the node header plus the key, there is no data.
921 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
923 /** Size of a node in a leaf page with a given key and data.
924 * This is node header plus key plus data size.
926 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
928 /** Address of node \b i in page \b p */
929 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
931 /** Address of the key for the node */
932 #define NODEKEY(node) (void *)((node)->mn_data)
934 /** Address of the data for a node */
935 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
937 /** Get the page number pointed to by a branch node */
938 #define NODEPGNO(node) \
939 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
940 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
941 /** Set the page number in a branch node */
942 #define SETPGNO(node,pgno) do { \
943 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
944 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
946 /** Get the size of the data in a leaf node */
947 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
948 /** Set the size of the data for a leaf node */
949 #define SETDSZ(node,size) do { \
950 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
951 /** The size of a key in a node */
952 #define NODEKSZ(node) ((node)->mn_ksize)
954 /** Copy a page number from src to dst */
956 #define COPY_PGNO(dst,src) dst = src
958 #if SIZE_MAX > 4294967295UL
959 #define COPY_PGNO(dst,src) do { \
960 unsigned short *s, *d; \
961 s = (unsigned short *)&(src); \
962 d = (unsigned short *)&(dst); \
969 #define COPY_PGNO(dst,src) do { \
970 unsigned short *s, *d; \
971 s = (unsigned short *)&(src); \
972 d = (unsigned short *)&(dst); \
978 /** The address of a key in a LEAF2 page.
979 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
980 * There are no node headers, keys are stored contiguously.
982 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
984 /** Set the \b node's key into \b keyptr, if requested. */
985 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
986 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
988 /** Set the \b node's key into \b key. */
989 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
991 /** Information about a single database in the environment. */
992 typedef struct MDB_db {
993 uint32_t md_pad; /**< also ksize for LEAF2 pages */
994 uint16_t md_flags; /**< @ref mdb_dbi_open */
995 uint16_t md_depth; /**< depth of this tree */
996 pgno_t md_branch_pages; /**< number of internal pages */
997 pgno_t md_leaf_pages; /**< number of leaf pages */
998 pgno_t md_overflow_pages; /**< number of overflow pages */
999 size_t md_entries; /**< number of data items */
1000 pgno_t md_root; /**< the root page of this tree */
1003 /** mdb_dbi_open flags */
1004 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1005 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1006 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1007 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1009 /** Handle for the DB used to track free pages. */
1011 /** Handle for the default DB. */
1013 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1016 /** Number of meta pages - also hardcoded elsewhere */
1019 /** Meta page content.
1020 * A meta page is the start point for accessing a database snapshot.
1021 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1023 typedef struct MDB_meta {
1024 /** Stamp identifying this as an LMDB file. It must be set
1027 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1028 uint32_t mm_version;
1029 void *mm_address; /**< address for fixed mapping */
1030 size_t mm_mapsize; /**< size of mmap region */
1031 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1032 /** The size of pages used in this DB */
1033 #define mm_psize mm_dbs[FREE_DBI].md_pad
1034 /** Any persistent environment flags. @ref mdb_env */
1035 #define mm_flags mm_dbs[FREE_DBI].md_flags
1036 pgno_t mm_last_pg; /**< last used page in file */
1037 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1040 /** Buffer for a stack-allocated meta page.
1041 * The members define size and alignment, and silence type
1042 * aliasing warnings. They are not used directly; that could
1043 * mean incorrectly using several union members in parallel.
1045 typedef union MDB_metabuf {
1048 char mm_pad[PAGEHDRSZ];
1053 /** Auxiliary DB info.
1054 * The information here is mostly static/read-only. There is
1055 * only a single copy of this record in the environment.
1057 typedef struct MDB_dbx {
1058 MDB_val md_name; /**< name of the database */
1059 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1060 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1061 MDB_rel_func *md_rel; /**< user relocate function */
1062 void *md_relctx; /**< user-provided context for md_rel */
1065 /** A database transaction.
1066 * Every operation requires a transaction handle.
1069 MDB_txn *mt_parent; /**< parent of a nested txn */
1070 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1072 pgno_t mt_next_pgno; /**< next unallocated page */
1073 /** The ID of this transaction. IDs are integers incrementing from 1.
1074 * Only committed write transactions increment the ID. If a transaction
1075 * aborts, the ID may be re-used by the next writer.
1078 MDB_env *mt_env; /**< the DB environment */
1079 /** The list of pages that became unused during this transaction.
1081 MDB_IDL mt_free_pgs;
1082 /** The list of loose pages that became unused and may be reused
1083 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1085 MDB_page *mt_loose_pgs;
1086 /* #Number of loose pages (#mt_loose_pgs) */
1088 /** The sorted list of dirty pages we temporarily wrote to disk
1089 * because the dirty list was full. page numbers in here are
1090 * shifted left by 1, deleted slots have the LSB set.
1092 MDB_IDL mt_spill_pgs;
1094 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1095 MDB_ID2L dirty_list;
1096 /** For read txns: This thread/txn's reader table slot, or NULL. */
1099 /** Array of records for each DB known in the environment. */
1101 /** Array of MDB_db records for each known DB */
1103 /** Array of sequence numbers for each DB handle */
1104 unsigned int *mt_dbiseqs;
1105 /** @defgroup mt_dbflag Transaction DB Flags
1109 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1110 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1111 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1112 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1113 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1115 /** In write txns, array of cursors for each DB */
1116 MDB_cursor **mt_cursors;
1117 /** Array of flags for each DB */
1118 unsigned char *mt_dbflags;
1119 /** Number of DB records in use, or 0 when the txn is finished.
1120 * This number only ever increments until the txn finishes; we
1121 * don't decrement it when individual DB handles are closed.
1125 /** @defgroup mdb_txn Transaction Flags
1129 /** #mdb_txn_begin() flags */
1130 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1131 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1132 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1133 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1134 /* internal txn flags */
1135 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1136 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1137 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1138 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1139 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1140 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1141 /** most operations on the txn are currently illegal */
1142 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1144 unsigned int mt_flags; /**< @ref mdb_txn */
1145 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1146 * Includes ancestor txns' dirty pages not hidden by other txns'
1147 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1148 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1150 unsigned int mt_dirty_room;
1153 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1154 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1155 * raise this on a 64 bit machine.
1157 #define CURSOR_STACK 32
1161 /** Cursors are used for all DB operations.
1162 * A cursor holds a path of (page pointer, key index) from the DB
1163 * root to a position in the DB, plus other state. #MDB_DUPSORT
1164 * cursors include an xcursor to the current data item. Write txns
1165 * track their cursors and keep them up to date when data moves.
1166 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1167 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1170 /** Next cursor on this DB in this txn */
1171 MDB_cursor *mc_next;
1172 /** Backup of the original cursor if this cursor is a shadow */
1173 MDB_cursor *mc_backup;
1174 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1175 struct MDB_xcursor *mc_xcursor;
1176 /** The transaction that owns this cursor */
1178 /** The database handle this cursor operates on */
1180 /** The database record for this cursor */
1182 /** The database auxiliary record for this cursor */
1184 /** The @ref mt_dbflag for this database */
1185 unsigned char *mc_dbflag;
1186 unsigned short mc_snum; /**< number of pushed pages */
1187 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1188 /** @defgroup mdb_cursor Cursor Flags
1190 * Cursor state flags.
1193 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1194 #define C_EOF 0x02 /**< No more data */
1195 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1196 #define C_DEL 0x08 /**< last op was a cursor_del */
1197 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1198 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1200 unsigned int mc_flags; /**< @ref mdb_cursor */
1201 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1202 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1205 /** Context for sorted-dup records.
1206 * We could have gone to a fully recursive design, with arbitrarily
1207 * deep nesting of sub-databases. But for now we only handle these
1208 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1210 typedef struct MDB_xcursor {
1211 /** A sub-cursor for traversing the Dup DB */
1212 MDB_cursor mx_cursor;
1213 /** The database record for this Dup DB */
1215 /** The auxiliary DB record for this Dup DB */
1217 /** The @ref mt_dbflag for this Dup DB */
1218 unsigned char mx_dbflag;
1221 /** State of FreeDB old pages, stored in the MDB_env */
1222 typedef struct MDB_pgstate {
1223 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1224 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1227 /** The database environment. */
1229 HANDLE me_fd; /**< The main data file */
1230 HANDLE me_lfd; /**< The lock file */
1231 HANDLE me_mfd; /**< just for writing the meta pages */
1232 /** Failed to update the meta page. Probably an I/O error. */
1233 #define MDB_FATAL_ERROR 0x80000000U
1234 /** Some fields are initialized. */
1235 #define MDB_ENV_ACTIVE 0x20000000U
1236 /** me_txkey is set */
1237 #define MDB_ENV_TXKEY 0x10000000U
1238 /** fdatasync is unreliable */
1239 #define MDB_FSYNCONLY 0x08000000U
1240 uint32_t me_flags; /**< @ref mdb_env */
1241 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1242 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1243 unsigned int me_maxreaders; /**< size of the reader table */
1244 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1245 volatile int me_close_readers;
1246 MDB_dbi me_numdbs; /**< number of DBs opened */
1247 MDB_dbi me_maxdbs; /**< size of the DB table */
1248 MDB_PID_T me_pid; /**< process ID of this env */
1249 char *me_path; /**< path to the DB files */
1250 char *me_map; /**< the memory map of the data file */
1251 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1252 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1253 void *me_pbuf; /**< scratch area for DUPSORT put() */
1254 MDB_txn *me_txn; /**< current write transaction */
1255 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1256 size_t me_mapsize; /**< size of the data memory map */
1257 off_t me_size; /**< current file size */
1258 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1259 MDB_dbx *me_dbxs; /**< array of static DB info */
1260 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1261 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1262 pthread_key_t me_txkey; /**< thread-key for readers */
1263 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1264 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1265 # define me_pglast me_pgstate.mf_pglast
1266 # define me_pghead me_pgstate.mf_pghead
1267 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1268 /** IDL of pages that became unused in a write txn */
1269 MDB_IDL me_free_pgs;
1270 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1271 MDB_ID2L me_dirty_list;
1272 /** Max number of freelist items that can fit in a single overflow page */
1274 /** Max size of a node on a page */
1275 unsigned int me_nodemax;
1276 #if !(MDB_MAXKEYSIZE)
1277 unsigned int me_maxkey; /**< max size of a key */
1279 int me_live_reader; /**< have liveness lock in reader table */
1281 int me_pidquery; /**< Used in OpenProcess */
1283 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1284 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1285 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1287 mdb_mutex_t me_rmutex;
1288 mdb_mutex_t me_wmutex;
1290 void *me_userctx; /**< User-settable context */
1291 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1294 /** Nested transaction */
1295 typedef struct MDB_ntxn {
1296 MDB_txn mnt_txn; /**< the transaction */
1297 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1300 /** max number of pages to commit in one writev() call */
1301 #define MDB_COMMIT_PAGES 64
1302 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1303 #undef MDB_COMMIT_PAGES
1304 #define MDB_COMMIT_PAGES IOV_MAX
1307 /** max bytes to write in one call */
1308 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1310 /** Check \b txn and \b dbi arguments to a function */
1311 #define TXN_DBI_EXIST(txn, dbi, validity) \
1312 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1314 /** Check for misused \b dbi handles */
1315 #define TXN_DBI_CHANGED(txn, dbi) \
1316 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1318 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1319 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1320 static int mdb_page_touch(MDB_cursor *mc);
1322 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1323 "reset-tmp", "fail-begin", "fail-beginchild"}
1325 /* mdb_txn_end operation number, for logging */
1326 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1327 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1329 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1330 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1331 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1332 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1333 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1335 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1336 static int mdb_page_search_root(MDB_cursor *mc,
1337 MDB_val *key, int modify);
1338 #define MDB_PS_MODIFY 1
1339 #define MDB_PS_ROOTONLY 2
1340 #define MDB_PS_FIRST 4
1341 #define MDB_PS_LAST 8
1342 static int mdb_page_search(MDB_cursor *mc,
1343 MDB_val *key, int flags);
1344 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1346 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1347 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1348 pgno_t newpgno, unsigned int nflags);
1350 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1351 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1352 static int mdb_env_write_meta(MDB_txn *txn);
1353 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1354 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1356 static void mdb_env_close0(MDB_env *env, int excl);
1358 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1359 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1360 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1361 static void mdb_node_del(MDB_cursor *mc, int ksize);
1362 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1363 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1364 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1365 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1366 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1368 static int mdb_rebalance(MDB_cursor *mc);
1369 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1371 static void mdb_cursor_pop(MDB_cursor *mc);
1372 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1374 static int mdb_cursor_del0(MDB_cursor *mc);
1375 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1376 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1377 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1378 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1379 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1381 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1382 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1384 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1385 static void mdb_xcursor_init0(MDB_cursor *mc);
1386 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1387 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_node *node);
1389 static int mdb_drop0(MDB_cursor *mc, int subs);
1390 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1391 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1394 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1397 /** Compare two items pointing at size_t's of unknown alignment. */
1398 #ifdef MISALIGNED_OK
1399 # define mdb_cmp_clong mdb_cmp_long
1401 # define mdb_cmp_clong mdb_cmp_cint
1405 static SECURITY_DESCRIPTOR mdb_null_sd;
1406 static SECURITY_ATTRIBUTES mdb_all_sa;
1407 static int mdb_sec_inited;
1410 /** Return the library version info. */
1412 mdb_version(int *major, int *minor, int *patch)
1414 if (major) *major = MDB_VERSION_MAJOR;
1415 if (minor) *minor = MDB_VERSION_MINOR;
1416 if (patch) *patch = MDB_VERSION_PATCH;
1417 return MDB_VERSION_STRING;
1420 /** Table of descriptions for LMDB @ref errors */
1421 static char *const mdb_errstr[] = {
1422 "MDB_KEYEXIST: Key/data pair already exists",
1423 "MDB_NOTFOUND: No matching key/data pair found",
1424 "MDB_PAGE_NOTFOUND: Requested page not found",
1425 "MDB_CORRUPTED: Located page was wrong type",
1426 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1427 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1428 "MDB_INVALID: File is not an LMDB file",
1429 "MDB_MAP_FULL: Environment mapsize limit reached",
1430 "MDB_DBS_FULL: Environment maxdbs limit reached",
1431 "MDB_READERS_FULL: Environment maxreaders limit reached",
1432 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1433 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1434 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1435 "MDB_PAGE_FULL: Internal error - page has no more space",
1436 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1437 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1438 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1439 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1440 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1441 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1445 mdb_strerror(int err)
1448 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1449 * This works as long as no function between the call to mdb_strerror
1450 * and the actual use of the message uses more than 4K of stack.
1453 char buf[1024], *ptr = buf;
1457 return ("Successful return: 0");
1459 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1460 i = err - MDB_KEYEXIST;
1461 return mdb_errstr[i];
1465 /* These are the C-runtime error codes we use. The comment indicates
1466 * their numeric value, and the Win32 error they would correspond to
1467 * if the error actually came from a Win32 API. A major mess, we should
1468 * have used LMDB-specific error codes for everything.
1471 case ENOENT: /* 2, FILE_NOT_FOUND */
1472 case EIO: /* 5, ACCESS_DENIED */
1473 case ENOMEM: /* 12, INVALID_ACCESS */
1474 case EACCES: /* 13, INVALID_DATA */
1475 case EBUSY: /* 16, CURRENT_DIRECTORY */
1476 case EINVAL: /* 22, BAD_COMMAND */
1477 case ENOSPC: /* 28, OUT_OF_PAPER */
1478 return strerror(err);
1483 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1484 FORMAT_MESSAGE_IGNORE_INSERTS,
1485 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1488 return strerror(err);
1492 /** assert(3) variant in cursor context */
1493 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1494 /** assert(3) variant in transaction context */
1495 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1496 /** assert(3) variant in environment context */
1497 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1500 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1501 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1504 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1505 const char *func, const char *file, int line)
1508 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1509 file, line, expr_txt, func);
1510 if (env->me_assert_func)
1511 env->me_assert_func(env, buf);
1512 fprintf(stderr, "%s\n", buf);
1516 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1520 /** Return the page number of \b mp which may be sub-page, for debug output */
1522 mdb_dbg_pgno(MDB_page *mp)
1525 COPY_PGNO(ret, mp->mp_pgno);
1529 /** Display a key in hexadecimal and return the address of the result.
1530 * @param[in] key the key to display
1531 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1532 * @return The key in hexadecimal form.
1535 mdb_dkey(MDB_val *key, char *buf)
1538 unsigned char *c = key->mv_data;
1544 if (key->mv_size > DKBUF_MAXKEYSIZE)
1545 return "MDB_MAXKEYSIZE";
1546 /* may want to make this a dynamic check: if the key is mostly
1547 * printable characters, print it as-is instead of converting to hex.
1551 for (i=0; i<key->mv_size; i++)
1552 ptr += sprintf(ptr, "%02x", *c++);
1554 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1560 mdb_leafnode_type(MDB_node *n)
1562 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1563 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1564 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1567 /** Display all the keys in the page. */
1569 mdb_page_list(MDB_page *mp)
1571 pgno_t pgno = mdb_dbg_pgno(mp);
1572 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1574 unsigned int i, nkeys, nsize, total = 0;
1578 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1579 case P_BRANCH: type = "Branch page"; break;
1580 case P_LEAF: type = "Leaf page"; break;
1581 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1582 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1583 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1585 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1586 pgno, mp->mp_pages, state);
1589 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1590 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1593 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1597 nkeys = NUMKEYS(mp);
1598 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1600 for (i=0; i<nkeys; i++) {
1601 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1602 key.mv_size = nsize = mp->mp_pad;
1603 key.mv_data = LEAF2KEY(mp, i, nsize);
1605 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1608 node = NODEPTR(mp, i);
1609 key.mv_size = node->mn_ksize;
1610 key.mv_data = node->mn_data;
1611 nsize = NODESIZE + key.mv_size;
1612 if (IS_BRANCH(mp)) {
1613 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1617 if (F_ISSET(node->mn_flags, F_BIGDATA))
1618 nsize += sizeof(pgno_t);
1620 nsize += NODEDSZ(node);
1622 nsize += sizeof(indx_t);
1623 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1624 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1626 total = EVEN(total);
1628 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1629 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1633 mdb_cursor_chk(MDB_cursor *mc)
1639 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1640 for (i=0; i<mc->mc_top; i++) {
1642 node = NODEPTR(mp, mc->mc_ki[i]);
1643 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1646 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1652 /** Count all the pages in each DB and in the freelist
1653 * and make sure it matches the actual number of pages
1655 * All named DBs must be open for a correct count.
1657 static void mdb_audit(MDB_txn *txn)
1661 MDB_ID freecount, count;
1666 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1667 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1668 freecount += *(MDB_ID *)data.mv_data;
1669 mdb_tassert(txn, rc == MDB_NOTFOUND);
1672 for (i = 0; i<txn->mt_numdbs; i++) {
1674 if (!(txn->mt_dbflags[i] & DB_VALID))
1676 mdb_cursor_init(&mc, txn, i, &mx);
1677 if (txn->mt_dbs[i].md_root == P_INVALID)
1679 count += txn->mt_dbs[i].md_branch_pages +
1680 txn->mt_dbs[i].md_leaf_pages +
1681 txn->mt_dbs[i].md_overflow_pages;
1682 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1683 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1684 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1687 mp = mc.mc_pg[mc.mc_top];
1688 for (j=0; j<NUMKEYS(mp); j++) {
1689 MDB_node *leaf = NODEPTR(mp, j);
1690 if (leaf->mn_flags & F_SUBDATA) {
1692 memcpy(&db, NODEDATA(leaf), sizeof(db));
1693 count += db.md_branch_pages + db.md_leaf_pages +
1694 db.md_overflow_pages;
1698 mdb_tassert(txn, rc == MDB_NOTFOUND);
1701 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1702 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1703 txn->mt_txnid, freecount, count+NUM_METAS,
1704 freecount+count+NUM_METAS, txn->mt_next_pgno);
1710 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1712 return txn->mt_dbxs[dbi].md_cmp(a, b);
1716 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1718 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1719 #if UINT_MAX < SIZE_MAX
1720 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1721 dcmp = mdb_cmp_clong;
1726 /** Allocate memory for a page.
1727 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1730 mdb_page_malloc(MDB_txn *txn, unsigned num)
1732 MDB_env *env = txn->mt_env;
1733 MDB_page *ret = env->me_dpages;
1734 size_t psize = env->me_psize, sz = psize, off;
1735 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1736 * For a single page alloc, we init everything after the page header.
1737 * For multi-page, we init the final page; if the caller needed that
1738 * many pages they will be filling in at least up to the last page.
1742 VGMEMP_ALLOC(env, ret, sz);
1743 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1744 env->me_dpages = ret->mp_next;
1747 psize -= off = PAGEHDRSZ;
1752 if ((ret = malloc(sz)) != NULL) {
1753 VGMEMP_ALLOC(env, ret, sz);
1754 if (!(env->me_flags & MDB_NOMEMINIT)) {
1755 memset((char *)ret + off, 0, psize);
1759 txn->mt_flags |= MDB_TXN_ERROR;
1763 /** Free a single page.
1764 * Saves single pages to a list, for future reuse.
1765 * (This is not used for multi-page overflow pages.)
1768 mdb_page_free(MDB_env *env, MDB_page *mp)
1770 mp->mp_next = env->me_dpages;
1771 VGMEMP_FREE(env, mp);
1772 env->me_dpages = mp;
1775 /** Free a dirty page */
1777 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1779 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1780 mdb_page_free(env, dp);
1782 /* large pages just get freed directly */
1783 VGMEMP_FREE(env, dp);
1788 /** Return all dirty pages to dpage list */
1790 mdb_dlist_free(MDB_txn *txn)
1792 MDB_env *env = txn->mt_env;
1793 MDB_ID2L dl = txn->mt_u.dirty_list;
1794 unsigned i, n = dl[0].mid;
1796 for (i = 1; i <= n; i++) {
1797 mdb_dpage_free(env, dl[i].mptr);
1802 /** Loosen or free a single page.
1803 * Saves single pages to a list for future reuse
1804 * in this same txn. It has been pulled from the freeDB
1805 * and already resides on the dirty list, but has been
1806 * deleted. Use these pages first before pulling again
1809 * If the page wasn't dirtied in this txn, just add it
1810 * to this txn's free list.
1813 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1816 pgno_t pgno = mp->mp_pgno;
1817 MDB_txn *txn = mc->mc_txn;
1819 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1820 if (txn->mt_parent) {
1821 MDB_ID2 *dl = txn->mt_u.dirty_list;
1822 /* If txn has a parent, make sure the page is in our
1826 unsigned x = mdb_mid2l_search(dl, pgno);
1827 if (x <= dl[0].mid && dl[x].mid == pgno) {
1828 if (mp != dl[x].mptr) { /* bad cursor? */
1829 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1830 txn->mt_flags |= MDB_TXN_ERROR;
1831 return MDB_CORRUPTED;
1838 /* no parent txn, so it's just ours */
1843 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1845 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1846 txn->mt_loose_pgs = mp;
1847 txn->mt_loose_count++;
1848 mp->mp_flags |= P_LOOSE;
1850 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1858 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1859 * @param[in] mc A cursor handle for the current operation.
1860 * @param[in] pflags Flags of the pages to update:
1861 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1862 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1863 * @return 0 on success, non-zero on failure.
1866 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1868 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1869 MDB_txn *txn = mc->mc_txn;
1875 int rc = MDB_SUCCESS, level;
1877 /* Mark pages seen by cursors */
1878 if (mc->mc_flags & C_UNTRACK)
1879 mc = NULL; /* will find mc in mt_cursors */
1880 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1881 for (; mc; mc=mc->mc_next) {
1882 if (!(mc->mc_flags & C_INITIALIZED))
1884 for (m3 = mc;; m3 = &mx->mx_cursor) {
1886 for (j=0; j<m3->mc_snum; j++) {
1888 if ((mp->mp_flags & Mask) == pflags)
1889 mp->mp_flags ^= P_KEEP;
1891 mx = m3->mc_xcursor;
1892 /* Proceed to mx if it is at a sub-database */
1893 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1895 if (! (mp && (mp->mp_flags & P_LEAF)))
1897 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1898 if (!(leaf->mn_flags & F_SUBDATA))
1907 /* Mark dirty root pages */
1908 for (i=0; i<txn->mt_numdbs; i++) {
1909 if (txn->mt_dbflags[i] & DB_DIRTY) {
1910 pgno_t pgno = txn->mt_dbs[i].md_root;
1911 if (pgno == P_INVALID)
1913 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1915 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1916 dp->mp_flags ^= P_KEEP;
1924 static int mdb_page_flush(MDB_txn *txn, int keep);
1926 /** Spill pages from the dirty list back to disk.
