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 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1061 pgno_t mt_next_pgno; /**< next unallocated page */
1062 /** The ID of this transaction. IDs are integers incrementing from 1.
1063 * Only committed write transactions increment the ID. If a transaction
1064 * aborts, the ID may be re-used by the next writer.
1067 MDB_env *mt_env; /**< the DB environment */
1068 /** The list of pages that became unused during this transaction.
1070 MDB_IDL mt_free_pgs;
1071 /** The list of loose pages that became unused and may be reused
1072 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1074 MDB_page *mt_loose_pgs;
1075 /* #Number of loose pages (#mt_loose_pgs) */
1077 /** The sorted list of dirty pages we temporarily wrote to disk
1078 * because the dirty list was full. page numbers in here are
1079 * shifted left by 1, deleted slots have the LSB set.
1081 MDB_IDL mt_spill_pgs;
1083 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1084 MDB_ID2L dirty_list;
1085 /** For read txns: This thread/txn's reader table slot, or NULL. */
1088 /** Array of records for each DB known in the environment. */
1090 /** Array of MDB_db records for each known DB */
1092 /** Array of sequence numbers for each DB handle */
1093 unsigned int *mt_dbiseqs;
1094 /** @defgroup mt_dbflag Transaction DB Flags
1098 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1099 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1100 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1101 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1102 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1104 /** In write txns, array of cursors for each DB */
1105 MDB_cursor **mt_cursors;
1106 /** Array of flags for each DB */
1107 unsigned char *mt_dbflags;
1108 /** Number of DB records in use, or 0 when the txn is finished.
1109 * This number only ever increments until the txn finishes; we
1110 * don't decrement it when individual DB handles are closed.
1114 /** @defgroup mdb_txn Transaction Flags
1118 /** #mdb_txn_begin() flags */
1119 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1120 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1121 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1122 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1123 /* internal txn flags */
1124 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1125 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1126 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1127 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1128 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1129 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1130 /** most operations on the txn are currently illegal */
1131 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1133 unsigned int mt_flags; /**< @ref mdb_txn */
1134 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1135 * Includes ancestor txns' dirty pages not hidden by other txns'
1136 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1137 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1139 unsigned int mt_dirty_room;
1142 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1143 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1144 * raise this on a 64 bit machine.
1146 #define CURSOR_STACK 32
1150 /** Cursors are used for all DB operations.
1151 * A cursor holds a path of (page pointer, key index) from the DB
1152 * root to a position in the DB, plus other state. #MDB_DUPSORT
1153 * cursors include an xcursor to the current data item. Write txns
1154 * track their cursors and keep them up to date when data moves.
1155 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1156 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1159 /** Next cursor on this DB in this txn */
1160 MDB_cursor *mc_next;
1161 /** Backup of the original cursor if this cursor is a shadow */
1162 MDB_cursor *mc_backup;
1163 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1164 struct MDB_xcursor *mc_xcursor;
1165 /** The transaction that owns this cursor */
1167 /** The database handle this cursor operates on */
1169 /** The database record for this cursor */
1171 /** The database auxiliary record for this cursor */
1173 /** The @ref mt_dbflag for this database */
1174 unsigned char *mc_dbflag;
1175 unsigned short mc_snum; /**< number of pushed pages */
1176 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1177 /** @defgroup mdb_cursor Cursor Flags
1179 * Cursor state flags.
1182 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1183 #define C_EOF 0x02 /**< No more data */
1184 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1185 #define C_DEL 0x08 /**< last op was a cursor_del */
1186 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1187 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1189 unsigned int mc_flags; /**< @ref mdb_cursor */
1190 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1191 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1194 /** Context for sorted-dup records.
1195 * We could have gone to a fully recursive design, with arbitrarily
1196 * deep nesting of sub-databases. But for now we only handle these
1197 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1199 typedef struct MDB_xcursor {
1200 /** A sub-cursor for traversing the Dup DB */
1201 MDB_cursor mx_cursor;
1202 /** The database record for this Dup DB */
1204 /** The auxiliary DB record for this Dup DB */
1206 /** The @ref mt_dbflag for this Dup DB */
1207 unsigned char mx_dbflag;
1210 /** State of FreeDB old pages, stored in the MDB_env */
1211 typedef struct MDB_pgstate {
1212 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1213 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1216 /** The database environment. */
1218 HANDLE me_fd; /**< The main data file */
1219 HANDLE me_lfd; /**< The lock file */
1220 HANDLE me_mfd; /**< just for writing the meta pages */
1221 /** Failed to update the meta page. Probably an I/O error. */
1222 #define MDB_FATAL_ERROR 0x80000000U
1223 /** Some fields are initialized. */
1224 #define MDB_ENV_ACTIVE 0x20000000U
1225 /** me_txkey is set */
1226 #define MDB_ENV_TXKEY 0x10000000U
1227 /** fdatasync is unreliable */
1228 #define MDB_FSYNCONLY 0x08000000U
1229 uint32_t me_flags; /**< @ref mdb_env */
1230 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1231 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1232 unsigned int me_maxreaders; /**< size of the reader table */
1233 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1234 volatile int me_close_readers;
1235 MDB_dbi me_numdbs; /**< number of DBs opened */
1236 MDB_dbi me_maxdbs; /**< size of the DB table */
1237 MDB_PID_T me_pid; /**< process ID of this env */
1238 char *me_path; /**< path to the DB files */
1239 char *me_map; /**< the memory map of the data file */
1240 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1241 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1242 void *me_pbuf; /**< scratch area for DUPSORT put() */
1243 MDB_txn *me_txn; /**< current write transaction */
1244 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1245 size_t me_mapsize; /**< size of the data memory map */
1246 off_t me_size; /**< current file size */
1247 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1248 MDB_dbx *me_dbxs; /**< array of static DB info */
1249 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1250 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1251 pthread_key_t me_txkey; /**< thread-key for readers */
1252 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1253 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1254 # define me_pglast me_pgstate.mf_pglast
1255 # define me_pghead me_pgstate.mf_pghead
1256 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1257 /** IDL of pages that became unused in a write txn */
1258 MDB_IDL me_free_pgs;
1259 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1260 MDB_ID2L me_dirty_list;
1261 /** Max number of freelist items that can fit in a single overflow page */
1263 /** Max size of a node on a page */
1264 unsigned int me_nodemax;
1265 #if !(MDB_MAXKEYSIZE)
1266 unsigned int me_maxkey; /**< max size of a key */
1268 int me_live_reader; /**< have liveness lock in reader table */
1270 int me_pidquery; /**< Used in OpenProcess */
1272 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1273 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1274 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1276 mdb_mutex_t me_rmutex;
1277 mdb_mutex_t me_wmutex;
1279 void *me_userctx; /**< User-settable context */
1280 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1283 /** Nested transaction */
1284 typedef struct MDB_ntxn {
1285 MDB_txn mnt_txn; /**< the transaction */
1286 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1289 /** max number of pages to commit in one writev() call */
1290 #define MDB_COMMIT_PAGES 64
1291 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1292 #undef MDB_COMMIT_PAGES
1293 #define MDB_COMMIT_PAGES IOV_MAX
1296 /** max bytes to write in one call */
1297 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1299 /** Check \b txn and \b dbi arguments to a function */
1300 #define TXN_DBI_EXIST(txn, dbi, validity) \
1301 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1303 /** Check for misused \b dbi handles */
1304 #define TXN_DBI_CHANGED(txn, dbi) \
1305 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1307 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1308 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1309 static int mdb_page_touch(MDB_cursor *mc);
1311 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1312 "reset-tmp", "fail-begin", "fail-beginchild"}
1314 /* mdb_txn_end operation number, for logging */
1315 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1316 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1318 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1319 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1320 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1321 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1322 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1324 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1325 static int mdb_page_search_root(MDB_cursor *mc,
1326 MDB_val *key, int modify);
1327 #define MDB_PS_MODIFY 1
1328 #define MDB_PS_ROOTONLY 2
1329 #define MDB_PS_FIRST 4
1330 #define MDB_PS_LAST 8
1331 static int mdb_page_search(MDB_cursor *mc,
1332 MDB_val *key, int flags);
1333 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1335 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1336 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1337 pgno_t newpgno, unsigned int nflags);
1339 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1340 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1341 static int mdb_env_write_meta(MDB_txn *txn);
1342 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1343 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1345 static void mdb_env_close0(MDB_env *env, int excl);
1347 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1348 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1349 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1350 static void mdb_node_del(MDB_cursor *mc, int ksize);
1351 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1352 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1353 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1354 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1355 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1357 static int mdb_rebalance(MDB_cursor *mc);
1358 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1360 static void mdb_cursor_pop(MDB_cursor *mc);
1361 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1363 static int mdb_cursor_del0(MDB_cursor *mc);
1364 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1365 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1366 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1367 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1368 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1370 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1371 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1373 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1374 static void mdb_xcursor_init0(MDB_cursor *mc);
1375 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1377 static int mdb_drop0(MDB_cursor *mc, int subs);
1378 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1379 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1382 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1385 /** Compare two items pointing at size_t's of unknown alignment. */
1386 #ifdef MISALIGNED_OK
1387 # define mdb_cmp_clong mdb_cmp_long
1389 # define mdb_cmp_clong mdb_cmp_cint
1393 static SECURITY_DESCRIPTOR mdb_null_sd;
1394 static SECURITY_ATTRIBUTES mdb_all_sa;
1395 static int mdb_sec_inited;
1398 /** Return the library version info. */
1400 mdb_version(int *major, int *minor, int *patch)
1402 if (major) *major = MDB_VERSION_MAJOR;
1403 if (minor) *minor = MDB_VERSION_MINOR;
1404 if (patch) *patch = MDB_VERSION_PATCH;
1405 return MDB_VERSION_STRING;
1408 /** Table of descriptions for LMDB @ref errors */
1409 static char *const mdb_errstr[] = {
1410 "MDB_KEYEXIST: Key/data pair already exists",
1411 "MDB_NOTFOUND: No matching key/data pair found",
1412 "MDB_PAGE_NOTFOUND: Requested page not found",
1413 "MDB_CORRUPTED: Located page was wrong type",
1414 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1415 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1416 "MDB_INVALID: File is not an LMDB file",
1417 "MDB_MAP_FULL: Environment mapsize limit reached",
1418 "MDB_DBS_FULL: Environment maxdbs limit reached",
1419 "MDB_READERS_FULL: Environment maxreaders limit reached",
1420 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1421 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1422 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1423 "MDB_PAGE_FULL: Internal error - page has no more space",
1424 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1425 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1426 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1427 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1428 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1429 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1433 mdb_strerror(int err)
1436 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1437 * This works as long as no function between the call to mdb_strerror
1438 * and the actual use of the message uses more than 4K of stack.
1441 char buf[1024], *ptr = buf;
1445 return ("Successful return: 0");
1447 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1448 i = err - MDB_KEYEXIST;
1449 return mdb_errstr[i];
1453 /* These are the C-runtime error codes we use. The comment indicates
1454 * their numeric value, and the Win32 error they would correspond to
1455 * if the error actually came from a Win32 API. A major mess, we should
1456 * have used LMDB-specific error codes for everything.
1459 case ENOENT: /* 2, FILE_NOT_FOUND */
1460 case EIO: /* 5, ACCESS_DENIED */
1461 case ENOMEM: /* 12, INVALID_ACCESS */
1462 case EACCES: /* 13, INVALID_DATA */
1463 case EBUSY: /* 16, CURRENT_DIRECTORY */
1464 case EINVAL: /* 22, BAD_COMMAND */
1465 case ENOSPC: /* 28, OUT_OF_PAPER */
1466 return strerror(err);
1471 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1472 FORMAT_MESSAGE_IGNORE_INSERTS,
1473 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1476 return strerror(err);
1480 /** assert(3) variant in cursor context */
1481 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1482 /** assert(3) variant in transaction context */
1483 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1484 /** assert(3) variant in environment context */
1485 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1488 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1489 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1492 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1493 const char *func, const char *file, int line)
1496 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1497 file, line, expr_txt, func);
1498 if (env->me_assert_func)
1499 env->me_assert_func(env, buf);
1500 fprintf(stderr, "%s\n", buf);
1504 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1508 /** Return the page number of \b mp which may be sub-page, for debug output */
1510 mdb_dbg_pgno(MDB_page *mp)
1513 COPY_PGNO(ret, mp->mp_pgno);
1517 /** Display a key in hexadecimal and return the address of the result.
1518 * @param[in] key the key to display
1519 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1520 * @return The key in hexadecimal form.
1523 mdb_dkey(MDB_val *key, char *buf)
1526 unsigned char *c = key->mv_data;
1532 if (key->mv_size > DKBUF_MAXKEYSIZE)
1533 return "MDB_MAXKEYSIZE";
1534 /* may want to make this a dynamic check: if the key is mostly
1535 * printable characters, print it as-is instead of converting to hex.
