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);
1388 static int mdb_drop0(MDB_cursor *mc, int subs);
1389 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1390 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1393 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1396 /** Compare two items pointing at size_t's of unknown alignment. */
1397 #ifdef MISALIGNED_OK
1398 # define mdb_cmp_clong mdb_cmp_long
1400 # define mdb_cmp_clong mdb_cmp_cint
1404 static SECURITY_DESCRIPTOR mdb_null_sd;
1405 static SECURITY_ATTRIBUTES mdb_all_sa;
1406 static int mdb_sec_inited;
1409 /** Return the library version info. */
1411 mdb_version(int *major, int *minor, int *patch)
1413 if (major) *major = MDB_VERSION_MAJOR;
1414 if (minor) *minor = MDB_VERSION_MINOR;
1415 if (patch) *patch = MDB_VERSION_PATCH;
1416 return MDB_VERSION_STRING;
1419 /** Table of descriptions for LMDB @ref errors */
1420 static char *const mdb_errstr[] = {
1421 "MDB_KEYEXIST: Key/data pair already exists",
1422 "MDB_NOTFOUND: No matching key/data pair found",
1423 "MDB_PAGE_NOTFOUND: Requested page not found",
1424 "MDB_CORRUPTED: Located page was wrong type",
1425 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1426 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1427 "MDB_INVALID: File is not an LMDB file",
1428 "MDB_MAP_FULL: Environment mapsize limit reached",
1429 "MDB_DBS_FULL: Environment maxdbs limit reached",
1430 "MDB_READERS_FULL: Environment maxreaders limit reached",
1431 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1432 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1433 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1434 "MDB_PAGE_FULL: Internal error - page has no more space",
1435 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1436 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1437 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1438 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1439 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1440 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1444 mdb_strerror(int err)
1447 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1448 * This works as long as no function between the call to mdb_strerror
1449 * and the actual use of the message uses more than 4K of stack.
1452 char buf[1024], *ptr = buf;
1456 return ("Successful return: 0");
1458 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1459 i = err - MDB_KEYEXIST;
1460 return mdb_errstr[i];
1464 /* These are the C-runtime error codes we use. The comment indicates
1465 * their numeric value, and the Win32 error they would correspond to
1466 * if the error actually came from a Win32 API. A major mess, we should
1467 * have used LMDB-specific error codes for everything.
1470 case ENOENT: /* 2, FILE_NOT_FOUND */
1471 case EIO: /* 5, ACCESS_DENIED */
1472 case ENOMEM: /* 12, INVALID_ACCESS */
1473 case EACCES: /* 13, INVALID_DATA */
1474 case EBUSY: /* 16, CURRENT_DIRECTORY */
1475 case EINVAL: /* 22, BAD_COMMAND */
1476 case ENOSPC: /* 28, OUT_OF_PAPER */
1477 return strerror(err);
1482 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1483 FORMAT_MESSAGE_IGNORE_INSERTS,
1484 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1487 return strerror(err);
1491 /** assert(3) variant in cursor context */
1492 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1493 /** assert(3) variant in transaction context */
1494 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1495 /** assert(3) variant in environment context */
1496 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1499 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1500 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1503 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1504 const char *func, const char *file, int line)
1507 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1508 file, line, expr_txt, func);
1509 if (env->me_assert_func)
1510 env->me_assert_func(env, buf);
1511 fprintf(stderr, "%s\n", buf);
1515 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1519 /** Return the page number of \b mp which may be sub-page, for debug output */
1521 mdb_dbg_pgno(MDB_page *mp)
1524 COPY_PGNO(ret, mp->mp_pgno);
1528 /** Display a key in hexadecimal and return the address of the result.
1529 * @param[in] key the key to display
1530 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1531 * @return The key in hexadecimal form.
1534 mdb_dkey(MDB_val *key, char *buf)
1537 unsigned char *c = key->mv_data;
1543 if (key->mv_size > DKBUF_MAXKEYSIZE)
1544 return "MDB_MAXKEYSIZE";
1545 /* may want to make this a dynamic check: if the key is mostly
1546 * printable characters, print it as-is instead of converting to hex.
1550 for (i=0; i<key->mv_size; i++)
1551 ptr += sprintf(ptr, "%02x", *c++);
1553 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1559 mdb_leafnode_type(MDB_node *n)
1561 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1562 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1563 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1566 /** Display all the keys in the page. */
1568 mdb_page_list(MDB_page *mp)
1570 pgno_t pgno = mdb_dbg_pgno(mp);
1571 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1573 unsigned int i, nkeys, nsize, total = 0;
1577 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1578 case P_BRANCH: type = "Branch page"; break;
1579 case P_LEAF: type = "Leaf page"; break;
1580 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1581 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1582 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1584 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1585 pgno, mp->mp_pages, state);
1588 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1589 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1592 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1596 nkeys = NUMKEYS(mp);
1597 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1599 for (i=0; i<nkeys; i++) {
1600 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1601 key.mv_size = nsize = mp->mp_pad;
1602 key.mv_data = LEAF2KEY(mp, i, nsize);
1604 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1607 node = NODEPTR(mp, i);
1608 key.mv_size = node->mn_ksize;
1609 key.mv_data = node->mn_data;
1610 nsize = NODESIZE + key.mv_size;
1611 if (IS_BRANCH(mp)) {
1612 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1616 if (F_ISSET(node->mn_flags, F_BIGDATA))
1617 nsize += sizeof(pgno_t);
1619 nsize += NODEDSZ(node);
1621 nsize += sizeof(indx_t);
1622 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1623 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1625 total = EVEN(total);
1627 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1628 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1632 mdb_cursor_chk(MDB_cursor *mc)
1638 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1639 for (i=0; i<mc->mc_top; i++) {
1641 node = NODEPTR(mp, mc->mc_ki[i]);
1642 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1645 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1651 /** Count all the pages in each DB and in the freelist
1652 * and make sure it matches the actual number of pages
1654 * All named DBs must be open for a correct count.
1656 static void mdb_audit(MDB_txn *txn)
1660 MDB_ID freecount, count;
1665 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1666 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1667 freecount += *(MDB_ID *)data.mv_data;
1668 mdb_tassert(txn, rc == MDB_NOTFOUND);
1671 for (i = 0; i<txn->mt_numdbs; i++) {
1673 if (!(txn->mt_dbflags[i] & DB_VALID))
1675 mdb_cursor_init(&mc, txn, i, &mx);
1676 if (txn->mt_dbs[i].md_root == P_INVALID)
1678 count += txn->mt_dbs[i].md_branch_pages +
1679 txn->mt_dbs[i].md_leaf_pages +
1680 txn->mt_dbs[i].md_overflow_pages;
1681 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1682 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1683 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1686 mp = mc.mc_pg[mc.mc_top];
1687 for (j=0; j<NUMKEYS(mp); j++) {
1688 MDB_node *leaf = NODEPTR(mp, j);
1689 if (leaf->mn_flags & F_SUBDATA) {
1691 memcpy(&db, NODEDATA(leaf), sizeof(db));
1692 count += db.md_branch_pages + db.md_leaf_pages +
1693 db.md_overflow_pages;
1697 mdb_tassert(txn, rc == MDB_NOTFOUND);
1700 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1701 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1702 txn->mt_txnid, freecount, count+NUM_METAS,
1703 freecount+count+NUM_METAS, txn->mt_next_pgno);
1709 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1711 return txn->mt_dbxs[dbi].md_cmp(a, b);
1715 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1717 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1718 #if UINT_MAX < SIZE_MAX
1719 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1720 dcmp = mdb_cmp_clong;
1725 /** Allocate memory for a page.
1726 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1729 mdb_page_malloc(MDB_txn *txn, unsigned num)
1731 MDB_env *env = txn->mt_env;
1732 MDB_page *ret = env->me_dpages;
1733 size_t psize = env->me_psize, sz = psize, off;
1734 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1735 * For a single page alloc, we init everything after the page header.
1736 * For multi-page, we init the final page; if the caller needed that
1737 * many pages they will be filling in at least up to the last page.
1741 VGMEMP_ALLOC(env, ret, sz);
1742 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1743 env->me_dpages = ret->mp_next;
1746 psize -= off = PAGEHDRSZ;
1751 if ((ret = malloc(sz)) != NULL) {
1752 VGMEMP_ALLOC(env, ret, sz);
1753 if (!(env->me_flags & MDB_NOMEMINIT)) {
1754 memset((char *)ret + off, 0, psize);
1758 txn->mt_flags |= MDB_TXN_ERROR;
1762 /** Free a single page.
1763 * Saves single pages to a list, for future reuse.
1764 * (This is not used for multi-page overflow pages.)
1767 mdb_page_free(MDB_env *env, MDB_page *mp)
1769 mp->mp_next = env->me_dpages;
1770 VGMEMP_FREE(env, mp);
1771 env->me_dpages = mp;
1774 /** Free a dirty page */
1776 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1778 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1779 mdb_page_free(env, dp);
1781 /* large pages just get freed directly */
1782 VGMEMP_FREE(env, dp);
1787 /** Return all dirty pages to dpage list */
1789 mdb_dlist_free(MDB_txn *txn)
1791 MDB_env *env = txn->mt_env;
1792 MDB_ID2L dl = txn->mt_u.dirty_list;
1793 unsigned i, n = dl[0].mid;
1795 for (i = 1; i <= n; i++) {
1796 mdb_dpage_free(env, dl[i].mptr);
1801 /** Loosen or free a single page.
1802 * Saves single pages to a list for future reuse
1803 * in this same txn. It has been pulled from the freeDB
1804 * and already resides on the dirty list, but has been
1805 * deleted. Use these pages first before pulling again
1808 * If the page wasn't dirtied in this txn, just add it
1809 * to this txn's free list.
1812 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1815 pgno_t pgno = mp->mp_pgno;
1816 MDB_txn *txn = mc->mc_txn;
1818 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1819 if (txn->mt_parent) {
1820 MDB_ID2 *dl = txn->mt_u.dirty_list;
1821 /* If txn has a parent, make sure the page is in our
1825 unsigned x = mdb_mid2l_search(dl, pgno);
1826 if (x <= dl[0].mid && dl[x].mid == pgno) {
1827 if (mp != dl[x].mptr) { /* bad cursor? */
1828 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1829 txn->mt_flags |= MDB_TXN_ERROR;
1830 return MDB_CORRUPTED;
1837 /* no parent txn, so it's just ours */
1842 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1844 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1845 txn->mt_loose_pgs = mp;
1846 txn->mt_loose_count++;
1847 mp->mp_flags |= P_LOOSE;
1849 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1857 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1858 * @param[in] mc A cursor handle for the current operation.
1859 * @param[in] pflags Flags of the pages to update:
1860 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1861 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1862 * @return 0 on success, non-zero on failure.
1865 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1867 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1868 MDB_txn *txn = mc->mc_txn;
1874 int rc = MDB_SUCCESS, level;
1876 /* Mark pages seen by cursors */
1877 if (mc->mc_flags & C_UNTRACK)
1878 mc = NULL; /* will find mc in mt_cursors */
1879 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1880 for (; mc; mc=mc->mc_next) {
1881 if (!(mc->mc_flags & C_INITIALIZED))
1883 for (m3 = mc;; m3 = &mx->mx_cursor) {
1885 for (j=0; j<m3->mc_snum; j++) {
1887 if ((mp->mp_flags & Mask) == pflags)
1888 mp->mp_flags ^= P_KEEP;
1890 mx = m3->mc_xcursor;
1891 /* Proceed to mx if it is at a sub-database */
1892 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1894 if (! (mp && (mp->mp_flags & P_LEAF)))
1896 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1897 if (!(leaf->mn_flags & F_SUBDATA))
1906 /* Mark dirty root pages */
1907 for (i=0; i<txn->mt_numdbs; i++) {
1908 if (txn->mt_dbflags[i] & DB_DIRTY) {
1909 pgno_t pgno = txn->mt_dbs[i].md_root;
1910 if (pgno == P_INVALID)
1912 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1914 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1915 dp->mp_flags ^= P_KEEP;
1923 static int mdb_page_flush(MDB_txn *txn, int keep);
1925 /** Spill pages from the dirty list back to disk.
1926 * This is intended to prevent running into #MDB_TXN_FULL situations,
1927 * but note that they may still occur in a few cases:
1928 * 1) our estimate of the txn size could be too small. Currently this
1929 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1930 * 2) child txns may run out of space if their parents dirtied a
1931 * lot of pages and never spilled them. TODO: we probably should do
1932 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1933 * the parent's dirty_room is below a given threshold.
1935 * Otherwise, if not using nested txns, it is expected that apps will
1936 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1937 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1938 * If the txn never references them again, they can be left alone.
1939 * If the txn only reads them, they can be used without any fuss.
1940 * If the txn writes them again, they can be dirtied immediately without
1941 * going thru all of the work of #mdb_page_touch(). Such references are
1942 * handled by #mdb_page_unspill().
1944 * Also note, we never spill DB root pages, nor pages of active cursors,
1945 * because we'll need these back again soon anyway. And in nested txns,
1946 * we can't spill a page in a child txn if it was already spilled in a
1947 * parent txn. That would alter the parent txns' data even though
1948 * the child hasn't committed yet, and we'd have no way to undo it if
1949 * the child aborted.
1951 * @param[in] m0 cursor A cursor handle identifying the transaction and
1952 * database for which we are checking space.
1953 * @param[in] key For a put operation, the key being stored.
1954 * @param[in] data For a put operation, the data being stored.
1955 * @return 0 on success, non-zero on failure.
1958 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1960 MDB_txn *txn = m0->mc_txn;
1962 MDB_ID2L dl = txn->mt_u.dirty_list;
1963 unsigned int i, j, need;
1966 if (m0->mc_flags & C_SUB)
1969 /* Estimate how much space this op will take */
1970 i = m0->mc_db->md_depth;
1971 /* Named DBs also dirty the main DB */
1972 if (m0->mc_dbi >= CORE_DBS)
1973 i += txn->mt_dbs[MAIN_DBI].md_depth;
1974 /* For puts, roughly factor in the key+data size */
1976 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1977 i += i; /* double it for good measure */
1980 if (txn->mt_dirty_room > i)
1983 if (!txn->mt_spill_pgs) {
1984 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1985 if (!txn->mt_spill_pgs)
1988 /* purge deleted slots */
1989 MDB_IDL sl = txn->mt_spill_pgs;
1990 unsigned int num = sl[0];
1992 for (i=1; i<=num; i++) {
1999 /* Preserve pages which may soon be dirtied again */
2000 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2003 /* Less aggressive spill - we originally spilled the entire dirty list,
2004 * with a few exceptions for cursor pages and DB root pages. But this
2005 * turns out to be a lot of wasted effort because in a large txn many
2006 * of those pages will need to be used again. So now we spill only 1/8th
2007 * of the dirty pages. Testing revealed this to be a good tradeoff,
2008 * better than 1/2, 1/4, or 1/10.
2010 if (need < MDB_IDL_UM_MAX / 8)
2011 need = MDB_IDL_UM_MAX / 8;
2013 /* Save the page IDs of all the pages we're flushing */
2014 /* flush from the tail forward, this saves a lot of shifting later on. */
2015 for (i=dl[0].mid; i && need; i--) {
2016 MDB_ID pn = dl[i].mid << 1;
2018 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2020 /* Can't spill twice, make sure it's not already in a parent's
2023 if (txn->mt_parent) {
2025 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2026 if (tx2->mt_spill_pgs) {
2027 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2028 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2029 dp->mp_flags |= P_KEEP;
2037 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2041 mdb_midl_sort(txn->mt_spill_pgs);
2043 /* Flush the spilled part of dirty list */
2044 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2047 /* Reset any dirty pages we kept that page_flush didn't see */
2048 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2051 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2055 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2057 mdb_find_oldest(MDB_txn *txn)
2060 txnid_t mr, oldest = txn->mt_txnid - 1;
2061 if (txn->mt_env->me_txns) {
2062 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2063 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2074 /** Add a page to the txn's dirty list */
2076 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2079 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2081 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2082 insert = mdb_mid2l_append;
2084 insert = mdb_mid2l_insert;
2086 mid.mid = mp->mp_pgno;
2088 rc = insert(txn->mt_u.dirty_list, &mid);
2089 mdb_tassert(txn, rc == 0);
2090 txn->mt_dirty_room--;
2093 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2094 * me_pghead and mt_next_pgno.
2096 * If there are free pages available from older transactions, they
2097 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2098 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2099 * and move me_pglast to say which records were consumed. Only this
2100 * function can create me_pghead and move me_pglast/mt_next_pgno.
2101 * @param[in] mc cursor A cursor handle identifying the transaction and
2102 * database for which we are allocating.
2103 * @param[in] num the number of pages to allocate.
2104 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2105 * will always be satisfied by a single contiguous chunk of memory.
2106 * @return 0 on success, non-zero on failure.
2109 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2111 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2112 /* Get at most <Max_retries> more freeDB records once me_pghead
2113 * has enough pages. If not enough, use new pages from the map.
2114 * If <Paranoid> and mc is updating the freeDB, only get new
2115 * records if me_pghead is empty. Then the freelist cannot play
2116 * catch-up with itself by growing while trying to save it.
2118 enum { Paranoid = 1, Max_retries = 500 };
2120 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2122 int rc, retry = num * 60;
2123 MDB_txn *txn = mc->mc_txn;
2124 MDB_env *env = txn->mt_env;
2125 pgno_t pgno, *mop = env->me_pghead;
2126 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2128 txnid_t oldest = 0, last;
2133 /* If there are any loose pages, just use them */
2134 if (num == 1 && txn->mt_loose_pgs) {
2135 np = txn->mt_loose_pgs;
2136 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2137 txn->mt_loose_count--;
2138 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2146 /* If our dirty list is already full, we can't do anything */
2147 if (txn->mt_dirty_room == 0) {
2152 for (op = MDB_FIRST;; op = MDB_NEXT) {
2157 /* Seek a big enough contiguous page range. Prefer
2158 * pages at the tail, just truncating the list.
2164 if (mop[i-n2] == pgno+n2)
2171 if (op == MDB_FIRST) { /* 1st iteration */
2172 /* Prepare to fetch more and coalesce */
2173 last = env->me_pglast;
2174 oldest = env->me_pgoldest;
2175 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2178 key.mv_data = &last; /* will look up last+1 */
2179 key.mv_size = sizeof(last);
2181 if (Paranoid && mc->mc_dbi == FREE_DBI)
2184 if (Paranoid && retry < 0 && mop_len)
2188 /* Do not fetch more if the record will be too recent */
2189 if (oldest <= last) {
2191 oldest = mdb_find_oldest(txn);
2192 env->me_pgoldest = oldest;
2198 rc = mdb_cursor_get(&m2, &key, NULL, op);
2200 if (rc == MDB_NOTFOUND)
2204 last = *(txnid_t*)key.mv_data;
2205 if (oldest <= last) {
2207 oldest = mdb_find_oldest(txn);
2208 env->me_pgoldest = oldest;
2214 np = m2.mc_pg[m2.mc_top];
2215 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2216 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2219 idl = (MDB_ID *) data.mv_data;
2222 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2227 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2229 mop = env->me_pghead;
2231 env->me_pglast = last;
2233 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2234 last, txn->mt_dbs[FREE_DBI].md_root, i));
2236 DPRINTF(("IDL %"Z"u", idl[j]));
2238 /* Merge in descending sorted order */
2239 mdb_midl_xmerge(mop, idl);
2243 /* Use new pages from the map when nothing suitable in the freeDB */
2245 pgno = txn->mt_next_pgno;
2246 if (pgno + num >= env->me_maxpg) {
2247 DPUTS("DB size maxed out");
2253 if (env->me_flags & MDB_WRITEMAP) {
2254 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2256 if (!(np = mdb_page_malloc(txn, num))) {
2262 mop[0] = mop_len -= num;
2263 /* Move any stragglers down */
2264 for (j = i-num; j < mop_len; )
2265 mop[++j] = mop[++i];
2267 txn->mt_next_pgno = pgno + num;
2270 mdb_page_dirty(txn, np);
2276 txn->mt_flags |= MDB_TXN_ERROR;
2280 /** Copy the used portions of a non-overflow page.
2281 * @param[in] dst page to copy into
2282 * @param[in] src page to copy from
2283 * @param[in] psize size of a page
2286 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2288 enum { Align = sizeof(pgno_t) };
2289 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2291 /* If page isn't full, just copy the used portion. Adjust
2292 * alignment so memcpy may copy words instead of bytes.
2294 if ((unused &= -Align) && !IS_LEAF2(src)) {
2295 upper = (upper + PAGEBASE) & -Align;
2296 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2297 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2300 memcpy(dst, src, psize - unused);
2304 /** Pull a page off the txn's spill list, if present.
2305 * If a page being referenced was spilled to disk in this txn, bring
2306 * it back and make it dirty/writable again.
2307 * @param[in] txn the transaction handle.
2308 * @param[in] mp the page being referenced. It must not be dirty.
2309 * @param[out] ret the writable page, if any. ret is unchanged if
2310 * mp wasn't spilled.
2313 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2315 MDB_env *env = txn->mt_env;
2318 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2320 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2321 if (!tx2->mt_spill_pgs)
2323 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2324 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2327 if (txn->mt_dirty_room == 0)
2328 return MDB_TXN_FULL;
2329 if (IS_OVERFLOW(mp))
2333 if (env->me_flags & MDB_WRITEMAP) {
2336 np = mdb_page_malloc(txn, num);
2340 memcpy(np, mp, num * env->me_psize);
2342 mdb_page_copy(np, mp, env->me_psize);
2345 /* If in current txn, this page is no longer spilled.
2346 * If it happens to be the last page, truncate the spill list.
2347 * Otherwise mark it as deleted by setting the LSB.
2349 if (x == txn->mt_spill_pgs[0])
2350 txn->mt_spill_pgs[0]--;
2352 txn->mt_spill_pgs[x] |= 1;
2353 } /* otherwise, if belonging to a parent txn, the
2354 * page remains spilled until child commits
2357 mdb_page_dirty(txn, np);
2358 np->mp_flags |= P_DIRTY;
2366 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2367 * @param[in] mc cursor pointing to the page to be touched
2368 * @return 0 on success, non-zero on failure.