1927 * This is intended to prevent running into #MDB_TXN_FULL situations,
1928 * but note that they may still occur in a few cases:
1929 * 1) our estimate of the txn size could be too small. Currently this
1930 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1931 * 2) child txns may run out of space if their parents dirtied a
1932 * lot of pages and never spilled them. TODO: we probably should do
1933 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1934 * the parent's dirty_room is below a given threshold.
1936 * Otherwise, if not using nested txns, it is expected that apps will
1937 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1938 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1939 * If the txn never references them again, they can be left alone.
1940 * If the txn only reads them, they can be used without any fuss.
1941 * If the txn writes them again, they can be dirtied immediately without
1942 * going thru all of the work of #mdb_page_touch(). Such references are
1943 * handled by #mdb_page_unspill().
1945 * Also note, we never spill DB root pages, nor pages of active cursors,
1946 * because we'll need these back again soon anyway. And in nested txns,
1947 * we can't spill a page in a child txn if it was already spilled in a
1948 * parent txn. That would alter the parent txns' data even though
1949 * the child hasn't committed yet, and we'd have no way to undo it if
1950 * the child aborted.
1952 * @param[in] m0 cursor A cursor handle identifying the transaction and
1953 * database for which we are checking space.
1954 * @param[in] key For a put operation, the key being stored.
1955 * @param[in] data For a put operation, the data being stored.
1956 * @return 0 on success, non-zero on failure.
1959 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1961 MDB_txn *txn = m0->mc_txn;
1963 MDB_ID2L dl = txn->mt_u.dirty_list;
1964 unsigned int i, j, need;
1967 if (m0->mc_flags & C_SUB)
1970 /* Estimate how much space this op will take */
1971 i = m0->mc_db->md_depth;
1972 /* Named DBs also dirty the main DB */
1973 if (m0->mc_dbi >= CORE_DBS)
1974 i += txn->mt_dbs[MAIN_DBI].md_depth;
1975 /* For puts, roughly factor in the key+data size */
1977 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1978 i += i; /* double it for good measure */
1981 if (txn->mt_dirty_room > i)
1984 if (!txn->mt_spill_pgs) {
1985 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1986 if (!txn->mt_spill_pgs)
1989 /* purge deleted slots */
1990 MDB_IDL sl = txn->mt_spill_pgs;
1991 unsigned int num = sl[0];
1993 for (i=1; i<=num; i++) {
2000 /* Preserve pages which may soon be dirtied again */
2001 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2004 /* Less aggressive spill - we originally spilled the entire dirty list,
2005 * with a few exceptions for cursor pages and DB root pages. But this
2006 * turns out to be a lot of wasted effort because in a large txn many
2007 * of those pages will need to be used again. So now we spill only 1/8th
2008 * of the dirty pages. Testing revealed this to be a good tradeoff,
2009 * better than 1/2, 1/4, or 1/10.
2011 if (need < MDB_IDL_UM_MAX / 8)
2012 need = MDB_IDL_UM_MAX / 8;
2014 /* Save the page IDs of all the pages we're flushing */
2015 /* flush from the tail forward, this saves a lot of shifting later on. */
2016 for (i=dl[0].mid; i && need; i--) {
2017 MDB_ID pn = dl[i].mid << 1;
2019 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2021 /* Can't spill twice, make sure it's not already in a parent's
2024 if (txn->mt_parent) {
2026 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2027 if (tx2->mt_spill_pgs) {
2028 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2029 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2030 dp->mp_flags |= P_KEEP;
2038 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2042 mdb_midl_sort(txn->mt_spill_pgs);
2044 /* Flush the spilled part of dirty list */
2045 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2048 /* Reset any dirty pages we kept that page_flush didn't see */
2049 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2052 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2056 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2058 mdb_find_oldest(MDB_txn *txn)
2061 txnid_t mr, oldest = txn->mt_txnid - 1;
2062 if (txn->mt_env->me_txns) {
2063 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2064 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2075 /** Add a page to the txn's dirty list */
2077 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2080 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2082 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2083 insert = mdb_mid2l_append;
2085 insert = mdb_mid2l_insert;
2087 mid.mid = mp->mp_pgno;
2089 rc = insert(txn->mt_u.dirty_list, &mid);
2090 mdb_tassert(txn, rc == 0);
2091 txn->mt_dirty_room--;
2094 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2095 * me_pghead and mt_next_pgno.
2097 * If there are free pages available from older transactions, they
2098 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2099 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2100 * and move me_pglast to say which records were consumed. Only this
2101 * function can create me_pghead and move me_pglast/mt_next_pgno.
2102 * @param[in] mc cursor A cursor handle identifying the transaction and
2103 * database for which we are allocating.
2104 * @param[in] num the number of pages to allocate.
2105 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2106 * will always be satisfied by a single contiguous chunk of memory.
2107 * @return 0 on success, non-zero on failure.
2110 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2112 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2113 /* Get at most <Max_retries> more freeDB records once me_pghead
2114 * has enough pages. If not enough, use new pages from the map.
2115 * If <Paranoid> and mc is updating the freeDB, only get new
2116 * records if me_pghead is empty. Then the freelist cannot play
2117 * catch-up with itself by growing while trying to save it.
2119 enum { Paranoid = 1, Max_retries = 500 };
2121 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2123 int rc, retry = num * 60;
2124 MDB_txn *txn = mc->mc_txn;
2125 MDB_env *env = txn->mt_env;
2126 pgno_t pgno, *mop = env->me_pghead;
2127 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2129 txnid_t oldest = 0, last;
2134 /* If there are any loose pages, just use them */
2135 if (num == 1 && txn->mt_loose_pgs) {
2136 np = txn->mt_loose_pgs;
2137 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2138 txn->mt_loose_count--;
2139 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2147 /* If our dirty list is already full, we can't do anything */
2148 if (txn->mt_dirty_room == 0) {
2153 for (op = MDB_FIRST;; op = MDB_NEXT) {
2158 /* Seek a big enough contiguous page range. Prefer
2159 * pages at the tail, just truncating the list.
2165 if (mop[i-n2] == pgno+n2)
2172 if (op == MDB_FIRST) { /* 1st iteration */
2173 /* Prepare to fetch more and coalesce */
2174 last = env->me_pglast;
2175 oldest = env->me_pgoldest;
2176 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2179 key.mv_data = &last; /* will look up last+1 */
2180 key.mv_size = sizeof(last);
2182 if (Paranoid && mc->mc_dbi == FREE_DBI)
2185 if (Paranoid && retry < 0 && mop_len)
2189 /* Do not fetch more if the record will be too recent */
2190 if (oldest <= last) {
2192 oldest = mdb_find_oldest(txn);
2193 env->me_pgoldest = oldest;
2199 rc = mdb_cursor_get(&m2, &key, NULL, op);
2201 if (rc == MDB_NOTFOUND)
2205 last = *(txnid_t*)key.mv_data;
2206 if (oldest <= last) {
2208 oldest = mdb_find_oldest(txn);
2209 env->me_pgoldest = oldest;
2215 np = m2.mc_pg[m2.mc_top];
2216 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2217 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2220 idl = (MDB_ID *) data.mv_data;
2223 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2228 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2230 mop = env->me_pghead;
2232 env->me_pglast = last;
2234 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2235 last, txn->mt_dbs[FREE_DBI].md_root, i));
2237 DPRINTF(("IDL %"Z"u", idl[j]));
2239 /* Merge in descending sorted order */
2240 mdb_midl_xmerge(mop, idl);
2244 /* Use new pages from the map when nothing suitable in the freeDB */
2246 pgno = txn->mt_next_pgno;
2247 if (pgno + num >= env->me_maxpg) {
2248 DPUTS("DB size maxed out");
2254 if (env->me_flags & MDB_WRITEMAP) {
2255 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2257 if (!(np = mdb_page_malloc(txn, num))) {
2263 mop[0] = mop_len -= num;
2264 /* Move any stragglers down */
2265 for (j = i-num; j < mop_len; )
2266 mop[++j] = mop[++i];
2268 txn->mt_next_pgno = pgno + num;
2271 mdb_page_dirty(txn, np);
2277 txn->mt_flags |= MDB_TXN_ERROR;
2281 /** Copy the used portions of a non-overflow page.
2282 * @param[in] dst page to copy into
2283 * @param[in] src page to copy from
2284 * @param[in] psize size of a page
2287 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2289 enum { Align = sizeof(pgno_t) };
2290 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2292 /* If page isn't full, just copy the used portion. Adjust
2293 * alignment so memcpy may copy words instead of bytes.
2295 if ((unused &= -Align) && !IS_LEAF2(src)) {
2296 upper = (upper + PAGEBASE) & -Align;
2297 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2298 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2301 memcpy(dst, src, psize - unused);
2305 /** Pull a page off the txn's spill list, if present.
2306 * If a page being referenced was spilled to disk in this txn, bring
2307 * it back and make it dirty/writable again.
2308 * @param[in] txn the transaction handle.
2309 * @param[in] mp the page being referenced. It must not be dirty.
2310 * @param[out] ret the writable page, if any. ret is unchanged if
2311 * mp wasn't spilled.
2314 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2316 MDB_env *env = txn->mt_env;
2319 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2321 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2322 if (!tx2->mt_spill_pgs)
2324 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2325 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2328 if (txn->mt_dirty_room == 0)
2329 return MDB_TXN_FULL;
2330 if (IS_OVERFLOW(mp))
2334 if (env->me_flags & MDB_WRITEMAP) {
2337 np = mdb_page_malloc(txn, num);
2341 memcpy(np, mp, num * env->me_psize);
2343 mdb_page_copy(np, mp, env->me_psize);
2346 /* If in current txn, this page is no longer spilled.
2347 * If it happens to be the last page, truncate the spill list.
2348 * Otherwise mark it as deleted by setting the LSB.
2350 if (x == txn->mt_spill_pgs[0])
2351 txn->mt_spill_pgs[0]--;
2353 txn->mt_spill_pgs[x] |= 1;
2354 } /* otherwise, if belonging to a parent txn, the
2355 * page remains spilled until child commits
2358 mdb_page_dirty(txn, np);
2359 np->mp_flags |= P_DIRTY;
2367 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2368 * @param[in] mc cursor pointing to the page to be touched
2369 * @return 0 on success, non-zero on failure.
2372 mdb_page_touch(MDB_cursor *mc)
2374 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2375 MDB_txn *txn = mc->mc_txn;
2376 MDB_cursor *m2, *m3;
2380 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2381 if (txn->mt_flags & MDB_TXN_SPILLS) {
2383 rc = mdb_page_unspill(txn, mp, &np);
2389 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2390 (rc = mdb_page_alloc(mc, 1, &np)))
2393 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2394 mp->mp_pgno, pgno));
2395 mdb_cassert(mc, mp->mp_pgno != pgno);
2396 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2397 /* Update the parent page, if any, to point to the new page */
2399 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2400 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2401 SETPGNO(node, pgno);
2403 mc->mc_db->md_root = pgno;
2405 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2406 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2408 /* If txn has a parent, make sure the page is in our
2412 unsigned x = mdb_mid2l_search(dl, pgno);
2413 if (x <= dl[0].mid && dl[x].mid == pgno) {
2414 if (mp != dl[x].mptr) { /* bad cursor? */
2415 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2416 txn->mt_flags |= MDB_TXN_ERROR;
2417 return MDB_CORRUPTED;
2422 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2424 np = mdb_page_malloc(txn, 1);
2429 rc = mdb_mid2l_insert(dl, &mid);
2430 mdb_cassert(mc, rc == 0);
2435 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2437 np->mp_flags |= P_DIRTY;
2440 /* Adjust cursors pointing to mp */
2441 mc->mc_pg[mc->mc_top] = np;
2442 m2 = txn->mt_cursors[mc->mc_dbi];
2443 if (mc->mc_flags & C_SUB) {
2444 for (; m2; m2=m2->mc_next) {
2445 m3 = &m2->mc_xcursor->mx_cursor;
2446 if (m3->mc_snum < mc->mc_snum) continue;
2447 if (m3->mc_pg[mc->mc_top] == mp)
2448 m3->mc_pg[mc->mc_top] = np;
2451 for (; m2; m2=m2->mc_next) {
2452 if (m2->mc_snum < mc->mc_snum) continue;
2453 if (m2->mc_pg[mc->mc_top] == mp) {
2454 m2->mc_pg[mc->mc_top] = np;
2455 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2457 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2459 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2460 if (!(leaf->mn_flags & F_SUBDATA))
2461 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2469 txn->mt_flags |= MDB_TXN_ERROR;
2474 mdb_env_sync(MDB_env *env, int force)
2477 if (env->me_flags & MDB_RDONLY)
2479 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2480 if (env->me_flags & MDB_WRITEMAP) {
2481 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2482 ? MS_ASYNC : MS_SYNC;
2483 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2486 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2490 #ifdef BROKEN_FDATASYNC
2491 if (env->me_flags & MDB_FSYNCONLY) {
2492 if (fsync(env->me_fd))
2496 if (MDB_FDATASYNC(env->me_fd))
2503 /** Back up parent txn's cursors, then grab the originals for tracking */
2505 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2507 MDB_cursor *mc, *bk;
2512 for (i = src->mt_numdbs; --i >= 0; ) {
2513 if ((mc = src->mt_cursors[i]) != NULL) {
2514 size = sizeof(MDB_cursor);
2516 size += sizeof(MDB_xcursor);
2517 for (; mc; mc = bk->mc_next) {
2523 mc->mc_db = &dst->mt_dbs[i];
2524 /* Kill pointers into src - and dst to reduce abuse: The
2525 * user may not use mc until dst ends. Otherwise we'd...
2527 mc->mc_txn = NULL; /* ...set this to dst */
2528 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2529 if ((mx = mc->mc_xcursor) != NULL) {
2530 *(MDB_xcursor *)(bk+1) = *mx;
2531 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2533 mc->mc_next = dst->mt_cursors[i];
2534 dst->mt_cursors[i] = mc;
2541 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2542 * @param[in] txn the transaction handle.
2543 * @param[in] merge true to keep changes to parent cursors, false to revert.
2544 * @return 0 on success, non-zero on failure.
2547 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2549 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2553 for (i = txn->mt_numdbs; --i >= 0; ) {
2554 for (mc = cursors[i]; mc; mc = next) {
2556 if ((bk = mc->mc_backup) != NULL) {
2558 /* Commit changes to parent txn */
2559 mc->mc_next = bk->mc_next;
2560 mc->mc_backup = bk->mc_backup;
2561 mc->mc_txn = bk->mc_txn;
2562 mc->mc_db = bk->mc_db;
2563 mc->mc_dbflag = bk->mc_dbflag;
2564 if ((mx = mc->mc_xcursor) != NULL)
2565 mx->mx_cursor.mc_txn = bk->mc_txn;
2567 /* Abort nested txn */
2569 if ((mx = mc->mc_xcursor) != NULL)
2570 *mx = *(MDB_xcursor *)(bk+1);
2574 /* Only malloced cursors are permanently tracked. */
2581 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2587 Pidset = F_SETLK, Pidcheck = F_GETLK
2591 /** Set or check a pid lock. Set returns 0 on success.
2592 * Check returns 0 if the process is certainly dead, nonzero if it may
2593 * be alive (the lock exists or an error happened so we do not know).
2595 * On Windows Pidset is a no-op, we merely check for the existence
2596 * of the process with the given pid. On POSIX we use a single byte
2597 * lock on the lockfile, set at an offset equal to the pid.
2600 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2602 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2605 if (op == Pidcheck) {
2606 h = OpenProcess(env->me_pidquery, FALSE, pid);
2607 /* No documented "no such process" code, but other program use this: */
2609 return ErrCode() != ERROR_INVALID_PARAMETER;
2610 /* A process exists until all handles to it close. Has it exited? */
2611 ret = WaitForSingleObject(h, 0) != 0;
2618 struct flock lock_info;
2619 memset(&lock_info, 0, sizeof(lock_info));
2620 lock_info.l_type = F_WRLCK;
2621 lock_info.l_whence = SEEK_SET;
2622 lock_info.l_start = pid;
2623 lock_info.l_len = 1;
2624 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2625 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2627 } else if ((rc = ErrCode()) == EINTR) {
2635 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2636 * @param[in] txn the transaction handle to initialize
2637 * @return 0 on success, non-zero on failure.
2640 mdb_txn_renew0(MDB_txn *txn)
2642 MDB_env *env = txn->mt_env;
2643 MDB_txninfo *ti = env->me_txns;
2645 unsigned int i, nr, flags = txn->mt_flags;
2647 int rc, new_notls = 0;
2649 if ((flags &= MDB_TXN_RDONLY) != 0) {
2651 meta = mdb_env_pick_meta(env);
2652 txn->mt_txnid = meta->mm_txnid;
2653 txn->mt_u.reader = NULL;
2655 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2656 pthread_getspecific(env->me_txkey);
2658 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2659 return MDB_BAD_RSLOT;
2661 MDB_PID_T pid = env->me_pid;
2662 MDB_THR_T tid = pthread_self();
2663 mdb_mutexref_t rmutex = env->me_rmutex;
2665 if (!env->me_live_reader) {
2666 rc = mdb_reader_pid(env, Pidset, pid);
2669 env->me_live_reader = 1;
2672 if (LOCK_MUTEX(rc, env, rmutex))
2674 nr = ti->mti_numreaders;
2675 for (i=0; i<nr; i++)
2676 if (ti->mti_readers[i].mr_pid == 0)
2678 if (i == env->me_maxreaders) {
2679 UNLOCK_MUTEX(rmutex);
2680 return MDB_READERS_FULL;
2682 r = &ti->mti_readers[i];
2683 /* Claim the reader slot, carefully since other code
2684 * uses the reader table un-mutexed: First reset the
2685 * slot, next publish it in mti_numreaders. After
2686 * that, it is safe for mdb_env_close() to touch it.
2687 * When it will be closed, we can finally claim it.
2690 r->mr_txnid = (txnid_t)-1;
2693 ti->mti_numreaders = ++nr;
2694 env->me_close_readers = nr;
2696 UNLOCK_MUTEX(rmutex);
2698 new_notls = (env->me_flags & MDB_NOTLS);
2699 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2704 do /* LY: Retry on a race, ITS#7970. */
2705 r->mr_txnid = ti->mti_txnid;
2706 while(r->mr_txnid != ti->mti_txnid);
2707 txn->mt_txnid = r->mr_txnid;
2708 txn->mt_u.reader = r;
2709 meta = env->me_metas[txn->mt_txnid & 1];
2713 /* Not yet touching txn == env->me_txn0, it may be active */
2715 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2717 txn->mt_txnid = ti->mti_txnid;
2718 meta = env->me_metas[txn->mt_txnid & 1];
2720 meta = mdb_env_pick_meta(env);
2721 txn->mt_txnid = meta->mm_txnid;
2725 if (txn->mt_txnid == mdb_debug_start)
2728 txn->mt_child = NULL;
2729 txn->mt_loose_pgs = NULL;
2730 txn->mt_loose_count = 0;
2731 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2732 txn->mt_u.dirty_list = env->me_dirty_list;
2733 txn->mt_u.dirty_list[0].mid = 0;
2734 txn->mt_free_pgs = env->me_free_pgs;
2735 txn->mt_free_pgs[0] = 0;
2736 txn->mt_spill_pgs = NULL;
2738 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2741 /* Copy the DB info and flags */
2742 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2744 /* Moved to here to avoid a data race in read TXNs */
2745 txn->mt_next_pgno = meta->mm_last_pg+1;
2747 txn->mt_flags = flags;
2750 txn->mt_numdbs = env->me_numdbs;
2751 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2752 x = env->me_dbflags[i];
2753 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2754 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2756 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2757 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2759 if (env->me_flags & MDB_FATAL_ERROR) {
2760 DPUTS("environment had fatal error, must shutdown!");
2762 } else if (env->me_maxpg < txn->mt_next_pgno) {
2763 rc = MDB_MAP_RESIZED;
2767 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2772 mdb_txn_renew(MDB_txn *txn)
2776 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2779 rc = mdb_txn_renew0(txn);
2780 if (rc == MDB_SUCCESS) {
2781 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2782 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2783 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2789 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2793 int rc, size, tsize;
2795 flags &= MDB_TXN_BEGIN_FLAGS;
2796 flags |= env->me_flags & MDB_WRITEMAP;
2798 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2802 /* Nested transactions: Max 1 child, write txns only, no writemap */
2803 flags |= parent->mt_flags;
2804 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2805 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2807 /* Child txns save MDB_pgstate and use own copy of cursors */
2808 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2809 size += tsize = sizeof(MDB_ntxn);
2810 } else if (flags & MDB_RDONLY) {
2811 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2812 size += tsize = sizeof(MDB_txn);
2814 /* Reuse preallocated write txn. However, do not touch it until
2815 * mdb_txn_renew0() succeeds, since it currently may be active.
2820 if ((txn = calloc(1, size)) == NULL) {
2821 DPRINTF(("calloc: %s", strerror(errno)));
2824 txn->mt_dbxs = env->me_dbxs; /* static */
2825 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2826 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2827 txn->mt_flags = flags;
2832 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2833 txn->mt_dbiseqs = parent->mt_dbiseqs;
2834 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2835 if (!txn->mt_u.dirty_list ||
2836 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2838 free(txn->mt_u.dirty_list);
2842 txn->mt_txnid = parent->mt_txnid;
2843 txn->mt_dirty_room = parent->mt_dirty_room;
2844 txn->mt_u.dirty_list[0].mid = 0;
2845 txn->mt_spill_pgs = NULL;
2846 txn->mt_next_pgno = parent->mt_next_pgno;
2847 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2848 parent->mt_child = txn;
2849 txn->mt_parent = parent;
2850 txn->mt_numdbs = parent->mt_numdbs;
2851 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2852 /* Copy parent's mt_dbflags, but clear DB_NEW */
2853 for (i=0; i<txn->mt_numdbs; i++)
2854 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2856 ntxn = (MDB_ntxn *)txn;
2857 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2858 if (env->me_pghead) {
2859 size = MDB_IDL_SIZEOF(env->me_pghead);
2860 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2862 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2867 rc = mdb_cursor_shadow(parent, txn);
2869 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2870 } else { /* MDB_RDONLY */
2871 txn->mt_dbiseqs = env->me_dbiseqs;
2873 rc = mdb_txn_renew0(txn);
2876 if (txn != env->me_txn0)
2879 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2881 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2882 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2883 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2890 mdb_txn_env(MDB_txn *txn)
2892 if(!txn) return NULL;
2897 mdb_txn_id(MDB_txn *txn)
2900 return txn->mt_txnid;
2903 /** Export or close DBI handles opened in this txn. */
2905 mdb_dbis_update(MDB_txn *txn, int keep)
2908 MDB_dbi n = txn->mt_numdbs;
2909 MDB_env *env = txn->mt_env;
2910 unsigned char *tdbflags = txn->mt_dbflags;
2912 for (i = n; --i >= CORE_DBS;) {
2913 if (tdbflags[i] & DB_NEW) {
2915 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2917 char *ptr = env->me_dbxs[i].md_name.mv_data;
2919 env->me_dbxs[i].md_name.mv_data = NULL;
2920 env->me_dbxs[i].md_name.mv_size = 0;
2921 env->me_dbflags[i] = 0;
2922 env->me_dbiseqs[i]++;
2928 if (keep && env->me_numdbs < n)
2932 /** End a transaction, except successful commit of a nested transaction.