1539 for (i=0; i<key->mv_size; i++)
1540 ptr += sprintf(ptr, "%02x", *c++);
1542 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1548 mdb_leafnode_type(MDB_node *n)
1550 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1551 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1552 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1555 /** Display all the keys in the page. */
1557 mdb_page_list(MDB_page *mp)
1559 pgno_t pgno = mdb_dbg_pgno(mp);
1560 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1562 unsigned int i, nkeys, nsize, total = 0;
1566 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1567 case P_BRANCH: type = "Branch page"; break;
1568 case P_LEAF: type = "Leaf page"; break;
1569 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1570 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1571 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1573 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1574 pgno, mp->mp_pages, state);
1577 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1578 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1581 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1585 nkeys = NUMKEYS(mp);
1586 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1588 for (i=0; i<nkeys; i++) {
1589 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1590 key.mv_size = nsize = mp->mp_pad;
1591 key.mv_data = LEAF2KEY(mp, i, nsize);
1593 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1596 node = NODEPTR(mp, i);
1597 key.mv_size = node->mn_ksize;
1598 key.mv_data = node->mn_data;
1599 nsize = NODESIZE + key.mv_size;
1600 if (IS_BRANCH(mp)) {
1601 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1605 if (F_ISSET(node->mn_flags, F_BIGDATA))
1606 nsize += sizeof(pgno_t);
1608 nsize += NODEDSZ(node);
1610 nsize += sizeof(indx_t);
1611 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1612 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1614 total = EVEN(total);
1616 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1617 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1621 mdb_cursor_chk(MDB_cursor *mc)
1627 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1628 for (i=0; i<mc->mc_top; i++) {
1630 node = NODEPTR(mp, mc->mc_ki[i]);
1631 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1634 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1640 /** Count all the pages in each DB and in the freelist
1641 * and make sure it matches the actual number of pages
1643 * All named DBs must be open for a correct count.
1645 static void mdb_audit(MDB_txn *txn)
1649 MDB_ID freecount, count;
1654 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1655 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1656 freecount += *(MDB_ID *)data.mv_data;
1657 mdb_tassert(txn, rc == MDB_NOTFOUND);
1660 for (i = 0; i<txn->mt_numdbs; i++) {
1662 if (!(txn->mt_dbflags[i] & DB_VALID))
1664 mdb_cursor_init(&mc, txn, i, &mx);
1665 if (txn->mt_dbs[i].md_root == P_INVALID)
1667 count += txn->mt_dbs[i].md_branch_pages +
1668 txn->mt_dbs[i].md_leaf_pages +
1669 txn->mt_dbs[i].md_overflow_pages;
1670 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1671 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1672 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1675 mp = mc.mc_pg[mc.mc_top];
1676 for (j=0; j<NUMKEYS(mp); j++) {
1677 MDB_node *leaf = NODEPTR(mp, j);
1678 if (leaf->mn_flags & F_SUBDATA) {
1680 memcpy(&db, NODEDATA(leaf), sizeof(db));
1681 count += db.md_branch_pages + db.md_leaf_pages +
1682 db.md_overflow_pages;
1686 mdb_tassert(txn, rc == MDB_NOTFOUND);
1689 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1690 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1691 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1697 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1699 return txn->mt_dbxs[dbi].md_cmp(a, b);
1703 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1705 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1706 #if UINT_MAX < SIZE_MAX
1707 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1708 dcmp = mdb_cmp_clong;
1713 /** Allocate memory for a page.
1714 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1717 mdb_page_malloc(MDB_txn *txn, unsigned num)
1719 MDB_env *env = txn->mt_env;
1720 MDB_page *ret = env->me_dpages;
1721 size_t psize = env->me_psize, sz = psize, off;
1722 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1723 * For a single page alloc, we init everything after the page header.
1724 * For multi-page, we init the final page; if the caller needed that
1725 * many pages they will be filling in at least up to the last page.
1729 VGMEMP_ALLOC(env, ret, sz);
1730 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1731 env->me_dpages = ret->mp_next;
1734 psize -= off = PAGEHDRSZ;
1739 if ((ret = malloc(sz)) != NULL) {
1740 VGMEMP_ALLOC(env, ret, sz);
1741 if (!(env->me_flags & MDB_NOMEMINIT)) {
1742 memset((char *)ret + off, 0, psize);
1746 txn->mt_flags |= MDB_TXN_ERROR;
1750 /** Free a single page.
1751 * Saves single pages to a list, for future reuse.
1752 * (This is not used for multi-page overflow pages.)
1755 mdb_page_free(MDB_env *env, MDB_page *mp)
1757 mp->mp_next = env->me_dpages;
1758 VGMEMP_FREE(env, mp);
1759 env->me_dpages = mp;
1762 /** Free a dirty page */
1764 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1766 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1767 mdb_page_free(env, dp);
1769 /* large pages just get freed directly */
1770 VGMEMP_FREE(env, dp);
1775 /** Return all dirty pages to dpage list */
1777 mdb_dlist_free(MDB_txn *txn)
1779 MDB_env *env = txn->mt_env;
1780 MDB_ID2L dl = txn->mt_u.dirty_list;
1781 unsigned i, n = dl[0].mid;
1783 for (i = 1; i <= n; i++) {
1784 mdb_dpage_free(env, dl[i].mptr);
1789 /** Loosen or free a single page.
1790 * Saves single pages to a list for future reuse
1791 * in this same txn. It has been pulled from the freeDB
1792 * and already resides on the dirty list, but has been
1793 * deleted. Use these pages first before pulling again
1796 * If the page wasn't dirtied in this txn, just add it
1797 * to this txn's free list.
1800 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1803 pgno_t pgno = mp->mp_pgno;
1804 MDB_txn *txn = mc->mc_txn;
1806 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1807 if (txn->mt_parent) {
1808 MDB_ID2 *dl = txn->mt_u.dirty_list;
1809 /* If txn has a parent, make sure the page is in our
1813 unsigned x = mdb_mid2l_search(dl, pgno);
1814 if (x <= dl[0].mid && dl[x].mid == pgno) {
1815 if (mp != dl[x].mptr) { /* bad cursor? */
1816 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1817 txn->mt_flags |= MDB_TXN_ERROR;
1818 return MDB_CORRUPTED;
1825 /* no parent txn, so it's just ours */
1830 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1832 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1833 txn->mt_loose_pgs = mp;
1834 txn->mt_loose_count++;
1835 mp->mp_flags |= P_LOOSE;
1837 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1845 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1846 * @param[in] mc A cursor handle for the current operation.
1847 * @param[in] pflags Flags of the pages to update:
1848 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1849 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1850 * @return 0 on success, non-zero on failure.
1853 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1855 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1856 MDB_txn *txn = mc->mc_txn;
1862 int rc = MDB_SUCCESS, level;
1864 /* Mark pages seen by cursors */
1865 if (mc->mc_flags & C_UNTRACK)
1866 mc = NULL; /* will find mc in mt_cursors */
1867 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1868 for (; mc; mc=mc->mc_next) {
1869 if (!(mc->mc_flags & C_INITIALIZED))
1871 for (m3 = mc;; m3 = &mx->mx_cursor) {
1873 for (j=0; j<m3->mc_snum; j++) {
1875 if ((mp->mp_flags & Mask) == pflags)
1876 mp->mp_flags ^= P_KEEP;
1878 mx = m3->mc_xcursor;
1879 /* Proceed to mx if it is at a sub-database */
1880 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1882 if (! (mp && (mp->mp_flags & P_LEAF)))
1884 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1885 if (!(leaf->mn_flags & F_SUBDATA))
1894 /* Mark dirty root pages */
1895 for (i=0; i<txn->mt_numdbs; i++) {
1896 if (txn->mt_dbflags[i] & DB_DIRTY) {
1897 pgno_t pgno = txn->mt_dbs[i].md_root;
1898 if (pgno == P_INVALID)
1900 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1902 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1903 dp->mp_flags ^= P_KEEP;
1911 static int mdb_page_flush(MDB_txn *txn, int keep);
1913 /** Spill pages from the dirty list back to disk.
1914 * This is intended to prevent running into #MDB_TXN_FULL situations,
1915 * but note that they may still occur in a few cases:
1916 * 1) our estimate of the txn size could be too small. Currently this
1917 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1918 * 2) child txns may run out of space if their parents dirtied a
1919 * lot of pages and never spilled them. TODO: we probably should do
1920 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1921 * the parent's dirty_room is below a given threshold.
1923 * Otherwise, if not using nested txns, it is expected that apps will
1924 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1925 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1926 * If the txn never references them again, they can be left alone.
1927 * If the txn only reads them, they can be used without any fuss.
1928 * If the txn writes them again, they can be dirtied immediately without
1929 * going thru all of the work of #mdb_page_touch(). Such references are
1930 * handled by #mdb_page_unspill().
1932 * Also note, we never spill DB root pages, nor pages of active cursors,
1933 * because we'll need these back again soon anyway. And in nested txns,
1934 * we can't spill a page in a child txn if it was already spilled in a
1935 * parent txn. That would alter the parent txns' data even though
1936 * the child hasn't committed yet, and we'd have no way to undo it if
1937 * the child aborted.
1939 * @param[in] m0 cursor A cursor handle identifying the transaction and
1940 * database for which we are checking space.
1941 * @param[in] key For a put operation, the key being stored.
1942 * @param[in] data For a put operation, the data being stored.
1943 * @return 0 on success, non-zero on failure.
1946 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1948 MDB_txn *txn = m0->mc_txn;
1950 MDB_ID2L dl = txn->mt_u.dirty_list;
1951 unsigned int i, j, need;
1954 if (m0->mc_flags & C_SUB)
1957 /* Estimate how much space this op will take */
1958 i = m0->mc_db->md_depth;
1959 /* Named DBs also dirty the main DB */
1960 if (m0->mc_dbi > MAIN_DBI)
1961 i += txn->mt_dbs[MAIN_DBI].md_depth;
1962 /* For puts, roughly factor in the key+data size */
1964 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1965 i += i; /* double it for good measure */
1968 if (txn->mt_dirty_room > i)
1971 if (!txn->mt_spill_pgs) {
1972 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1973 if (!txn->mt_spill_pgs)
1976 /* purge deleted slots */
1977 MDB_IDL sl = txn->mt_spill_pgs;
1978 unsigned int num = sl[0];
1980 for (i=1; i<=num; i++) {
1987 /* Preserve pages which may soon be dirtied again */
1988 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1991 /* Less aggressive spill - we originally spilled the entire dirty list,
1992 * with a few exceptions for cursor pages and DB root pages. But this
1993 * turns out to be a lot of wasted effort because in a large txn many
1994 * of those pages will need to be used again. So now we spill only 1/8th
1995 * of the dirty pages. Testing revealed this to be a good tradeoff,
1996 * better than 1/2, 1/4, or 1/10.
1998 if (need < MDB_IDL_UM_MAX / 8)
1999 need = MDB_IDL_UM_MAX / 8;
2001 /* Save the page IDs of all the pages we're flushing */
2002 /* flush from the tail forward, this saves a lot of shifting later on. */
2003 for (i=dl[0].mid; i && need; i--) {
2004 MDB_ID pn = dl[i].mid << 1;
2006 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2008 /* Can't spill twice, make sure it's not already in a parent's
2011 if (txn->mt_parent) {
2013 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2014 if (tx2->mt_spill_pgs) {
2015 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2016 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2017 dp->mp_flags |= P_KEEP;
2025 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2029 mdb_midl_sort(txn->mt_spill_pgs);
2031 /* Flush the spilled part of dirty list */
2032 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2035 /* Reset any dirty pages we kept that page_flush didn't see */
2036 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2039 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2043 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2045 mdb_find_oldest(MDB_txn *txn)
2048 txnid_t mr, oldest = txn->mt_txnid - 1;
2049 if (txn->mt_env->me_txns) {
2050 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2051 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2062 /** Add a page to the txn's dirty list */
2064 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2067 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2069 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2070 insert = mdb_mid2l_append;
2072 insert = mdb_mid2l_insert;
2074 mid.mid = mp->mp_pgno;
2076 rc = insert(txn->mt_u.dirty_list, &mid);
2077 mdb_tassert(txn, rc == 0);
2078 txn->mt_dirty_room--;
2081 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2082 * me_pghead and mt_next_pgno.
2084 * If there are free pages available from older transactions, they
2085 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2086 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2087 * and move me_pglast to say which records were consumed. Only this
2088 * function can create me_pghead and move me_pglast/mt_next_pgno.
2089 * @param[in] mc cursor A cursor handle identifying the transaction and
2090 * database for which we are allocating.
2091 * @param[in] num the number of pages to allocate.
2092 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2093 * will always be satisfied by a single contiguous chunk of memory.
2094 * @return 0 on success, non-zero on failure.