2371 mdb_page_touch(MDB_cursor *mc)
2373 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2374 MDB_txn *txn = mc->mc_txn;
2375 MDB_cursor *m2, *m3;
2379 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2380 if (txn->mt_flags & MDB_TXN_SPILLS) {
2382 rc = mdb_page_unspill(txn, mp, &np);
2388 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2389 (rc = mdb_page_alloc(mc, 1, &np)))
2392 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2393 mp->mp_pgno, pgno));
2394 mdb_cassert(mc, mp->mp_pgno != pgno);
2395 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2396 /* Update the parent page, if any, to point to the new page */
2398 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2399 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2400 SETPGNO(node, pgno);
2402 mc->mc_db->md_root = pgno;
2404 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2405 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2407 /* If txn has a parent, make sure the page is in our
2411 unsigned x = mdb_mid2l_search(dl, pgno);
2412 if (x <= dl[0].mid && dl[x].mid == pgno) {
2413 if (mp != dl[x].mptr) { /* bad cursor? */
2414 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2415 txn->mt_flags |= MDB_TXN_ERROR;
2416 return MDB_CORRUPTED;
2421 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2423 np = mdb_page_malloc(txn, 1);
2428 rc = mdb_mid2l_insert(dl, &mid);
2429 mdb_cassert(mc, rc == 0);
2434 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2436 np->mp_flags |= P_DIRTY;
2439 /* Adjust cursors pointing to mp */
2440 mc->mc_pg[mc->mc_top] = np;
2441 m2 = txn->mt_cursors[mc->mc_dbi];
2442 if (mc->mc_flags & C_SUB) {
2443 for (; m2; m2=m2->mc_next) {
2444 m3 = &m2->mc_xcursor->mx_cursor;
2445 if (m3->mc_snum < mc->mc_snum) continue;
2446 if (m3->mc_pg[mc->mc_top] == mp)
2447 m3->mc_pg[mc->mc_top] = np;
2450 for (; m2; m2=m2->mc_next) {
2451 if (m2->mc_snum < mc->mc_snum) continue;
2452 if (m2->mc_pg[mc->mc_top] == mp) {
2453 m2->mc_pg[mc->mc_top] = np;
2454 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2456 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2458 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2459 if (!(leaf->mn_flags & F_SUBDATA))
2460 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2468 txn->mt_flags |= MDB_TXN_ERROR;
2473 mdb_env_sync(MDB_env *env, int force)
2476 if (env->me_flags & MDB_RDONLY)
2478 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2479 if (env->me_flags & MDB_WRITEMAP) {
2480 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2481 ? MS_ASYNC : MS_SYNC;
2482 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2485 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2489 #ifdef BROKEN_FDATASYNC
2490 if (env->me_flags & MDB_FSYNCONLY) {
2491 if (fsync(env->me_fd))
2495 if (MDB_FDATASYNC(env->me_fd))
2502 /** Back up parent txn's cursors, then grab the originals for tracking */
2504 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2506 MDB_cursor *mc, *bk;
2511 for (i = src->mt_numdbs; --i >= 0; ) {
2512 if ((mc = src->mt_cursors[i]) != NULL) {
2513 size = sizeof(MDB_cursor);
2515 size += sizeof(MDB_xcursor);
2516 for (; mc; mc = bk->mc_next) {
2522 mc->mc_db = &dst->mt_dbs[i];
2523 /* Kill pointers into src - and dst to reduce abuse: The
2524 * user may not use mc until dst ends. Otherwise we'd...
2526 mc->mc_txn = NULL; /* ...set this to dst */
2527 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2528 if ((mx = mc->mc_xcursor) != NULL) {
2529 *(MDB_xcursor *)(bk+1) = *mx;
2530 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2532 mc->mc_next = dst->mt_cursors[i];
2533 dst->mt_cursors[i] = mc;
2540 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2541 * @param[in] txn the transaction handle.
2542 * @param[in] merge true to keep changes to parent cursors, false to revert.
2543 * @return 0 on success, non-zero on failure.
2546 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2548 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2552 for (i = txn->mt_numdbs; --i >= 0; ) {
2553 for (mc = cursors[i]; mc; mc = next) {
2555 if ((bk = mc->mc_backup) != NULL) {
2557 /* Commit changes to parent txn */
2558 mc->mc_next = bk->mc_next;
2559 mc->mc_backup = bk->mc_backup;
2560 mc->mc_txn = bk->mc_txn;
2561 mc->mc_db = bk->mc_db;
2562 mc->mc_dbflag = bk->mc_dbflag;
2563 if ((mx = mc->mc_xcursor) != NULL)
2564 mx->mx_cursor.mc_txn = bk->mc_txn;
2566 /* Abort nested txn */
2568 if ((mx = mc->mc_xcursor) != NULL)
2569 *mx = *(MDB_xcursor *)(bk+1);
2573 /* Only malloced cursors are permanently tracked. */
2580 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2586 Pidset = F_SETLK, Pidcheck = F_GETLK
2590 /** Set or check a pid lock. Set returns 0 on success.
2591 * Check returns 0 if the process is certainly dead, nonzero if it may
2592 * be alive (the lock exists or an error happened so we do not know).
2594 * On Windows Pidset is a no-op, we merely check for the existence
2595 * of the process with the given pid. On POSIX we use a single byte
2596 * lock on the lockfile, set at an offset equal to the pid.
2599 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2601 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2604 if (op == Pidcheck) {
2605 h = OpenProcess(env->me_pidquery, FALSE, pid);
2606 /* No documented "no such process" code, but other program use this: */
2608 return ErrCode() != ERROR_INVALID_PARAMETER;
2609 /* A process exists until all handles to it close. Has it exited? */
2610 ret = WaitForSingleObject(h, 0) != 0;
2617 struct flock lock_info;
2618 memset(&lock_info, 0, sizeof(lock_info));
2619 lock_info.l_type = F_WRLCK;
2620 lock_info.l_whence = SEEK_SET;
2621 lock_info.l_start = pid;
2622 lock_info.l_len = 1;
2623 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2624 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2626 } else if ((rc = ErrCode()) == EINTR) {
2634 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2635 * @param[in] txn the transaction handle to initialize
2636 * @return 0 on success, non-zero on failure.
2639 mdb_txn_renew0(MDB_txn *txn)
2641 MDB_env *env = txn->mt_env;
2642 MDB_txninfo *ti = env->me_txns;
2644 unsigned int i, nr, flags = txn->mt_flags;
2646 int rc, new_notls = 0;
2648 if ((flags &= MDB_TXN_RDONLY) != 0) {
2650 meta = mdb_env_pick_meta(env);
2651 txn->mt_txnid = meta->mm_txnid;
2652 txn->mt_u.reader = NULL;
2654 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2655 pthread_getspecific(env->me_txkey);
2657 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2658 return MDB_BAD_RSLOT;
2660 MDB_PID_T pid = env->me_pid;
2661 MDB_THR_T tid = pthread_self();
2662 mdb_mutexref_t rmutex = env->me_rmutex;
2664 if (!env->me_live_reader) {
2665 rc = mdb_reader_pid(env, Pidset, pid);
2668 env->me_live_reader = 1;
2671 if (LOCK_MUTEX(rc, env, rmutex))
2673 nr = ti->mti_numreaders;
2674 for (i=0; i<nr; i++)
2675 if (ti->mti_readers[i].mr_pid == 0)
2677 if (i == env->me_maxreaders) {
2678 UNLOCK_MUTEX(rmutex);
2679 return MDB_READERS_FULL;
2681 r = &ti->mti_readers[i];
2682 /* Claim the reader slot, carefully since other code
2683 * uses the reader table un-mutexed: First reset the
2684 * slot, next publish it in mti_numreaders. After
2685 * that, it is safe for mdb_env_close() to touch it.
2686 * When it will be closed, we can finally claim it.
2689 r->mr_txnid = (txnid_t)-1;
2692 ti->mti_numreaders = ++nr;
2693 env->me_close_readers = nr;
2695 UNLOCK_MUTEX(rmutex);
2697 new_notls = (env->me_flags & MDB_NOTLS);
2698 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2703 do /* LY: Retry on a race, ITS#7970. */
2704 r->mr_txnid = ti->mti_txnid;
2705 while(r->mr_txnid != ti->mti_txnid);
2706 txn->mt_txnid = r->mr_txnid;
2707 txn->mt_u.reader = r;
2708 meta = env->me_metas[txn->mt_txnid & 1];
2712 /* Not yet touching txn == env->me_txn0, it may be active */
2714 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2716 txn->mt_txnid = ti->mti_txnid;
2717 meta = env->me_metas[txn->mt_txnid & 1];
2719 meta = mdb_env_pick_meta(env);
2720 txn->mt_txnid = meta->mm_txnid;
2724 if (txn->mt_txnid == mdb_debug_start)
2727 txn->mt_child = NULL;
2728 txn->mt_loose_pgs = NULL;
2729 txn->mt_loose_count = 0;
2730 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2731 txn->mt_u.dirty_list = env->me_dirty_list;
2732 txn->mt_u.dirty_list[0].mid = 0;
2733 txn->mt_free_pgs = env->me_free_pgs;
2734 txn->mt_free_pgs[0] = 0;
2735 txn->mt_spill_pgs = NULL;
2737 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2740 /* Copy the DB info and flags */
2741 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2743 /* Moved to here to avoid a data race in read TXNs */
2744 txn->mt_next_pgno = meta->mm_last_pg+1;
2746 txn->mt_flags = flags;
2749 txn->mt_numdbs = env->me_numdbs;
2750 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2751 x = env->me_dbflags[i];
2752 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2753 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2755 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2756 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2758 if (env->me_flags & MDB_FATAL_ERROR) {
2759 DPUTS("environment had fatal error, must shutdown!");
2761 } else if (env->me_maxpg < txn->mt_next_pgno) {
2762 rc = MDB_MAP_RESIZED;
2766 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2771 mdb_txn_renew(MDB_txn *txn)
2775 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2778 rc = mdb_txn_renew0(txn);
2779 if (rc == MDB_SUCCESS) {
2780 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2781 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2782 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2788 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2792 int rc, size, tsize;
2794 flags &= MDB_TXN_BEGIN_FLAGS;
2795 flags |= env->me_flags & MDB_WRITEMAP;
2797 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2801 /* Nested transactions: Max 1 child, write txns only, no writemap */
2802 flags |= parent->mt_flags;
2803 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2804 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2806 /* Child txns save MDB_pgstate and use own copy of cursors */
2807 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2808 size += tsize = sizeof(MDB_ntxn);
2809 } else if (flags & MDB_RDONLY) {
2810 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2811 size += tsize = sizeof(MDB_txn);
2813 /* Reuse preallocated write txn. However, do not touch it until
2814 * mdb_txn_renew0() succeeds, since it currently may be active.
2819 if ((txn = calloc(1, size)) == NULL) {
2820 DPRINTF(("calloc: %s", strerror(errno)));
2823 txn->mt_dbxs = env->me_dbxs; /* static */
2824 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2825 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2826 txn->mt_flags = flags;
2831 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2832 txn->mt_dbiseqs = parent->mt_dbiseqs;
2833 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2834 if (!txn->mt_u.dirty_list ||
2835 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2837 free(txn->mt_u.dirty_list);
2841 txn->mt_txnid = parent->mt_txnid;
2842 txn->mt_dirty_room = parent->mt_dirty_room;
2843 txn->mt_u.dirty_list[0].mid = 0;
2844 txn->mt_spill_pgs = NULL;
2845 txn->mt_next_pgno = parent->mt_next_pgno;
2846 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2847 parent->mt_child = txn;
2848 txn->mt_parent = parent;
2849 txn->mt_numdbs = parent->mt_numdbs;
2850 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2851 /* Copy parent's mt_dbflags, but clear DB_NEW */
2852 for (i=0; i<txn->mt_numdbs; i++)
2853 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2855 ntxn = (MDB_ntxn *)txn;
2856 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2857 if (env->me_pghead) {
2858 size = MDB_IDL_SIZEOF(env->me_pghead);
2859 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2861 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2866 rc = mdb_cursor_shadow(parent, txn);
2868 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2869 } else { /* MDB_RDONLY */
2870 txn->mt_dbiseqs = env->me_dbiseqs;
2872 rc = mdb_txn_renew0(txn);
2875 if (txn != env->me_txn0)
2878 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2880 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2881 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2882 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2889 mdb_txn_env(MDB_txn *txn)
2891 if(!txn) return NULL;
2896 mdb_txn_id(MDB_txn *txn)
2899 return txn->mt_txnid;
2902 /** Export or close DBI handles opened in this txn. */
2904 mdb_dbis_update(MDB_txn *txn, int keep)
2907 MDB_dbi n = txn->mt_numdbs;
2908 MDB_env *env = txn->mt_env;
2909 unsigned char *tdbflags = txn->mt_dbflags;
2911 for (i = n; --i >= CORE_DBS;) {
2912 if (tdbflags[i] & DB_NEW) {
2914 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2916 char *ptr = env->me_dbxs[i].md_name.mv_data;
2918 env->me_dbxs[i].md_name.mv_data = NULL;
2919 env->me_dbxs[i].md_name.mv_size = 0;
2920 env->me_dbflags[i] = 0;
2921 env->me_dbiseqs[i]++;
2927 if (keep && env->me_numdbs < n)
2931 /** End a transaction, except successful commit of a nested transaction.
2932 * May be called twice for readonly txns: First reset it, then abort.
2933 * @param[in] txn the transaction handle to end
2934 * @param[in] mode why and how to end the transaction
2937 mdb_txn_end(MDB_txn *txn, unsigned mode)
2939 MDB_env *env = txn->mt_env;
2941 static const char *const names[] = MDB_END_NAMES;
2944 /* Export or close DBI handles opened in this txn */
2945 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2947 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2948 names[mode & MDB_END_OPMASK],
2949 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2950 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2952 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2953 if (txn->mt_u.reader) {
2954 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2955 if (!(env->me_flags & MDB_NOTLS)) {
2956 txn->mt_u.reader = NULL; /* txn does not own reader */
2957 } else if (mode & MDB_END_SLOT) {
2958 txn->mt_u.reader->mr_pid = 0;
2959 txn->mt_u.reader = NULL;
2960 } /* else txn owns the slot until it does MDB_END_SLOT */
2962 txn->mt_numdbs = 0; /* prevent further DBI activity */
2963 txn->mt_flags |= MDB_TXN_FINISHED;
2965 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2966 pgno_t *pghead = env->me_pghead;
2968 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2969 mdb_cursors_close(txn, 0);
2970 if (!(env->me_flags & MDB_WRITEMAP)) {
2971 mdb_dlist_free(txn);
2975 txn->mt_flags = MDB_TXN_FINISHED;
2977 if (!txn->mt_parent) {
2978 mdb_midl_shrink(&txn->mt_free_pgs);
2979 env->me_free_pgs = txn->mt_free_pgs;
2981 env->me_pghead = NULL;
2985 mode = 0; /* txn == env->me_txn0, do not free() it */
2987 /* The writer mutex was locked in mdb_txn_begin. */
2989 UNLOCK_MUTEX(env->me_wmutex);
2991 txn->mt_parent->mt_child = NULL;
2992 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2993 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2994 mdb_midl_free(txn->mt_free_pgs);
2995 mdb_midl_free(txn->mt_spill_pgs);
2996 free(txn->mt_u.dirty_list);
2999 mdb_midl_free(pghead);
3002 if (mode & MDB_END_FREE)
3007 mdb_txn_reset(MDB_txn *txn)
3012 /* This call is only valid for read-only txns */
3013 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3016 mdb_txn_end(txn, MDB_END_RESET);
3020 mdb_txn_abort(MDB_txn *txn)
3026 mdb_txn_abort(txn->mt_child);
3028 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3031 /** Save the freelist as of this transaction to the freeDB.
3032 * This changes the freelist. Keep trying until it stabilizes.
3035 mdb_freelist_save(MDB_txn *txn)
3037 /* env->me_pghead[] can grow and shrink during this call.
3038 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3039 * Page numbers cannot disappear from txn->mt_free_pgs[].
3042 MDB_env *env = txn->mt_env;
3043 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3044 txnid_t pglast = 0, head_id = 0;
3045 pgno_t freecnt = 0, *free_pgs, *mop;
3046 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3048 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3050 if (env->me_pghead) {
3051 /* Make sure first page of freeDB is touched and on freelist */
3052 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3053 if (rc && rc != MDB_NOTFOUND)
3057 if (!env->me_pghead && txn->mt_loose_pgs) {
3058 /* Put loose page numbers in mt_free_pgs, since
3059 * we may be unable to return them to me_pghead.
3061 MDB_page *mp = txn->mt_loose_pgs;
3062 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3064 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3065 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3066 txn->mt_loose_pgs = NULL;
3067 txn->mt_loose_count = 0;
3070 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3071 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3072 ? SSIZE_MAX : maxfree_1pg;
3075 /* Come back here after each Put() in case freelist changed */
3080 /* If using records from freeDB which we have not yet
3081 * deleted, delete them and any we reserved for me_pghead.
3083 while (pglast < env->me_pglast) {
3084 rc = mdb_cursor_first(&mc, &key, NULL);
3087 pglast = head_id = *(txnid_t *)key.mv_data;
3088 total_room = head_room = 0;
3089 mdb_tassert(txn, pglast <= env->me_pglast);
3090 rc = mdb_cursor_del(&mc, 0);
3095 /* Save the IDL of pages freed by this txn, to a single record */
3096 if (freecnt < txn->mt_free_pgs[0]) {
3098 /* Make sure last page of freeDB is touched and on freelist */
3099 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3100 if (rc && rc != MDB_NOTFOUND)
3103 free_pgs = txn->mt_free_pgs;
3104 /* Write to last page of freeDB */
3105 key.mv_size = sizeof(txn->mt_txnid);
3106 key.mv_data = &txn->mt_txnid;
3108 freecnt = free_pgs[0];
3109 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3110 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3113 /* Retry if mt_free_pgs[] grew during the Put() */
3114 free_pgs = txn->mt_free_pgs;
3115 } while (freecnt < free_pgs[0]);
3116 mdb_midl_sort(free_pgs);
3117 memcpy(data.mv_data, free_pgs, data.mv_size);
3120 unsigned int i = free_pgs[0];
3121 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3122 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3124 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3130 mop = env->me_pghead;
3131 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3133 /* Reserve records for me_pghead[]. Split it if multi-page,
3134 * to avoid searching freeDB for a page range. Use keys in
3135 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3137 if (total_room >= mop_len) {
3138 if (total_room == mop_len || --more < 0)
3140 } else if (head_room >= maxfree_1pg && head_id > 1) {
3141 /* Keep current record (overflow page), add a new one */
3145 /* (Re)write {key = head_id, IDL length = head_room} */
3146 total_room -= head_room;
3147 head_room = mop_len - total_room;
3148 if (head_room > maxfree_1pg && head_id > 1) {
3149 /* Overflow multi-page for part of me_pghead */
3150 head_room /= head_id; /* amortize page sizes */
3151 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3152 } else if (head_room < 0) {
3153 /* Rare case, not bothering to delete this record */
3156 key.mv_size = sizeof(head_id);
3157 key.mv_data = &head_id;
3158 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3159 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3162 /* IDL is initially empty, zero out at least the length */
3163 pgs = (pgno_t *)data.mv_data;
3164 j = head_room > clean_limit ? head_room : 0;
3168 total_room += head_room;
3171 /* Return loose page numbers to me_pghead, though usually none are
3172 * left at this point. The pages themselves remain in dirty_list.
3174 if (txn->mt_loose_pgs) {
3175 MDB_page *mp = txn->mt_loose_pgs;
3176 unsigned count = txn->mt_loose_count;
3178 /* Room for loose pages + temp IDL with same */
3179 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3181 mop = env->me_pghead;
3182 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3183 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3184 loose[ ++count ] = mp->mp_pgno;
3186 mdb_midl_sort(loose);
3187 mdb_midl_xmerge(mop, loose);
3188 txn->mt_loose_pgs = NULL;
3189 txn->mt_loose_count = 0;
3193 /* Fill in the reserved me_pghead records */
3199 rc = mdb_cursor_first(&mc, &key, &data);
3200 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3201 txnid_t id = *(txnid_t *)key.mv_data;
3202 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3205 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3207 if (len > mop_len) {
3209 data.mv_size = (len + 1) * sizeof(MDB_ID);
3211 data.mv_data = mop -= len;
3214 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3216 if (rc || !(mop_len -= len))
3223 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3224 * @param[in] txn the transaction that's being committed
3225 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3226 * @return 0 on success, non-zero on failure.
