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
101 #if defined(__sun) || defined(ANDROID)
102 /* Most platforms have posix_memalign, older may only have memalign */
103 #define HAVE_MEMALIGN 1
107 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
108 #include <netinet/in.h>
109 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
112 #if defined(__APPLE__) || defined (BSD)
113 # if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
114 # define MDB_USE_SYSV_SEM 1
116 # define MDB_FDATASYNC fsync
117 #elif defined(ANDROID)
118 # define MDB_FDATASYNC fsync
123 #ifdef MDB_USE_POSIX_SEM
124 # define MDB_USE_HASH 1
125 #include <semaphore.h>
126 #elif defined(MDB_USE_SYSV_SEM)
129 #ifdef _SEM_SEMUN_UNDEFINED
132 struct semid_ds *buf;
133 unsigned short *array;
135 #endif /* _SEM_SEMUN_UNDEFINED */
137 #define MDB_USE_POSIX_MUTEX 1
138 #endif /* MDB_USE_POSIX_SEM */
141 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
142 + defined(MDB_USE_POSIX_MUTEX) != 1
143 # error "Ambiguous shared-lock implementation"
147 #include <valgrind/memcheck.h>
148 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
149 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
150 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
151 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
152 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
154 #define VGMEMP_CREATE(h,r,z)
155 #define VGMEMP_ALLOC(h,a,s)
156 #define VGMEMP_FREE(h,a)
157 #define VGMEMP_DESTROY(h)
158 #define VGMEMP_DEFINED(a,s)
162 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
163 /* Solaris just defines one or the other */
164 # define LITTLE_ENDIAN 1234
165 # define BIG_ENDIAN 4321
166 # ifdef _LITTLE_ENDIAN
167 # define BYTE_ORDER LITTLE_ENDIAN
169 # define BYTE_ORDER BIG_ENDIAN
172 # define BYTE_ORDER __BYTE_ORDER
176 #ifndef LITTLE_ENDIAN
177 #define LITTLE_ENDIAN __LITTLE_ENDIAN
180 #define BIG_ENDIAN __BIG_ENDIAN
183 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
184 #define MISALIGNED_OK 1
190 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
191 # error "Unknown or unsupported endianness (BYTE_ORDER)"
192 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
193 # error "Two's complement, reasonably sized integer types, please"
197 /** Put infrequently used env functions in separate section */
199 # define ESECT __attribute__ ((section("__TEXT,text_env")))
201 # define ESECT __attribute__ ((section("text_env")))
207 /** @defgroup internal LMDB Internals
210 /** @defgroup compat Compatibility Macros
211 * A bunch of macros to minimize the amount of platform-specific ifdefs
212 * needed throughout the rest of the code. When the features this library
213 * needs are similar enough to POSIX to be hidden in a one-or-two line
214 * replacement, this macro approach is used.
218 /** Features under development */
223 /** Wrapper around __func__, which is a C99 feature */
224 #if __STDC_VERSION__ >= 199901L
225 # define mdb_func_ __func__
226 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
227 # define mdb_func_ __FUNCTION__
229 /* If a debug message says <mdb_unknown>(), update the #if statements above */
230 # define mdb_func_ "<mdb_unknown>"
233 /* Internal error codes, not exposed outside liblmdb */
234 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
236 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
237 #elif defined MDB_USE_SYSV_SEM
238 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
239 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
240 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
244 #define MDB_ROBUST_SUPPORTED 1
248 #define MDB_USE_HASH 1
249 #define MDB_PIDLOCK 0
250 #define THREAD_RET DWORD
251 #define pthread_t HANDLE
252 #define pthread_mutex_t HANDLE
253 #define pthread_cond_t HANDLE
254 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
255 #define pthread_key_t DWORD
256 #define pthread_self() GetCurrentThreadId()
257 #define pthread_key_create(x,y) \
258 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
259 #define pthread_key_delete(x) TlsFree(x)
260 #define pthread_getspecific(x) TlsGetValue(x)
261 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
262 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
263 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
264 #define pthread_cond_signal(x) SetEvent(*x)
265 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
266 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
267 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
268 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
269 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
270 #define mdb_mutex_consistent(mutex) 0
271 #define getpid() GetCurrentProcessId()
272 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
273 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
274 #define ErrCode() GetLastError()
275 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
276 #define close(fd) (CloseHandle(fd) ? 0 : -1)
277 #define munmap(ptr,len) UnmapViewOfFile(ptr)
278 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
279 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
281 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
285 #define THREAD_RET void *
286 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
287 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
288 #define Z "z" /**< printf format modifier for size_t */
290 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
291 #define MDB_PIDLOCK 1
293 #ifdef MDB_USE_POSIX_SEM
295 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
296 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
297 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
300 mdb_sem_wait(sem_t *sem)
303 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
307 #elif defined MDB_USE_SYSV_SEM
309 typedef struct mdb_mutex {
313 } mdb_mutex_t[1], *mdb_mutexref_t;
315 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
316 #define UNLOCK_MUTEX(mutex) do { \
317 struct sembuf sb = { 0, 1, SEM_UNDO }; \
318 sb.sem_num = (mutex)->semnum; \
319 *(mutex)->locked = 0; \
320 semop((mutex)->semid, &sb, 1); \
324 mdb_sem_wait(mdb_mutexref_t sem)
326 int rc, *locked = sem->locked;
327 struct sembuf sb = { 0, -1, SEM_UNDO };
328 sb.sem_num = sem->semnum;
330 if (!semop(sem->semid, &sb, 1)) {
331 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
335 } while ((rc = errno) == EINTR);
339 #define mdb_mutex_consistent(mutex) 0
341 #else /* MDB_USE_POSIX_MUTEX: */
342 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
343 * local variables keep it (mdb_mutexref_t).
345 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
346 * be the same, or an array[size 1] and a pointer.
349 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
351 /** Lock the reader or writer mutex.
352 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
354 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
355 /** Unlock the reader or writer mutex.
357 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
358 /** Mark mutex-protected data as repaired, after death of previous owner.
360 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
361 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
363 /** Get the error code for the last failed system function.
365 #define ErrCode() errno
367 /** An abstraction for a file handle.
368 * On POSIX systems file handles are small integers. On Windows
369 * they're opaque pointers.
373 /** A value for an invalid file handle.
374 * Mainly used to initialize file variables and signify that they are
377 #define INVALID_HANDLE_VALUE (-1)
379 /** Get the size of a memory page for the system.
380 * This is the basic size that the platform's memory manager uses, and is
381 * fundamental to the use of memory-mapped files.
383 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
386 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
388 #elif defined(MDB_USE_SYSV_SEM)
389 #define MNAME_LEN (sizeof(int))
391 #define MNAME_LEN (sizeof(pthread_mutex_t))
394 #ifdef MDB_USE_SYSV_SEM
395 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
397 #define SYSV_SEM_FLAG 0
402 #ifdef MDB_ROBUST_SUPPORTED
403 /** Lock mutex, handle any error, set rc = result.
404 * Return 0 on success, nonzero (not rc) on error.
406 #define LOCK_MUTEX(rc, env, mutex) \
407 (((rc) = LOCK_MUTEX0(mutex)) && \
408 ((rc) = mdb_mutex_failed(env, mutex, rc)))
409 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
411 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
412 #define mdb_mutex_failed(env, mutex, rc) (rc)
416 /** A flag for opening a file and requesting synchronous data writes.
417 * This is only used when writing a meta page. It's not strictly needed;
418 * we could just do a normal write and then immediately perform a flush.
419 * But if this flag is available it saves us an extra system call.
421 * @note If O_DSYNC is undefined but exists in /usr/include,
422 * preferably set some compiler flag to get the definition.
423 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
426 # define MDB_DSYNC O_DSYNC
430 /** Function for flushing the data of a file. Define this to fsync
431 * if fdatasync() is not supported.
433 #ifndef MDB_FDATASYNC
434 # define MDB_FDATASYNC fdatasync
438 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
449 /** A page number in the database.
450 * Note that 64 bit page numbers are overkill, since pages themselves
451 * already represent 12-13 bits of addressable memory, and the OS will
452 * always limit applications to a maximum of 63 bits of address space.
454 * @note In the #MDB_node structure, we only store 48 bits of this value,
455 * which thus limits us to only 60 bits of addressable data.
457 typedef MDB_ID pgno_t;
459 /** A transaction ID.
460 * See struct MDB_txn.mt_txnid for details.
462 typedef MDB_ID txnid_t;
464 /** @defgroup debug Debug Macros
468 /** Enable debug output. Needs variable argument macros (a C99 feature).
469 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
470 * read from and written to the database (used for free space management).
476 static int mdb_debug;
477 static txnid_t mdb_debug_start;
479 /** Print a debug message with printf formatting.
480 * Requires double parenthesis around 2 or more args.
482 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
483 # define DPRINTF0(fmt, ...) \
484 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
486 # define DPRINTF(args) ((void) 0)
488 /** Print a debug string.
489 * The string is printed literally, with no format processing.
491 #define DPUTS(arg) DPRINTF(("%s", arg))
492 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
494 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
497 /** @brief The maximum size of a database page.
499 * It is 32k or 64k, since value-PAGEBASE must fit in
500 * #MDB_page.%mp_upper.
502 * LMDB will use database pages < OS pages if needed.
503 * That causes more I/O in write transactions: The OS must
504 * know (read) the whole page before writing a partial page.
506 * Note that we don't currently support Huge pages. On Linux,
507 * regular data files cannot use Huge pages, and in general
508 * Huge pages aren't actually pageable. We rely on the OS
509 * demand-pager to read our data and page it out when memory
510 * pressure from other processes is high. So until OSs have
511 * actual paging support for Huge pages, they're not viable.
513 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
515 /** The minimum number of keys required in a database page.
516 * Setting this to a larger value will place a smaller bound on the
517 * maximum size of a data item. Data items larger than this size will
518 * be pushed into overflow pages instead of being stored directly in
519 * the B-tree node. This value used to default to 4. With a page size
520 * of 4096 bytes that meant that any item larger than 1024 bytes would
521 * go into an overflow page. That also meant that on average 2-3KB of
522 * each overflow page was wasted space. The value cannot be lower than
523 * 2 because then there would no longer be a tree structure. With this
524 * value, items larger than 2KB will go into overflow pages, and on
525 * average only 1KB will be wasted.
527 #define MDB_MINKEYS 2
529 /** A stamp that identifies a file as an LMDB file.
530 * There's nothing special about this value other than that it is easily
531 * recognizable, and it will reflect any byte order mismatches.
533 #define MDB_MAGIC 0xBEEFC0DE
535 /** The version number for a database's datafile format. */
536 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
537 /** The version number for a database's lockfile format. */
538 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
540 /** @brief The max size of a key we can write, or 0 for computed max.
542 * This macro should normally be left alone or set to 0.
543 * Note that a database with big keys or dupsort data cannot be
544 * reliably modified by a liblmdb which uses a smaller max.
545 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
547 * Other values are allowed, for backwards compat. However:
548 * A value bigger than the computed max can break if you do not
549 * know what you are doing, and liblmdb <= 0.9.10 can break when
550 * modifying a DB with keys/dupsort data bigger than its max.
552 * Data items in an #MDB_DUPSORT database are also limited to
553 * this size, since they're actually keys of a sub-DB. Keys and
554 * #MDB_DUPSORT data items must fit on a node in a regular page.
556 #ifndef MDB_MAXKEYSIZE
557 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
560 /** The maximum size of a key we can write to the environment. */
562 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
564 #define ENV_MAXKEY(env) ((env)->me_maxkey)
567 /** @brief The maximum size of a data item.
569 * We only store a 32 bit value for node sizes.
571 #define MAXDATASIZE 0xffffffffUL
574 /** Key size which fits in a #DKBUF.
577 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
580 * This is used for printing a hex dump of a key's contents.
582 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
583 /** Display a key in hex.
585 * Invoke a function to display a key in hex.
587 #define DKEY(x) mdb_dkey(x, kbuf)
593 /** An invalid page number.
594 * Mainly used to denote an empty tree.
596 #define P_INVALID (~(pgno_t)0)
598 /** Test if the flags \b f are set in a flag word \b w. */
599 #define F_ISSET(w, f) (((w) & (f)) == (f))
601 /** Round \b n up to an even number. */
602 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
604 /** Used for offsets within a single page.
605 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
608 typedef uint16_t indx_t;
610 /** Default size of memory map.
611 * This is certainly too small for any actual applications. Apps should always set
612 * the size explicitly using #mdb_env_set_mapsize().
614 #define DEFAULT_MAPSIZE 1048576
616 /** @defgroup readers Reader Lock Table
617 * Readers don't acquire any locks for their data access. Instead, they
618 * simply record their transaction ID in the reader table. The reader
619 * mutex is needed just to find an empty slot in the reader table. The
620 * slot's address is saved in thread-specific data so that subsequent read
621 * transactions started by the same thread need no further locking to proceed.
623 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
625 * No reader table is used if the database is on a read-only filesystem, or
626 * if #MDB_NOLOCK is set.
628 * Since the database uses multi-version concurrency control, readers don't
629 * actually need any locking. This table is used to keep track of which
630 * readers are using data from which old transactions, so that we'll know
631 * when a particular old transaction is no longer in use. Old transactions
632 * that have discarded any data pages can then have those pages reclaimed
633 * for use by a later write transaction.
635 * The lock table is constructed such that reader slots are aligned with the
636 * processor's cache line size. Any slot is only ever used by one thread.
637 * This alignment guarantees that there will be no contention or cache
638 * thrashing as threads update their own slot info, and also eliminates
639 * any need for locking when accessing a slot.
641 * A writer thread will scan every slot in the table to determine the oldest
642 * outstanding reader transaction. Any freed pages older than this will be
643 * reclaimed by the writer. The writer doesn't use any locks when scanning
644 * this table. This means that there's no guarantee that the writer will
645 * see the most up-to-date reader info, but that's not required for correct
646 * operation - all we need is to know the upper bound on the oldest reader,
647 * we don't care at all about the newest reader. So the only consequence of
648 * reading stale information here is that old pages might hang around a
649 * while longer before being reclaimed. That's actually good anyway, because
650 * the longer we delay reclaiming old pages, the more likely it is that a
651 * string of contiguous pages can be found after coalescing old pages from
652 * many old transactions together.
655 /** Number of slots in the reader table.
656 * This value was chosen somewhat arbitrarily. 126 readers plus a
657 * couple mutexes fit exactly into 8KB on my development machine.
658 * Applications should set the table size using #mdb_env_set_maxreaders().
660 #define DEFAULT_READERS 126
662 /** The size of a CPU cache line in bytes. We want our lock structures
663 * aligned to this size to avoid false cache line sharing in the
665 * This value works for most CPUs. For Itanium this should be 128.
671 /** The information we store in a single slot of the reader table.
672 * In addition to a transaction ID, we also record the process and
673 * thread ID that owns a slot, so that we can detect stale information,
674 * e.g. threads or processes that went away without cleaning up.
675 * @note We currently don't check for stale records. We simply re-init
676 * the table when we know that we're the only process opening the
679 typedef struct MDB_rxbody {
680 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
681 * Multiple readers that start at the same time will probably have the
682 * same ID here. Again, it's not important to exclude them from
683 * anything; all we need to know is which version of the DB they
684 * started from so we can avoid overwriting any data used in that
685 * particular version.
687 volatile txnid_t mrb_txnid;
688 /** The process ID of the process owning this reader txn. */
689 volatile MDB_PID_T mrb_pid;
690 /** The thread ID of the thread owning this txn. */
691 volatile MDB_THR_T mrb_tid;
694 /** The actual reader record, with cacheline padding. */
695 typedef struct MDB_reader {
698 /** shorthand for mrb_txnid */
699 #define mr_txnid mru.mrx.mrb_txnid
700 #define mr_pid mru.mrx.mrb_pid
701 #define mr_tid mru.mrx.mrb_tid
702 /** cache line alignment */
703 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
707 /** The header for the reader table.
708 * The table resides in a memory-mapped file. (This is a different file
709 * than is used for the main database.)
711 * For POSIX the actual mutexes reside in the shared memory of this
712 * mapped file. On Windows, mutexes are named objects allocated by the
713 * kernel; we store the mutex names in this mapped file so that other
714 * processes can grab them. This same approach is also used on
715 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
716 * process-shared POSIX mutexes. For these cases where a named object
717 * is used, the object name is derived from a 64 bit FNV hash of the
718 * environment pathname. As such, naming collisions are extremely
719 * unlikely. If a collision occurs, the results are unpredictable.
721 typedef struct MDB_txbody {
722 /** Stamp identifying this as an LMDB file. It must be set
725 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
727 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
728 char mtb_rmname[MNAME_LEN];
729 #elif defined(MDB_USE_SYSV_SEM)
733 /** Mutex protecting access to this table.
734 * This is the reader table lock used with LOCK_MUTEX().
736 mdb_mutex_t mtb_rmutex;
738 /** The ID of the last transaction committed to the database.
739 * This is recorded here only for convenience; the value can always
740 * be determined by reading the main database meta pages.
742 volatile txnid_t mtb_txnid;
743 /** The number of slots that have been used in the reader table.
744 * This always records the maximum count, it is not decremented
745 * when readers release their slots.
747 volatile unsigned mtb_numreaders;
750 /** The actual reader table definition. */
751 typedef struct MDB_txninfo {
754 #define mti_magic mt1.mtb.mtb_magic
755 #define mti_format mt1.mtb.mtb_format
756 #define mti_rmutex mt1.mtb.mtb_rmutex
757 #define mti_rmname mt1.mtb.mtb_rmname
758 #define mti_txnid mt1.mtb.mtb_txnid
759 #define mti_numreaders mt1.mtb.mtb_numreaders
760 #ifdef MDB_USE_SYSV_SEM
761 #define mti_semid mt1.mtb.mtb_semid
762 #define mti_rlocked mt1.mtb.mtb_rlocked
764 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
767 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
768 char mt2_wmname[MNAME_LEN];
769 #define mti_wmname mt2.mt2_wmname
770 #elif defined MDB_USE_SYSV_SEM
772 #define mti_wlocked mt2.mt2_wlocked
774 mdb_mutex_t mt2_wmutex;
775 #define mti_wmutex mt2.mt2_wmutex
777 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
779 MDB_reader mti_readers[1];
782 /** Lockfile format signature: version, features and field layout */
783 #define MDB_LOCK_FORMAT \
785 ((MDB_LOCK_VERSION) \
786 /* Flags which describe functionality */ \
787 + (SYSV_SEM_FLAG << 18) \
788 + (((MDB_PIDLOCK) != 0) << 16)))
791 /** Common header for all page types.
792 * Overflow records occupy a number of contiguous pages with no
793 * headers on any page after the first.
795 typedef struct MDB_page {
796 #define mp_pgno mp_p.p_pgno
797 #define mp_next mp_p.p_next
799 pgno_t p_pgno; /**< page number */
800 struct MDB_page *p_next; /**< for in-memory list of freed pages */
803 /** @defgroup mdb_page Page Flags
805 * Flags for the page headers.
808 #define P_BRANCH 0x01 /**< branch page */
809 #define P_LEAF 0x02 /**< leaf page */
810 #define P_OVERFLOW 0x04 /**< overflow page */
811 #define P_META 0x08 /**< meta page */
812 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
813 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
814 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
815 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
816 #define P_KEEP 0x8000 /**< leave this page alone during spill */
818 uint16_t mp_flags; /**< @ref mdb_page */
819 #define mp_lower mp_pb.pb.pb_lower
820 #define mp_upper mp_pb.pb.pb_upper
821 #define mp_pages mp_pb.pb_pages
824 indx_t pb_lower; /**< lower bound of free space */
825 indx_t pb_upper; /**< upper bound of free space */
827 uint32_t pb_pages; /**< number of overflow pages */
829 indx_t mp_ptrs[1]; /**< dynamic size */
832 /** Size of the page header, excluding dynamic data at the end */
833 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
835 /** Address of first usable data byte in a page, after the header */
836 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
838 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
839 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
841 /** Number of nodes on a page */
842 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
844 /** The amount of space remaining in the page */
845 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
847 /** The percentage of space used in the page, in tenths of a percent. */
848 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
849 ((env)->me_psize - PAGEHDRSZ))
850 /** The minimum page fill factor, in tenths of a percent.
851 * Pages emptier than this are candidates for merging.
853 #define FILL_THRESHOLD 250
855 /** Test if a page is a leaf page */
856 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
857 /** Test if a page is a LEAF2 page */
858 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
859 /** Test if a page is a branch page */
860 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
861 /** Test if a page is an overflow page */
862 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
863 /** Test if a page is a sub page */
864 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
866 /** The number of overflow pages needed to store the given size. */
867 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
869 /** Link in #MDB_txn.%mt_loose_pgs list */
870 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
872 /** Header for a single key/data pair within a page.
873 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
874 * We guarantee 2-byte alignment for 'MDB_node's.
876 typedef struct MDB_node {
877 /** lo and hi are used for data size on leaf nodes and for
878 * child pgno on branch nodes. On 64 bit platforms, flags
879 * is also used for pgno. (Branch nodes have no flags).
880 * They are in host byte order in case that lets some
881 * accesses be optimized into a 32-bit word access.
883 #if BYTE_ORDER == LITTLE_ENDIAN
884 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
886 unsigned short mn_hi, mn_lo;
888 /** @defgroup mdb_node Node Flags
890 * Flags for node headers.
893 #define F_BIGDATA 0x01 /**< data put on overflow page */
894 #define F_SUBDATA 0x02 /**< data is a sub-database */
895 #define F_DUPDATA 0x04 /**< data has duplicates */
897 /** valid flags for #mdb_node_add() */
898 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
901 unsigned short mn_flags; /**< @ref mdb_node */
902 unsigned short mn_ksize; /**< key size */
903 char mn_data[1]; /**< key and data are appended here */
906 /** Size of the node header, excluding dynamic data at the end */
907 #define NODESIZE offsetof(MDB_node, mn_data)
909 /** Bit position of top word in page number, for shifting mn_flags */
910 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
912 /** Size of a node in a branch page with a given key.
913 * This is just the node header plus the key, there is no data.
915 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
917 /** Size of a node in a leaf page with a given key and data.
918 * This is node header plus key plus data size.
920 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
922 /** Address of node \b i in page \b p */
923 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
925 /** Address of the key for the node */
926 #define NODEKEY(node) (void *)((node)->mn_data)
928 /** Address of the data for a node */
929 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
931 /** Get the page number pointed to by a branch node */
932 #define NODEPGNO(node) \
933 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
934 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
935 /** Set the page number in a branch node */
936 #define SETPGNO(node,pgno) do { \
937 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
938 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
940 /** Get the size of the data in a leaf node */
941 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
942 /** Set the size of the data for a leaf node */
943 #define SETDSZ(node,size) do { \
944 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
945 /** The size of a key in a node */
946 #define NODEKSZ(node) ((node)->mn_ksize)
948 /** Copy a page number from src to dst */
950 #define COPY_PGNO(dst,src) dst = src
952 #if SIZE_MAX > 4294967295UL
953 #define COPY_PGNO(dst,src) do { \
954 unsigned short *s, *d; \
955 s = (unsigned short *)&(src); \
956 d = (unsigned short *)&(dst); \
963 #define COPY_PGNO(dst,src) do { \
964 unsigned short *s, *d; \
965 s = (unsigned short *)&(src); \
966 d = (unsigned short *)&(dst); \
972 /** The address of a key in a LEAF2 page.
973 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
974 * There are no node headers, keys are stored contiguously.
976 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
978 /** Set the \b node's key into \b keyptr, if requested. */
979 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
980 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
982 /** Set the \b node's key into \b key. */
983 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
985 /** Information about a single database in the environment. */
986 typedef struct MDB_db {
987 uint32_t md_pad; /**< also ksize for LEAF2 pages */
988 uint16_t md_flags; /**< @ref mdb_dbi_open */
989 uint16_t md_depth; /**< depth of this tree */
990 pgno_t md_branch_pages; /**< number of internal pages */
991 pgno_t md_leaf_pages; /**< number of leaf pages */
992 pgno_t md_overflow_pages; /**< number of overflow pages */
993 size_t md_entries; /**< number of data items */
994 pgno_t md_root; /**< the root page of this tree */
997 /** mdb_dbi_open flags */
998 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
999 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1000 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1001 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1003 /** Handle for the DB used to track free pages. */
1005 /** Handle for the default DB. */
1008 /** Meta page content.
1009 * A meta page is the start point for accessing a database snapshot.
1010 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1012 typedef struct MDB_meta {
1013 /** Stamp identifying this as an LMDB file. It must be set
1016 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1017 uint32_t mm_version;
1018 void *mm_address; /**< address for fixed mapping */
1019 size_t mm_mapsize; /**< size of mmap region */
1020 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
1021 /** The size of pages used in this DB */
1022 #define mm_psize mm_dbs[0].md_pad
1023 /** Any persistent environment flags. @ref mdb_env */
1024 #define mm_flags mm_dbs[0].md_flags
1025 pgno_t mm_last_pg; /**< last used page in file */
1026 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1029 /** Buffer for a stack-allocated meta page.
1030 * The members define size and alignment, and silence type
1031 * aliasing warnings. They are not used directly; that could
1032 * mean incorrectly using several union members in parallel.
1034 typedef union MDB_metabuf {
1037 char mm_pad[PAGEHDRSZ];
1042 /** Auxiliary DB info.
1043 * The information here is mostly static/read-only. There is
1044 * only a single copy of this record in the environment.
1046 typedef struct MDB_dbx {
1047 MDB_val md_name; /**< name of the database */
1048 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1049 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1050 MDB_rel_func *md_rel; /**< user relocate function */
1051 void *md_relctx; /**< user-provided context for md_rel */
1054 /** A database transaction.
1055 * Every operation requires a transaction handle.
1058 MDB_txn *mt_parent; /**< parent of a nested txn */
1059 MDB_txn *mt_child; /**< nested txn under this txn */
1060 pgno_t mt_next_pgno; /**< next unallocated page */
1061 /** The ID of this transaction. IDs are integers incrementing from 1.
1062 * Only committed write transactions increment the ID. If a transaction
1063 * aborts, the ID may be re-used by the next writer.
1066 MDB_env *mt_env; /**< the DB environment */
1067 /** The list of pages that became unused during this transaction.
1069 MDB_IDL mt_free_pgs;
1070 /** The list of loose pages that became unused and may be reused
1071 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1073 MDB_page *mt_loose_pgs;
1074 /* #Number of loose pages (#mt_loose_pgs) */
1076 /** The sorted list of dirty pages we temporarily wrote to disk
1077 * because the dirty list was full. page numbers in here are
1078 * shifted left by 1, deleted slots have the LSB set.