2097 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2099 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2100 /* Get at most <Max_retries> more freeDB records once me_pghead
2101 * has enough pages. If not enough, use new pages from the map.
2102 * If <Paranoid> and mc is updating the freeDB, only get new
2103 * records if me_pghead is empty. Then the freelist cannot play
2104 * catch-up with itself by growing while trying to save it.
2106 enum { Paranoid = 1, Max_retries = 500 };
2108 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2110 int rc, retry = num * 60;
2111 MDB_txn *txn = mc->mc_txn;
2112 MDB_env *env = txn->mt_env;
2113 pgno_t pgno, *mop = env->me_pghead;
2114 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2116 txnid_t oldest = 0, last;
2121 /* If there are any loose pages, just use them */
2122 if (num == 1 && txn->mt_loose_pgs) {
2123 np = txn->mt_loose_pgs;
2124 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2125 txn->mt_loose_count--;
2126 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2134 /* If our dirty list is already full, we can't do anything */
2135 if (txn->mt_dirty_room == 0) {
2140 for (op = MDB_FIRST;; op = MDB_NEXT) {
2145 /* Seek a big enough contiguous page range. Prefer
2146 * pages at the tail, just truncating the list.
2152 if (mop[i-n2] == pgno+n2)
2159 if (op == MDB_FIRST) { /* 1st iteration */
2160 /* Prepare to fetch more and coalesce */
2161 last = env->me_pglast;
2162 oldest = env->me_pgoldest;
2163 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2166 key.mv_data = &last; /* will look up last+1 */
2167 key.mv_size = sizeof(last);
2169 if (Paranoid && mc->mc_dbi == FREE_DBI)
2172 if (Paranoid && retry < 0 && mop_len)
2176 /* Do not fetch more if the record will be too recent */
2177 if (oldest <= last) {
2179 oldest = mdb_find_oldest(txn);
2180 env->me_pgoldest = oldest;
2186 rc = mdb_cursor_get(&m2, &key, NULL, op);
2188 if (rc == MDB_NOTFOUND)
2192 last = *(txnid_t*)key.mv_data;
2193 if (oldest <= last) {
2195 oldest = mdb_find_oldest(txn);
2196 env->me_pgoldest = oldest;
2202 np = m2.mc_pg[m2.mc_top];
2203 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2204 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2207 idl = (MDB_ID *) data.mv_data;
2210 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2215 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2217 mop = env->me_pghead;
2219 env->me_pglast = last;
2221 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2222 last, txn->mt_dbs[FREE_DBI].md_root, i));
2224 DPRINTF(("IDL %"Z"u", idl[j]));
2226 /* Merge in descending sorted order */
2227 mdb_midl_xmerge(mop, idl);
2231 /* Use new pages from the map when nothing suitable in the freeDB */
2233 pgno = txn->mt_next_pgno;
2234 if (pgno + num >= env->me_maxpg) {
2235 DPUTS("DB size maxed out");
2241 if (env->me_flags & MDB_WRITEMAP) {
2242 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2244 if (!(np = mdb_page_malloc(txn, num))) {
2250 mop[0] = mop_len -= num;
2251 /* Move any stragglers down */
2252 for (j = i-num; j < mop_len; )
2253 mop[++j] = mop[++i];
2255 txn->mt_next_pgno = pgno + num;
2258 mdb_page_dirty(txn, np);
2264 txn->mt_flags |= MDB_TXN_ERROR;
2268 /** Copy the used portions of a non-overflow page.
2269 * @param[in] dst page to copy into
2270 * @param[in] src page to copy from
2271 * @param[in] psize size of a page
2274 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2276 enum { Align = sizeof(pgno_t) };
2277 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2279 /* If page isn't full, just copy the used portion. Adjust
2280 * alignment so memcpy may copy words instead of bytes.
2282 if ((unused &= -Align) && !IS_LEAF2(src)) {
2283 upper = (upper + PAGEBASE) & -Align;
2284 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2285 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2288 memcpy(dst, src, psize - unused);
2292 /** Pull a page off the txn's spill list, if present.
2293 * If a page being referenced was spilled to disk in this txn, bring
2294 * it back and make it dirty/writable again.
2295 * @param[in] txn the transaction handle.
2296 * @param[in] mp the page being referenced. It must not be dirty.
2297 * @param[out] ret the writable page, if any. ret is unchanged if
2298 * mp wasn't spilled.
2301 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2303 MDB_env *env = txn->mt_env;
2306 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2308 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2309 if (!tx2->mt_spill_pgs)
2311 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2312 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2315 if (txn->mt_dirty_room == 0)
2316 return MDB_TXN_FULL;
2317 if (IS_OVERFLOW(mp))
2321 if (env->me_flags & MDB_WRITEMAP) {
2324 np = mdb_page_malloc(txn, num);
2328 memcpy(np, mp, num * env->me_psize);
2330 mdb_page_copy(np, mp, env->me_psize);
2333 /* If in current txn, this page is no longer spilled.
2334 * If it happens to be the last page, truncate the spill list.
2335 * Otherwise mark it as deleted by setting the LSB.
2337 if (x == txn->mt_spill_pgs[0])
2338 txn->mt_spill_pgs[0]--;
2340 txn->mt_spill_pgs[x] |= 1;
2341 } /* otherwise, if belonging to a parent txn, the
2342 * page remains spilled until child commits
2345 mdb_page_dirty(txn, np);
2346 np->mp_flags |= P_DIRTY;
2354 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2355 * @param[in] mc cursor pointing to the page to be touched
2356 * @return 0 on success, non-zero on failure.
2359 mdb_page_touch(MDB_cursor *mc)
2361 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2362 MDB_txn *txn = mc->mc_txn;
2363 MDB_cursor *m2, *m3;
2367 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2368 if (txn->mt_flags & MDB_TXN_SPILLS) {
2370 rc = mdb_page_unspill(txn, mp, &np);
2376 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2377 (rc = mdb_page_alloc(mc, 1, &np)))
2380 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2381 mp->mp_pgno, pgno));
2382 mdb_cassert(mc, mp->mp_pgno != pgno);
2383 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2384 /* Update the parent page, if any, to point to the new page */
2386 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2387 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2388 SETPGNO(node, pgno);
2390 mc->mc_db->md_root = pgno;
2392 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2393 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2395 /* If txn has a parent, make sure the page is in our
2399 unsigned x = mdb_mid2l_search(dl, pgno);
2400 if (x <= dl[0].mid && dl[x].mid == pgno) {
2401 if (mp != dl[x].mptr) { /* bad cursor? */
2402 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2403 txn->mt_flags |= MDB_TXN_ERROR;
2404 return MDB_CORRUPTED;
2409 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2411 np = mdb_page_malloc(txn, 1);
2416 rc = mdb_mid2l_insert(dl, &mid);
2417 mdb_cassert(mc, rc == 0);
2422 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2424 np->mp_flags |= P_DIRTY;
2427 /* Adjust cursors pointing to mp */
2428 mc->mc_pg[mc->mc_top] = np;
2429 m2 = txn->mt_cursors[mc->mc_dbi];
2430 if (mc->mc_flags & C_SUB) {
2431 for (; m2; m2=m2->mc_next) {
2432 m3 = &m2->mc_xcursor->mx_cursor;
2433 if (m3->mc_snum < mc->mc_snum) continue;
2434 if (m3->mc_pg[mc->mc_top] == mp)
2435 m3->mc_pg[mc->mc_top] = np;
2438 for (; m2; m2=m2->mc_next) {
2439 if (m2->mc_snum < mc->mc_snum) continue;
2440 if (m2->mc_pg[mc->mc_top] == mp) {
2441 m2->mc_pg[mc->mc_top] = np;
2442 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2444 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2446 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2447 if (!(leaf->mn_flags & F_SUBDATA))
2448 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2456 txn->mt_flags |= MDB_TXN_ERROR;
2461 mdb_env_sync(MDB_env *env, int force)
2464 if (env->me_flags & MDB_RDONLY)
2466 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2467 if (env->me_flags & MDB_WRITEMAP) {
2468 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2469 ? MS_ASYNC : MS_SYNC;
2470 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2473 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2477 #ifdef BROKEN_FDATASYNC
2478 if (env->me_flags & MDB_FSYNCONLY) {
2479 if (fsync(env->me_fd))
2483 if (MDB_FDATASYNC(env->me_fd))
2490 /** Back up parent txn's cursors, then grab the originals for tracking */
2492 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2494 MDB_cursor *mc, *bk;
2499 for (i = src->mt_numdbs; --i >= 0; ) {
2500 if ((mc = src->mt_cursors[i]) != NULL) {
2501 size = sizeof(MDB_cursor);
2503 size += sizeof(MDB_xcursor);
2504 for (; mc; mc = bk->mc_next) {
2510 mc->mc_db = &dst->mt_dbs[i];
2511 /* Kill pointers into src - and dst to reduce abuse: The
2512 * user may not use mc until dst ends. Otherwise we'd...
2514 mc->mc_txn = NULL; /* ...set this to dst */
2515 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2516 if ((mx = mc->mc_xcursor) != NULL) {
2517 *(MDB_xcursor *)(bk+1) = *mx;
2518 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2520 mc->mc_next = dst->mt_cursors[i];
2521 dst->mt_cursors[i] = mc;
2528 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2529 * @param[in] txn the transaction handle.
2530 * @param[in] merge true to keep changes to parent cursors, false to revert.
2531 * @return 0 on success, non-zero on failure.
2534 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2536 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2540 for (i = txn->mt_numdbs; --i >= 0; ) {
2541 for (mc = cursors[i]; mc; mc = next) {
2543 if ((bk = mc->mc_backup) != NULL) {
2545 /* Commit changes to parent txn */
2546 mc->mc_next = bk->mc_next;
2547 mc->mc_backup = bk->mc_backup;
2548 mc->mc_txn = bk->mc_txn;
2549 mc->mc_db = bk->mc_db;
2550 mc->mc_dbflag = bk->mc_dbflag;
2551 if ((mx = mc->mc_xcursor) != NULL)
2552 mx->mx_cursor.mc_txn = bk->mc_txn;
2554 /* Abort nested txn */
2556 if ((mx = mc->mc_xcursor) != NULL)
2557 *mx = *(MDB_xcursor *)(bk+1);
2561 /* Only malloced cursors are permanently tracked. */
2568 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2574 Pidset = F_SETLK, Pidcheck = F_GETLK
2578 /** Set or check a pid lock. Set returns 0 on success.
2579 * Check returns 0 if the process is certainly dead, nonzero if it may
2580 * be alive (the lock exists or an error happened so we do not know).
2582 * On Windows Pidset is a no-op, we merely check for the existence
2583 * of the process with the given pid. On POSIX we use a single byte
2584 * lock on the lockfile, set at an offset equal to the pid.
2587 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2589 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2592 if (op == Pidcheck) {
2593 h = OpenProcess(env->me_pidquery, FALSE, pid);
2594 /* No documented "no such process" code, but other program use this: */
2596 return ErrCode() != ERROR_INVALID_PARAMETER;
2597 /* A process exists until all handles to it close. Has it exited? */
2598 ret = WaitForSingleObject(h, 0) != 0;
2605 struct flock lock_info;
2606 memset(&lock_info, 0, sizeof(lock_info));
2607 lock_info.l_type = F_WRLCK;
2608 lock_info.l_whence = SEEK_SET;
2609 lock_info.l_start = pid;
2610 lock_info.l_len = 1;
2611 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2612 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2614 } else if ((rc = ErrCode()) == EINTR) {
2622 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2623 * @param[in] txn the transaction handle to initialize
2624 * @return 0 on success, non-zero on failure.
2627 mdb_txn_renew0(MDB_txn *txn)
2629 MDB_env *env = txn->mt_env;
2630 MDB_txninfo *ti = env->me_txns;
2632 unsigned int i, nr, flags = txn->mt_flags;
2634 int rc, new_notls = 0;
2636 if ((flags &= MDB_TXN_RDONLY) != 0) {
2638 meta = mdb_env_pick_meta(env);
2639 txn->mt_txnid = meta->mm_txnid;
2640 txn->mt_u.reader = NULL;
2642 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2643 pthread_getspecific(env->me_txkey);
2645 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2646 return MDB_BAD_RSLOT;
2648 MDB_PID_T pid = env->me_pid;
2649 MDB_THR_T tid = pthread_self();
2650 mdb_mutexref_t rmutex = env->me_rmutex;
2652 if (!env->me_live_reader) {
2653 rc = mdb_reader_pid(env, Pidset, pid);
2656 env->me_live_reader = 1;
2659 if (LOCK_MUTEX(rc, env, rmutex))
2661 nr = ti->mti_numreaders;
2662 for (i=0; i<nr; i++)
2663 if (ti->mti_readers[i].mr_pid == 0)
2665 if (i == env->me_maxreaders) {
2666 UNLOCK_MUTEX(rmutex);
2667 return MDB_READERS_FULL;
2669 r = &ti->mti_readers[i];
2670 /* Claim the reader slot, carefully since other code
2671 * uses the reader table un-mutexed: First reset the
2672 * slot, next publish it in mti_numreaders. After
2673 * that, it is safe for mdb_env_close() to touch it.