2933 * May be called twice for readonly txns: First reset it, then abort.
2934 * @param[in] txn the transaction handle to end
2935 * @param[in] mode why and how to end the transaction
2938 mdb_txn_end(MDB_txn *txn, unsigned mode)
2940 MDB_env *env = txn->mt_env;
2942 static const char *const names[] = MDB_END_NAMES;
2945 /* Export or close DBI handles opened in this txn */
2946 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2948 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2949 names[mode & MDB_END_OPMASK],
2950 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2951 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2953 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2954 if (txn->mt_u.reader) {
2955 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2956 if (!(env->me_flags & MDB_NOTLS)) {
2957 txn->mt_u.reader = NULL; /* txn does not own reader */
2958 } else if (mode & MDB_END_SLOT) {
2959 txn->mt_u.reader->mr_pid = 0;
2960 txn->mt_u.reader = NULL;
2961 } /* else txn owns the slot until it does MDB_END_SLOT */
2963 txn->mt_numdbs = 0; /* prevent further DBI activity */
2964 txn->mt_flags |= MDB_TXN_FINISHED;
2966 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2967 pgno_t *pghead = env->me_pghead;
2969 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2970 mdb_cursors_close(txn, 0);
2971 if (!(env->me_flags & MDB_WRITEMAP)) {
2972 mdb_dlist_free(txn);
2976 txn->mt_flags = MDB_TXN_FINISHED;
2978 if (!txn->mt_parent) {
2979 mdb_midl_shrink(&txn->mt_free_pgs);
2980 env->me_free_pgs = txn->mt_free_pgs;
2982 env->me_pghead = NULL;
2986 mode = 0; /* txn == env->me_txn0, do not free() it */
2988 /* The writer mutex was locked in mdb_txn_begin. */
2990 UNLOCK_MUTEX(env->me_wmutex);
2992 txn->mt_parent->mt_child = NULL;
2993 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2994 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2995 mdb_midl_free(txn->mt_free_pgs);
2996 mdb_midl_free(txn->mt_spill_pgs);
2997 free(txn->mt_u.dirty_list);
3000 mdb_midl_free(pghead);
3003 if (mode & MDB_END_FREE)
3008 mdb_txn_reset(MDB_txn *txn)
3013 /* This call is only valid for read-only txns */
3014 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3017 mdb_txn_end(txn, MDB_END_RESET);
3021 mdb_txn_abort(MDB_txn *txn)
3027 mdb_txn_abort(txn->mt_child);
3029 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3032 /** Save the freelist as of this transaction to the freeDB.
3033 * This changes the freelist. Keep trying until it stabilizes.
3036 mdb_freelist_save(MDB_txn *txn)
3038 /* env->me_pghead[] can grow and shrink during this call.
3039 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3040 * Page numbers cannot disappear from txn->mt_free_pgs[].
3043 MDB_env *env = txn->mt_env;
3044 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3045 txnid_t pglast = 0, head_id = 0;
3046 pgno_t freecnt = 0, *free_pgs, *mop;
3047 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3049 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3051 if (env->me_pghead) {
3052 /* Make sure first page of freeDB is touched and on freelist */
3053 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3054 if (rc && rc != MDB_NOTFOUND)
3058 if (!env->me_pghead && txn->mt_loose_pgs) {
3059 /* Put loose page numbers in mt_free_pgs, since
3060 * we may be unable to return them to me_pghead.
3062 MDB_page *mp = txn->mt_loose_pgs;
3063 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3065 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3066 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3067 txn->mt_loose_pgs = NULL;
3068 txn->mt_loose_count = 0;
3071 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3072 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3073 ? SSIZE_MAX : maxfree_1pg;
3076 /* Come back here after each Put() in case freelist changed */
3081 /* If using records from freeDB which we have not yet
3082 * deleted, delete them and any we reserved for me_pghead.
3084 while (pglast < env->me_pglast) {
3085 rc = mdb_cursor_first(&mc, &key, NULL);
3088 pglast = head_id = *(txnid_t *)key.mv_data;
3089 total_room = head_room = 0;
3090 mdb_tassert(txn, pglast <= env->me_pglast);
3091 rc = mdb_cursor_del(&mc, 0);
3096 /* Save the IDL of pages freed by this txn, to a single record */
3097 if (freecnt < txn->mt_free_pgs[0]) {
3099 /* Make sure last page of freeDB is touched and on freelist */
3100 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3101 if (rc && rc != MDB_NOTFOUND)
3104 free_pgs = txn->mt_free_pgs;
3105 /* Write to last page of freeDB */
3106 key.mv_size = sizeof(txn->mt_txnid);
3107 key.mv_data = &txn->mt_txnid;
3109 freecnt = free_pgs[0];
3110 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3111 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3114 /* Retry if mt_free_pgs[] grew during the Put() */
3115 free_pgs = txn->mt_free_pgs;
3116 } while (freecnt < free_pgs[0]);
3117 mdb_midl_sort(free_pgs);
3118 memcpy(data.mv_data, free_pgs, data.mv_size);
3121 unsigned int i = free_pgs[0];
3122 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3123 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3125 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3131 mop = env->me_pghead;
3132 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3134 /* Reserve records for me_pghead[]. Split it if multi-page,
3135 * to avoid searching freeDB for a page range. Use keys in
3136 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3138 if (total_room >= mop_len) {
3139 if (total_room == mop_len || --more < 0)
3141 } else if (head_room >= maxfree_1pg && head_id > 1) {
3142 /* Keep current record (overflow page), add a new one */
3146 /* (Re)write {key = head_id, IDL length = head_room} */
3147 total_room -= head_room;
3148 head_room = mop_len - total_room;
3149 if (head_room > maxfree_1pg && head_id > 1) {
3150 /* Overflow multi-page for part of me_pghead */
3151 head_room /= head_id; /* amortize page sizes */
3152 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3153 } else if (head_room < 0) {
3154 /* Rare case, not bothering to delete this record */
3157 key.mv_size = sizeof(head_id);
3158 key.mv_data = &head_id;
3159 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3160 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3163 /* IDL is initially empty, zero out at least the length */
3164 pgs = (pgno_t *)data.mv_data;
3165 j = head_room > clean_limit ? head_room : 0;
3169 total_room += head_room;
3172 /* Return loose page numbers to me_pghead, though usually none are
3173 * left at this point. The pages themselves remain in dirty_list.
3175 if (txn->mt_loose_pgs) {
3176 MDB_page *mp = txn->mt_loose_pgs;
3177 unsigned count = txn->mt_loose_count;
3179 /* Room for loose pages + temp IDL with same */
3180 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3182 mop = env->me_pghead;
3183 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3184 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3185 loose[ ++count ] = mp->mp_pgno;
3187 mdb_midl_sort(loose);
3188 mdb_midl_xmerge(mop, loose);
3189 txn->mt_loose_pgs = NULL;
3190 txn->mt_loose_count = 0;
3194 /* Fill in the reserved me_pghead records */
3200 rc = mdb_cursor_first(&mc, &key, &data);
3201 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3202 txnid_t id = *(txnid_t *)key.mv_data;
3203 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3206 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3208 if (len > mop_len) {
3210 data.mv_size = (len + 1) * sizeof(MDB_ID);
3212 data.mv_data = mop -= len;
3215 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3217 if (rc || !(mop_len -= len))
3224 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3225 * @param[in] txn the transaction that's being committed
3226 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3227 * @return 0 on success, non-zero on failure.
3230 mdb_page_flush(MDB_txn *txn, int keep)
3232 MDB_env *env = txn->mt_env;
3233 MDB_ID2L dl = txn->mt_u.dirty_list;
3234 unsigned psize = env->me_psize, j;
3235 int i, pagecount = dl[0].mid, rc;
3236 size_t size = 0, pos = 0;
3238 MDB_page *dp = NULL;
3242 struct iovec iov[MDB_COMMIT_PAGES];
3243 ssize_t wpos = 0, wsize = 0, wres;
3244 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3250 if (env->me_flags & MDB_WRITEMAP) {
3251 /* Clear dirty flags */
3252 while (++i <= pagecount) {
3254 /* Don't flush this page yet */
3255 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3256 dp->mp_flags &= ~P_KEEP;
3260 dp->mp_flags &= ~P_DIRTY;
3265 /* Write the pages */
3267 if (++i <= pagecount) {
3269 /* Don't flush this page yet */
3270 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3271 dp->mp_flags &= ~P_KEEP;
3276 /* clear dirty flag */
3277 dp->mp_flags &= ~P_DIRTY;
3280 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3285 /* Windows actually supports scatter/gather I/O, but only on
3286 * unbuffered file handles. Since we're relying on the OS page
3287 * cache for all our data, that's self-defeating. So we just
3288 * write pages one at a time. We use the ov structure to set
3289 * the write offset, to at least save the overhead of a Seek
3292 DPRINTF(("committing page %"Z"u", pgno));
3293 memset(&ov, 0, sizeof(ov));
3294 ov.Offset = pos & 0xffffffff;
3295 ov.OffsetHigh = pos >> 16 >> 16;
3296 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3298 DPRINTF(("WriteFile: %d", rc));
3302 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3303 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3306 /* Write previous page(s) */
3307 #ifdef MDB_USE_PWRITEV
3308 wres = pwritev(env->me_fd, iov, n, wpos);
3311 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3314 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3318 DPRINTF(("lseek: %s", strerror(rc)));
3321 wres = writev(env->me_fd, iov, n);
3324 if (wres != wsize) {
3329 DPRINTF(("Write error: %s", strerror(rc)));
3331 rc = EIO; /* TODO: Use which error code? */
3332 DPUTS("short write, filesystem full?");
3343 DPRINTF(("committing page %"Z"u", pgno));
3344 next_pos = pos + size;
3345 iov[n].iov_len = size;
3346 iov[n].iov_base = (char *)dp;
3352 /* MIPS has cache coherency issues, this is a no-op everywhere else
3353 * Note: for any size >= on-chip cache size, entire on-chip cache is
3356 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3358 for (i = keep; ++i <= pagecount; ) {
3360 /* This is a page we skipped above */
3363 dl[j].mid = dp->mp_pgno;
3366 mdb_dpage_free(env, dp);
3371 txn->mt_dirty_room += i - j;
3377 mdb_txn_commit(MDB_txn *txn)
3380 unsigned int i, end_mode;
3386 /* mdb_txn_end() mode for a commit which writes nothing */
3387 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3389 if (txn->mt_child) {
3390 rc = mdb_txn_commit(txn->mt_child);
3397 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3401 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3402 DPUTS("txn has failed/finished, can't commit");
3404 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3409 if (txn->mt_parent) {
3410 MDB_txn *parent = txn->mt_parent;
3414 unsigned x, y, len, ps_len;
3416 /* Append our free list to parent's */
3417 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3420 mdb_midl_free(txn->mt_free_pgs);
3421 /* Failures after this must either undo the changes
3422 * to the parent or set MDB_TXN_ERROR in the parent.
3425 parent->mt_next_pgno = txn->mt_next_pgno;
3426 parent->mt_flags = txn->mt_flags;
3428 /* Merge our cursors into parent's and close them */
3429 mdb_cursors_close(txn, 1);
3431 /* Update parent's DB table. */
3432 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3433 parent->mt_numdbs = txn->mt_numdbs;
3434 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3435 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3436 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3437 /* preserve parent's DB_NEW status */
3438 x = parent->mt_dbflags[i] & DB_NEW;
3439 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3442 dst = parent->mt_u.dirty_list;
3443 src = txn->mt_u.dirty_list;
3444 /* Remove anything in our dirty list from parent's spill list */
3445 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3447 pspill[0] = (pgno_t)-1;
3448 /* Mark our dirty pages as deleted in parent spill list */
3449 for (i=0, len=src[0].mid; ++i <= len; ) {
3450 MDB_ID pn = src[i].mid << 1;
3451 while (pn > pspill[x])
3453 if (pn == pspill[x]) {
3458 /* Squash deleted pagenums if we deleted any */
3459 for (x=y; ++x <= ps_len; )
3460 if (!(pspill[x] & 1))
3461 pspill[++y] = pspill[x];
3465 /* Find len = length of merging our dirty list with parent's */
3467 dst[0].mid = 0; /* simplify loops */
3468 if (parent->mt_parent) {
3469 len = x + src[0].mid;
3470 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3471 for (i = x; y && i; y--) {
3472 pgno_t yp = src[y].mid;
3473 while (yp < dst[i].mid)
3475 if (yp == dst[i].mid) {
3480 } else { /* Simplify the above for single-ancestor case */
3481 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3483 /* Merge our dirty list with parent's */
3485 for (i = len; y; dst[i--] = src[y--]) {
3486 pgno_t yp = src[y].mid;
3487 while (yp < dst[x].mid)
3488 dst[i--] = dst[x--];
3489 if (yp == dst[x].mid)
3490 free(dst[x--].mptr);
3492 mdb_tassert(txn, i == x);
3494 free(txn->mt_u.dirty_list);
3495 parent->mt_dirty_room = txn->mt_dirty_room;
3496 if (txn->mt_spill_pgs) {
3497 if (parent->mt_spill_pgs) {
3498 /* TODO: Prevent failure here, so parent does not fail */
3499 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3501 parent->mt_flags |= MDB_TXN_ERROR;
3502 mdb_midl_free(txn->mt_spill_pgs);
3503 mdb_midl_sort(parent->mt_spill_pgs);
3505 parent->mt_spill_pgs = txn->mt_spill_pgs;
3509 /* Append our loose page list to parent's */
3510 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3512 *lp = txn->mt_loose_pgs;
3513 parent->mt_loose_count += txn->mt_loose_count;
3515 parent->mt_child = NULL;
3516 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3521 if (txn != env->me_txn) {
3522 DPUTS("attempt to commit unknown transaction");
3527 mdb_cursors_close(txn, 0);
3529 if (!txn->mt_u.dirty_list[0].mid &&
3530 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3533 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3534 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3536 /* Update DB root pointers */
3537 if (txn->mt_numdbs > CORE_DBS) {
3541 data.mv_size = sizeof(MDB_db);
3543 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3544 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3545 if (txn->mt_dbflags[i] & DB_DIRTY) {
3546 if (TXN_DBI_CHANGED(txn, i)) {
3550 data.mv_data = &txn->mt_dbs[i];
3551 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3559 rc = mdb_freelist_save(txn);
3563 mdb_midl_free(env->me_pghead);
3564 env->me_pghead = NULL;
3565 mdb_midl_shrink(&txn->mt_free_pgs);
3571 if ((rc = mdb_page_flush(txn, 0)))
3573 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3574 (rc = mdb_env_sync(env, 0)))
3576 if ((rc = mdb_env_write_meta(txn)))
3578 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3581 mdb_txn_end(txn, end_mode);
3589 /** Read the environment parameters of a DB environment before
3590 * mapping it into memory.
3591 * @param[in] env the environment handle
3592 * @param[out] meta address of where to store the meta information
3593 * @return 0 on success, non-zero on failure.
3596 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3602 enum { Size = sizeof(pbuf) };
3604 /* We don't know the page size yet, so use a minimum value.
3605 * Read both meta pages so we can use the latest one.
3608 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3612 memset(&ov, 0, sizeof(ov));
3614 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3615 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3618 rc = pread(env->me_fd, &pbuf, Size, off);
3621 if (rc == 0 && off == 0)
3623 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3624 DPRINTF(("read: %s", mdb_strerror(rc)));
3628 p = (MDB_page *)&pbuf;
3630 if (!F_ISSET(p->mp_flags, P_META)) {
3631 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3636 if (m->mm_magic != MDB_MAGIC) {
3637 DPUTS("meta has invalid magic");
3641 if (m->mm_version != MDB_DATA_VERSION) {
3642 DPRINTF(("database is version %u, expected version %u",
3643 m->mm_version, MDB_DATA_VERSION));
3644 return MDB_VERSION_MISMATCH;
3647 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3653 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3655 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3657 meta->mm_magic = MDB_MAGIC;
3658 meta->mm_version = MDB_DATA_VERSION;
3659 meta->mm_mapsize = env->me_mapsize;
3660 meta->mm_psize = env->me_psize;
3661 meta->mm_last_pg = NUM_METAS-1;
3662 meta->mm_flags = env->me_flags & 0xffff;
3663 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3664 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3665 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3668 /** Write the environment parameters of a freshly created DB environment.
3669 * @param[in] env the environment handle
3670 * @param[in] meta the #MDB_meta to write
3671 * @return 0 on success, non-zero on failure.
3674 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3682 memset(&ov, 0, sizeof(ov));
3683 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3685 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3688 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3689 len = pwrite(fd, ptr, size, pos); \
3690 if (len == -1 && ErrCode() == EINTR) continue; \
3691 rc = (len >= 0); break; } while(1)
3694 DPUTS("writing new meta page");
3696 psize = env->me_psize;
3698 p = calloc(NUM_METAS, psize);
3702 p->mp_flags = P_META;
3703 *(MDB_meta *)METADATA(p) = *meta;
3705 q = (MDB_page *)((char *)p + psize);
3707 q->mp_flags = P_META;
3708 *(MDB_meta *)METADATA(q) = *meta;
3710 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3713 else if ((unsigned) len == psize * NUM_METAS)
3721 /** Update the environment info to commit a transaction.
3722 * @param[in] txn the transaction that's being committed
3723 * @return 0 on success, non-zero on failure.
3726 mdb_env_write_meta(MDB_txn *txn)
3729 MDB_meta meta, metab, *mp;
3733 int rc, len, toggle;
3742 toggle = txn->mt_txnid & 1;
3743 DPRINTF(("writing meta page %d for root page %"Z"u",
3744 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3747 flags = txn->mt_flags & env->me_flags;
3748 mp = env->me_metas[toggle];
3749 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3750 /* Persist any increases of mapsize config */
3751 if (mapsize < env->me_mapsize)
3752 mapsize = env->me_mapsize;
3754 if (flags & MDB_WRITEMAP) {
3755 mp->mm_mapsize = mapsize;
3756 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3757 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3758 mp->mm_last_pg = txn->mt_next_pgno - 1;
3759 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3760 !(defined(__i386__) || defined(__x86_64__))
3761 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3762 __sync_synchronize();
3764 mp->mm_txnid = txn->mt_txnid;
3765 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3766 unsigned meta_size = env->me_psize;
3767 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3768 ptr = (char *)mp - PAGEHDRSZ;
3769 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3770 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3774 if (MDB_MSYNC(ptr, meta_size, rc)) {
3781 metab.mm_txnid = mp->mm_txnid;
3782 metab.mm_last_pg = mp->mm_last_pg;
3784 meta.mm_mapsize = mapsize;
3785 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3786 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3787 meta.mm_last_pg = txn->mt_next_pgno - 1;
3788 meta.mm_txnid = txn->mt_txnid;
3790 off = offsetof(MDB_meta, mm_mapsize);
3791 ptr = (char *)&meta + off;
3792 len = sizeof(MDB_meta) - off;
3793 off += (char *)mp - env->me_map;
3795 /* Write to the SYNC fd */
3796 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3799 memset(&ov, 0, sizeof(ov));
3801 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3806 rc = pwrite(mfd, ptr, len, off);
3809 rc = rc < 0 ? ErrCode() : EIO;
3814 DPUTS("write failed, disk error?");
3815 /* On a failure, the pagecache still contains the new data.
3816 * Write some old data back, to prevent it from being used.
3817 * Use the non-SYNC fd; we know it will fail anyway.
3819 meta.mm_last_pg = metab.mm_last_pg;
3820 meta.mm_txnid = metab.mm_txnid;
3822 memset(&ov, 0, sizeof(ov));
3824 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3826 r2 = pwrite(env->me_fd, ptr, len, off);
3827 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3830 env->me_flags |= MDB_FATAL_ERROR;
3833 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3834 CACHEFLUSH(env->me_map + off, len, DCACHE);
3836 /* Memory ordering issues are irrelevant; since the entire writer
3837 * is wrapped by wmutex, all of these changes will become visible
3838 * after the wmutex is unlocked. Since the DB is multi-version,
3839 * readers will get consistent data regardless of how fresh or
3840 * how stale their view of these values is.
3843 env->me_txns->mti_txnid = txn->mt_txnid;
3848 /** Check both meta pages to see which one is newer.
3849 * @param[in] env the environment handle
3850 * @return newest #MDB_meta.
3853 mdb_env_pick_meta(const MDB_env *env)
3855 MDB_meta *const *metas = env->me_metas;
3856 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3860 mdb_env_create(MDB_env **env)
3864 e = calloc(1, sizeof(MDB_env));
3868 e->me_maxreaders = DEFAULT_READERS;
3869 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3870 e->me_fd = INVALID_HANDLE_VALUE;
3871 e->me_lfd = INVALID_HANDLE_VALUE;
3872 e->me_mfd = INVALID_HANDLE_VALUE;
3873 #ifdef MDB_USE_POSIX_SEM
3874 e->me_rmutex = SEM_FAILED;
3875 e->me_wmutex = SEM_FAILED;
3876 #elif defined MDB_USE_SYSV_SEM
3877 e->me_rmutex->semid = -1;
3878 e->me_wmutex->semid = -1;
3880 e->me_pid = getpid();
3881 GET_PAGESIZE(e->me_os_psize);
3882 VGMEMP_CREATE(e,0,0);
3888 mdb_env_map(MDB_env *env, void *addr)
3891 unsigned int flags = env->me_flags;
3895 LONG sizelo, sizehi;
3898 if (flags & MDB_RDONLY) {
3899 /* Don't set explicit map size, use whatever exists */
3904 msize = env->me_mapsize;
3905 sizelo = msize & 0xffffffff;
3906 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3908 /* Windows won't create mappings for zero length files.
3909 * and won't map more than the file size.
3910 * Just set the maxsize right now.
3912 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3913 || !SetEndOfFile(env->me_fd)
3914 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3918 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3919 PAGE_READWRITE : PAGE_READONLY,
3920 sizehi, sizelo, NULL);
3923 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3924 FILE_MAP_WRITE : FILE_MAP_READ,
3926 rc = env->me_map ? 0 : ErrCode();
3931 int prot = PROT_READ;
3932 if (flags & MDB_WRITEMAP) {
3934 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3937 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3939 if (env->me_map == MAP_FAILED) {
3944 if (flags & MDB_NORDAHEAD) {
3945 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3947 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3949 #ifdef POSIX_MADV_RANDOM
3950 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3951 #endif /* POSIX_MADV_RANDOM */
3952 #endif /* MADV_RANDOM */
3956 /* Can happen because the address argument to mmap() is just a
3957 * hint. mmap() can pick another, e.g. if the range is in use.
3958 * The MAP_FIXED flag would prevent that, but then mmap could
3959 * instead unmap existing pages to make room for the new map.
3961 if (addr && env->me_map != addr)
3962 return EBUSY; /* TODO: Make a new MDB_* error code? */
3964 p = (MDB_page *)env->me_map;
3965 env->me_metas[0] = METADATA(p);
3966 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3972 mdb_env_set_mapsize(MDB_env *env, size_t size)
3974 /* If env is already open, caller is responsible for making
3975 * sure there are no active txns.
3983 meta = mdb_env_pick_meta(env);
3985 size = meta->mm_mapsize;
3987 /* Silently round up to minimum if the size is too small */
3988 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3992 munmap(env->me_map, env->me_mapsize);
3993 env->me_mapsize = size;
3994 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3995 rc = mdb_env_map(env, old);
3999 env->me_mapsize = size;
4001 env->me_maxpg = env->me_mapsize / env->me_psize;
4006 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4010 env->me_maxdbs = dbs + CORE_DBS;
4015 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4017 if (env->me_map || readers < 1)
4019 env->me_maxreaders = readers;
4024 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4026 if (!env || !readers)
4028 *readers = env->me_maxreaders;
4033 mdb_fsize(HANDLE fd, size_t *size)
4036 LARGE_INTEGER fsize;
4038 if (!GetFileSizeEx(fd, &fsize))
4041 *size = fsize.QuadPart;
4053 #ifdef BROKEN_FDATASYNC
4054 #include <sys/utsname.h>
4055 #include <sys/vfs.h>
4058 /** Further setup required for opening an LMDB environment
4061 mdb_env_open2(MDB_env *env)
4063 unsigned int flags = env->me_flags;
4064 int i, newenv = 0, rc;
4068 /* See if we should use QueryLimited */
4070 if ((rc & 0xff) > 5)
4071 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4073 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4076 #ifdef BROKEN_FDATASYNC
4077 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4078 * https://lkml.org/lkml/2012/9/3/83
4079 * Kernels after 3.6-rc6 are known good.