3229 mdb_page_flush(MDB_txn *txn, int keep)
3231 MDB_env *env = txn->mt_env;
3232 MDB_ID2L dl = txn->mt_u.dirty_list;
3233 unsigned psize = env->me_psize, j;
3234 int i, pagecount = dl[0].mid, rc;
3235 size_t size = 0, pos = 0;
3237 MDB_page *dp = NULL;
3241 struct iovec iov[MDB_COMMIT_PAGES];
3242 ssize_t wpos = 0, wsize = 0, wres;
3243 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3249 if (env->me_flags & MDB_WRITEMAP) {
3250 /* Clear dirty flags */
3251 while (++i <= pagecount) {
3253 /* Don't flush this page yet */
3254 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3255 dp->mp_flags &= ~P_KEEP;
3259 dp->mp_flags &= ~P_DIRTY;
3264 /* Write the pages */
3266 if (++i <= pagecount) {
3268 /* Don't flush this page yet */
3269 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3270 dp->mp_flags &= ~P_KEEP;
3275 /* clear dirty flag */
3276 dp->mp_flags &= ~P_DIRTY;
3279 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3284 /* Windows actually supports scatter/gather I/O, but only on
3285 * unbuffered file handles. Since we're relying on the OS page
3286 * cache for all our data, that's self-defeating. So we just
3287 * write pages one at a time. We use the ov structure to set
3288 * the write offset, to at least save the overhead of a Seek
3291 DPRINTF(("committing page %"Z"u", pgno));
3292 memset(&ov, 0, sizeof(ov));
3293 ov.Offset = pos & 0xffffffff;
3294 ov.OffsetHigh = pos >> 16 >> 16;
3295 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3297 DPRINTF(("WriteFile: %d", rc));
3301 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3302 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3305 /* Write previous page(s) */
3306 #ifdef MDB_USE_PWRITEV
3307 wres = pwritev(env->me_fd, iov, n, wpos);
3310 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3313 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3317 DPRINTF(("lseek: %s", strerror(rc)));
3320 wres = writev(env->me_fd, iov, n);
3323 if (wres != wsize) {
3328 DPRINTF(("Write error: %s", strerror(rc)));
3330 rc = EIO; /* TODO: Use which error code? */
3331 DPUTS("short write, filesystem full?");
3342 DPRINTF(("committing page %"Z"u", pgno));
3343 next_pos = pos + size;
3344 iov[n].iov_len = size;
3345 iov[n].iov_base = (char *)dp;
3351 /* MIPS has cache coherency issues, this is a no-op everywhere else
3352 * Note: for any size >= on-chip cache size, entire on-chip cache is
3355 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3357 for (i = keep; ++i <= pagecount; ) {
3359 /* This is a page we skipped above */
3362 dl[j].mid = dp->mp_pgno;
3365 mdb_dpage_free(env, dp);
3370 txn->mt_dirty_room += i - j;
3376 mdb_txn_commit(MDB_txn *txn)
3379 unsigned int i, end_mode;
3385 /* mdb_txn_end() mode for a commit which writes nothing */
3386 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3388 if (txn->mt_child) {
3389 rc = mdb_txn_commit(txn->mt_child);
3396 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3400 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3401 DPUTS("txn has failed/finished, can't commit");
3403 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3408 if (txn->mt_parent) {
3409 MDB_txn *parent = txn->mt_parent;
3413 unsigned x, y, len, ps_len;
3415 /* Append our free list to parent's */
3416 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3419 mdb_midl_free(txn->mt_free_pgs);
3420 /* Failures after this must either undo the changes
3421 * to the parent or set MDB_TXN_ERROR in the parent.
3424 parent->mt_next_pgno = txn->mt_next_pgno;
3425 parent->mt_flags = txn->mt_flags;
3427 /* Merge our cursors into parent's and close them */
3428 mdb_cursors_close(txn, 1);
3430 /* Update parent's DB table. */
3431 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3432 parent->mt_numdbs = txn->mt_numdbs;
3433 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3434 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3435 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3436 /* preserve parent's DB_NEW status */
3437 x = parent->mt_dbflags[i] & DB_NEW;
3438 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3441 dst = parent->mt_u.dirty_list;
3442 src = txn->mt_u.dirty_list;
3443 /* Remove anything in our dirty list from parent's spill list */
3444 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3446 pspill[0] = (pgno_t)-1;
3447 /* Mark our dirty pages as deleted in parent spill list */
3448 for (i=0, len=src[0].mid; ++i <= len; ) {
3449 MDB_ID pn = src[i].mid << 1;
3450 while (pn > pspill[x])
3452 if (pn == pspill[x]) {
3457 /* Squash deleted pagenums if we deleted any */
3458 for (x=y; ++x <= ps_len; )
3459 if (!(pspill[x] & 1))
3460 pspill[++y] = pspill[x];
3464 /* Find len = length of merging our dirty list with parent's */
3466 dst[0].mid = 0; /* simplify loops */
3467 if (parent->mt_parent) {
3468 len = x + src[0].mid;
3469 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3470 for (i = x; y && i; y--) {
3471 pgno_t yp = src[y].mid;
3472 while (yp < dst[i].mid)
3474 if (yp == dst[i].mid) {
3479 } else { /* Simplify the above for single-ancestor case */
3480 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3482 /* Merge our dirty list with parent's */
3484 for (i = len; y; dst[i--] = src[y--]) {
3485 pgno_t yp = src[y].mid;
3486 while (yp < dst[x].mid)
3487 dst[i--] = dst[x--];
3488 if (yp == dst[x].mid)
3489 free(dst[x--].mptr);
3491 mdb_tassert(txn, i == x);
3493 free(txn->mt_u.dirty_list);
3494 parent->mt_dirty_room = txn->mt_dirty_room;
3495 if (txn->mt_spill_pgs) {
3496 if (parent->mt_spill_pgs) {
3497 /* TODO: Prevent failure here, so parent does not fail */
3498 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3500 parent->mt_flags |= MDB_TXN_ERROR;
3501 mdb_midl_free(txn->mt_spill_pgs);
3502 mdb_midl_sort(parent->mt_spill_pgs);
3504 parent->mt_spill_pgs = txn->mt_spill_pgs;
3508 /* Append our loose page list to parent's */
3509 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3511 *lp = txn->mt_loose_pgs;
3512 parent->mt_loose_count += txn->mt_loose_count;
3514 parent->mt_child = NULL;
3515 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3520 if (txn != env->me_txn) {
3521 DPUTS("attempt to commit unknown transaction");
3526 mdb_cursors_close(txn, 0);
3528 if (!txn->mt_u.dirty_list[0].mid &&
3529 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3532 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3533 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3535 /* Update DB root pointers */
3536 if (txn->mt_numdbs > CORE_DBS) {
3540 data.mv_size = sizeof(MDB_db);
3542 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3543 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3544 if (txn->mt_dbflags[i] & DB_DIRTY) {
3545 if (TXN_DBI_CHANGED(txn, i)) {
3549 data.mv_data = &txn->mt_dbs[i];
3550 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3558 rc = mdb_freelist_save(txn);
3562 mdb_midl_free(env->me_pghead);
3563 env->me_pghead = NULL;
3564 mdb_midl_shrink(&txn->mt_free_pgs);
3570 if ((rc = mdb_page_flush(txn, 0)))
3572 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3573 (rc = mdb_env_sync(env, 0)))
3575 if ((rc = mdb_env_write_meta(txn)))
3577 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3580 mdb_txn_end(txn, end_mode);
3588 /** Read the environment parameters of a DB environment before
3589 * mapping it into memory.
3590 * @param[in] env the environment handle
3591 * @param[out] meta address of where to store the meta information
3592 * @return 0 on success, non-zero on failure.
3595 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3601 enum { Size = sizeof(pbuf) };
3603 /* We don't know the page size yet, so use a minimum value.
3604 * Read both meta pages so we can use the latest one.
3607 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3611 memset(&ov, 0, sizeof(ov));
3613 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3614 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3617 rc = pread(env->me_fd, &pbuf, Size, off);
3620 if (rc == 0 && off == 0)
3622 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3623 DPRINTF(("read: %s", mdb_strerror(rc)));
3627 p = (MDB_page *)&pbuf;
3629 if (!F_ISSET(p->mp_flags, P_META)) {
3630 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3635 if (m->mm_magic != MDB_MAGIC) {
3636 DPUTS("meta has invalid magic");
3640 if (m->mm_version != MDB_DATA_VERSION) {
3641 DPRINTF(("database is version %u, expected version %u",
3642 m->mm_version, MDB_DATA_VERSION));
3643 return MDB_VERSION_MISMATCH;
3646 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3652 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3654 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3656 meta->mm_magic = MDB_MAGIC;
3657 meta->mm_version = MDB_DATA_VERSION;
3658 meta->mm_mapsize = env->me_mapsize;
3659 meta->mm_psize = env->me_psize;
3660 meta->mm_last_pg = NUM_METAS-1;
3661 meta->mm_flags = env->me_flags & 0xffff;
3662 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3663 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3664 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3667 /** Write the environment parameters of a freshly created DB environment.
3668 * @param[in] env the environment handle
3669 * @param[in] meta the #MDB_meta to write
3670 * @return 0 on success, non-zero on failure.
3673 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3681 memset(&ov, 0, sizeof(ov));
3682 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3684 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3687 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3688 len = pwrite(fd, ptr, size, pos); \
3689 if (len == -1 && ErrCode() == EINTR) continue; \
3690 rc = (len >= 0); break; } while(1)
3693 DPUTS("writing new meta page");
3695 psize = env->me_psize;
3697 p = calloc(NUM_METAS, psize);
3699 p->mp_flags = P_META;
3700 *(MDB_meta *)METADATA(p) = *meta;
3702 q = (MDB_page *)((char *)p + psize);
3704 q->mp_flags = P_META;
3705 *(MDB_meta *)METADATA(q) = *meta;
3707 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3710 else if ((unsigned) len == psize * NUM_METAS)
3718 /** Update the environment info to commit a transaction.
3719 * @param[in] txn the transaction that's being committed
3720 * @return 0 on success, non-zero on failure.
3723 mdb_env_write_meta(MDB_txn *txn)
3726 MDB_meta meta, metab, *mp;
3730 int rc, len, toggle;
3739 toggle = txn->mt_txnid & 1;
3740 DPRINTF(("writing meta page %d for root page %"Z"u",
3741 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3744 flags = txn->mt_flags & env->me_flags;
3745 mp = env->me_metas[toggle];
3746 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3747 /* Persist any increases of mapsize config */
3748 if (mapsize < env->me_mapsize)
3749 mapsize = env->me_mapsize;
3751 if (flags & MDB_WRITEMAP) {
3752 mp->mm_mapsize = mapsize;
3753 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3754 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3755 mp->mm_last_pg = txn->mt_next_pgno - 1;
3756 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3757 !(defined(__i386__) || defined(__x86_64__))
3758 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3759 __sync_synchronize();
3761 mp->mm_txnid = txn->mt_txnid;
3762 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3763 unsigned meta_size = env->me_psize;
3764 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3765 ptr = (char *)mp - PAGEHDRSZ;
3766 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3767 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3771 if (MDB_MSYNC(ptr, meta_size, rc)) {
3778 metab.mm_txnid = mp->mm_txnid;
3779 metab.mm_last_pg = mp->mm_last_pg;
3781 meta.mm_mapsize = mapsize;
3782 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3783 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3784 meta.mm_last_pg = txn->mt_next_pgno - 1;
3785 meta.mm_txnid = txn->mt_txnid;
3787 off = offsetof(MDB_meta, mm_mapsize);
3788 ptr = (char *)&meta + off;
3789 len = sizeof(MDB_meta) - off;
3790 off += (char *)mp - env->me_map;
3792 /* Write to the SYNC fd */
3793 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3796 memset(&ov, 0, sizeof(ov));
3798 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3803 rc = pwrite(mfd, ptr, len, off);
3806 rc = rc < 0 ? ErrCode() : EIO;
3811 DPUTS("write failed, disk error?");
3812 /* On a failure, the pagecache still contains the new data.
3813 * Write some old data back, to prevent it from being used.
3814 * Use the non-SYNC fd; we know it will fail anyway.
3816 meta.mm_last_pg = metab.mm_last_pg;
3817 meta.mm_txnid = metab.mm_txnid;
3819 memset(&ov, 0, sizeof(ov));
3821 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3823 r2 = pwrite(env->me_fd, ptr, len, off);
3824 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3827 env->me_flags |= MDB_FATAL_ERROR;
3830 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3831 CACHEFLUSH(env->me_map + off, len, DCACHE);
3833 /* Memory ordering issues are irrelevant; since the entire writer
3834 * is wrapped by wmutex, all of these changes will become visible
3835 * after the wmutex is unlocked. Since the DB is multi-version,
3836 * readers will get consistent data regardless of how fresh or
3837 * how stale their view of these values is.
3840 env->me_txns->mti_txnid = txn->mt_txnid;
3845 /** Check both meta pages to see which one is newer.
3846 * @param[in] env the environment handle
3847 * @return newest #MDB_meta.
3850 mdb_env_pick_meta(const MDB_env *env)
3852 MDB_meta *const *metas = env->me_metas;
3853 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3857 mdb_env_create(MDB_env **env)
3861 e = calloc(1, sizeof(MDB_env));
3865 e->me_maxreaders = DEFAULT_READERS;
3866 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3867 e->me_fd = INVALID_HANDLE_VALUE;
3868 e->me_lfd = INVALID_HANDLE_VALUE;
3869 e->me_mfd = INVALID_HANDLE_VALUE;
3870 #ifdef MDB_USE_POSIX_SEM
3871 e->me_rmutex = SEM_FAILED;
3872 e->me_wmutex = SEM_FAILED;
3873 #elif defined MDB_USE_SYSV_SEM
3874 e->me_rmutex->semid = -1;
3875 e->me_wmutex->semid = -1;
3877 e->me_pid = getpid();
3878 GET_PAGESIZE(e->me_os_psize);
3879 VGMEMP_CREATE(e,0,0);
3885 mdb_env_map(MDB_env *env, void *addr)
3888 unsigned int flags = env->me_flags;
3892 LONG sizelo, sizehi;
3895 if (flags & MDB_RDONLY) {
3896 /* Don't set explicit map size, use whatever exists */
3901 msize = env->me_mapsize;
3902 sizelo = msize & 0xffffffff;
3903 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3905 /* Windows won't create mappings for zero length files.
3906 * and won't map more than the file size.
3907 * Just set the maxsize right now.
3909 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3910 || !SetEndOfFile(env->me_fd)
3911 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3915 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3916 PAGE_READWRITE : PAGE_READONLY,
3917 sizehi, sizelo, NULL);
3920 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3921 FILE_MAP_WRITE : FILE_MAP_READ,
3923 rc = env->me_map ? 0 : ErrCode();
3928 int prot = PROT_READ;
3929 if (flags & MDB_WRITEMAP) {
3931 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3934 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3936 if (env->me_map == MAP_FAILED) {
3941 if (flags & MDB_NORDAHEAD) {
3942 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3944 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3946 #ifdef POSIX_MADV_RANDOM
3947 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3948 #endif /* POSIX_MADV_RANDOM */
3949 #endif /* MADV_RANDOM */
3953 /* Can happen because the address argument to mmap() is just a
3954 * hint. mmap() can pick another, e.g. if the range is in use.
3955 * The MAP_FIXED flag would prevent that, but then mmap could
3956 * instead unmap existing pages to make room for the new map.
3958 if (addr && env->me_map != addr)
3959 return EBUSY; /* TODO: Make a new MDB_* error code? */
3961 p = (MDB_page *)env->me_map;
3962 env->me_metas[0] = METADATA(p);
3963 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3969 mdb_env_set_mapsize(MDB_env *env, size_t size)
3971 /* If env is already open, caller is responsible for making
3972 * sure there are no active txns.
3980 meta = mdb_env_pick_meta(env);
3982 size = meta->mm_mapsize;
3984 /* Silently round up to minimum if the size is too small */
3985 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3989 munmap(env->me_map, env->me_mapsize);
3990 env->me_mapsize = size;
3991 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3992 rc = mdb_env_map(env, old);
3996 env->me_mapsize = size;
3998 env->me_maxpg = env->me_mapsize / env->me_psize;
4003 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4007 env->me_maxdbs = dbs + CORE_DBS;
4012 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4014 if (env->me_map || readers < 1)
4016 env->me_maxreaders = readers;
4021 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4023 if (!env || !readers)
4025 *readers = env->me_maxreaders;
4030 mdb_fsize(HANDLE fd, size_t *size)
4033 LARGE_INTEGER fsize;
4035 if (!GetFileSizeEx(fd, &fsize))
4038 *size = fsize.QuadPart;
4050 #ifdef BROKEN_FDATASYNC
4051 #include <sys/utsname.h>
4052 #include <sys/vfs.h>
4055 /** Further setup required for opening an LMDB environment
4058 mdb_env_open2(MDB_env *env)
4060 unsigned int flags = env->me_flags;
4061 int i, newenv = 0, rc;
4065 /* See if we should use QueryLimited */
4067 if ((rc & 0xff) > 5)
4068 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4070 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4073 #ifdef BROKEN_FDATASYNC
4074 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4075 * https://lkml.org/lkml/2012/9/3/83
4076 * Kernels after 3.6-rc6 are known good.
4077 * https://lkml.org/lkml/2012/9/10/556
4078 * See if the DB is on ext3/ext4, then check for new enough kernel
4079 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4084 fstatfs(env->me_fd, &st);
4085 while (st.f_type == 0xEF53) {
4089 if (uts.release[0] < '3') {
4090 if (!strncmp(uts.release, "2.6.32.", 7)) {
4091 i = atoi(uts.release+7);
4093 break; /* 2.6.32.60 and newer is OK */
4094 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4095 i = atoi(uts.release+7);
4097 break; /* 2.6.34.15 and newer is OK */
4099 } else if (uts.release[0] == '3') {
4100 i = atoi(uts.release+2);
4102 break; /* 3.6 and newer is OK */
4104 i = atoi(uts.release+4);
4106 break; /* 3.5.4 and newer is OK */
4107 } else if (i == 2) {
4108 i = atoi(uts.release+4);
4110 break; /* 3.2.30 and newer is OK */
4112 } else { /* 4.x and newer is OK */
4115 env->me_flags |= MDB_FSYNCONLY;
4121 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4124 DPUTS("new mdbenv");
4126 env->me_psize = env->me_os_psize;
4127 if (env->me_psize > MAX_PAGESIZE)
4128 env->me_psize = MAX_PAGESIZE;
4129 memset(&meta, 0, sizeof(meta));
4130 mdb_env_init_meta0(env, &meta);
4131 meta.mm_mapsize = DEFAULT_MAPSIZE;
4133 env->me_psize = meta.mm_psize;
4136 /* Was a mapsize configured? */
4137 if (!env->me_mapsize) {
4138 env->me_mapsize = meta.mm_mapsize;
4141 /* Make sure mapsize >= committed data size. Even when using
4142 * mm_mapsize, which could be broken in old files (ITS#7789).
4144 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4145 if (env->me_mapsize < minsize)
4146 env->me_mapsize = minsize;
4148 meta.mm_mapsize = env->me_mapsize;
4150 if (newenv && !(flags & MDB_FIXEDMAP)) {
4151 /* mdb_env_map() may grow the datafile. Write the metapages
4152 * first, so the file will be valid if initialization fails.
4153 * Except with FIXEDMAP, since we do not yet know mm_address.
4154 * We could fill in mm_address later, but then a different
4155 * program might end up doing that - one with a memory layout
4156 * and map address which does not suit the main program.
4158 rc = mdb_env_init_meta(env, &meta);
4164 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4169 if (flags & MDB_FIXEDMAP)
4170 meta.mm_address = env->me_map;
4171 i = mdb_env_init_meta(env, &meta);
4172 if (i != MDB_SUCCESS) {
4177 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4178 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4180 #if !(MDB_MAXKEYSIZE)
4181 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4183 env->me_maxpg = env->me_mapsize / env->me_psize;
4187 MDB_meta *meta = mdb_env_pick_meta(env);
4188 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4190 DPRINTF(("opened database version %u, pagesize %u",
4191 meta->mm_version, env->me_psize));
4192 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4193 DPRINTF(("depth: %u", db->md_depth));
4194 DPRINTF(("entries: %"Z"u", db->md_entries));
4195 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4196 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4197 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4198 DPRINTF(("root: %"Z"u", db->md_root));
4206 /** Release a reader thread's slot in the reader lock table.
4207 * This function is called automatically when a thread exits.
4208 * @param[in] ptr This points to the slot in the reader lock table.
4211 mdb_env_reader_dest(void *ptr)
4213 MDB_reader *reader = ptr;
4219 /** Junk for arranging thread-specific callbacks on Windows. This is
4220 * necessarily platform and compiler-specific. Windows supports up
4221 * to 1088 keys. Let's assume nobody opens more than 64 environments
4222 * in a single process, for now. They can override this if needed.
4224 #ifndef MAX_TLS_KEYS
4225 #define MAX_TLS_KEYS 64
4227 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4228 static int mdb_tls_nkeys;
4230 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4234 case DLL_PROCESS_ATTACH: break;
4235 case DLL_THREAD_ATTACH: break;
4236 case DLL_THREAD_DETACH:
4237 for (i=0; i<mdb_tls_nkeys; i++) {
4238 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4240 mdb_env_reader_dest(r);
4244 case DLL_PROCESS_DETACH: break;
4249 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4251 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4255 /* Force some symbol references.
4256 * _tls_used forces the linker to create the TLS directory if not already done
4257 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4259 #pragma comment(linker, "/INCLUDE:_tls_used")
4260 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4261 #pragma const_seg(".CRT$XLB")
4262 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4263 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4266 #pragma comment(linker, "/INCLUDE:__tls_used")
4267 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4268 #pragma data_seg(".CRT$XLB")
4269 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4271 #endif /* WIN 32/64 */
4272 #endif /* !__GNUC__ */
4275 /** Downgrade the exclusive lock on the region back to shared */
4277 mdb_env_share_locks(MDB_env *env, int *excl)
4280 MDB_meta *meta = mdb_env_pick_meta(env);
4282 env->me_txns->mti_txnid = meta->mm_txnid;
4287 /* First acquire a shared lock. The Unlock will
4288 * then release the existing exclusive lock.
4290 memset(&ov, 0, sizeof(ov));
4291 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4294 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4300 struct flock lock_info;
4301 /* The shared lock replaces the existing lock */
4302 memset((void *)&lock_info, 0, sizeof(lock_info));
4303 lock_info.l_type = F_RDLCK;
4304 lock_info.l_whence = SEEK_SET;
4305 lock_info.l_start = 0;
4306 lock_info.l_len = 1;
4307 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4308 (rc = ErrCode()) == EINTR) ;
4309 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4316 /** Try to get exclusive lock, otherwise shared.
4317 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4320 mdb_env_excl_lock(MDB_env *env, int *excl)
4324 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4328 memset(&ov, 0, sizeof(ov));
4329 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4336 struct flock lock_info;
4337 memset((void *)&lock_info, 0, sizeof(lock_info));
4338 lock_info.l_type = F_WRLCK;
4339 lock_info.l_whence = SEEK_SET;
4340 lock_info.l_start = 0;
4341 lock_info.l_len = 1;
4342 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4343 (rc = ErrCode()) == EINTR) ;
4347 # ifndef MDB_USE_POSIX_MUTEX
4348 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4351 lock_info.l_type = F_RDLCK;
4352 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4353 (rc = ErrCode()) == EINTR) ;
4363 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4365 * @(#) $Revision: 5.1 $
4366 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4367 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4369 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4373 * Please do not copyright this code. This code is in the public domain.
4375 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4376 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4377 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4378 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4379 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4380 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4381 * PERFORMANCE OF THIS SOFTWARE.