1080 MDB_IDL mt_spill_pgs;
1082 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1083 MDB_ID2L dirty_list;
1084 /** For read txns: This thread/txn's reader table slot, or NULL. */
1087 /** Array of records for each DB known in the environment. */
1089 /** Array of MDB_db records for each known DB */
1091 /** Array of sequence numbers for each DB handle */
1092 unsigned int *mt_dbiseqs;
1093 /** @defgroup mt_dbflag Transaction DB Flags
1097 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1098 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1099 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1100 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1102 /** In write txns, array of cursors for each DB */
1103 MDB_cursor **mt_cursors;
1104 /** Array of flags for each DB */
1105 unsigned char *mt_dbflags;
1106 /** Number of DB records in use. This number only ever increments;
1107 * we don't decrement it when individual DB handles are closed.
1111 /** @defgroup mdb_txn Transaction Flags
1115 /** #mdb_txn_begin() flags */
1116 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1117 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1118 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1119 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1120 /* internal txn flags */
1121 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1122 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1123 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1124 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1126 unsigned int mt_flags; /**< @ref mdb_txn */
1127 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1128 * Includes ancestor txns' dirty pages not hidden by other txns'
1129 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1130 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1132 unsigned int mt_dirty_room;
1135 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1136 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1137 * raise this on a 64 bit machine.
1139 #define CURSOR_STACK 32
1143 /** Cursors are used for all DB operations.
1144 * A cursor holds a path of (page pointer, key index) from the DB
1145 * root to a position in the DB, plus other state. #MDB_DUPSORT
1146 * cursors include an xcursor to the current data item. Write txns
1147 * track their cursors and keep them up to date when data moves.
1148 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1149 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1152 /** Next cursor on this DB in this txn */
1153 MDB_cursor *mc_next;
1154 /** Backup of the original cursor if this cursor is a shadow */
1155 MDB_cursor *mc_backup;
1156 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1157 struct MDB_xcursor *mc_xcursor;
1158 /** The transaction that owns this cursor */
1160 /** The database handle this cursor operates on */
1162 /** The database record for this cursor */
1164 /** The database auxiliary record for this cursor */
1166 /** The @ref mt_dbflag for this database */
1167 unsigned char *mc_dbflag;
1168 unsigned short mc_snum; /**< number of pushed pages */
1169 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1170 /** @defgroup mdb_cursor Cursor Flags
1172 * Cursor state flags.
1175 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1176 #define C_EOF 0x02 /**< No more data */
1177 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1178 #define C_DEL 0x08 /**< last op was a cursor_del */
1179 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1180 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1182 unsigned int mc_flags; /**< @ref mdb_cursor */
1183 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1184 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1187 /** Context for sorted-dup records.
1188 * We could have gone to a fully recursive design, with arbitrarily
1189 * deep nesting of sub-databases. But for now we only handle these
1190 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1192 typedef struct MDB_xcursor {
1193 /** A sub-cursor for traversing the Dup DB */
1194 MDB_cursor mx_cursor;
1195 /** The database record for this Dup DB */
1197 /** The auxiliary DB record for this Dup DB */
1199 /** The @ref mt_dbflag for this Dup DB */
1200 unsigned char mx_dbflag;
1203 /** State of FreeDB old pages, stored in the MDB_env */
1204 typedef struct MDB_pgstate {
1205 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1206 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1209 /** The database environment. */
1211 HANDLE me_fd; /**< The main data file */
1212 HANDLE me_lfd; /**< The lock file */
1213 HANDLE me_mfd; /**< just for writing the meta pages */
1214 /** Failed to update the meta page. Probably an I/O error. */
1215 #define MDB_FATAL_ERROR 0x80000000U
1216 /** Some fields are initialized. */
1217 #define MDB_ENV_ACTIVE 0x20000000U
1218 /** me_txkey is set */
1219 #define MDB_ENV_TXKEY 0x10000000U
1220 /** fdatasync is unreliable */
1221 #define MDB_FSYNCONLY 0x08000000U
1222 uint32_t me_flags; /**< @ref mdb_env */
1223 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1224 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1225 unsigned int me_maxreaders; /**< size of the reader table */
1226 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1227 volatile int me_close_readers;
1228 MDB_dbi me_numdbs; /**< number of DBs opened */
1229 MDB_dbi me_maxdbs; /**< size of the DB table */
1230 MDB_PID_T me_pid; /**< process ID of this env */
1231 char *me_path; /**< path to the DB files */
1232 char *me_map; /**< the memory map of the data file */
1233 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1234 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1235 void *me_pbuf; /**< scratch area for DUPSORT put() */
1236 MDB_txn *me_txn; /**< current write transaction */
1237 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1238 size_t me_mapsize; /**< size of the data memory map */
1239 off_t me_size; /**< current file size */
1240 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1241 MDB_dbx *me_dbxs; /**< array of static DB info */
1242 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1243 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1244 pthread_key_t me_txkey; /**< thread-key for readers */
1245 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1246 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1247 # define me_pglast me_pgstate.mf_pglast
1248 # define me_pghead me_pgstate.mf_pghead
1249 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1250 /** IDL of pages that became unused in a write txn */
1251 MDB_IDL me_free_pgs;
1252 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1253 MDB_ID2L me_dirty_list;
1254 /** Max number of freelist items that can fit in a single overflow page */
1256 /** Max size of a node on a page */
1257 unsigned int me_nodemax;
1258 #if !(MDB_MAXKEYSIZE)
1259 unsigned int me_maxkey; /**< max size of a key */
1261 int me_live_reader; /**< have liveness lock in reader table */
1263 int me_pidquery; /**< Used in OpenProcess */
1265 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1266 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1267 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1269 mdb_mutex_t me_rmutex;
1270 mdb_mutex_t me_wmutex;
1272 void *me_userctx; /**< User-settable context */
1273 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1276 /** Nested transaction */
1277 typedef struct MDB_ntxn {
1278 MDB_txn mnt_txn; /**< the transaction */
1279 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1282 /** max number of pages to commit in one writev() call */
1283 #define MDB_COMMIT_PAGES 64
1284 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1285 #undef MDB_COMMIT_PAGES
1286 #define MDB_COMMIT_PAGES IOV_MAX
1289 /** max bytes to write in one call */
1290 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1292 /** Check \b txn and \b dbi arguments to a function */
1293 #define TXN_DBI_EXIST(txn, dbi) \
1294 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1296 /** Check for misused \b dbi handles */
1297 #define TXN_DBI_CHANGED(txn, dbi) \
1298 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1300 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1301 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1302 static int mdb_page_touch(MDB_cursor *mc);
1304 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1305 static int mdb_page_search_root(MDB_cursor *mc,
1306 MDB_val *key, int modify);
1307 #define MDB_PS_MODIFY 1
1308 #define MDB_PS_ROOTONLY 2
1309 #define MDB_PS_FIRST 4
1310 #define MDB_PS_LAST 8
1311 static int mdb_page_search(MDB_cursor *mc,
1312 MDB_val *key, int flags);
1313 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1315 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1316 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1317 pgno_t newpgno, unsigned int nflags);
1319 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1320 static int mdb_env_pick_meta(const MDB_env *env);
1321 static int mdb_env_write_meta(MDB_txn *txn);
1322 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1323 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1325 static void mdb_env_close0(MDB_env *env, int excl);
1327 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1328 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1329 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1330 static void mdb_node_del(MDB_cursor *mc, int ksize);
1331 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1332 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1333 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1334 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1335 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1337 static int mdb_rebalance(MDB_cursor *mc);
1338 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1340 static void mdb_cursor_pop(MDB_cursor *mc);
1341 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1343 static int mdb_cursor_del0(MDB_cursor *mc);
1344 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1345 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1346 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1347 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1348 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1350 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1351 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1353 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1354 static void mdb_xcursor_init0(MDB_cursor *mc);
1355 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1357 static int mdb_drop0(MDB_cursor *mc, int subs);
1358 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1359 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1362 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1365 /** Compare two items pointing at size_t's of unknown alignment. */
1366 #ifdef MISALIGNED_OK
1367 # define mdb_cmp_clong mdb_cmp_long
1369 # define mdb_cmp_clong mdb_cmp_cint
1373 static SECURITY_DESCRIPTOR mdb_null_sd;
1374 static SECURITY_ATTRIBUTES mdb_all_sa;
1375 static int mdb_sec_inited;
1378 /** Return the library version info. */
1380 mdb_version(int *major, int *minor, int *patch)
1382 if (major) *major = MDB_VERSION_MAJOR;
1383 if (minor) *minor = MDB_VERSION_MINOR;
1384 if (patch) *patch = MDB_VERSION_PATCH;
1385 return MDB_VERSION_STRING;
1388 /** Table of descriptions for LMDB @ref errors */
1389 static char *const mdb_errstr[] = {
1390 "MDB_KEYEXIST: Key/data pair already exists",
1391 "MDB_NOTFOUND: No matching key/data pair found",
1392 "MDB_PAGE_NOTFOUND: Requested page not found",
1393 "MDB_CORRUPTED: Located page was wrong type",
1394 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1395 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1396 "MDB_INVALID: File is not an LMDB file",
1397 "MDB_MAP_FULL: Environment mapsize limit reached",
1398 "MDB_DBS_FULL: Environment maxdbs limit reached",
1399 "MDB_READERS_FULL: Environment maxreaders limit reached",
1400 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1401 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1402 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1403 "MDB_PAGE_FULL: Internal error - page has no more space",
1404 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1405 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1406 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1407 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1408 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1409 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1413 mdb_strerror(int err)
1416 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1417 * This works as long as no function between the call to mdb_strerror
1418 * and the actual use of the message uses more than 4K of stack.
1421 char buf[1024], *ptr = buf;
1425 return ("Successful return: 0");
1427 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1428 i = err - MDB_KEYEXIST;
1429 return mdb_errstr[i];
1433 /* These are the C-runtime error codes we use. The comment indicates
1434 * their numeric value, and the Win32 error they would correspond to
1435 * if the error actually came from a Win32 API. A major mess, we should
1436 * have used LMDB-specific error codes for everything.
1439 case ENOENT: /* 2, FILE_NOT_FOUND */
1440 case EIO: /* 5, ACCESS_DENIED */
1441 case ENOMEM: /* 12, INVALID_ACCESS */
1442 case EACCES: /* 13, INVALID_DATA */
1443 case EBUSY: /* 16, CURRENT_DIRECTORY */
1444 case EINVAL: /* 22, BAD_COMMAND */
1445 case ENOSPC: /* 28, OUT_OF_PAPER */
1446 return strerror(err);
1451 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1452 FORMAT_MESSAGE_IGNORE_INSERTS,
1453 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1456 return strerror(err);
1460 /** assert(3) variant in cursor context */
1461 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1462 /** assert(3) variant in transaction context */
1463 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1464 /** assert(3) variant in environment context */
1465 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1468 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1469 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1472 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1473 const char *func, const char *file, int line)
1476 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1477 file, line, expr_txt, func);
1478 if (env->me_assert_func)
1479 env->me_assert_func(env, buf);
1480 fprintf(stderr, "%s\n", buf);
1484 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1488 /** Return the page number of \b mp which may be sub-page, for debug output */
1490 mdb_dbg_pgno(MDB_page *mp)
1493 COPY_PGNO(ret, mp->mp_pgno);
1497 /** Display a key in hexadecimal and return the address of the result.
1498 * @param[in] key the key to display
1499 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1500 * @return The key in hexadecimal form.
1503 mdb_dkey(MDB_val *key, char *buf)
1506 unsigned char *c = key->mv_data;
1512 if (key->mv_size > DKBUF_MAXKEYSIZE)
1513 return "MDB_MAXKEYSIZE";
1514 /* may want to make this a dynamic check: if the key is mostly
1515 * printable characters, print it as-is instead of converting to hex.
1519 for (i=0; i<key->mv_size; i++)
1520 ptr += sprintf(ptr, "%02x", *c++);
1522 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1528 mdb_leafnode_type(MDB_node *n)
1530 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1531 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1532 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1535 /** Display all the keys in the page. */
1537 mdb_page_list(MDB_page *mp)
1539 pgno_t pgno = mdb_dbg_pgno(mp);
1540 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1542 unsigned int i, nkeys, nsize, total = 0;
1546 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1547 case P_BRANCH: type = "Branch page"; break;
1548 case P_LEAF: type = "Leaf page"; break;
1549 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1550 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1551 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1553 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1554 pgno, mp->mp_pages, state);
1557 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1558 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1561 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1565 nkeys = NUMKEYS(mp);
1566 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1568 for (i=0; i<nkeys; i++) {
1569 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1570 key.mv_size = nsize = mp->mp_pad;
1571 key.mv_data = LEAF2KEY(mp, i, nsize);
1573 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1576 node = NODEPTR(mp, i);
1577 key.mv_size = node->mn_ksize;
1578 key.mv_data = node->mn_data;
1579 nsize = NODESIZE + key.mv_size;
1580 if (IS_BRANCH(mp)) {
1581 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1585 if (F_ISSET(node->mn_flags, F_BIGDATA))
1586 nsize += sizeof(pgno_t);
1588 nsize += NODEDSZ(node);
1590 nsize += sizeof(indx_t);
1591 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1592 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1594 total = EVEN(total);
1596 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1597 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1601 mdb_cursor_chk(MDB_cursor *mc)
1607 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1608 for (i=0; i<mc->mc_top; i++) {
1610 node = NODEPTR(mp, mc->mc_ki[i]);
1611 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1614 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1620 /** Count all the pages in each DB and in the freelist
1621 * and make sure it matches the actual number of pages
1623 * All named DBs must be open for a correct count.
1625 static void mdb_audit(MDB_txn *txn)
1629 MDB_ID freecount, count;
1634 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1635 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1636 freecount += *(MDB_ID *)data.mv_data;
1637 mdb_tassert(txn, rc == MDB_NOTFOUND);
1640 for (i = 0; i<txn->mt_numdbs; i++) {
1642 if (!(txn->mt_dbflags[i] & DB_VALID))
1644 mdb_cursor_init(&mc, txn, i, &mx);
1645 if (txn->mt_dbs[i].md_root == P_INVALID)
1647 count += txn->mt_dbs[i].md_branch_pages +
1648 txn->mt_dbs[i].md_leaf_pages +
1649 txn->mt_dbs[i].md_overflow_pages;
1650 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1651 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1652 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1655 mp = mc.mc_pg[mc.mc_top];
1656 for (j=0; j<NUMKEYS(mp); j++) {
1657 MDB_node *leaf = NODEPTR(mp, j);
1658 if (leaf->mn_flags & F_SUBDATA) {
1660 memcpy(&db, NODEDATA(leaf), sizeof(db));
1661 count += db.md_branch_pages + db.md_leaf_pages +
1662 db.md_overflow_pages;
1666 mdb_tassert(txn, rc == MDB_NOTFOUND);
1669 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1670 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1671 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1677 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1679 return txn->mt_dbxs[dbi].md_cmp(a, b);
1683 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1685 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1686 #if UINT_MAX < SIZE_MAX
1687 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1688 dcmp = mdb_cmp_clong;
1693 /** Allocate memory for a page.
1694 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1697 mdb_page_malloc(MDB_txn *txn, unsigned num)
1699 MDB_env *env = txn->mt_env;
1700 MDB_page *ret = env->me_dpages;
1701 size_t psize = env->me_psize, sz = psize, off;
1702 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1703 * For a single page alloc, we init everything after the page header.
1704 * For multi-page, we init the final page; if the caller needed that
1705 * many pages they will be filling in at least up to the last page.
1709 VGMEMP_ALLOC(env, ret, sz);
1710 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1711 env->me_dpages = ret->mp_next;
1714 psize -= off = PAGEHDRSZ;
1719 if ((ret = malloc(sz)) != NULL) {
1720 VGMEMP_ALLOC(env, ret, sz);
1721 if (!(env->me_flags & MDB_NOMEMINIT)) {
1722 memset((char *)ret + off, 0, psize);
1726 txn->mt_flags |= MDB_TXN_ERROR;
1730 /** Free a single page.
1731 * Saves single pages to a list, for future reuse.
1732 * (This is not used for multi-page overflow pages.)
1735 mdb_page_free(MDB_env *env, MDB_page *mp)
1737 mp->mp_next = env->me_dpages;
1738 VGMEMP_FREE(env, mp);
1739 env->me_dpages = mp;
1742 /** Free a dirty page */
1744 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1746 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1747 mdb_page_free(env, dp);
1749 /* large pages just get freed directly */
1750 VGMEMP_FREE(env, dp);
1755 /** Return all dirty pages to dpage list */
1757 mdb_dlist_free(MDB_txn *txn)
1759 MDB_env *env = txn->mt_env;
1760 MDB_ID2L dl = txn->mt_u.dirty_list;
1761 unsigned i, n = dl[0].mid;
1763 for (i = 1; i <= n; i++) {
1764 mdb_dpage_free(env, dl[i].mptr);
1769 /** Loosen or free a single page.
1770 * Saves single pages to a list for future reuse
1771 * in this same txn. It has been pulled from the freeDB
1772 * and already resides on the dirty list, but has been
1773 * deleted. Use these pages first before pulling again
1776 * If the page wasn't dirtied in this txn, just add it
1777 * to this txn's free list.
1780 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1783 pgno_t pgno = mp->mp_pgno;
1784 MDB_txn *txn = mc->mc_txn;
1786 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1787 if (txn->mt_parent) {
1788 MDB_ID2 *dl = txn->mt_u.dirty_list;
1789 /* If txn has a parent, make sure the page is in our
1793 unsigned x = mdb_mid2l_search(dl, pgno);
1794 if (x <= dl[0].mid && dl[x].mid == pgno) {
1795 if (mp != dl[x].mptr) { /* bad cursor? */
1796 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1797 txn->mt_flags |= MDB_TXN_ERROR;
1798 return MDB_CORRUPTED;
1805 /* no parent txn, so it's just ours */
1810 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1812 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1813 txn->mt_loose_pgs = mp;
1814 txn->mt_loose_count++;
1815 mp->mp_flags |= P_LOOSE;
1817 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1825 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1826 * @param[in] mc A cursor handle for the current operation.
1827 * @param[in] pflags Flags of the pages to update:
1828 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1829 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1830 * @return 0 on success, non-zero on failure.
1833 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1835 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1836 MDB_txn *txn = mc->mc_txn;
1842 int rc = MDB_SUCCESS, level;
1844 /* Mark pages seen by cursors */
1845 if (mc->mc_flags & C_UNTRACK)
1846 mc = NULL; /* will find mc in mt_cursors */
1847 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1848 for (; mc; mc=mc->mc_next) {
1849 if (!(mc->mc_flags & C_INITIALIZED))
1851 for (m3 = mc;; m3 = &mx->mx_cursor) {
1853 for (j=0; j<m3->mc_snum; j++) {
1855 if ((mp->mp_flags & Mask) == pflags)
1856 mp->mp_flags ^= P_KEEP;
1858 mx = m3->mc_xcursor;
1859 /* Proceed to mx if it is at a sub-database */
1860 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1862 if (! (mp && (mp->mp_flags & P_LEAF)))
1864 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1865 if (!(leaf->mn_flags & F_SUBDATA))
1874 /* Mark dirty root pages */
1875 for (i=0; i<txn->mt_numdbs; i++) {
1876 if (txn->mt_dbflags[i] & DB_DIRTY) {
1877 pgno_t pgno = txn->mt_dbs[i].md_root;
1878 if (pgno == P_INVALID)
1880 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1882 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1883 dp->mp_flags ^= P_KEEP;
1891 static int mdb_page_flush(MDB_txn *txn, int keep);
1893 /** Spill pages from the dirty list back to disk.
1894 * This is intended to prevent running into #MDB_TXN_FULL situations,
1895 * but note that they may still occur in a few cases:
1896 * 1) our estimate of the txn size could be too small. Currently this
1897 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1898 * 2) child txns may run out of space if their parents dirtied a
1899 * lot of pages and never spilled them. TODO: we probably should do
1900 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1901 * the parent's dirty_room is below a given threshold.
1903 * Otherwise, if not using nested txns, it is expected that apps will
1904 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1905 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1906 * If the txn never references them again, they can be left alone.
1907 * If the txn only reads them, they can be used without any fuss.
1908 * If the txn writes them again, they can be dirtied immediately without
1909 * going thru all of the work of #mdb_page_touch(). Such references are
1910 * handled by #mdb_page_unspill().
1912 * Also note, we never spill DB root pages, nor pages of active cursors,
1913 * because we'll need these back again soon anyway. And in nested txns,
1914 * we can't spill a page in a child txn if it was already spilled in a
1915 * parent txn. That would alter the parent txns' data even though
1916 * the child hasn't committed yet, and we'd have no way to undo it if
1917 * the child aborted.
1919 * @param[in] m0 cursor A cursor handle identifying the transaction and
1920 * database for which we are checking space.
1921 * @param[in] key For a put operation, the key being stored.
1922 * @param[in] data For a put operation, the data being stored.
1923 * @return 0 on success, non-zero on failure.
1926 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1928 MDB_txn *txn = m0->mc_txn;
1930 MDB_ID2L dl = txn->mt_u.dirty_list;
1931 unsigned int i, j, need;
1934 if (m0->mc_flags & C_SUB)
1937 /* Estimate how much space this op will take */
1938 i = m0->mc_db->md_depth;
1939 /* Named DBs also dirty the main DB */
1940 if (m0->mc_dbi > MAIN_DBI)
1941 i += txn->mt_dbs[MAIN_DBI].md_depth;
1942 /* For puts, roughly factor in the key+data size */
1944 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1945 i += i; /* double it for good measure */
1948 if (txn->mt_dirty_room > i)
1951 if (!txn->mt_spill_pgs) {
1952 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1953 if (!txn->mt_spill_pgs)
1956 /* purge deleted slots */
1957 MDB_IDL sl = txn->mt_spill_pgs;
1958 unsigned int num = sl[0];
1960 for (i=1; i<=num; i++) {
1967 /* Preserve pages which may soon be dirtied again */
1968 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1971 /* Less aggressive spill - we originally spilled the entire dirty list,
1972 * with a few exceptions for cursor pages and DB root pages. But this
1973 * turns out to be a lot of wasted effort because in a large txn many
1974 * of those pages will need to be used again. So now we spill only 1/8th
1975 * of the dirty pages. Testing revealed this to be a good tradeoff,
1976 * better than 1/2, 1/4, or 1/10.
1978 if (need < MDB_IDL_UM_MAX / 8)
1979 need = MDB_IDL_UM_MAX / 8;
1981 /* Save the page IDs of all the pages we're flushing */
1982 /* flush from the tail forward, this saves a lot of shifting later on. */
1983 for (i=dl[0].mid; i && need; i--) {
1984 MDB_ID pn = dl[i].mid << 1;
1986 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1988 /* Can't spill twice, make sure it's not already in a parent's
1991 if (txn->mt_parent) {
1993 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1994 if (tx2->mt_spill_pgs) {
1995 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1996 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1997 dp->mp_flags |= P_KEEP;
2005 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2009 mdb_midl_sort(txn->mt_spill_pgs);
2011 /* Flush the spilled part of dirty list */
2012 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2015 /* Reset any dirty pages we kept that page_flush didn't see */
2016 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2019 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2023 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2025 mdb_find_oldest(MDB_txn *txn)
2028 txnid_t mr, oldest = txn->mt_txnid - 1;
2029 if (txn->mt_env->me_txns) {
2030 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2031 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2042 /** Add a page to the txn's dirty list */
2044 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2047 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2049 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2050 insert = mdb_mid2l_append;
2052 insert = mdb_mid2l_insert;
2054 mid.mid = mp->mp_pgno;
2056 rc = insert(txn->mt_u.dirty_list, &mid);
2057 mdb_tassert(txn, rc == 0);
2058 txn->mt_dirty_room--;
2061 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2062 * me_pghead and mt_next_pgno.
2064 * If there are free pages available from older transactions, they
2065 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2066 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2067 * and move me_pglast to say which records were consumed. Only this
2068 * function can create me_pghead and move me_pglast/mt_next_pgno.
2069 * @param[in] mc cursor A cursor handle identifying the transaction and
2070 * database for which we are allocating.
2071 * @param[in] num the number of pages to allocate.
2072 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2073 * will always be satisfied by a single contiguous chunk of memory.
2074 * @return 0 on success, non-zero on failure.
2077 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2079 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2080 /* Get at most <Max_retries> more freeDB records once me_pghead
2081 * has enough pages. If not enough, use new pages from the map.
2082 * If <Paranoid> and mc is updating the freeDB, only get new
2083 * records if me_pghead is empty. Then the freelist cannot play
2084 * catch-up with itself by growing while trying to save it.
2086 enum { Paranoid = 1, Max_retries = 500 };
2088 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2090 int rc, retry = num * 60;
2091 MDB_txn *txn = mc->mc_txn;
2092 MDB_env *env = txn->mt_env;
2093 pgno_t pgno, *mop = env->me_pghead;
2094 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2096 txnid_t oldest = 0, last;
2101 /* If there are any loose pages, just use them */
2102 if (num == 1 && txn->mt_loose_pgs) {
2103 np = txn->mt_loose_pgs;
2104 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2105 txn->mt_loose_count--;
2106 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2114 /* If our dirty list is already full, we can't do anything */
2115 if (txn->mt_dirty_room == 0) {
2120 for (op = MDB_FIRST;; op = MDB_NEXT) {
2125 /* Seek a big enough contiguous page range. Prefer
2126 * pages at the tail, just truncating the list.