2674 * When it will be closed, we can finally claim it.
2677 r->mr_txnid = (txnid_t)-1;
2680 ti->mti_numreaders = ++nr;
2681 env->me_close_readers = nr;
2683 UNLOCK_MUTEX(rmutex);
2685 new_notls = (env->me_flags & MDB_NOTLS);
2686 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2691 do /* LY: Retry on a race, ITS#7970. */
2692 r->mr_txnid = ti->mti_txnid;
2693 while(r->mr_txnid != ti->mti_txnid);
2694 txn->mt_txnid = r->mr_txnid;
2695 txn->mt_u.reader = r;
2696 meta = env->me_metas[txn->mt_txnid & 1];
2698 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2700 /* Not yet touching txn == env->me_txn0, it may be active */
2702 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2704 txn->mt_txnid = ti->mti_txnid;
2705 meta = env->me_metas[txn->mt_txnid & 1];
2707 meta = mdb_env_pick_meta(env);
2708 txn->mt_txnid = meta->mm_txnid;
2712 if (txn->mt_txnid == mdb_debug_start)
2715 txn->mt_child = NULL;
2716 txn->mt_loose_pgs = NULL;
2717 txn->mt_loose_count = 0;
2718 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2719 txn->mt_u.dirty_list = env->me_dirty_list;
2720 txn->mt_u.dirty_list[0].mid = 0;
2721 txn->mt_free_pgs = env->me_free_pgs;
2722 txn->mt_free_pgs[0] = 0;
2723 txn->mt_spill_pgs = NULL;
2725 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2728 /* Copy the DB info and flags */
2729 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2731 /* Moved to here to avoid a data race in read TXNs */
2732 txn->mt_next_pgno = meta->mm_last_pg+1;
2734 txn->mt_flags = flags;
2737 txn->mt_numdbs = env->me_numdbs;
2738 for (i=2; i<txn->mt_numdbs; i++) {
2739 x = env->me_dbflags[i];
2740 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2741 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2743 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2744 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2746 if (env->me_maxpg < txn->mt_next_pgno) {
2747 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2748 return MDB_MAP_RESIZED;
2755 mdb_txn_renew(MDB_txn *txn)
2759 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2762 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2763 DPUTS("environment had fatal error, must shutdown!");
2767 rc = mdb_txn_renew0(txn);
2768 if (rc == MDB_SUCCESS) {
2769 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2770 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2771 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2777 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2781 int rc, size, tsize;
2783 flags &= MDB_TXN_BEGIN_FLAGS;
2784 flags |= env->me_flags & MDB_WRITEMAP;
2786 if (env->me_flags & MDB_FATAL_ERROR) {
2787 DPUTS("environment had fatal error, must shutdown!");
2790 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2794 /* Nested transactions: Max 1 child, write txns only, no writemap */
2795 flags |= parent->mt_flags;
2796 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2797 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2799 /* Child txns save MDB_pgstate and use own copy of cursors */
2800 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2801 size += tsize = sizeof(MDB_ntxn);
2802 } else if (flags & MDB_RDONLY) {
2803 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2804 size += tsize = sizeof(MDB_txn);
2806 /* Reuse preallocated write txn. However, do not touch it until
2807 * mdb_txn_renew0() succeeds, since it currently may be active.
2812 if ((txn = calloc(1, size)) == NULL) {
2813 DPRINTF(("calloc: %s", strerror(errno)));
2816 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2817 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2818 txn->mt_flags = flags;
2823 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2824 txn->mt_dbiseqs = parent->mt_dbiseqs;
2825 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2826 if (!txn->mt_u.dirty_list ||
2827 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2829 free(txn->mt_u.dirty_list);
2833 txn->mt_txnid = parent->mt_txnid;
2834 txn->mt_dirty_room = parent->mt_dirty_room;
2835 txn->mt_u.dirty_list[0].mid = 0;
2836 txn->mt_spill_pgs = NULL;
2837 txn->mt_next_pgno = parent->mt_next_pgno;
2838 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2839 parent->mt_child = txn;
2840 txn->mt_parent = parent;
2841 txn->mt_numdbs = parent->mt_numdbs;
2842 txn->mt_dbxs = parent->mt_dbxs;
2843 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2844 /* Copy parent's mt_dbflags, but clear DB_NEW */
2845 for (i=0; i<txn->mt_numdbs; i++)
2846 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2848 ntxn = (MDB_ntxn *)txn;
2849 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2850 if (env->me_pghead) {
2851 size = MDB_IDL_SIZEOF(env->me_pghead);
2852 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2854 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2859 rc = mdb_cursor_shadow(parent, txn);
2861 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2862 } else { /* MDB_RDONLY */
2863 txn->mt_dbiseqs = env->me_dbiseqs;
2865 rc = mdb_txn_renew0(txn);
2868 if (txn != env->me_txn0)
2871 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2873 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2874 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2875 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2882 mdb_txn_env(MDB_txn *txn)
2884 if(!txn) return NULL;
2889 mdb_txn_id(MDB_txn *txn)
2892 return txn->mt_txnid;
2895 /** Export or close DBI handles opened in this txn. */
2897 mdb_dbis_update(MDB_txn *txn, int keep)
2900 MDB_dbi n = txn->mt_numdbs;
2901 MDB_env *env = txn->mt_env;
2902 unsigned char *tdbflags = txn->mt_dbflags;
2904 for (i = n; --i >= 2;) {
2905 if (tdbflags[i] & DB_NEW) {
2907 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2909 char *ptr = env->me_dbxs[i].md_name.mv_data;
2911 env->me_dbxs[i].md_name.mv_data = NULL;
2912 env->me_dbxs[i].md_name.mv_size = 0;
2913 env->me_dbflags[i] = 0;
2914 env->me_dbiseqs[i]++;
2920 if (keep && env->me_numdbs < n)
2924 /** End a transaction, except successful commit of a nested transaction.
2925 * May be called twice for readonly txns: First reset it, then abort.
2926 * @param[in] txn the transaction handle to end
2927 * @param[in] mode why and how to end the transaction
2930 mdb_txn_end(MDB_txn *txn, unsigned mode)
2932 MDB_env *env = txn->mt_env;
2934 static const char *const names[] = MDB_END_NAMES;
2937 /* Export or close DBI handles opened in this txn */
2938 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2940 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2941 names[mode & MDB_END_OPMASK],
2942 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2943 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2945 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2946 if (txn->mt_u.reader) {
2947 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2948 if (!(env->me_flags & MDB_NOTLS)) {
2949 txn->mt_u.reader = NULL; /* txn does not own reader */
2950 } else if (mode & MDB_END_SLOT) {
2951 txn->mt_u.reader->mr_pid = 0;
2952 txn->mt_u.reader = NULL;
2953 } /* else txn owns the slot until it does MDB_END_SLOT */
2955 txn->mt_numdbs = 0; /* prevent further DBI activity */
2956 txn->mt_flags |= MDB_TXN_FINISHED;
2957 txn->mt_dbxs = NULL; /* mark txn as reset */
2958 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2959 pgno_t *pghead = env->me_pghead;
2961 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2962 mdb_cursors_close(txn, 0);
2963 if (!(env->me_flags & MDB_WRITEMAP)) {
2964 mdb_dlist_free(txn);
2968 txn->mt_flags = MDB_TXN_FINISHED;
2970 if (!txn->mt_parent) {
2971 mdb_midl_shrink(&txn->mt_free_pgs);
2972 env->me_free_pgs = txn->mt_free_pgs;
2974 env->me_pghead = NULL;
2978 mode = 0; /* txn == env->me_txn0, do not free() it */
2980 /* The writer mutex was locked in mdb_txn_begin. */
2982 UNLOCK_MUTEX(env->me_wmutex);
2984 txn->mt_parent->mt_child = NULL;
2985 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2986 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2987 mdb_midl_free(txn->mt_free_pgs);
2988 mdb_midl_free(txn->mt_spill_pgs);
2989 free(txn->mt_u.dirty_list);
2992 mdb_midl_free(pghead);
2995 if (mode & MDB_END_FREE)
3000 mdb_txn_reset(MDB_txn *txn)
3005 /* This call is only valid for read-only txns */
3006 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3009 mdb_txn_end(txn, MDB_END_RESET);
3013 mdb_txn_abort(MDB_txn *txn)
3019 mdb_txn_abort(txn->mt_child);
3021 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3024 /** Save the freelist as of this transaction to the freeDB.
3025 * This changes the freelist. Keep trying until it stabilizes.
3028 mdb_freelist_save(MDB_txn *txn)
3030 /* env->me_pghead[] can grow and shrink during this call.
3031 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3032 * Page numbers cannot disappear from txn->mt_free_pgs[].
3035 MDB_env *env = txn->mt_env;
3036 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3037 txnid_t pglast = 0, head_id = 0;
3038 pgno_t freecnt = 0, *free_pgs, *mop;
3039 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3041 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3043 if (env->me_pghead) {
3044 /* Make sure first page of freeDB is touched and on freelist */
3045 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3046 if (rc && rc != MDB_NOTFOUND)
3050 if (!env->me_pghead && txn->mt_loose_pgs) {
3051 /* Put loose page numbers in mt_free_pgs, since
3052 * we may be unable to return them to me_pghead.
3054 MDB_page *mp = txn->mt_loose_pgs;
3055 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3057 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3058 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3059 txn->mt_loose_pgs = NULL;
3060 txn->mt_loose_count = 0;
3063 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3064 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3065 ? SSIZE_MAX : maxfree_1pg;
3068 /* Come back here after each Put() in case freelist changed */
3073 /* If using records from freeDB which we have not yet
3074 * deleted, delete them and any we reserved for me_pghead.
3076 while (pglast < env->me_pglast) {
3077 rc = mdb_cursor_first(&mc, &key, NULL);
3080 pglast = head_id = *(txnid_t *)key.mv_data;
3081 total_room = head_room = 0;
3082 mdb_tassert(txn, pglast <= env->me_pglast);
3083 rc = mdb_cursor_del(&mc, 0);
3088 /* Save the IDL of pages freed by this txn, to a single record */
3089 if (freecnt < txn->mt_free_pgs[0]) {
3091 /* Make sure last page of freeDB is touched and on freelist */
3092 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3093 if (rc && rc != MDB_NOTFOUND)
3096 free_pgs = txn->mt_free_pgs;
3097 /* Write to last page of freeDB */
3098 key.mv_size = sizeof(txn->mt_txnid);
3099 key.mv_data = &txn->mt_txnid;
3101 freecnt = free_pgs[0];
3102 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3103 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3106 /* Retry if mt_free_pgs[] grew during the Put() */
3107 free_pgs = txn->mt_free_pgs;
3108 } while (freecnt < free_pgs[0]);
3109 mdb_midl_sort(free_pgs);
3110 memcpy(data.mv_data, free_pgs, data.mv_size);
3113 unsigned int i = free_pgs[0];
3114 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3115 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3117 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3123 mop = env->me_pghead;
3124 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3126 /* Reserve records for me_pghead[]. Split it if multi-page,
3127 * to avoid searching freeDB for a page range. Use keys in
3128 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3130 if (total_room >= mop_len) {
3131 if (total_room == mop_len || --more < 0)
3133 } else if (head_room >= maxfree_1pg && head_id > 1) {
3134 /* Keep current record (overflow page), add a new one */
3138 /* (Re)write {key = head_id, IDL length = head_room} */
3139 total_room -= head_room;
3140 head_room = mop_len - total_room;
3141 if (head_room > maxfree_1pg && head_id > 1) {
3142 /* Overflow multi-page for part of me_pghead */
3143 head_room /= head_id; /* amortize page sizes */
3144 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3145 } else if (head_room < 0) {
3146 /* Rare case, not bothering to delete this record */
3149 key.mv_size = sizeof(head_id);
3150 key.mv_data = &head_id;
3151 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3152 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3155 /* IDL is initially empty, zero out at least the length */
3156 pgs = (pgno_t *)data.mv_data;
3157 j = head_room > clean_limit ? head_room : 0;
3161 total_room += head_room;
3164 /* Return loose page numbers to me_pghead, though usually none are
3165 * left at this point. The pages themselves remain in dirty_list.