4080 * https://lkml.org/lkml/2012/9/10/556
4081 * See if the DB is on ext3/ext4, then check for new enough kernel
4082 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4087 fstatfs(env->me_fd, &st);
4088 while (st.f_type == 0xEF53) {
4092 if (uts.release[0] < '3') {
4093 if (!strncmp(uts.release, "2.6.32.", 7)) {
4094 i = atoi(uts.release+7);
4096 break; /* 2.6.32.60 and newer is OK */
4097 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4098 i = atoi(uts.release+7);
4100 break; /* 2.6.34.15 and newer is OK */
4102 } else if (uts.release[0] == '3') {
4103 i = atoi(uts.release+2);
4105 break; /* 3.6 and newer is OK */
4107 i = atoi(uts.release+4);
4109 break; /* 3.5.4 and newer is OK */
4110 } else if (i == 2) {
4111 i = atoi(uts.release+4);
4113 break; /* 3.2.30 and newer is OK */
4115 } else { /* 4.x and newer is OK */
4118 env->me_flags |= MDB_FSYNCONLY;
4124 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4127 DPUTS("new mdbenv");
4129 env->me_psize = env->me_os_psize;
4130 if (env->me_psize > MAX_PAGESIZE)
4131 env->me_psize = MAX_PAGESIZE;
4132 memset(&meta, 0, sizeof(meta));
4133 mdb_env_init_meta0(env, &meta);
4134 meta.mm_mapsize = DEFAULT_MAPSIZE;
4136 env->me_psize = meta.mm_psize;
4139 /* Was a mapsize configured? */
4140 if (!env->me_mapsize) {
4141 env->me_mapsize = meta.mm_mapsize;
4144 /* Make sure mapsize >= committed data size. Even when using
4145 * mm_mapsize, which could be broken in old files (ITS#7789).
4147 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4148 if (env->me_mapsize < minsize)
4149 env->me_mapsize = minsize;
4151 meta.mm_mapsize = env->me_mapsize;
4153 if (newenv && !(flags & MDB_FIXEDMAP)) {
4154 /* mdb_env_map() may grow the datafile. Write the metapages
4155 * first, so the file will be valid if initialization fails.
4156 * Except with FIXEDMAP, since we do not yet know mm_address.
4157 * We could fill in mm_address later, but then a different
4158 * program might end up doing that - one with a memory layout
4159 * and map address which does not suit the main program.
4161 rc = mdb_env_init_meta(env, &meta);
4167 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4172 if (flags & MDB_FIXEDMAP)
4173 meta.mm_address = env->me_map;
4174 i = mdb_env_init_meta(env, &meta);
4175 if (i != MDB_SUCCESS) {
4180 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4181 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4183 #if !(MDB_MAXKEYSIZE)
4184 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4186 env->me_maxpg = env->me_mapsize / env->me_psize;
4190 MDB_meta *meta = mdb_env_pick_meta(env);
4191 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4193 DPRINTF(("opened database version %u, pagesize %u",
4194 meta->mm_version, env->me_psize));
4195 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4196 DPRINTF(("depth: %u", db->md_depth));
4197 DPRINTF(("entries: %"Z"u", db->md_entries));
4198 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4199 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4200 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4201 DPRINTF(("root: %"Z"u", db->md_root));
4209 /** Release a reader thread's slot in the reader lock table.
4210 * This function is called automatically when a thread exits.
4211 * @param[in] ptr This points to the slot in the reader lock table.
4214 mdb_env_reader_dest(void *ptr)
4216 MDB_reader *reader = ptr;
4222 /** Junk for arranging thread-specific callbacks on Windows. This is
4223 * necessarily platform and compiler-specific. Windows supports up
4224 * to 1088 keys. Let's assume nobody opens more than 64 environments
4225 * in a single process, for now. They can override this if needed.
4227 #ifndef MAX_TLS_KEYS
4228 #define MAX_TLS_KEYS 64
4230 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4231 static int mdb_tls_nkeys;
4233 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4237 case DLL_PROCESS_ATTACH: break;
4238 case DLL_THREAD_ATTACH: break;
4239 case DLL_THREAD_DETACH:
4240 for (i=0; i<mdb_tls_nkeys; i++) {
4241 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4243 mdb_env_reader_dest(r);
4247 case DLL_PROCESS_DETACH: break;
4252 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4254 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4258 /* Force some symbol references.
4259 * _tls_used forces the linker to create the TLS directory if not already done
4260 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4262 #pragma comment(linker, "/INCLUDE:_tls_used")
4263 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4264 #pragma const_seg(".CRT$XLB")
4265 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4266 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4269 #pragma comment(linker, "/INCLUDE:__tls_used")
4270 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4271 #pragma data_seg(".CRT$XLB")
4272 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4274 #endif /* WIN 32/64 */
4275 #endif /* !__GNUC__ */
4278 /** Downgrade the exclusive lock on the region back to shared */
4280 mdb_env_share_locks(MDB_env *env, int *excl)
4283 MDB_meta *meta = mdb_env_pick_meta(env);
4285 env->me_txns->mti_txnid = meta->mm_txnid;
4290 /* First acquire a shared lock. The Unlock will
4291 * then release the existing exclusive lock.
4293 memset(&ov, 0, sizeof(ov));
4294 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4297 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4303 struct flock lock_info;
4304 /* The shared lock replaces the existing lock */
4305 memset((void *)&lock_info, 0, sizeof(lock_info));
4306 lock_info.l_type = F_RDLCK;
4307 lock_info.l_whence = SEEK_SET;
4308 lock_info.l_start = 0;
4309 lock_info.l_len = 1;
4310 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4311 (rc = ErrCode()) == EINTR) ;
4312 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4319 /** Try to get exclusive lock, otherwise shared.
4320 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4323 mdb_env_excl_lock(MDB_env *env, int *excl)
4327 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4331 memset(&ov, 0, sizeof(ov));
4332 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4339 struct flock lock_info;
4340 memset((void *)&lock_info, 0, sizeof(lock_info));
4341 lock_info.l_type = F_WRLCK;
4342 lock_info.l_whence = SEEK_SET;
4343 lock_info.l_start = 0;
4344 lock_info.l_len = 1;
4345 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4346 (rc = ErrCode()) == EINTR) ;
4350 # ifndef MDB_USE_POSIX_MUTEX
4351 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4354 lock_info.l_type = F_RDLCK;
4355 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4356 (rc = ErrCode()) == EINTR) ;
4366 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4368 * @(#) $Revision: 5.1 $
4369 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4370 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4372 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4376 * Please do not copyright this code. This code is in the public domain.
4378 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4379 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4380 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4381 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4382 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4383 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4384 * PERFORMANCE OF THIS SOFTWARE.
4387 * chongo <Landon Curt Noll> /\oo/\
4388 * http://www.isthe.com/chongo/
4390 * Share and Enjoy! :-)
4393 typedef unsigned long long mdb_hash_t;
4394 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4396 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4397 * @param[in] val value to hash
4398 * @param[in] hval initial value for hash
4399 * @return 64 bit hash
4401 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4402 * hval arg on the first call.
4405 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4407 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4408 unsigned char *end = s + val->mv_size;
4410 * FNV-1a hash each octet of the string
4413 /* xor the bottom with the current octet */
4414 hval ^= (mdb_hash_t)*s++;
4416 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4417 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4418 (hval << 7) + (hval << 8) + (hval << 40);
4420 /* return our new hash value */
4424 /** Hash the string and output the encoded hash.
4425 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4426 * very short name limits. We don't care about the encoding being reversible,
4427 * we just want to preserve as many bits of the input as possible in a
4428 * small printable string.
4429 * @param[in] str string to hash
4430 * @param[out] encbuf an array of 11 chars to hold the hash
4432 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4435 mdb_pack85(unsigned long l, char *out)
4439 for (i=0; i<5; i++) {
4440 *out++ = mdb_a85[l % 85];
4446 mdb_hash_enc(MDB_val *val, char *encbuf)
4448 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4450 mdb_pack85(h, encbuf);
4451 mdb_pack85(h>>32, encbuf+5);
4456 /** Open and/or initialize the lock region for the environment.
4457 * @param[in] env The LMDB environment.
4458 * @param[in] lpath The pathname of the file used for the lock region.
4459 * @param[in] mode The Unix permissions for the file, if we create it.
4460 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4461 * @return 0 on success, non-zero on failure.
4464 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4467 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4469 # define MDB_ERRCODE_ROFS EROFS
4470 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4471 # define MDB_CLOEXEC O_CLOEXEC
4474 # define MDB_CLOEXEC 0
4477 #ifdef MDB_USE_SYSV_SEM
4485 env->me_lfd = CreateFileA(lpath, GENERIC_READ|GENERIC_WRITE,
4486 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4487 FILE_ATTRIBUTE_NORMAL, NULL);
4489 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4491 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4493 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4498 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4499 /* Lose record locks when exec*() */
4500 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4501 fcntl(env->me_lfd, F_SETFD, fdflags);
4504 if (!(env->me_flags & MDB_NOTLS)) {
4505 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4508 env->me_flags |= MDB_ENV_TXKEY;
4510 /* Windows TLS callbacks need help finding their TLS info. */
4511 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4515 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4519 /* Try to get exclusive lock. If we succeed, then
4520 * nobody is using the lock region and we should initialize it.
4522 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4525 size = GetFileSize(env->me_lfd, NULL);
4527 size = lseek(env->me_lfd, 0, SEEK_END);
4528 if (size == -1) goto fail_errno;
4530 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4531 if (size < rsize && *excl > 0) {
4533 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4534 || !SetEndOfFile(env->me_lfd))
4537 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4541 size = rsize - sizeof(MDB_txninfo);
4542 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4547 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4549 if (!mh) goto fail_errno;
4550 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4552 if (!env->me_txns) goto fail_errno;
4554 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4556 if (m == MAP_FAILED) goto fail_errno;
4562 BY_HANDLE_FILE_INFORMATION stbuf;
4571 if (!mdb_sec_inited) {
4572 InitializeSecurityDescriptor(&mdb_null_sd,
4573 SECURITY_DESCRIPTOR_REVISION);
4574 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4575 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4576 mdb_all_sa.bInheritHandle = FALSE;
4577 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4580 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4581 idbuf.volume = stbuf.dwVolumeSerialNumber;
4582 idbuf.nhigh = stbuf.nFileIndexHigh;
4583 idbuf.nlow = stbuf.nFileIndexLow;
4584 val.mv_data = &idbuf;
4585 val.mv_size = sizeof(idbuf);
4586 mdb_hash_enc(&val, encbuf);
4587 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4588 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4589 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4590 if (!env->me_rmutex) goto fail_errno;
4591 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4592 if (!env->me_wmutex) goto fail_errno;
4593 #elif defined(MDB_USE_POSIX_SEM)
4602 #if defined(__NetBSD__)
4603 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4605 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4606 idbuf.dev = stbuf.st_dev;
4607 idbuf.ino = stbuf.st_ino;
4608 val.mv_data = &idbuf;
4609 val.mv_size = sizeof(idbuf);
4610 mdb_hash_enc(&val, encbuf);
4611 #ifdef MDB_SHORT_SEMNAMES
4612 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4614 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4615 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4616 /* Clean up after a previous run, if needed: Try to
4617 * remove both semaphores before doing anything else.
4619 sem_unlink(env->me_txns->mti_rmname);
4620 sem_unlink(env->me_txns->mti_wmname);
4621 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4622 O_CREAT|O_EXCL, mode, 1);
4623 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4624 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4625 O_CREAT|O_EXCL, mode, 1);
4626 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4627 #elif defined(MDB_USE_SYSV_SEM)
4628 unsigned short vals[2] = {1, 1};
4629 key_t key = ftok(lpath, 'M');
4632 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4636 if (semctl(semid, 0, SETALL, semu) < 0)
4638 env->me_txns->mti_semid = semid;
4639 #else /* MDB_USE_POSIX_MUTEX: */
4640 pthread_mutexattr_t mattr;
4642 if ((rc = pthread_mutexattr_init(&mattr))
4643 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4644 #ifdef MDB_ROBUST_SUPPORTED
4645 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4647 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4648 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4650 pthread_mutexattr_destroy(&mattr);
4651 #endif /* _WIN32 || ... */
4653 env->me_txns->mti_magic = MDB_MAGIC;
4654 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4655 env->me_txns->mti_txnid = 0;
4656 env->me_txns->mti_numreaders = 0;
4659 #ifdef MDB_USE_SYSV_SEM
4660 struct semid_ds buf;
4662 if (env->me_txns->mti_magic != MDB_MAGIC) {
4663 DPUTS("lock region has invalid magic");
4667 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4668 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4669 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4670 rc = MDB_VERSION_MISMATCH;
4674 if (rc && rc != EACCES && rc != EAGAIN) {
4678 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4679 if (!env->me_rmutex) goto fail_errno;
4680 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4681 if (!env->me_wmutex) goto fail_errno;
4682 #elif defined(MDB_USE_POSIX_SEM)
4683 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4684 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4685 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4686 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4687 #elif defined(MDB_USE_SYSV_SEM)
4688 semid = env->me_txns->mti_semid;
4690 /* check for read access */
4691 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4693 /* check for write access */
4694 if (semctl(semid, 0, IPC_SET, semu) < 0)
4698 #ifdef MDB_USE_SYSV_SEM
4699 env->me_rmutex->semid = semid;
4700 env->me_wmutex->semid = semid;
4701 env->me_rmutex->semnum = 0;
4702 env->me_wmutex->semnum = 1;
4703 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4704 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4715 /** The name of the lock file in the DB environment */
4716 #define LOCKNAME "/lock.mdb"
4717 /** The name of the data file in the DB environment */
4718 #define DATANAME "/data.mdb"
4719 /** The suffix of the lock file when no subdir is used */
4720 #define LOCKSUFF "-lock"
4721 /** Only a subset of the @ref mdb_env flags can be changed
4722 * at runtime. Changing other flags requires closing the
4723 * environment and re-opening it with the new flags.
4725 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4726 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4727 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4729 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4730 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4734 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4736 int oflags, rc, len, excl = -1;
4737 char *lpath, *dpath;
4739 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4743 if (flags & MDB_NOSUBDIR) {
4744 rc = len + sizeof(LOCKSUFF) + len + 1;
4746 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4751 if (flags & MDB_NOSUBDIR) {
4752 dpath = lpath + len + sizeof(LOCKSUFF);
4753 sprintf(lpath, "%s" LOCKSUFF, path);
4754 strcpy(dpath, path);
4756 dpath = lpath + len + sizeof(LOCKNAME);
4757 sprintf(lpath, "%s" LOCKNAME, path);
4758 sprintf(dpath, "%s" DATANAME, path);
4762 flags |= env->me_flags;
4763 if (flags & MDB_RDONLY) {
4764 /* silently ignore WRITEMAP when we're only getting read access */
4765 flags &= ~MDB_WRITEMAP;
4767 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4768 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4771 env->me_flags = flags |= MDB_ENV_ACTIVE;
4775 env->me_path = strdup(path);
4776 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4777 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4778 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4779 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4783 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4785 /* For RDONLY, get lockfile after we know datafile exists */
4786 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4787 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4793 if (F_ISSET(flags, MDB_RDONLY)) {
4794 oflags = GENERIC_READ;
4795 len = OPEN_EXISTING;
4797 oflags = GENERIC_READ|GENERIC_WRITE;
4800 mode = FILE_ATTRIBUTE_NORMAL;
4801 env->me_fd = CreateFileA(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4802 NULL, len, mode, NULL);
4804 if (F_ISSET(flags, MDB_RDONLY))
4807 oflags = O_RDWR | O_CREAT;
4809 env->me_fd = open(dpath, oflags, mode);
4811 if (env->me_fd == INVALID_HANDLE_VALUE) {
4816 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4817 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4822 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4823 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4824 env->me_mfd = env->me_fd;
4826 /* Synchronous fd for meta writes. Needed even with
4827 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4830 len = OPEN_EXISTING;
4831 env->me_mfd = CreateFileA(dpath, oflags,
4832 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4833 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4836 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4838 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4843 DPRINTF(("opened dbenv %p", (void *) env));
4845 rc = mdb_env_share_locks(env, &excl);
4849 if (!(flags & MDB_RDONLY)) {
4851 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4852 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4853 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4854 (txn = calloc(1, size)))
4856 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4857 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4858 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4859 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4861 txn->mt_dbxs = env->me_dbxs;
4862 txn->mt_flags = MDB_TXN_FINISHED;
4872 mdb_env_close0(env, excl);
4878 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4880 mdb_env_close0(MDB_env *env, int excl)
4884 if (!(env->me_flags & MDB_ENV_ACTIVE))
4887 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4889 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4890 free(env->me_dbxs[i].md_name.mv_data);
4895 free(env->me_dbiseqs);
4896 free(env->me_dbflags);
4898 free(env->me_dirty_list);
4900 mdb_midl_free(env->me_free_pgs);
4902 if (env->me_flags & MDB_ENV_TXKEY) {
4903 pthread_key_delete(env->me_txkey);
4905 /* Delete our key from the global list */
4906 for (i=0; i<mdb_tls_nkeys; i++)
4907 if (mdb_tls_keys[i] == env->me_txkey) {
4908 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4916 munmap(env->me_map, env->me_mapsize);
4918 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4919 (void) close(env->me_mfd);
4920 if (env->me_fd != INVALID_HANDLE_VALUE)
4921 (void) close(env->me_fd);
4923 MDB_PID_T pid = env->me_pid;
4924 /* Clearing readers is done in this function because
4925 * me_txkey with its destructor must be disabled first.
4927 * We skip the the reader mutex, so we touch only
4928 * data owned by this process (me_close_readers and
4929 * our readers), and clear each reader atomically.
4931 for (i = env->me_close_readers; --i >= 0; )
4932 if (env->me_txns->mti_readers[i].mr_pid == pid)
4933 env->me_txns->mti_readers[i].mr_pid = 0;
4935 if (env->me_rmutex) {
4936 CloseHandle(env->me_rmutex);
4937 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4939 /* Windows automatically destroys the mutexes when
4940 * the last handle closes.
4942 #elif defined(MDB_USE_POSIX_SEM)
4943 if (env->me_rmutex != SEM_FAILED) {
4944 sem_close(env->me_rmutex);
4945 if (env->me_wmutex != SEM_FAILED)
4946 sem_close(env->me_wmutex);
4947 /* If we have the filelock: If we are the
4948 * only remaining user, clean up semaphores.
4951 mdb_env_excl_lock(env, &excl);
4953 sem_unlink(env->me_txns->mti_rmname);
4954 sem_unlink(env->me_txns->mti_wmname);
4957 #elif defined(MDB_USE_SYSV_SEM)
4958 if (env->me_rmutex->semid != -1) {
4959 /* If we have the filelock: If we are the
4960 * only remaining user, clean up semaphores.
4963 mdb_env_excl_lock(env, &excl);
4965 semctl(env->me_rmutex->semid, 0, IPC_RMID);
4968 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4970 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4973 /* Unlock the lockfile. Windows would have unlocked it
4974 * after closing anyway, but not necessarily at once.
4976 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4979 (void) close(env->me_lfd);
4982 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4986 mdb_env_close(MDB_env *env)
4993 VGMEMP_DESTROY(env);
4994 while ((dp = env->me_dpages) != NULL) {
4995 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4996 env->me_dpages = dp->mp_next;
5000 mdb_env_close0(env, 0);
5004 /** Compare two items pointing at aligned size_t's */
5006 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5008 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5009 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5012 /** Compare two items pointing at aligned unsigned int's.
5014 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5015 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5018 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5020 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5021 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5024 /** Compare two items pointing at unsigned ints of unknown alignment.
5025 * Nodes and keys are guaranteed to be 2-byte aligned.
5028 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5030 #if BYTE_ORDER == LITTLE_ENDIAN
5031 unsigned short *u, *c;
5034 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5035 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5038 } while(!x && u > (unsigned short *)a->mv_data);
5041 unsigned short *u, *c, *end;
5044 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5045 u = (unsigned short *)a->mv_data;
5046 c = (unsigned short *)b->mv_data;
5049 } while(!x && u < end);
5054 /** Compare two items lexically */
5056 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5063 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5069 diff = memcmp(a->mv_data, b->mv_data, len);
5070 return diff ? diff : len_diff<0 ? -1 : len_diff;
5073 /** Compare two items in reverse byte order */
5075 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5077 const unsigned char *p1, *p2, *p1_lim;
5081 p1_lim = (const unsigned char *)a->mv_data;
5082 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5083 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5085 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5091 while (p1 > p1_lim) {
5092 diff = *--p1 - *--p2;
5096 return len_diff<0 ? -1 : len_diff;
5099 /** Search for key within a page, using binary search.
5100 * Returns the smallest entry larger or equal to the key.
5101 * If exactp is non-null, stores whether the found entry was an exact match
5102 * in *exactp (1 or 0).
5103 * Updates the cursor index with the index of the found entry.
5104 * If no entry larger or equal to the key is found, returns NULL.
5107 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5109 unsigned int i = 0, nkeys;
5112 MDB_page *mp = mc->mc_pg[mc->mc_top];
5113 MDB_node *node = NULL;
5118 nkeys = NUMKEYS(mp);
5120 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5121 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5124 low = IS_LEAF(mp) ? 0 : 1;
5126 cmp = mc->mc_dbx->md_cmp;
5128 /* Branch pages have no data, so if using integer keys,
5129 * alignment is guaranteed. Use faster mdb_cmp_int.
5131 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5132 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5139 nodekey.mv_size = mc->mc_db->md_pad;
5140 node = NODEPTR(mp, 0); /* fake */
5141 while (low <= high) {
5142 i = (low + high) >> 1;
5143 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5144 rc = cmp(key, &nodekey);
5145 DPRINTF(("found leaf index %u [%s], rc = %i",
5146 i, DKEY(&nodekey), rc));
5155 while (low <= high) {
5156 i = (low + high) >> 1;
5158 node = NODEPTR(mp, i);
5159 nodekey.mv_size = NODEKSZ(node);
5160 nodekey.mv_data = NODEKEY(node);
5162 rc = cmp(key, &nodekey);
5165 DPRINTF(("found leaf index %u [%s], rc = %i",
5166 i, DKEY(&nodekey), rc));
5168 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5169 i, DKEY(&nodekey), NODEPGNO(node), rc));
5180 if (rc > 0) { /* Found entry is less than the key. */
5181 i++; /* Skip to get the smallest entry larger than key. */
5183 node = NODEPTR(mp, i);
5186 *exactp = (rc == 0 && nkeys > 0);
5187 /* store the key index */
5188 mc->mc_ki[mc->mc_top] = i;
5190 /* There is no entry larger or equal to the key. */
5193 /* nodeptr is fake for LEAF2 */
5199 mdb_cursor_adjust(MDB_cursor *mc, func)
5203 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5204 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5211 /** Pop a page off the top of the cursor's stack. */
5213 mdb_cursor_pop(MDB_cursor *mc)
5216 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5217 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5225 /** Push a page onto the top of the cursor's stack. */
5227 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5229 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5230 DDBI(mc), (void *) mc));
5232 if (mc->mc_snum >= CURSOR_STACK) {
5233 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5234 return MDB_CURSOR_FULL;
5237 mc->mc_top = mc->mc_snum++;
5238 mc->mc_pg[mc->mc_top] = mp;
5239 mc->mc_ki[mc->mc_top] = 0;
5244 /** Find the address of the page corresponding to a given page number.
5245 * @param[in] txn the transaction for this access.
5246 * @param[in] pgno the page number for the page to retrieve.
5247 * @param[out] ret address of a pointer where the page's address will be stored.