2132 if (mop[i-n2] == pgno+n2)
2139 if (op == MDB_FIRST) { /* 1st iteration */
2140 /* Prepare to fetch more and coalesce */
2141 last = env->me_pglast;
2142 oldest = env->me_pgoldest;
2143 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2146 key.mv_data = &last; /* will look up last+1 */
2147 key.mv_size = sizeof(last);
2149 if (Paranoid && mc->mc_dbi == FREE_DBI)
2152 if (Paranoid && retry < 0 && mop_len)
2156 /* Do not fetch more if the record will be too recent */
2157 if (oldest <= last) {
2159 oldest = mdb_find_oldest(txn);
2160 env->me_pgoldest = oldest;
2166 rc = mdb_cursor_get(&m2, &key, NULL, op);
2168 if (rc == MDB_NOTFOUND)
2172 last = *(txnid_t*)key.mv_data;
2173 if (oldest <= last) {
2175 oldest = mdb_find_oldest(txn);
2176 env->me_pgoldest = oldest;
2182 np = m2.mc_pg[m2.mc_top];
2183 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2184 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2187 idl = (MDB_ID *) data.mv_data;
2190 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2195 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2197 mop = env->me_pghead;
2199 env->me_pglast = last;
2201 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2202 last, txn->mt_dbs[FREE_DBI].md_root, i));
2204 DPRINTF(("IDL %"Z"u", idl[j]));
2206 /* Merge in descending sorted order */
2207 mdb_midl_xmerge(mop, idl);
2211 /* Use new pages from the map when nothing suitable in the freeDB */
2213 pgno = txn->mt_next_pgno;
2214 if (pgno + num >= env->me_maxpg) {
2215 DPUTS("DB size maxed out");
2221 if (env->me_flags & MDB_WRITEMAP) {
2222 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2224 if (!(np = mdb_page_malloc(txn, num))) {
2230 mop[0] = mop_len -= num;
2231 /* Move any stragglers down */
2232 for (j = i-num; j < mop_len; )
2233 mop[++j] = mop[++i];
2235 txn->mt_next_pgno = pgno + num;
2238 mdb_page_dirty(txn, np);
2244 txn->mt_flags |= MDB_TXN_ERROR;
2248 /** Copy the used portions of a non-overflow page.
2249 * @param[in] dst page to copy into
2250 * @param[in] src page to copy from
2251 * @param[in] psize size of a page
2254 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2256 enum { Align = sizeof(pgno_t) };
2257 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2259 /* If page isn't full, just copy the used portion. Adjust
2260 * alignment so memcpy may copy words instead of bytes.
2262 if ((unused &= -Align) && !IS_LEAF2(src)) {
2263 upper = (upper + PAGEBASE) & -Align;
2264 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2265 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2268 memcpy(dst, src, psize - unused);
2272 /** Pull a page off the txn's spill list, if present.
2273 * If a page being referenced was spilled to disk in this txn, bring
2274 * it back and make it dirty/writable again.
2275 * @param[in] txn the transaction handle.
2276 * @param[in] mp the page being referenced. It must not be dirty.
2277 * @param[out] ret the writable page, if any. ret is unchanged if
2278 * mp wasn't spilled.
2281 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2283 MDB_env *env = txn->mt_env;
2286 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2288 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2289 if (!tx2->mt_spill_pgs)
2291 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2292 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2295 if (txn->mt_dirty_room == 0)
2296 return MDB_TXN_FULL;
2297 if (IS_OVERFLOW(mp))
2301 if (env->me_flags & MDB_WRITEMAP) {
2304 np = mdb_page_malloc(txn, num);
2308 memcpy(np, mp, num * env->me_psize);
2310 mdb_page_copy(np, mp, env->me_psize);
2313 /* If in current txn, this page is no longer spilled.
2314 * If it happens to be the last page, truncate the spill list.
2315 * Otherwise mark it as deleted by setting the LSB.
2317 if (x == txn->mt_spill_pgs[0])
2318 txn->mt_spill_pgs[0]--;
2320 txn->mt_spill_pgs[x] |= 1;
2321 } /* otherwise, if belonging to a parent txn, the
2322 * page remains spilled until child commits
2325 mdb_page_dirty(txn, np);
2326 np->mp_flags |= P_DIRTY;
2334 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2335 * @param[in] mc cursor pointing to the page to be touched
2336 * @return 0 on success, non-zero on failure.
2339 mdb_page_touch(MDB_cursor *mc)
2341 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2342 MDB_txn *txn = mc->mc_txn;
2343 MDB_cursor *m2, *m3;
2347 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2348 if (txn->mt_flags & MDB_TXN_SPILLS) {
2350 rc = mdb_page_unspill(txn, mp, &np);
2356 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2357 (rc = mdb_page_alloc(mc, 1, &np)))
2360 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2361 mp->mp_pgno, pgno));
2362 mdb_cassert(mc, mp->mp_pgno != pgno);
2363 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2364 /* Update the parent page, if any, to point to the new page */
2366 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2367 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2368 SETPGNO(node, pgno);
2370 mc->mc_db->md_root = pgno;
2372 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2373 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2375 /* If txn has a parent, make sure the page is in our
2379 unsigned x = mdb_mid2l_search(dl, pgno);
2380 if (x <= dl[0].mid && dl[x].mid == pgno) {
2381 if (mp != dl[x].mptr) { /* bad cursor? */
2382 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2383 txn->mt_flags |= MDB_TXN_ERROR;
2384 return MDB_CORRUPTED;
2389 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2391 np = mdb_page_malloc(txn, 1);
2396 rc = mdb_mid2l_insert(dl, &mid);
2397 mdb_cassert(mc, rc == 0);
2402 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2404 np->mp_flags |= P_DIRTY;
2407 /* Adjust cursors pointing to mp */
2408 mc->mc_pg[mc->mc_top] = np;
2409 m2 = txn->mt_cursors[mc->mc_dbi];
2410 if (mc->mc_flags & C_SUB) {
2411 for (; m2; m2=m2->mc_next) {
2412 m3 = &m2->mc_xcursor->mx_cursor;
2413 if (m3->mc_snum < mc->mc_snum) continue;
2414 if (m3->mc_pg[mc->mc_top] == mp)
2415 m3->mc_pg[mc->mc_top] = np;
2418 for (; m2; m2=m2->mc_next) {
2419 if (m2->mc_snum < mc->mc_snum) continue;
2420 if (m2->mc_pg[mc->mc_top] == mp) {
2421 m2->mc_pg[mc->mc_top] = np;
2422 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2424 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2426 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2427 if (!(leaf->mn_flags & F_SUBDATA))
2428 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2436 txn->mt_flags |= MDB_TXN_ERROR;
2441 mdb_env_sync(MDB_env *env, int force)
2444 if (env->me_flags & MDB_RDONLY)
2446 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2447 if (env->me_flags & MDB_WRITEMAP) {
2448 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2449 ? MS_ASYNC : MS_SYNC;
2450 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2453 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2457 #ifdef BROKEN_FDATASYNC
2458 if (env->me_flags & MDB_FSYNCONLY) {
2459 if (fsync(env->me_fd))
2463 if (MDB_FDATASYNC(env->me_fd))
2470 /** Back up parent txn's cursors, then grab the originals for tracking */
2472 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2474 MDB_cursor *mc, *bk;
2479 for (i = src->mt_numdbs; --i >= 0; ) {
2480 if ((mc = src->mt_cursors[i]) != NULL) {
2481 size = sizeof(MDB_cursor);
2483 size += sizeof(MDB_xcursor);
2484 for (; mc; mc = bk->mc_next) {
2490 mc->mc_db = &dst->mt_dbs[i];
2491 /* Kill pointers into src - and dst to reduce abuse: The
2492 * user may not use mc until dst ends. Otherwise we'd...
2494 mc->mc_txn = NULL; /* ...set this to dst */
2495 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2496 if ((mx = mc->mc_xcursor) != NULL) {
2497 *(MDB_xcursor *)(bk+1) = *mx;
2498 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2500 mc->mc_next = dst->mt_cursors[i];
2501 dst->mt_cursors[i] = mc;
2508 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2509 * @param[in] txn the transaction handle.
2510 * @param[in] merge true to keep changes to parent cursors, false to revert.
2511 * @return 0 on success, non-zero on failure.
2514 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2516 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2520 for (i = txn->mt_numdbs; --i >= 0; ) {
2521 for (mc = cursors[i]; mc; mc = next) {
2523 if ((bk = mc->mc_backup) != NULL) {
2525 /* Commit changes to parent txn */
2526 mc->mc_next = bk->mc_next;
2527 mc->mc_backup = bk->mc_backup;
2528 mc->mc_txn = bk->mc_txn;
2529 mc->mc_db = bk->mc_db;
2530 mc->mc_dbflag = bk->mc_dbflag;
2531 if ((mx = mc->mc_xcursor) != NULL)
2532 mx->mx_cursor.mc_txn = bk->mc_txn;
2534 /* Abort nested txn */
2536 if ((mx = mc->mc_xcursor) != NULL)
2537 *mx = *(MDB_xcursor *)(bk+1);
2541 /* Only malloced cursors are permanently tracked. */
2549 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2552 mdb_txn_reset0(MDB_txn *txn, const char *act);
2554 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2560 Pidset = F_SETLK, Pidcheck = F_GETLK
2564 /** Set or check a pid lock. Set returns 0 on success.
2565 * Check returns 0 if the process is certainly dead, nonzero if it may
2566 * be alive (the lock exists or an error happened so we do not know).
2568 * On Windows Pidset is a no-op, we merely check for the existence
2569 * of the process with the given pid. On POSIX we use a single byte
2570 * lock on the lockfile, set at an offset equal to the pid.
2573 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2575 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2578 if (op == Pidcheck) {
2579 h = OpenProcess(env->me_pidquery, FALSE, pid);
2580 /* No documented "no such process" code, but other program use this: */
2582 return ErrCode() != ERROR_INVALID_PARAMETER;
2583 /* A process exists until all handles to it close. Has it exited? */
2584 ret = WaitForSingleObject(h, 0) != 0;
2591 struct flock lock_info;
2592 memset(&lock_info, 0, sizeof(lock_info));
2593 lock_info.l_type = F_WRLCK;
2594 lock_info.l_whence = SEEK_SET;
2595 lock_info.l_start = pid;
2596 lock_info.l_len = 1;
2597 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2598 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2600 } else if ((rc = ErrCode()) == EINTR) {
2608 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2609 * @param[in] txn the transaction handle to initialize
2610 * @return 0 on success, non-zero on failure.
2613 mdb_txn_renew0(MDB_txn *txn)
2615 MDB_env *env = txn->mt_env;
2616 MDB_txninfo *ti = env->me_txns;
2618 unsigned int i, nr, flags = txn->mt_flags;
2620 int rc, new_notls = 0;
2622 if ((flags &= MDB_TXN_RDONLY) != 0) {
2624 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2625 txn->mt_txnid = meta->mm_txnid;
2626 txn->mt_u.reader = NULL;
2628 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2629 pthread_getspecific(env->me_txkey);
2631 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2632 return MDB_BAD_RSLOT;
2634 MDB_PID_T pid = env->me_pid;
2635 MDB_THR_T tid = pthread_self();
2636 mdb_mutexref_t rmutex = env->me_rmutex;
2638 if (!env->me_live_reader) {
2639 rc = mdb_reader_pid(env, Pidset, pid);
2642 env->me_live_reader = 1;
2645 if (LOCK_MUTEX(rc, env, rmutex))
2647 nr = ti->mti_numreaders;
2648 for (i=0; i<nr; i++)
2649 if (ti->mti_readers[i].mr_pid == 0)
2651 if (i == env->me_maxreaders) {
2652 UNLOCK_MUTEX(rmutex);
2653 return MDB_READERS_FULL;
2655 r = &ti->mti_readers[i];
2656 /* Claim the reader slot, carefully since other code
2657 * uses the reader table un-mutexed: First reset the
2658 * slot, next publish it in mti_numreaders. After
2659 * that, it is safe for mdb_env_close() to touch it.
2660 * When it will be closed, we can finally claim it.
2663 r->mr_txnid = (txnid_t)-1;
2666 ti->mti_numreaders = ++nr;
2667 env->me_close_readers = nr;
2669 UNLOCK_MUTEX(rmutex);
2671 new_notls = (env->me_flags & MDB_NOTLS);
2672 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2677 do /* LY: Retry on a race, ITS#7970. */
2678 r->mr_txnid = ti->mti_txnid;
2679 while(r->mr_txnid != ti->mti_txnid);
2680 txn->mt_txnid = r->mr_txnid;
2681 txn->mt_u.reader = r;
2682 meta = env->me_metas[txn->mt_txnid & 1];
2684 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2686 /* Not yet touching txn == env->me_txn0, it may be active */
2688 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2690 txn->mt_txnid = ti->mti_txnid;
2691 meta = env->me_metas[txn->mt_txnid & 1];
2693 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2694 txn->mt_txnid = meta->mm_txnid;
2698 if (txn->mt_txnid == mdb_debug_start)
2701 txn->mt_child = NULL;
2702 txn->mt_loose_pgs = NULL;
2703 txn->mt_loose_count = 0;
2704 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2705 txn->mt_u.dirty_list = env->me_dirty_list;
2706 txn->mt_u.dirty_list[0].mid = 0;
2707 txn->mt_free_pgs = env->me_free_pgs;
2708 txn->mt_free_pgs[0] = 0;
2709 txn->mt_spill_pgs = NULL;
2711 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2714 /* Copy the DB info and flags */
2715 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2717 /* Moved to here to avoid a data race in read TXNs */
2718 txn->mt_next_pgno = meta->mm_last_pg+1;
2720 txn->mt_flags = flags;
2723 txn->mt_numdbs = env->me_numdbs;
2724 for (i=2; i<txn->mt_numdbs; i++) {
2725 x = env->me_dbflags[i];
2726 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2727 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2729 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2731 if (env->me_maxpg < txn->mt_next_pgno) {
2732 mdb_txn_reset0(txn, "renew0-mapfail");
2734 txn->mt_u.reader->mr_pid = 0;
2735 txn->mt_u.reader = NULL;
2737 return MDB_MAP_RESIZED;
2744 mdb_txn_renew(MDB_txn *txn)
2748 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2751 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2752 DPUTS("environment had fatal error, must shutdown!");
2756 rc = mdb_txn_renew0(txn);
2757 if (rc == MDB_SUCCESS) {
2758 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2759 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2760 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2766 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2770 int rc, size, tsize;
2772 flags &= MDB_TXN_BEGIN_FLAGS;
2773 flags |= env->me_flags & MDB_WRITEMAP;
2775 if (env->me_flags & MDB_FATAL_ERROR) {
2776 DPUTS("environment had fatal error, must shutdown!");
2779 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2783 /* Nested transactions: Max 1 child, write txns only, no writemap */
2784 flags |= parent->mt_flags;
2785 if (parent->mt_child ||
2786 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2788 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2790 /* Child txns save MDB_pgstate and use own copy of cursors */
2791 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2792 size += tsize = sizeof(MDB_ntxn);
2793 } else if (flags & MDB_RDONLY) {
2794 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2795 size += tsize = sizeof(MDB_txn);
2797 /* Reuse preallocated write txn. However, do not touch it until
2798 * mdb_txn_renew0() succeeds, since it currently may be active.
2803 if ((txn = calloc(1, size)) == NULL) {
2804 DPRINTF(("calloc: %s", strerror(errno)));
2807 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2808 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2809 txn->mt_flags = flags;
2814 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2815 txn->mt_dbiseqs = parent->mt_dbiseqs;
2816 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2817 if (!txn->mt_u.dirty_list ||
2818 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2820 free(txn->mt_u.dirty_list);
2824 txn->mt_txnid = parent->mt_txnid;
2825 txn->mt_dirty_room = parent->mt_dirty_room;
2826 txn->mt_u.dirty_list[0].mid = 0;
2827 txn->mt_spill_pgs = NULL;
2828 txn->mt_next_pgno = parent->mt_next_pgno;
2829 parent->mt_child = txn;
2830 txn->mt_parent = parent;
2831 txn->mt_numdbs = parent->mt_numdbs;
2832 txn->mt_dbxs = parent->mt_dbxs;
2833 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2834 /* Copy parent's mt_dbflags, but clear DB_NEW */
2835 for (i=0; i<txn->mt_numdbs; i++)
2836 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2838 ntxn = (MDB_ntxn *)txn;
2839 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2840 if (env->me_pghead) {
2841 size = MDB_IDL_SIZEOF(env->me_pghead);
2842 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2844 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2849 rc = mdb_cursor_shadow(parent, txn);
2851 mdb_txn_reset0(txn, "beginchild-fail");
2852 } else { /* MDB_RDONLY */
2853 txn->mt_dbiseqs = env->me_dbiseqs;
2855 rc = mdb_txn_renew0(txn);
2858 if (txn != env->me_txn0)
2861 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2863 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2864 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2865 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2872 mdb_txn_env(MDB_txn *txn)
2874 if(!txn) return NULL;
2879 mdb_txn_id(MDB_txn *txn)
2882 return txn->mt_txnid;
2885 /** Export or close DBI handles opened in this txn. */
2887 mdb_dbis_update(MDB_txn *txn, int keep)
2890 MDB_dbi n = txn->mt_numdbs;
2891 MDB_env *env = txn->mt_env;
2892 unsigned char *tdbflags = txn->mt_dbflags;
2894 for (i = n; --i >= 2;) {
2895 if (tdbflags[i] & DB_NEW) {
2897 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2899 char *ptr = env->me_dbxs[i].md_name.mv_data;
2901 env->me_dbxs[i].md_name.mv_data = NULL;
2902 env->me_dbxs[i].md_name.mv_size = 0;
2903 env->me_dbflags[i] = 0;
2904 env->me_dbiseqs[i]++;
2910 if (keep && env->me_numdbs < n)
2914 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2915 * May be called twice for readonly txns: First reset it, then abort.
2916 * @param[in] txn the transaction handle to reset
2917 * @param[in] act why the transaction is being reset
2920 mdb_txn_reset0(MDB_txn *txn, const char *act)
2922 MDB_env *env = txn->mt_env;
2924 /* Close any DBI handles opened in this txn */
2925 mdb_dbis_update(txn, 0);
2927 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2928 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2929 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2931 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2932 if (txn->mt_u.reader) {
2933 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2934 if (!(env->me_flags & MDB_NOTLS))
2935 txn->mt_u.reader = NULL; /* txn does not own reader */
2937 txn->mt_numdbs = 0; /* close nothing if called again */
2938 txn->mt_dbxs = NULL; /* mark txn as reset */
2940 pgno_t *pghead = env->me_pghead;
2942 mdb_cursors_close(txn, 0);
2943 if (!(env->me_flags & MDB_WRITEMAP)) {
2944 mdb_dlist_free(txn);
2947 if (!txn->mt_parent) {
2948 if (mdb_midl_shrink(&txn->mt_free_pgs))
2949 env->me_free_pgs = txn->mt_free_pgs;
2951 env->me_pghead = NULL;
2955 /* The writer mutex was locked in mdb_txn_begin. */
2957 UNLOCK_MUTEX(env->me_wmutex);
2959 txn->mt_parent->mt_child = NULL;
2960 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2961 mdb_midl_free(txn->mt_free_pgs);
2962 mdb_midl_free(txn->mt_spill_pgs);
2963 free(txn->mt_u.dirty_list);
2966 mdb_midl_free(pghead);
2971 mdb_txn_reset(MDB_txn *txn)
2976 /* This call is only valid for read-only txns */
2977 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2980 mdb_txn_reset0(txn, "reset");
2984 mdb_txn_abort(MDB_txn *txn)
2990 mdb_txn_abort(txn->mt_child);
2992 mdb_txn_reset0(txn, "abort");
2993 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2994 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2995 txn->mt_u.reader->mr_pid = 0;
2997 if (txn != txn->mt_env->me_txn0)
3001 /** Save the freelist as of this transaction to the freeDB.
3002 * This changes the freelist. Keep trying until it stabilizes.
3005 mdb_freelist_save(MDB_txn *txn)
3007 /* env->me_pghead[] can grow and shrink during this call.
3008 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3009 * Page numbers cannot disappear from txn->mt_free_pgs[].
3012 MDB_env *env = txn->mt_env;
3013 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3014 txnid_t pglast = 0, head_id = 0;
3015 pgno_t freecnt = 0, *free_pgs, *mop;
3016 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3018 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3020 if (env->me_pghead) {
3021 /* Make sure first page of freeDB is touched and on freelist */
3022 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3023 if (rc && rc != MDB_NOTFOUND)
3027 if (!env->me_pghead && txn->mt_loose_pgs) {
3028 /* Put loose page numbers in mt_free_pgs, since
3029 * we may be unable to return them to me_pghead.
3031 MDB_page *mp = txn->mt_loose_pgs;
3032 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3034 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3035 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3036 txn->mt_loose_pgs = NULL;
3037 txn->mt_loose_count = 0;
3040 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3041 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3042 ? SSIZE_MAX : maxfree_1pg;
3045 /* Come back here after each Put() in case freelist changed */
3050 /* If using records from freeDB which we have not yet
3051 * deleted, delete them and any we reserved for me_pghead.
3053 while (pglast < env->me_pglast) {
3054 rc = mdb_cursor_first(&mc, &key, NULL);
3057 pglast = head_id = *(txnid_t *)key.mv_data;
3058 total_room = head_room = 0;
3059 mdb_tassert(txn, pglast <= env->me_pglast);
3060 rc = mdb_cursor_del(&mc, 0);
3065 /* Save the IDL of pages freed by this txn, to a single record */
3066 if (freecnt < txn->mt_free_pgs[0]) {
3068 /* Make sure last page of freeDB is touched and on freelist */
3069 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3070 if (rc && rc != MDB_NOTFOUND)
3073 free_pgs = txn->mt_free_pgs;
3074 /* Write to last page of freeDB */
3075 key.mv_size = sizeof(txn->mt_txnid);
3076 key.mv_data = &txn->mt_txnid;
3078 freecnt = free_pgs[0];
3079 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3080 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3083 /* Retry if mt_free_pgs[] grew during the Put() */
3084 free_pgs = txn->mt_free_pgs;
3085 } while (freecnt < free_pgs[0]);
3086 mdb_midl_sort(free_pgs);
3087 memcpy(data.mv_data, free_pgs, data.mv_size);
3090 unsigned int i = free_pgs[0];
3091 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3092 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3094 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3100 mop = env->me_pghead;
3101 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3103 /* Reserve records for me_pghead[]. Split it if multi-page,
3104 * to avoid searching freeDB for a page range. Use keys in
3105 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3107 if (total_room >= mop_len) {
3108 if (total_room == mop_len || --more < 0)
3110 } else if (head_room >= maxfree_1pg && head_id > 1) {
3111 /* Keep current record (overflow page), add a new one */
3115 /* (Re)write {key = head_id, IDL length = head_room} */
3116 total_room -= head_room;
3117 head_room = mop_len - total_room;
3118 if (head_room > maxfree_1pg && head_id > 1) {
3119 /* Overflow multi-page for part of me_pghead */
3120 head_room /= head_id; /* amortize page sizes */
3121 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3122 } else if (head_room < 0) {
3123 /* Rare case, not bothering to delete this record */
3126 key.mv_size = sizeof(head_id);
3127 key.mv_data = &head_id;
3128 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3129 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3132 /* IDL is initially empty, zero out at least the length */
3133 pgs = (pgno_t *)data.mv_data;
3134 j = head_room > clean_limit ? head_room : 0;
3138 total_room += head_room;
3141 /* Return loose page numbers to me_pghead, though usually none are
3142 * left at this point. The pages themselves remain in dirty_list.
3144 if (txn->mt_loose_pgs) {
3145 MDB_page *mp = txn->mt_loose_pgs;
3146 unsigned count = txn->mt_loose_count;
3148 /* Room for loose pages + temp IDL with same */
3149 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3151 mop = env->me_pghead;
3152 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3153 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3154 loose[ ++count ] = mp->mp_pgno;
3156 mdb_midl_sort(loose);
3157 mdb_midl_xmerge(mop, loose);
3158 txn->mt_loose_pgs = NULL;
3159 txn->mt_loose_count = 0;
3163 /* Fill in the reserved me_pghead records */
3169 rc = mdb_cursor_first(&mc, &key, &data);
3170 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3171 txnid_t id = *(txnid_t *)key.mv_data;
3172 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3175 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3177 if (len > mop_len) {
3179 data.mv_size = (len + 1) * sizeof(MDB_ID);
3181 data.mv_data = mop -= len;
3184 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3186 if (rc || !(mop_len -= len))
3193 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3194 * @param[in] txn the transaction that's being committed
3195 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3196 * @return 0 on success, non-zero on failure.
3199 mdb_page_flush(MDB_txn *txn, int keep)
3201 MDB_env *env = txn->mt_env;
3202 MDB_ID2L dl = txn->mt_u.dirty_list;
3203 unsigned psize = env->me_psize, j;
3204 int i, pagecount = dl[0].mid, rc;
3205 size_t size = 0, pos = 0;
3207 MDB_page *dp = NULL;
3211 struct iovec iov[MDB_COMMIT_PAGES];
3212 ssize_t wpos = 0, wsize = 0, wres;
3213 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3219 if (env->me_flags & MDB_WRITEMAP) {
3220 /* Clear dirty flags */
3221 while (++i <= pagecount) {
3223 /* Don't flush this page yet */
3224 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3225 dp->mp_flags &= ~P_KEEP;
3229 dp->mp_flags &= ~P_DIRTY;
3234 /* Write the pages */
3236 if (++i <= pagecount) {
3238 /* Don't flush this page yet */
3239 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3240 dp->mp_flags &= ~P_KEEP;
3245 /* clear dirty flag */
3246 dp->mp_flags &= ~P_DIRTY;
3249 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3254 /* Windows actually supports scatter/gather I/O, but only on
3255 * unbuffered file handles. Since we're relying on the OS page
3256 * cache for all our data, that's self-defeating. So we just
3257 * write pages one at a time. We use the ov structure to set
3258 * the write offset, to at least save the overhead of a Seek
3261 DPRINTF(("committing page %"Z"u", pgno));
3262 memset(&ov, 0, sizeof(ov));
3263 ov.Offset = pos & 0xffffffff;
3264 ov.OffsetHigh = pos >> 16 >> 16;
3265 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3267 DPRINTF(("WriteFile: %d", rc));
3271 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3272 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3275 /* Write previous page(s) */
3276 #ifdef MDB_USE_PWRITEV
3277 wres = pwritev(env->me_fd, iov, n, wpos);
3280 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3283 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3287 DPRINTF(("lseek: %s", strerror(rc)));
3290 wres = writev(env->me_fd, iov, n);
3293 if (wres != wsize) {
3298 DPRINTF(("Write error: %s", strerror(rc)));
3300 rc = EIO; /* TODO: Use which error code? */
3301 DPUTS("short write, filesystem full?");
3312 DPRINTF(("committing page %"Z"u", pgno));
3313 next_pos = pos + size;
3314 iov[n].iov_len = size;
3315 iov[n].iov_base = (char *)dp;
3321 /* MIPS has cache coherency issues, this is a no-op everywhere else
3322 * Note: for any size >= on-chip cache size, entire on-chip cache is
3325 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3327 for (i = keep; ++i <= pagecount; ) {
3329 /* This is a page we skipped above */
3332 dl[j].mid = dp->mp_pgno;
3335 mdb_dpage_free(env, dp);
3340 txn->mt_dirty_room += i - j;
3346 mdb_txn_commit(MDB_txn *txn)
3355 if (txn->mt_child) {
3356 rc = mdb_txn_commit(txn->mt_child);
3363 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3364 mdb_dbis_update(txn, 1);
3365 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3370 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3371 DPUTS("error flag is set, can't commit");
3373 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3378 if (txn->mt_parent) {
3379 MDB_txn *parent = txn->mt_parent;
3383 unsigned x, y, len, ps_len;
3385 /* Append our free list to parent's */
3386 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3389 mdb_midl_free(txn->mt_free_pgs);
3390 /* Failures after this must either undo the changes
3391 * to the parent or set MDB_TXN_ERROR in the parent.