4384 * chongo <Landon Curt Noll> /\oo/\
4385 * http://www.isthe.com/chongo/
4387 * Share and Enjoy! :-)
4390 typedef unsigned long long mdb_hash_t;
4391 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4393 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4394 * @param[in] val value to hash
4395 * @param[in] hval initial value for hash
4396 * @return 64 bit hash
4398 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4399 * hval arg on the first call.
4402 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4404 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4405 unsigned char *end = s + val->mv_size;
4407 * FNV-1a hash each octet of the string
4410 /* xor the bottom with the current octet */
4411 hval ^= (mdb_hash_t)*s++;
4413 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4414 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4415 (hval << 7) + (hval << 8) + (hval << 40);
4417 /* return our new hash value */
4421 /** Hash the string and output the encoded hash.
4422 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4423 * very short name limits. We don't care about the encoding being reversible,
4424 * we just want to preserve as many bits of the input as possible in a
4425 * small printable string.
4426 * @param[in] str string to hash
4427 * @param[out] encbuf an array of 11 chars to hold the hash
4429 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4432 mdb_pack85(unsigned long l, char *out)
4436 for (i=0; i<5; i++) {
4437 *out++ = mdb_a85[l % 85];
4443 mdb_hash_enc(MDB_val *val, char *encbuf)
4445 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4447 mdb_pack85(h, encbuf);
4448 mdb_pack85(h>>32, encbuf+5);
4453 /** Open and/or initialize the lock region for the environment.
4454 * @param[in] env The LMDB environment.
4455 * @param[in] lpath The pathname of the file used for the lock region.
4456 * @param[in] mode The Unix permissions for the file, if we create it.
4457 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4458 * @return 0 on success, non-zero on failure.
4461 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4464 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4466 # define MDB_ERRCODE_ROFS EROFS
4467 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4468 # define MDB_CLOEXEC O_CLOEXEC
4471 # define MDB_CLOEXEC 0
4474 #ifdef MDB_USE_SYSV_SEM
4482 env->me_lfd = CreateFileA(lpath, GENERIC_READ|GENERIC_WRITE,
4483 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4484 FILE_ATTRIBUTE_NORMAL, NULL);
4486 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4488 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4490 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4495 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4496 /* Lose record locks when exec*() */
4497 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4498 fcntl(env->me_lfd, F_SETFD, fdflags);
4501 if (!(env->me_flags & MDB_NOTLS)) {
4502 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4505 env->me_flags |= MDB_ENV_TXKEY;
4507 /* Windows TLS callbacks need help finding their TLS info. */
4508 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4512 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4516 /* Try to get exclusive lock. If we succeed, then
4517 * nobody is using the lock region and we should initialize it.
4519 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4522 size = GetFileSize(env->me_lfd, NULL);
4524 size = lseek(env->me_lfd, 0, SEEK_END);
4525 if (size == -1) goto fail_errno;
4527 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4528 if (size < rsize && *excl > 0) {
4530 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4531 || !SetEndOfFile(env->me_lfd))
4534 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4538 size = rsize - sizeof(MDB_txninfo);
4539 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4544 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4546 if (!mh) goto fail_errno;
4547 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4549 if (!env->me_txns) goto fail_errno;
4551 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4553 if (m == MAP_FAILED) goto fail_errno;
4559 BY_HANDLE_FILE_INFORMATION stbuf;
4568 if (!mdb_sec_inited) {
4569 InitializeSecurityDescriptor(&mdb_null_sd,
4570 SECURITY_DESCRIPTOR_REVISION);
4571 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4572 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4573 mdb_all_sa.bInheritHandle = FALSE;
4574 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4577 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4578 idbuf.volume = stbuf.dwVolumeSerialNumber;
4579 idbuf.nhigh = stbuf.nFileIndexHigh;
4580 idbuf.nlow = stbuf.nFileIndexLow;
4581 val.mv_data = &idbuf;
4582 val.mv_size = sizeof(idbuf);
4583 mdb_hash_enc(&val, encbuf);
4584 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4585 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4586 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4587 if (!env->me_rmutex) goto fail_errno;
4588 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4589 if (!env->me_wmutex) goto fail_errno;
4590 #elif defined(MDB_USE_POSIX_SEM)
4599 #if defined(__NetBSD__)
4600 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4602 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4603 idbuf.dev = stbuf.st_dev;
4604 idbuf.ino = stbuf.st_ino;
4605 val.mv_data = &idbuf;
4606 val.mv_size = sizeof(idbuf);
4607 mdb_hash_enc(&val, encbuf);
4608 #ifdef MDB_SHORT_SEMNAMES
4609 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4611 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4612 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4613 /* Clean up after a previous run, if needed: Try to
4614 * remove both semaphores before doing anything else.
4616 sem_unlink(env->me_txns->mti_rmname);
4617 sem_unlink(env->me_txns->mti_wmname);
4618 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4619 O_CREAT|O_EXCL, mode, 1);
4620 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4621 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4622 O_CREAT|O_EXCL, mode, 1);
4623 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4624 #elif defined(MDB_USE_SYSV_SEM)
4625 unsigned short vals[2] = {1, 1};
4626 key_t key = ftok(lpath, 'M');
4629 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4633 if (semctl(semid, 0, SETALL, semu) < 0)
4635 env->me_txns->mti_semid = semid;
4636 #else /* MDB_USE_POSIX_MUTEX: */
4637 pthread_mutexattr_t mattr;
4639 if ((rc = pthread_mutexattr_init(&mattr))
4640 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4641 #ifdef MDB_ROBUST_SUPPORTED
4642 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4644 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4645 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4647 pthread_mutexattr_destroy(&mattr);
4648 #endif /* _WIN32 || ... */
4650 env->me_txns->mti_magic = MDB_MAGIC;
4651 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4652 env->me_txns->mti_txnid = 0;
4653 env->me_txns->mti_numreaders = 0;
4656 #ifdef MDB_USE_SYSV_SEM
4657 struct semid_ds buf;
4659 if (env->me_txns->mti_magic != MDB_MAGIC) {
4660 DPUTS("lock region has invalid magic");
4664 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4665 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4666 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4667 rc = MDB_VERSION_MISMATCH;
4671 if (rc && rc != EACCES && rc != EAGAIN) {
4675 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4676 if (!env->me_rmutex) goto fail_errno;
4677 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4678 if (!env->me_wmutex) goto fail_errno;
4679 #elif defined(MDB_USE_POSIX_SEM)
4680 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4681 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4682 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4683 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4684 #elif defined(MDB_USE_SYSV_SEM)
4685 semid = env->me_txns->mti_semid;
4687 /* check for read access */
4688 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4690 /* check for write access */
4691 if (semctl(semid, 0, IPC_SET, semu) < 0)
4695 #ifdef MDB_USE_SYSV_SEM
4696 env->me_rmutex->semid = semid;
4697 env->me_wmutex->semid = semid;
4698 env->me_rmutex->semnum = 0;
4699 env->me_wmutex->semnum = 1;
4700 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4701 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4712 /** The name of the lock file in the DB environment */
4713 #define LOCKNAME "/lock.mdb"
4714 /** The name of the data file in the DB environment */
4715 #define DATANAME "/data.mdb"
4716 /** The suffix of the lock file when no subdir is used */
4717 #define LOCKSUFF "-lock"
4718 /** Only a subset of the @ref mdb_env flags can be changed
4719 * at runtime. Changing other flags requires closing the
4720 * environment and re-opening it with the new flags.
4722 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4723 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4724 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4726 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4727 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4731 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4733 int oflags, rc, len, excl = -1;
4734 char *lpath, *dpath;
4736 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4740 if (flags & MDB_NOSUBDIR) {
4741 rc = len + sizeof(LOCKSUFF) + len + 1;
4743 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4748 if (flags & MDB_NOSUBDIR) {
4749 dpath = lpath + len + sizeof(LOCKSUFF);
4750 sprintf(lpath, "%s" LOCKSUFF, path);
4751 strcpy(dpath, path);
4753 dpath = lpath + len + sizeof(LOCKNAME);
4754 sprintf(lpath, "%s" LOCKNAME, path);
4755 sprintf(dpath, "%s" DATANAME, path);
4759 flags |= env->me_flags;
4760 if (flags & MDB_RDONLY) {
4761 /* silently ignore WRITEMAP when we're only getting read access */
4762 flags &= ~MDB_WRITEMAP;
4764 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4765 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4768 env->me_flags = flags |= MDB_ENV_ACTIVE;
4772 env->me_path = strdup(path);
4773 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4774 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4775 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4776 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4780 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4782 /* For RDONLY, get lockfile after we know datafile exists */
4783 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4784 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4790 if (F_ISSET(flags, MDB_RDONLY)) {
4791 oflags = GENERIC_READ;
4792 len = OPEN_EXISTING;
4794 oflags = GENERIC_READ|GENERIC_WRITE;
4797 mode = FILE_ATTRIBUTE_NORMAL;
4798 env->me_fd = CreateFileA(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4799 NULL, len, mode, NULL);
4801 if (F_ISSET(flags, MDB_RDONLY))
4804 oflags = O_RDWR | O_CREAT;
4806 env->me_fd = open(dpath, oflags, mode);
4808 if (env->me_fd == INVALID_HANDLE_VALUE) {
4813 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4814 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4819 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4820 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4821 env->me_mfd = env->me_fd;
4823 /* Synchronous fd for meta writes. Needed even with
4824 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4827 len = OPEN_EXISTING;
4828 env->me_mfd = CreateFileA(dpath, oflags,
4829 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4830 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4833 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4835 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4840 DPRINTF(("opened dbenv %p", (void *) env));
4842 rc = mdb_env_share_locks(env, &excl);
4846 if (!(flags & MDB_RDONLY)) {
4848 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4849 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4850 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4851 (txn = calloc(1, size)))
4853 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4854 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4855 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4856 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4858 txn->mt_dbxs = env->me_dbxs;
4859 txn->mt_flags = MDB_TXN_FINISHED;
4869 mdb_env_close0(env, excl);
4875 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4877 mdb_env_close0(MDB_env *env, int excl)
4881 if (!(env->me_flags & MDB_ENV_ACTIVE))
4884 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4886 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4887 free(env->me_dbxs[i].md_name.mv_data);
4892 free(env->me_dbiseqs);
4893 free(env->me_dbflags);
4895 free(env->me_dirty_list);
4897 mdb_midl_free(env->me_free_pgs);
4899 if (env->me_flags & MDB_ENV_TXKEY) {
4900 pthread_key_delete(env->me_txkey);
4902 /* Delete our key from the global list */
4903 for (i=0; i<mdb_tls_nkeys; i++)
4904 if (mdb_tls_keys[i] == env->me_txkey) {
4905 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4913 munmap(env->me_map, env->me_mapsize);
4915 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4916 (void) close(env->me_mfd);
4917 if (env->me_fd != INVALID_HANDLE_VALUE)
4918 (void) close(env->me_fd);
4920 MDB_PID_T pid = env->me_pid;
4921 /* Clearing readers is done in this function because
4922 * me_txkey with its destructor must be disabled first.
4924 * We skip the the reader mutex, so we touch only
4925 * data owned by this process (me_close_readers and
4926 * our readers), and clear each reader atomically.
4928 for (i = env->me_close_readers; --i >= 0; )
4929 if (env->me_txns->mti_readers[i].mr_pid == pid)
4930 env->me_txns->mti_readers[i].mr_pid = 0;
4932 if (env->me_rmutex) {
4933 CloseHandle(env->me_rmutex);
4934 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4936 /* Windows automatically destroys the mutexes when
4937 * the last handle closes.
4939 #elif defined(MDB_USE_POSIX_SEM)
4940 if (env->me_rmutex != SEM_FAILED) {
4941 sem_close(env->me_rmutex);
4942 if (env->me_wmutex != SEM_FAILED)
4943 sem_close(env->me_wmutex);
4944 /* If we have the filelock: If we are the
4945 * only remaining user, clean up semaphores.
4948 mdb_env_excl_lock(env, &excl);
4950 sem_unlink(env->me_txns->mti_rmname);
4951 sem_unlink(env->me_txns->mti_wmname);
4954 #elif defined(MDB_USE_SYSV_SEM)
4955 if (env->me_rmutex->semid != -1) {
4956 /* If we have the filelock: If we are the
4957 * only remaining user, clean up semaphores.
4960 mdb_env_excl_lock(env, &excl);
4962 semctl(env->me_rmutex->semid, 0, IPC_RMID);
4965 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4967 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4970 /* Unlock the lockfile. Windows would have unlocked it
4971 * after closing anyway, but not necessarily at once.
4973 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4976 (void) close(env->me_lfd);
4979 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4983 mdb_env_close(MDB_env *env)
4990 VGMEMP_DESTROY(env);
4991 while ((dp = env->me_dpages) != NULL) {
4992 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4993 env->me_dpages = dp->mp_next;
4997 mdb_env_close0(env, 0);
5001 /** Compare two items pointing at aligned size_t's */
5003 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5005 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5006 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5009 /** Compare two items pointing at aligned unsigned int's.
5011 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5012 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5015 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5017 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5018 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5021 /** Compare two items pointing at unsigned ints of unknown alignment.
5022 * Nodes and keys are guaranteed to be 2-byte aligned.
5025 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5027 #if BYTE_ORDER == LITTLE_ENDIAN
5028 unsigned short *u, *c;
5031 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5032 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5035 } while(!x && u > (unsigned short *)a->mv_data);
5038 unsigned short *u, *c, *end;
5041 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5042 u = (unsigned short *)a->mv_data;
5043 c = (unsigned short *)b->mv_data;
5046 } while(!x && u < end);
5051 /** Compare two items lexically */
5053 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5060 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5066 diff = memcmp(a->mv_data, b->mv_data, len);
5067 return diff ? diff : len_diff<0 ? -1 : len_diff;
5070 /** Compare two items in reverse byte order */
5072 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5074 const unsigned char *p1, *p2, *p1_lim;
5078 p1_lim = (const unsigned char *)a->mv_data;
5079 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5080 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5082 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5088 while (p1 > p1_lim) {
5089 diff = *--p1 - *--p2;
5093 return len_diff<0 ? -1 : len_diff;
5096 /** Search for key within a page, using binary search.
5097 * Returns the smallest entry larger or equal to the key.
5098 * If exactp is non-null, stores whether the found entry was an exact match
5099 * in *exactp (1 or 0).
5100 * Updates the cursor index with the index of the found entry.
5101 * If no entry larger or equal to the key is found, returns NULL.
5104 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5106 unsigned int i = 0, nkeys;
5109 MDB_page *mp = mc->mc_pg[mc->mc_top];
5110 MDB_node *node = NULL;
5115 nkeys = NUMKEYS(mp);
5117 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5118 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5121 low = IS_LEAF(mp) ? 0 : 1;
5123 cmp = mc->mc_dbx->md_cmp;
5125 /* Branch pages have no data, so if using integer keys,
5126 * alignment is guaranteed. Use faster mdb_cmp_int.
5128 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5129 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5136 nodekey.mv_size = mc->mc_db->md_pad;
5137 node = NODEPTR(mp, 0); /* fake */
5138 while (low <= high) {
5139 i = (low + high) >> 1;
5140 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5141 rc = cmp(key, &nodekey);
5142 DPRINTF(("found leaf index %u [%s], rc = %i",
5143 i, DKEY(&nodekey), rc));
5152 while (low <= high) {
5153 i = (low + high) >> 1;
5155 node = NODEPTR(mp, i);
5156 nodekey.mv_size = NODEKSZ(node);
5157 nodekey.mv_data = NODEKEY(node);
5159 rc = cmp(key, &nodekey);
5162 DPRINTF(("found leaf index %u [%s], rc = %i",
5163 i, DKEY(&nodekey), rc));
5165 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5166 i, DKEY(&nodekey), NODEPGNO(node), rc));
5177 if (rc > 0) { /* Found entry is less than the key. */
5178 i++; /* Skip to get the smallest entry larger than key. */
5180 node = NODEPTR(mp, i);
5183 *exactp = (rc == 0 && nkeys > 0);
5184 /* store the key index */
5185 mc->mc_ki[mc->mc_top] = i;
5187 /* There is no entry larger or equal to the key. */
5190 /* nodeptr is fake for LEAF2 */
5196 mdb_cursor_adjust(MDB_cursor *mc, func)
5200 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5201 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5208 /** Pop a page off the top of the cursor's stack. */
5210 mdb_cursor_pop(MDB_cursor *mc)
5213 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5214 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5222 /** Push a page onto the top of the cursor's stack. */
5224 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5226 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5227 DDBI(mc), (void *) mc));
5229 if (mc->mc_snum >= CURSOR_STACK) {
5230 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5231 return MDB_CURSOR_FULL;
5234 mc->mc_top = mc->mc_snum++;
5235 mc->mc_pg[mc->mc_top] = mp;
5236 mc->mc_ki[mc->mc_top] = 0;
5241 /** Find the address of the page corresponding to a given page number.
5242 * @param[in] txn the transaction for this access.
5243 * @param[in] pgno the page number for the page to retrieve.
5244 * @param[out] ret address of a pointer where the page's address will be stored.
5245 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5246 * @return 0 on success, non-zero on failure.
5249 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5251 MDB_env *env = txn->mt_env;
5255 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5259 MDB_ID2L dl = tx2->mt_u.dirty_list;
5261 /* Spilled pages were dirtied in this txn and flushed
5262 * because the dirty list got full. Bring this page
5263 * back in from the map (but don't unspill it here,
5264 * leave that unless page_touch happens again).
5266 if (tx2->mt_spill_pgs) {
5267 MDB_ID pn = pgno << 1;
5268 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5269 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5270 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5275 unsigned x = mdb_mid2l_search(dl, pgno);
5276 if (x <= dl[0].mid && dl[x].mid == pgno) {
5282 } while ((tx2 = tx2->mt_parent) != NULL);
5285 if (pgno < txn->mt_next_pgno) {
5287 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5289 DPRINTF(("page %"Z"u not found", pgno));
5290 txn->mt_flags |= MDB_TXN_ERROR;
5291 return MDB_PAGE_NOTFOUND;
5301 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5302 * The cursor is at the root page, set up the rest of it.
5305 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5307 MDB_page *mp = mc->mc_pg[mc->mc_top];
5311 while (IS_BRANCH(mp)) {
5315 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5316 mdb_cassert(mc, NUMKEYS(mp) > 1);
5317 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5319 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5321 if (flags & MDB_PS_LAST)
5322 i = NUMKEYS(mp) - 1;
5325 node = mdb_node_search(mc, key, &exact);
5327 i = NUMKEYS(mp) - 1;
5329 i = mc->mc_ki[mc->mc_top];
5331 mdb_cassert(mc, i > 0);
5335 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5338 mdb_cassert(mc, i < NUMKEYS(mp));
5339 node = NODEPTR(mp, i);
5341 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5344 mc->mc_ki[mc->mc_top] = i;
5345 if ((rc = mdb_cursor_push(mc, mp)))
5348 if (flags & MDB_PS_MODIFY) {
5349 if ((rc = mdb_page_touch(mc)) != 0)
5351 mp = mc->mc_pg[mc->mc_top];
5356 DPRINTF(("internal error, index points to a %02X page!?",
5358 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5359 return MDB_CORRUPTED;
5362 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5363 key ? DKEY(key) : "null"));
5364 mc->mc_flags |= C_INITIALIZED;
5365 mc->mc_flags &= ~C_EOF;
5370 /** Search for the lowest key under the current branch page.
5371 * This just bypasses a NUMKEYS check in the current page
5372 * before calling mdb_page_search_root(), because the callers
5373 * are all in situations where the current page is known to
5377 mdb_page_search_lowest(MDB_cursor *mc)
5379 MDB_page *mp = mc->mc_pg[mc->mc_top];
5380 MDB_node *node = NODEPTR(mp, 0);
5383 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5386 mc->mc_ki[mc->mc_top] = 0;
5387 if ((rc = mdb_cursor_push(mc, mp)))
5389 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5392 /** Search for the page a given key should be in.
5393 * Push it and its parent pages on the cursor stack.
5394 * @param[in,out] mc the cursor for this operation.
5395 * @param[in] key the key to search for, or NULL for first/last page.
5396 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5397 * are touched (updated with new page numbers).
5398 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5399 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5400 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5401 * @return 0 on success, non-zero on failure.
5404 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5409 /* Make sure the txn is still viable, then find the root from
5410 * the txn's db table and set it as the root of the cursor's stack.
5412 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5413 DPUTS("transaction may not be used now");
5416 /* Make sure we're using an up-to-date root */
5417 if (*mc->mc_dbflag & DB_STALE) {
5419 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5421 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5422 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5429 MDB_node *leaf = mdb_node_search(&mc2,
5430 &mc->mc_dbx->md_name, &exact);
5432 return MDB_NOTFOUND;
5433 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5434 return MDB_INCOMPATIBLE; /* not a named DB */
5435 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5438 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5440 /* The txn may not know this DBI, or another process may
5441 * have dropped and recreated the DB with other flags.
5443 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5444 return MDB_INCOMPATIBLE;
5445 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5447 *mc->mc_dbflag &= ~DB_STALE;
5449 root = mc->mc_db->md_root;
5451 if (root == P_INVALID) { /* Tree is empty. */
5452 DPUTS("tree is empty");
5453 return MDB_NOTFOUND;
5457 mdb_cassert(mc, root > 1);
5458 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5459 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5465 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5466 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5468 if (flags & MDB_PS_MODIFY) {
5469 if ((rc = mdb_page_touch(mc)))
5473 if (flags & MDB_PS_ROOTONLY)
5476 return mdb_page_search_root(mc, key, flags);
5480 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5482 MDB_txn *txn = mc->mc_txn;
5483 pgno_t pg = mp->mp_pgno;
5484 unsigned x = 0, ovpages = mp->mp_pages;
5485 MDB_env *env = txn->mt_env;
5486 MDB_IDL sl = txn->mt_spill_pgs;
5487 MDB_ID pn = pg << 1;
5490 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5491 /* If the page is dirty or on the spill list we just acquired it,
5492 * so we should give it back to our current free list, if any.
5493 * Otherwise put it onto the list of pages we freed in this txn.
5495 * Won't create me_pghead: me_pglast must be inited along with it.
5496 * Unsupported in nested txns: They would need to hide the page
5497 * range in ancestor txns' dirty and spilled lists.