5248 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5249 * @return 0 on success, non-zero on failure.
5252 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5254 MDB_env *env = txn->mt_env;
5258 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5262 MDB_ID2L dl = tx2->mt_u.dirty_list;
5264 /* Spilled pages were dirtied in this txn and flushed
5265 * because the dirty list got full. Bring this page
5266 * back in from the map (but don't unspill it here,
5267 * leave that unless page_touch happens again).
5269 if (tx2->mt_spill_pgs) {
5270 MDB_ID pn = pgno << 1;
5271 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5272 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5273 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5278 unsigned x = mdb_mid2l_search(dl, pgno);
5279 if (x <= dl[0].mid && dl[x].mid == pgno) {
5285 } while ((tx2 = tx2->mt_parent) != NULL);
5288 if (pgno < txn->mt_next_pgno) {
5290 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5292 DPRINTF(("page %"Z"u not found", pgno));
5293 txn->mt_flags |= MDB_TXN_ERROR;
5294 return MDB_PAGE_NOTFOUND;
5304 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5305 * The cursor is at the root page, set up the rest of it.
5308 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5310 MDB_page *mp = mc->mc_pg[mc->mc_top];
5314 while (IS_BRANCH(mp)) {
5318 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5319 mdb_cassert(mc, NUMKEYS(mp) > 1);
5320 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5322 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5324 if (flags & MDB_PS_LAST)
5325 i = NUMKEYS(mp) - 1;
5328 node = mdb_node_search(mc, key, &exact);
5330 i = NUMKEYS(mp) - 1;
5332 i = mc->mc_ki[mc->mc_top];
5334 mdb_cassert(mc, i > 0);
5338 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5341 mdb_cassert(mc, i < NUMKEYS(mp));
5342 node = NODEPTR(mp, i);
5344 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5347 mc->mc_ki[mc->mc_top] = i;
5348 if ((rc = mdb_cursor_push(mc, mp)))
5351 if (flags & MDB_PS_MODIFY) {
5352 if ((rc = mdb_page_touch(mc)) != 0)
5354 mp = mc->mc_pg[mc->mc_top];
5359 DPRINTF(("internal error, index points to a %02X page!?",
5361 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5362 return MDB_CORRUPTED;
5365 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5366 key ? DKEY(key) : "null"));
5367 mc->mc_flags |= C_INITIALIZED;
5368 mc->mc_flags &= ~C_EOF;
5373 /** Search for the lowest key under the current branch page.
5374 * This just bypasses a NUMKEYS check in the current page
5375 * before calling mdb_page_search_root(), because the callers
5376 * are all in situations where the current page is known to
5380 mdb_page_search_lowest(MDB_cursor *mc)
5382 MDB_page *mp = mc->mc_pg[mc->mc_top];
5383 MDB_node *node = NODEPTR(mp, 0);
5386 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5389 mc->mc_ki[mc->mc_top] = 0;
5390 if ((rc = mdb_cursor_push(mc, mp)))
5392 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5395 /** Search for the page a given key should be in.
5396 * Push it and its parent pages on the cursor stack.
5397 * @param[in,out] mc the cursor for this operation.
5398 * @param[in] key the key to search for, or NULL for first/last page.
5399 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5400 * are touched (updated with new page numbers).
5401 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5402 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5403 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5404 * @return 0 on success, non-zero on failure.
5407 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5412 /* Make sure the txn is still viable, then find the root from
5413 * the txn's db table and set it as the root of the cursor's stack.
5415 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5416 DPUTS("transaction may not be used now");
5419 /* Make sure we're using an up-to-date root */
5420 if (*mc->mc_dbflag & DB_STALE) {
5422 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5424 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5425 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5432 MDB_node *leaf = mdb_node_search(&mc2,
5433 &mc->mc_dbx->md_name, &exact);
5435 return MDB_NOTFOUND;
5436 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5437 return MDB_INCOMPATIBLE; /* not a named DB */
5438 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5441 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5443 /* The txn may not know this DBI, or another process may
5444 * have dropped and recreated the DB with other flags.
5446 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5447 return MDB_INCOMPATIBLE;
5448 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5450 *mc->mc_dbflag &= ~DB_STALE;
5452 root = mc->mc_db->md_root;
5454 if (root == P_INVALID) { /* Tree is empty. */
5455 DPUTS("tree is empty");
5456 return MDB_NOTFOUND;
5460 mdb_cassert(mc, root > 1);
5461 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5462 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5468 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5469 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5471 if (flags & MDB_PS_MODIFY) {
5472 if ((rc = mdb_page_touch(mc)))
5476 if (flags & MDB_PS_ROOTONLY)
5479 return mdb_page_search_root(mc, key, flags);
5483 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5485 MDB_txn *txn = mc->mc_txn;
5486 pgno_t pg = mp->mp_pgno;
5487 unsigned x = 0, ovpages = mp->mp_pages;
5488 MDB_env *env = txn->mt_env;
5489 MDB_IDL sl = txn->mt_spill_pgs;
5490 MDB_ID pn = pg << 1;
5493 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5494 /* If the page is dirty or on the spill list we just acquired it,
5495 * so we should give it back to our current free list, if any.
5496 * Otherwise put it onto the list of pages we freed in this txn.
5498 * Won't create me_pghead: me_pglast must be inited along with it.
5499 * Unsupported in nested txns: They would need to hide the page
5500 * range in ancestor txns' dirty and spilled lists.
5502 if (env->me_pghead &&
5504 ((mp->mp_flags & P_DIRTY) ||
5505 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5509 MDB_ID2 *dl, ix, iy;
5510 rc = mdb_midl_need(&env->me_pghead, ovpages);
5513 if (!(mp->mp_flags & P_DIRTY)) {
5514 /* This page is no longer spilled */
5521 /* Remove from dirty list */
5522 dl = txn->mt_u.dirty_list;
5524 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5530 mdb_cassert(mc, x > 1);
5532 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5533 txn->mt_flags |= MDB_TXN_ERROR;
5534 return MDB_CORRUPTED;
5537 if (!(env->me_flags & MDB_WRITEMAP))
5538 mdb_dpage_free(env, mp);
5540 /* Insert in me_pghead */
5541 mop = env->me_pghead;
5542 j = mop[0] + ovpages;
5543 for (i = mop[0]; i && mop[i] < pg; i--)
5549 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5553 mc->mc_db->md_overflow_pages -= ovpages;
5557 /** Return the data associated with a given node.
5558 * @param[in] txn The transaction for this operation.
5559 * @param[in] leaf The node being read.
5560 * @param[out] data Updated to point to the node's data.
5561 * @return 0 on success, non-zero on failure.
5564 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5566 MDB_page *omp; /* overflow page */
5570 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5571 data->mv_size = NODEDSZ(leaf);
5572 data->mv_data = NODEDATA(leaf);
5576 /* Read overflow data.
5578 data->mv_size = NODEDSZ(leaf);
5579 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5580 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5581 DPRINTF(("read overflow page %"Z"u failed", pgno));
5584 data->mv_data = METADATA(omp);
5590 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5591 MDB_val *key, MDB_val *data)
5598 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5600 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5603 if (txn->mt_flags & MDB_TXN_BLOCKED)
5606 mdb_cursor_init(&mc, txn, dbi, &mx);
5607 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5610 /** Find a sibling for a page.
5611 * Replaces the page at the top of the cursor's stack with the
5612 * specified sibling, if one exists.
5613 * @param[in] mc The cursor for this operation.
5614 * @param[in] move_right Non-zero if the right sibling is requested,
5615 * otherwise the left sibling.
5616 * @return 0 on success, non-zero on failure.
5619 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5625 if (mc->mc_snum < 2) {
5626 return MDB_NOTFOUND; /* root has no siblings */
5630 DPRINTF(("parent page is page %"Z"u, index %u",
5631 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5633 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5634 : (mc->mc_ki[mc->mc_top] == 0)) {
5635 DPRINTF(("no more keys left, moving to %s sibling",
5636 move_right ? "right" : "left"));
5637 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5638 /* undo cursor_pop before returning */
5645 mc->mc_ki[mc->mc_top]++;
5647 mc->mc_ki[mc->mc_top]--;
5648 DPRINTF(("just moving to %s index key %u",
5649 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5651 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5653 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5654 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5655 /* mc will be inconsistent if caller does mc_snum++ as above */
5656 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5660 mdb_cursor_push(mc, mp);
5662 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5667 /** Move the cursor to the next data item. */
5669 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5675 if (mc->mc_flags & C_EOF) {
5676 return MDB_NOTFOUND;
5679 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5681 mp = mc->mc_pg[mc->mc_top];
5683 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5684 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5685 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5686 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5687 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5688 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5689 if (rc == MDB_SUCCESS)
5690 MDB_GET_KEY(leaf, key);
5695 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5696 if (op == MDB_NEXT_DUP)
5697 return MDB_NOTFOUND;
5701 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5702 mdb_dbg_pgno(mp), (void *) mc));
5703 if (mc->mc_flags & C_DEL)
5706 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5707 DPUTS("=====> move to next sibling page");
5708 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5709 mc->mc_flags |= C_EOF;
5712 mp = mc->mc_pg[mc->mc_top];
5713 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5715 mc->mc_ki[mc->mc_top]++;
5718 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5719 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5722 key->mv_size = mc->mc_db->md_pad;
5723 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5727 mdb_cassert(mc, IS_LEAF(mp));
5728 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5730 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5731 mdb_xcursor_init1(mc, leaf);
5734 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5737 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5738 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5739 if (rc != MDB_SUCCESS)
5744 MDB_GET_KEY(leaf, key);
5748 /** Move the cursor to the previous data item. */
5750 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5756 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5758 mp = mc->mc_pg[mc->mc_top];
5760 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5761 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5762 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5763 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5764 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5765 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5766 if (rc == MDB_SUCCESS) {
5767 MDB_GET_KEY(leaf, key);
5768 mc->mc_flags &= ~C_EOF;
5774 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5775 if (op == MDB_PREV_DUP)
5776 return MDB_NOTFOUND;
5780 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5781 mdb_dbg_pgno(mp), (void *) mc));
5783 if (mc->mc_ki[mc->mc_top] == 0) {
5784 DPUTS("=====> move to prev sibling page");
5785 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5788 mp = mc->mc_pg[mc->mc_top];
5789 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5790 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5792 mc->mc_ki[mc->mc_top]--;
5794 mc->mc_flags &= ~C_EOF;
5796 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5797 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5800 key->mv_size = mc->mc_db->md_pad;
5801 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5805 mdb_cassert(mc, IS_LEAF(mp));
5806 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5808 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5809 mdb_xcursor_init1(mc, leaf);
5812 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5815 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5816 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5817 if (rc != MDB_SUCCESS)
5822 MDB_GET_KEY(leaf, key);
5826 /** Set the cursor on a specific data item. */
5828 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5829 MDB_cursor_op op, int *exactp)
5833 MDB_node *leaf = NULL;
5836 if (key->mv_size == 0)
5837 return MDB_BAD_VALSIZE;
5840 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5842 /* See if we're already on the right page */
5843 if (mc->mc_flags & C_INITIALIZED) {
5846 mp = mc->mc_pg[mc->mc_top];
5848 mc->mc_ki[mc->mc_top] = 0;
5849 return MDB_NOTFOUND;
5851 if (mp->mp_flags & P_LEAF2) {
5852 nodekey.mv_size = mc->mc_db->md_pad;
5853 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5855 leaf = NODEPTR(mp, 0);
5856 MDB_GET_KEY2(leaf, nodekey);
5858 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5860 /* Probably happens rarely, but first node on the page
5861 * was the one we wanted.
5863 mc->mc_ki[mc->mc_top] = 0;
5870 unsigned int nkeys = NUMKEYS(mp);
5872 if (mp->mp_flags & P_LEAF2) {
5873 nodekey.mv_data = LEAF2KEY(mp,
5874 nkeys-1, nodekey.mv_size);
5876 leaf = NODEPTR(mp, nkeys-1);
5877 MDB_GET_KEY2(leaf, nodekey);
5879 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5881 /* last node was the one we wanted */
5882 mc->mc_ki[mc->mc_top] = nkeys-1;
5888 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5889 /* This is definitely the right page, skip search_page */
5890 if (mp->mp_flags & P_LEAF2) {
5891 nodekey.mv_data = LEAF2KEY(mp,
5892 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5894 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5895 MDB_GET_KEY2(leaf, nodekey);
5897 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5899 /* current node was the one we wanted */
5909 /* If any parents have right-sibs, search.
5910 * Otherwise, there's nothing further.
5912 for (i=0; i<mc->mc_top; i++)
5914 NUMKEYS(mc->mc_pg[i])-1)
5916 if (i == mc->mc_top) {
5917 /* There are no other pages */
5918 mc->mc_ki[mc->mc_top] = nkeys;
5919 return MDB_NOTFOUND;
5923 /* There are no other pages */
5924 mc->mc_ki[mc->mc_top] = 0;
5925 if (op == MDB_SET_RANGE && !exactp) {
5929 return MDB_NOTFOUND;
5933 rc = mdb_page_search(mc, key, 0);
5934 if (rc != MDB_SUCCESS)
5937 mp = mc->mc_pg[mc->mc_top];
5938 mdb_cassert(mc, IS_LEAF(mp));
5941 leaf = mdb_node_search(mc, key, exactp);
5942 if (exactp != NULL && !*exactp) {
5943 /* MDB_SET specified and not an exact match. */
5944 return MDB_NOTFOUND;
5948 DPUTS("===> inexact leaf not found, goto sibling");
5949 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5950 mc->mc_flags |= C_EOF;
5951 return rc; /* no entries matched */
5953 mp = mc->mc_pg[mc->mc_top];
5954 mdb_cassert(mc, IS_LEAF(mp));
5955 leaf = NODEPTR(mp, 0);
5959 mc->mc_flags |= C_INITIALIZED;
5960 mc->mc_flags &= ~C_EOF;
5963 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5964 key->mv_size = mc->mc_db->md_pad;
5965 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5970 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5971 mdb_xcursor_init1(mc, leaf);
5974 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5975 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5976 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5979 if (op == MDB_GET_BOTH) {
5985 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5986 if (rc != MDB_SUCCESS)
5989 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5992 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5994 dcmp = mc->mc_dbx->md_dcmp;
5995 #if UINT_MAX < SIZE_MAX
5996 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5997 dcmp = mdb_cmp_clong;
5999 rc = dcmp(data, &olddata);
6001 if (op == MDB_GET_BOTH || rc > 0)
6002 return MDB_NOTFOUND;
6009 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6010 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6015 /* The key already matches in all other cases */
6016 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6017 MDB_GET_KEY(leaf, key);
6018 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6023 /** Move the cursor to the first item in the database. */
6025 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6031 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6033 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6034 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6035 if (rc != MDB_SUCCESS)
6038 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6040 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6041 mc->mc_flags |= C_INITIALIZED;
6042 mc->mc_flags &= ~C_EOF;
6044 mc->mc_ki[mc->mc_top] = 0;
6046 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6047 key->mv_size = mc->mc_db->md_pad;
6048 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6053 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6054 mdb_xcursor_init1(mc, leaf);
6055 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6059 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6063 MDB_GET_KEY(leaf, key);
6067 /** Move the cursor to the last item in the database. */
6069 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6075 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6077 if (!(mc->mc_flags & C_EOF)) {
6079 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6080 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6081 if (rc != MDB_SUCCESS)
6084 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6087 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6088 mc->mc_flags |= C_INITIALIZED|C_EOF;
6089 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6091 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6092 key->mv_size = mc->mc_db->md_pad;
6093 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6098 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6099 mdb_xcursor_init1(mc, leaf);
6100 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6104 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6109 MDB_GET_KEY(leaf, key);
6114 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6119 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6124 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6128 case MDB_GET_CURRENT:
6129 if (!(mc->mc_flags & C_INITIALIZED)) {
6132 MDB_page *mp = mc->mc_pg[mc->mc_top];
6133 int nkeys = NUMKEYS(mp);
6134 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6135 mc->mc_ki[mc->mc_top] = nkeys;
6141 key->mv_size = mc->mc_db->md_pad;
6142 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6144 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6145 MDB_GET_KEY(leaf, key);
6147 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6148 if (mc->mc_flags & C_DEL)
6149 mdb_xcursor_init1(mc, leaf);
6150 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6152 rc = mdb_node_read(mc->mc_txn, leaf, data);
6159 case MDB_GET_BOTH_RANGE:
6164 if (mc->mc_xcursor == NULL) {
6165 rc = MDB_INCOMPATIBLE;
6175 rc = mdb_cursor_set(mc, key, data, op,
6176 op == MDB_SET_RANGE ? NULL : &exact);
6179 case MDB_GET_MULTIPLE:
6180 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6184 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6185 rc = MDB_INCOMPATIBLE;
6189 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6190 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6193 case MDB_NEXT_MULTIPLE:
6198 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6199 rc = MDB_INCOMPATIBLE;
6202 if (!(mc->mc_flags & C_INITIALIZED))
6203 rc = mdb_cursor_first(mc, key, data);
6205 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6206 if (rc == MDB_SUCCESS) {
6207 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6210 mx = &mc->mc_xcursor->mx_cursor;
6211 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6213 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6214 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6222 case MDB_NEXT_NODUP:
6223 if (!(mc->mc_flags & C_INITIALIZED))
6224 rc = mdb_cursor_first(mc, key, data);
6226 rc = mdb_cursor_next(mc, key, data, op);
6230 case MDB_PREV_NODUP:
6231 if (!(mc->mc_flags & C_INITIALIZED)) {
6232 rc = mdb_cursor_last(mc, key, data);
6235 mc->mc_flags |= C_INITIALIZED;
6236 mc->mc_ki[mc->mc_top]++;
6238 rc = mdb_cursor_prev(mc, key, data, op);
6241 rc = mdb_cursor_first(mc, key, data);
6244 mfunc = mdb_cursor_first;
6246 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6250 if (mc->mc_xcursor == NULL) {
6251 rc = MDB_INCOMPATIBLE;
6255 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6256 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6257 MDB_GET_KEY(leaf, key);
6258 rc = mdb_node_read(mc->mc_txn, leaf, data);
6262 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6266 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6269 rc = mdb_cursor_last(mc, key, data);
6272 mfunc = mdb_cursor_last;
6275 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6280 if (mc->mc_flags & C_DEL)
6281 mc->mc_flags ^= C_DEL;
6286 /** Touch all the pages in the cursor stack. Set mc_top.
6287 * Makes sure all the pages are writable, before attempting a write operation.
6288 * @param[in] mc The cursor to operate on.
6291 mdb_cursor_touch(MDB_cursor *mc)
6293 int rc = MDB_SUCCESS;
6295 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6298 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6300 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6301 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6304 *mc->mc_dbflag |= DB_DIRTY;
6309 rc = mdb_page_touch(mc);
6310 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6311 mc->mc_top = mc->mc_snum-1;
6316 /** Do not spill pages to disk if txn is getting full, may fail instead */
6317 #define MDB_NOSPILL 0x8000
6320 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6324 MDB_node *leaf = NULL;
6327 MDB_val xdata, *rdata, dkey, olddata;
6329 int do_sub = 0, insert_key, insert_data;
6330 unsigned int mcount = 0, dcount = 0, nospill;
6333 unsigned int nflags;
6336 if (mc == NULL || key == NULL)
6339 env = mc->mc_txn->mt_env;
6341 /* Check this first so counter will always be zero on any
6344 if (flags & MDB_MULTIPLE) {
6345 dcount = data[1].mv_size;
6346 data[1].mv_size = 0;
6347 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6348 return MDB_INCOMPATIBLE;
6351 nospill = flags & MDB_NOSPILL;
6352 flags &= ~MDB_NOSPILL;
6354 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6355 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6357 if (key->mv_size-1 >= ENV_MAXKEY(env))
6358 return MDB_BAD_VALSIZE;
6360 #if SIZE_MAX > MAXDATASIZE
6361 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6362 return MDB_BAD_VALSIZE;
6364 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6365 return MDB_BAD_VALSIZE;
6368 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6369 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6373 if (flags == MDB_CURRENT) {
6374 if (!(mc->mc_flags & C_INITIALIZED))
6377 } else if (mc->mc_db->md_root == P_INVALID) {
6378 /* new database, cursor has nothing to point to */
6381 mc->mc_flags &= ~C_INITIALIZED;
6386 if (flags & MDB_APPEND) {
6388 rc = mdb_cursor_last(mc, &k2, &d2);
6390 rc = mc->mc_dbx->md_cmp(key, &k2);
6393 mc->mc_ki[mc->mc_top]++;
6395 /* new key is <= last key */
6400 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6402 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6403 DPRINTF(("duplicate key [%s]", DKEY(key)));
6405 return MDB_KEYEXIST;
6407 if (rc && rc != MDB_NOTFOUND)
6411 if (mc->mc_flags & C_DEL)
6412 mc->mc_flags ^= C_DEL;
6414 /* Cursor is positioned, check for room in the dirty list */
6416 if (flags & MDB_MULTIPLE) {
6418 xdata.mv_size = data->mv_size * dcount;
6422 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6426 if (rc == MDB_NO_ROOT) {
6428 /* new database, write a root leaf page */
6429 DPUTS("allocating new root leaf page");
6430 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6433 mdb_cursor_push(mc, np);
6434 mc->mc_db->md_root = np->mp_pgno;
6435 mc->mc_db->md_depth++;
6436 *mc->mc_dbflag |= DB_DIRTY;
6437 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6439 np->mp_flags |= P_LEAF2;
6440 mc->mc_flags |= C_INITIALIZED;
6442 /* make sure all cursor pages are writable */
6443 rc2 = mdb_cursor_touch(mc);
6448 insert_key = insert_data = rc;
6450 /* The key does not exist */
6451 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6452 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6453 LEAFSIZE(key, data) > env->me_nodemax)
6455 /* Too big for a node, insert in sub-DB. Set up an empty
6456 * "old sub-page" for prep_subDB to expand to a full page.
6458 fp_flags = P_LEAF|P_DIRTY;
6460 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6461 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6462 olddata.mv_size = PAGEHDRSZ;
6466 /* there's only a key anyway, so this is a no-op */
6467 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6469 unsigned int ksize = mc->mc_db->md_pad;
6470 if (key->mv_size != ksize)
6471 return MDB_BAD_VALSIZE;
6472 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6473 memcpy(ptr, key->mv_data, ksize);
6475 /* if overwriting slot 0 of leaf, need to
6476 * update branch key if there is a parent page
6478 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6479 unsigned short top = mc->mc_top;
6481 /* slot 0 is always an empty key, find real slot */
6482 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6484 if (mc->mc_ki[mc->mc_top])
6485 rc2 = mdb_update_key(mc, key);
6496 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6497 olddata.mv_size = NODEDSZ(leaf);
6498 olddata.mv_data = NODEDATA(leaf);
6501 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6502 /* Prepare (sub-)page/sub-DB to accept the new item,
6503 * if needed. fp: old sub-page or a header faking
6504 * it. mp: new (sub-)page. offset: growth in page
6505 * size. xdata: node data with new page or DB.