3394 parent->mt_next_pgno = txn->mt_next_pgno;
3395 parent->mt_flags = txn->mt_flags;
3397 /* Merge our cursors into parent's and close them */
3398 mdb_cursors_close(txn, 1);
3400 /* Update parent's DB table. */
3401 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3402 parent->mt_numdbs = txn->mt_numdbs;
3403 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3404 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3405 for (i=2; i<txn->mt_numdbs; i++) {
3406 /* preserve parent's DB_NEW status */
3407 x = parent->mt_dbflags[i] & DB_NEW;
3408 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3411 dst = parent->mt_u.dirty_list;
3412 src = txn->mt_u.dirty_list;
3413 /* Remove anything in our dirty list from parent's spill list */
3414 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3416 pspill[0] = (pgno_t)-1;
3417 /* Mark our dirty pages as deleted in parent spill list */
3418 for (i=0, len=src[0].mid; ++i <= len; ) {
3419 MDB_ID pn = src[i].mid << 1;
3420 while (pn > pspill[x])
3422 if (pn == pspill[x]) {
3427 /* Squash deleted pagenums if we deleted any */
3428 for (x=y; ++x <= ps_len; )
3429 if (!(pspill[x] & 1))
3430 pspill[++y] = pspill[x];
3434 /* Find len = length of merging our dirty list with parent's */
3436 dst[0].mid = 0; /* simplify loops */
3437 if (parent->mt_parent) {
3438 len = x + src[0].mid;
3439 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3440 for (i = x; y && i; y--) {
3441 pgno_t yp = src[y].mid;
3442 while (yp < dst[i].mid)
3444 if (yp == dst[i].mid) {
3449 } else { /* Simplify the above for single-ancestor case */
3450 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3452 /* Merge our dirty list with parent's */
3454 for (i = len; y; dst[i--] = src[y--]) {
3455 pgno_t yp = src[y].mid;
3456 while (yp < dst[x].mid)
3457 dst[i--] = dst[x--];
3458 if (yp == dst[x].mid)
3459 free(dst[x--].mptr);
3461 mdb_tassert(txn, i == x);
3463 free(txn->mt_u.dirty_list);
3464 parent->mt_dirty_room = txn->mt_dirty_room;
3465 if (txn->mt_spill_pgs) {
3466 if (parent->mt_spill_pgs) {
3467 /* TODO: Prevent failure here, so parent does not fail */
3468 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3470 parent->mt_flags |= MDB_TXN_ERROR;
3471 mdb_midl_free(txn->mt_spill_pgs);
3472 mdb_midl_sort(parent->mt_spill_pgs);
3474 parent->mt_spill_pgs = txn->mt_spill_pgs;
3478 /* Append our loose page list to parent's */
3479 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3481 *lp = txn->mt_loose_pgs;
3482 parent->mt_loose_count += txn->mt_loose_count;
3484 parent->mt_child = NULL;
3485 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3490 if (txn != env->me_txn) {
3491 DPUTS("attempt to commit unknown transaction");
3496 mdb_cursors_close(txn, 0);
3498 if (!txn->mt_u.dirty_list[0].mid &&
3499 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3502 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3503 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3505 /* Update DB root pointers */
3506 if (txn->mt_numdbs > 2) {
3510 data.mv_size = sizeof(MDB_db);
3512 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3513 for (i = 2; i < txn->mt_numdbs; i++) {
3514 if (txn->mt_dbflags[i] & DB_DIRTY) {
3515 if (TXN_DBI_CHANGED(txn, i)) {
3519 data.mv_data = &txn->mt_dbs[i];
3520 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3527 rc = mdb_freelist_save(txn);
3531 mdb_midl_free(env->me_pghead);
3532 env->me_pghead = NULL;
3533 if (mdb_midl_shrink(&txn->mt_free_pgs))
3534 env->me_free_pgs = txn->mt_free_pgs;
3540 if ((rc = mdb_page_flush(txn, 0)))
3542 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3543 (rc = mdb_env_sync(env, 0)))
3545 if ((rc = mdb_env_write_meta(txn)))
3548 /* Free P_LOOSE pages left behind in dirty_list */
3549 if (!(env->me_flags & MDB_WRITEMAP))
3550 mdb_dlist_free(txn);
3555 mdb_dbis_update(txn, 1);
3558 UNLOCK_MUTEX(env->me_wmutex);
3559 if (txn != env->me_txn0)
3569 /** Read the environment parameters of a DB environment before
3570 * mapping it into memory.
3571 * @param[in] env the environment handle
3572 * @param[out] meta address of where to store the meta information
3573 * @return 0 on success, non-zero on failure.
3576 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3582 enum { Size = sizeof(pbuf) };
3584 /* We don't know the page size yet, so use a minimum value.
3585 * Read both meta pages so we can use the latest one.
3588 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3592 memset(&ov, 0, sizeof(ov));
3594 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3595 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3598 rc = pread(env->me_fd, &pbuf, Size, off);
3601 if (rc == 0 && off == 0)
3603 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3604 DPRINTF(("read: %s", mdb_strerror(rc)));
3608 p = (MDB_page *)&pbuf;
3610 if (!F_ISSET(p->mp_flags, P_META)) {
3611 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3616 if (m->mm_magic != MDB_MAGIC) {
3617 DPUTS("meta has invalid magic");
3621 if (m->mm_version != MDB_DATA_VERSION) {
3622 DPRINTF(("database is version %u, expected version %u",
3623 m->mm_version, MDB_DATA_VERSION));
3624 return MDB_VERSION_MISMATCH;
3627 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3633 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3635 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3637 meta->mm_magic = MDB_MAGIC;
3638 meta->mm_version = MDB_DATA_VERSION;
3639 meta->mm_mapsize = env->me_mapsize;
3640 meta->mm_psize = env->me_psize;
3641 meta->mm_last_pg = 1;
3642 meta->mm_flags = env->me_flags & 0xffff;
3643 meta->mm_flags |= MDB_INTEGERKEY;
3644 meta->mm_dbs[0].md_root = P_INVALID;
3645 meta->mm_dbs[1].md_root = P_INVALID;
3648 /** Write the environment parameters of a freshly created DB environment.
3649 * @param[in] env the environment handle
3650 * @param[in] meta the #MDB_meta to write
3651 * @return 0 on success, non-zero on failure.
3654 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3662 memset(&ov, 0, sizeof(ov));
3663 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3665 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3668 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3669 len = pwrite(fd, ptr, size, pos); \
3670 if (len == -1 && ErrCode() == EINTR) continue; \
3671 rc = (len >= 0); break; } while(1)
3674 DPUTS("writing new meta page");
3676 psize = env->me_psize;
3678 p = calloc(2, psize);
3680 p->mp_flags = P_META;
3681 *(MDB_meta *)METADATA(p) = *meta;
3683 q = (MDB_page *)((char *)p + psize);
3685 q->mp_flags = P_META;
3686 *(MDB_meta *)METADATA(q) = *meta;
3688 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3691 else if ((unsigned) len == psize * 2)
3699 /** Update the environment info to commit a transaction.
3700 * @param[in] txn the transaction that's being committed
3701 * @return 0 on success, non-zero on failure.
3704 mdb_env_write_meta(MDB_txn *txn)
3707 MDB_meta meta, metab, *mp;
3711 int rc, len, toggle;
3720 toggle = txn->mt_txnid & 1;
3721 DPRINTF(("writing meta page %d for root page %"Z"u",
3722 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3725 flags = txn->mt_flags & env->me_flags;
3726 mp = env->me_metas[toggle];
3727 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3728 /* Persist any increases of mapsize config */
3729 if (mapsize < env->me_mapsize)
3730 mapsize = env->me_mapsize;
3732 if (flags & MDB_WRITEMAP) {
3733 mp->mm_mapsize = mapsize;
3734 mp->mm_dbs[0] = txn->mt_dbs[0];
3735 mp->mm_dbs[1] = txn->mt_dbs[1];
3736 mp->mm_last_pg = txn->mt_next_pgno - 1;
3737 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3738 !(defined(__i386__) || defined(__x86_64__))
3739 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3740 __sync_synchronize();
3742 mp->mm_txnid = txn->mt_txnid;
3743 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3744 unsigned meta_size = env->me_psize;
3745 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3748 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3749 if (meta_size < env->me_os_psize)
3750 meta_size += meta_size;
3755 if (MDB_MSYNC(ptr, meta_size, rc)) {
3762 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3763 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3765 meta.mm_mapsize = mapsize;
3766 meta.mm_dbs[0] = txn->mt_dbs[0];
3767 meta.mm_dbs[1] = txn->mt_dbs[1];
3768 meta.mm_last_pg = txn->mt_next_pgno - 1;
3769 meta.mm_txnid = txn->mt_txnid;
3771 off = offsetof(MDB_meta, mm_mapsize);
3772 ptr = (char *)&meta + off;
3773 len = sizeof(MDB_meta) - off;
3775 off += env->me_psize;
3778 /* Write to the SYNC fd */
3779 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3783 memset(&ov, 0, sizeof(ov));
3785 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3789 rc = pwrite(mfd, ptr, len, off);
3792 rc = rc < 0 ? ErrCode() : EIO;
3795 DPUTS("write failed, disk error?");
3796 /* On a failure, the pagecache still contains the new data.
3797 * Write some old data back, to prevent it from being used.
3798 * Use the non-SYNC fd; we know it will fail anyway.
3800 meta.mm_last_pg = metab.mm_last_pg;
3801 meta.mm_txnid = metab.mm_txnid;
3803 memset(&ov, 0, sizeof(ov));
3805 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3807 r2 = pwrite(env->me_fd, ptr, len, off);
3808 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3811 env->me_flags |= MDB_FATAL_ERROR;
3814 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3815 CACHEFLUSH(env->me_map + off, len, DCACHE);
3817 /* Memory ordering issues are irrelevant; since the entire writer
3818 * is wrapped by wmutex, all of these changes will become visible
3819 * after the wmutex is unlocked. Since the DB is multi-version,
3820 * readers will get consistent data regardless of how fresh or
3821 * how stale their view of these values is.
3824 env->me_txns->mti_txnid = txn->mt_txnid;
3829 /** Check both meta pages to see which one is newer.
3830 * @param[in] env the environment handle
3831 * @return meta toggle (0 or 1).
3834 mdb_env_pick_meta(const MDB_env *env)
3836 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3840 mdb_env_create(MDB_env **env)
3844 e = calloc(1, sizeof(MDB_env));
3848 e->me_maxreaders = DEFAULT_READERS;
3849 e->me_maxdbs = e->me_numdbs = 2;
3850 e->me_fd = INVALID_HANDLE_VALUE;
3851 e->me_lfd = INVALID_HANDLE_VALUE;
3852 e->me_mfd = INVALID_HANDLE_VALUE;
3853 #ifdef MDB_USE_POSIX_SEM
3854 e->me_rmutex = SEM_FAILED;
3855 e->me_wmutex = SEM_FAILED;
3856 #elif defined MDB_USE_SYSV_SEM
3857 e->me_rmutex->semid = -1;
3858 e->me_wmutex->semid = -1;
3860 e->me_pid = getpid();
3861 GET_PAGESIZE(e->me_os_psize);
3862 VGMEMP_CREATE(e,0,0);
3868 mdb_env_map(MDB_env *env, void *addr)
3871 unsigned int flags = env->me_flags;
3875 LONG sizelo, sizehi;
3878 if (flags & MDB_RDONLY) {
3879 /* Don't set explicit map size, use whatever exists */
3884 msize = env->me_mapsize;
3885 sizelo = msize & 0xffffffff;
3886 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3888 /* Windows won't create mappings for zero length files.
3889 * and won't map more than the file size.
3890 * Just set the maxsize right now.
3892 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3893 || !SetEndOfFile(env->me_fd)
3894 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3898 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3899 PAGE_READWRITE : PAGE_READONLY,
3900 sizehi, sizelo, NULL);
3903 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3904 FILE_MAP_WRITE : FILE_MAP_READ,
3906 rc = env->me_map ? 0 : ErrCode();
3911 int prot = PROT_READ;
3912 if (flags & MDB_WRITEMAP) {
3914 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3917 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3919 if (env->me_map == MAP_FAILED) {
3924 if (flags & MDB_NORDAHEAD) {
3925 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3927 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3929 #ifdef POSIX_MADV_RANDOM
3930 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3931 #endif /* POSIX_MADV_RANDOM */
3932 #endif /* MADV_RANDOM */
3936 /* Can happen because the address argument to mmap() is just a
3937 * hint. mmap() can pick another, e.g. if the range is in use.
3938 * The MAP_FIXED flag would prevent that, but then mmap could
3939 * instead unmap existing pages to make room for the new map.
3941 if (addr && env->me_map != addr)
3942 return EBUSY; /* TODO: Make a new MDB_* error code? */
3944 p = (MDB_page *)env->me_map;
3945 env->me_metas[0] = METADATA(p);
3946 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3952 mdb_env_set_mapsize(MDB_env *env, size_t size)
3954 /* If env is already open, caller is responsible for making
3955 * sure there are no active txns.
3963 meta = env->me_metas[mdb_env_pick_meta(env)];
3965 size = meta->mm_mapsize;
3967 /* Silently round up to minimum if the size is too small */
3968 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3972 munmap(env->me_map, env->me_mapsize);
3973 env->me_mapsize = size;
3974 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3975 rc = mdb_env_map(env, old);
3979 env->me_mapsize = size;
3981 env->me_maxpg = env->me_mapsize / env->me_psize;
3986 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3990 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3995 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3997 if (env->me_map || readers < 1)
3999 env->me_maxreaders = readers;
4004 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4006 if (!env || !readers)
4008 *readers = env->me_maxreaders;
4013 mdb_fsize(HANDLE fd, size_t *size)
4016 LARGE_INTEGER fsize;
4018 if (!GetFileSizeEx(fd, &fsize))
4021 *size = fsize.QuadPart;
4033 #ifdef BROKEN_FDATASYNC
4034 #include <sys/utsname.h>
4035 #include <sys/vfs.h>
4038 /** Further setup required for opening an LMDB environment
4041 mdb_env_open2(MDB_env *env)
4043 unsigned int flags = env->me_flags;
4044 int i, newenv = 0, rc;
4048 /* See if we should use QueryLimited */
4050 if ((rc & 0xff) > 5)
4051 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4053 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4056 #ifdef BROKEN_FDATASYNC
4057 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4058 * https://lkml.org/lkml/2012/9/3/83
4059 * Kernels after 3.6-rc6 are known good.
4060 * https://lkml.org/lkml/2012/9/10/556
4061 * See if the DB is on ext3/ext4, then check for new enough kernel
4062 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4067 fstatfs(env->me_fd, &st);
4068 while (st.f_type == 0xEF53) {
4072 if (uts.release[0] < '3') {
4073 if (!strncmp(uts.release, "2.6.32.", 7)) {
4074 i = atoi(uts.release+7);
4076 break; /* 2.6.32.60 and newer is OK */
4077 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4078 i = atoi(uts.release+7);
4080 break; /* 2.6.34.15 and newer is OK */
4082 } else if (uts.release[0] == '3') {
4083 i = atoi(uts.release+2);
4085 break; /* 3.6 and newer is OK */
4087 i = atoi(uts.release+4);
4089 break; /* 3.5.4 and newer is OK */
4090 } else if (i == 2) {
4091 i = atoi(uts.release+4);
4093 break; /* 3.2.30 and newer is OK */
4095 } else { /* 4.x and newer is OK */
4098 env->me_flags |= MDB_FSYNCONLY;
4104 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4107 DPUTS("new mdbenv");
4109 env->me_psize = env->me_os_psize;
4110 if (env->me_psize > MAX_PAGESIZE)
4111 env->me_psize = MAX_PAGESIZE;
4112 memset(&meta, 0, sizeof(meta));
4113 mdb_env_init_meta0(env, &meta);
4114 meta.mm_mapsize = DEFAULT_MAPSIZE;
4116 env->me_psize = meta.mm_psize;
4119 /* Was a mapsize configured? */
4120 if (!env->me_mapsize) {
4121 env->me_mapsize = meta.mm_mapsize;
4124 /* Make sure mapsize >= committed data size. Even when using
4125 * mm_mapsize, which could be broken in old files (ITS#7789).
4127 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4128 if (env->me_mapsize < minsize)
4129 env->me_mapsize = minsize;
4131 meta.mm_mapsize = env->me_mapsize;
4133 if (newenv && !(flags & MDB_FIXEDMAP)) {
4134 /* mdb_env_map() may grow the datafile. Write the metapages
4135 * first, so the file will be valid if initialization fails.
4136 * Except with FIXEDMAP, since we do not yet know mm_address.
4137 * We could fill in mm_address later, but then a different
4138 * program might end up doing that - one with a memory layout
4139 * and map address which does not suit the main program.
4141 rc = mdb_env_init_meta(env, &meta);
4147 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4152 if (flags & MDB_FIXEDMAP)
4153 meta.mm_address = env->me_map;
4154 i = mdb_env_init_meta(env, &meta);
4155 if (i != MDB_SUCCESS) {
4160 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4161 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4163 #if !(MDB_MAXKEYSIZE)
4164 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4166 env->me_maxpg = env->me_mapsize / env->me_psize;
4170 int toggle = mdb_env_pick_meta(env);
4171 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4173 DPRINTF(("opened database version %u, pagesize %u",
4174 env->me_metas[0]->mm_version, env->me_psize));
4175 DPRINTF(("using meta page %d", toggle));
4176 DPRINTF(("depth: %u", db->md_depth));
4177 DPRINTF(("entries: %"Z"u", db->md_entries));
4178 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4179 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4180 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4181 DPRINTF(("root: %"Z"u", db->md_root));
4189 /** Release a reader thread's slot in the reader lock table.
4190 * This function is called automatically when a thread exits.
4191 * @param[in] ptr This points to the slot in the reader lock table.
4194 mdb_env_reader_dest(void *ptr)
4196 MDB_reader *reader = ptr;
4202 /** Junk for arranging thread-specific callbacks on Windows. This is
4203 * necessarily platform and compiler-specific. Windows supports up
4204 * to 1088 keys. Let's assume nobody opens more than 64 environments
4205 * in a single process, for now. They can override this if needed.
4207 #ifndef MAX_TLS_KEYS
4208 #define MAX_TLS_KEYS 64
4210 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4211 static int mdb_tls_nkeys;
4213 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4217 case DLL_PROCESS_ATTACH: break;
4218 case DLL_THREAD_ATTACH: break;
4219 case DLL_THREAD_DETACH:
4220 for (i=0; i<mdb_tls_nkeys; i++) {
4221 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4223 mdb_env_reader_dest(r);
4227 case DLL_PROCESS_DETACH: break;
4232 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4234 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4238 /* Force some symbol references.
4239 * _tls_used forces the linker to create the TLS directory if not already done
4240 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4242 #pragma comment(linker, "/INCLUDE:_tls_used")
4243 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4244 #pragma const_seg(".CRT$XLB")
4245 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4246 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4249 #pragma comment(linker, "/INCLUDE:__tls_used")
4250 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4251 #pragma data_seg(".CRT$XLB")
4252 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4254 #endif /* WIN 32/64 */
4255 #endif /* !__GNUC__ */
4258 /** Downgrade the exclusive lock on the region back to shared */
4260 mdb_env_share_locks(MDB_env *env, int *excl)
4262 int rc = 0, toggle = mdb_env_pick_meta(env);
4264 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4269 /* First acquire a shared lock. The Unlock will
4270 * then release the existing exclusive lock.
4272 memset(&ov, 0, sizeof(ov));
4273 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4276 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4282 struct flock lock_info;
4283 /* The shared lock replaces the existing lock */
4284 memset((void *)&lock_info, 0, sizeof(lock_info));
4285 lock_info.l_type = F_RDLCK;
4286 lock_info.l_whence = SEEK_SET;
4287 lock_info.l_start = 0;
4288 lock_info.l_len = 1;
4289 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4290 (rc = ErrCode()) == EINTR) ;
4291 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4298 /** Try to get exclusive lock, otherwise shared.
4299 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4302 mdb_env_excl_lock(MDB_env *env, int *excl)
4306 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4310 memset(&ov, 0, sizeof(ov));
4311 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4318 struct flock lock_info;
4319 memset((void *)&lock_info, 0, sizeof(lock_info));
4320 lock_info.l_type = F_WRLCK;
4321 lock_info.l_whence = SEEK_SET;
4322 lock_info.l_start = 0;
4323 lock_info.l_len = 1;
4324 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4325 (rc = ErrCode()) == EINTR) ;
4329 # ifndef MDB_USE_POSIX_MUTEX
4330 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4333 lock_info.l_type = F_RDLCK;
4334 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4335 (rc = ErrCode()) == EINTR) ;
4345 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4347 * @(#) $Revision: 5.1 $
4348 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4349 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4351 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4355 * Please do not copyright this code. This code is in the public domain.
4357 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4358 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4359 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4360 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4361 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4362 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4363 * PERFORMANCE OF THIS SOFTWARE.
4366 * chongo <Landon Curt Noll> /\oo/\
4367 * http://www.isthe.com/chongo/
4369 * Share and Enjoy! :-)
4372 typedef unsigned long long mdb_hash_t;
4373 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4375 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4376 * @param[in] val value to hash
4377 * @param[in] hval initial value for hash
4378 * @return 64 bit hash
4380 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4381 * hval arg on the first call.
4384 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4386 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4387 unsigned char *end = s + val->mv_size;
4389 * FNV-1a hash each octet of the string
4392 /* xor the bottom with the current octet */
4393 hval ^= (mdb_hash_t)*s++;
4395 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4396 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4397 (hval << 7) + (hval << 8) + (hval << 40);
4399 /* return our new hash value */
4403 /** Hash the string and output the encoded hash.
4404 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4405 * very short name limits. We don't care about the encoding being reversible,
4406 * we just want to preserve as many bits of the input as possible in a
4407 * small printable string.
4408 * @param[in] str string to hash
4409 * @param[out] encbuf an array of 11 chars to hold the hash
4411 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4414 mdb_pack85(unsigned long l, char *out)
4418 for (i=0; i<5; i++) {
4419 *out++ = mdb_a85[l % 85];
4425 mdb_hash_enc(MDB_val *val, char *encbuf)
4427 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4429 mdb_pack85(h, encbuf);
4430 mdb_pack85(h>>32, encbuf+5);
4435 /** Open and/or initialize the lock region for the environment.
4436 * @param[in] env The LMDB environment.
4437 * @param[in] lpath The pathname of the file used for the lock region.
4438 * @param[in] mode The Unix permissions for the file, if we create it.
4439 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4440 * @return 0 on success, non-zero on failure.
4443 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4446 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4448 # define MDB_ERRCODE_ROFS EROFS
4449 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4450 # define MDB_CLOEXEC O_CLOEXEC
4453 # define MDB_CLOEXEC 0
4456 #ifdef MDB_USE_SYSV_SEM
4464 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4465 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4466 FILE_ATTRIBUTE_NORMAL, NULL);
4468 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4470 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4472 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4477 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4478 /* Lose record locks when exec*() */
4479 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4480 fcntl(env->me_lfd, F_SETFD, fdflags);
4483 if (!(env->me_flags & MDB_NOTLS)) {
4484 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4487 env->me_flags |= MDB_ENV_TXKEY;
4489 /* Windows TLS callbacks need help finding their TLS info. */
4490 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4494 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4498 /* Try to get exclusive lock. If we succeed, then
4499 * nobody is using the lock region and we should initialize it.
4501 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4504 size = GetFileSize(env->me_lfd, NULL);
4506 size = lseek(env->me_lfd, 0, SEEK_END);
4507 if (size == -1) goto fail_errno;
4509 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4510 if (size < rsize && *excl > 0) {
4512 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4513 || !SetEndOfFile(env->me_lfd))
4516 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4520 size = rsize - sizeof(MDB_txninfo);
4521 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4526 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4528 if (!mh) goto fail_errno;
4529 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4531 if (!env->me_txns) goto fail_errno;
4533 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4535 if (m == MAP_FAILED) goto fail_errno;
4541 BY_HANDLE_FILE_INFORMATION stbuf;
4550 if (!mdb_sec_inited) {
4551 InitializeSecurityDescriptor(&mdb_null_sd,
4552 SECURITY_DESCRIPTOR_REVISION);
4553 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4554 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4555 mdb_all_sa.bInheritHandle = FALSE;
4556 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4559 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4560 idbuf.volume = stbuf.dwVolumeSerialNumber;
4561 idbuf.nhigh = stbuf.nFileIndexHigh;
4562 idbuf.nlow = stbuf.nFileIndexLow;
4563 val.mv_data = &idbuf;
4564 val.mv_size = sizeof(idbuf);
4565 mdb_hash_enc(&val, encbuf);
4566 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4567 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4568 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4569 if (!env->me_rmutex) goto fail_errno;
4570 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4571 if (!env->me_wmutex) goto fail_errno;
4572 #elif defined(MDB_USE_POSIX_SEM)
4581 #if defined(__NetBSD__)
4582 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4584 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4585 idbuf.dev = stbuf.st_dev;
4586 idbuf.ino = stbuf.st_ino;
4587 val.mv_data = &idbuf;
4588 val.mv_size = sizeof(idbuf);
4589 mdb_hash_enc(&val, encbuf);
4590 #ifdef MDB_SHORT_SEMNAMES
4591 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4593 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4594 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4595 /* Clean up after a previous run, if needed: Try to
4596 * remove both semaphores before doing anything else.