3167 if (txn->mt_loose_pgs) {
3168 MDB_page *mp = txn->mt_loose_pgs;
3169 unsigned count = txn->mt_loose_count;
3171 /* Room for loose pages + temp IDL with same */
3172 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3174 mop = env->me_pghead;
3175 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3176 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3177 loose[ ++count ] = mp->mp_pgno;
3179 mdb_midl_sort(loose);
3180 mdb_midl_xmerge(mop, loose);
3181 txn->mt_loose_pgs = NULL;
3182 txn->mt_loose_count = 0;
3186 /* Fill in the reserved me_pghead records */
3192 rc = mdb_cursor_first(&mc, &key, &data);
3193 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3194 txnid_t id = *(txnid_t *)key.mv_data;
3195 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3198 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3200 if (len > mop_len) {
3202 data.mv_size = (len + 1) * sizeof(MDB_ID);
3204 data.mv_data = mop -= len;
3207 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3209 if (rc || !(mop_len -= len))
3216 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3217 * @param[in] txn the transaction that's being committed
3218 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3219 * @return 0 on success, non-zero on failure.
3222 mdb_page_flush(MDB_txn *txn, int keep)
3224 MDB_env *env = txn->mt_env;
3225 MDB_ID2L dl = txn->mt_u.dirty_list;
3226 unsigned psize = env->me_psize, j;
3227 int i, pagecount = dl[0].mid, rc;
3228 size_t size = 0, pos = 0;
3230 MDB_page *dp = NULL;
3234 struct iovec iov[MDB_COMMIT_PAGES];
3235 ssize_t wpos = 0, wsize = 0, wres;
3236 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3242 if (env->me_flags & MDB_WRITEMAP) {
3243 /* Clear dirty flags */
3244 while (++i <= pagecount) {
3246 /* Don't flush this page yet */
3247 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3248 dp->mp_flags &= ~P_KEEP;
3252 dp->mp_flags &= ~P_DIRTY;
3257 /* Write the pages */
3259 if (++i <= pagecount) {
3261 /* Don't flush this page yet */
3262 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3263 dp->mp_flags &= ~P_KEEP;
3268 /* clear dirty flag */
3269 dp->mp_flags &= ~P_DIRTY;
3272 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3277 /* Windows actually supports scatter/gather I/O, but only on
3278 * unbuffered file handles. Since we're relying on the OS page
3279 * cache for all our data, that's self-defeating. So we just
3280 * write pages one at a time. We use the ov structure to set
3281 * the write offset, to at least save the overhead of a Seek
3284 DPRINTF(("committing page %"Z"u", pgno));
3285 memset(&ov, 0, sizeof(ov));
3286 ov.Offset = pos & 0xffffffff;
3287 ov.OffsetHigh = pos >> 16 >> 16;
3288 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3290 DPRINTF(("WriteFile: %d", rc));
3294 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3295 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3298 /* Write previous page(s) */
3299 #ifdef MDB_USE_PWRITEV
3300 wres = pwritev(env->me_fd, iov, n, wpos);
3303 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3306 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3310 DPRINTF(("lseek: %s", strerror(rc)));
3313 wres = writev(env->me_fd, iov, n);
3316 if (wres != wsize) {
3321 DPRINTF(("Write error: %s", strerror(rc)));
3323 rc = EIO; /* TODO: Use which error code? */
3324 DPUTS("short write, filesystem full?");
3335 DPRINTF(("committing page %"Z"u", pgno));
3336 next_pos = pos + size;
3337 iov[n].iov_len = size;
3338 iov[n].iov_base = (char *)dp;
3344 /* MIPS has cache coherency issues, this is a no-op everywhere else
3345 * Note: for any size >= on-chip cache size, entire on-chip cache is
3348 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3350 for (i = keep; ++i <= pagecount; ) {
3352 /* This is a page we skipped above */
3355 dl[j].mid = dp->mp_pgno;
3358 mdb_dpage_free(env, dp);
3363 txn->mt_dirty_room += i - j;
3369 mdb_txn_commit(MDB_txn *txn)
3372 unsigned int i, end_mode;
3378 /* mdb_txn_end() mode for a commit which writes nothing */
3379 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3381 if (txn->mt_child) {
3382 rc = mdb_txn_commit(txn->mt_child);
3389 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3393 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3394 DPUTS("txn has failed/finished, can't commit");
3396 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3401 if (txn->mt_parent) {
3402 MDB_txn *parent = txn->mt_parent;
3406 unsigned x, y, len, ps_len;
3408 /* Append our free list to parent's */
3409 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3412 mdb_midl_free(txn->mt_free_pgs);
3413 /* Failures after this must either undo the changes
3414 * to the parent or set MDB_TXN_ERROR in the parent.
3417 parent->mt_next_pgno = txn->mt_next_pgno;
3418 parent->mt_flags = txn->mt_flags;
3420 /* Merge our cursors into parent's and close them */
3421 mdb_cursors_close(txn, 1);
3423 /* Update parent's DB table. */
3424 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3425 parent->mt_numdbs = txn->mt_numdbs;
3426 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3427 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3428 for (i=2; i<txn->mt_numdbs; i++) {
3429 /* preserve parent's DB_NEW status */
3430 x = parent->mt_dbflags[i] & DB_NEW;
3431 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3434 dst = parent->mt_u.dirty_list;
3435 src = txn->mt_u.dirty_list;
3436 /* Remove anything in our dirty list from parent's spill list */
3437 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3439 pspill[0] = (pgno_t)-1;
3440 /* Mark our dirty pages as deleted in parent spill list */
3441 for (i=0, len=src[0].mid; ++i <= len; ) {
3442 MDB_ID pn = src[i].mid << 1;
3443 while (pn > pspill[x])
3445 if (pn == pspill[x]) {
3450 /* Squash deleted pagenums if we deleted any */
3451 for (x=y; ++x <= ps_len; )
3452 if (!(pspill[x] & 1))
3453 pspill[++y] = pspill[x];
3457 /* Find len = length of merging our dirty list with parent's */
3459 dst[0].mid = 0; /* simplify loops */
3460 if (parent->mt_parent) {
3461 len = x + src[0].mid;
3462 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3463 for (i = x; y && i; y--) {
3464 pgno_t yp = src[y].mid;
3465 while (yp < dst[i].mid)
3467 if (yp == dst[i].mid) {
3472 } else { /* Simplify the above for single-ancestor case */
3473 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3475 /* Merge our dirty list with parent's */
3477 for (i = len; y; dst[i--] = src[y--]) {
3478 pgno_t yp = src[y].mid;
3479 while (yp < dst[x].mid)
3480 dst[i--] = dst[x--];
3481 if (yp == dst[x].mid)
3482 free(dst[x--].mptr);
3484 mdb_tassert(txn, i == x);
3486 free(txn->mt_u.dirty_list);
3487 parent->mt_dirty_room = txn->mt_dirty_room;
3488 if (txn->mt_spill_pgs) {
3489 if (parent->mt_spill_pgs) {
3490 /* TODO: Prevent failure here, so parent does not fail */
3491 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3493 parent->mt_flags |= MDB_TXN_ERROR;
3494 mdb_midl_free(txn->mt_spill_pgs);
3495 mdb_midl_sort(parent->mt_spill_pgs);
3497 parent->mt_spill_pgs = txn->mt_spill_pgs;
3501 /* Append our loose page list to parent's */
3502 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3504 *lp = txn->mt_loose_pgs;
3505 parent->mt_loose_count += txn->mt_loose_count;
3507 parent->mt_child = NULL;
3508 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3513 if (txn != env->me_txn) {
3514 DPUTS("attempt to commit unknown transaction");
3519 mdb_cursors_close(txn, 0);
3521 if (!txn->mt_u.dirty_list[0].mid &&
3522 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3525 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3526 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3528 /* Update DB root pointers */
3529 if (txn->mt_numdbs > 2) {
3533 data.mv_size = sizeof(MDB_db);
3535 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3536 for (i = 2; i < txn->mt_numdbs; i++) {
3537 if (txn->mt_dbflags[i] & DB_DIRTY) {
3538 if (TXN_DBI_CHANGED(txn, i)) {
3542 data.mv_data = &txn->mt_dbs[i];
3543 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3551 rc = mdb_freelist_save(txn);
3555 mdb_midl_free(env->me_pghead);
3556 env->me_pghead = NULL;
3557 mdb_midl_shrink(&txn->mt_free_pgs);
3563 if ((rc = mdb_page_flush(txn, 0)))
3565 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3566 (rc = mdb_env_sync(env, 0)))
3568 if ((rc = mdb_env_write_meta(txn)))
3570 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3573 mdb_txn_end(txn, end_mode);
3581 /** Read the environment parameters of a DB environment before
3582 * mapping it into memory.
3583 * @param[in] env the environment handle
3584 * @param[out] meta address of where to store the meta information
3585 * @return 0 on success, non-zero on failure.
3588 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3594 enum { Size = sizeof(pbuf) };
3596 /* We don't know the page size yet, so use a minimum value.
3597 * Read both meta pages so we can use the latest one.
3600 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3604 memset(&ov, 0, sizeof(ov));
3606 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3607 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3610 rc = pread(env->me_fd, &pbuf, Size, off);
3613 if (rc == 0 && off == 0)
3615 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3616 DPRINTF(("read: %s", mdb_strerror(rc)));
3620 p = (MDB_page *)&pbuf;
3622 if (!F_ISSET(p->mp_flags, P_META)) {
3623 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3628 if (m->mm_magic != MDB_MAGIC) {
3629 DPUTS("meta has invalid magic");
3633 if (m->mm_version != MDB_DATA_VERSION) {
3634 DPRINTF(("database is version %u, expected version %u",
3635 m->mm_version, MDB_DATA_VERSION));
3636 return MDB_VERSION_MISMATCH;
3639 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3645 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3647 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3649 meta->mm_magic = MDB_MAGIC;
3650 meta->mm_version = MDB_DATA_VERSION;
3651 meta->mm_mapsize = env->me_mapsize;
3652 meta->mm_psize = env->me_psize;
3653 meta->mm_last_pg = 1;
3654 meta->mm_flags = env->me_flags & 0xffff;
3655 meta->mm_flags |= MDB_INTEGERKEY;
3656 meta->mm_dbs[0].md_root = P_INVALID;
3657 meta->mm_dbs[1].md_root = P_INVALID;
3660 /** Write the environment parameters of a freshly created DB environment.
3661 * @param[in] env the environment handle
3662 * @param[in] meta the #MDB_meta to write
3663 * @return 0 on success, non-zero on failure.
3666 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3674 memset(&ov, 0, sizeof(ov));
3675 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3677 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3680 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3681 len = pwrite(fd, ptr, size, pos); \
3682 if (len == -1 && ErrCode() == EINTR) continue; \
3683 rc = (len >= 0); break; } while(1)
3686 DPUTS("writing new meta page");
3688 psize = env->me_psize;
3690 p = calloc(2, psize);
3692 p->mp_flags = P_META;
3693 *(MDB_meta *)METADATA(p) = *meta;
3695 q = (MDB_page *)((char *)p + psize);
3697 q->mp_flags = P_META;
3698 *(MDB_meta *)METADATA(q) = *meta;
3700 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3703 else if ((unsigned) len == psize * 2)
3711 /** Update the environment info to commit a transaction.
3712 * @param[in] txn the transaction that's being committed
3713 * @return 0 on success, non-zero on failure.
3716 mdb_env_write_meta(MDB_txn *txn)
3719 MDB_meta meta, metab, *mp;
3723 int rc, len, toggle;
3732 toggle = txn->mt_txnid & 1;
3733 DPRINTF(("writing meta page %d for root page %"Z"u",
3734 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3737 flags = txn->mt_flags & env->me_flags;
3738 mp = env->me_metas[toggle];
3739 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3740 /* Persist any increases of mapsize config */
3741 if (mapsize < env->me_mapsize)
3742 mapsize = env->me_mapsize;
3744 if (flags & MDB_WRITEMAP) {
3745 mp->mm_mapsize = mapsize;
3746 mp->mm_dbs[0] = txn->mt_dbs[0];
3747 mp->mm_dbs[1] = txn->mt_dbs[1];
3748 mp->mm_last_pg = txn->mt_next_pgno - 1;
3749 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3750 !(defined(__i386__) || defined(__x86_64__))
3751 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3752 __sync_synchronize();
3754 mp->mm_txnid = txn->mt_txnid;
3755 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3756 unsigned meta_size = env->me_psize;
3757 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3760 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3761 if (meta_size < env->me_os_psize)
3762 meta_size += meta_size;
3767 if (MDB_MSYNC(ptr, meta_size, rc)) {
3774 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3775 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3777 meta.mm_mapsize = mapsize;
3778 meta.mm_dbs[0] = txn->mt_dbs[0];
3779 meta.mm_dbs[1] = txn->mt_dbs[1];
3780 meta.mm_last_pg = txn->mt_next_pgno - 1;
3781 meta.mm_txnid = txn->mt_txnid;
3783 off = offsetof(MDB_meta, mm_mapsize);
3784 ptr = (char *)&meta + off;
3785 len = sizeof(MDB_meta) - off;
3787 off += env->me_psize;
3790 /* Write to the SYNC fd */
3791 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3795 memset(&ov, 0, sizeof(ov));
3797 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3801 rc = pwrite(mfd, ptr, len, off);
3804 rc = rc < 0 ? ErrCode() : EIO;
3807 DPUTS("write failed, disk error?");
3808 /* On a failure, the pagecache still contains the new data.