5499 if (env->me_pghead &&
5501 ((mp->mp_flags & P_DIRTY) ||
5502 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5506 MDB_ID2 *dl, ix, iy;
5507 rc = mdb_midl_need(&env->me_pghead, ovpages);
5510 if (!(mp->mp_flags & P_DIRTY)) {
5511 /* This page is no longer spilled */
5518 /* Remove from dirty list */
5519 dl = txn->mt_u.dirty_list;
5521 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5527 mdb_cassert(mc, x > 1);
5529 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5530 txn->mt_flags |= MDB_TXN_ERROR;
5531 return MDB_CORRUPTED;
5534 if (!(env->me_flags & MDB_WRITEMAP))
5535 mdb_dpage_free(env, mp);
5537 /* Insert in me_pghead */
5538 mop = env->me_pghead;
5539 j = mop[0] + ovpages;
5540 for (i = mop[0]; i && mop[i] < pg; i--)
5546 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5550 mc->mc_db->md_overflow_pages -= ovpages;
5554 /** Return the data associated with a given node.
5555 * @param[in] txn The transaction for this operation.
5556 * @param[in] leaf The node being read.
5557 * @param[out] data Updated to point to the node's data.
5558 * @return 0 on success, non-zero on failure.
5561 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5563 MDB_page *omp; /* overflow page */
5567 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5568 data->mv_size = NODEDSZ(leaf);
5569 data->mv_data = NODEDATA(leaf);
5573 /* Read overflow data.
5575 data->mv_size = NODEDSZ(leaf);
5576 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5577 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5578 DPRINTF(("read overflow page %"Z"u failed", pgno));
5581 data->mv_data = METADATA(omp);
5587 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5588 MDB_val *key, MDB_val *data)
5595 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5597 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5600 if (txn->mt_flags & MDB_TXN_BLOCKED)
5603 mdb_cursor_init(&mc, txn, dbi, &mx);
5604 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5607 /** Find a sibling for a page.
5608 * Replaces the page at the top of the cursor's stack with the
5609 * specified sibling, if one exists.
5610 * @param[in] mc The cursor for this operation.
5611 * @param[in] move_right Non-zero if the right sibling is requested,
5612 * otherwise the left sibling.
5613 * @return 0 on success, non-zero on failure.
5616 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5622 if (mc->mc_snum < 2) {
5623 return MDB_NOTFOUND; /* root has no siblings */
5627 DPRINTF(("parent page is page %"Z"u, index %u",
5628 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5630 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5631 : (mc->mc_ki[mc->mc_top] == 0)) {
5632 DPRINTF(("no more keys left, moving to %s sibling",
5633 move_right ? "right" : "left"));
5634 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5635 /* undo cursor_pop before returning */
5642 mc->mc_ki[mc->mc_top]++;
5644 mc->mc_ki[mc->mc_top]--;
5645 DPRINTF(("just moving to %s index key %u",
5646 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5648 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5650 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5651 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5652 /* mc will be inconsistent if caller does mc_snum++ as above */
5653 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5657 mdb_cursor_push(mc, mp);
5659 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5664 /** Move the cursor to the next data item. */
5666 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5672 if (mc->mc_flags & C_EOF) {
5673 return MDB_NOTFOUND;
5676 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5678 mp = mc->mc_pg[mc->mc_top];
5680 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5681 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5682 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5683 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5684 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5685 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5686 if (rc == MDB_SUCCESS)
5687 MDB_GET_KEY(leaf, key);
5692 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5693 if (op == MDB_NEXT_DUP)
5694 return MDB_NOTFOUND;
5698 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5699 mdb_dbg_pgno(mp), (void *) mc));
5700 if (mc->mc_flags & C_DEL)
5703 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5704 DPUTS("=====> move to next sibling page");
5705 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5706 mc->mc_flags |= C_EOF;
5709 mp = mc->mc_pg[mc->mc_top];
5710 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5712 mc->mc_ki[mc->mc_top]++;
5715 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5716 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5719 key->mv_size = mc->mc_db->md_pad;
5720 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5724 mdb_cassert(mc, IS_LEAF(mp));
5725 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5727 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5728 mdb_xcursor_init1(mc, leaf);
5731 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5734 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5735 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5736 if (rc != MDB_SUCCESS)
5741 MDB_GET_KEY(leaf, key);
5745 /** Move the cursor to the previous data item. */
5747 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5753 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5755 mp = mc->mc_pg[mc->mc_top];
5757 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5758 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5759 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5760 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5761 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5762 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5763 if (rc == MDB_SUCCESS) {
5764 MDB_GET_KEY(leaf, key);
5765 mc->mc_flags &= ~C_EOF;
5771 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5772 if (op == MDB_PREV_DUP)
5773 return MDB_NOTFOUND;
5777 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5778 mdb_dbg_pgno(mp), (void *) mc));
5780 if (mc->mc_ki[mc->mc_top] == 0) {
5781 DPUTS("=====> move to prev sibling page");
5782 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5785 mp = mc->mc_pg[mc->mc_top];
5786 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5787 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5789 mc->mc_ki[mc->mc_top]--;
5791 mc->mc_flags &= ~C_EOF;
5793 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5794 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5797 key->mv_size = mc->mc_db->md_pad;
5798 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5802 mdb_cassert(mc, IS_LEAF(mp));
5803 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5805 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5806 mdb_xcursor_init1(mc, leaf);
5809 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5812 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5813 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5814 if (rc != MDB_SUCCESS)
5819 MDB_GET_KEY(leaf, key);
5823 /** Set the cursor on a specific data item. */
5825 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5826 MDB_cursor_op op, int *exactp)
5830 MDB_node *leaf = NULL;
5833 if (key->mv_size == 0)
5834 return MDB_BAD_VALSIZE;
5837 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5839 /* See if we're already on the right page */
5840 if (mc->mc_flags & C_INITIALIZED) {
5843 mp = mc->mc_pg[mc->mc_top];
5845 mc->mc_ki[mc->mc_top] = 0;
5846 return MDB_NOTFOUND;
5848 if (mp->mp_flags & P_LEAF2) {
5849 nodekey.mv_size = mc->mc_db->md_pad;
5850 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5852 leaf = NODEPTR(mp, 0);
5853 MDB_GET_KEY2(leaf, nodekey);
5855 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5857 /* Probably happens rarely, but first node on the page
5858 * was the one we wanted.
5860 mc->mc_ki[mc->mc_top] = 0;
5867 unsigned int nkeys = NUMKEYS(mp);
5869 if (mp->mp_flags & P_LEAF2) {
5870 nodekey.mv_data = LEAF2KEY(mp,
5871 nkeys-1, nodekey.mv_size);
5873 leaf = NODEPTR(mp, nkeys-1);
5874 MDB_GET_KEY2(leaf, nodekey);
5876 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5878 /* last node was the one we wanted */
5879 mc->mc_ki[mc->mc_top] = nkeys-1;
5885 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5886 /* This is definitely the right page, skip search_page */
5887 if (mp->mp_flags & P_LEAF2) {
5888 nodekey.mv_data = LEAF2KEY(mp,
5889 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5891 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5892 MDB_GET_KEY2(leaf, nodekey);
5894 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5896 /* current node was the one we wanted */
5906 /* If any parents have right-sibs, search.
5907 * Otherwise, there's nothing further.
5909 for (i=0; i<mc->mc_top; i++)
5911 NUMKEYS(mc->mc_pg[i])-1)
5913 if (i == mc->mc_top) {
5914 /* There are no other pages */
5915 mc->mc_ki[mc->mc_top] = nkeys;
5916 return MDB_NOTFOUND;
5920 /* There are no other pages */
5921 mc->mc_ki[mc->mc_top] = 0;
5922 if (op == MDB_SET_RANGE && !exactp) {
5926 return MDB_NOTFOUND;
5930 rc = mdb_page_search(mc, key, 0);
5931 if (rc != MDB_SUCCESS)
5934 mp = mc->mc_pg[mc->mc_top];
5935 mdb_cassert(mc, IS_LEAF(mp));
5938 leaf = mdb_node_search(mc, key, exactp);
5939 if (exactp != NULL && !*exactp) {
5940 /* MDB_SET specified and not an exact match. */
5941 return MDB_NOTFOUND;
5945 DPUTS("===> inexact leaf not found, goto sibling");
5946 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5947 mc->mc_flags |= C_EOF;
5948 return rc; /* no entries matched */
5950 mp = mc->mc_pg[mc->mc_top];
5951 mdb_cassert(mc, IS_LEAF(mp));
5952 leaf = NODEPTR(mp, 0);
5956 mc->mc_flags |= C_INITIALIZED;
5957 mc->mc_flags &= ~C_EOF;
5960 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5961 key->mv_size = mc->mc_db->md_pad;
5962 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5967 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5968 mdb_xcursor_init1(mc, leaf);
5971 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5972 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5973 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5976 if (op == MDB_GET_BOTH) {
5982 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5983 if (rc != MDB_SUCCESS)
5986 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5989 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5991 dcmp = mc->mc_dbx->md_dcmp;
5992 #if UINT_MAX < SIZE_MAX
5993 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5994 dcmp = mdb_cmp_clong;
5996 rc = dcmp(data, &olddata);
5998 if (op == MDB_GET_BOTH || rc > 0)
5999 return MDB_NOTFOUND;
6006 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6007 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6012 /* The key already matches in all other cases */
6013 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6014 MDB_GET_KEY(leaf, key);
6015 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6020 /** Move the cursor to the first item in the database. */
6022 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6028 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6030 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6031 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6032 if (rc != MDB_SUCCESS)
6035 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6037 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6038 mc->mc_flags |= C_INITIALIZED;
6039 mc->mc_flags &= ~C_EOF;
6041 mc->mc_ki[mc->mc_top] = 0;
6043 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6044 key->mv_size = mc->mc_db->md_pad;
6045 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6050 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6051 mdb_xcursor_init1(mc, leaf);
6052 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6056 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6060 MDB_GET_KEY(leaf, key);
6064 /** Move the cursor to the last item in the database. */
6066 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6072 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6074 if (!(mc->mc_flags & C_EOF)) {
6076 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6077 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6078 if (rc != MDB_SUCCESS)
6081 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6084 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6085 mc->mc_flags |= C_INITIALIZED|C_EOF;
6086 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6088 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6089 key->mv_size = mc->mc_db->md_pad;
6090 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6095 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6096 mdb_xcursor_init1(mc, leaf);
6097 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6101 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6106 MDB_GET_KEY(leaf, key);
6111 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6116 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6121 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6125 case MDB_GET_CURRENT:
6126 if (!(mc->mc_flags & C_INITIALIZED)) {
6129 MDB_page *mp = mc->mc_pg[mc->mc_top];
6130 int nkeys = NUMKEYS(mp);
6131 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6132 mc->mc_ki[mc->mc_top] = nkeys;
6138 key->mv_size = mc->mc_db->md_pad;
6139 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6141 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6142 MDB_GET_KEY(leaf, key);
6144 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6145 if (mc->mc_flags & C_DEL)
6146 mdb_xcursor_init1(mc, leaf);
6147 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6149 rc = mdb_node_read(mc->mc_txn, leaf, data);
6156 case MDB_GET_BOTH_RANGE:
6161 if (mc->mc_xcursor == NULL) {
6162 rc = MDB_INCOMPATIBLE;
6172 rc = mdb_cursor_set(mc, key, data, op,
6173 op == MDB_SET_RANGE ? NULL : &exact);
6176 case MDB_GET_MULTIPLE:
6177 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6181 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6182 rc = MDB_INCOMPATIBLE;
6186 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6187 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6190 case MDB_NEXT_MULTIPLE:
6195 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6196 rc = MDB_INCOMPATIBLE;
6199 if (!(mc->mc_flags & C_INITIALIZED))
6200 rc = mdb_cursor_first(mc, key, data);
6202 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6203 if (rc == MDB_SUCCESS) {
6204 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6207 mx = &mc->mc_xcursor->mx_cursor;
6208 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6210 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6211 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6219 case MDB_NEXT_NODUP:
6220 if (!(mc->mc_flags & C_INITIALIZED))
6221 rc = mdb_cursor_first(mc, key, data);
6223 rc = mdb_cursor_next(mc, key, data, op);
6227 case MDB_PREV_NODUP:
6228 if (!(mc->mc_flags & C_INITIALIZED)) {
6229 rc = mdb_cursor_last(mc, key, data);
6232 mc->mc_flags |= C_INITIALIZED;
6233 mc->mc_ki[mc->mc_top]++;
6235 rc = mdb_cursor_prev(mc, key, data, op);
6238 rc = mdb_cursor_first(mc, key, data);
6241 mfunc = mdb_cursor_first;
6243 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6247 if (mc->mc_xcursor == NULL) {
6248 rc = MDB_INCOMPATIBLE;
6252 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6253 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6254 MDB_GET_KEY(leaf, key);
6255 rc = mdb_node_read(mc->mc_txn, leaf, data);
6259 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6263 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6266 rc = mdb_cursor_last(mc, key, data);
6269 mfunc = mdb_cursor_last;
6272 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6277 if (mc->mc_flags & C_DEL)
6278 mc->mc_flags ^= C_DEL;
6283 /** Touch all the pages in the cursor stack. Set mc_top.
6284 * Makes sure all the pages are writable, before attempting a write operation.
6285 * @param[in] mc The cursor to operate on.
6288 mdb_cursor_touch(MDB_cursor *mc)
6290 int rc = MDB_SUCCESS;
6292 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6295 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6297 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6298 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6301 *mc->mc_dbflag |= DB_DIRTY;
6306 rc = mdb_page_touch(mc);
6307 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6308 mc->mc_top = mc->mc_snum-1;
6313 /** Do not spill pages to disk if txn is getting full, may fail instead */
6314 #define MDB_NOSPILL 0x8000
6317 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6321 MDB_node *leaf = NULL;
6324 MDB_val xdata, *rdata, dkey, olddata;
6326 int do_sub = 0, insert_key, insert_data;
6327 unsigned int mcount = 0, dcount = 0, nospill;
6330 unsigned int nflags;
6333 if (mc == NULL || key == NULL)
6336 env = mc->mc_txn->mt_env;
6338 /* Check this first so counter will always be zero on any
6341 if (flags & MDB_MULTIPLE) {
6342 dcount = data[1].mv_size;
6343 data[1].mv_size = 0;
6344 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6345 return MDB_INCOMPATIBLE;
6348 nospill = flags & MDB_NOSPILL;
6349 flags &= ~MDB_NOSPILL;
6351 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6352 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6354 if (key->mv_size-1 >= ENV_MAXKEY(env))
6355 return MDB_BAD_VALSIZE;
6357 #if SIZE_MAX > MAXDATASIZE
6358 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6359 return MDB_BAD_VALSIZE;
6361 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6362 return MDB_BAD_VALSIZE;
6365 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6366 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6370 if (flags == MDB_CURRENT) {
6371 if (!(mc->mc_flags & C_INITIALIZED))
6374 } else if (mc->mc_db->md_root == P_INVALID) {
6375 /* new database, cursor has nothing to point to */
6378 mc->mc_flags &= ~C_INITIALIZED;
6383 if (flags & MDB_APPEND) {
6385 rc = mdb_cursor_last(mc, &k2, &d2);
6387 rc = mc->mc_dbx->md_cmp(key, &k2);
6390 mc->mc_ki[mc->mc_top]++;
6392 /* new key is <= last key */
6397 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6399 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6400 DPRINTF(("duplicate key [%s]", DKEY(key)));
6402 return MDB_KEYEXIST;
6404 if (rc && rc != MDB_NOTFOUND)
6408 if (mc->mc_flags & C_DEL)
6409 mc->mc_flags ^= C_DEL;
6411 /* Cursor is positioned, check for room in the dirty list */
6413 if (flags & MDB_MULTIPLE) {
6415 xdata.mv_size = data->mv_size * dcount;
6419 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6423 if (rc == MDB_NO_ROOT) {
6425 /* new database, write a root leaf page */
6426 DPUTS("allocating new root leaf page");
6427 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6430 mdb_cursor_push(mc, np);
6431 mc->mc_db->md_root = np->mp_pgno;
6432 mc->mc_db->md_depth++;
6433 *mc->mc_dbflag |= DB_DIRTY;
6434 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6436 np->mp_flags |= P_LEAF2;
6437 mc->mc_flags |= C_INITIALIZED;
6439 /* make sure all cursor pages are writable */
6440 rc2 = mdb_cursor_touch(mc);
6445 insert_key = insert_data = rc;
6447 /* The key does not exist */
6448 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6449 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6450 LEAFSIZE(key, data) > env->me_nodemax)
6452 /* Too big for a node, insert in sub-DB. Set up an empty
6453 * "old sub-page" for prep_subDB to expand to a full page.
6455 fp_flags = P_LEAF|P_DIRTY;
6457 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6458 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6459 olddata.mv_size = PAGEHDRSZ;
6463 /* there's only a key anyway, so this is a no-op */
6464 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6466 unsigned int ksize = mc->mc_db->md_pad;
6467 if (key->mv_size != ksize)
6468 return MDB_BAD_VALSIZE;
6469 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6470 memcpy(ptr, key->mv_data, ksize);
6472 /* if overwriting slot 0 of leaf, need to
6473 * update branch key if there is a parent page
6475 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6476 unsigned short top = mc->mc_top;
6478 /* slot 0 is always an empty key, find real slot */
6479 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6481 if (mc->mc_ki[mc->mc_top])
6482 rc2 = mdb_update_key(mc, key);
6493 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6494 olddata.mv_size = NODEDSZ(leaf);
6495 olddata.mv_data = NODEDATA(leaf);
6498 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6499 /* Prepare (sub-)page/sub-DB to accept the new item,
6500 * if needed. fp: old sub-page or a header faking
6501 * it. mp: new (sub-)page. offset: growth in page
6502 * size. xdata: node data with new page or DB.
6504 unsigned i, offset = 0;
6505 mp = fp = xdata.mv_data = env->me_pbuf;
6506 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6508 /* Was a single item before, must convert now */
6509 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6511 /* Just overwrite the current item */
6512 if (flags == MDB_CURRENT)
6514 dcmp = mc->mc_dbx->md_dcmp;
6515 #if UINT_MAX < SIZE_MAX
6516 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6517 dcmp = mdb_cmp_clong;
6519 /* does data match? */
6520 if (!dcmp(data, &olddata)) {
6521 if (flags & MDB_NODUPDATA)
6522 return MDB_KEYEXIST;
6527 /* Back up original data item */
6528 dkey.mv_size = olddata.mv_size;
6529 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6531 /* Make sub-page header for the dup items, with dummy body */
6532 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6533 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6534 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6535 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6536 fp->mp_flags |= P_LEAF2;
6537 fp->mp_pad = data->mv_size;
6538 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6540 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6541 (dkey.mv_size & 1) + (data->mv_size & 1);
6543 fp->mp_upper = xdata.mv_size - PAGEBASE;
6544 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6545 } else if (leaf->mn_flags & F_SUBDATA) {
6546 /* Data is on sub-DB, just store it */
6547 flags |= F_DUPDATA|F_SUBDATA;
6550 /* Data is on sub-page */
6551 fp = olddata.mv_data;
6554 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6555 offset = EVEN(NODESIZE + sizeof(indx_t) +
6559 offset = fp->mp_pad;
6560 if (SIZELEFT(fp) < offset) {
6561 offset *= 4; /* space for 4 more */
6564 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6566 fp->mp_flags |= P_DIRTY;
6567 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6568 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6572 xdata.mv_size = olddata.mv_size + offset;
6575 fp_flags = fp->mp_flags;
6576 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6577 /* Too big for a sub-page, convert to sub-DB */
6578 fp_flags &= ~P_SUBP;
6580 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6581 fp_flags |= P_LEAF2;
6582 dummy.md_pad = fp->mp_pad;
6583 dummy.md_flags = MDB_DUPFIXED;
6584 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6585 dummy.md_flags |= MDB_INTEGERKEY;
6591 dummy.md_branch_pages = 0;
6592 dummy.md_leaf_pages = 1;
6593 dummy.md_overflow_pages = 0;
6594 dummy.md_entries = NUMKEYS(fp);
6595 xdata.mv_size = sizeof(MDB_db);
6596 xdata.mv_data = &dummy;
6597 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6599 offset = env->me_psize - olddata.mv_size;
6600 flags |= F_DUPDATA|F_SUBDATA;
6601 dummy.md_root = mp->mp_pgno;
6604 mp->mp_flags = fp_flags | P_DIRTY;
6605 mp->mp_pad = fp->mp_pad;
6606 mp->mp_lower = fp->mp_lower;
6607 mp->mp_upper = fp->mp_upper + offset;
6608 if (fp_flags & P_LEAF2) {
6609 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6611 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6612 olddata.mv_size - fp->mp_upper - PAGEBASE);
6613 for (i=0; i<NUMKEYS(fp); i++)
6614 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6622 mdb_node_del(mc, 0);
6626 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6627 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6628 return MDB_INCOMPATIBLE;
6629 /* overflow page overwrites need special handling */
6630 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6633 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6635 memcpy(&pg, olddata.mv_data, sizeof(pg));
6636 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6638 ovpages = omp->mp_pages;
6640 /* Is the ov page large enough? */
6641 if (ovpages >= dpages) {
6642 if (!(omp->mp_flags & P_DIRTY) &&
6643 (level || (env->me_flags & MDB_WRITEMAP)))
6645 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6648 level = 0; /* dirty in this txn or clean */
6651 if (omp->mp_flags & P_DIRTY) {
6652 /* yes, overwrite it. Note in this case we don't
6653 * bother to try shrinking the page if the new data
6654 * is smaller than the overflow threshold.
6657 /* It is writable only in a parent txn */
6658 size_t sz = (size_t) env->me_psize * ovpages, off;
6659 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6665 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6666 mdb_cassert(mc, rc2 == 0);
6667 if (!(flags & MDB_RESERVE)) {
6668 /* Copy end of page, adjusting alignment so
6669 * compiler may copy words instead of bytes.
6671 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6672 memcpy((size_t *)((char *)np + off),
6673 (size_t *)((char *)omp + off), sz - off);
6676 memcpy(np, omp, sz); /* Copy beginning of page */
6679 SETDSZ(leaf, data->mv_size);
6680 if (F_ISSET(flags, MDB_RESERVE))
6681 data->mv_data = METADATA(omp);
6683 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6687 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6689 } else if (data->mv_size == olddata.mv_size) {
6690 /* same size, just replace it. Note that we could
6691 * also reuse this node if the new data is smaller,
6692 * but instead we opt to shrink the node in that case.