6507 unsigned i, offset = 0;
6508 mp = fp = xdata.mv_data = env->me_pbuf;
6509 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6511 /* Was a single item before, must convert now */
6512 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6514 /* Just overwrite the current item */
6515 if (flags == MDB_CURRENT)
6517 dcmp = mc->mc_dbx->md_dcmp;
6518 #if UINT_MAX < SIZE_MAX
6519 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6520 dcmp = mdb_cmp_clong;
6522 /* does data match? */
6523 if (!dcmp(data, &olddata)) {
6524 if (flags & MDB_NODUPDATA)
6525 return MDB_KEYEXIST;
6530 /* Back up original data item */
6531 dkey.mv_size = olddata.mv_size;
6532 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6534 /* Make sub-page header for the dup items, with dummy body */
6535 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6536 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6537 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6538 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6539 fp->mp_flags |= P_LEAF2;
6540 fp->mp_pad = data->mv_size;
6541 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6543 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6544 (dkey.mv_size & 1) + (data->mv_size & 1);
6546 fp->mp_upper = xdata.mv_size - PAGEBASE;
6547 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6548 } else if (leaf->mn_flags & F_SUBDATA) {
6549 /* Data is on sub-DB, just store it */
6550 flags |= F_DUPDATA|F_SUBDATA;
6553 /* Data is on sub-page */
6554 fp = olddata.mv_data;
6557 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6558 offset = EVEN(NODESIZE + sizeof(indx_t) +
6562 offset = fp->mp_pad;
6563 if (SIZELEFT(fp) < offset) {
6564 offset *= 4; /* space for 4 more */
6567 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6569 fp->mp_flags |= P_DIRTY;
6570 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6571 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6575 xdata.mv_size = olddata.mv_size + offset;
6578 fp_flags = fp->mp_flags;
6579 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6580 /* Too big for a sub-page, convert to sub-DB */
6581 fp_flags &= ~P_SUBP;
6583 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6584 fp_flags |= P_LEAF2;
6585 dummy.md_pad = fp->mp_pad;
6586 dummy.md_flags = MDB_DUPFIXED;
6587 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6588 dummy.md_flags |= MDB_INTEGERKEY;
6594 dummy.md_branch_pages = 0;
6595 dummy.md_leaf_pages = 1;
6596 dummy.md_overflow_pages = 0;
6597 dummy.md_entries = NUMKEYS(fp);
6598 xdata.mv_size = sizeof(MDB_db);
6599 xdata.mv_data = &dummy;
6600 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6602 offset = env->me_psize - olddata.mv_size;
6603 flags |= F_DUPDATA|F_SUBDATA;
6604 dummy.md_root = mp->mp_pgno;
6607 mp->mp_flags = fp_flags | P_DIRTY;
6608 mp->mp_pad = fp->mp_pad;
6609 mp->mp_lower = fp->mp_lower;
6610 mp->mp_upper = fp->mp_upper + offset;
6611 if (fp_flags & P_LEAF2) {
6612 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6614 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6615 olddata.mv_size - fp->mp_upper - PAGEBASE);
6616 for (i=0; i<NUMKEYS(fp); i++)
6617 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6625 mdb_node_del(mc, 0);
6629 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6630 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6631 return MDB_INCOMPATIBLE;
6632 /* overflow page overwrites need special handling */
6633 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6636 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6638 memcpy(&pg, olddata.mv_data, sizeof(pg));
6639 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6641 ovpages = omp->mp_pages;
6643 /* Is the ov page large enough? */
6644 if (ovpages >= dpages) {
6645 if (!(omp->mp_flags & P_DIRTY) &&
6646 (level || (env->me_flags & MDB_WRITEMAP)))
6648 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6651 level = 0; /* dirty in this txn or clean */
6654 if (omp->mp_flags & P_DIRTY) {
6655 /* yes, overwrite it. Note in this case we don't
6656 * bother to try shrinking the page if the new data
6657 * is smaller than the overflow threshold.
6660 /* It is writable only in a parent txn */
6661 size_t sz = (size_t) env->me_psize * ovpages, off;
6662 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6668 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6669 mdb_cassert(mc, rc2 == 0);
6670 if (!(flags & MDB_RESERVE)) {
6671 /* Copy end of page, adjusting alignment so
6672 * compiler may copy words instead of bytes.
6674 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6675 memcpy((size_t *)((char *)np + off),
6676 (size_t *)((char *)omp + off), sz - off);
6679 memcpy(np, omp, sz); /* Copy beginning of page */
6682 SETDSZ(leaf, data->mv_size);
6683 if (F_ISSET(flags, MDB_RESERVE))
6684 data->mv_data = METADATA(omp);
6686 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6690 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6692 } else if (data->mv_size == olddata.mv_size) {
6693 /* same size, just replace it. Note that we could
6694 * also reuse this node if the new data is smaller,
6695 * but instead we opt to shrink the node in that case.
6697 if (F_ISSET(flags, MDB_RESERVE))
6698 data->mv_data = olddata.mv_data;
6699 else if (!(mc->mc_flags & C_SUB))
6700 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6702 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6707 mdb_node_del(mc, 0);
6713 nflags = flags & NODE_ADD_FLAGS;
6714 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6715 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6716 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6717 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6719 nflags |= MDB_SPLIT_REPLACE;
6720 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6722 /* There is room already in this leaf page. */
6723 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6724 if (rc == 0 && insert_key) {
6725 /* Adjust other cursors pointing to mp */
6726 MDB_cursor *m2, *m3;
6727 MDB_dbi dbi = mc->mc_dbi;
6728 unsigned i = mc->mc_top;
6729 MDB_page *mp = mc->mc_pg[i];
6731 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6732 if (mc->mc_flags & C_SUB)
6733 m3 = &m2->mc_xcursor->mx_cursor;
6736 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6737 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6744 if (rc == MDB_SUCCESS) {
6745 /* Now store the actual data in the child DB. Note that we're
6746 * storing the user data in the keys field, so there are strict
6747 * size limits on dupdata. The actual data fields of the child
6748 * DB are all zero size.
6756 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6757 if (flags & MDB_CURRENT) {
6758 xflags = MDB_CURRENT|MDB_NOSPILL;
6760 mdb_xcursor_init1(mc, leaf);
6761 xflags = (flags & MDB_NODUPDATA) ?
6762 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6764 /* converted, write the original data first */
6766 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6769 /* we've done our job */
6773 /* Adjust other cursors pointing to mp */
6775 unsigned i = mc->mc_top;
6776 MDB_page *mp = mc->mc_pg[i];
6778 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6779 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6780 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6781 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6782 mdb_xcursor_init2(m2, leaf);
6786 ecount = mc->mc_xcursor->mx_db.md_entries;
6787 if (flags & MDB_APPENDDUP)
6788 xflags |= MDB_APPEND;
6789 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6790 if (flags & F_SUBDATA) {
6791 void *db = NODEDATA(leaf);
6792 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6794 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6796 /* Increment count unless we just replaced an existing item. */
6798 mc->mc_db->md_entries++;
6800 /* Invalidate txn if we created an empty sub-DB */
6803 /* If we succeeded and the key didn't exist before,
6804 * make sure the cursor is marked valid.
6806 mc->mc_flags |= C_INITIALIZED;
6808 if (flags & MDB_MULTIPLE) {
6811 /* let caller know how many succeeded, if any */
6812 data[1].mv_size = mcount;
6813 if (mcount < dcount) {
6814 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6815 insert_key = insert_data = 0;
6822 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6825 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6830 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6836 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6837 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6839 if (!(mc->mc_flags & C_INITIALIZED))
6842 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6843 return MDB_NOTFOUND;
6845 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6848 rc = mdb_cursor_touch(mc);
6852 mp = mc->mc_pg[mc->mc_top];
6855 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6857 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6858 if (flags & MDB_NODUPDATA) {
6859 /* mdb_cursor_del0() will subtract the final entry */
6860 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6862 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6863 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6865 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6868 /* If sub-DB still has entries, we're done */
6869 if (mc->mc_xcursor->mx_db.md_entries) {
6870 if (leaf->mn_flags & F_SUBDATA) {
6871 /* update subDB info */
6872 void *db = NODEDATA(leaf);
6873 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6876 /* shrink fake page */
6877 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6878 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6879 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6880 /* fix other sub-DB cursors pointed at this fake page */
6881 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6882 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6883 if (m2->mc_pg[mc->mc_top] == mp &&
6884 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6885 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6888 mc->mc_db->md_entries--;
6889 mc->mc_flags |= C_DEL;
6892 /* otherwise fall thru and delete the sub-DB */
6895 if (leaf->mn_flags & F_SUBDATA) {
6896 /* add all the child DB's pages to the free list */
6897 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6902 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6903 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6904 rc = MDB_INCOMPATIBLE;
6908 /* add overflow pages to free list */
6909 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6913 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6914 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6915 (rc = mdb_ovpage_free(mc, omp)))
6920 return mdb_cursor_del0(mc);
6923 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6927 /** Allocate and initialize new pages for a database.
6928 * @param[in] mc a cursor on the database being added to.
6929 * @param[in] flags flags defining what type of page is being allocated.
6930 * @param[in] num the number of pages to allocate. This is usually 1,
6931 * unless allocating overflow pages for a large record.
6932 * @param[out] mp Address of a page, or NULL on failure.
6933 * @return 0 on success, non-zero on failure.
6936 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6941 if ((rc = mdb_page_alloc(mc, num, &np)))
6943 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6944 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6945 np->mp_flags = flags | P_DIRTY;
6946 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6947 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6950 mc->mc_db->md_branch_pages++;
6951 else if (IS_LEAF(np))
6952 mc->mc_db->md_leaf_pages++;
6953 else if (IS_OVERFLOW(np)) {
6954 mc->mc_db->md_overflow_pages += num;
6962 /** Calculate the size of a leaf node.
6963 * The size depends on the environment's page size; if a data item
6964 * is too large it will be put onto an overflow page and the node
6965 * size will only include the key and not the data. Sizes are always
6966 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6967 * of the #MDB_node headers.
6968 * @param[in] env The environment handle.
6969 * @param[in] key The key for the node.
6970 * @param[in] data The data for the node.
6971 * @return The number of bytes needed to store the node.
6974 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6978 sz = LEAFSIZE(key, data);
6979 if (sz > env->me_nodemax) {
6980 /* put on overflow page */
6981 sz -= data->mv_size - sizeof(pgno_t);
6984 return EVEN(sz + sizeof(indx_t));
6987 /** Calculate the size of a branch node.
6988 * The size should depend on the environment's page size but since
6989 * we currently don't support spilling large keys onto overflow
6990 * pages, it's simply the size of the #MDB_node header plus the
6991 * size of the key. Sizes are always rounded up to an even number
6992 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6993 * @param[in] env The environment handle.
6994 * @param[in] key The key for the node.
6995 * @return The number of bytes needed to store the node.
6998 mdb_branch_size(MDB_env *env, MDB_val *key)
7003 if (sz > env->me_nodemax) {
7004 /* put on overflow page */
7005 /* not implemented */
7006 /* sz -= key->size - sizeof(pgno_t); */
7009 return sz + sizeof(indx_t);
7012 /** Add a node to the page pointed to by the cursor.
7013 * @param[in] mc The cursor for this operation.
7014 * @param[in] indx The index on the page where the new node should be added.
7015 * @param[in] key The key for the new node.
7016 * @param[in] data The data for the new node, if any.
7017 * @param[in] pgno The page number, if adding a branch node.
7018 * @param[in] flags Flags for the node.
7019 * @return 0 on success, non-zero on failure. Possible errors are:
7021 * <li>ENOMEM - failed to allocate overflow pages for the node.
7022 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7023 * should never happen since all callers already calculate the
7024 * page's free space before calling this function.
7028 mdb_node_add(MDB_cursor *mc, indx_t indx,
7029 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7032 size_t node_size = NODESIZE;
7036 MDB_page *mp = mc->mc_pg[mc->mc_top];
7037 MDB_page *ofp = NULL; /* overflow page */
7041 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7043 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7044 IS_LEAF(mp) ? "leaf" : "branch",
7045 IS_SUBP(mp) ? "sub-" : "",
7046 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7047 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7050 /* Move higher keys up one slot. */
7051 int ksize = mc->mc_db->md_pad, dif;
7052 char *ptr = LEAF2KEY(mp, indx, ksize);
7053 dif = NUMKEYS(mp) - indx;
7055 memmove(ptr+ksize, ptr, dif*ksize);
7056 /* insert new key */
7057 memcpy(ptr, key->mv_data, ksize);
7059 /* Just using these for counting */
7060 mp->mp_lower += sizeof(indx_t);
7061 mp->mp_upper -= ksize - sizeof(indx_t);
7065 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7067 node_size += key->mv_size;
7069 mdb_cassert(mc, key && data);
7070 if (F_ISSET(flags, F_BIGDATA)) {
7071 /* Data already on overflow page. */
7072 node_size += sizeof(pgno_t);
7073 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7074 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7076 /* Put data on overflow page. */
7077 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7078 data->mv_size, node_size+data->mv_size));
7079 node_size = EVEN(node_size + sizeof(pgno_t));
7080 if ((ssize_t)node_size > room)
7082 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7084 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7088 node_size += data->mv_size;
7091 node_size = EVEN(node_size);
7092 if ((ssize_t)node_size > room)
7096 /* Move higher pointers up one slot. */
7097 for (i = NUMKEYS(mp); i > indx; i--)
7098 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7100 /* Adjust free space offsets. */
7101 ofs = mp->mp_upper - node_size;
7102 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7103 mp->mp_ptrs[indx] = ofs;
7105 mp->mp_lower += sizeof(indx_t);
7107 /* Write the node data. */
7108 node = NODEPTR(mp, indx);
7109 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7110 node->mn_flags = flags;
7112 SETDSZ(node,data->mv_size);
7117 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7120 ndata = NODEDATA(node);
7122 if (F_ISSET(flags, F_BIGDATA))
7123 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7124 else if (F_ISSET(flags, MDB_RESERVE))
7125 data->mv_data = ndata;
7127 memcpy(ndata, data->mv_data, data->mv_size);
7129 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7130 ndata = METADATA(ofp);
7131 if (F_ISSET(flags, MDB_RESERVE))
7132 data->mv_data = ndata;
7134 memcpy(ndata, data->mv_data, data->mv_size);
7141 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7142 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7143 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7144 DPRINTF(("node size = %"Z"u", node_size));
7145 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7146 return MDB_PAGE_FULL;
7149 /** Delete the specified node from a page.
7150 * @param[in] mc Cursor pointing to the node to delete.
7151 * @param[in] ksize The size of a node. Only used if the page is
7152 * part of a #MDB_DUPFIXED database.
7155 mdb_node_del(MDB_cursor *mc, int ksize)
7157 MDB_page *mp = mc->mc_pg[mc->mc_top];
7158 indx_t indx = mc->mc_ki[mc->mc_top];
7160 indx_t i, j, numkeys, ptr;
7164 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7165 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7166 numkeys = NUMKEYS(mp);
7167 mdb_cassert(mc, indx < numkeys);
7170 int x = numkeys - 1 - indx;
7171 base = LEAF2KEY(mp, indx, ksize);
7173 memmove(base, base + ksize, x * ksize);
7174 mp->mp_lower -= sizeof(indx_t);
7175 mp->mp_upper += ksize - sizeof(indx_t);
7179 node = NODEPTR(mp, indx);
7180 sz = NODESIZE + node->mn_ksize;
7182 if (F_ISSET(node->mn_flags, F_BIGDATA))
7183 sz += sizeof(pgno_t);
7185 sz += NODEDSZ(node);
7189 ptr = mp->mp_ptrs[indx];
7190 for (i = j = 0; i < numkeys; i++) {
7192 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7193 if (mp->mp_ptrs[i] < ptr)
7194 mp->mp_ptrs[j] += sz;
7199 base = (char *)mp + mp->mp_upper + PAGEBASE;
7200 memmove(base + sz, base, ptr - mp->mp_upper);
7202 mp->mp_lower -= sizeof(indx_t);
7206 /** Compact the main page after deleting a node on a subpage.
7207 * @param[in] mp The main page to operate on.
7208 * @param[in] indx The index of the subpage on the main page.
7211 mdb_node_shrink(MDB_page *mp, indx_t indx)
7216 indx_t delta, nsize, len, ptr;
7219 node = NODEPTR(mp, indx);
7220 sp = (MDB_page *)NODEDATA(node);
7221 delta = SIZELEFT(sp);
7222 nsize = NODEDSZ(node) - delta;
7224 /* Prepare to shift upward, set len = length(subpage part to shift) */
7228 return; /* do not make the node uneven-sized */
7230 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7231 for (i = NUMKEYS(sp); --i >= 0; )
7232 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7235 sp->mp_upper = sp->mp_lower;
7236 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7237 SETDSZ(node, nsize);
7239 /* Shift <lower nodes...initial part of subpage> upward */
7240 base = (char *)mp + mp->mp_upper + PAGEBASE;
7241 memmove(base + delta, base, (char *)sp + len - base);
7243 ptr = mp->mp_ptrs[indx];
7244 for (i = NUMKEYS(mp); --i >= 0; ) {
7245 if (mp->mp_ptrs[i] <= ptr)
7246 mp->mp_ptrs[i] += delta;
7248 mp->mp_upper += delta;
7251 /** Initial setup of a sorted-dups cursor.
7252 * Sorted duplicates are implemented as a sub-database for the given key.
7253 * The duplicate data items are actually keys of the sub-database.
7254 * Operations on the duplicate data items are performed using a sub-cursor
7255 * initialized when the sub-database is first accessed. This function does
7256 * the preliminary setup of the sub-cursor, filling in the fields that
7257 * depend only on the parent DB.
7258 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7261 mdb_xcursor_init0(MDB_cursor *mc)
7263 MDB_xcursor *mx = mc->mc_xcursor;
7265 mx->mx_cursor.mc_xcursor = NULL;
7266 mx->mx_cursor.mc_txn = mc->mc_txn;
7267 mx->mx_cursor.mc_db = &mx->mx_db;
7268 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7269 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7270 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7271 mx->mx_cursor.mc_snum = 0;
7272 mx->mx_cursor.mc_top = 0;
7273 mx->mx_cursor.mc_flags = C_SUB;
7274 mx->mx_dbx.md_name.mv_size = 0;
7275 mx->mx_dbx.md_name.mv_data = NULL;
7276 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7277 mx->mx_dbx.md_dcmp = NULL;
7278 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7281 /** Final setup of a sorted-dups cursor.
7282 * Sets up the fields that depend on the data from the main cursor.
7283 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7284 * @param[in] node The data containing the #MDB_db record for the
7285 * sorted-dup database.
7288 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7290 MDB_xcursor *mx = mc->mc_xcursor;
7292 if (node->mn_flags & F_SUBDATA) {
7293 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7294 mx->mx_cursor.mc_pg[0] = 0;
7295 mx->mx_cursor.mc_snum = 0;
7296 mx->mx_cursor.mc_top = 0;
7297 mx->mx_cursor.mc_flags = C_SUB;
7299 MDB_page *fp = NODEDATA(node);
7300 mx->mx_db.md_pad = 0;
7301 mx->mx_db.md_flags = 0;
7302 mx->mx_db.md_depth = 1;
7303 mx->mx_db.md_branch_pages = 0;
7304 mx->mx_db.md_leaf_pages = 1;
7305 mx->mx_db.md_overflow_pages = 0;
7306 mx->mx_db.md_entries = NUMKEYS(fp);
7307 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7308 mx->mx_cursor.mc_snum = 1;
7309 mx->mx_cursor.mc_top = 0;
7310 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7311 mx->mx_cursor.mc_pg[0] = fp;
7312 mx->mx_cursor.mc_ki[0] = 0;
7313 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7314 mx->mx_db.md_flags = MDB_DUPFIXED;
7315 mx->mx_db.md_pad = fp->mp_pad;
7316 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7317 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7320 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7321 mx->mx_db.md_root));
7322 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7323 #if UINT_MAX < SIZE_MAX
7324 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7325 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7330 /** Fixup a sorted-dups cursor due to underlying update.
7331 * Sets up some fields that depend on the data from the main cursor.
7332 * Almost the same as init1, but skips initialization steps if the
7333 * xcursor had already been used.
7334 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7335 * @param[in] node The data containing the #MDB_db record for the
7336 * sorted-dup database.
7339 mdb_xcursor_init2(MDB_cursor *mc, MDB_node *node)
7341 MDB_xcursor *mx = mc->mc_xcursor;
7343 if (node->mn_flags & F_SUBDATA) {
7344 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7345 mdb_page_get(mc->mc_txn,mx->mx_db.md_root,&mx->mx_cursor.mc_pg[0],NULL);
7347 MDB_page *fp = NODEDATA(node);
7348 mx->mx_db.md_entries = NUMKEYS(fp);
7349 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7350 mx->mx_cursor.mc_pg[0] = fp;
7352 if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7353 mx->mx_cursor.mc_snum = 1;
7354 mx->mx_cursor.mc_top = 0;
7355 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7356 mx->mx_cursor.mc_ki[0] = 0;
7357 if (!(node->mn_flags & F_SUBDATA)) {
7358 mx->mx_db.md_pad = 0;
7359 mx->mx_db.md_flags = 0;
7360 mx->mx_db.md_depth = 1;
7361 mx->mx_db.md_branch_pages = 0;
7362 mx->mx_db.md_leaf_pages = 1;
7363 mx->mx_db.md_overflow_pages = 0;
7364 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7365 mx->mx_db.md_flags = MDB_DUPFIXED;
7366 mx->mx_db.md_pad = mx->mx_cursor.mc_pg[0]->mp_pad;
7367 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7368 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7372 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7373 mx->mx_db.md_root));
7374 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7377 /** Initialize a cursor for a given transaction and database. */
7379 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7382 mc->mc_backup = NULL;
7385 mc->mc_db = &txn->mt_dbs[dbi];
7386 mc->mc_dbx = &txn->mt_dbxs[dbi];
7387 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7393 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7394 mdb_tassert(txn, mx != NULL);
7395 mc->mc_xcursor = mx;
7396 mdb_xcursor_init0(mc);
7398 mc->mc_xcursor = NULL;
7400 if (*mc->mc_dbflag & DB_STALE) {
7401 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7406 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7409 size_t size = sizeof(MDB_cursor);
7411 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7414 if (txn->mt_flags & MDB_TXN_BLOCKED)
7417 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7420 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7421 size += sizeof(MDB_xcursor);
7423 if ((mc = malloc(size)) != NULL) {
7424 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7425 if (txn->mt_cursors) {
7426 mc->mc_next = txn->mt_cursors[dbi];
7427 txn->mt_cursors[dbi] = mc;
7428 mc->mc_flags |= C_UNTRACK;
7440 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7442 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7445 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7448 if (txn->mt_flags & MDB_TXN_BLOCKED)
7451 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7455 /* Return the count of duplicate data items for the current key */
7457 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7461 if (mc == NULL || countp == NULL)
7464 if (mc->mc_xcursor == NULL)
7465 return MDB_INCOMPATIBLE;
7467 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7470 if (!(mc->mc_flags & C_INITIALIZED))
7473 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7474 return MDB_NOTFOUND;
7476 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7477 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7480 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7483 *countp = mc->mc_xcursor->mx_db.md_entries;
7489 mdb_cursor_close(MDB_cursor *mc)
7491 if (mc && !mc->mc_backup) {
7492 /* remove from txn, if tracked */
7493 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7494 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7495 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7497 *prev = mc->mc_next;
7504 mdb_cursor_txn(MDB_cursor *mc)
7506 if (!mc) return NULL;
7511 mdb_cursor_dbi(MDB_cursor *mc)
7516 /** Replace the key for a branch node with a new key.
7517 * @param[in] mc Cursor pointing to the node to operate on.
7518 * @param[in] key The new key to use.
7519 * @return 0 on success, non-zero on failure.
7522 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7528 int delta, ksize, oksize;
7529 indx_t ptr, i, numkeys, indx;
7532 indx = mc->mc_ki[mc->mc_top];
7533 mp = mc->mc_pg[mc->mc_top];
7534 node = NODEPTR(mp, indx);
7535 ptr = mp->mp_ptrs[indx];
7539 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7540 k2.mv_data = NODEKEY(node);
7541 k2.mv_size = node->mn_ksize;
7542 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7544 mdb_dkey(&k2, kbuf2),
7550 /* Sizes must be 2-byte aligned. */
7551 ksize = EVEN(key->mv_size);
7552 oksize = EVEN(node->mn_ksize);
7553 delta = ksize - oksize;
7555 /* Shift node contents if EVEN(key length) changed. */
7557 if (delta > 0 && SIZELEFT(mp) < delta) {
7559 /* not enough space left, do a delete and split */
7560 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7561 pgno = NODEPGNO(node);
7562 mdb_node_del(mc, 0);
7563 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7566 numkeys = NUMKEYS(mp);
7567 for (i = 0; i < numkeys; i++) {
7568 if (mp->mp_ptrs[i] <= ptr)
7569 mp->mp_ptrs[i] -= delta;
7572 base = (char *)mp + mp->mp_upper + PAGEBASE;
7573 len = ptr - mp->mp_upper + NODESIZE;
7574 memmove(base - delta, base, len);
7575 mp->mp_upper -= delta;
7577 node = NODEPTR(mp, indx);
7580 /* But even if no shift was needed, update ksize */
7581 if (node->mn_ksize != key->mv_size)
7582 node->mn_ksize = key->mv_size;
7585 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7591 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7593 /** Move a node from csrc to cdst.