3809 * Write some old data back, to prevent it from being used.
3810 * Use the non-SYNC fd; we know it will fail anyway.
3812 meta.mm_last_pg = metab.mm_last_pg;
3813 meta.mm_txnid = metab.mm_txnid;
3815 memset(&ov, 0, sizeof(ov));
3817 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3819 r2 = pwrite(env->me_fd, ptr, len, off);
3820 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3823 env->me_flags |= MDB_FATAL_ERROR;
3826 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3827 CACHEFLUSH(env->me_map + off, len, DCACHE);
3829 /* Memory ordering issues are irrelevant; since the entire writer
3830 * is wrapped by wmutex, all of these changes will become visible
3831 * after the wmutex is unlocked. Since the DB is multi-version,
3832 * readers will get consistent data regardless of how fresh or
3833 * how stale their view of these values is.
3836 env->me_txns->mti_txnid = txn->mt_txnid;
3841 /** Check both meta pages to see which one is newer.
3842 * @param[in] env the environment handle
3843 * @return newest #MDB_meta.
3846 mdb_env_pick_meta(const MDB_env *env)
3848 MDB_meta *const *metas = env->me_metas;
3849 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3853 mdb_env_create(MDB_env **env)
3857 e = calloc(1, sizeof(MDB_env));
3861 e->me_maxreaders = DEFAULT_READERS;
3862 e->me_maxdbs = e->me_numdbs = 2;
3863 e->me_fd = INVALID_HANDLE_VALUE;
3864 e->me_lfd = INVALID_HANDLE_VALUE;
3865 e->me_mfd = INVALID_HANDLE_VALUE;
3866 #ifdef MDB_USE_POSIX_SEM
3867 e->me_rmutex = SEM_FAILED;
3868 e->me_wmutex = SEM_FAILED;
3869 #elif defined MDB_USE_SYSV_SEM
3870 e->me_rmutex->semid = -1;
3871 e->me_wmutex->semid = -1;
3873 e->me_pid = getpid();
3874 GET_PAGESIZE(e->me_os_psize);
3875 VGMEMP_CREATE(e,0,0);
3881 mdb_env_map(MDB_env *env, void *addr)
3884 unsigned int flags = env->me_flags;
3888 LONG sizelo, sizehi;
3891 if (flags & MDB_RDONLY) {
3892 /* Don't set explicit map size, use whatever exists */
3897 msize = env->me_mapsize;
3898 sizelo = msize & 0xffffffff;
3899 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3901 /* Windows won't create mappings for zero length files.
3902 * and won't map more than the file size.
3903 * Just set the maxsize right now.
3905 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3906 || !SetEndOfFile(env->me_fd)
3907 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3911 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3912 PAGE_READWRITE : PAGE_READONLY,
3913 sizehi, sizelo, NULL);
3916 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3917 FILE_MAP_WRITE : FILE_MAP_READ,
3919 rc = env->me_map ? 0 : ErrCode();
3924 int prot = PROT_READ;
3925 if (flags & MDB_WRITEMAP) {
3927 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3930 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3932 if (env->me_map == MAP_FAILED) {
3937 if (flags & MDB_NORDAHEAD) {
3938 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3940 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3942 #ifdef POSIX_MADV_RANDOM
3943 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3944 #endif /* POSIX_MADV_RANDOM */
3945 #endif /* MADV_RANDOM */
3949 /* Can happen because the address argument to mmap() is just a
3950 * hint. mmap() can pick another, e.g. if the range is in use.
3951 * The MAP_FIXED flag would prevent that, but then mmap could
3952 * instead unmap existing pages to make room for the new map.
3954 if (addr && env->me_map != addr)
3955 return EBUSY; /* TODO: Make a new MDB_* error code? */
3957 p = (MDB_page *)env->me_map;
3958 env->me_metas[0] = METADATA(p);
3959 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3965 mdb_env_set_mapsize(MDB_env *env, size_t size)
3967 /* If env is already open, caller is responsible for making
3968 * sure there are no active txns.
3976 meta = mdb_env_pick_meta(env);
3978 size = meta->mm_mapsize;
3980 /* Silently round up to minimum if the size is too small */
3981 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3985 munmap(env->me_map, env->me_mapsize);
3986 env->me_mapsize = size;
3987 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3988 rc = mdb_env_map(env, old);
3992 env->me_mapsize = size;
3994 env->me_maxpg = env->me_mapsize / env->me_psize;
3999 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4003 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
4008 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4010 if (env->me_map || readers < 1)
4012 env->me_maxreaders = readers;
4017 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4019 if (!env || !readers)
4021 *readers = env->me_maxreaders;
4026 mdb_fsize(HANDLE fd, size_t *size)
4029 LARGE_INTEGER fsize;
4031 if (!GetFileSizeEx(fd, &fsize))
4034 *size = fsize.QuadPart;
4046 #ifdef BROKEN_FDATASYNC
4047 #include <sys/utsname.h>
4048 #include <sys/vfs.h>
4051 /** Further setup required for opening an LMDB environment
4054 mdb_env_open2(MDB_env *env)
4056 unsigned int flags = env->me_flags;
4057 int i, newenv = 0, rc;
4061 /* See if we should use QueryLimited */
4063 if ((rc & 0xff) > 5)
4064 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4066 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4069 #ifdef BROKEN_FDATASYNC
4070 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4071 * https://lkml.org/lkml/2012/9/3/83
4072 * Kernels after 3.6-rc6 are known good.
4073 * https://lkml.org/lkml/2012/9/10/556
4074 * See if the DB is on ext3/ext4, then check for new enough kernel
4075 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4080 fstatfs(env->me_fd, &st);
4081 while (st.f_type == 0xEF53) {
4085 if (uts.release[0] < '3') {
4086 if (!strncmp(uts.release, "2.6.32.", 7)) {
4087 i = atoi(uts.release+7);
4089 break; /* 2.6.32.60 and newer is OK */
4090 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4091 i = atoi(uts.release+7);
4093 break; /* 2.6.34.15 and newer is OK */
4095 } else if (uts.release[0] == '3') {
4096 i = atoi(uts.release+2);
4098 break; /* 3.6 and newer is OK */
4100 i = atoi(uts.release+4);
4102 break; /* 3.5.4 and newer is OK */
4103 } else if (i == 2) {
4104 i = atoi(uts.release+4);
4106 break; /* 3.2.30 and newer is OK */
4108 } else { /* 4.x and newer is OK */
4111 env->me_flags |= MDB_FSYNCONLY;
4117 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4120 DPUTS("new mdbenv");
4122 env->me_psize = env->me_os_psize;
4123 if (env->me_psize > MAX_PAGESIZE)
4124 env->me_psize = MAX_PAGESIZE;
4125 memset(&meta, 0, sizeof(meta));
4126 mdb_env_init_meta0(env, &meta);
4127 meta.mm_mapsize = DEFAULT_MAPSIZE;
4129 env->me_psize = meta.mm_psize;
4132 /* Was a mapsize configured? */
4133 if (!env->me_mapsize) {
4134 env->me_mapsize = meta.mm_mapsize;
4137 /* Make sure mapsize >= committed data size. Even when using
4138 * mm_mapsize, which could be broken in old files (ITS#7789).
4140 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4141 if (env->me_mapsize < minsize)
4142 env->me_mapsize = minsize;
4144 meta.mm_mapsize = env->me_mapsize;
4146 if (newenv && !(flags & MDB_FIXEDMAP)) {
4147 /* mdb_env_map() may grow the datafile. Write the metapages
4148 * first, so the file will be valid if initialization fails.
4149 * Except with FIXEDMAP, since we do not yet know mm_address.
4150 * We could fill in mm_address later, but then a different
4151 * program might end up doing that - one with a memory layout
4152 * and map address which does not suit the main program.
4154 rc = mdb_env_init_meta(env, &meta);
4160 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4165 if (flags & MDB_FIXEDMAP)
4166 meta.mm_address = env->me_map;
4167 i = mdb_env_init_meta(env, &meta);
4168 if (i != MDB_SUCCESS) {
4173 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4174 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4176 #if !(MDB_MAXKEYSIZE)
4177 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4179 env->me_maxpg = env->me_mapsize / env->me_psize;
4183 MDB_meta *meta = mdb_env_pick_meta(env);
4184 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4186 DPRINTF(("opened database version %u, pagesize %u",
4187 meta->mm_version, env->me_psize));
4188 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4189 DPRINTF(("depth: %u", db->md_depth));
4190 DPRINTF(("entries: %"Z"u", db->md_entries));
4191 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4192 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4193 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4194 DPRINTF(("root: %"Z"u", db->md_root));
4202 /** Release a reader thread's slot in the reader lock table.
4203 * This function is called automatically when a thread exits.
4204 * @param[in] ptr This points to the slot in the reader lock table.
4207 mdb_env_reader_dest(void *ptr)
4209 MDB_reader *reader = ptr;
4215 /** Junk for arranging thread-specific callbacks on Windows. This is
4216 * necessarily platform and compiler-specific. Windows supports up
4217 * to 1088 keys. Let's assume nobody opens more than 64 environments
4218 * in a single process, for now. They can override this if needed.
4220 #ifndef MAX_TLS_KEYS
4221 #define MAX_TLS_KEYS 64
4223 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4224 static int mdb_tls_nkeys;
4226 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4230 case DLL_PROCESS_ATTACH: break;
4231 case DLL_THREAD_ATTACH: break;
4232 case DLL_THREAD_DETACH:
4233 for (i=0; i<mdb_tls_nkeys; i++) {
4234 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4236 mdb_env_reader_dest(r);
4240 case DLL_PROCESS_DETACH: break;
4245 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4247 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4251 /* Force some symbol references.
4252 * _tls_used forces the linker to create the TLS directory if not already done
4253 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4255 #pragma comment(linker, "/INCLUDE:_tls_used")
4256 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4257 #pragma const_seg(".CRT$XLB")
4258 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4259 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4262 #pragma comment(linker, "/INCLUDE:__tls_used")
4263 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4264 #pragma data_seg(".CRT$XLB")
4265 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4267 #endif /* WIN 32/64 */
4268 #endif /* !__GNUC__ */
4271 /** Downgrade the exclusive lock on the region back to shared */
4273 mdb_env_share_locks(MDB_env *env, int *excl)
4276 MDB_meta *meta = mdb_env_pick_meta(env);
4278 env->me_txns->mti_txnid = meta->mm_txnid;
4283 /* First acquire a shared lock. The Unlock will
4284 * then release the existing exclusive lock.
4286 memset(&ov, 0, sizeof(ov));
4287 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4290 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4296 struct flock lock_info;
4297 /* The shared lock replaces the existing lock */
4298 memset((void *)&lock_info, 0, sizeof(lock_info));
4299 lock_info.l_type = F_RDLCK;
4300 lock_info.l_whence = SEEK_SET;
4301 lock_info.l_start = 0;
4302 lock_info.l_len = 1;
4303 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4304 (rc = ErrCode()) == EINTR) ;
4305 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4312 /** Try to get exclusive lock, otherwise shared.
4313 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4316 mdb_env_excl_lock(MDB_env *env, int *excl)
4320 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4324 memset(&ov, 0, sizeof(ov));
4325 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4332 struct flock lock_info;
4333 memset((void *)&lock_info, 0, sizeof(lock_info));
4334 lock_info.l_type = F_WRLCK;
4335 lock_info.l_whence = SEEK_SET;
4336 lock_info.l_start = 0;
4337 lock_info.l_len = 1;
4338 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4339 (rc = ErrCode()) == EINTR) ;
4343 # ifndef MDB_USE_POSIX_MUTEX
4344 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4347 lock_info.l_type = F_RDLCK;
4348 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4349 (rc = ErrCode()) == EINTR) ;
4359 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4361 * @(#) $Revision: 5.1 $
4362 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4363 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4365 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4369 * Please do not copyright this code. This code is in the public domain.
4371 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4372 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4373 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4374 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4375 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4376 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4377 * PERFORMANCE OF THIS SOFTWARE.
4380 * chongo <Landon Curt Noll> /\oo/\
4381 * http://www.isthe.com/chongo/
4383 * Share and Enjoy! :-)
4386 typedef unsigned long long mdb_hash_t;
4387 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4389 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4390 * @param[in] val value to hash
4391 * @param[in] hval initial value for hash
4392 * @return 64 bit hash
4394 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4395 * hval arg on the first call.
4398 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4400 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4401 unsigned char *end = s + val->mv_size;
4403 * FNV-1a hash each octet of the string
4406 /* xor the bottom with the current octet */
4407 hval ^= (mdb_hash_t)*s++;
4409 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4410 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4411 (hval << 7) + (hval << 8) + (hval << 40);
4413 /* return our new hash value */
4417 /** Hash the string and output the encoded hash.