6694 if (F_ISSET(flags, MDB_RESERVE))
6695 data->mv_data = olddata.mv_data;
6696 else if (!(mc->mc_flags & C_SUB))
6697 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6699 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6704 mdb_node_del(mc, 0);
6710 nflags = flags & NODE_ADD_FLAGS;
6711 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6712 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6713 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6714 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6716 nflags |= MDB_SPLIT_REPLACE;
6717 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6719 /* There is room already in this leaf page. */
6720 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6721 if (rc == 0 && insert_key) {
6722 /* Adjust other cursors pointing to mp */
6723 MDB_cursor *m2, *m3;
6724 MDB_dbi dbi = mc->mc_dbi;
6725 unsigned i = mc->mc_top;
6726 MDB_page *mp = mc->mc_pg[i];
6728 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6729 if (mc->mc_flags & C_SUB)
6730 m3 = &m2->mc_xcursor->mx_cursor;
6733 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6734 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6741 if (rc == MDB_SUCCESS) {
6742 /* Now store the actual data in the child DB. Note that we're
6743 * storing the user data in the keys field, so there are strict
6744 * size limits on dupdata. The actual data fields of the child
6745 * DB are all zero size.
6753 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6754 if (flags & MDB_CURRENT) {
6755 xflags = MDB_CURRENT|MDB_NOSPILL;
6757 mdb_xcursor_init1(mc, leaf);
6758 xflags = (flags & MDB_NODUPDATA) ?
6759 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6761 /* converted, write the original data first */
6763 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6767 /* Adjust other cursors pointing to mp */
6769 unsigned i = mc->mc_top;
6770 MDB_page *mp = mc->mc_pg[i];
6772 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6773 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6774 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6775 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6776 mdb_xcursor_init1(m2, leaf);
6780 /* we've done our job */
6783 ecount = mc->mc_xcursor->mx_db.md_entries;
6784 if (flags & MDB_APPENDDUP)
6785 xflags |= MDB_APPEND;
6786 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6787 if (flags & F_SUBDATA) {
6788 void *db = NODEDATA(leaf);
6789 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6791 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6793 /* Increment count unless we just replaced an existing item. */
6795 mc->mc_db->md_entries++;
6797 /* Invalidate txn if we created an empty sub-DB */
6800 /* If we succeeded and the key didn't exist before,
6801 * make sure the cursor is marked valid.
6803 mc->mc_flags |= C_INITIALIZED;
6805 if (flags & MDB_MULTIPLE) {
6808 /* let caller know how many succeeded, if any */
6809 data[1].mv_size = mcount;
6810 if (mcount < dcount) {
6811 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6812 insert_key = insert_data = 0;
6819 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6822 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6827 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6833 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6834 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6836 if (!(mc->mc_flags & C_INITIALIZED))
6839 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6840 return MDB_NOTFOUND;
6842 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6845 rc = mdb_cursor_touch(mc);
6849 mp = mc->mc_pg[mc->mc_top];
6852 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6854 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6855 if (flags & MDB_NODUPDATA) {
6856 /* mdb_cursor_del0() will subtract the final entry */
6857 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6859 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6860 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6862 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6865 /* If sub-DB still has entries, we're done */
6866 if (mc->mc_xcursor->mx_db.md_entries) {
6867 if (leaf->mn_flags & F_SUBDATA) {
6868 /* update subDB info */
6869 void *db = NODEDATA(leaf);
6870 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6873 /* shrink fake page */
6874 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6875 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6876 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6877 /* fix other sub-DB cursors pointed at this fake page */
6878 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6879 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6880 if (m2->mc_pg[mc->mc_top] == mp &&
6881 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6882 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6885 mc->mc_db->md_entries--;
6886 mc->mc_flags |= C_DEL;
6889 /* otherwise fall thru and delete the sub-DB */
6892 if (leaf->mn_flags & F_SUBDATA) {
6893 /* add all the child DB's pages to the free list */
6894 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6899 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6900 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6901 rc = MDB_INCOMPATIBLE;
6905 /* add overflow pages to free list */
6906 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6910 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6911 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6912 (rc = mdb_ovpage_free(mc, omp)))
6917 return mdb_cursor_del0(mc);
6920 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6924 /** Allocate and initialize new pages for a database.
6925 * @param[in] mc a cursor on the database being added to.
6926 * @param[in] flags flags defining what type of page is being allocated.
6927 * @param[in] num the number of pages to allocate. This is usually 1,
6928 * unless allocating overflow pages for a large record.
6929 * @param[out] mp Address of a page, or NULL on failure.
6930 * @return 0 on success, non-zero on failure.
6933 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6938 if ((rc = mdb_page_alloc(mc, num, &np)))
6940 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6941 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6942 np->mp_flags = flags | P_DIRTY;
6943 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6944 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6947 mc->mc_db->md_branch_pages++;
6948 else if (IS_LEAF(np))
6949 mc->mc_db->md_leaf_pages++;
6950 else if (IS_OVERFLOW(np)) {
6951 mc->mc_db->md_overflow_pages += num;
6959 /** Calculate the size of a leaf node.
6960 * The size depends on the environment's page size; if a data item
6961 * is too large it will be put onto an overflow page and the node
6962 * size will only include the key and not the data. Sizes are always
6963 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6964 * of the #MDB_node headers.
6965 * @param[in] env The environment handle.
6966 * @param[in] key The key for the node.
6967 * @param[in] data The data for the node.
6968 * @return The number of bytes needed to store the node.
6971 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6975 sz = LEAFSIZE(key, data);
6976 if (sz > env->me_nodemax) {
6977 /* put on overflow page */
6978 sz -= data->mv_size - sizeof(pgno_t);
6981 return EVEN(sz + sizeof(indx_t));
6984 /** Calculate the size of a branch node.
6985 * The size should depend on the environment's page size but since
6986 * we currently don't support spilling large keys onto overflow
6987 * pages, it's simply the size of the #MDB_node header plus the
6988 * size of the key. Sizes are always rounded up to an even number
6989 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6990 * @param[in] env The environment handle.
6991 * @param[in] key The key for the node.
6992 * @return The number of bytes needed to store the node.
6995 mdb_branch_size(MDB_env *env, MDB_val *key)
7000 if (sz > env->me_nodemax) {
7001 /* put on overflow page */
7002 /* not implemented */
7003 /* sz -= key->size - sizeof(pgno_t); */
7006 return sz + sizeof(indx_t);
7009 /** Add a node to the page pointed to by the cursor.
7010 * @param[in] mc The cursor for this operation.
7011 * @param[in] indx The index on the page where the new node should be added.
7012 * @param[in] key The key for the new node.
7013 * @param[in] data The data for the new node, if any.
7014 * @param[in] pgno The page number, if adding a branch node.
7015 * @param[in] flags Flags for the node.
7016 * @return 0 on success, non-zero on failure. Possible errors are:
7018 * <li>ENOMEM - failed to allocate overflow pages for the node.
7019 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7020 * should never happen since all callers already calculate the
7021 * page's free space before calling this function.
7025 mdb_node_add(MDB_cursor *mc, indx_t indx,
7026 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7029 size_t node_size = NODESIZE;
7033 MDB_page *mp = mc->mc_pg[mc->mc_top];
7034 MDB_page *ofp = NULL; /* overflow page */
7038 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7040 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7041 IS_LEAF(mp) ? "leaf" : "branch",
7042 IS_SUBP(mp) ? "sub-" : "",
7043 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7044 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7047 /* Move higher keys up one slot. */
7048 int ksize = mc->mc_db->md_pad, dif;
7049 char *ptr = LEAF2KEY(mp, indx, ksize);
7050 dif = NUMKEYS(mp) - indx;
7052 memmove(ptr+ksize, ptr, dif*ksize);
7053 /* insert new key */
7054 memcpy(ptr, key->mv_data, ksize);
7056 /* Just using these for counting */
7057 mp->mp_lower += sizeof(indx_t);
7058 mp->mp_upper -= ksize - sizeof(indx_t);
7062 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7064 node_size += key->mv_size;
7066 mdb_cassert(mc, key && data);
7067 if (F_ISSET(flags, F_BIGDATA)) {
7068 /* Data already on overflow page. */
7069 node_size += sizeof(pgno_t);
7070 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7071 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7073 /* Put data on overflow page. */
7074 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7075 data->mv_size, node_size+data->mv_size));
7076 node_size = EVEN(node_size + sizeof(pgno_t));
7077 if ((ssize_t)node_size > room)
7079 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7081 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7085 node_size += data->mv_size;
7088 node_size = EVEN(node_size);
7089 if ((ssize_t)node_size > room)
7093 /* Move higher pointers up one slot. */
7094 for (i = NUMKEYS(mp); i > indx; i--)
7095 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7097 /* Adjust free space offsets. */
7098 ofs = mp->mp_upper - node_size;
7099 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7100 mp->mp_ptrs[indx] = ofs;
7102 mp->mp_lower += sizeof(indx_t);
7104 /* Write the node data. */
7105 node = NODEPTR(mp, indx);
7106 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7107 node->mn_flags = flags;
7109 SETDSZ(node,data->mv_size);
7114 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7117 ndata = NODEDATA(node);
7119 if (F_ISSET(flags, F_BIGDATA))
7120 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7121 else if (F_ISSET(flags, MDB_RESERVE))
7122 data->mv_data = ndata;
7124 memcpy(ndata, data->mv_data, data->mv_size);
7126 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7127 ndata = METADATA(ofp);
7128 if (F_ISSET(flags, MDB_RESERVE))
7129 data->mv_data = ndata;
7131 memcpy(ndata, data->mv_data, data->mv_size);
7138 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7139 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7140 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7141 DPRINTF(("node size = %"Z"u", node_size));
7142 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7143 return MDB_PAGE_FULL;
7146 /** Delete the specified node from a page.
7147 * @param[in] mc Cursor pointing to the node to delete.
7148 * @param[in] ksize The size of a node. Only used if the page is
7149 * part of a #MDB_DUPFIXED database.
7152 mdb_node_del(MDB_cursor *mc, int ksize)
7154 MDB_page *mp = mc->mc_pg[mc->mc_top];
7155 indx_t indx = mc->mc_ki[mc->mc_top];
7157 indx_t i, j, numkeys, ptr;
7161 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7162 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7163 numkeys = NUMKEYS(mp);
7164 mdb_cassert(mc, indx < numkeys);
7167 int x = numkeys - 1 - indx;
7168 base = LEAF2KEY(mp, indx, ksize);
7170 memmove(base, base + ksize, x * ksize);
7171 mp->mp_lower -= sizeof(indx_t);
7172 mp->mp_upper += ksize - sizeof(indx_t);
7176 node = NODEPTR(mp, indx);
7177 sz = NODESIZE + node->mn_ksize;
7179 if (F_ISSET(node->mn_flags, F_BIGDATA))
7180 sz += sizeof(pgno_t);
7182 sz += NODEDSZ(node);
7186 ptr = mp->mp_ptrs[indx];
7187 for (i = j = 0; i < numkeys; i++) {
7189 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7190 if (mp->mp_ptrs[i] < ptr)
7191 mp->mp_ptrs[j] += sz;
7196 base = (char *)mp + mp->mp_upper + PAGEBASE;
7197 memmove(base + sz, base, ptr - mp->mp_upper);
7199 mp->mp_lower -= sizeof(indx_t);
7203 /** Compact the main page after deleting a node on a subpage.
7204 * @param[in] mp The main page to operate on.
7205 * @param[in] indx The index of the subpage on the main page.
7208 mdb_node_shrink(MDB_page *mp, indx_t indx)
7213 indx_t delta, nsize, len, ptr;
7216 node = NODEPTR(mp, indx);
7217 sp = (MDB_page *)NODEDATA(node);
7218 delta = SIZELEFT(sp);
7219 nsize = NODEDSZ(node) - delta;
7221 /* Prepare to shift upward, set len = length(subpage part to shift) */
7225 return; /* do not make the node uneven-sized */
7227 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7228 for (i = NUMKEYS(sp); --i >= 0; )
7229 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7232 sp->mp_upper = sp->mp_lower;
7233 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7234 SETDSZ(node, nsize);
7236 /* Shift <lower nodes...initial part of subpage> upward */
7237 base = (char *)mp + mp->mp_upper + PAGEBASE;
7238 memmove(base + delta, base, (char *)sp + len - base);
7240 ptr = mp->mp_ptrs[indx];
7241 for (i = NUMKEYS(mp); --i >= 0; ) {
7242 if (mp->mp_ptrs[i] <= ptr)
7243 mp->mp_ptrs[i] += delta;
7245 mp->mp_upper += delta;
7248 /** Initial setup of a sorted-dups cursor.
7249 * Sorted duplicates are implemented as a sub-database for the given key.
7250 * The duplicate data items are actually keys of the sub-database.
7251 * Operations on the duplicate data items are performed using a sub-cursor
7252 * initialized when the sub-database is first accessed. This function does
7253 * the preliminary setup of the sub-cursor, filling in the fields that
7254 * depend only on the parent DB.
7255 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7258 mdb_xcursor_init0(MDB_cursor *mc)
7260 MDB_xcursor *mx = mc->mc_xcursor;
7262 mx->mx_cursor.mc_xcursor = NULL;
7263 mx->mx_cursor.mc_txn = mc->mc_txn;
7264 mx->mx_cursor.mc_db = &mx->mx_db;
7265 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7266 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7267 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7268 mx->mx_cursor.mc_snum = 0;
7269 mx->mx_cursor.mc_top = 0;
7270 mx->mx_cursor.mc_flags = C_SUB;
7271 mx->mx_dbx.md_name.mv_size = 0;
7272 mx->mx_dbx.md_name.mv_data = NULL;
7273 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7274 mx->mx_dbx.md_dcmp = NULL;
7275 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7278 /** Final setup of a sorted-dups cursor.
7279 * Sets up the fields that depend on the data from the main cursor.
7280 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7281 * @param[in] node The data containing the #MDB_db record for the
7282 * sorted-dup database.
7285 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7287 MDB_xcursor *mx = mc->mc_xcursor;
7289 if (node->mn_flags & F_SUBDATA) {
7290 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7291 mx->mx_cursor.mc_pg[0] = 0;
7292 mx->mx_cursor.mc_snum = 0;
7293 mx->mx_cursor.mc_top = 0;
7294 mx->mx_cursor.mc_flags = C_SUB;
7296 MDB_page *fp = NODEDATA(node);
7297 mx->mx_db.md_pad = 0;
7298 mx->mx_db.md_flags = 0;
7299 mx->mx_db.md_depth = 1;
7300 mx->mx_db.md_branch_pages = 0;
7301 mx->mx_db.md_leaf_pages = 1;
7302 mx->mx_db.md_overflow_pages = 0;
7303 mx->mx_db.md_entries = NUMKEYS(fp);
7304 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7305 mx->mx_cursor.mc_snum = 1;
7306 mx->mx_cursor.mc_top = 0;
7307 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7308 mx->mx_cursor.mc_pg[0] = fp;
7309 mx->mx_cursor.mc_ki[0] = 0;
7310 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7311 mx->mx_db.md_flags = MDB_DUPFIXED;
7312 mx->mx_db.md_pad = fp->mp_pad;
7313 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7314 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7317 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7318 mx->mx_db.md_root));
7319 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7320 #if UINT_MAX < SIZE_MAX
7321 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7322 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7326 /** Initialize a cursor for a given transaction and database. */
7328 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7331 mc->mc_backup = NULL;
7334 mc->mc_db = &txn->mt_dbs[dbi];
7335 mc->mc_dbx = &txn->mt_dbxs[dbi];
7336 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7342 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7343 mdb_tassert(txn, mx != NULL);
7344 mc->mc_xcursor = mx;
7345 mdb_xcursor_init0(mc);
7347 mc->mc_xcursor = NULL;
7349 if (*mc->mc_dbflag & DB_STALE) {
7350 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7355 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7358 size_t size = sizeof(MDB_cursor);
7360 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7363 if (txn->mt_flags & MDB_TXN_BLOCKED)
7366 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7369 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7370 size += sizeof(MDB_xcursor);
7372 if ((mc = malloc(size)) != NULL) {
7373 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7374 if (txn->mt_cursors) {
7375 mc->mc_next = txn->mt_cursors[dbi];
7376 txn->mt_cursors[dbi] = mc;
7377 mc->mc_flags |= C_UNTRACK;
7389 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7391 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7394 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7397 if (txn->mt_flags & MDB_TXN_BLOCKED)
7400 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7404 /* Return the count of duplicate data items for the current key */
7406 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7410 if (mc == NULL || countp == NULL)
7413 if (mc->mc_xcursor == NULL)
7414 return MDB_INCOMPATIBLE;
7416 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7419 if (!(mc->mc_flags & C_INITIALIZED))
7422 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7423 return MDB_NOTFOUND;
7425 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7426 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7429 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7432 *countp = mc->mc_xcursor->mx_db.md_entries;
7438 mdb_cursor_close(MDB_cursor *mc)
7440 if (mc && !mc->mc_backup) {
7441 /* remove from txn, if tracked */
7442 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7443 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7444 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7446 *prev = mc->mc_next;
7453 mdb_cursor_txn(MDB_cursor *mc)
7455 if (!mc) return NULL;
7460 mdb_cursor_dbi(MDB_cursor *mc)
7465 /** Replace the key for a branch node with a new key.
7466 * @param[in] mc Cursor pointing to the node to operate on.
7467 * @param[in] key The new key to use.
7468 * @return 0 on success, non-zero on failure.
7471 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7477 int delta, ksize, oksize;
7478 indx_t ptr, i, numkeys, indx;
7481 indx = mc->mc_ki[mc->mc_top];
7482 mp = mc->mc_pg[mc->mc_top];
7483 node = NODEPTR(mp, indx);
7484 ptr = mp->mp_ptrs[indx];
7488 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7489 k2.mv_data = NODEKEY(node);
7490 k2.mv_size = node->mn_ksize;
7491 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7493 mdb_dkey(&k2, kbuf2),
7499 /* Sizes must be 2-byte aligned. */
7500 ksize = EVEN(key->mv_size);
7501 oksize = EVEN(node->mn_ksize);
7502 delta = ksize - oksize;
7504 /* Shift node contents if EVEN(key length) changed. */
7506 if (delta > 0 && SIZELEFT(mp) < delta) {
7508 /* not enough space left, do a delete and split */
7509 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7510 pgno = NODEPGNO(node);
7511 mdb_node_del(mc, 0);
7512 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7515 numkeys = NUMKEYS(mp);
7516 for (i = 0; i < numkeys; i++) {
7517 if (mp->mp_ptrs[i] <= ptr)
7518 mp->mp_ptrs[i] -= delta;
7521 base = (char *)mp + mp->mp_upper + PAGEBASE;
7522 len = ptr - mp->mp_upper + NODESIZE;
7523 memmove(base - delta, base, len);
7524 mp->mp_upper -= delta;
7526 node = NODEPTR(mp, indx);
7529 /* But even if no shift was needed, update ksize */
7530 if (node->mn_ksize != key->mv_size)
7531 node->mn_ksize = key->mv_size;
7534 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7540 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7542 /** Move a node from csrc to cdst.
7545 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7552 unsigned short flags;
7556 /* Mark src and dst as dirty. */
7557 if ((rc = mdb_page_touch(csrc)) ||
7558 (rc = mdb_page_touch(cdst)))
7561 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7562 key.mv_size = csrc->mc_db->md_pad;
7563 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7565 data.mv_data = NULL;
7569 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7570 mdb_cassert(csrc, !((size_t)srcnode & 1));
7571 srcpg = NODEPGNO(srcnode);
7572 flags = srcnode->mn_flags;
7573 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7574 unsigned int snum = csrc->mc_snum;
7576 /* must find the lowest key below src */
7577 rc = mdb_page_search_lowest(csrc);
7580 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7581 key.mv_size = csrc->mc_db->md_pad;
7582 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7584 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7585 key.mv_size = NODEKSZ(s2);
7586 key.mv_data = NODEKEY(s2);
7588 csrc->mc_snum = snum--;
7589 csrc->mc_top = snum;
7591 key.mv_size = NODEKSZ(srcnode);
7592 key.mv_data = NODEKEY(srcnode);
7594 data.mv_size = NODEDSZ(srcnode);
7595 data.mv_data = NODEDATA(srcnode);
7597 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7598 unsigned int snum = cdst->mc_snum;
7601 /* must find the lowest key below dst */
7602 mdb_cursor_copy(cdst, &mn);
7603 rc = mdb_page_search_lowest(&mn);
7606 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7607 bkey.mv_size = mn.mc_db->md_pad;
7608 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7610 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7611 bkey.mv_size = NODEKSZ(s2);
7612 bkey.mv_data = NODEKEY(s2);
7614 mn.mc_snum = snum--;
7617 rc = mdb_update_key(&mn, &bkey);
7622 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7623 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7624 csrc->mc_ki[csrc->mc_top],
7626 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7627 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7629 /* Add the node to the destination page.
7631 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7632 if (rc != MDB_SUCCESS)
7635 /* Delete the node from the source page.
7637 mdb_node_del(csrc, key.mv_size);
7640 /* Adjust other cursors pointing to mp */
7641 MDB_cursor *m2, *m3;
7642 MDB_dbi dbi = csrc->mc_dbi;
7643 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7645 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7646 if (csrc->mc_flags & C_SUB)
7647 m3 = &m2->mc_xcursor->mx_cursor;
7650 if (m3 == csrc) continue;
7651 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7652 csrc->mc_ki[csrc->mc_top]) {
7653 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7654 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7659 /* Update the parent separators.