4598 sem_unlink(env->me_txns->mti_rmname);
4599 sem_unlink(env->me_txns->mti_wmname);
4600 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4601 O_CREAT|O_EXCL, mode, 1);
4602 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4603 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4604 O_CREAT|O_EXCL, mode, 1);
4605 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4606 #elif defined(MDB_USE_SYSV_SEM)
4607 unsigned short vals[2] = {1, 1};
4608 key_t key = ftok(lpath, 'M');
4611 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4615 if (semctl(semid, 0, SETALL, semu) < 0)
4617 env->me_txns->mti_semid = semid;
4618 #else /* MDB_USE_POSIX_MUTEX: */
4619 pthread_mutexattr_t mattr;
4621 if ((rc = pthread_mutexattr_init(&mattr))
4622 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4623 #ifdef MDB_ROBUST_SUPPORTED
4624 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4626 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4627 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4629 pthread_mutexattr_destroy(&mattr);
4630 #endif /* _WIN32 || ... */
4632 env->me_txns->mti_magic = MDB_MAGIC;
4633 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4634 env->me_txns->mti_txnid = 0;
4635 env->me_txns->mti_numreaders = 0;
4638 #ifdef MDB_USE_SYSV_SEM
4639 struct semid_ds buf;
4641 if (env->me_txns->mti_magic != MDB_MAGIC) {
4642 DPUTS("lock region has invalid magic");
4646 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4647 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4648 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4649 rc = MDB_VERSION_MISMATCH;
4653 if (rc && rc != EACCES && rc != EAGAIN) {
4657 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4658 if (!env->me_rmutex) goto fail_errno;
4659 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4660 if (!env->me_wmutex) goto fail_errno;
4661 #elif defined(MDB_USE_POSIX_SEM)
4662 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4663 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4664 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4665 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4666 #elif defined(MDB_USE_SYSV_SEM)
4667 semid = env->me_txns->mti_semid;
4669 /* check for read access */
4670 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4672 /* check for write access */
4673 if (semctl(semid, 0, IPC_SET, semu) < 0)
4677 #ifdef MDB_USE_SYSV_SEM
4678 env->me_rmutex->semid = semid;
4679 env->me_wmutex->semid = semid;
4680 env->me_rmutex->semnum = 0;
4681 env->me_wmutex->semnum = 1;
4682 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4683 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4694 /** The name of the lock file in the DB environment */
4695 #define LOCKNAME "/lock.mdb"
4696 /** The name of the data file in the DB environment */
4697 #define DATANAME "/data.mdb"
4698 /** The suffix of the lock file when no subdir is used */
4699 #define LOCKSUFF "-lock"
4700 /** Only a subset of the @ref mdb_env flags can be changed
4701 * at runtime. Changing other flags requires closing the
4702 * environment and re-opening it with the new flags.
4704 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4705 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4706 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4708 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4709 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4713 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4715 int oflags, rc, len, excl = -1;
4716 char *lpath, *dpath;
4718 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4722 if (flags & MDB_NOSUBDIR) {
4723 rc = len + sizeof(LOCKSUFF) + len + 1;
4725 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4730 if (flags & MDB_NOSUBDIR) {
4731 dpath = lpath + len + sizeof(LOCKSUFF);
4732 sprintf(lpath, "%s" LOCKSUFF, path);
4733 strcpy(dpath, path);
4735 dpath = lpath + len + sizeof(LOCKNAME);
4736 sprintf(lpath, "%s" LOCKNAME, path);
4737 sprintf(dpath, "%s" DATANAME, path);
4741 flags |= env->me_flags;
4742 if (flags & MDB_RDONLY) {
4743 /* silently ignore WRITEMAP when we're only getting read access */
4744 flags &= ~MDB_WRITEMAP;
4746 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4747 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4750 env->me_flags = flags |= MDB_ENV_ACTIVE;
4754 env->me_path = strdup(path);
4755 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4756 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4757 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4758 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4762 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4764 /* For RDONLY, get lockfile after we know datafile exists */
4765 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4766 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4772 if (F_ISSET(flags, MDB_RDONLY)) {
4773 oflags = GENERIC_READ;
4774 len = OPEN_EXISTING;
4776 oflags = GENERIC_READ|GENERIC_WRITE;
4779 mode = FILE_ATTRIBUTE_NORMAL;
4780 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4781 NULL, len, mode, NULL);
4783 if (F_ISSET(flags, MDB_RDONLY))
4786 oflags = O_RDWR | O_CREAT;
4788 env->me_fd = open(dpath, oflags, mode);
4790 if (env->me_fd == INVALID_HANDLE_VALUE) {
4795 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4796 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4801 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4802 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4803 env->me_mfd = env->me_fd;
4805 /* Synchronous fd for meta writes. Needed even with
4806 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4809 len = OPEN_EXISTING;
4810 env->me_mfd = CreateFile(dpath, oflags,
4811 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4812 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4815 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4817 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4822 DPRINTF(("opened dbenv %p", (void *) env));
4824 rc = mdb_env_share_locks(env, &excl);
4828 if (!(flags & MDB_RDONLY)) {
4830 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4831 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4832 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4833 (txn = calloc(1, size)))
4835 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4836 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4837 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4838 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4840 txn->mt_dbxs = env->me_dbxs;
4850 mdb_env_close0(env, excl);
4856 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4858 mdb_env_close0(MDB_env *env, int excl)
4862 if (!(env->me_flags & MDB_ENV_ACTIVE))
4865 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4867 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4868 free(env->me_dbxs[i].md_name.mv_data);
4873 free(env->me_dbiseqs);
4874 free(env->me_dbflags);
4876 free(env->me_dirty_list);
4878 mdb_midl_free(env->me_free_pgs);
4880 if (env->me_flags & MDB_ENV_TXKEY) {
4881 pthread_key_delete(env->me_txkey);
4883 /* Delete our key from the global list */
4884 for (i=0; i<mdb_tls_nkeys; i++)
4885 if (mdb_tls_keys[i] == env->me_txkey) {
4886 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4894 munmap(env->me_map, env->me_mapsize);
4896 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4897 (void) close(env->me_mfd);
4898 if (env->me_fd != INVALID_HANDLE_VALUE)
4899 (void) close(env->me_fd);
4901 MDB_PID_T pid = env->me_pid;
4902 /* Clearing readers is done in this function because
4903 * me_txkey with its destructor must be disabled first.
4905 * We skip the the reader mutex, so we touch only
4906 * data owned by this process (me_close_readers and
4907 * our readers), and clear each reader atomically.
4909 for (i = env->me_close_readers; --i >= 0; )
4910 if (env->me_txns->mti_readers[i].mr_pid == pid)
4911 env->me_txns->mti_readers[i].mr_pid = 0;
4913 if (env->me_rmutex) {
4914 CloseHandle(env->me_rmutex);
4915 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4917 /* Windows automatically destroys the mutexes when
4918 * the last handle closes.
4920 #elif defined(MDB_USE_POSIX_SEM)
4921 if (env->me_rmutex != SEM_FAILED) {
4922 sem_close(env->me_rmutex);
4923 if (env->me_wmutex != SEM_FAILED)
4924 sem_close(env->me_wmutex);
4925 /* If we have the filelock: If we are the
4926 * only remaining user, clean up semaphores.
4929 mdb_env_excl_lock(env, &excl);
4931 sem_unlink(env->me_txns->mti_rmname);
4932 sem_unlink(env->me_txns->mti_wmname);
4935 #elif defined(MDB_USE_SYSV_SEM)
4936 if (env->me_rmutex->semid != -1) {
4937 /* If we have the filelock: If we are the
4938 * only remaining user, clean up semaphores.
4941 mdb_env_excl_lock(env, &excl);
4943 semctl(env->me_rmutex->semid, 0, IPC_RMID);
4946 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4948 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4951 /* Unlock the lockfile. Windows would have unlocked it
4952 * after closing anyway, but not necessarily at once.
4954 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4957 (void) close(env->me_lfd);
4960 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4964 mdb_env_close(MDB_env *env)
4971 VGMEMP_DESTROY(env);
4972 while ((dp = env->me_dpages) != NULL) {
4973 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4974 env->me_dpages = dp->mp_next;
4978 mdb_env_close0(env, 0);
4982 /** Compare two items pointing at aligned size_t's */
4984 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4986 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4987 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4990 /** Compare two items pointing at aligned unsigned int's.
4992 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4993 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4996 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4998 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4999 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5002 /** Compare two items pointing at unsigned ints of unknown alignment.
5003 * Nodes and keys are guaranteed to be 2-byte aligned.
5006 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5008 #if BYTE_ORDER == LITTLE_ENDIAN
5009 unsigned short *u, *c;
5012 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5013 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5016 } while(!x && u > (unsigned short *)a->mv_data);
5019 unsigned short *u, *c, *end;
5022 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5023 u = (unsigned short *)a->mv_data;
5024 c = (unsigned short *)b->mv_data;
5027 } while(!x && u < end);
5032 /** Compare two items lexically */
5034 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5041 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5047 diff = memcmp(a->mv_data, b->mv_data, len);
5048 return diff ? diff : len_diff<0 ? -1 : len_diff;
5051 /** Compare two items in reverse byte order */
5053 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5055 const unsigned char *p1, *p2, *p1_lim;
5059 p1_lim = (const unsigned char *)a->mv_data;
5060 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5061 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5063 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5069 while (p1 > p1_lim) {
5070 diff = *--p1 - *--p2;
5074 return len_diff<0 ? -1 : len_diff;
5077 /** Search for key within a page, using binary search.
5078 * Returns the smallest entry larger or equal to the key.
5079 * If exactp is non-null, stores whether the found entry was an exact match
5080 * in *exactp (1 or 0).
5081 * Updates the cursor index with the index of the found entry.
5082 * If no entry larger or equal to the key is found, returns NULL.
5085 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5087 unsigned int i = 0, nkeys;
5090 MDB_page *mp = mc->mc_pg[mc->mc_top];
5091 MDB_node *node = NULL;
5096 nkeys = NUMKEYS(mp);
5098 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5099 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5102 low = IS_LEAF(mp) ? 0 : 1;
5104 cmp = mc->mc_dbx->md_cmp;
5106 /* Branch pages have no data, so if using integer keys,
5107 * alignment is guaranteed. Use faster mdb_cmp_int.
5109 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5110 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5117 nodekey.mv_size = mc->mc_db->md_pad;
5118 node = NODEPTR(mp, 0); /* fake */
5119 while (low <= high) {
5120 i = (low + high) >> 1;
5121 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5122 rc = cmp(key, &nodekey);
5123 DPRINTF(("found leaf index %u [%s], rc = %i",
5124 i, DKEY(&nodekey), rc));
5133 while (low <= high) {
5134 i = (low + high) >> 1;
5136 node = NODEPTR(mp, i);
5137 nodekey.mv_size = NODEKSZ(node);
5138 nodekey.mv_data = NODEKEY(node);
5140 rc = cmp(key, &nodekey);
5143 DPRINTF(("found leaf index %u [%s], rc = %i",
5144 i, DKEY(&nodekey), rc));
5146 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5147 i, DKEY(&nodekey), NODEPGNO(node), rc));
5158 if (rc > 0) { /* Found entry is less than the key. */
5159 i++; /* Skip to get the smallest entry larger than key. */
5161 node = NODEPTR(mp, i);
5164 *exactp = (rc == 0 && nkeys > 0);
5165 /* store the key index */
5166 mc->mc_ki[mc->mc_top] = i;
5168 /* There is no entry larger or equal to the key. */
5171 /* nodeptr is fake for LEAF2 */
5177 mdb_cursor_adjust(MDB_cursor *mc, func)
5181 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5182 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5189 /** Pop a page off the top of the cursor's stack. */
5191 mdb_cursor_pop(MDB_cursor *mc)
5195 MDB_page *top = mc->mc_pg[mc->mc_top];
5201 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5202 DDBI(mc), (void *) mc));
5206 /** Push a page onto the top of the cursor's stack. */
5208 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5210 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5211 DDBI(mc), (void *) mc));
5213 if (mc->mc_snum >= CURSOR_STACK) {
5214 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5215 return MDB_CURSOR_FULL;
5218 mc->mc_top = mc->mc_snum++;
5219 mc->mc_pg[mc->mc_top] = mp;
5220 mc->mc_ki[mc->mc_top] = 0;
5225 /** Find the address of the page corresponding to a given page number.
5226 * @param[in] txn the transaction for this access.
5227 * @param[in] pgno the page number for the page to retrieve.
5228 * @param[out] ret address of a pointer where the page's address will be stored.
5229 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5230 * @return 0 on success, non-zero on failure.
5233 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5235 MDB_env *env = txn->mt_env;
5239 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5243 MDB_ID2L dl = tx2->mt_u.dirty_list;
5245 /* Spilled pages were dirtied in this txn and flushed
5246 * because the dirty list got full. Bring this page
5247 * back in from the map (but don't unspill it here,
5248 * leave that unless page_touch happens again).
5250 if (tx2->mt_spill_pgs) {
5251 MDB_ID pn = pgno << 1;
5252 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5253 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5254 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5259 unsigned x = mdb_mid2l_search(dl, pgno);
5260 if (x <= dl[0].mid && dl[x].mid == pgno) {
5266 } while ((tx2 = tx2->mt_parent) != NULL);
5269 if (pgno < txn->mt_next_pgno) {
5271 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5273 DPRINTF(("page %"Z"u not found", pgno));
5274 txn->mt_flags |= MDB_TXN_ERROR;
5275 return MDB_PAGE_NOTFOUND;
5285 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5286 * The cursor is at the root page, set up the rest of it.
5289 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5291 MDB_page *mp = mc->mc_pg[mc->mc_top];
5295 while (IS_BRANCH(mp)) {
5299 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5300 mdb_cassert(mc, NUMKEYS(mp) > 1);
5301 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5303 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5305 if (flags & MDB_PS_LAST)
5306 i = NUMKEYS(mp) - 1;
5309 node = mdb_node_search(mc, key, &exact);
5311 i = NUMKEYS(mp) - 1;
5313 i = mc->mc_ki[mc->mc_top];
5315 mdb_cassert(mc, i > 0);
5319 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5322 mdb_cassert(mc, i < NUMKEYS(mp));
5323 node = NODEPTR(mp, i);
5325 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5328 mc->mc_ki[mc->mc_top] = i;
5329 if ((rc = mdb_cursor_push(mc, mp)))
5332 if (flags & MDB_PS_MODIFY) {
5333 if ((rc = mdb_page_touch(mc)) != 0)
5335 mp = mc->mc_pg[mc->mc_top];
5340 DPRINTF(("internal error, index points to a %02X page!?",
5342 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5343 return MDB_CORRUPTED;
5346 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5347 key ? DKEY(key) : "null"));
5348 mc->mc_flags |= C_INITIALIZED;
5349 mc->mc_flags &= ~C_EOF;
5354 /** Search for the lowest key under the current branch page.
5355 * This just bypasses a NUMKEYS check in the current page
5356 * before calling mdb_page_search_root(), because the callers
5357 * are all in situations where the current page is known to
5361 mdb_page_search_lowest(MDB_cursor *mc)
5363 MDB_page *mp = mc->mc_pg[mc->mc_top];
5364 MDB_node *node = NODEPTR(mp, 0);
5367 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5370 mc->mc_ki[mc->mc_top] = 0;
5371 if ((rc = mdb_cursor_push(mc, mp)))
5373 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5376 /** Search for the page a given key should be in.
5377 * Push it and its parent pages on the cursor stack.
5378 * @param[in,out] mc the cursor for this operation.
5379 * @param[in] key the key to search for, or NULL for first/last page.
5380 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5381 * are touched (updated with new page numbers).
5382 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5383 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5384 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5385 * @return 0 on success, non-zero on failure.
5388 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5393 /* Make sure the txn is still viable, then find the root from
5394 * the txn's db table and set it as the root of the cursor's stack.
5396 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5397 DPUTS("transaction has failed, must abort");
5400 /* Make sure we're using an up-to-date root */
5401 if (*mc->mc_dbflag & DB_STALE) {
5403 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5405 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5406 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5413 MDB_node *leaf = mdb_node_search(&mc2,
5414 &mc->mc_dbx->md_name, &exact);
5416 return MDB_NOTFOUND;
5417 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5420 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5422 /* The txn may not know this DBI, or another process may
5423 * have dropped and recreated the DB with other flags.
5425 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5426 return MDB_INCOMPATIBLE;
5427 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5429 *mc->mc_dbflag &= ~DB_STALE;
5431 root = mc->mc_db->md_root;
5433 if (root == P_INVALID) { /* Tree is empty. */
5434 DPUTS("tree is empty");
5435 return MDB_NOTFOUND;
5439 mdb_cassert(mc, root > 1);
5440 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5441 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5447 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5448 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5450 if (flags & MDB_PS_MODIFY) {
5451 if ((rc = mdb_page_touch(mc)))
5455 if (flags & MDB_PS_ROOTONLY)
5458 return mdb_page_search_root(mc, key, flags);
5462 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5464 MDB_txn *txn = mc->mc_txn;
5465 pgno_t pg = mp->mp_pgno;
5466 unsigned x = 0, ovpages = mp->mp_pages;
5467 MDB_env *env = txn->mt_env;
5468 MDB_IDL sl = txn->mt_spill_pgs;
5469 MDB_ID pn = pg << 1;
5472 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5473 /* If the page is dirty or on the spill list we just acquired it,
5474 * so we should give it back to our current free list, if any.
5475 * Otherwise put it onto the list of pages we freed in this txn.
5477 * Won't create me_pghead: me_pglast must be inited along with it.
5478 * Unsupported in nested txns: They would need to hide the page
5479 * range in ancestor txns' dirty and spilled lists.
5481 if (env->me_pghead &&
5483 ((mp->mp_flags & P_DIRTY) ||
5484 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5488 MDB_ID2 *dl, ix, iy;
5489 rc = mdb_midl_need(&env->me_pghead, ovpages);
5492 if (!(mp->mp_flags & P_DIRTY)) {
5493 /* This page is no longer spilled */
5500 /* Remove from dirty list */
5501 dl = txn->mt_u.dirty_list;
5503 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5509 mdb_cassert(mc, x > 1);
5511 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5512 txn->mt_flags |= MDB_TXN_ERROR;
5513 return MDB_CORRUPTED;
5516 if (!(env->me_flags & MDB_WRITEMAP))
5517 mdb_dpage_free(env, mp);
5519 /* Insert in me_pghead */
5520 mop = env->me_pghead;
5521 j = mop[0] + ovpages;
5522 for (i = mop[0]; i && mop[i] < pg; i--)
5528 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5532 mc->mc_db->md_overflow_pages -= ovpages;
5536 /** Return the data associated with a given node.
5537 * @param[in] txn The transaction for this operation.
5538 * @param[in] leaf The node being read.
5539 * @param[out] data Updated to point to the node's data.
5540 * @return 0 on success, non-zero on failure.
5543 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5545 MDB_page *omp; /* overflow page */
5549 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5550 data->mv_size = NODEDSZ(leaf);
5551 data->mv_data = NODEDATA(leaf);
5555 /* Read overflow data.
5557 data->mv_size = NODEDSZ(leaf);
5558 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5559 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5560 DPRINTF(("read overflow page %"Z"u failed", pgno));
5563 data->mv_data = METADATA(omp);
5569 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5570 MDB_val *key, MDB_val *data)
5577 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5579 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5582 if (txn->mt_flags & MDB_TXN_ERROR)
5585 mdb_cursor_init(&mc, txn, dbi, &mx);
5586 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5589 /** Find a sibling for a page.
5590 * Replaces the page at the top of the cursor's stack with the
5591 * specified sibling, if one exists.
5592 * @param[in] mc The cursor for this operation.
5593 * @param[in] move_right Non-zero if the right sibling is requested,
5594 * otherwise the left sibling.
5595 * @return 0 on success, non-zero on failure.
5598 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5604 if (mc->mc_snum < 2) {
5605 return MDB_NOTFOUND; /* root has no siblings */
5609 DPRINTF(("parent page is page %"Z"u, index %u",
5610 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5612 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5613 : (mc->mc_ki[mc->mc_top] == 0)) {
5614 DPRINTF(("no more keys left, moving to %s sibling",
5615 move_right ? "right" : "left"));
5616 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5617 /* undo cursor_pop before returning */
5624 mc->mc_ki[mc->mc_top]++;
5626 mc->mc_ki[mc->mc_top]--;
5627 DPRINTF(("just moving to %s index key %u",
5628 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5630 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5632 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5633 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5634 /* mc will be inconsistent if caller does mc_snum++ as above */
5635 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5639 mdb_cursor_push(mc, mp);
5641 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5646 /** Move the cursor to the next data item. */
5648 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5654 if (mc->mc_flags & C_EOF) {
5655 return MDB_NOTFOUND;
5658 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5660 mp = mc->mc_pg[mc->mc_top];
5662 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5663 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5664 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5665 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5666 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5667 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5668 if (rc == MDB_SUCCESS)
5669 MDB_GET_KEY(leaf, key);
5674 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5675 if (op == MDB_NEXT_DUP)
5676 return MDB_NOTFOUND;
5680 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5681 mdb_dbg_pgno(mp), (void *) mc));
5682 if (mc->mc_flags & C_DEL)
5685 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5686 DPUTS("=====> move to next sibling page");
5687 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5688 mc->mc_flags |= C_EOF;
5691 mp = mc->mc_pg[mc->mc_top];
5692 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5694 mc->mc_ki[mc->mc_top]++;
5697 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5698 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5701 key->mv_size = mc->mc_db->md_pad;
5702 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5706 mdb_cassert(mc, IS_LEAF(mp));
5707 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5709 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5710 mdb_xcursor_init1(mc, leaf);
5713 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5716 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5717 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5718 if (rc != MDB_SUCCESS)
5723 MDB_GET_KEY(leaf, key);
5727 /** Move the cursor to the previous data item. */
5729 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5735 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5737 mp = mc->mc_pg[mc->mc_top];
5739 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5740 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5741 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5742 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5743 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5744 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5745 if (rc == MDB_SUCCESS) {
5746 MDB_GET_KEY(leaf, key);
5747 mc->mc_flags &= ~C_EOF;
5753 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5754 if (op == MDB_PREV_DUP)
5755 return MDB_NOTFOUND;
5759 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5760 mdb_dbg_pgno(mp), (void *) mc));
5762 if (mc->mc_ki[mc->mc_top] == 0) {
5763 DPUTS("=====> move to prev sibling page");
5764 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5767 mp = mc->mc_pg[mc->mc_top];
5768 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5769 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5771 mc->mc_ki[mc->mc_top]--;
5773 mc->mc_flags &= ~C_EOF;
5775 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5776 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5779 key->mv_size = mc->mc_db->md_pad;
5780 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5784 mdb_cassert(mc, IS_LEAF(mp));
5785 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5787 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5788 mdb_xcursor_init1(mc, leaf);
5791 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5794 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5795 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5796 if (rc != MDB_SUCCESS)
5801 MDB_GET_KEY(leaf, key);
5805 /** Set the cursor on a specific data item. */
5807 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5808 MDB_cursor_op op, int *exactp)
5812 MDB_node *leaf = NULL;
5815 if (key->mv_size == 0)
5816 return MDB_BAD_VALSIZE;
5819 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5821 /* See if we're already on the right page */
5822 if (mc->mc_flags & C_INITIALIZED) {
5825 mp = mc->mc_pg[mc->mc_top];
5827 mc->mc_ki[mc->mc_top] = 0;
5828 return MDB_NOTFOUND;
5830 if (mp->mp_flags & P_LEAF2) {
5831 nodekey.mv_size = mc->mc_db->md_pad;
5832 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5834 leaf = NODEPTR(mp, 0);
5835 MDB_GET_KEY2(leaf, nodekey);
5837 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5839 /* Probably happens rarely, but first node on the page
5840 * was the one we wanted.
5842 mc->mc_ki[mc->mc_top] = 0;
5849 unsigned int nkeys = NUMKEYS(mp);
5851 if (mp->mp_flags & P_LEAF2) {
5852 nodekey.mv_data = LEAF2KEY(mp,
5853 nkeys-1, nodekey.mv_size);
5855 leaf = NODEPTR(mp, nkeys-1);
5856 MDB_GET_KEY2(leaf, nodekey);
5858 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5860 /* last node was the one we wanted */
5861 mc->mc_ki[mc->mc_top] = nkeys-1;
5867 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5868 /* This is definitely the right page, skip search_page */
5869 if (mp->mp_flags & P_LEAF2) {
5870 nodekey.mv_data = LEAF2KEY(mp,
5871 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5873 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5874 MDB_GET_KEY2(leaf, nodekey);
5876 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5878 /* current node was the one we wanted */
5888 /* If any parents have right-sibs, search.
5889 * Otherwise, there's nothing further.