4418 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4419 * very short name limits. We don't care about the encoding being reversible,
4420 * we just want to preserve as many bits of the input as possible in a
4421 * small printable string.
4422 * @param[in] str string to hash
4423 * @param[out] encbuf an array of 11 chars to hold the hash
4425 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4428 mdb_pack85(unsigned long l, char *out)
4432 for (i=0; i<5; i++) {
4433 *out++ = mdb_a85[l % 85];
4439 mdb_hash_enc(MDB_val *val, char *encbuf)
4441 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4443 mdb_pack85(h, encbuf);
4444 mdb_pack85(h>>32, encbuf+5);
4449 /** Open and/or initialize the lock region for the environment.
4450 * @param[in] env The LMDB environment.
4451 * @param[in] lpath The pathname of the file used for the lock region.
4452 * @param[in] mode The Unix permissions for the file, if we create it.
4453 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4454 * @return 0 on success, non-zero on failure.
4457 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4460 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4462 # define MDB_ERRCODE_ROFS EROFS
4463 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4464 # define MDB_CLOEXEC O_CLOEXEC
4467 # define MDB_CLOEXEC 0
4470 #ifdef MDB_USE_SYSV_SEM
4478 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4479 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4480 FILE_ATTRIBUTE_NORMAL, NULL);
4482 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4484 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4486 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4491 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4492 /* Lose record locks when exec*() */
4493 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4494 fcntl(env->me_lfd, F_SETFD, fdflags);
4497 if (!(env->me_flags & MDB_NOTLS)) {
4498 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4501 env->me_flags |= MDB_ENV_TXKEY;
4503 /* Windows TLS callbacks need help finding their TLS info. */
4504 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4508 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4512 /* Try to get exclusive lock. If we succeed, then
4513 * nobody is using the lock region and we should initialize it.
4515 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4518 size = GetFileSize(env->me_lfd, NULL);
4520 size = lseek(env->me_lfd, 0, SEEK_END);
4521 if (size == -1) goto fail_errno;
4523 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4524 if (size < rsize && *excl > 0) {
4526 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4527 || !SetEndOfFile(env->me_lfd))
4530 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4534 size = rsize - sizeof(MDB_txninfo);
4535 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4540 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4542 if (!mh) goto fail_errno;
4543 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4545 if (!env->me_txns) goto fail_errno;
4547 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4549 if (m == MAP_FAILED) goto fail_errno;
4555 BY_HANDLE_FILE_INFORMATION stbuf;
4564 if (!mdb_sec_inited) {
4565 InitializeSecurityDescriptor(&mdb_null_sd,
4566 SECURITY_DESCRIPTOR_REVISION);
4567 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4568 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4569 mdb_all_sa.bInheritHandle = FALSE;
4570 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4573 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4574 idbuf.volume = stbuf.dwVolumeSerialNumber;
4575 idbuf.nhigh = stbuf.nFileIndexHigh;
4576 idbuf.nlow = stbuf.nFileIndexLow;
4577 val.mv_data = &idbuf;
4578 val.mv_size = sizeof(idbuf);
4579 mdb_hash_enc(&val, encbuf);
4580 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4581 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4582 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4583 if (!env->me_rmutex) goto fail_errno;
4584 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4585 if (!env->me_wmutex) goto fail_errno;
4586 #elif defined(MDB_USE_POSIX_SEM)
4595 #if defined(__NetBSD__)
4596 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4598 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4599 idbuf.dev = stbuf.st_dev;
4600 idbuf.ino = stbuf.st_ino;
4601 val.mv_data = &idbuf;
4602 val.mv_size = sizeof(idbuf);
4603 mdb_hash_enc(&val, encbuf);
4604 #ifdef MDB_SHORT_SEMNAMES
4605 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4607 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4608 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4609 /* Clean up after a previous run, if needed: Try to
4610 * remove both semaphores before doing anything else.
4612 sem_unlink(env->me_txns->mti_rmname);
4613 sem_unlink(env->me_txns->mti_wmname);
4614 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4615 O_CREAT|O_EXCL, mode, 1);
4616 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4617 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4618 O_CREAT|O_EXCL, mode, 1);
4619 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4620 #elif defined(MDB_USE_SYSV_SEM)
4621 unsigned short vals[2] = {1, 1};
4622 key_t key = ftok(lpath, 'M');
4625 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4629 if (semctl(semid, 0, SETALL, semu) < 0)
4631 env->me_txns->mti_semid = semid;
4632 #else /* MDB_USE_POSIX_MUTEX: */
4633 pthread_mutexattr_t mattr;
4635 if ((rc = pthread_mutexattr_init(&mattr))
4636 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4637 #ifdef MDB_ROBUST_SUPPORTED
4638 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4640 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4641 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4643 pthread_mutexattr_destroy(&mattr);
4644 #endif /* _WIN32 || ... */
4646 env->me_txns->mti_magic = MDB_MAGIC;
4647 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4648 env->me_txns->mti_txnid = 0;
4649 env->me_txns->mti_numreaders = 0;
4652 #ifdef MDB_USE_SYSV_SEM
4653 struct semid_ds buf;
4655 if (env->me_txns->mti_magic != MDB_MAGIC) {
4656 DPUTS("lock region has invalid magic");
4660 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4661 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4662 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4663 rc = MDB_VERSION_MISMATCH;
4667 if (rc && rc != EACCES && rc != EAGAIN) {
4671 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4672 if (!env->me_rmutex) goto fail_errno;
4673 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4674 if (!env->me_wmutex) goto fail_errno;
4675 #elif defined(MDB_USE_POSIX_SEM)
4676 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4677 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4678 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4679 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4680 #elif defined(MDB_USE_SYSV_SEM)
4681 semid = env->me_txns->mti_semid;
4683 /* check for read access */
4684 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4686 /* check for write access */
4687 if (semctl(semid, 0, IPC_SET, semu) < 0)
4691 #ifdef MDB_USE_SYSV_SEM
4692 env->me_rmutex->semid = semid;
4693 env->me_wmutex->semid = semid;
4694 env->me_rmutex->semnum = 0;
4695 env->me_wmutex->semnum = 1;
4696 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4697 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4708 /** The name of the lock file in the DB environment */
4709 #define LOCKNAME "/lock.mdb"
4710 /** The name of the data file in the DB environment */
4711 #define DATANAME "/data.mdb"
4712 /** The suffix of the lock file when no subdir is used */
4713 #define LOCKSUFF "-lock"
4714 /** Only a subset of the @ref mdb_env flags can be changed
4715 * at runtime. Changing other flags requires closing the
4716 * environment and re-opening it with the new flags.
4718 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4719 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4720 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4722 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4723 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4727 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4729 int oflags, rc, len, excl = -1;
4730 char *lpath, *dpath;
4732 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4736 if (flags & MDB_NOSUBDIR) {
4737 rc = len + sizeof(LOCKSUFF) + len + 1;
4739 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4744 if (flags & MDB_NOSUBDIR) {
4745 dpath = lpath + len + sizeof(LOCKSUFF);
4746 sprintf(lpath, "%s" LOCKSUFF, path);
4747 strcpy(dpath, path);
4749 dpath = lpath + len + sizeof(LOCKNAME);
4750 sprintf(lpath, "%s" LOCKNAME, path);
4751 sprintf(dpath, "%s" DATANAME, path);
4755 flags |= env->me_flags;
4756 if (flags & MDB_RDONLY) {
4757 /* silently ignore WRITEMAP when we're only getting read access */
4758 flags &= ~MDB_WRITEMAP;
4760 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4761 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4764 env->me_flags = flags |= MDB_ENV_ACTIVE;
4768 env->me_path = strdup(path);
4769 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4770 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4771 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4772 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4776 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4778 /* For RDONLY, get lockfile after we know datafile exists */
4779 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4780 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4786 if (F_ISSET(flags, MDB_RDONLY)) {
4787 oflags = GENERIC_READ;
4788 len = OPEN_EXISTING;
4790 oflags = GENERIC_READ|GENERIC_WRITE;
4793 mode = FILE_ATTRIBUTE_NORMAL;
4794 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4795 NULL, len, mode, NULL);
4797 if (F_ISSET(flags, MDB_RDONLY))
4800 oflags = O_RDWR | O_CREAT;
4802 env->me_fd = open(dpath, oflags, mode);
4804 if (env->me_fd == INVALID_HANDLE_VALUE) {
4809 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4810 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4815 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4816 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4817 env->me_mfd = env->me_fd;
4819 /* Synchronous fd for meta writes. Needed even with
4820 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4823 len = OPEN_EXISTING;
4824 env->me_mfd = CreateFile(dpath, oflags,
4825 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4826 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4829 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4831 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4836 DPRINTF(("opened dbenv %p", (void *) env));
4838 rc = mdb_env_share_locks(env, &excl);
4842 if (!(flags & MDB_RDONLY)) {
4844 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4845 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4846 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4847 (txn = calloc(1, size)))
4849 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4850 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4851 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4852 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4854 txn->mt_dbxs = env->me_dbxs;
4855 txn->mt_flags = MDB_TXN_FINISHED;
4865 mdb_env_close0(env, excl);
4871 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4873 mdb_env_close0(MDB_env *env, int excl)
4877 if (!(env->me_flags & MDB_ENV_ACTIVE))
4880 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4882 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4883 free(env->me_dbxs[i].md_name.mv_data);
4888 free(env->me_dbiseqs);
4889 free(env->me_dbflags);
4891 free(env->me_dirty_list);
4893 mdb_midl_free(env->me_free_pgs);
4895 if (env->me_flags & MDB_ENV_TXKEY) {
4896 pthread_key_delete(env->me_txkey);
4898 /* Delete our key from the global list */
4899 for (i=0; i<mdb_tls_nkeys; i++)
4900 if (mdb_tls_keys[i] == env->me_txkey) {
4901 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4909 munmap(env->me_map, env->me_mapsize);
4911 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4912 (void) close(env->me_mfd);
4913 if (env->me_fd != INVALID_HANDLE_VALUE)
4914 (void) close(env->me_fd);
4916 MDB_PID_T pid = env->me_pid;
4917 /* Clearing readers is done in this function because
4918 * me_txkey with its destructor must be disabled first.
4920 * We skip the the reader mutex, so we touch only
4921 * data owned by this process (me_close_readers and
4922 * our readers), and clear each reader atomically.
4924 for (i = env->me_close_readers; --i >= 0; )
4925 if (env->me_txns->mti_readers[i].mr_pid == pid)
4926 env->me_txns->mti_readers[i].mr_pid = 0;
4928 if (env->me_rmutex) {
4929 CloseHandle(env->me_rmutex);
4930 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4932 /* Windows automatically destroys the mutexes when
4933 * the last handle closes.
4935 #elif defined(MDB_USE_POSIX_SEM)
4936 if (env->me_rmutex != SEM_FAILED) {
4937 sem_close(env->me_rmutex);
4938 if (env->me_wmutex != SEM_FAILED)
4939 sem_close(env->me_wmutex);
4940 /* If we have the filelock: If we are the
4941 * only remaining user, clean up semaphores.
4944 mdb_env_excl_lock(env, &excl);
4946 sem_unlink(env->me_txns->mti_rmname);
4947 sem_unlink(env->me_txns->mti_wmname);
4950 #elif defined(MDB_USE_SYSV_SEM)
4951 if (env->me_rmutex->semid != -1) {
4952 /* If we have the filelock: If we are the
4953 * only remaining user, clean up semaphores.
4956 mdb_env_excl_lock(env, &excl);
4958 semctl(env->me_rmutex->semid, 0, IPC_RMID);
4961 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4963 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4966 /* Unlock the lockfile. Windows would have unlocked it
4967 * after closing anyway, but not necessarily at once.
4969 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4972 (void) close(env->me_lfd);
4975 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4979 mdb_env_close(MDB_env *env)
4986 VGMEMP_DESTROY(env);
4987 while ((dp = env->me_dpages) != NULL) {
4988 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4989 env->me_dpages = dp->mp_next;
4993 mdb_env_close0(env, 0);
4997 /** Compare two items pointing at aligned size_t's */
4999 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5001 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5002 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5005 /** Compare two items pointing at aligned unsigned int's.
5007 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5008 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5011 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5013 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5014 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5017 /** Compare two items pointing at unsigned ints of unknown alignment.
5018 * Nodes and keys are guaranteed to be 2-byte aligned.