7596 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7603 unsigned short flags;
7607 /* Mark src and dst as dirty. */
7608 if ((rc = mdb_page_touch(csrc)) ||
7609 (rc = mdb_page_touch(cdst)))
7612 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7613 key.mv_size = csrc->mc_db->md_pad;
7614 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7616 data.mv_data = NULL;
7620 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7621 mdb_cassert(csrc, !((size_t)srcnode & 1));
7622 srcpg = NODEPGNO(srcnode);
7623 flags = srcnode->mn_flags;
7624 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7625 unsigned int snum = csrc->mc_snum;
7627 /* must find the lowest key below src */
7628 rc = mdb_page_search_lowest(csrc);
7631 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7632 key.mv_size = csrc->mc_db->md_pad;
7633 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7635 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7636 key.mv_size = NODEKSZ(s2);
7637 key.mv_data = NODEKEY(s2);
7639 csrc->mc_snum = snum--;
7640 csrc->mc_top = snum;
7642 key.mv_size = NODEKSZ(srcnode);
7643 key.mv_data = NODEKEY(srcnode);
7645 data.mv_size = NODEDSZ(srcnode);
7646 data.mv_data = NODEDATA(srcnode);
7648 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7649 unsigned int snum = cdst->mc_snum;
7652 /* must find the lowest key below dst */
7653 mdb_cursor_copy(cdst, &mn);
7654 rc = mdb_page_search_lowest(&mn);
7657 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7658 bkey.mv_size = mn.mc_db->md_pad;
7659 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7661 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7662 bkey.mv_size = NODEKSZ(s2);
7663 bkey.mv_data = NODEKEY(s2);
7665 mn.mc_snum = snum--;
7668 rc = mdb_update_key(&mn, &bkey);
7673 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7674 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7675 csrc->mc_ki[csrc->mc_top],
7677 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7678 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7680 /* Add the node to the destination page.
7682 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7683 if (rc != MDB_SUCCESS)
7686 /* Delete the node from the source page.
7688 mdb_node_del(csrc, key.mv_size);
7691 /* Adjust other cursors pointing to mp */
7692 MDB_cursor *m2, *m3;
7693 MDB_dbi dbi = csrc->mc_dbi;
7696 mp = cdst->mc_pg[csrc->mc_top];
7697 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7698 if (csrc->mc_flags & C_SUB)
7699 m3 = &m2->mc_xcursor->mx_cursor;
7702 if (m3 == cdst) continue;
7703 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] >=
7704 cdst->mc_ki[csrc->mc_top]) {
7705 m3->mc_ki[csrc->mc_top]++;
7709 mp = csrc->mc_pg[csrc->mc_top];
7710 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7711 if (csrc->mc_flags & C_SUB)
7712 m3 = &m2->mc_xcursor->mx_cursor;
7715 if (m3 == csrc) continue;
7716 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7717 csrc->mc_ki[csrc->mc_top]) {
7718 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7719 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7724 /* Update the parent separators.
7726 if (csrc->mc_ki[csrc->mc_top] == 0) {
7727 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7728 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7729 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7731 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7732 key.mv_size = NODEKSZ(srcnode);
7733 key.mv_data = NODEKEY(srcnode);
7735 DPRINTF(("update separator for source page %"Z"u to [%s]",
7736 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7737 mdb_cursor_copy(csrc, &mn);
7740 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7743 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7745 indx_t ix = csrc->mc_ki[csrc->mc_top];
7746 nullkey.mv_size = 0;
7747 csrc->mc_ki[csrc->mc_top] = 0;
7748 rc = mdb_update_key(csrc, &nullkey);
7749 csrc->mc_ki[csrc->mc_top] = ix;
7750 mdb_cassert(csrc, rc == MDB_SUCCESS);
7754 if (cdst->mc_ki[cdst->mc_top] == 0) {
7755 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7756 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7757 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7759 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7760 key.mv_size = NODEKSZ(srcnode);
7761 key.mv_data = NODEKEY(srcnode);
7763 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7764 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7765 mdb_cursor_copy(cdst, &mn);
7768 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7771 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7773 indx_t ix = cdst->mc_ki[cdst->mc_top];
7774 nullkey.mv_size = 0;
7775 cdst->mc_ki[cdst->mc_top] = 0;
7776 rc = mdb_update_key(cdst, &nullkey);
7777 cdst->mc_ki[cdst->mc_top] = ix;
7778 mdb_cassert(cdst, rc == MDB_SUCCESS);
7785 /** Merge one page into another.
7786 * The nodes from the page pointed to by \b csrc will
7787 * be copied to the page pointed to by \b cdst and then
7788 * the \b csrc page will be freed.
7789 * @param[in] csrc Cursor pointing to the source page.
7790 * @param[in] cdst Cursor pointing to the destination page.
7791 * @return 0 on success, non-zero on failure.
7794 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7796 MDB_page *psrc, *pdst;
7803 psrc = csrc->mc_pg[csrc->mc_top];
7804 pdst = cdst->mc_pg[cdst->mc_top];
7806 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7808 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7809 mdb_cassert(csrc, cdst->mc_snum > 1);
7811 /* Mark dst as dirty. */
7812 if ((rc = mdb_page_touch(cdst)))
7815 /* Move all nodes from src to dst.
7817 j = nkeys = NUMKEYS(pdst);
7818 if (IS_LEAF2(psrc)) {
7819 key.mv_size = csrc->mc_db->md_pad;
7820 key.mv_data = METADATA(psrc);
7821 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7822 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7823 if (rc != MDB_SUCCESS)
7825 key.mv_data = (char *)key.mv_data + key.mv_size;
7828 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7829 srcnode = NODEPTR(psrc, i);
7830 if (i == 0 && IS_BRANCH(psrc)) {
7833 mdb_cursor_copy(csrc, &mn);
7834 /* must find the lowest key below src */
7835 rc = mdb_page_search_lowest(&mn);
7838 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7839 key.mv_size = mn.mc_db->md_pad;
7840 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7842 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7843 key.mv_size = NODEKSZ(s2);
7844 key.mv_data = NODEKEY(s2);
7847 key.mv_size = srcnode->mn_ksize;
7848 key.mv_data = NODEKEY(srcnode);
7851 data.mv_size = NODEDSZ(srcnode);
7852 data.mv_data = NODEDATA(srcnode);
7853 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7854 if (rc != MDB_SUCCESS)
7859 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7860 pdst->mp_pgno, NUMKEYS(pdst),
7861 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7863 /* Unlink the src page from parent and add to free list.
7866 mdb_node_del(csrc, 0);
7867 if (csrc->mc_ki[csrc->mc_top] == 0) {
7869 rc = mdb_update_key(csrc, &key);
7877 psrc = csrc->mc_pg[csrc->mc_top];
7878 /* If not operating on FreeDB, allow this page to be reused
7879 * in this txn. Otherwise just add to free list.
7881 rc = mdb_page_loose(csrc, psrc);
7885 csrc->mc_db->md_leaf_pages--;
7887 csrc->mc_db->md_branch_pages--;
7889 /* Adjust other cursors pointing to mp */
7890 MDB_cursor *m2, *m3;
7891 MDB_dbi dbi = csrc->mc_dbi;
7893 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7894 if (csrc->mc_flags & C_SUB)
7895 m3 = &m2->mc_xcursor->mx_cursor;
7898 if (m3 == csrc) continue;
7899 if (m3->mc_snum < csrc->mc_snum) continue;
7900 if (m3->mc_pg[csrc->mc_top] == psrc) {
7901 m3->mc_pg[csrc->mc_top] = pdst;
7902 m3->mc_ki[csrc->mc_top] += nkeys;
7907 unsigned int snum = cdst->mc_snum;
7908 uint16_t depth = cdst->mc_db->md_depth;
7909 mdb_cursor_pop(cdst);
7910 rc = mdb_rebalance(cdst);
7911 /* Did the tree height change? */
7912 if (depth != cdst->mc_db->md_depth)
7913 snum += cdst->mc_db->md_depth - depth;
7914 cdst->mc_snum = snum;
7915 cdst->mc_top = snum-1;
7920 /** Copy the contents of a cursor.
7921 * @param[in] csrc The cursor to copy from.
7922 * @param[out] cdst The cursor to copy to.
7925 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7929 cdst->mc_txn = csrc->mc_txn;
7930 cdst->mc_dbi = csrc->mc_dbi;
7931 cdst->mc_db = csrc->mc_db;
7932 cdst->mc_dbx = csrc->mc_dbx;
7933 cdst->mc_snum = csrc->mc_snum;
7934 cdst->mc_top = csrc->mc_top;
7935 cdst->mc_flags = csrc->mc_flags;
7937 for (i=0; i<csrc->mc_snum; i++) {
7938 cdst->mc_pg[i] = csrc->mc_pg[i];
7939 cdst->mc_ki[i] = csrc->mc_ki[i];
7943 /** Rebalance the tree after a delete operation.
7944 * @param[in] mc Cursor pointing to the page where rebalancing
7946 * @return 0 on success, non-zero on failure.
7949 mdb_rebalance(MDB_cursor *mc)
7953 unsigned int ptop, minkeys, thresh;
7957 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7962 thresh = FILL_THRESHOLD;
7964 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7965 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7966 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7967 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7969 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7970 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7971 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7972 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7976 if (mc->mc_snum < 2) {
7977 MDB_page *mp = mc->mc_pg[0];
7979 DPUTS("Can't rebalance a subpage, ignoring");
7982 if (NUMKEYS(mp) == 0) {
7983 DPUTS("tree is completely empty");
7984 mc->mc_db->md_root = P_INVALID;
7985 mc->mc_db->md_depth = 0;
7986 mc->mc_db->md_leaf_pages = 0;
7987 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7990 /* Adjust cursors pointing to mp */
7993 mc->mc_flags &= ~C_INITIALIZED;
7995 MDB_cursor *m2, *m3;
7996 MDB_dbi dbi = mc->mc_dbi;
7998 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7999 if (mc->mc_flags & C_SUB)
8000 m3 = &m2->mc_xcursor->mx_cursor;
8003 if (m3->mc_snum < mc->mc_snum) continue;
8004 if (m3->mc_pg[0] == mp) {
8007 m3->mc_flags &= ~C_INITIALIZED;
8011 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8013 DPUTS("collapsing root page!");
8014 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8017 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8018 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
8021 mc->mc_db->md_depth--;
8022 mc->mc_db->md_branch_pages--;
8023 mc->mc_ki[0] = mc->mc_ki[1];
8024 for (i = 1; i<mc->mc_db->md_depth; i++) {
8025 mc->mc_pg[i] = mc->mc_pg[i+1];
8026 mc->mc_ki[i] = mc->mc_ki[i+1];
8029 /* Adjust other cursors pointing to mp */
8030 MDB_cursor *m2, *m3;
8031 MDB_dbi dbi = mc->mc_dbi;
8033 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8034 if (mc->mc_flags & C_SUB)
8035 m3 = &m2->mc_xcursor->mx_cursor;
8038 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
8039 if (m3->mc_pg[0] == mp) {
8040 for (i=0; i<m3->mc_snum; i++) {
8041 m3->mc_pg[i] = m3->mc_pg[i+1];
8042 m3->mc_ki[i] = m3->mc_ki[i+1];
8050 DPUTS("root page doesn't need rebalancing");
8054 /* The parent (branch page) must have at least 2 pointers,
8055 * otherwise the tree is invalid.
8057 ptop = mc->mc_top-1;
8058 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8060 /* Leaf page fill factor is below the threshold.
8061 * Try to move keys from left or right neighbor, or
8062 * merge with a neighbor page.
8067 mdb_cursor_copy(mc, &mn);
8068 mn.mc_xcursor = NULL;
8070 oldki = mc->mc_ki[mc->mc_top];
8071 if (mc->mc_ki[ptop] == 0) {
8072 /* We're the leftmost leaf in our parent.
8074 DPUTS("reading right neighbor");
8076 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8077 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8080 mn.mc_ki[mn.mc_top] = 0;
8081 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8083 /* There is at least one neighbor to the left.
8085 DPUTS("reading left neighbor");
8087 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8088 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8091 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8092 mc->mc_ki[mc->mc_top] = 0;
8095 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8096 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8097 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8099 /* If the neighbor page is above threshold and has enough keys,
8100 * move one key from it. Otherwise we should try to merge them.
8101 * (A branch page must never have less than 2 keys.)
8103 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8104 rc = mdb_node_move(&mn, mc);
8105 if (mc->mc_ki[mc->mc_top-1]) {
8109 if (mc->mc_ki[ptop] == 0) {
8110 rc = mdb_page_merge(&mn, mc);
8113 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8114 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8115 /* We want mdb_rebalance to find mn when doing fixups */
8116 if (mc->mc_flags & C_SUB) {
8117 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8118 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8119 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8121 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8122 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8124 rc = mdb_page_merge(mc, &mn);
8125 if (mc->mc_flags & C_SUB)
8126 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8128 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8129 mdb_cursor_copy(&mn, mc);
8131 mc->mc_flags &= ~C_EOF;
8133 mc->mc_ki[mc->mc_top] = oldki;
8137 /** Complete a delete operation started by #mdb_cursor_del(). */
8139 mdb_cursor_del0(MDB_cursor *mc)
8145 MDB_cursor *m2, *m3;
8146 MDB_dbi dbi = mc->mc_dbi;
8148 ki = mc->mc_ki[mc->mc_top];
8149 mp = mc->mc_pg[mc->mc_top];
8150 mdb_node_del(mc, mc->mc_db->md_pad);
8151 mc->mc_db->md_entries--;
8153 /* Adjust other cursors pointing to mp */
8154 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8155 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8156 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8158 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8160 if (m3->mc_pg[mc->mc_top] == mp) {
8161 if (m3->mc_ki[mc->mc_top] >= ki) {
8162 m3->mc_flags |= C_DEL;
8163 if (m3->mc_ki[mc->mc_top] > ki)
8164 m3->mc_ki[mc->mc_top]--;
8165 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8166 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8171 rc = mdb_rebalance(mc);
8173 if (rc == MDB_SUCCESS) {
8174 /* DB is totally empty now, just bail out.
8175 * Other cursors adjustments were already done
8176 * by mdb_rebalance and aren't needed here.
8181 mp = mc->mc_pg[mc->mc_top];
8182 nkeys = NUMKEYS(mp);
8184 /* Adjust other cursors pointing to mp */
8185 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8186 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8187 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8189 if (m3->mc_snum < mc->mc_snum)
8191 if (m3->mc_pg[mc->mc_top] == mp) {
8192 /* if m3 points past last node in page, find next sibling */
8193 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8194 rc = mdb_cursor_sibling(m3, 1);
8195 if (rc == MDB_NOTFOUND) {
8196 m3->mc_flags |= C_EOF;
8202 mc->mc_flags |= C_DEL;
8206 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8211 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8212 MDB_val *key, MDB_val *data)
8214 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8217 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8218 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8220 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8221 /* must ignore any data */
8225 return mdb_del0(txn, dbi, key, data, 0);
8229 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8230 MDB_val *key, MDB_val *data, unsigned flags)
8235 MDB_val rdata, *xdata;
8239 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8241 mdb_cursor_init(&mc, txn, dbi, &mx);
8250 flags |= MDB_NODUPDATA;
8252 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8254 /* let mdb_page_split know about this cursor if needed:
8255 * delete will trigger a rebalance; if it needs to move
8256 * a node from one page to another, it will have to
8257 * update the parent's separator key(s). If the new sepkey
8258 * is larger than the current one, the parent page may
8259 * run out of space, triggering a split. We need this
8260 * cursor to be consistent until the end of the rebalance.
8262 mc.mc_flags |= C_UNTRACK;
8263 mc.mc_next = txn->mt_cursors[dbi];
8264 txn->mt_cursors[dbi] = &mc;
8265 rc = mdb_cursor_del(&mc, flags);
8266 txn->mt_cursors[dbi] = mc.mc_next;
8271 /** Split a page and insert a new node.
8272 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8273 * The cursor will be updated to point to the actual page and index where
8274 * the node got inserted after the split.
8275 * @param[in] newkey The key for the newly inserted node.
8276 * @param[in] newdata The data for the newly inserted node.
8277 * @param[in] newpgno The page number, if the new node is a branch node.
8278 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8279 * @return 0 on success, non-zero on failure.
8282 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8283 unsigned int nflags)
8286 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8289 int i, j, split_indx, nkeys, pmax;
8290 MDB_env *env = mc->mc_txn->mt_env;
8292 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8293 MDB_page *copy = NULL;
8294 MDB_page *mp, *rp, *pp;
8299 mp = mc->mc_pg[mc->mc_top];
8300 newindx = mc->mc_ki[mc->mc_top];
8301 nkeys = NUMKEYS(mp);
8303 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8304 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8305 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8307 /* Create a right sibling. */
8308 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8310 rp->mp_pad = mp->mp_pad;
8311 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8313 if (mc->mc_snum < 2) {
8314 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8316 /* shift current top to make room for new parent */
8317 mc->mc_pg[1] = mc->mc_pg[0];
8318 mc->mc_ki[1] = mc->mc_ki[0];
8321 mc->mc_db->md_root = pp->mp_pgno;
8322 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8323 new_root = mc->mc_db->md_depth++;
8325 /* Add left (implicit) pointer. */
8326 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8327 /* undo the pre-push */
8328 mc->mc_pg[0] = mc->mc_pg[1];
8329 mc->mc_ki[0] = mc->mc_ki[1];
8330 mc->mc_db->md_root = mp->mp_pgno;
8331 mc->mc_db->md_depth--;
8338 ptop = mc->mc_top-1;
8339 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8342 mc->mc_flags |= C_SPLITTING;
8343 mdb_cursor_copy(mc, &mn);
8344 mn.mc_pg[mn.mc_top] = rp;
8345 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8347 if (nflags & MDB_APPEND) {
8348 mn.mc_ki[mn.mc_top] = 0;
8350 split_indx = newindx;
8354 split_indx = (nkeys+1) / 2;
8359 unsigned int lsize, rsize, ksize;
8360 /* Move half of the keys to the right sibling */
8361 x = mc->mc_ki[mc->mc_top] - split_indx;
8362 ksize = mc->mc_db->md_pad;
8363 split = LEAF2KEY(mp, split_indx, ksize);
8364 rsize = (nkeys - split_indx) * ksize;
8365 lsize = (nkeys - split_indx) * sizeof(indx_t);
8366 mp->mp_lower -= lsize;
8367 rp->mp_lower += lsize;
8368 mp->mp_upper += rsize - lsize;
8369 rp->mp_upper -= rsize - lsize;
8370 sepkey.mv_size = ksize;
8371 if (newindx == split_indx) {
8372 sepkey.mv_data = newkey->mv_data;
8374 sepkey.mv_data = split;
8377 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8378 memcpy(rp->mp_ptrs, split, rsize);
8379 sepkey.mv_data = rp->mp_ptrs;
8380 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8381 memcpy(ins, newkey->mv_data, ksize);
8382 mp->mp_lower += sizeof(indx_t);
8383 mp->mp_upper -= ksize - sizeof(indx_t);
8386 memcpy(rp->mp_ptrs, split, x * ksize);
8387 ins = LEAF2KEY(rp, x, ksize);
8388 memcpy(ins, newkey->mv_data, ksize);
8389 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8390 rp->mp_lower += sizeof(indx_t);
8391 rp->mp_upper -= ksize - sizeof(indx_t);
8392 mc->mc_ki[mc->mc_top] = x;
8393 mc->mc_pg[mc->mc_top] = rp;
8396 int psize, nsize, k;
8397 /* Maximum free space in an empty page */
8398 pmax = env->me_psize - PAGEHDRSZ;
8400 nsize = mdb_leaf_size(env, newkey, newdata);
8402 nsize = mdb_branch_size(env, newkey);
8403 nsize = EVEN(nsize);
8405 /* grab a page to hold a temporary copy */
8406 copy = mdb_page_malloc(mc->mc_txn, 1);
8411 copy->mp_pgno = mp->mp_pgno;
8412 copy->mp_flags = mp->mp_flags;
8413 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8414 copy->mp_upper = env->me_psize - PAGEBASE;
8416 /* prepare to insert */
8417 for (i=0, j=0; i<nkeys; i++) {
8419 copy->mp_ptrs[j++] = 0;
8421 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8424 /* When items are relatively large the split point needs
8425 * to be checked, because being off-by-one will make the
8426 * difference between success or failure in mdb_node_add.
8428 * It's also relevant if a page happens to be laid out
8429 * such that one half of its nodes are all "small" and
8430 * the other half of its nodes are "large." If the new
8431 * item is also "large" and falls on the half with
8432 * "large" nodes, it also may not fit.
8434 * As a final tweak, if the new item goes on the last
8435 * spot on the page (and thus, onto the new page), bias
8436 * the split so the new page is emptier than the old page.
8437 * This yields better packing during sequential inserts.
8439 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8440 /* Find split point */
8442 if (newindx <= split_indx || newindx >= nkeys) {
8444 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8449 for (; i!=k; i+=j) {
8454 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8455 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8457 if (F_ISSET(node->mn_flags, F_BIGDATA))
8458 psize += sizeof(pgno_t);
8460 psize += NODEDSZ(node);
8462 psize = EVEN(psize);
8464 if (psize > pmax || i == k-j) {
8465 split_indx = i + (j<0);
8470 if (split_indx == newindx) {
8471 sepkey.mv_size = newkey->mv_size;
8472 sepkey.mv_data = newkey->mv_data;
8474 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8475 sepkey.mv_size = node->mn_ksize;
8476 sepkey.mv_data = NODEKEY(node);
8481 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8483 /* Copy separator key to the parent.
8485 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8489 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8494 if (mn.mc_snum == mc->mc_snum) {
8495 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8496 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8497 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8498 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8503 /* Right page might now have changed parent.
8504 * Check if left page also changed parent.
8506 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8507 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8508 for (i=0; i<ptop; i++) {
8509 mc->mc_pg[i] = mn.mc_pg[i];
8510 mc->mc_ki[i] = mn.mc_ki[i];
8512 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8513 if (mn.mc_ki[ptop]) {
8514 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8516 /* find right page's left sibling */
8517 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8518 mdb_cursor_sibling(mc, 0);
8523 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8526 mc->mc_flags ^= C_SPLITTING;
8527 if (rc != MDB_SUCCESS) {
8530 if (nflags & MDB_APPEND) {
8531 mc->mc_pg[mc->mc_top] = rp;
8532 mc->mc_ki[mc->mc_top] = 0;
8533 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8536 for (i=0; i<mc->mc_top; i++)
8537 mc->mc_ki[i] = mn.mc_ki[i];
8538 } else if (!IS_LEAF2(mp)) {
8540 mc->mc_pg[mc->mc_top] = rp;
8545 rkey.mv_data = newkey->mv_data;
8546 rkey.mv_size = newkey->mv_size;
8552 /* Update index for the new key. */
8553 mc->mc_ki[mc->mc_top] = j;
8555 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8556 rkey.mv_data = NODEKEY(node);
8557 rkey.mv_size = node->mn_ksize;
8559 xdata.mv_data = NODEDATA(node);
8560 xdata.mv_size = NODEDSZ(node);
8563 pgno = NODEPGNO(node);
8564 flags = node->mn_flags;
8567 if (!IS_LEAF(mp) && j == 0) {
8568 /* First branch index doesn't need key data. */
8572 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8578 mc->mc_pg[mc->mc_top] = copy;
8583 } while (i != split_indx);
8585 nkeys = NUMKEYS(copy);
8586 for (i=0; i<nkeys; i++)
8587 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8588 mp->mp_lower = copy->mp_lower;
8589 mp->mp_upper = copy->mp_upper;
8590 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8591 env->me_psize - copy->mp_upper - PAGEBASE);
8593 /* reset back to original page */
8594 if (newindx < split_indx) {
8595 mc->mc_pg[mc->mc_top] = mp;
8596 if (nflags & MDB_RESERVE) {
8597 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8598 if (!(node->mn_flags & F_BIGDATA))
8599 newdata->mv_data = NODEDATA(node);
8602 mc->mc_pg[mc->mc_top] = rp;
8604 /* Make sure mc_ki is still valid.