7661 if (csrc->mc_ki[csrc->mc_top] == 0) {
7662 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7663 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7664 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7666 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7667 key.mv_size = NODEKSZ(srcnode);
7668 key.mv_data = NODEKEY(srcnode);
7670 DPRINTF(("update separator for source page %"Z"u to [%s]",
7671 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7672 mdb_cursor_copy(csrc, &mn);
7675 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7678 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7680 indx_t ix = csrc->mc_ki[csrc->mc_top];
7681 nullkey.mv_size = 0;
7682 csrc->mc_ki[csrc->mc_top] = 0;
7683 rc = mdb_update_key(csrc, &nullkey);
7684 csrc->mc_ki[csrc->mc_top] = ix;
7685 mdb_cassert(csrc, rc == MDB_SUCCESS);
7689 if (cdst->mc_ki[cdst->mc_top] == 0) {
7690 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7691 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7692 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7694 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7695 key.mv_size = NODEKSZ(srcnode);
7696 key.mv_data = NODEKEY(srcnode);
7698 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7699 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7700 mdb_cursor_copy(cdst, &mn);
7703 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7706 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7708 indx_t ix = cdst->mc_ki[cdst->mc_top];
7709 nullkey.mv_size = 0;
7710 cdst->mc_ki[cdst->mc_top] = 0;
7711 rc = mdb_update_key(cdst, &nullkey);
7712 cdst->mc_ki[cdst->mc_top] = ix;
7713 mdb_cassert(cdst, rc == MDB_SUCCESS);
7720 /** Merge one page into another.
7721 * The nodes from the page pointed to by \b csrc will
7722 * be copied to the page pointed to by \b cdst and then
7723 * the \b csrc page will be freed.
7724 * @param[in] csrc Cursor pointing to the source page.
7725 * @param[in] cdst Cursor pointing to the destination page.
7726 * @return 0 on success, non-zero on failure.
7729 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7731 MDB_page *psrc, *pdst;
7738 psrc = csrc->mc_pg[csrc->mc_top];
7739 pdst = cdst->mc_pg[cdst->mc_top];
7741 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7743 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7744 mdb_cassert(csrc, cdst->mc_snum > 1);
7746 /* Mark dst as dirty. */
7747 if ((rc = mdb_page_touch(cdst)))
7750 /* Move all nodes from src to dst.
7752 j = nkeys = NUMKEYS(pdst);
7753 if (IS_LEAF2(psrc)) {
7754 key.mv_size = csrc->mc_db->md_pad;
7755 key.mv_data = METADATA(psrc);
7756 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7757 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7758 if (rc != MDB_SUCCESS)
7760 key.mv_data = (char *)key.mv_data + key.mv_size;
7763 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7764 srcnode = NODEPTR(psrc, i);
7765 if (i == 0 && IS_BRANCH(psrc)) {
7768 mdb_cursor_copy(csrc, &mn);
7769 /* must find the lowest key below src */
7770 rc = mdb_page_search_lowest(&mn);
7773 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7774 key.mv_size = mn.mc_db->md_pad;
7775 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7777 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7778 key.mv_size = NODEKSZ(s2);
7779 key.mv_data = NODEKEY(s2);
7782 key.mv_size = srcnode->mn_ksize;
7783 key.mv_data = NODEKEY(srcnode);
7786 data.mv_size = NODEDSZ(srcnode);
7787 data.mv_data = NODEDATA(srcnode);
7788 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7789 if (rc != MDB_SUCCESS)
7794 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7795 pdst->mp_pgno, NUMKEYS(pdst),
7796 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7798 /* Unlink the src page from parent and add to free list.
7801 mdb_node_del(csrc, 0);
7802 if (csrc->mc_ki[csrc->mc_top] == 0) {
7804 rc = mdb_update_key(csrc, &key);
7812 psrc = csrc->mc_pg[csrc->mc_top];
7813 /* If not operating on FreeDB, allow this page to be reused
7814 * in this txn. Otherwise just add to free list.
7816 rc = mdb_page_loose(csrc, psrc);
7820 csrc->mc_db->md_leaf_pages--;
7822 csrc->mc_db->md_branch_pages--;
7824 /* Adjust other cursors pointing to mp */
7825 MDB_cursor *m2, *m3;
7826 MDB_dbi dbi = csrc->mc_dbi;
7828 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7829 if (csrc->mc_flags & C_SUB)
7830 m3 = &m2->mc_xcursor->mx_cursor;
7833 if (m3 == csrc) continue;
7834 if (m3->mc_snum < csrc->mc_snum) continue;
7835 if (m3->mc_pg[csrc->mc_top] == psrc) {
7836 m3->mc_pg[csrc->mc_top] = pdst;
7837 m3->mc_ki[csrc->mc_top] += nkeys;
7842 unsigned int snum = cdst->mc_snum;
7843 uint16_t depth = cdst->mc_db->md_depth;
7844 mdb_cursor_pop(cdst);
7845 rc = mdb_rebalance(cdst);
7846 /* Did the tree shrink? */
7847 if (depth > cdst->mc_db->md_depth)
7849 cdst->mc_snum = snum;
7850 cdst->mc_top = snum-1;
7855 /** Copy the contents of a cursor.
7856 * @param[in] csrc The cursor to copy from.
7857 * @param[out] cdst The cursor to copy to.
7860 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7864 cdst->mc_txn = csrc->mc_txn;
7865 cdst->mc_dbi = csrc->mc_dbi;
7866 cdst->mc_db = csrc->mc_db;
7867 cdst->mc_dbx = csrc->mc_dbx;
7868 cdst->mc_snum = csrc->mc_snum;
7869 cdst->mc_top = csrc->mc_top;
7870 cdst->mc_flags = csrc->mc_flags;
7872 for (i=0; i<csrc->mc_snum; i++) {
7873 cdst->mc_pg[i] = csrc->mc_pg[i];
7874 cdst->mc_ki[i] = csrc->mc_ki[i];
7878 /** Rebalance the tree after a delete operation.
7879 * @param[in] mc Cursor pointing to the page where rebalancing
7881 * @return 0 on success, non-zero on failure.
7884 mdb_rebalance(MDB_cursor *mc)
7888 unsigned int ptop, minkeys;
7892 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7893 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7894 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7895 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7896 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7898 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7899 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7900 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7901 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7905 if (mc->mc_snum < 2) {
7906 MDB_page *mp = mc->mc_pg[0];
7908 DPUTS("Can't rebalance a subpage, ignoring");
7911 if (NUMKEYS(mp) == 0) {
7912 DPUTS("tree is completely empty");
7913 mc->mc_db->md_root = P_INVALID;
7914 mc->mc_db->md_depth = 0;
7915 mc->mc_db->md_leaf_pages = 0;
7916 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7919 /* Adjust cursors pointing to mp */
7922 mc->mc_flags &= ~C_INITIALIZED;
7924 MDB_cursor *m2, *m3;
7925 MDB_dbi dbi = mc->mc_dbi;
7927 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7928 if (mc->mc_flags & C_SUB)
7929 m3 = &m2->mc_xcursor->mx_cursor;
7932 if (m3->mc_snum < mc->mc_snum) continue;
7933 if (m3->mc_pg[0] == mp) {
7936 m3->mc_flags &= ~C_INITIALIZED;
7940 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7942 DPUTS("collapsing root page!");
7943 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7946 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7947 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7950 mc->mc_db->md_depth--;
7951 mc->mc_db->md_branch_pages--;
7952 mc->mc_ki[0] = mc->mc_ki[1];
7953 for (i = 1; i<mc->mc_db->md_depth; i++) {
7954 mc->mc_pg[i] = mc->mc_pg[i+1];
7955 mc->mc_ki[i] = mc->mc_ki[i+1];
7958 /* Adjust other cursors pointing to mp */
7959 MDB_cursor *m2, *m3;
7960 MDB_dbi dbi = mc->mc_dbi;
7962 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7963 if (mc->mc_flags & C_SUB)
7964 m3 = &m2->mc_xcursor->mx_cursor;
7967 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7968 if (m3->mc_pg[0] == mp) {
7969 for (i=0; i<m3->mc_snum; i++) {
7970 m3->mc_pg[i] = m3->mc_pg[i+1];
7971 m3->mc_ki[i] = m3->mc_ki[i+1];
7979 DPUTS("root page doesn't need rebalancing");
7983 /* The parent (branch page) must have at least 2 pointers,
7984 * otherwise the tree is invalid.
7986 ptop = mc->mc_top-1;
7987 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7989 /* Leaf page fill factor is below the threshold.
7990 * Try to move keys from left or right neighbor, or
7991 * merge with a neighbor page.
7996 mdb_cursor_copy(mc, &mn);
7997 mn.mc_xcursor = NULL;
7999 oldki = mc->mc_ki[mc->mc_top];
8000 if (mc->mc_ki[ptop] == 0) {
8001 /* We're the leftmost leaf in our parent.
8003 DPUTS("reading right neighbor");
8005 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8006 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8009 mn.mc_ki[mn.mc_top] = 0;
8010 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8012 /* There is at least one neighbor to the left.
8014 DPUTS("reading left neighbor");
8016 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8017 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8020 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8021 mc->mc_ki[mc->mc_top] = 0;
8024 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8025 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8026 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8028 /* If the neighbor page is above threshold and has enough keys,
8029 * move one key from it. Otherwise we should try to merge them.
8030 * (A branch page must never have less than 2 keys.)
8032 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
8033 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8034 rc = mdb_node_move(&mn, mc);
8035 if (mc->mc_ki[ptop]) {
8039 if (mc->mc_ki[ptop] == 0) {
8040 rc = mdb_page_merge(&mn, mc);
8043 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8044 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8045 /* We want mdb_rebalance to find mn when doing fixups */
8046 if (mc->mc_flags & C_SUB) {
8047 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8048 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8049 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8051 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8052 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8054 rc = mdb_page_merge(mc, &mn);
8055 if (mc->mc_flags & C_SUB)
8056 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8058 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8059 mdb_cursor_copy(&mn, mc);
8061 mc->mc_flags &= ~C_EOF;
8063 mc->mc_ki[mc->mc_top] = oldki;
8067 /** Complete a delete operation started by #mdb_cursor_del(). */
8069 mdb_cursor_del0(MDB_cursor *mc)
8076 ki = mc->mc_ki[mc->mc_top];
8077 mdb_node_del(mc, mc->mc_db->md_pad);
8078 mc->mc_db->md_entries--;
8079 rc = mdb_rebalance(mc);
8081 if (rc == MDB_SUCCESS) {
8082 MDB_cursor *m2, *m3;
8083 MDB_dbi dbi = mc->mc_dbi;
8085 /* DB is totally empty now, just bail out.
8086 * Other cursors adjustments were already done
8087 * by mdb_rebalance and aren't needed here.
8092 mp = mc->mc_pg[mc->mc_top];
8093 nkeys = NUMKEYS(mp);
8095 /* if mc points past last node in page, find next sibling */
8096 if (mc->mc_ki[mc->mc_top] >= nkeys) {
8097 rc = mdb_cursor_sibling(mc, 1);
8098 if (rc == MDB_NOTFOUND) {
8099 mc->mc_flags |= C_EOF;
8104 /* Adjust other cursors pointing to mp */
8105 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8106 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8107 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8109 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8111 if (m3->mc_pg[mc->mc_top] == mp) {
8112 if (m3->mc_ki[mc->mc_top] >= ki) {
8113 m3->mc_flags |= C_DEL;
8114 if (m3->mc_ki[mc->mc_top] > ki)
8115 m3->mc_ki[mc->mc_top]--;
8116 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8117 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8119 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8120 rc = mdb_cursor_sibling(m3, 1);
8121 if (rc == MDB_NOTFOUND) {
8122 m3->mc_flags |= C_EOF;
8128 mc->mc_flags |= C_DEL;
8132 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8137 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8138 MDB_val *key, MDB_val *data)
8140 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8143 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8144 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8146 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8147 /* must ignore any data */
8151 return mdb_del0(txn, dbi, key, data, 0);
8155 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8156 MDB_val *key, MDB_val *data, unsigned flags)
8161 MDB_val rdata, *xdata;
8165 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8167 mdb_cursor_init(&mc, txn, dbi, &mx);
8176 flags |= MDB_NODUPDATA;
8178 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8180 /* let mdb_page_split know about this cursor if needed:
8181 * delete will trigger a rebalance; if it needs to move
8182 * a node from one page to another, it will have to
8183 * update the parent's separator key(s). If the new sepkey
8184 * is larger than the current one, the parent page may
8185 * run out of space, triggering a split. We need this
8186 * cursor to be consistent until the end of the rebalance.
8188 mc.mc_flags |= C_UNTRACK;
8189 mc.mc_next = txn->mt_cursors[dbi];
8190 txn->mt_cursors[dbi] = &mc;
8191 rc = mdb_cursor_del(&mc, flags);
8192 txn->mt_cursors[dbi] = mc.mc_next;
8197 /** Split a page and insert a new node.
8198 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8199 * The cursor will be updated to point to the actual page and index where
8200 * the node got inserted after the split.
8201 * @param[in] newkey The key for the newly inserted node.
8202 * @param[in] newdata The data for the newly inserted node.
8203 * @param[in] newpgno The page number, if the new node is a branch node.
8204 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8205 * @return 0 on success, non-zero on failure.
8208 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8209 unsigned int nflags)
8212 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8215 int i, j, split_indx, nkeys, pmax;
8216 MDB_env *env = mc->mc_txn->mt_env;
8218 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8219 MDB_page *copy = NULL;
8220 MDB_page *mp, *rp, *pp;
8225 mp = mc->mc_pg[mc->mc_top];
8226 newindx = mc->mc_ki[mc->mc_top];
8227 nkeys = NUMKEYS(mp);
8229 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8230 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8231 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8233 /* Create a right sibling. */
8234 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8236 rp->mp_pad = mp->mp_pad;
8237 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8239 if (mc->mc_snum < 2) {
8240 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8242 /* shift current top to make room for new parent */
8243 mc->mc_pg[1] = mc->mc_pg[0];
8244 mc->mc_ki[1] = mc->mc_ki[0];
8247 mc->mc_db->md_root = pp->mp_pgno;
8248 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8249 mc->mc_db->md_depth++;
8252 /* Add left (implicit) pointer. */
8253 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8254 /* undo the pre-push */
8255 mc->mc_pg[0] = mc->mc_pg[1];
8256 mc->mc_ki[0] = mc->mc_ki[1];
8257 mc->mc_db->md_root = mp->mp_pgno;
8258 mc->mc_db->md_depth--;
8265 ptop = mc->mc_top-1;
8266 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8269 mc->mc_flags |= C_SPLITTING;
8270 mdb_cursor_copy(mc, &mn);
8271 mn.mc_pg[mn.mc_top] = rp;
8272 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8274 if (nflags & MDB_APPEND) {
8275 mn.mc_ki[mn.mc_top] = 0;
8277 split_indx = newindx;
8281 split_indx = (nkeys+1) / 2;
8286 unsigned int lsize, rsize, ksize;
8287 /* Move half of the keys to the right sibling */
8288 x = mc->mc_ki[mc->mc_top] - split_indx;
8289 ksize = mc->mc_db->md_pad;
8290 split = LEAF2KEY(mp, split_indx, ksize);
8291 rsize = (nkeys - split_indx) * ksize;
8292 lsize = (nkeys - split_indx) * sizeof(indx_t);
8293 mp->mp_lower -= lsize;
8294 rp->mp_lower += lsize;
8295 mp->mp_upper += rsize - lsize;
8296 rp->mp_upper -= rsize - lsize;
8297 sepkey.mv_size = ksize;
8298 if (newindx == split_indx) {
8299 sepkey.mv_data = newkey->mv_data;
8301 sepkey.mv_data = split;
8304 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8305 memcpy(rp->mp_ptrs, split, rsize);
8306 sepkey.mv_data = rp->mp_ptrs;
8307 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8308 memcpy(ins, newkey->mv_data, ksize);
8309 mp->mp_lower += sizeof(indx_t);
8310 mp->mp_upper -= ksize - sizeof(indx_t);
8313 memcpy(rp->mp_ptrs, split, x * ksize);
8314 ins = LEAF2KEY(rp, x, ksize);
8315 memcpy(ins, newkey->mv_data, ksize);
8316 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8317 rp->mp_lower += sizeof(indx_t);
8318 rp->mp_upper -= ksize - sizeof(indx_t);
8319 mc->mc_ki[mc->mc_top] = x;
8320 mc->mc_pg[mc->mc_top] = rp;
8323 int psize, nsize, k;
8324 /* Maximum free space in an empty page */
8325 pmax = env->me_psize - PAGEHDRSZ;
8327 nsize = mdb_leaf_size(env, newkey, newdata);
8329 nsize = mdb_branch_size(env, newkey);
8330 nsize = EVEN(nsize);
8332 /* grab a page to hold a temporary copy */
8333 copy = mdb_page_malloc(mc->mc_txn, 1);
8338 copy->mp_pgno = mp->mp_pgno;
8339 copy->mp_flags = mp->mp_flags;
8340 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8341 copy->mp_upper = env->me_psize - PAGEBASE;
8343 /* prepare to insert */
8344 for (i=0, j=0; i<nkeys; i++) {
8346 copy->mp_ptrs[j++] = 0;
8348 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8351 /* When items are relatively large the split point needs
8352 * to be checked, because being off-by-one will make the
8353 * difference between success or failure in mdb_node_add.
8355 * It's also relevant if a page happens to be laid out
8356 * such that one half of its nodes are all "small" and
8357 * the other half of its nodes are "large." If the new
8358 * item is also "large" and falls on the half with
8359 * "large" nodes, it also may not fit.
8361 * As a final tweak, if the new item goes on the last
8362 * spot on the page (and thus, onto the new page), bias
8363 * the split so the new page is emptier than the old page.
8364 * This yields better packing during sequential inserts.
8366 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8367 /* Find split point */
8369 if (newindx <= split_indx || newindx >= nkeys) {
8371 k = newindx >= nkeys ? nkeys : split_indx+2;
8376 for (; i!=k; i+=j) {
8381 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8382 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8384 if (F_ISSET(node->mn_flags, F_BIGDATA))
8385 psize += sizeof(pgno_t);
8387 psize += NODEDSZ(node);
8389 psize = EVEN(psize);
8391 if (psize > pmax || i == k-j) {
8392 split_indx = i + (j<0);
8397 if (split_indx == newindx) {
8398 sepkey.mv_size = newkey->mv_size;
8399 sepkey.mv_data = newkey->mv_data;
8401 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8402 sepkey.mv_size = node->mn_ksize;
8403 sepkey.mv_data = NODEKEY(node);
8408 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8410 /* Copy separator key to the parent.
8412 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8416 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8421 if (mn.mc_snum == mc->mc_snum) {
8422 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8423 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8424 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8425 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8430 /* Right page might now have changed parent.
8431 * Check if left page also changed parent.
8433 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8434 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8435 for (i=0; i<ptop; i++) {
8436 mc->mc_pg[i] = mn.mc_pg[i];
8437 mc->mc_ki[i] = mn.mc_ki[i];
8439 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8440 if (mn.mc_ki[ptop]) {
8441 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8443 /* find right page's left sibling */
8444 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8445 mdb_cursor_sibling(mc, 0);
8450 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8453 mc->mc_flags ^= C_SPLITTING;
8454 if (rc != MDB_SUCCESS) {
8457 if (nflags & MDB_APPEND) {
8458 mc->mc_pg[mc->mc_top] = rp;
8459 mc->mc_ki[mc->mc_top] = 0;
8460 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8463 for (i=0; i<mc->mc_top; i++)
8464 mc->mc_ki[i] = mn.mc_ki[i];
8465 } else if (!IS_LEAF2(mp)) {
8467 mc->mc_pg[mc->mc_top] = rp;
8472 rkey.mv_data = newkey->mv_data;
8473 rkey.mv_size = newkey->mv_size;
8479 /* Update index for the new key. */
8480 mc->mc_ki[mc->mc_top] = j;
8482 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8483 rkey.mv_data = NODEKEY(node);
8484 rkey.mv_size = node->mn_ksize;
8486 xdata.mv_data = NODEDATA(node);
8487 xdata.mv_size = NODEDSZ(node);
8490 pgno = NODEPGNO(node);
8491 flags = node->mn_flags;
8494 if (!IS_LEAF(mp) && j == 0) {
8495 /* First branch index doesn't need key data. */
8499 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8505 mc->mc_pg[mc->mc_top] = copy;
8510 } while (i != split_indx);
8512 nkeys = NUMKEYS(copy);
8513 for (i=0; i<nkeys; i++)
8514 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8515 mp->mp_lower = copy->mp_lower;
8516 mp->mp_upper = copy->mp_upper;
8517 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8518 env->me_psize - copy->mp_upper - PAGEBASE);
8520 /* reset back to original page */
8521 if (newindx < split_indx) {
8522 mc->mc_pg[mc->mc_top] = mp;
8523 if (nflags & MDB_RESERVE) {
8524 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8525 if (!(node->mn_flags & F_BIGDATA))
8526 newdata->mv_data = NODEDATA(node);
8529 mc->mc_pg[mc->mc_top] = rp;
8531 /* Make sure mc_ki is still valid.