5891 for (i=0; i<mc->mc_top; i++)
5893 NUMKEYS(mc->mc_pg[i])-1)
5895 if (i == mc->mc_top) {
5896 /* There are no other pages */
5897 mc->mc_ki[mc->mc_top] = nkeys;
5898 return MDB_NOTFOUND;
5902 /* There are no other pages */
5903 mc->mc_ki[mc->mc_top] = 0;
5904 if (op == MDB_SET_RANGE && !exactp) {
5908 return MDB_NOTFOUND;
5912 rc = mdb_page_search(mc, key, 0);
5913 if (rc != MDB_SUCCESS)
5916 mp = mc->mc_pg[mc->mc_top];
5917 mdb_cassert(mc, IS_LEAF(mp));
5920 leaf = mdb_node_search(mc, key, exactp);
5921 if (exactp != NULL && !*exactp) {
5922 /* MDB_SET specified and not an exact match. */
5923 return MDB_NOTFOUND;
5927 DPUTS("===> inexact leaf not found, goto sibling");
5928 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5929 return rc; /* no entries matched */
5930 mp = mc->mc_pg[mc->mc_top];
5931 mdb_cassert(mc, IS_LEAF(mp));
5932 leaf = NODEPTR(mp, 0);
5936 mc->mc_flags |= C_INITIALIZED;
5937 mc->mc_flags &= ~C_EOF;
5940 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5941 key->mv_size = mc->mc_db->md_pad;
5942 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5947 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5948 mdb_xcursor_init1(mc, leaf);
5951 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5952 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5953 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5956 if (op == MDB_GET_BOTH) {
5962 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5963 if (rc != MDB_SUCCESS)
5966 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5969 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5971 dcmp = mc->mc_dbx->md_dcmp;
5972 #if UINT_MAX < SIZE_MAX
5973 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5974 dcmp = mdb_cmp_clong;
5976 rc = dcmp(data, &olddata);
5978 if (op == MDB_GET_BOTH || rc > 0)
5979 return MDB_NOTFOUND;
5986 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5987 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5992 /* The key already matches in all other cases */
5993 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5994 MDB_GET_KEY(leaf, key);
5995 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6000 /** Move the cursor to the first item in the database. */
6002 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6008 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6010 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6011 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6012 if (rc != MDB_SUCCESS)
6015 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6017 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6018 mc->mc_flags |= C_INITIALIZED;
6019 mc->mc_flags &= ~C_EOF;
6021 mc->mc_ki[mc->mc_top] = 0;
6023 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6024 key->mv_size = mc->mc_db->md_pad;
6025 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6030 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6031 mdb_xcursor_init1(mc, leaf);
6032 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6036 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6040 MDB_GET_KEY(leaf, key);
6044 /** Move the cursor to the last item in the database. */
6046 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6052 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6054 if (!(mc->mc_flags & C_EOF)) {
6056 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6057 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6058 if (rc != MDB_SUCCESS)
6061 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6064 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6065 mc->mc_flags |= C_INITIALIZED|C_EOF;
6066 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6068 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6069 key->mv_size = mc->mc_db->md_pad;
6070 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6075 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6076 mdb_xcursor_init1(mc, leaf);
6077 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6081 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6086 MDB_GET_KEY(leaf, key);
6091 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6096 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6101 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6105 case MDB_GET_CURRENT:
6106 if (!(mc->mc_flags & C_INITIALIZED)) {
6109 MDB_page *mp = mc->mc_pg[mc->mc_top];
6110 int nkeys = NUMKEYS(mp);
6111 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6112 mc->mc_ki[mc->mc_top] = nkeys;
6118 key->mv_size = mc->mc_db->md_pad;
6119 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6121 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6122 MDB_GET_KEY(leaf, key);
6124 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6125 if (mc->mc_flags & C_DEL)
6126 mdb_xcursor_init1(mc, leaf);
6127 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6129 rc = mdb_node_read(mc->mc_txn, leaf, data);
6136 case MDB_GET_BOTH_RANGE:
6141 if (mc->mc_xcursor == NULL) {
6142 rc = MDB_INCOMPATIBLE;
6152 rc = mdb_cursor_set(mc, key, data, op,
6153 op == MDB_SET_RANGE ? NULL : &exact);
6156 case MDB_GET_MULTIPLE:
6157 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6161 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6162 rc = MDB_INCOMPATIBLE;
6166 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6167 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6170 case MDB_NEXT_MULTIPLE:
6175 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6176 rc = MDB_INCOMPATIBLE;
6179 if (!(mc->mc_flags & C_INITIALIZED))
6180 rc = mdb_cursor_first(mc, key, data);
6182 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6183 if (rc == MDB_SUCCESS) {
6184 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6187 mx = &mc->mc_xcursor->mx_cursor;
6188 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6190 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6191 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6199 case MDB_NEXT_NODUP:
6200 if (!(mc->mc_flags & C_INITIALIZED))
6201 rc = mdb_cursor_first(mc, key, data);
6203 rc = mdb_cursor_next(mc, key, data, op);
6207 case MDB_PREV_NODUP:
6208 if (!(mc->mc_flags & C_INITIALIZED)) {
6209 rc = mdb_cursor_last(mc, key, data);
6212 mc->mc_flags |= C_INITIALIZED;
6213 mc->mc_ki[mc->mc_top]++;
6215 rc = mdb_cursor_prev(mc, key, data, op);
6218 rc = mdb_cursor_first(mc, key, data);
6221 mfunc = mdb_cursor_first;
6223 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6227 if (mc->mc_xcursor == NULL) {
6228 rc = MDB_INCOMPATIBLE;
6232 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6233 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6234 MDB_GET_KEY(leaf, key);
6235 rc = mdb_node_read(mc->mc_txn, leaf, data);
6239 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6243 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6246 rc = mdb_cursor_last(mc, key, data);
6249 mfunc = mdb_cursor_last;
6252 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6257 if (mc->mc_flags & C_DEL)
6258 mc->mc_flags ^= C_DEL;
6263 /** Touch all the pages in the cursor stack. Set mc_top.
6264 * Makes sure all the pages are writable, before attempting a write operation.
6265 * @param[in] mc The cursor to operate on.
6268 mdb_cursor_touch(MDB_cursor *mc)
6270 int rc = MDB_SUCCESS;
6272 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6275 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6277 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6278 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6281 *mc->mc_dbflag |= DB_DIRTY;
6286 rc = mdb_page_touch(mc);
6287 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6288 mc->mc_top = mc->mc_snum-1;
6293 /** Do not spill pages to disk if txn is getting full, may fail instead */
6294 #define MDB_NOSPILL 0x8000
6297 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6301 MDB_node *leaf = NULL;
6304 MDB_val xdata, *rdata, dkey, olddata;
6306 int do_sub = 0, insert_key, insert_data;
6307 unsigned int mcount = 0, dcount = 0, nospill;
6310 unsigned int nflags;
6313 if (mc == NULL || key == NULL)
6316 env = mc->mc_txn->mt_env;
6318 /* Check this first so counter will always be zero on any
6321 if (flags & MDB_MULTIPLE) {
6322 dcount = data[1].mv_size;
6323 data[1].mv_size = 0;
6324 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6325 return MDB_INCOMPATIBLE;
6328 nospill = flags & MDB_NOSPILL;
6329 flags &= ~MDB_NOSPILL;
6331 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6332 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6334 if (key->mv_size-1 >= ENV_MAXKEY(env))
6335 return MDB_BAD_VALSIZE;
6337 #if SIZE_MAX > MAXDATASIZE
6338 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6339 return MDB_BAD_VALSIZE;
6341 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6342 return MDB_BAD_VALSIZE;
6345 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6346 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6350 if (flags == MDB_CURRENT) {
6351 if (!(mc->mc_flags & C_INITIALIZED))
6354 } else if (mc->mc_db->md_root == P_INVALID) {
6355 /* new database, cursor has nothing to point to */
6358 mc->mc_flags &= ~C_INITIALIZED;
6363 if (flags & MDB_APPEND) {
6365 rc = mdb_cursor_last(mc, &k2, &d2);
6367 rc = mc->mc_dbx->md_cmp(key, &k2);
6370 mc->mc_ki[mc->mc_top]++;
6372 /* new key is <= last key */
6377 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6379 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6380 DPRINTF(("duplicate key [%s]", DKEY(key)));
6382 return MDB_KEYEXIST;
6384 if (rc && rc != MDB_NOTFOUND)
6388 if (mc->mc_flags & C_DEL)
6389 mc->mc_flags ^= C_DEL;
6391 /* Cursor is positioned, check for room in the dirty list */
6393 if (flags & MDB_MULTIPLE) {
6395 xdata.mv_size = data->mv_size * dcount;
6399 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6403 if (rc == MDB_NO_ROOT) {
6405 /* new database, write a root leaf page */
6406 DPUTS("allocating new root leaf page");
6407 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6410 mdb_cursor_push(mc, np);
6411 mc->mc_db->md_root = np->mp_pgno;
6412 mc->mc_db->md_depth++;
6413 *mc->mc_dbflag |= DB_DIRTY;
6414 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6416 np->mp_flags |= P_LEAF2;
6417 mc->mc_flags |= C_INITIALIZED;
6419 /* make sure all cursor pages are writable */
6420 rc2 = mdb_cursor_touch(mc);
6425 insert_key = insert_data = rc;
6427 /* The key does not exist */
6428 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6429 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6430 LEAFSIZE(key, data) > env->me_nodemax)
6432 /* Too big for a node, insert in sub-DB. Set up an empty
6433 * "old sub-page" for prep_subDB to expand to a full page.
6435 fp_flags = P_LEAF|P_DIRTY;
6437 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6438 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6439 olddata.mv_size = PAGEHDRSZ;
6443 /* there's only a key anyway, so this is a no-op */
6444 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6446 unsigned int ksize = mc->mc_db->md_pad;
6447 if (key->mv_size != ksize)
6448 return MDB_BAD_VALSIZE;
6449 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6450 memcpy(ptr, key->mv_data, ksize);
6452 /* if overwriting slot 0 of leaf, need to
6453 * update branch key if there is a parent page
6455 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6456 unsigned short top = mc->mc_top;
6458 /* slot 0 is always an empty key, find real slot */
6459 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6461 if (mc->mc_ki[mc->mc_top])
6462 rc2 = mdb_update_key(mc, key);
6473 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6474 olddata.mv_size = NODEDSZ(leaf);
6475 olddata.mv_data = NODEDATA(leaf);
6478 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6479 /* Prepare (sub-)page/sub-DB to accept the new item,
6480 * if needed. fp: old sub-page or a header faking
6481 * it. mp: new (sub-)page. offset: growth in page
6482 * size. xdata: node data with new page or DB.
6484 unsigned i, offset = 0;
6485 mp = fp = xdata.mv_data = env->me_pbuf;
6486 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6488 /* Was a single item before, must convert now */
6489 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6491 /* Just overwrite the current item */
6492 if (flags == MDB_CURRENT)
6494 dcmp = mc->mc_dbx->md_dcmp;
6495 #if UINT_MAX < SIZE_MAX
6496 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6497 dcmp = mdb_cmp_clong;
6499 /* does data match? */
6500 if (!dcmp(data, &olddata)) {
6501 if (flags & MDB_NODUPDATA)
6502 return MDB_KEYEXIST;
6507 /* Back up original data item */
6508 dkey.mv_size = olddata.mv_size;
6509 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6511 /* Make sub-page header for the dup items, with dummy body */
6512 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6513 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6514 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6515 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6516 fp->mp_flags |= P_LEAF2;
6517 fp->mp_pad = data->mv_size;
6518 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6520 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6521 (dkey.mv_size & 1) + (data->mv_size & 1);
6523 fp->mp_upper = xdata.mv_size - PAGEBASE;
6524 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6525 } else if (leaf->mn_flags & F_SUBDATA) {
6526 /* Data is on sub-DB, just store it */
6527 flags |= F_DUPDATA|F_SUBDATA;
6530 /* Data is on sub-page */
6531 fp = olddata.mv_data;
6534 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6535 offset = EVEN(NODESIZE + sizeof(indx_t) +
6539 offset = fp->mp_pad;
6540 if (SIZELEFT(fp) < offset) {
6541 offset *= 4; /* space for 4 more */
6544 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6546 fp->mp_flags |= P_DIRTY;
6547 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6548 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6552 xdata.mv_size = olddata.mv_size + offset;
6555 fp_flags = fp->mp_flags;
6556 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6557 /* Too big for a sub-page, convert to sub-DB */
6558 fp_flags &= ~P_SUBP;
6560 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6561 fp_flags |= P_LEAF2;
6562 dummy.md_pad = fp->mp_pad;
6563 dummy.md_flags = MDB_DUPFIXED;
6564 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6565 dummy.md_flags |= MDB_INTEGERKEY;
6571 dummy.md_branch_pages = 0;
6572 dummy.md_leaf_pages = 1;
6573 dummy.md_overflow_pages = 0;
6574 dummy.md_entries = NUMKEYS(fp);
6575 xdata.mv_size = sizeof(MDB_db);
6576 xdata.mv_data = &dummy;
6577 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6579 offset = env->me_psize - olddata.mv_size;
6580 flags |= F_DUPDATA|F_SUBDATA;
6581 dummy.md_root = mp->mp_pgno;
6584 mp->mp_flags = fp_flags | P_DIRTY;
6585 mp->mp_pad = fp->mp_pad;
6586 mp->mp_lower = fp->mp_lower;
6587 mp->mp_upper = fp->mp_upper + offset;
6588 if (fp_flags & P_LEAF2) {
6589 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6591 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6592 olddata.mv_size - fp->mp_upper - PAGEBASE);
6593 for (i=0; i<NUMKEYS(fp); i++)
6594 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6602 mdb_node_del(mc, 0);
6606 /* overflow page overwrites need special handling */
6607 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6610 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6612 memcpy(&pg, olddata.mv_data, sizeof(pg));
6613 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6615 ovpages = omp->mp_pages;
6617 /* Is the ov page large enough? */
6618 if (ovpages >= dpages) {
6619 if (!(omp->mp_flags & P_DIRTY) &&
6620 (level || (env->me_flags & MDB_WRITEMAP)))
6622 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6625 level = 0; /* dirty in this txn or clean */
6628 if (omp->mp_flags & P_DIRTY) {
6629 /* yes, overwrite it. Note in this case we don't
6630 * bother to try shrinking the page if the new data
6631 * is smaller than the overflow threshold.
6634 /* It is writable only in a parent txn */
6635 size_t sz = (size_t) env->me_psize * ovpages, off;
6636 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6642 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6643 mdb_cassert(mc, rc2 == 0);
6644 if (!(flags & MDB_RESERVE)) {
6645 /* Copy end of page, adjusting alignment so
6646 * compiler may copy words instead of bytes.
6648 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6649 memcpy((size_t *)((char *)np + off),
6650 (size_t *)((char *)omp + off), sz - off);
6653 memcpy(np, omp, sz); /* Copy beginning of page */
6656 SETDSZ(leaf, data->mv_size);
6657 if (F_ISSET(flags, MDB_RESERVE))
6658 data->mv_data = METADATA(omp);
6660 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6664 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6666 } else if (data->mv_size == olddata.mv_size) {
6667 /* same size, just replace it. Note that we could
6668 * also reuse this node if the new data is smaller,
6669 * but instead we opt to shrink the node in that case.
6671 if (F_ISSET(flags, MDB_RESERVE))
6672 data->mv_data = olddata.mv_data;
6673 else if (!(mc->mc_flags & C_SUB))
6674 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6676 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6681 mdb_node_del(mc, 0);
6687 nflags = flags & NODE_ADD_FLAGS;
6688 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6689 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6690 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6691 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6693 nflags |= MDB_SPLIT_REPLACE;
6694 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6696 /* There is room already in this leaf page. */
6697 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6698 if (rc == 0 && insert_key) {
6699 /* Adjust other cursors pointing to mp */
6700 MDB_cursor *m2, *m3;
6701 MDB_dbi dbi = mc->mc_dbi;
6702 unsigned i = mc->mc_top;
6703 MDB_page *mp = mc->mc_pg[i];
6705 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6706 if (mc->mc_flags & C_SUB)
6707 m3 = &m2->mc_xcursor->mx_cursor;
6710 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6711 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6718 if (rc == MDB_SUCCESS) {
6719 /* Now store the actual data in the child DB. Note that we're
6720 * storing the user data in the keys field, so there are strict
6721 * size limits on dupdata. The actual data fields of the child
6722 * DB are all zero size.
6730 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6731 if (flags & MDB_CURRENT) {
6732 xflags = MDB_CURRENT|MDB_NOSPILL;
6734 mdb_xcursor_init1(mc, leaf);
6735 xflags = (flags & MDB_NODUPDATA) ?
6736 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6738 /* converted, write the original data first */
6740 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6744 /* Adjust other cursors pointing to mp */
6746 unsigned i = mc->mc_top;
6747 MDB_page *mp = mc->mc_pg[i];
6749 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6750 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6751 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6752 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6753 mdb_xcursor_init1(m2, leaf);
6757 /* we've done our job */
6760 ecount = mc->mc_xcursor->mx_db.md_entries;
6761 if (flags & MDB_APPENDDUP)
6762 xflags |= MDB_APPEND;
6763 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6764 if (flags & F_SUBDATA) {
6765 void *db = NODEDATA(leaf);
6766 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6768 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6770 /* Increment count unless we just replaced an existing item. */
6772 mc->mc_db->md_entries++;
6774 /* Invalidate txn if we created an empty sub-DB */
6777 /* If we succeeded and the key didn't exist before,
6778 * make sure the cursor is marked valid.
6780 mc->mc_flags |= C_INITIALIZED;
6782 if (flags & MDB_MULTIPLE) {
6785 /* let caller know how many succeeded, if any */
6786 data[1].mv_size = mcount;
6787 if (mcount < dcount) {
6788 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6789 insert_key = insert_data = 0;
6796 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6799 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6804 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6810 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6811 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6813 if (!(mc->mc_flags & C_INITIALIZED))
6816 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6817 return MDB_NOTFOUND;
6819 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6822 rc = mdb_cursor_touch(mc);
6826 mp = mc->mc_pg[mc->mc_top];
6829 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6831 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6832 if (flags & MDB_NODUPDATA) {
6833 /* mdb_cursor_del0() will subtract the final entry */
6834 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6836 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6837 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6839 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6842 /* If sub-DB still has entries, we're done */
6843 if (mc->mc_xcursor->mx_db.md_entries) {
6844 if (leaf->mn_flags & F_SUBDATA) {
6845 /* update subDB info */
6846 void *db = NODEDATA(leaf);
6847 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6850 /* shrink fake page */
6851 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6852 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6853 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6854 /* fix other sub-DB cursors pointed at this fake page */
6855 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6856 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6857 if (m2->mc_pg[mc->mc_top] == mp &&
6858 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6859 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6862 mc->mc_db->md_entries--;
6863 mc->mc_flags |= C_DEL;
6866 /* otherwise fall thru and delete the sub-DB */
6869 if (leaf->mn_flags & F_SUBDATA) {
6870 /* add all the child DB's pages to the free list */
6871 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6877 /* add overflow pages to free list */
6878 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6882 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6883 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6884 (rc = mdb_ovpage_free(mc, omp)))
6889 return mdb_cursor_del0(mc);
6892 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6896 /** Allocate and initialize new pages for a database.
6897 * @param[in] mc a cursor on the database being added to.
6898 * @param[in] flags flags defining what type of page is being allocated.
6899 * @param[in] num the number of pages to allocate. This is usually 1,
6900 * unless allocating overflow pages for a large record.
6901 * @param[out] mp Address of a page, or NULL on failure.
6902 * @return 0 on success, non-zero on failure.
6905 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6910 if ((rc = mdb_page_alloc(mc, num, &np)))
6912 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6913 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6914 np->mp_flags = flags | P_DIRTY;
6915 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6916 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6919 mc->mc_db->md_branch_pages++;
6920 else if (IS_LEAF(np))
6921 mc->mc_db->md_leaf_pages++;
6922 else if (IS_OVERFLOW(np)) {
6923 mc->mc_db->md_overflow_pages += num;
6931 /** Calculate the size of a leaf node.
6932 * The size depends on the environment's page size; if a data item
6933 * is too large it will be put onto an overflow page and the node
6934 * size will only include the key and not the data. Sizes are always
6935 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6936 * of the #MDB_node headers.
6937 * @param[in] env The environment handle.
6938 * @param[in] key The key for the node.
6939 * @param[in] data The data for the node.
6940 * @return The number of bytes needed to store the node.
6943 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6947 sz = LEAFSIZE(key, data);
6948 if (sz > env->me_nodemax) {
6949 /* put on overflow page */
6950 sz -= data->mv_size - sizeof(pgno_t);
6953 return EVEN(sz + sizeof(indx_t));
6956 /** Calculate the size of a branch node.
6957 * The size should depend on the environment's page size but since
6958 * we currently don't support spilling large keys onto overflow
6959 * pages, it's simply the size of the #MDB_node header plus the
6960 * size of the key. Sizes are always rounded up to an even number
6961 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6962 * @param[in] env The environment handle.
6963 * @param[in] key The key for the node.
6964 * @return The number of bytes needed to store the node.
6967 mdb_branch_size(MDB_env *env, MDB_val *key)
6972 if (sz > env->me_nodemax) {
6973 /* put on overflow page */
6974 /* not implemented */
6975 /* sz -= key->size - sizeof(pgno_t); */
6978 return sz + sizeof(indx_t);
6981 /** Add a node to the page pointed to by the cursor.
6982 * @param[in] mc The cursor for this operation.
6983 * @param[in] indx The index on the page where the new node should be added.
6984 * @param[in] key The key for the new node.
6985 * @param[in] data The data for the new node, if any.
6986 * @param[in] pgno The page number, if adding a branch node.
6987 * @param[in] flags Flags for the node.
6988 * @return 0 on success, non-zero on failure. Possible errors are:
6990 * <li>ENOMEM - failed to allocate overflow pages for the node.
6991 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6992 * should never happen since all callers already calculate the
6993 * page's free space before calling this function.
6997 mdb_node_add(MDB_cursor *mc, indx_t indx,
6998 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7001 size_t node_size = NODESIZE;
7005 MDB_page *mp = mc->mc_pg[mc->mc_top];
7006 MDB_page *ofp = NULL; /* overflow page */
7009 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7011 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7012 IS_LEAF(mp) ? "leaf" : "branch",
7013 IS_SUBP(mp) ? "sub-" : "",
7014 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7015 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7018 /* Move higher keys up one slot. */
7019 int ksize = mc->mc_db->md_pad, dif;
7020 char *ptr = LEAF2KEY(mp, indx, ksize);
7021 dif = NUMKEYS(mp) - indx;
7023 memmove(ptr+ksize, ptr, dif*ksize);
7024 /* insert new key */
7025 memcpy(ptr, key->mv_data, ksize);
7027 /* Just using these for counting */
7028 mp->mp_lower += sizeof(indx_t);
7029 mp->mp_upper -= ksize - sizeof(indx_t);
7033 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7035 node_size += key->mv_size;
7037 mdb_cassert(mc, data);
7038 if (F_ISSET(flags, F_BIGDATA)) {
7039 /* Data already on overflow page. */
7040 node_size += sizeof(pgno_t);
7041 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7042 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7044 /* Put data on overflow page. */
7045 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7046 data->mv_size, node_size+data->mv_size));
7047 node_size = EVEN(node_size + sizeof(pgno_t));
7048 if ((ssize_t)node_size > room)
7050 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7052 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7056 node_size += data->mv_size;
7059 node_size = EVEN(node_size);
7060 if ((ssize_t)node_size > room)
7064 /* Move higher pointers up one slot. */
7065 for (i = NUMKEYS(mp); i > indx; i--)
7066 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7068 /* Adjust free space offsets. */
7069 ofs = mp->mp_upper - node_size;
7070 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7071 mp->mp_ptrs[indx] = ofs;
7073 mp->mp_lower += sizeof(indx_t);
7075 /* Write the node data. */
7076 node = NODEPTR(mp, indx);
7077 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7078 node->mn_flags = flags;
7080 SETDSZ(node,data->mv_size);
7085 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7088 mdb_cassert(mc, key);
7090 if (F_ISSET(flags, F_BIGDATA))
7091 memcpy(node->mn_data + key->mv_size, data->mv_data,
7093 else if (F_ISSET(flags, MDB_RESERVE))
7094 data->mv_data = node->mn_data + key->mv_size;
7096 memcpy(node->mn_data + key->mv_size, data->mv_data,
7099 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
7101 if (F_ISSET(flags, MDB_RESERVE))
7102 data->mv_data = METADATA(ofp);
7104 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
7111 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7112 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7113 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7114 DPRINTF(("node size = %"Z"u", node_size));
7115 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7116 return MDB_PAGE_FULL;
7119 /** Delete the specified node from a page.
7120 * @param[in] mc Cursor pointing to the node to delete.
7121 * @param[in] ksize The size of a node. Only used if the page is
7122 * part of a #MDB_DUPFIXED database.
7125 mdb_node_del(MDB_cursor *mc, int ksize)
7127 MDB_page *mp = mc->mc_pg[mc->mc_top];
7128 indx_t indx = mc->mc_ki[mc->mc_top];
7130 indx_t i, j, numkeys, ptr;
7134 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7135 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7136 numkeys = NUMKEYS(mp);
7137 mdb_cassert(mc, indx < numkeys);
7140 int x = numkeys - 1 - indx;
7141 base = LEAF2KEY(mp, indx, ksize);
7143 memmove(base, base + ksize, x * ksize);
7144 mp->mp_lower -= sizeof(indx_t);
7145 mp->mp_upper += ksize - sizeof(indx_t);
7149 node = NODEPTR(mp, indx);
7150 sz = NODESIZE + node->mn_ksize;
7152 if (F_ISSET(node->mn_flags, F_BIGDATA))
7153 sz += sizeof(pgno_t);
7155 sz += NODEDSZ(node);
7159 ptr = mp->mp_ptrs[indx];
7160 for (i = j = 0; i < numkeys; i++) {
7162 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7163 if (mp->mp_ptrs[i] < ptr)
7164 mp->mp_ptrs[j] += sz;
7169 base = (char *)mp + mp->mp_upper + PAGEBASE;
7170 memmove(base + sz, base, ptr - mp->mp_upper);
7172 mp->mp_lower -= sizeof(indx_t);
7176 /** Compact the main page after deleting a node on a subpage.
7177 * @param[in] mp The main page to operate on.
7178 * @param[in] indx The index of the subpage on the main page.
7181 mdb_node_shrink(MDB_page *mp, indx_t indx)
7187 indx_t i, numkeys, ptr;
7189 node = NODEPTR(mp, indx);
7190 sp = (MDB_page *)NODEDATA(node);
7191 delta = SIZELEFT(sp);
7192 xp = (MDB_page *)((char *)sp + delta);
7194 /* shift subpage upward */
7196 nsize = NUMKEYS(sp) * sp->mp_pad;
7198 return; /* do not make the node uneven-sized */
7199 memmove(METADATA(xp), METADATA(sp), nsize);
7202 numkeys = NUMKEYS(sp);
7203 for (i=numkeys-1; i>=0; i--)
7204 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7206 xp->mp_upper = sp->mp_lower;
7207 xp->mp_lower = sp->mp_lower;
7208 xp->mp_flags = sp->mp_flags;
7209 xp->mp_pad = sp->mp_pad;
7210 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7212 nsize = NODEDSZ(node) - delta;
7213 SETDSZ(node, nsize);
7215 /* shift lower nodes upward */
7216 ptr = mp->mp_ptrs[indx];
7217 numkeys = NUMKEYS(mp);
7218 for (i = 0; i < numkeys; i++) {
7219 if (mp->mp_ptrs[i] <= ptr)
7220 mp->mp_ptrs[i] += delta;
7223 base = (char *)mp + mp->mp_upper + PAGEBASE;
7224 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7225 mp->mp_upper += delta;
7228 /** Initial setup of a sorted-dups cursor.
7229 * Sorted duplicates are implemented as a sub-database for the given key.
7230 * The duplicate data items are actually keys of the sub-database.
7231 * Operations on the duplicate data items are performed using a sub-cursor
7232 * initialized when the sub-database is first accessed. This function does
7233 * the preliminary setup of the sub-cursor, filling in the fields that
7234 * depend only on the parent DB.