8606 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8607 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8608 for (i=0; i<=ptop; i++) {
8609 mc->mc_pg[i] = mn.mc_pg[i];
8610 mc->mc_ki[i] = mn.mc_ki[i];
8617 /* Adjust other cursors pointing to mp */
8618 MDB_cursor *m2, *m3;
8619 MDB_dbi dbi = mc->mc_dbi;
8620 int fixup = NUMKEYS(mp);
8622 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8623 if (mc->mc_flags & C_SUB)
8624 m3 = &m2->mc_xcursor->mx_cursor;
8629 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8631 if (m3->mc_flags & C_SPLITTING)
8636 for (k=new_root; k>=0; k--) {
8637 m3->mc_ki[k+1] = m3->mc_ki[k];
8638 m3->mc_pg[k+1] = m3->mc_pg[k];
8640 if (m3->mc_ki[0] >= split_indx) {
8645 m3->mc_pg[0] = mc->mc_pg[0];
8649 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8650 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8651 m3->mc_ki[mc->mc_top]++;
8652 if (m3->mc_ki[mc->mc_top] >= fixup) {
8653 m3->mc_pg[mc->mc_top] = rp;
8654 m3->mc_ki[mc->mc_top] -= fixup;
8655 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8657 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8658 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8663 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8666 if (copy) /* tmp page */
8667 mdb_page_free(env, copy);
8669 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8674 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8675 MDB_val *key, MDB_val *data, unsigned int flags)
8680 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8683 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8686 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8687 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8689 mdb_cursor_init(&mc, txn, dbi, &mx);
8690 return mdb_cursor_put(&mc, key, data, flags);
8694 #define MDB_WBUF (1024*1024)
8697 /** State needed for a compacting copy. */
8698 typedef struct mdb_copy {
8699 pthread_mutex_t mc_mutex;
8700 pthread_cond_t mc_cond;
8707 pgno_t mc_next_pgno;
8710 volatile int mc_new;
8715 /** Dedicated writer thread for compacting copy. */
8716 static THREAD_RET ESECT
8717 mdb_env_copythr(void *arg)
8721 int toggle = 0, wsize, rc;
8724 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8727 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8730 pthread_mutex_lock(&my->mc_mutex);
8732 pthread_cond_signal(&my->mc_cond);
8735 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8736 if (my->mc_new < 0) {
8741 wsize = my->mc_wlen[toggle];
8742 ptr = my->mc_wbuf[toggle];
8745 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8749 } else if (len > 0) {
8763 /* If there's an overflow page tail, write it too */
8764 if (my->mc_olen[toggle]) {
8765 wsize = my->mc_olen[toggle];
8766 ptr = my->mc_over[toggle];
8767 my->mc_olen[toggle] = 0;
8770 my->mc_wlen[toggle] = 0;
8772 pthread_cond_signal(&my->mc_cond);
8774 pthread_cond_signal(&my->mc_cond);
8775 pthread_mutex_unlock(&my->mc_mutex);
8776 return (THREAD_RET)0;
8780 /** Tell the writer thread there's a buffer ready to write */
8782 mdb_env_cthr_toggle(mdb_copy *my, int st)
8784 int toggle = my->mc_toggle ^ 1;
8785 pthread_mutex_lock(&my->mc_mutex);
8786 if (my->mc_status) {
8787 pthread_mutex_unlock(&my->mc_mutex);
8788 return my->mc_status;
8790 while (my->mc_new == 1)
8791 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8793 my->mc_toggle = toggle;
8794 pthread_cond_signal(&my->mc_cond);
8795 pthread_mutex_unlock(&my->mc_mutex);
8799 /** Depth-first tree traversal for compacting copy. */
8801 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8804 MDB_txn *txn = my->mc_txn;
8806 MDB_page *mo, *mp, *leaf;
8811 /* Empty DB, nothing to do */
8812 if (*pg == P_INVALID)
8819 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8822 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8826 /* Make cursor pages writable */
8827 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8831 for (i=0; i<mc.mc_top; i++) {
8832 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8833 mc.mc_pg[i] = (MDB_page *)ptr;
8834 ptr += my->mc_env->me_psize;
8837 /* This is writable space for a leaf page. Usually not needed. */
8838 leaf = (MDB_page *)ptr;
8840 toggle = my->mc_toggle;
8841 while (mc.mc_snum > 0) {
8843 mp = mc.mc_pg[mc.mc_top];
8847 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8848 for (i=0; i<n; i++) {
8849 ni = NODEPTR(mp, i);
8850 if (ni->mn_flags & F_BIGDATA) {
8854 /* Need writable leaf */
8856 mc.mc_pg[mc.mc_top] = leaf;
8857 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8859 ni = NODEPTR(mp, i);
8862 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8863 rc = mdb_page_get(txn, pg, &omp, NULL);
8866 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8867 rc = mdb_env_cthr_toggle(my, 1);
8870 toggle = my->mc_toggle;
8872 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8873 memcpy(mo, omp, my->mc_env->me_psize);
8874 mo->mp_pgno = my->mc_next_pgno;
8875 my->mc_next_pgno += omp->mp_pages;
8876 my->mc_wlen[toggle] += my->mc_env->me_psize;
8877 if (omp->mp_pages > 1) {
8878 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8879 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8880 rc = mdb_env_cthr_toggle(my, 1);
8883 toggle = my->mc_toggle;
8885 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8886 } else if (ni->mn_flags & F_SUBDATA) {
8889 /* Need writable leaf */
8891 mc.mc_pg[mc.mc_top] = leaf;
8892 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8894 ni = NODEPTR(mp, i);
8897 memcpy(&db, NODEDATA(ni), sizeof(db));
8898 my->mc_toggle = toggle;
8899 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8902 toggle = my->mc_toggle;
8903 memcpy(NODEDATA(ni), &db, sizeof(db));
8908 mc.mc_ki[mc.mc_top]++;
8909 if (mc.mc_ki[mc.mc_top] < n) {
8912 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8914 rc = mdb_page_get(txn, pg, &mp, NULL);
8919 mc.mc_ki[mc.mc_top] = 0;
8920 if (IS_BRANCH(mp)) {
8921 /* Whenever we advance to a sibling branch page,
8922 * we must proceed all the way down to its first leaf.
8924 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8927 mc.mc_pg[mc.mc_top] = mp;
8931 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8932 rc = mdb_env_cthr_toggle(my, 1);
8935 toggle = my->mc_toggle;
8937 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8938 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8939 mo->mp_pgno = my->mc_next_pgno++;
8940 my->mc_wlen[toggle] += my->mc_env->me_psize;
8942 /* Update parent if there is one */
8943 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8944 SETPGNO(ni, mo->mp_pgno);
8945 mdb_cursor_pop(&mc);
8947 /* Otherwise we're done */
8957 /** Copy environment with compaction. */
8959 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8964 MDB_txn *txn = NULL;
8969 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8970 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8971 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8972 if (my.mc_wbuf[0] == NULL)
8975 pthread_mutex_init(&my.mc_mutex, NULL);
8976 pthread_cond_init(&my.mc_cond, NULL);
8977 #ifdef HAVE_MEMALIGN
8978 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8979 if (my.mc_wbuf[0] == NULL)
8982 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8987 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8988 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8993 my.mc_next_pgno = NUM_METAS;
8999 THREAD_CREATE(thr, mdb_env_copythr, &my);
9001 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9005 mp = (MDB_page *)my.mc_wbuf[0];
9006 memset(mp, 0, NUM_METAS * env->me_psize);
9008 mp->mp_flags = P_META;
9009 mm = (MDB_meta *)METADATA(mp);
9010 mdb_env_init_meta0(env, mm);
9011 mm->mm_address = env->me_metas[0]->mm_address;
9013 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9015 mp->mp_flags = P_META;
9016 *(MDB_meta *)METADATA(mp) = *mm;
9017 mm = (MDB_meta *)METADATA(mp);
9019 /* Count the number of free pages, subtract from lastpg to find
9020 * number of active pages
9023 MDB_ID freecount = 0;
9026 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9027 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9028 freecount += *(MDB_ID *)data.mv_data;
9029 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9030 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9031 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9033 /* Set metapage 1 */
9034 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
9035 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9036 if (mm->mm_last_pg > NUM_METAS-1) {
9037 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
9040 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
9043 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9045 pthread_mutex_lock(&my.mc_mutex);
9047 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9048 pthread_mutex_unlock(&my.mc_mutex);
9049 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
9050 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
9051 rc = mdb_env_cthr_toggle(&my, 1);
9052 mdb_env_cthr_toggle(&my, -1);
9053 pthread_mutex_lock(&my.mc_mutex);
9055 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9056 pthread_mutex_unlock(&my.mc_mutex);
9061 CloseHandle(my.mc_cond);
9062 CloseHandle(my.mc_mutex);
9063 _aligned_free(my.mc_wbuf[0]);
9065 pthread_cond_destroy(&my.mc_cond);
9066 pthread_mutex_destroy(&my.mc_mutex);
9067 free(my.mc_wbuf[0]);
9072 /** Copy environment as-is. */
9074 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9076 MDB_txn *txn = NULL;
9077 mdb_mutexref_t wmutex = NULL;
9083 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9087 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9090 /* Do the lock/unlock of the reader mutex before starting the
9091 * write txn. Otherwise other read txns could block writers.
9093 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9098 /* We must start the actual read txn after blocking writers */
9099 mdb_txn_end(txn, MDB_END_RESET_TMP);
9101 /* Temporarily block writers until we snapshot the meta pages */
9102 wmutex = env->me_wmutex;
9103 if (LOCK_MUTEX(rc, env, wmutex))
9106 rc = mdb_txn_renew0(txn);
9108 UNLOCK_MUTEX(wmutex);
9113 wsize = env->me_psize * NUM_METAS;
9117 DO_WRITE(rc, fd, ptr, w2, len);
9121 } else if (len > 0) {
9127 /* Non-blocking or async handles are not supported */
9133 UNLOCK_MUTEX(wmutex);
9138 w2 = txn->mt_next_pgno * env->me_psize;
9141 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9148 if (wsize > MAX_WRITE)
9152 DO_WRITE(rc, fd, ptr, w2, len);
9156 } else if (len > 0) {
9173 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9175 if (flags & MDB_CP_COMPACT)
9176 return mdb_env_copyfd1(env, fd);
9178 return mdb_env_copyfd0(env, fd);
9182 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9184 return mdb_env_copyfd2(env, fd, 0);
9188 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9192 HANDLE newfd = INVALID_HANDLE_VALUE;
9194 if (env->me_flags & MDB_NOSUBDIR) {
9195 lpath = (char *)path;
9198 len += sizeof(DATANAME);
9199 lpath = malloc(len);
9202 sprintf(lpath, "%s" DATANAME, path);
9205 /* The destination path must exist, but the destination file must not.
9206 * We don't want the OS to cache the writes, since the source data is
9207 * already in the OS cache.
9210 newfd = CreateFileA(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9211 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9213 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9215 if (newfd == INVALID_HANDLE_VALUE) {
9220 if (env->me_psize >= env->me_os_psize) {
9222 /* Set O_DIRECT if the file system supports it */
9223 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9224 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9226 #ifdef F_NOCACHE /* __APPLE__ */
9227 rc = fcntl(newfd, F_NOCACHE, 1);
9235 rc = mdb_env_copyfd2(env, newfd, flags);
9238 if (!(env->me_flags & MDB_NOSUBDIR))
9240 if (newfd != INVALID_HANDLE_VALUE)
9241 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9248 mdb_env_copy(MDB_env *env, const char *path)
9250 return mdb_env_copy2(env, path, 0);
9254 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9256 if (flag & ~CHANGEABLE)
9259 env->me_flags |= flag;
9261 env->me_flags &= ~flag;
9266 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9271 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9276 mdb_env_set_userctx(MDB_env *env, void *ctx)
9280 env->me_userctx = ctx;
9285 mdb_env_get_userctx(MDB_env *env)
9287 return env ? env->me_userctx : NULL;
9291 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9296 env->me_assert_func = func;
9302 mdb_env_get_path(MDB_env *env, const char **arg)
9307 *arg = env->me_path;
9312 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9321 /** Common code for #mdb_stat() and #mdb_env_stat().
9322 * @param[in] env the environment to operate in.
9323 * @param[in] db the #MDB_db record containing the stats to return.
9324 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9325 * @return 0, this function always succeeds.
9328 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9330 arg->ms_psize = env->me_psize;
9331 arg->ms_depth = db->md_depth;
9332 arg->ms_branch_pages = db->md_branch_pages;
9333 arg->ms_leaf_pages = db->md_leaf_pages;
9334 arg->ms_overflow_pages = db->md_overflow_pages;
9335 arg->ms_entries = db->md_entries;
9341 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9345 if (env == NULL || arg == NULL)
9348 meta = mdb_env_pick_meta(env);
9350 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9354 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9358 if (env == NULL || arg == NULL)
9361 meta = mdb_env_pick_meta(env);
9362 arg->me_mapaddr = meta->mm_address;
9363 arg->me_last_pgno = meta->mm_last_pg;
9364 arg->me_last_txnid = meta->mm_txnid;
9366 arg->me_mapsize = env->me_mapsize;
9367 arg->me_maxreaders = env->me_maxreaders;
9368 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9372 /** Set the default comparison functions for a database.
9373 * Called immediately after a database is opened to set the defaults.
9374 * The user can then override them with #mdb_set_compare() or
9375 * #mdb_set_dupsort().
9376 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9377 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9380 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9382 uint16_t f = txn->mt_dbs[dbi].md_flags;
9384 txn->mt_dbxs[dbi].md_cmp =
9385 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9386 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9388 txn->mt_dbxs[dbi].md_dcmp =
9389 !(f & MDB_DUPSORT) ? 0 :
9390 ((f & MDB_INTEGERDUP)
9391 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9392 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9395 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9401 int rc, dbflag, exact;
9402 unsigned int unused = 0, seq;
9405 if (flags & ~VALID_FLAGS)
9407 if (txn->mt_flags & MDB_TXN_BLOCKED)
9413 if (flags & PERSISTENT_FLAGS) {
9414 uint16_t f2 = flags & PERSISTENT_FLAGS;
9415 /* make sure flag changes get committed */
9416 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9417 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9418 txn->mt_flags |= MDB_TXN_DIRTY;
9421 mdb_default_cmp(txn, MAIN_DBI);
9425 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9426 mdb_default_cmp(txn, MAIN_DBI);
9429 /* Is the DB already open? */
9431 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9432 if (!txn->mt_dbxs[i].md_name.mv_size) {
9433 /* Remember this free slot */
9434 if (!unused) unused = i;
9437 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9438 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9444 /* If no free slot and max hit, fail */
9445 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9446 return MDB_DBS_FULL;
9448 /* Cannot mix named databases with some mainDB flags */
9449 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9450 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9452 /* Find the DB info */
9453 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9456 key.mv_data = (void *)name;
9457 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9458 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9459 if (rc == MDB_SUCCESS) {
9460 /* make sure this is actually a DB */
9461 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9462 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9463 return MDB_INCOMPATIBLE;
9464 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9465 /* Create if requested */
9466 data.mv_size = sizeof(MDB_db);
9467 data.mv_data = &dummy;
9468 memset(&dummy, 0, sizeof(dummy));
9469 dummy.md_root = P_INVALID;
9470 dummy.md_flags = flags & PERSISTENT_FLAGS;
9471 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9475 /* OK, got info, add to table */
9476 if (rc == MDB_SUCCESS) {
9477 unsigned int slot = unused ? unused : txn->mt_numdbs;
9478 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9479 txn->mt_dbxs[slot].md_name.mv_size = len;
9480 txn->mt_dbxs[slot].md_rel = NULL;
9481 txn->mt_dbflags[slot] = dbflag;
9482 /* txn-> and env-> are the same in read txns, use
9483 * tmp variable to avoid undefined assignment
9485 seq = ++txn->mt_env->me_dbiseqs[slot];
9486 txn->mt_dbiseqs[slot] = seq;
9488 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9490 mdb_default_cmp(txn, slot);
9500 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9502 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9505 if (txn->mt_flags & MDB_TXN_BLOCKED)
9508 if (txn->mt_dbflags[dbi] & DB_STALE) {
9511 /* Stale, must read the DB's root. cursor_init does it for us. */
9512 mdb_cursor_init(&mc, txn, dbi, &mx);
9514 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9517 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9520 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9522 ptr = env->me_dbxs[dbi].md_name.mv_data;
9523 /* If there was no name, this was already closed */
9525 env->me_dbxs[dbi].md_name.mv_data = NULL;
9526 env->me_dbxs[dbi].md_name.mv_size = 0;
9527 env->me_dbflags[dbi] = 0;
9528 env->me_dbiseqs[dbi]++;
9533 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9535 /* We could return the flags for the FREE_DBI too but what's the point? */
9536 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9538 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9542 /** Add all the DB's pages to the free list.
9543 * @param[in] mc Cursor on the DB to free.
9544 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9545 * @return 0 on success, non-zero on failure.
9548 mdb_drop0(MDB_cursor *mc, int subs)
9552 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9553 if (rc == MDB_SUCCESS) {
9554 MDB_txn *txn = mc->mc_txn;
9559 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9560 * This also avoids any P_LEAF2 pages, which have no nodes.
9562 if (mc->mc_flags & C_SUB)
9565 mdb_cursor_copy(mc, &mx);
9566 while (mc->mc_snum > 0) {
9567 MDB_page *mp = mc->mc_pg[mc->mc_top];
9568 unsigned n = NUMKEYS(mp);
9570 for (i=0; i<n; i++) {
9571 ni = NODEPTR(mp, i);
9572 if (ni->mn_flags & F_BIGDATA) {
9575 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9576 rc = mdb_page_get(txn, pg, &omp, NULL);
9579 mdb_cassert(mc, IS_OVERFLOW(omp));
9580 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9584 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9585 mdb_xcursor_init1(mc, ni);
9586 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9592 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9594 for (i=0; i<n; i++) {
9596 ni = NODEPTR(mp, i);
9599 mdb_midl_xappend(txn->mt_free_pgs, pg);
9604 mc->mc_ki[mc->mc_top] = i;
9605 rc = mdb_cursor_sibling(mc, 1);
9607 if (rc != MDB_NOTFOUND)
9609 /* no more siblings, go back to beginning
9610 * of previous level.
9614 for (i=1; i<mc->mc_snum; i++) {
9616 mc->mc_pg[i] = mx.mc_pg[i];
9621 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9624 txn->mt_flags |= MDB_TXN_ERROR;
9625 } else if (rc == MDB_NOTFOUND) {
9631 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9633 MDB_cursor *mc, *m2;
9636 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9639 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9642 if (TXN_DBI_CHANGED(txn, dbi))
9645 rc = mdb_cursor_open(txn, dbi, &mc);
9649 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9650 /* Invalidate the dropped DB's cursors */
9651 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9652 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9656 /* Can't delete the main DB */
9657 if (del && dbi >= CORE_DBS) {
9658 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9660 txn->mt_dbflags[dbi] = DB_STALE;
9661 mdb_dbi_close(txn->mt_env, dbi);
9663 txn->mt_flags |= MDB_TXN_ERROR;
9666 /* reset the DB record, mark it dirty */
9667 txn->mt_dbflags[dbi] |= DB_DIRTY;
9668 txn->mt_dbs[dbi].md_depth = 0;
9669 txn->mt_dbs[dbi].md_branch_pages = 0;
9670 txn->mt_dbs[dbi].md_leaf_pages = 0;
9671 txn->mt_dbs[dbi].md_overflow_pages = 0;
9672 txn->mt_dbs[dbi].md_entries = 0;
9673 txn->mt_dbs[dbi].md_root = P_INVALID;
9675 txn->mt_flags |= MDB_TXN_DIRTY;
9678 mdb_cursor_close(mc);
9682 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9684 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9687 txn->mt_dbxs[dbi].md_cmp = cmp;
9691 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9693 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9696 txn->mt_dbxs[dbi].md_dcmp = cmp;
9700 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9702 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9705 txn->mt_dbxs[dbi].md_rel = rel;
9709 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9711 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9714 txn->mt_dbxs[dbi].md_relctx = ctx;
9719 mdb_env_get_maxkeysize(MDB_env *env)
9721 return ENV_MAXKEY(env);
9725 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9727 unsigned int i, rdrs;
9730 int rc = 0, first = 1;
9734 if (!env->me_txns) {
9735 return func("(no reader locks)\n", ctx);
9737 rdrs = env->me_txns->mti_numreaders;
9738 mr = env->me_txns->mti_readers;
9739 for (i=0; i<rdrs; i++) {
9741 txnid_t txnid = mr[i].mr_txnid;
9742 sprintf(buf, txnid == (txnid_t)-1 ?
9743 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9744 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9747 rc = func(" pid thread txnid\n", ctx);
9751 rc = func(buf, ctx);
9757 rc = func("(no active readers)\n", ctx);
9762 /** Insert pid into list if not already present.
9763 * return -1 if already present.
9766 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9768 /* binary search of pid in list */
9770 unsigned cursor = 1;
9772 unsigned n = ids[0];
9775 unsigned pivot = n >> 1;
9776 cursor = base + pivot + 1;
9777 val = pid - ids[cursor];
9782 } else if ( val > 0 ) {
9787 /* found, so it's a duplicate */
9796 for (n = ids[0]; n > cursor; n--)
9803 mdb_reader_check(MDB_env *env, int *dead)
9809 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9812 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9814 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9816 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9817 unsigned int i, j, rdrs;
9819 MDB_PID_T *pids, pid;
9820 int rc = MDB_SUCCESS, count = 0;
9822 rdrs = env->me_txns->mti_numreaders;
9823 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9827 mr = env->me_txns->mti_readers;
9828 for (i=0; i<rdrs; i++) {
9830 if (pid && pid != env->me_pid) {
9831 if (mdb_pid_insert(pids, pid) == 0) {
9832 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9833 /* Stale reader found */
9836 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9837 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9839 rdrs = 0; /* the above checked all readers */
9841 /* Recheck, a new process may have reused pid */
9842 if (mdb_reader_pid(env, Pidcheck, pid))
9847 if (mr[j].mr_pid == pid) {
9848 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9849 (unsigned) pid, mr[j].mr_txnid));
9854 UNLOCK_MUTEX(rmutex);
9865 #ifdef MDB_ROBUST_SUPPORTED
9866 /** Handle #LOCK_MUTEX0() failure.
9867 * Try to repair the lock file if the mutex owner died.
9868 * @param[in] env the environment handle
9869 * @param[in] mutex LOCK_MUTEX0() mutex
9870 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9871 * @return 0 on success with the mutex locked, or an error code on failure.
9874 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9879 if (rc == MDB_OWNERDEAD) {
9880 /* We own the mutex. Clean up after dead previous owner. */
9882 rlocked = (mutex == env->me_rmutex);
9884 /* Keep mti_txnid updated, otherwise next writer can
9885 * overwrite data which latest meta page refers to.
9887 meta = mdb_env_pick_meta(env);
9888 env->me_txns->mti_txnid = meta->mm_txnid;
9889 /* env is hosed if the dead thread was ours */
9891 env->me_flags |= MDB_FATAL_ERROR;
9896 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9897 (rc ? "this process' env is hosed" : "recovering")));
9898 rc2 = mdb_reader_check0(env, rlocked, NULL);
9900 rc2 = mdb_mutex_consistent(mutex);
9901 if (rc || (rc = rc2)) {
9902 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9903 UNLOCK_MUTEX(mutex);
9909 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9914 #endif /* MDB_ROBUST_SUPPORTED */