8533 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8534 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8535 for (i=0; i<=ptop; i++) {
8536 mc->mc_pg[i] = mn.mc_pg[i];
8537 mc->mc_ki[i] = mn.mc_ki[i];
8544 /* Adjust other cursors pointing to mp */
8545 MDB_cursor *m2, *m3;
8546 MDB_dbi dbi = mc->mc_dbi;
8547 int fixup = NUMKEYS(mp);
8549 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8550 if (mc->mc_flags & C_SUB)
8551 m3 = &m2->mc_xcursor->mx_cursor;
8556 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8558 if (m3->mc_flags & C_SPLITTING)
8563 for (k=m3->mc_top; k>=0; k--) {
8564 m3->mc_ki[k+1] = m3->mc_ki[k];
8565 m3->mc_pg[k+1] = m3->mc_pg[k];
8567 if (m3->mc_ki[0] >= split_indx) {
8572 m3->mc_pg[0] = mc->mc_pg[0];
8576 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8577 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8578 m3->mc_ki[mc->mc_top]++;
8579 if (m3->mc_ki[mc->mc_top] >= fixup) {
8580 m3->mc_pg[mc->mc_top] = rp;
8581 m3->mc_ki[mc->mc_top] -= fixup;
8582 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8584 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8585 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8590 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8593 if (copy) /* tmp page */
8594 mdb_page_free(env, copy);
8596 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8601 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8602 MDB_val *key, MDB_val *data, unsigned int flags)
8607 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8610 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8613 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8614 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8616 mdb_cursor_init(&mc, txn, dbi, &mx);
8617 return mdb_cursor_put(&mc, key, data, flags);
8621 #define MDB_WBUF (1024*1024)
8624 /** State needed for a compacting copy. */
8625 typedef struct mdb_copy {
8626 pthread_mutex_t mc_mutex;
8627 pthread_cond_t mc_cond;
8634 pgno_t mc_next_pgno;
8637 volatile int mc_new;
8642 /** Dedicated writer thread for compacting copy. */
8643 static THREAD_RET ESECT
8644 mdb_env_copythr(void *arg)
8648 int toggle = 0, wsize, rc;
8651 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8654 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8657 pthread_mutex_lock(&my->mc_mutex);
8659 pthread_cond_signal(&my->mc_cond);
8662 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8663 if (my->mc_new < 0) {
8668 wsize = my->mc_wlen[toggle];
8669 ptr = my->mc_wbuf[toggle];
8672 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8676 } else if (len > 0) {
8690 /* If there's an overflow page tail, write it too */
8691 if (my->mc_olen[toggle]) {
8692 wsize = my->mc_olen[toggle];
8693 ptr = my->mc_over[toggle];
8694 my->mc_olen[toggle] = 0;
8697 my->mc_wlen[toggle] = 0;
8699 pthread_cond_signal(&my->mc_cond);
8701 pthread_cond_signal(&my->mc_cond);
8702 pthread_mutex_unlock(&my->mc_mutex);
8703 return (THREAD_RET)0;
8707 /** Tell the writer thread there's a buffer ready to write */
8709 mdb_env_cthr_toggle(mdb_copy *my, int st)
8711 int toggle = my->mc_toggle ^ 1;
8712 pthread_mutex_lock(&my->mc_mutex);
8713 if (my->mc_status) {
8714 pthread_mutex_unlock(&my->mc_mutex);
8715 return my->mc_status;
8717 while (my->mc_new == 1)
8718 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8720 my->mc_toggle = toggle;
8721 pthread_cond_signal(&my->mc_cond);
8722 pthread_mutex_unlock(&my->mc_mutex);
8726 /** Depth-first tree traversal for compacting copy. */
8728 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8731 MDB_txn *txn = my->mc_txn;
8733 MDB_page *mo, *mp, *leaf;
8738 /* Empty DB, nothing to do */
8739 if (*pg == P_INVALID)
8746 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8749 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8753 /* Make cursor pages writable */
8754 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8758 for (i=0; i<mc.mc_top; i++) {
8759 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8760 mc.mc_pg[i] = (MDB_page *)ptr;
8761 ptr += my->mc_env->me_psize;
8764 /* This is writable space for a leaf page. Usually not needed. */
8765 leaf = (MDB_page *)ptr;
8767 toggle = my->mc_toggle;
8768 while (mc.mc_snum > 0) {
8770 mp = mc.mc_pg[mc.mc_top];
8774 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8775 for (i=0; i<n; i++) {
8776 ni = NODEPTR(mp, i);
8777 if (ni->mn_flags & F_BIGDATA) {
8781 /* Need writable leaf */
8783 mc.mc_pg[mc.mc_top] = leaf;
8784 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8786 ni = NODEPTR(mp, i);
8789 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8790 rc = mdb_page_get(txn, pg, &omp, NULL);
8793 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8794 rc = mdb_env_cthr_toggle(my, 1);
8797 toggle = my->mc_toggle;
8799 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8800 memcpy(mo, omp, my->mc_env->me_psize);
8801 mo->mp_pgno = my->mc_next_pgno;
8802 my->mc_next_pgno += omp->mp_pages;
8803 my->mc_wlen[toggle] += my->mc_env->me_psize;
8804 if (omp->mp_pages > 1) {
8805 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8806 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8807 rc = mdb_env_cthr_toggle(my, 1);
8810 toggle = my->mc_toggle;
8812 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8813 } else if (ni->mn_flags & F_SUBDATA) {
8816 /* Need writable leaf */
8818 mc.mc_pg[mc.mc_top] = leaf;
8819 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8821 ni = NODEPTR(mp, i);
8824 memcpy(&db, NODEDATA(ni), sizeof(db));
8825 my->mc_toggle = toggle;
8826 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8829 toggle = my->mc_toggle;
8830 memcpy(NODEDATA(ni), &db, sizeof(db));
8835 mc.mc_ki[mc.mc_top]++;
8836 if (mc.mc_ki[mc.mc_top] < n) {
8839 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8841 rc = mdb_page_get(txn, pg, &mp, NULL);
8846 mc.mc_ki[mc.mc_top] = 0;
8847 if (IS_BRANCH(mp)) {
8848 /* Whenever we advance to a sibling branch page,
8849 * we must proceed all the way down to its first leaf.
8851 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8854 mc.mc_pg[mc.mc_top] = mp;
8858 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8859 rc = mdb_env_cthr_toggle(my, 1);
8862 toggle = my->mc_toggle;
8864 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8865 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8866 mo->mp_pgno = my->mc_next_pgno++;
8867 my->mc_wlen[toggle] += my->mc_env->me_psize;
8869 /* Update parent if there is one */
8870 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8871 SETPGNO(ni, mo->mp_pgno);
8872 mdb_cursor_pop(&mc);
8874 /* Otherwise we're done */
8884 /** Copy environment with compaction. */
8886 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8891 MDB_txn *txn = NULL;
8896 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8897 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8898 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8899 if (my.mc_wbuf[0] == NULL)
8902 pthread_mutex_init(&my.mc_mutex, NULL);
8903 pthread_cond_init(&my.mc_cond, NULL);
8904 #ifdef HAVE_MEMALIGN
8905 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8906 if (my.mc_wbuf[0] == NULL)
8909 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8914 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8915 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8920 my.mc_next_pgno = NUM_METAS;
8926 THREAD_CREATE(thr, mdb_env_copythr, &my);
8928 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8932 mp = (MDB_page *)my.mc_wbuf[0];
8933 memset(mp, 0, NUM_METAS * env->me_psize);
8935 mp->mp_flags = P_META;
8936 mm = (MDB_meta *)METADATA(mp);
8937 mdb_env_init_meta0(env, mm);
8938 mm->mm_address = env->me_metas[0]->mm_address;
8940 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8942 mp->mp_flags = P_META;
8943 *(MDB_meta *)METADATA(mp) = *mm;
8944 mm = (MDB_meta *)METADATA(mp);
8946 /* Count the number of free pages, subtract from lastpg to find
8947 * number of active pages
8950 MDB_ID freecount = 0;
8953 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8954 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8955 freecount += *(MDB_ID *)data.mv_data;
8956 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
8957 txn->mt_dbs[FREE_DBI].md_leaf_pages +
8958 txn->mt_dbs[FREE_DBI].md_overflow_pages;
8960 /* Set metapage 1 */
8961 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8962 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
8963 if (mm->mm_last_pg > NUM_METAS-1) {
8964 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
8967 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
8970 my.mc_wlen[0] = env->me_psize * NUM_METAS;
8972 pthread_mutex_lock(&my.mc_mutex);
8974 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8975 pthread_mutex_unlock(&my.mc_mutex);
8976 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
8977 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8978 rc = mdb_env_cthr_toggle(&my, 1);
8979 mdb_env_cthr_toggle(&my, -1);
8980 pthread_mutex_lock(&my.mc_mutex);
8982 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8983 pthread_mutex_unlock(&my.mc_mutex);
8988 CloseHandle(my.mc_cond);
8989 CloseHandle(my.mc_mutex);
8990 _aligned_free(my.mc_wbuf[0]);
8992 pthread_cond_destroy(&my.mc_cond);
8993 pthread_mutex_destroy(&my.mc_mutex);
8994 free(my.mc_wbuf[0]);
8999 /** Copy environment as-is. */
9001 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9003 MDB_txn *txn = NULL;
9004 mdb_mutexref_t wmutex = NULL;
9010 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9014 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9017 /* Do the lock/unlock of the reader mutex before starting the
9018 * write txn. Otherwise other read txns could block writers.
9020 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9025 /* We must start the actual read txn after blocking writers */
9026 mdb_txn_end(txn, MDB_END_RESET_TMP);
9028 /* Temporarily block writers until we snapshot the meta pages */
9029 wmutex = env->me_wmutex;
9030 if (LOCK_MUTEX(rc, env, wmutex))
9033 rc = mdb_txn_renew0(txn);
9035 UNLOCK_MUTEX(wmutex);
9040 wsize = env->me_psize * NUM_METAS;
9044 DO_WRITE(rc, fd, ptr, w2, len);
9048 } else if (len > 0) {
9054 /* Non-blocking or async handles are not supported */
9060 UNLOCK_MUTEX(wmutex);
9065 w2 = txn->mt_next_pgno * env->me_psize;
9068 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9075 if (wsize > MAX_WRITE)
9079 DO_WRITE(rc, fd, ptr, w2, len);
9083 } else if (len > 0) {
9100 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9102 if (flags & MDB_CP_COMPACT)
9103 return mdb_env_copyfd1(env, fd);
9105 return mdb_env_copyfd0(env, fd);
9109 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9111 return mdb_env_copyfd2(env, fd, 0);
9115 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9119 HANDLE newfd = INVALID_HANDLE_VALUE;
9121 if (env->me_flags & MDB_NOSUBDIR) {
9122 lpath = (char *)path;
9125 len += sizeof(DATANAME);
9126 lpath = malloc(len);
9129 sprintf(lpath, "%s" DATANAME, path);
9132 /* The destination path must exist, but the destination file must not.
9133 * We don't want the OS to cache the writes, since the source data is
9134 * already in the OS cache.
9137 newfd = CreateFileA(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9138 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9140 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9142 if (newfd == INVALID_HANDLE_VALUE) {
9147 if (env->me_psize >= env->me_os_psize) {
9149 /* Set O_DIRECT if the file system supports it */
9150 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9151 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9153 #ifdef F_NOCACHE /* __APPLE__ */
9154 rc = fcntl(newfd, F_NOCACHE, 1);
9162 rc = mdb_env_copyfd2(env, newfd, flags);
9165 if (!(env->me_flags & MDB_NOSUBDIR))
9167 if (newfd != INVALID_HANDLE_VALUE)
9168 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9175 mdb_env_copy(MDB_env *env, const char *path)
9177 return mdb_env_copy2(env, path, 0);
9181 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9183 if (flag & ~CHANGEABLE)
9186 env->me_flags |= flag;
9188 env->me_flags &= ~flag;
9193 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9198 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9203 mdb_env_set_userctx(MDB_env *env, void *ctx)
9207 env->me_userctx = ctx;
9212 mdb_env_get_userctx(MDB_env *env)
9214 return env ? env->me_userctx : NULL;
9218 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9223 env->me_assert_func = func;
9229 mdb_env_get_path(MDB_env *env, const char **arg)
9234 *arg = env->me_path;
9239 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9248 /** Common code for #mdb_stat() and #mdb_env_stat().
9249 * @param[in] env the environment to operate in.
9250 * @param[in] db the #MDB_db record containing the stats to return.
9251 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9252 * @return 0, this function always succeeds.
9255 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9257 arg->ms_psize = env->me_psize;
9258 arg->ms_depth = db->md_depth;
9259 arg->ms_branch_pages = db->md_branch_pages;
9260 arg->ms_leaf_pages = db->md_leaf_pages;
9261 arg->ms_overflow_pages = db->md_overflow_pages;
9262 arg->ms_entries = db->md_entries;
9268 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9272 if (env == NULL || arg == NULL)
9275 meta = mdb_env_pick_meta(env);
9277 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9281 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9285 if (env == NULL || arg == NULL)
9288 meta = mdb_env_pick_meta(env);
9289 arg->me_mapaddr = meta->mm_address;
9290 arg->me_last_pgno = meta->mm_last_pg;
9291 arg->me_last_txnid = meta->mm_txnid;
9293 arg->me_mapsize = env->me_mapsize;
9294 arg->me_maxreaders = env->me_maxreaders;
9295 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9299 /** Set the default comparison functions for a database.
9300 * Called immediately after a database is opened to set the defaults.
9301 * The user can then override them with #mdb_set_compare() or
9302 * #mdb_set_dupsort().
9303 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9304 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9307 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9309 uint16_t f = txn->mt_dbs[dbi].md_flags;
9311 txn->mt_dbxs[dbi].md_cmp =
9312 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9313 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9315 txn->mt_dbxs[dbi].md_dcmp =
9316 !(f & MDB_DUPSORT) ? 0 :
9317 ((f & MDB_INTEGERDUP)
9318 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9319 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9322 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9328 int rc, dbflag, exact;
9329 unsigned int unused = 0, seq;
9332 if (flags & ~VALID_FLAGS)
9334 if (txn->mt_flags & MDB_TXN_BLOCKED)
9340 if (flags & PERSISTENT_FLAGS) {
9341 uint16_t f2 = flags & PERSISTENT_FLAGS;
9342 /* make sure flag changes get committed */
9343 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9344 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9345 txn->mt_flags |= MDB_TXN_DIRTY;
9348 mdb_default_cmp(txn, MAIN_DBI);
9352 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9353 mdb_default_cmp(txn, MAIN_DBI);
9356 /* Is the DB already open? */
9358 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9359 if (!txn->mt_dbxs[i].md_name.mv_size) {
9360 /* Remember this free slot */
9361 if (!unused) unused = i;
9364 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9365 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9371 /* If no free slot and max hit, fail */
9372 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9373 return MDB_DBS_FULL;
9375 /* Cannot mix named databases with some mainDB flags */
9376 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9377 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9379 /* Find the DB info */
9380 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9383 key.mv_data = (void *)name;
9384 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9385 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9386 if (rc == MDB_SUCCESS) {
9387 /* make sure this is actually a DB */
9388 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9389 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9390 return MDB_INCOMPATIBLE;
9391 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9392 /* Create if requested */
9393 data.mv_size = sizeof(MDB_db);
9394 data.mv_data = &dummy;
9395 memset(&dummy, 0, sizeof(dummy));
9396 dummy.md_root = P_INVALID;
9397 dummy.md_flags = flags & PERSISTENT_FLAGS;
9398 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9402 /* OK, got info, add to table */
9403 if (rc == MDB_SUCCESS) {
9404 unsigned int slot = unused ? unused : txn->mt_numdbs;
9405 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9406 txn->mt_dbxs[slot].md_name.mv_size = len;
9407 txn->mt_dbxs[slot].md_rel = NULL;
9408 txn->mt_dbflags[slot] = dbflag;
9409 /* txn-> and env-> are the same in read txns, use
9410 * tmp variable to avoid undefined assignment
9412 seq = ++txn->mt_env->me_dbiseqs[slot];
9413 txn->mt_dbiseqs[slot] = seq;
9415 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9417 mdb_default_cmp(txn, slot);
9427 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9429 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9432 if (txn->mt_flags & MDB_TXN_BLOCKED)
9435 if (txn->mt_dbflags[dbi] & DB_STALE) {
9438 /* Stale, must read the DB's root. cursor_init does it for us. */
9439 mdb_cursor_init(&mc, txn, dbi, &mx);
9441 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9444 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9447 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9449 ptr = env->me_dbxs[dbi].md_name.mv_data;
9450 /* If there was no name, this was already closed */
9452 env->me_dbxs[dbi].md_name.mv_data = NULL;
9453 env->me_dbxs[dbi].md_name.mv_size = 0;
9454 env->me_dbflags[dbi] = 0;
9455 env->me_dbiseqs[dbi]++;
9460 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9462 /* We could return the flags for the FREE_DBI too but what's the point? */
9463 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9465 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9469 /** Add all the DB's pages to the free list.
9470 * @param[in] mc Cursor on the DB to free.
9471 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9472 * @return 0 on success, non-zero on failure.
9475 mdb_drop0(MDB_cursor *mc, int subs)
9479 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9480 if (rc == MDB_SUCCESS) {
9481 MDB_txn *txn = mc->mc_txn;
9486 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9487 * This also avoids any P_LEAF2 pages, which have no nodes.
9489 if (mc->mc_flags & C_SUB)
9492 mdb_cursor_copy(mc, &mx);
9493 while (mc->mc_snum > 0) {
9494 MDB_page *mp = mc->mc_pg[mc->mc_top];
9495 unsigned n = NUMKEYS(mp);
9497 for (i=0; i<n; i++) {
9498 ni = NODEPTR(mp, i);
9499 if (ni->mn_flags & F_BIGDATA) {
9502 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9503 rc = mdb_page_get(txn, pg, &omp, NULL);
9506 mdb_cassert(mc, IS_OVERFLOW(omp));
9507 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9511 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9512 mdb_xcursor_init1(mc, ni);
9513 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9519 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9521 for (i=0; i<n; i++) {
9523 ni = NODEPTR(mp, i);
9526 mdb_midl_xappend(txn->mt_free_pgs, pg);
9531 mc->mc_ki[mc->mc_top] = i;
9532 rc = mdb_cursor_sibling(mc, 1);
9534 if (rc != MDB_NOTFOUND)
9536 /* no more siblings, go back to beginning
9537 * of previous level.
9541 for (i=1; i<mc->mc_snum; i++) {
9543 mc->mc_pg[i] = mx.mc_pg[i];
9548 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9551 txn->mt_flags |= MDB_TXN_ERROR;
9552 } else if (rc == MDB_NOTFOUND) {
9558 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9560 MDB_cursor *mc, *m2;
9563 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9566 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9569 if (TXN_DBI_CHANGED(txn, dbi))
9572 rc = mdb_cursor_open(txn, dbi, &mc);
9576 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9577 /* Invalidate the dropped DB's cursors */
9578 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9579 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9583 /* Can't delete the main DB */
9584 if (del && dbi >= CORE_DBS) {
9585 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9587 txn->mt_dbflags[dbi] = DB_STALE;
9588 mdb_dbi_close(txn->mt_env, dbi);
9590 txn->mt_flags |= MDB_TXN_ERROR;
9593 /* reset the DB record, mark it dirty */
9594 txn->mt_dbflags[dbi] |= DB_DIRTY;
9595 txn->mt_dbs[dbi].md_depth = 0;
9596 txn->mt_dbs[dbi].md_branch_pages = 0;
9597 txn->mt_dbs[dbi].md_leaf_pages = 0;
9598 txn->mt_dbs[dbi].md_overflow_pages = 0;
9599 txn->mt_dbs[dbi].md_entries = 0;
9600 txn->mt_dbs[dbi].md_root = P_INVALID;
9602 txn->mt_flags |= MDB_TXN_DIRTY;
9605 mdb_cursor_close(mc);
9609 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9611 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9614 txn->mt_dbxs[dbi].md_cmp = cmp;
9618 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9620 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9623 txn->mt_dbxs[dbi].md_dcmp = cmp;
9627 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9629 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9632 txn->mt_dbxs[dbi].md_rel = rel;
9636 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9638 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9641 txn->mt_dbxs[dbi].md_relctx = ctx;
9646 mdb_env_get_maxkeysize(MDB_env *env)
9648 return ENV_MAXKEY(env);
9652 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9654 unsigned int i, rdrs;
9657 int rc = 0, first = 1;
9661 if (!env->me_txns) {
9662 return func("(no reader locks)\n", ctx);
9664 rdrs = env->me_txns->mti_numreaders;
9665 mr = env->me_txns->mti_readers;
9666 for (i=0; i<rdrs; i++) {
9668 txnid_t txnid = mr[i].mr_txnid;
9669 sprintf(buf, txnid == (txnid_t)-1 ?
9670 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9671 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9674 rc = func(" pid thread txnid\n", ctx);
9678 rc = func(buf, ctx);
9684 rc = func("(no active readers)\n", ctx);
9689 /** Insert pid into list if not already present.
9690 * return -1 if already present.
9693 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9695 /* binary search of pid in list */
9697 unsigned cursor = 1;
9699 unsigned n = ids[0];
9702 unsigned pivot = n >> 1;
9703 cursor = base + pivot + 1;
9704 val = pid - ids[cursor];
9709 } else if ( val > 0 ) {
9714 /* found, so it's a duplicate */
9723 for (n = ids[0]; n > cursor; n--)
9730 mdb_reader_check(MDB_env *env, int *dead)
9736 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9739 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9741 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9743 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9744 unsigned int i, j, rdrs;
9746 MDB_PID_T *pids, pid;
9747 int rc = MDB_SUCCESS, count = 0;
9749 rdrs = env->me_txns->mti_numreaders;
9750 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9754 mr = env->me_txns->mti_readers;
9755 for (i=0; i<rdrs; i++) {
9757 if (pid && pid != env->me_pid) {
9758 if (mdb_pid_insert(pids, pid) == 0) {
9759 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9760 /* Stale reader found */
9763 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9764 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9766 rdrs = 0; /* the above checked all readers */
9768 /* Recheck, a new process may have reused pid */
9769 if (mdb_reader_pid(env, Pidcheck, pid))
9774 if (mr[j].mr_pid == pid) {
9775 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9776 (unsigned) pid, mr[j].mr_txnid));
9781 UNLOCK_MUTEX(rmutex);
9792 #ifdef MDB_ROBUST_SUPPORTED
9793 /** Handle #LOCK_MUTEX0() failure.
9794 * Try to repair the lock file if the mutex owner died.
9795 * @param[in] env the environment handle
9796 * @param[in] mutex LOCK_MUTEX0() mutex
9797 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9798 * @return 0 on success with the mutex locked, or an error code on failure.
9801 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9806 if (rc == MDB_OWNERDEAD) {
9807 /* We own the mutex. Clean up after dead previous owner. */
9809 rlocked = (mutex == env->me_rmutex);
9811 /* Keep mti_txnid updated, otherwise next writer can
9812 * overwrite data which latest meta page refers to.
9814 meta = mdb_env_pick_meta(env);
9815 env->me_txns->mti_txnid = meta->mm_txnid;
9816 /* env is hosed if the dead thread was ours */
9818 env->me_flags |= MDB_FATAL_ERROR;
9823 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9824 (rc ? "this process' env is hosed" : "recovering")));
9825 rc2 = mdb_reader_check0(env, rlocked, NULL);
9827 rc2 = mdb_mutex_consistent(mutex);
9828 if (rc || (rc = rc2)) {
9829 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9830 UNLOCK_MUTEX(mutex);
9836 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9841 #endif /* MDB_ROBUST_SUPPORTED */
projects / platform / upstream / lmdb.git / blob