7235 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7238 mdb_xcursor_init0(MDB_cursor *mc)
7240 MDB_xcursor *mx = mc->mc_xcursor;
7242 mx->mx_cursor.mc_xcursor = NULL;
7243 mx->mx_cursor.mc_txn = mc->mc_txn;
7244 mx->mx_cursor.mc_db = &mx->mx_db;
7245 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7246 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7247 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7248 mx->mx_cursor.mc_snum = 0;
7249 mx->mx_cursor.mc_top = 0;
7250 mx->mx_cursor.mc_flags = C_SUB;
7251 mx->mx_dbx.md_name.mv_size = 0;
7252 mx->mx_dbx.md_name.mv_data = NULL;
7253 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7254 mx->mx_dbx.md_dcmp = NULL;
7255 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7258 /** Final setup of a sorted-dups cursor.
7259 * Sets up the fields that depend on the data from the main cursor.
7260 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7261 * @param[in] node The data containing the #MDB_db record for the
7262 * sorted-dup database.
7265 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7267 MDB_xcursor *mx = mc->mc_xcursor;
7269 if (node->mn_flags & F_SUBDATA) {
7270 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7271 mx->mx_cursor.mc_pg[0] = 0;
7272 mx->mx_cursor.mc_snum = 0;
7273 mx->mx_cursor.mc_top = 0;
7274 mx->mx_cursor.mc_flags = C_SUB;
7276 MDB_page *fp = NODEDATA(node);
7277 mx->mx_db.md_pad = 0;
7278 mx->mx_db.md_flags = 0;
7279 mx->mx_db.md_depth = 1;
7280 mx->mx_db.md_branch_pages = 0;
7281 mx->mx_db.md_leaf_pages = 1;
7282 mx->mx_db.md_overflow_pages = 0;
7283 mx->mx_db.md_entries = NUMKEYS(fp);
7284 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7285 mx->mx_cursor.mc_snum = 1;
7286 mx->mx_cursor.mc_top = 0;
7287 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7288 mx->mx_cursor.mc_pg[0] = fp;
7289 mx->mx_cursor.mc_ki[0] = 0;
7290 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7291 mx->mx_db.md_flags = MDB_DUPFIXED;
7292 mx->mx_db.md_pad = fp->mp_pad;
7293 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7294 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7297 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7298 mx->mx_db.md_root));
7299 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7300 #if UINT_MAX < SIZE_MAX
7301 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7302 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7306 /** Initialize a cursor for a given transaction and database. */
7308 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7311 mc->mc_backup = NULL;
7314 mc->mc_db = &txn->mt_dbs[dbi];
7315 mc->mc_dbx = &txn->mt_dbxs[dbi];
7316 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7322 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7323 mdb_tassert(txn, mx != NULL);
7324 mc->mc_xcursor = mx;
7325 mdb_xcursor_init0(mc);
7327 mc->mc_xcursor = NULL;
7329 if (*mc->mc_dbflag & DB_STALE) {
7330 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7335 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7338 size_t size = sizeof(MDB_cursor);
7340 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7343 if (txn->mt_flags & MDB_TXN_ERROR)
7346 /* Allow read access to the freelist */
7347 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7350 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7351 size += sizeof(MDB_xcursor);
7353 if ((mc = malloc(size)) != NULL) {
7354 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7355 if (txn->mt_cursors) {
7356 mc->mc_next = txn->mt_cursors[dbi];
7357 txn->mt_cursors[dbi] = mc;
7358 mc->mc_flags |= C_UNTRACK;
7370 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7372 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7375 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7378 if (txn->mt_flags & MDB_TXN_ERROR)
7381 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7385 /* Return the count of duplicate data items for the current key */
7387 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7391 if (mc == NULL || countp == NULL)
7394 if (mc->mc_xcursor == NULL)
7395 return MDB_INCOMPATIBLE;
7397 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7400 if (!(mc->mc_flags & C_INITIALIZED))
7403 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7404 return MDB_NOTFOUND;
7406 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7407 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7410 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7413 *countp = mc->mc_xcursor->mx_db.md_entries;
7419 mdb_cursor_close(MDB_cursor *mc)
7421 if (mc && !mc->mc_backup) {
7422 /* remove from txn, if tracked */
7423 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7424 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7425 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7427 *prev = mc->mc_next;
7434 mdb_cursor_txn(MDB_cursor *mc)
7436 if (!mc) return NULL;
7441 mdb_cursor_dbi(MDB_cursor *mc)
7446 /** Replace the key for a branch node with a new key.
7447 * @param[in] mc Cursor pointing to the node to operate on.
7448 * @param[in] key The new key to use.
7449 * @return 0 on success, non-zero on failure.
7452 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7458 int delta, ksize, oksize;
7459 indx_t ptr, i, numkeys, indx;
7462 indx = mc->mc_ki[mc->mc_top];
7463 mp = mc->mc_pg[mc->mc_top];
7464 node = NODEPTR(mp, indx);
7465 ptr = mp->mp_ptrs[indx];
7469 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7470 k2.mv_data = NODEKEY(node);
7471 k2.mv_size = node->mn_ksize;
7472 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7474 mdb_dkey(&k2, kbuf2),
7480 /* Sizes must be 2-byte aligned. */
7481 ksize = EVEN(key->mv_size);
7482 oksize = EVEN(node->mn_ksize);
7483 delta = ksize - oksize;
7485 /* Shift node contents if EVEN(key length) changed. */
7487 if (delta > 0 && SIZELEFT(mp) < delta) {
7489 /* not enough space left, do a delete and split */
7490 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7491 pgno = NODEPGNO(node);
7492 mdb_node_del(mc, 0);
7493 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7496 numkeys = NUMKEYS(mp);
7497 for (i = 0; i < numkeys; i++) {
7498 if (mp->mp_ptrs[i] <= ptr)
7499 mp->mp_ptrs[i] -= delta;
7502 base = (char *)mp + mp->mp_upper + PAGEBASE;
7503 len = ptr - mp->mp_upper + NODESIZE;
7504 memmove(base - delta, base, len);
7505 mp->mp_upper -= delta;
7507 node = NODEPTR(mp, indx);
7510 /* But even if no shift was needed, update ksize */
7511 if (node->mn_ksize != key->mv_size)
7512 node->mn_ksize = key->mv_size;
7515 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7521 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7523 /** Move a node from csrc to cdst.
7526 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7533 unsigned short flags;
7537 /* Mark src and dst as dirty. */
7538 if ((rc = mdb_page_touch(csrc)) ||
7539 (rc = mdb_page_touch(cdst)))
7542 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7543 key.mv_size = csrc->mc_db->md_pad;
7544 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7546 data.mv_data = NULL;
7550 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7551 mdb_cassert(csrc, !((size_t)srcnode & 1));
7552 srcpg = NODEPGNO(srcnode);
7553 flags = srcnode->mn_flags;
7554 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7555 unsigned int snum = csrc->mc_snum;
7557 /* must find the lowest key below src */
7558 rc = mdb_page_search_lowest(csrc);
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], 0, key.mv_size);
7565 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7566 key.mv_size = NODEKSZ(s2);
7567 key.mv_data = NODEKEY(s2);
7569 csrc->mc_snum = snum--;
7570 csrc->mc_top = snum;
7572 key.mv_size = NODEKSZ(srcnode);
7573 key.mv_data = NODEKEY(srcnode);
7575 data.mv_size = NODEDSZ(srcnode);
7576 data.mv_data = NODEDATA(srcnode);
7578 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7579 unsigned int snum = cdst->mc_snum;
7582 /* must find the lowest key below dst */
7583 mdb_cursor_copy(cdst, &mn);
7584 rc = mdb_page_search_lowest(&mn);
7587 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7588 bkey.mv_size = mn.mc_db->md_pad;
7589 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7591 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7592 bkey.mv_size = NODEKSZ(s2);
7593 bkey.mv_data = NODEKEY(s2);
7595 mn.mc_snum = snum--;
7598 rc = mdb_update_key(&mn, &bkey);
7603 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7604 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7605 csrc->mc_ki[csrc->mc_top],
7607 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7608 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7610 /* Add the node to the destination page.
7612 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7613 if (rc != MDB_SUCCESS)
7616 /* Delete the node from the source page.
7618 mdb_node_del(csrc, key.mv_size);
7621 /* Adjust other cursors pointing to mp */
7622 MDB_cursor *m2, *m3;
7623 MDB_dbi dbi = csrc->mc_dbi;
7624 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7626 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7627 if (csrc->mc_flags & C_SUB)
7628 m3 = &m2->mc_xcursor->mx_cursor;
7631 if (m3 == csrc) continue;
7632 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7633 csrc->mc_ki[csrc->mc_top]) {
7634 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7635 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7640 /* Update the parent separators.
7642 if (csrc->mc_ki[csrc->mc_top] == 0) {
7643 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7644 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7645 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7647 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7648 key.mv_size = NODEKSZ(srcnode);
7649 key.mv_data = NODEKEY(srcnode);
7651 DPRINTF(("update separator for source page %"Z"u to [%s]",
7652 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7653 mdb_cursor_copy(csrc, &mn);
7656 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7659 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7661 indx_t ix = csrc->mc_ki[csrc->mc_top];
7662 nullkey.mv_size = 0;
7663 csrc->mc_ki[csrc->mc_top] = 0;
7664 rc = mdb_update_key(csrc, &nullkey);
7665 csrc->mc_ki[csrc->mc_top] = ix;
7666 mdb_cassert(csrc, rc == MDB_SUCCESS);
7670 if (cdst->mc_ki[cdst->mc_top] == 0) {
7671 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7672 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7673 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7675 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7676 key.mv_size = NODEKSZ(srcnode);
7677 key.mv_data = NODEKEY(srcnode);
7679 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7680 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7681 mdb_cursor_copy(cdst, &mn);
7684 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7687 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7689 indx_t ix = cdst->mc_ki[cdst->mc_top];
7690 nullkey.mv_size = 0;
7691 cdst->mc_ki[cdst->mc_top] = 0;
7692 rc = mdb_update_key(cdst, &nullkey);
7693 cdst->mc_ki[cdst->mc_top] = ix;
7694 mdb_cassert(cdst, rc == MDB_SUCCESS);
7701 /** Merge one page into another.
7702 * The nodes from the page pointed to by \b csrc will
7703 * be copied to the page pointed to by \b cdst and then
7704 * the \b csrc page will be freed.
7705 * @param[in] csrc Cursor pointing to the source page.
7706 * @param[in] cdst Cursor pointing to the destination page.
7707 * @return 0 on success, non-zero on failure.
7710 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7712 MDB_page *psrc, *pdst;
7719 psrc = csrc->mc_pg[csrc->mc_top];
7720 pdst = cdst->mc_pg[cdst->mc_top];
7722 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7724 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7725 mdb_cassert(csrc, cdst->mc_snum > 1);
7727 /* Mark dst as dirty. */
7728 if ((rc = mdb_page_touch(cdst)))
7731 /* Move all nodes from src to dst.
7733 j = nkeys = NUMKEYS(pdst);
7734 if (IS_LEAF2(psrc)) {
7735 key.mv_size = csrc->mc_db->md_pad;
7736 key.mv_data = METADATA(psrc);
7737 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7738 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7739 if (rc != MDB_SUCCESS)
7741 key.mv_data = (char *)key.mv_data + key.mv_size;
7744 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7745 srcnode = NODEPTR(psrc, i);
7746 if (i == 0 && IS_BRANCH(psrc)) {
7749 mdb_cursor_copy(csrc, &mn);
7750 /* must find the lowest key below src */
7751 rc = mdb_page_search_lowest(&mn);
7754 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7755 key.mv_size = mn.mc_db->md_pad;
7756 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7758 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7759 key.mv_size = NODEKSZ(s2);
7760 key.mv_data = NODEKEY(s2);
7763 key.mv_size = srcnode->mn_ksize;
7764 key.mv_data = NODEKEY(srcnode);
7767 data.mv_size = NODEDSZ(srcnode);
7768 data.mv_data = NODEDATA(srcnode);
7769 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7770 if (rc != MDB_SUCCESS)
7775 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7776 pdst->mp_pgno, NUMKEYS(pdst),
7777 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7779 /* Unlink the src page from parent and add to free list.
7782 mdb_node_del(csrc, 0);
7783 if (csrc->mc_ki[csrc->mc_top] == 0) {
7785 rc = mdb_update_key(csrc, &key);
7793 psrc = csrc->mc_pg[csrc->mc_top];
7794 /* If not operating on FreeDB, allow this page to be reused
7795 * in this txn. Otherwise just add to free list.
7797 rc = mdb_page_loose(csrc, psrc);
7801 csrc->mc_db->md_leaf_pages--;
7803 csrc->mc_db->md_branch_pages--;
7805 /* Adjust other cursors pointing to mp */
7806 MDB_cursor *m2, *m3;
7807 MDB_dbi dbi = csrc->mc_dbi;
7809 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7810 if (csrc->mc_flags & C_SUB)
7811 m3 = &m2->mc_xcursor->mx_cursor;
7814 if (m3 == csrc) continue;
7815 if (m3->mc_snum < csrc->mc_snum) continue;
7816 if (m3->mc_pg[csrc->mc_top] == psrc) {
7817 m3->mc_pg[csrc->mc_top] = pdst;
7818 m3->mc_ki[csrc->mc_top] += nkeys;
7823 unsigned int snum = cdst->mc_snum;
7824 uint16_t depth = cdst->mc_db->md_depth;
7825 mdb_cursor_pop(cdst);
7826 rc = mdb_rebalance(cdst);
7827 /* Did the tree shrink? */
7828 if (depth > cdst->mc_db->md_depth)
7830 cdst->mc_snum = snum;
7831 cdst->mc_top = snum-1;
7836 /** Copy the contents of a cursor.
7837 * @param[in] csrc The cursor to copy from.
7838 * @param[out] cdst The cursor to copy to.
7841 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7845 cdst->mc_txn = csrc->mc_txn;
7846 cdst->mc_dbi = csrc->mc_dbi;
7847 cdst->mc_db = csrc->mc_db;
7848 cdst->mc_dbx = csrc->mc_dbx;
7849 cdst->mc_snum = csrc->mc_snum;
7850 cdst->mc_top = csrc->mc_top;
7851 cdst->mc_flags = csrc->mc_flags;
7853 for (i=0; i<csrc->mc_snum; i++) {
7854 cdst->mc_pg[i] = csrc->mc_pg[i];
7855 cdst->mc_ki[i] = csrc->mc_ki[i];
7859 /** Rebalance the tree after a delete operation.
7860 * @param[in] mc Cursor pointing to the page where rebalancing
7862 * @return 0 on success, non-zero on failure.
7865 mdb_rebalance(MDB_cursor *mc)
7869 unsigned int ptop, minkeys;
7873 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7874 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7875 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7876 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7877 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7879 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7880 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7881 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7882 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7886 if (mc->mc_snum < 2) {
7887 MDB_page *mp = mc->mc_pg[0];
7889 DPUTS("Can't rebalance a subpage, ignoring");
7892 if (NUMKEYS(mp) == 0) {
7893 DPUTS("tree is completely empty");
7894 mc->mc_db->md_root = P_INVALID;
7895 mc->mc_db->md_depth = 0;
7896 mc->mc_db->md_leaf_pages = 0;
7897 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7900 /* Adjust cursors pointing to mp */
7903 mc->mc_flags &= ~C_INITIALIZED;
7905 MDB_cursor *m2, *m3;
7906 MDB_dbi dbi = mc->mc_dbi;
7908 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7909 if (mc->mc_flags & C_SUB)
7910 m3 = &m2->mc_xcursor->mx_cursor;
7913 if (m3->mc_snum < mc->mc_snum) continue;
7914 if (m3->mc_pg[0] == mp) {
7917 m3->mc_flags &= ~C_INITIALIZED;
7921 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7923 DPUTS("collapsing root page!");
7924 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7927 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7928 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7931 mc->mc_db->md_depth--;
7932 mc->mc_db->md_branch_pages--;
7933 mc->mc_ki[0] = mc->mc_ki[1];
7934 for (i = 1; i<mc->mc_db->md_depth; i++) {
7935 mc->mc_pg[i] = mc->mc_pg[i+1];
7936 mc->mc_ki[i] = mc->mc_ki[i+1];
7939 /* Adjust other cursors pointing to mp */
7940 MDB_cursor *m2, *m3;
7941 MDB_dbi dbi = mc->mc_dbi;
7943 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7944 if (mc->mc_flags & C_SUB)
7945 m3 = &m2->mc_xcursor->mx_cursor;
7948 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7949 if (m3->mc_pg[0] == mp) {
7950 for (i=0; i<m3->mc_snum; i++) {
7951 m3->mc_pg[i] = m3->mc_pg[i+1];
7952 m3->mc_ki[i] = m3->mc_ki[i+1];
7960 DPUTS("root page doesn't need rebalancing");
7964 /* The parent (branch page) must have at least 2 pointers,
7965 * otherwise the tree is invalid.
7967 ptop = mc->mc_top-1;
7968 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7970 /* Leaf page fill factor is below the threshold.
7971 * Try to move keys from left or right neighbor, or
7972 * merge with a neighbor page.
7977 mdb_cursor_copy(mc, &mn);
7978 mn.mc_xcursor = NULL;
7980 oldki = mc->mc_ki[mc->mc_top];
7981 if (mc->mc_ki[ptop] == 0) {
7982 /* We're the leftmost leaf in our parent.
7984 DPUTS("reading right neighbor");
7986 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7987 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7990 mn.mc_ki[mn.mc_top] = 0;
7991 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7993 /* There is at least one neighbor to the left.
7995 DPUTS("reading left neighbor");
7997 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7998 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8001 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8002 mc->mc_ki[mc->mc_top] = 0;
8005 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8006 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8007 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8009 /* If the neighbor page is above threshold and has enough keys,
8010 * move one key from it. Otherwise we should try to merge them.
8011 * (A branch page must never have less than 2 keys.)
8013 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
8014 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8015 rc = mdb_node_move(&mn, mc);
8016 if (mc->mc_ki[ptop]) {
8020 if (mc->mc_ki[ptop] == 0) {
8021 rc = mdb_page_merge(&mn, mc);
8024 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8025 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8026 /* We want mdb_rebalance to find mn when doing fixups */
8027 if (mc->mc_flags & C_SUB) {
8028 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8029 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8030 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8032 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8033 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8035 rc = mdb_page_merge(mc, &mn);
8036 if (mc->mc_flags & C_SUB)
8037 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8039 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8040 mdb_cursor_copy(&mn, mc);
8042 mc->mc_flags &= ~C_EOF;
8044 mc->mc_ki[mc->mc_top] = oldki;
8048 /** Complete a delete operation started by #mdb_cursor_del(). */
8050 mdb_cursor_del0(MDB_cursor *mc)
8057 ki = mc->mc_ki[mc->mc_top];
8058 mdb_node_del(mc, mc->mc_db->md_pad);
8059 mc->mc_db->md_entries--;
8060 rc = mdb_rebalance(mc);
8062 if (rc == MDB_SUCCESS) {
8063 MDB_cursor *m2, *m3;
8064 MDB_dbi dbi = mc->mc_dbi;
8066 /* DB is totally empty now, just bail out.
8067 * Other cursors adjustments were already done
8068 * by mdb_rebalance and aren't needed here.
8073 mp = mc->mc_pg[mc->mc_top];
8074 nkeys = NUMKEYS(mp);
8076 /* if mc points past last node in page, find next sibling */
8077 if (mc->mc_ki[mc->mc_top] >= nkeys) {
8078 rc = mdb_cursor_sibling(mc, 1);
8079 if (rc == MDB_NOTFOUND) {
8080 mc->mc_flags |= C_EOF;
8085 /* Adjust other cursors pointing to mp */
8086 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8087 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8088 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8090 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8092 if (m3->mc_pg[mc->mc_top] == mp) {
8093 if (m3->mc_ki[mc->mc_top] >= ki) {
8094 m3->mc_flags |= C_DEL;
8095 if (m3->mc_ki[mc->mc_top] > ki)
8096 m3->mc_ki[mc->mc_top]--;
8097 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8098 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8100 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8101 rc = mdb_cursor_sibling(m3, 1);
8102 if (rc == MDB_NOTFOUND) {
8103 m3->mc_flags |= C_EOF;
8109 mc->mc_flags |= C_DEL;
8113 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8118 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8119 MDB_val *key, MDB_val *data)
8121 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8124 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
8125 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8127 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8128 /* must ignore any data */
8132 return mdb_del0(txn, dbi, key, data, 0);
8136 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8137 MDB_val *key, MDB_val *data, unsigned flags)
8142 MDB_val rdata, *xdata;
8146 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8148 mdb_cursor_init(&mc, txn, dbi, &mx);
8157 flags |= MDB_NODUPDATA;
8159 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8161 /* let mdb_page_split know about this cursor if needed:
8162 * delete will trigger a rebalance; if it needs to move
8163 * a node from one page to another, it will have to
8164 * update the parent's separator key(s). If the new sepkey
8165 * is larger than the current one, the parent page may
8166 * run out of space, triggering a split. We need this
8167 * cursor to be consistent until the end of the rebalance.
8169 mc.mc_flags |= C_UNTRACK;
8170 mc.mc_next = txn->mt_cursors[dbi];
8171 txn->mt_cursors[dbi] = &mc;
8172 rc = mdb_cursor_del(&mc, flags);
8173 txn->mt_cursors[dbi] = mc.mc_next;
8178 /** Split a page and insert a new node.
8179 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8180 * The cursor will be updated to point to the actual page and index where
8181 * the node got inserted after the split.
8182 * @param[in] newkey The key for the newly inserted node.
8183 * @param[in] newdata The data for the newly inserted node.
8184 * @param[in] newpgno The page number, if the new node is a branch node.
8185 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8186 * @return 0 on success, non-zero on failure.
8189 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8190 unsigned int nflags)
8193 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8196 int i, j, split_indx, nkeys, pmax;
8197 MDB_env *env = mc->mc_txn->mt_env;
8199 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8200 MDB_page *copy = NULL;
8201 MDB_page *mp, *rp, *pp;
8206 mp = mc->mc_pg[mc->mc_top];
8207 newindx = mc->mc_ki[mc->mc_top];
8208 nkeys = NUMKEYS(mp);
8210 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8211 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8212 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8214 /* Create a right sibling. */
8215 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8217 rp->mp_pad = mp->mp_pad;
8218 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8220 if (mc->mc_snum < 2) {
8221 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8223 /* shift current top to make room for new parent */
8224 mc->mc_pg[1] = mc->mc_pg[0];
8225 mc->mc_ki[1] = mc->mc_ki[0];
8228 mc->mc_db->md_root = pp->mp_pgno;
8229 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8230 mc->mc_db->md_depth++;
8233 /* Add left (implicit) pointer. */
8234 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8235 /* undo the pre-push */
8236 mc->mc_pg[0] = mc->mc_pg[1];
8237 mc->mc_ki[0] = mc->mc_ki[1];
8238 mc->mc_db->md_root = mp->mp_pgno;
8239 mc->mc_db->md_depth--;
8246 ptop = mc->mc_top-1;
8247 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8250 mc->mc_flags |= C_SPLITTING;
8251 mdb_cursor_copy(mc, &mn);
8252 mn.mc_pg[mn.mc_top] = rp;
8253 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8255 if (nflags & MDB_APPEND) {
8256 mn.mc_ki[mn.mc_top] = 0;
8258 split_indx = newindx;
8262 split_indx = (nkeys+1) / 2;
8267 unsigned int lsize, rsize, ksize;
8268 /* Move half of the keys to the right sibling */
8269 x = mc->mc_ki[mc->mc_top] - split_indx;
8270 ksize = mc->mc_db->md_pad;
8271 split = LEAF2KEY(mp, split_indx, ksize);
8272 rsize = (nkeys - split_indx) * ksize;
8273 lsize = (nkeys - split_indx) * sizeof(indx_t);
8274 mp->mp_lower -= lsize;
8275 rp->mp_lower += lsize;
8276 mp->mp_upper += rsize - lsize;
8277 rp->mp_upper -= rsize - lsize;
8278 sepkey.mv_size = ksize;
8279 if (newindx == split_indx) {
8280 sepkey.mv_data = newkey->mv_data;
8282 sepkey.mv_data = split;
8285 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8286 memcpy(rp->mp_ptrs, split, rsize);
8287 sepkey.mv_data = rp->mp_ptrs;
8288 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8289 memcpy(ins, newkey->mv_data, ksize);
8290 mp->mp_lower += sizeof(indx_t);
8291 mp->mp_upper -= ksize - sizeof(indx_t);
8294 memcpy(rp->mp_ptrs, split, x * ksize);
8295 ins = LEAF2KEY(rp, x, ksize);
8296 memcpy(ins, newkey->mv_data, ksize);
8297 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8298 rp->mp_lower += sizeof(indx_t);
8299 rp->mp_upper -= ksize - sizeof(indx_t);
8300 mc->mc_ki[mc->mc_top] = x;
8301 mc->mc_pg[mc->mc_top] = rp;
8304 int psize, nsize, k;
8305 /* Maximum free space in an empty page */
8306 pmax = env->me_psize - PAGEHDRSZ;
8308 nsize = mdb_leaf_size(env, newkey, newdata);
8310 nsize = mdb_branch_size(env, newkey);
8311 nsize = EVEN(nsize);
8313 /* grab a page to hold a temporary copy */
8314 copy = mdb_page_malloc(mc->mc_txn, 1);
8319 copy->mp_pgno = mp->mp_pgno;
8320 copy->mp_flags = mp->mp_flags;
8321 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8322 copy->mp_upper = env->me_psize - PAGEBASE;
8324 /* prepare to insert */
8325 for (i=0, j=0; i<nkeys; i++) {
8327 copy->mp_ptrs[j++] = 0;
8329 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8332 /* When items are relatively large the split point needs
8333 * to be checked, because being off-by-one will make the
8334 * difference between success or failure in mdb_node_add.
8336 * It's also relevant if a page happens to be laid out
8337 * such that one half of its nodes are all "small" and
8338 * the other half of its nodes are "large." If the new
8339 * item is also "large" and falls on the half with
8340 * "large" nodes, it also may not fit.
8342 * As a final tweak, if the new item goes on the last
8343 * spot on the page (and thus, onto the new page), bias
8344 * the split so the new page is emptier than the old page.
8345 * This yields better packing during sequential inserts.
8347 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8348 /* Find split point */
8350 if (newindx <= split_indx || newindx >= nkeys) {
8352 k = newindx >= nkeys ? nkeys : split_indx+2;
8357 for (; i!=k; i+=j) {
8362 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8363 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8365 if (F_ISSET(node->mn_flags, F_BIGDATA))
8366 psize += sizeof(pgno_t);
8368 psize += NODEDSZ(node);
8370 psize = EVEN(psize);
8372 if (psize > pmax || i == k-j) {
8373 split_indx = i + (j<0);
8378 if (split_indx == newindx) {
8379 sepkey.mv_size = newkey->mv_size;
8380 sepkey.mv_data = newkey->mv_data;
8382 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8383 sepkey.mv_size = node->mn_ksize;
8384 sepkey.mv_data = NODEKEY(node);
8389 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8391 /* Copy separator key to the parent.
8393 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8397 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8402 if (mn.mc_snum == mc->mc_snum) {
8403 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8404 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8405 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8406 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8411 /* Right page might now have changed parent.
8412 * Check if left page also changed parent.
8414 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8415 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8416 for (i=0; i<ptop; i++) {
8417 mc->mc_pg[i] = mn.mc_pg[i];
8418 mc->mc_ki[i] = mn.mc_ki[i];
8420 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8421 if (mn.mc_ki[ptop]) {
8422 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8424 /* find right page's left sibling */
8425 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8426 mdb_cursor_sibling(mc, 0);
8431 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8434 mc->mc_flags ^= C_SPLITTING;
8435 if (rc != MDB_SUCCESS) {
8438 if (nflags & MDB_APPEND) {
8439 mc->mc_pg[mc->mc_top] = rp;
8440 mc->mc_ki[mc->mc_top] = 0;
8441 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8444 for (i=0; i<mc->mc_top; i++)
8445 mc->mc_ki[i] = mn.mc_ki[i];
8446 } else if (!IS_LEAF2(mp)) {
8448 mc->mc_pg[mc->mc_top] = rp;
8453 rkey.mv_data = newkey->mv_data;
8454 rkey.mv_size = newkey->mv_size;
8460 /* Update index for the new key. */
8461 mc->mc_ki[mc->mc_top] = j;
8463 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8464 rkey.mv_data = NODEKEY(node);
8465 rkey.mv_size = node->mn_ksize;
8467 xdata.mv_data = NODEDATA(node);
8468 xdata.mv_size = NODEDSZ(node);
8471 pgno = NODEPGNO(node);
8472 flags = node->mn_flags;
8475 if (!IS_LEAF(mp) && j == 0) {
8476 /* First branch index doesn't need key data. */
8480 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8486 mc->mc_pg[mc->mc_top] = copy;
8491 } while (i != split_indx);
8493 nkeys = NUMKEYS(copy);
8494 for (i=0; i<nkeys; i++)
8495 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8496 mp->mp_lower = copy->mp_lower;
8497 mp->mp_upper = copy->mp_upper;
8498 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8499 env->me_psize - copy->mp_upper - PAGEBASE);
8501 /* reset back to original page */
8502 if (newindx < split_indx) {
8503 mc->mc_pg[mc->mc_top] = mp;
8504 if (nflags & MDB_RESERVE) {
8505 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8506 if (!(node->mn_flags & F_BIGDATA))
8507 newdata->mv_data = NODEDATA(node);
8510 mc->mc_pg[mc->mc_top] = rp;
8512 /* Make sure mc_ki is still valid.
8514 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8515 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8516 for (i=0; i<=ptop; i++) {
8517 mc->mc_pg[i] = mn.mc_pg[i];
8518 mc->mc_ki[i] = mn.mc_ki[i];
8525 /* Adjust other cursors pointing to mp */
8526 MDB_cursor *m2, *m3;
8527 MDB_dbi dbi = mc->mc_dbi;
8528 int fixup = NUMKEYS(mp);
8530 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8531 if (mc->mc_flags & C_SUB)
8532 m3 = &m2->mc_xcursor->mx_cursor;
8537 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8539 if (m3->mc_flags & C_SPLITTING)
8544 for (k=m3->mc_top; k>=0; k--) {
8545 m3->mc_ki[k+1] = m3->mc_ki[k];
8546 m3->mc_pg[k+1] = m3->mc_pg[k];
8548 if (m3->mc_ki[0] >= split_indx) {
8553 m3->mc_pg[0] = mc->mc_pg[0];
8557 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8558 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8559 m3->mc_ki[mc->mc_top]++;
8560 if (m3->mc_ki[mc->mc_top] >= fixup) {
8561 m3->mc_pg[mc->mc_top] = rp;
8562 m3->mc_ki[mc->mc_top] -= fixup;
8563 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8565 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8566 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8571 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8574 if (copy) /* tmp page */
8575 mdb_page_free(env, copy);
8577 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8582 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8583 MDB_val *key, MDB_val *data, unsigned int flags)
8588 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8591 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8594 mdb_cursor_init(&mc, txn, dbi, &mx);
8595 return mdb_cursor_put(&mc, key, data, flags);
8599 #define MDB_WBUF (1024*1024)
8602 /** State needed for a compacting copy. */
8603 typedef struct mdb_copy {
8604 pthread_mutex_t mc_mutex;
8605 pthread_cond_t mc_cond;
8612 pgno_t mc_next_pgno;
8615 volatile int mc_new;
8620 /** Dedicated writer thread for compacting copy. */
8621 static THREAD_RET ESECT
8622 mdb_env_copythr(void *arg)
8626 int toggle = 0, wsize, rc;
8629 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8632 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8635 pthread_mutex_lock(&my->mc_mutex);
8637 pthread_cond_signal(&my->mc_cond);
8640 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8641 if (my->mc_new < 0) {
8646 wsize = my->mc_wlen[toggle];
8647 ptr = my->mc_wbuf[toggle];
8650 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8654 } else if (len > 0) {
8668 /* If there's an overflow page tail, write it too */
8669 if (my->mc_olen[toggle]) {
8670 wsize = my->mc_olen[toggle];
8671 ptr = my->mc_over[toggle];
8672 my->mc_olen[toggle] = 0;
8675 my->mc_wlen[toggle] = 0;
8677 pthread_cond_signal(&my->mc_cond);
8679 pthread_cond_signal(&my->mc_cond);
8680 pthread_mutex_unlock(&my->mc_mutex);
8681 return (THREAD_RET)0;
8685 /** Tell the writer thread there's a buffer ready to write */
8687 mdb_env_cthr_toggle(mdb_copy *my, int st)
8689 int toggle = my->mc_toggle ^ 1;
8690 pthread_mutex_lock(&my->mc_mutex);
8691 if (my->mc_status) {
8692 pthread_mutex_unlock(&my->mc_mutex);
8693 return my->mc_status;
8695 while (my->mc_new == 1)
8696 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8698 my->mc_toggle = toggle;
8699 pthread_cond_signal(&my->mc_cond);
8700 pthread_mutex_unlock(&my->mc_mutex);
8704 /** Depth-first tree traversal for compacting copy. */
8706 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8709 MDB_txn *txn = my->mc_txn;
8711 MDB_page *mo, *mp, *leaf;
8716 /* Empty DB, nothing to do */
8717 if (*pg == P_INVALID)
8724 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8727 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8731 /* Make cursor pages writable */
8732 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8736 for (i=0; i<mc.mc_top; i++) {
8737 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8738 mc.mc_pg[i] = (MDB_page *)ptr;
8739 ptr += my->mc_env->me_psize;
8742 /* This is writable space for a leaf page. Usually not needed. */
8743 leaf = (MDB_page *)ptr;
8745 toggle = my->mc_toggle;
8746 while (mc.mc_snum > 0) {
8748 mp = mc.mc_pg[mc.mc_top];
8752 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8753 for (i=0; i<n; i++) {
8754 ni = NODEPTR(mp, i);
8755 if (ni->mn_flags & F_BIGDATA) {
8759 /* Need writable leaf */
8761 mc.mc_pg[mc.mc_top] = leaf;
8762 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8764 ni = NODEPTR(mp, i);
8767 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8768 rc = mdb_page_get(txn, pg, &omp, NULL);
8771 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8772 rc = mdb_env_cthr_toggle(my, 1);
8775 toggle = my->mc_toggle;
8777 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8778 memcpy(mo, omp, my->mc_env->me_psize);
8779 mo->mp_pgno = my->mc_next_pgno;
8780 my->mc_next_pgno += omp->mp_pages;
8781 my->mc_wlen[toggle] += my->mc_env->me_psize;
8782 if (omp->mp_pages > 1) {
8783 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8784 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8785 rc = mdb_env_cthr_toggle(my, 1);
8788 toggle = my->mc_toggle;
8790 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8791 } else if (ni->mn_flags & F_SUBDATA) {
8794 /* Need writable leaf */
8796 mc.mc_pg[mc.mc_top] = leaf;
8797 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8799 ni = NODEPTR(mp, i);
8802 memcpy(&db, NODEDATA(ni), sizeof(db));
8803 my->mc_toggle = toggle;
8804 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8807 toggle = my->mc_toggle;
8808 memcpy(NODEDATA(ni), &db, sizeof(db));
8813 mc.mc_ki[mc.mc_top]++;
8814 if (mc.mc_ki[mc.mc_top] < n) {
8817 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8819 rc = mdb_page_get(txn, pg, &mp, NULL);
8824 mc.mc_ki[mc.mc_top] = 0;
8825 if (IS_BRANCH(mp)) {
8826 /* Whenever we advance to a sibling branch page,
8827 * we must proceed all the way down to its first leaf.
8829 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8832 mc.mc_pg[mc.mc_top] = mp;
8836 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8837 rc = mdb_env_cthr_toggle(my, 1);
8840 toggle = my->mc_toggle;
8842 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8843 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8844 mo->mp_pgno = my->mc_next_pgno++;
8845 my->mc_wlen[toggle] += my->mc_env->me_psize;
8847 /* Update parent if there is one */
8848 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8849 SETPGNO(ni, mo->mp_pgno);
8850 mdb_cursor_pop(&mc);
8852 /* Otherwise we're done */
8862 /** Copy environment with compaction. */
8864 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8869 MDB_txn *txn = NULL;
8874 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8875 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8876 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8877 if (my.mc_wbuf[0] == NULL)
8880 pthread_mutex_init(&my.mc_mutex, NULL);
8881 pthread_cond_init(&my.mc_cond, NULL);
8882 #ifdef HAVE_MEMALIGN
8883 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8884 if (my.mc_wbuf[0] == NULL)
8887 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8892 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8893 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8898 my.mc_next_pgno = 2;
8904 THREAD_CREATE(thr, mdb_env_copythr, &my);
8906 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8910 mp = (MDB_page *)my.mc_wbuf[0];
8911 memset(mp, 0, 2*env->me_psize);
8913 mp->mp_flags = P_META;
8914 mm = (MDB_meta *)METADATA(mp);
8915 mdb_env_init_meta0(env, mm);
8916 mm->mm_address = env->me_metas[0]->mm_address;
8918 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8920 mp->mp_flags = P_META;
8921 *(MDB_meta *)METADATA(mp) = *mm;
8922 mm = (MDB_meta *)METADATA(mp);
8924 /* Count the number of free pages, subtract from lastpg to find
8925 * number of active pages
8928 MDB_ID freecount = 0;
8931 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8932 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8933 freecount += *(MDB_ID *)data.mv_data;
8934 freecount += txn->mt_dbs[0].md_branch_pages +
8935 txn->mt_dbs[0].md_leaf_pages +
8936 txn->mt_dbs[0].md_overflow_pages;
8938 /* Set metapage 1 */
8939 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8940 mm->mm_dbs[1] = txn->mt_dbs[1];
8941 if (mm->mm_last_pg > 1) {
8942 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8945 mm->mm_dbs[1].md_root = P_INVALID;
8948 my.mc_wlen[0] = env->me_psize * 2;
8950 pthread_mutex_lock(&my.mc_mutex);
8952 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8953 pthread_mutex_unlock(&my.mc_mutex);
8954 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8955 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8956 rc = mdb_env_cthr_toggle(&my, 1);
8957 mdb_env_cthr_toggle(&my, -1);
8958 pthread_mutex_lock(&my.mc_mutex);
8960 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8961 pthread_mutex_unlock(&my.mc_mutex);
8966 CloseHandle(my.mc_cond);
8967 CloseHandle(my.mc_mutex);
8968 _aligned_free(my.mc_wbuf[0]);
8970 pthread_cond_destroy(&my.mc_cond);
8971 pthread_mutex_destroy(&my.mc_mutex);
8972 free(my.mc_wbuf[0]);
8977 /** Copy environment as-is. */
8979 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8981 MDB_txn *txn = NULL;
8982 mdb_mutexref_t wmutex = NULL;
8988 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8992 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8995 /* Do the lock/unlock of the reader mutex before starting the
8996 * write txn. Otherwise other read txns could block writers.
8998 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9003 /* We must start the actual read txn after blocking writers */
9004 mdb_txn_reset0(txn, "reset-stage1");
9006 /* Temporarily block writers until we snapshot the meta pages */
9007 wmutex = env->me_wmutex;
9008 if (LOCK_MUTEX(rc, env, wmutex))
9011 rc = mdb_txn_renew0(txn);
9013 UNLOCK_MUTEX(wmutex);
9018 wsize = env->me_psize * 2;
9022 DO_WRITE(rc, fd, ptr, w2, len);
9026 } else if (len > 0) {
9032 /* Non-blocking or async handles are not supported */
9038 UNLOCK_MUTEX(wmutex);
9043 w2 = txn->mt_next_pgno * env->me_psize;
9046 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9053 if (wsize > MAX_WRITE)
9057 DO_WRITE(rc, fd, ptr, w2, len);
9061 } else if (len > 0) {
9078 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9080 if (flags & MDB_CP_COMPACT)
9081 return mdb_env_copyfd1(env, fd);
9083 return mdb_env_copyfd0(env, fd);
9087 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9089 return mdb_env_copyfd2(env, fd, 0);
9093 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9097 HANDLE newfd = INVALID_HANDLE_VALUE;
9099 if (env->me_flags & MDB_NOSUBDIR) {
9100 lpath = (char *)path;
9103 len += sizeof(DATANAME);
9104 lpath = malloc(len);
9107 sprintf(lpath, "%s" DATANAME, path);
9110 /* The destination path must exist, but the destination file must not.
9111 * We don't want the OS to cache the writes, since the source data is
9112 * already in the OS cache.
9115 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9116 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9118 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9120 if (newfd == INVALID_HANDLE_VALUE) {
9125 if (env->me_psize >= env->me_os_psize) {
9127 /* Set O_DIRECT if the file system supports it */
9128 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9129 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9131 #ifdef F_NOCACHE /* __APPLE__ */
9132 rc = fcntl(newfd, F_NOCACHE, 1);
9140 rc = mdb_env_copyfd2(env, newfd, flags);
9143 if (!(env->me_flags & MDB_NOSUBDIR))
9145 if (newfd != INVALID_HANDLE_VALUE)
9146 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9153 mdb_env_copy(MDB_env *env, const char *path)
9155 return mdb_env_copy2(env, path, 0);
9159 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9161 if ((flag & CHANGEABLE) != flag)
9164 env->me_flags |= flag;
9166 env->me_flags &= ~flag;
9171 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9176 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9181 mdb_env_set_userctx(MDB_env *env, void *ctx)
9185 env->me_userctx = ctx;
9190 mdb_env_get_userctx(MDB_env *env)
9192 return env ? env->me_userctx : NULL;
9196 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9201 env->me_assert_func = func;
9207 mdb_env_get_path(MDB_env *env, const char **arg)
9212 *arg = env->me_path;
9217 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9226 /** Common code for #mdb_stat() and #mdb_env_stat().
9227 * @param[in] env the environment to operate in.
9228 * @param[in] db the #MDB_db record containing the stats to return.
9229 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9230 * @return 0, this function always succeeds.
9233 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9235 arg->ms_psize = env->me_psize;
9236 arg->ms_depth = db->md_depth;
9237 arg->ms_branch_pages = db->md_branch_pages;
9238 arg->ms_leaf_pages = db->md_leaf_pages;
9239 arg->ms_overflow_pages = db->md_overflow_pages;
9240 arg->ms_entries = db->md_entries;
9246 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9250 if (env == NULL || arg == NULL)
9253 toggle = mdb_env_pick_meta(env);
9255 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9259 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9263 if (env == NULL || arg == NULL)
9266 toggle = mdb_env_pick_meta(env);
9267 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9268 arg->me_mapsize = env->me_mapsize;
9269 arg->me_maxreaders = env->me_maxreaders;
9270 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9272 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9273 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9277 /** Set the default comparison functions for a database.
9278 * Called immediately after a database is opened to set the defaults.
9279 * The user can then override them with #mdb_set_compare() or
9280 * #mdb_set_dupsort().
9281 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9282 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9285 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9287 uint16_t f = txn->mt_dbs[dbi].md_flags;
9289 txn->mt_dbxs[dbi].md_cmp =
9290 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9291 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9293 txn->mt_dbxs[dbi].md_dcmp =
9294 !(f & MDB_DUPSORT) ? 0 :
9295 ((f & MDB_INTEGERDUP)
9296 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9297 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9300 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9306 int rc, dbflag, exact;
9307 unsigned int unused = 0, seq;
9310 if ((flags & VALID_FLAGS) != flags)
9312 if (txn->mt_flags & MDB_TXN_ERROR)
9318 if (flags & PERSISTENT_FLAGS) {
9319 uint16_t f2 = flags & PERSISTENT_FLAGS;
9320 /* make sure flag changes get committed */
9321 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9322 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9323 txn->mt_flags |= MDB_TXN_DIRTY;
9326 mdb_default_cmp(txn, MAIN_DBI);
9330 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9331 mdb_default_cmp(txn, MAIN_DBI);
9334 /* Is the DB already open? */
9336 for (i=2; i<txn->mt_numdbs; i++) {
9337 if (!txn->mt_dbxs[i].md_name.mv_size) {
9338 /* Remember this free slot */
9339 if (!unused) unused = i;
9342 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9343 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9349 /* If no free slot and max hit, fail */
9350 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9351 return MDB_DBS_FULL;
9353 /* Cannot mix named databases with some mainDB flags */
9354 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9355 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9357 /* Find the DB info */
9358 dbflag = DB_NEW|DB_VALID;
9361 key.mv_data = (void *)name;
9362 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9363 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9364 if (rc == MDB_SUCCESS) {
9365 /* make sure this is actually a DB */
9366 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9367 if (!(node->mn_flags & F_SUBDATA))
9368 return MDB_INCOMPATIBLE;
9369 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9370 /* Create if requested */
9371 data.mv_size = sizeof(MDB_db);
9372 data.mv_data = &dummy;
9373 memset(&dummy, 0, sizeof(dummy));
9374 dummy.md_root = P_INVALID;
9375 dummy.md_flags = flags & PERSISTENT_FLAGS;
9376 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9380 /* OK, got info, add to table */
9381 if (rc == MDB_SUCCESS) {
9382 unsigned int slot = unused ? unused : txn->mt_numdbs;
9383 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9384 txn->mt_dbxs[slot].md_name.mv_size = len;
9385 txn->mt_dbxs[slot].md_rel = NULL;
9386 txn->mt_dbflags[slot] = dbflag;
9387 /* txn-> and env-> are the same in read txns, use
9388 * tmp variable to avoid undefined assignment
9390 seq = ++txn->mt_env->me_dbiseqs[slot];
9391 txn->mt_dbiseqs[slot] = seq;
9393 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9395 mdb_default_cmp(txn, slot);
9404 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9406 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9409 if (txn->mt_flags & MDB_TXN_ERROR)
9412 if (txn->mt_dbflags[dbi] & DB_STALE) {
9415 /* Stale, must read the DB's root. cursor_init does it for us. */
9416 mdb_cursor_init(&mc, txn, dbi, &mx);
9418 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9421 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9424 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9426 ptr = env->me_dbxs[dbi].md_name.mv_data;
9427 /* If there was no name, this was already closed */
9429 env->me_dbxs[dbi].md_name.mv_data = NULL;
9430 env->me_dbxs[dbi].md_name.mv_size = 0;
9431 env->me_dbflags[dbi] = 0;
9432 env->me_dbiseqs[dbi]++;
9437 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9439 /* We could return the flags for the FREE_DBI too but what's the point? */
9440 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9442 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9446 /** Add all the DB's pages to the free list.
9447 * @param[in] mc Cursor on the DB to free.
9448 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9449 * @return 0 on success, non-zero on failure.
9452 mdb_drop0(MDB_cursor *mc, int subs)
9456 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9457 if (rc == MDB_SUCCESS) {
9458 MDB_txn *txn = mc->mc_txn;
9463 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9464 * This also avoids any P_LEAF2 pages, which have no nodes.
9466 if (mc->mc_flags & C_SUB)
9469 mdb_cursor_copy(mc, &mx);
9470 while (mc->mc_snum > 0) {
9471 MDB_page *mp = mc->mc_pg[mc->mc_top];
9472 unsigned n = NUMKEYS(mp);
9474 for (i=0; i<n; i++) {
9475 ni = NODEPTR(mp, i);
9476 if (ni->mn_flags & F_BIGDATA) {
9479 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9480 rc = mdb_page_get(txn, pg, &omp, NULL);
9483 mdb_cassert(mc, IS_OVERFLOW(omp));
9484 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9488 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9489 mdb_xcursor_init1(mc, ni);
9490 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9496 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9498 for (i=0; i<n; i++) {
9500 ni = NODEPTR(mp, i);
9503 mdb_midl_xappend(txn->mt_free_pgs, pg);
9508 mc->mc_ki[mc->mc_top] = i;
9509 rc = mdb_cursor_sibling(mc, 1);
9511 if (rc != MDB_NOTFOUND)
9513 /* no more siblings, go back to beginning
9514 * of previous level.
9518 for (i=1; i<mc->mc_snum; i++) {
9520 mc->mc_pg[i] = mx.mc_pg[i];
9525 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9528 txn->mt_flags |= MDB_TXN_ERROR;
9529 } else if (rc == MDB_NOTFOUND) {
9535 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9537 MDB_cursor *mc, *m2;
9540 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9543 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9546 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9549 rc = mdb_cursor_open(txn, dbi, &mc);
9553 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9554 /* Invalidate the dropped DB's cursors */
9555 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9556 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9560 /* Can't delete the main DB */
9561 if (del && dbi > MAIN_DBI) {
9562 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9564 txn->mt_dbflags[dbi] = DB_STALE;
9565 mdb_dbi_close(txn->mt_env, dbi);
9567 txn->mt_flags |= MDB_TXN_ERROR;
9570 /* reset the DB record, mark it dirty */
9571 txn->mt_dbflags[dbi] |= DB_DIRTY;
9572 txn->mt_dbs[dbi].md_depth = 0;
9573 txn->mt_dbs[dbi].md_branch_pages = 0;
9574 txn->mt_dbs[dbi].md_leaf_pages = 0;
9575 txn->mt_dbs[dbi].md_overflow_pages = 0;
9576 txn->mt_dbs[dbi].md_entries = 0;
9577 txn->mt_dbs[dbi].md_root = P_INVALID;
9579 txn->mt_flags |= MDB_TXN_DIRTY;
9582 mdb_cursor_close(mc);
9586 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9588 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9591 txn->mt_dbxs[dbi].md_cmp = cmp;
9595 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9597 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9600 txn->mt_dbxs[dbi].md_dcmp = cmp;
9604 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9606 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9609 txn->mt_dbxs[dbi].md_rel = rel;
9613 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9615 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9618 txn->mt_dbxs[dbi].md_relctx = ctx;
9623 mdb_env_get_maxkeysize(MDB_env *env)
9625 return ENV_MAXKEY(env);
9629 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9631 unsigned int i, rdrs;
9634 int rc = 0, first = 1;
9638 if (!env->me_txns) {
9639 return func("(no reader locks)\n", ctx);
9641 rdrs = env->me_txns->mti_numreaders;
9642 mr = env->me_txns->mti_readers;
9643 for (i=0; i<rdrs; i++) {
9645 txnid_t txnid = mr[i].mr_txnid;
9646 sprintf(buf, txnid == (txnid_t)-1 ?
9647 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9648 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9651 rc = func(" pid thread txnid\n", ctx);
9655 rc = func(buf, ctx);
9661 rc = func("(no active readers)\n", ctx);
9666 /** Insert pid into list if not already present.
9667 * return -1 if already present.
9670 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9672 /* binary search of pid in list */
9674 unsigned cursor = 1;
9676 unsigned n = ids[0];
9679 unsigned pivot = n >> 1;
9680 cursor = base + pivot + 1;
9681 val = pid - ids[cursor];
9686 } else if ( val > 0 ) {
9691 /* found, so it's a duplicate */
9700 for (n = ids[0]; n > cursor; n--)
9707 mdb_reader_check(MDB_env *env, int *dead)
9713 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9716 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9717 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9719 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9720 unsigned int i, j, rdrs;
9722 MDB_PID_T *pids, pid;
9723 int rc = MDB_SUCCESS, count = 0;
9725 rdrs = env->me_txns->mti_numreaders;
9726 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9730 mr = env->me_txns->mti_readers;
9731 for (i=0; i<rdrs; i++) {
9733 if (pid && pid != env->me_pid) {
9734 if (mdb_pid_insert(pids, pid) == 0) {
9735 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9736 /* Stale reader found */
9739 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9740 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9742 rdrs = 0; /* the above checked all readers */
9744 /* Recheck, a new process may have reused pid */
9745 if (mdb_reader_pid(env, Pidcheck, pid))
9750 if (mr[j].mr_pid == pid) {
9751 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9752 (unsigned) pid, mr[j].mr_txnid));
9757 UNLOCK_MUTEX(rmutex);
9768 #ifdef MDB_ROBUST_SUPPORTED
9769 /** Handle #LOCK_MUTEX0() failure.
9770 * Try to repair the lock file if the mutex owner died.
9771 * @param[in] env the environment handle
9772 * @param[in] mutex LOCK_MUTEX0() mutex
9773 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9774 * @return 0 on success with the mutex locked, or an error code on failure.
9776 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9778 int toggle, rlocked, rc2;
9780 if (rc == MDB_OWNERDEAD) {
9781 /* We own the mutex. Clean up after dead previous owner. */
9783 rlocked = (mutex == env->me_rmutex);
9785 /* Keep mti_txnid updated, otherwise next writer can
9786 * overwrite data which latest meta page refers to.
9788 toggle = mdb_env_pick_meta(env);
9789 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9790 /* env is hosed if the dead thread was ours */
9792 env->me_flags |= MDB_FATAL_ERROR;
9797 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9798 (rc ? "this process' env is hosed" : "recovering")));
9799 rc2 = mdb_reader_check0(env, rlocked, NULL);
9801 rc2 = mdb_mutex_consistent(mutex);
9802 if (rc || (rc = rc2)) {
9803 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9804 UNLOCK_MUTEX(mutex);
9810 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9815 #endif /* MDB_ROBUST_SUPPORTED */