5021 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5023 #if BYTE_ORDER == LITTLE_ENDIAN
5024 unsigned short *u, *c;
5027 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5028 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5031 } while(!x && u > (unsigned short *)a->mv_data);
5034 unsigned short *u, *c, *end;
5037 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5038 u = (unsigned short *)a->mv_data;
5039 c = (unsigned short *)b->mv_data;
5042 } while(!x && u < end);
5047 /** Compare two items lexically */
5049 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5056 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5062 diff = memcmp(a->mv_data, b->mv_data, len);
5063 return diff ? diff : len_diff<0 ? -1 : len_diff;
5066 /** Compare two items in reverse byte order */
5068 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5070 const unsigned char *p1, *p2, *p1_lim;
5074 p1_lim = (const unsigned char *)a->mv_data;
5075 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5076 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5078 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5084 while (p1 > p1_lim) {
5085 diff = *--p1 - *--p2;
5089 return len_diff<0 ? -1 : len_diff;
5092 /** Search for key within a page, using binary search.
5093 * Returns the smallest entry larger or equal to the key.
5094 * If exactp is non-null, stores whether the found entry was an exact match
5095 * in *exactp (1 or 0).
5096 * Updates the cursor index with the index of the found entry.
5097 * If no entry larger or equal to the key is found, returns NULL.
5100 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5102 unsigned int i = 0, nkeys;
5105 MDB_page *mp = mc->mc_pg[mc->mc_top];
5106 MDB_node *node = NULL;
5111 nkeys = NUMKEYS(mp);
5113 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5114 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5117 low = IS_LEAF(mp) ? 0 : 1;
5119 cmp = mc->mc_dbx->md_cmp;
5121 /* Branch pages have no data, so if using integer keys,
5122 * alignment is guaranteed. Use faster mdb_cmp_int.
5124 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5125 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5132 nodekey.mv_size = mc->mc_db->md_pad;
5133 node = NODEPTR(mp, 0); /* fake */
5134 while (low <= high) {
5135 i = (low + high) >> 1;
5136 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5137 rc = cmp(key, &nodekey);
5138 DPRINTF(("found leaf index %u [%s], rc = %i",
5139 i, DKEY(&nodekey), rc));
5148 while (low <= high) {
5149 i = (low + high) >> 1;
5151 node = NODEPTR(mp, i);
5152 nodekey.mv_size = NODEKSZ(node);
5153 nodekey.mv_data = NODEKEY(node);
5155 rc = cmp(key, &nodekey);
5158 DPRINTF(("found leaf index %u [%s], rc = %i",
5159 i, DKEY(&nodekey), rc));
5161 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5162 i, DKEY(&nodekey), NODEPGNO(node), rc));
5173 if (rc > 0) { /* Found entry is less than the key. */
5174 i++; /* Skip to get the smallest entry larger than key. */
5176 node = NODEPTR(mp, i);
5179 *exactp = (rc == 0 && nkeys > 0);
5180 /* store the key index */
5181 mc->mc_ki[mc->mc_top] = i;
5183 /* There is no entry larger or equal to the key. */
5186 /* nodeptr is fake for LEAF2 */
5192 mdb_cursor_adjust(MDB_cursor *mc, func)
5196 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5197 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5204 /** Pop a page off the top of the cursor's stack. */
5206 mdb_cursor_pop(MDB_cursor *mc)
5209 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5210 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5218 /** Push a page onto the top of the cursor's stack. */
5220 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5222 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5223 DDBI(mc), (void *) mc));
5225 if (mc->mc_snum >= CURSOR_STACK) {
5226 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5227 return MDB_CURSOR_FULL;
5230 mc->mc_top = mc->mc_snum++;
5231 mc->mc_pg[mc->mc_top] = mp;
5232 mc->mc_ki[mc->mc_top] = 0;
5237 /** Find the address of the page corresponding to a given page number.
5238 * @param[in] txn the transaction for this access.
5239 * @param[in] pgno the page number for the page to retrieve.
5240 * @param[out] ret address of a pointer where the page's address will be stored.
5241 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5242 * @return 0 on success, non-zero on failure.
5245 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5247 MDB_env *env = txn->mt_env;
5251 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5255 MDB_ID2L dl = tx2->mt_u.dirty_list;
5257 /* Spilled pages were dirtied in this txn and flushed
5258 * because the dirty list got full. Bring this page
5259 * back in from the map (but don't unspill it here,
5260 * leave that unless page_touch happens again).
5262 if (tx2->mt_spill_pgs) {
5263 MDB_ID pn = pgno << 1;
5264 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5265 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5266 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5271 unsigned x = mdb_mid2l_search(dl, pgno);
5272 if (x <= dl[0].mid && dl[x].mid == pgno) {
5278 } while ((tx2 = tx2->mt_parent) != NULL);
5281 if (pgno < txn->mt_next_pgno) {
5283 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5285 DPRINTF(("page %"Z"u not found", pgno));
5286 txn->mt_flags |= MDB_TXN_ERROR;
5287 return MDB_PAGE_NOTFOUND;
5297 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5298 * The cursor is at the root page, set up the rest of it.
5301 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5303 MDB_page *mp = mc->mc_pg[mc->mc_top];
5307 while (IS_BRANCH(mp)) {
5311 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5312 mdb_cassert(mc, NUMKEYS(mp) > 1);
5313 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5315 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5317 if (flags & MDB_PS_LAST)
5318 i = NUMKEYS(mp) - 1;
5321 node = mdb_node_search(mc, key, &exact);
5323 i = NUMKEYS(mp) - 1;
5325 i = mc->mc_ki[mc->mc_top];
5327 mdb_cassert(mc, i > 0);
5331 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5334 mdb_cassert(mc, i < NUMKEYS(mp));
5335 node = NODEPTR(mp, i);
5337 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5340 mc->mc_ki[mc->mc_top] = i;
5341 if ((rc = mdb_cursor_push(mc, mp)))
5344 if (flags & MDB_PS_MODIFY) {
5345 if ((rc = mdb_page_touch(mc)) != 0)
5347 mp = mc->mc_pg[mc->mc_top];
5352 DPRINTF(("internal error, index points to a %02X page!?",
5354 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5355 return MDB_CORRUPTED;
5358 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5359 key ? DKEY(key) : "null"));
5360 mc->mc_flags |= C_INITIALIZED;
5361 mc->mc_flags &= ~C_EOF;
5366 /** Search for the lowest key under the current branch page.
5367 * This just bypasses a NUMKEYS check in the current page
5368 * before calling mdb_page_search_root(), because the callers
5369 * are all in situations where the current page is known to
5373 mdb_page_search_lowest(MDB_cursor *mc)
5375 MDB_page *mp = mc->mc_pg[mc->mc_top];
5376 MDB_node *node = NODEPTR(mp, 0);
5379 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5382 mc->mc_ki[mc->mc_top] = 0;
5383 if ((rc = mdb_cursor_push(mc, mp)))
5385 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5388 /** Search for the page a given key should be in.
5389 * Push it and its parent pages on the cursor stack.
5390 * @param[in,out] mc the cursor for this operation.
5391 * @param[in] key the key to search for, or NULL for first/last page.
5392 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5393 * are touched (updated with new page numbers).
5394 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5395 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5396 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5397 * @return 0 on success, non-zero on failure.
5400 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5405 /* Make sure the txn is still viable, then find the root from
5406 * the txn's db table and set it as the root of the cursor's stack.
5408 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5409 DPUTS("transaction may not be used now");
5412 /* Make sure we're using an up-to-date root */
5413 if (*mc->mc_dbflag & DB_STALE) {
5415 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5417 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5418 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5425 MDB_node *leaf = mdb_node_search(&mc2,
5426 &mc->mc_dbx->md_name, &exact);
5428 return MDB_NOTFOUND;
5429 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5430 return MDB_INCOMPATIBLE; /* not a named DB */
5431 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5434 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5436 /* The txn may not know this DBI, or another process may
5437 * have dropped and recreated the DB with other flags.
5439 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5440 return MDB_INCOMPATIBLE;
5441 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5443 *mc->mc_dbflag &= ~DB_STALE;
5445 root = mc->mc_db->md_root;
5447 if (root == P_INVALID) { /* Tree is empty. */
5448 DPUTS("tree is empty");
5449 return MDB_NOTFOUND;
5453 mdb_cassert(mc, root > 1);
5454 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5455 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5461 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5462 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5464 if (flags & MDB_PS_MODIFY) {
5465 if ((rc = mdb_page_touch(mc)))
5469 if (flags & MDB_PS_ROOTONLY)
5472 return mdb_page_search_root(mc, key, flags);
5476 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5478 MDB_txn *txn = mc->mc_txn;
5479 pgno_t pg = mp->mp_pgno;
5480 unsigned x = 0, ovpages = mp->mp_pages;
5481 MDB_env *env = txn->mt_env;
5482 MDB_IDL sl = txn->mt_spill_pgs;
5483 MDB_ID pn = pg << 1;
5486 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5487 /* If the page is dirty or on the spill list we just acquired it,
5488 * so we should give it back to our current free list, if any.
5489 * Otherwise put it onto the list of pages we freed in this txn.
5491 * Won't create me_pghead: me_pglast must be inited along with it.
5492 * Unsupported in nested txns: They would need to hide the page
5493 * range in ancestor txns' dirty and spilled lists.
5495 if (env->me_pghead &&
5497 ((mp->mp_flags & P_DIRTY) ||
5498 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5502 MDB_ID2 *dl, ix, iy;
5503 rc = mdb_midl_need(&env->me_pghead, ovpages);
5506 if (!(mp->mp_flags & P_DIRTY)) {
5507 /* This page is no longer spilled */
5514 /* Remove from dirty list */
5515 dl = txn->mt_u.dirty_list;
5517 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5523 mdb_cassert(mc, x > 1);
5525 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5526 txn->mt_flags |= MDB_TXN_ERROR;
5527 return MDB_CORRUPTED;
5530 if (!(env->me_flags & MDB_WRITEMAP))
5531 mdb_dpage_free(env, mp);
5533 /* Insert in me_pghead */
5534 mop = env->me_pghead;
5535 j = mop[0] + ovpages;
5536 for (i = mop[0]; i && mop[i] < pg; i--)
5542 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5546 mc->mc_db->md_overflow_pages -= ovpages;
5550 /** Return the data associated with a given node.
5551 * @param[in] txn The transaction for this operation.
5552 * @param[in] leaf The node being read.
5553 * @param[out] data Updated to point to the node's data.
5554 * @return 0 on success, non-zero on failure.
5557 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5559 MDB_page *omp; /* overflow page */
5563 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5564 data->mv_size = NODEDSZ(leaf);
5565 data->mv_data = NODEDATA(leaf);
5569 /* Read overflow data.
5571 data->mv_size = NODEDSZ(leaf);
5572 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5573 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5574 DPRINTF(("read overflow page %"Z"u failed", pgno));
5577 data->mv_data = METADATA(omp);
5583 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5584 MDB_val *key, MDB_val *data)
5591 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5593 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5596 if (txn->mt_flags & MDB_TXN_BLOCKED)
5599 mdb_cursor_init(&mc, txn, dbi, &mx);
5600 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5603 /** Find a sibling for a page.
5604 * Replaces the page at the top of the cursor's stack with the
5605 * specified sibling, if one exists.
5606 * @param[in] mc The cursor for this operation.
5607 * @param[in] move_right Non-zero if the right sibling is requested,
5608 * otherwise the left sibling.
5609 * @return 0 on success, non-zero on failure.
5612 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5618 if (mc->mc_snum < 2) {
5619 return MDB_NOTFOUND; /* root has no siblings */
5623 DPRINTF(("parent page is page %"Z"u, index %u",
5624 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5626 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5627 : (mc->mc_ki[mc->mc_top] == 0)) {
5628 DPRINTF(("no more keys left, moving to %s sibling",
5629 move_right ? "right" : "left"));
5630 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5631 /* undo cursor_pop before returning */
5638 mc->mc_ki[mc->mc_top]++;
5640 mc->mc_ki[mc->mc_top]--;
5641 DPRINTF(("just moving to %s index key %u",
5642 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5644 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5646 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5647 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5648 /* mc will be inconsistent if caller does mc_snum++ as above */
5649 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5653 mdb_cursor_push(mc, mp);
5655 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5660 /** Move the cursor to the next data item. */
5662 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5668 if (mc->mc_flags & C_EOF) {
5669 return MDB_NOTFOUND;
5672 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5674 mp = mc->mc_pg[mc->mc_top];
5676 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5677 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5678 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5679 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5680 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5681 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5682 if (rc == MDB_SUCCESS)
5683 MDB_GET_KEY(leaf, key);
5688 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5689 if (op == MDB_NEXT_DUP)
5690 return MDB_NOTFOUND;
5694 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5695 mdb_dbg_pgno(mp), (void *) mc));
5696 if (mc->mc_flags & C_DEL)
5699 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5700 DPUTS("=====> move to next sibling page");
5701 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5702 mc->mc_flags |= C_EOF;
5705 mp = mc->mc_pg[mc->mc_top];
5706 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5708 mc->mc_ki[mc->mc_top]++;
5711 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5712 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5715 key->mv_size = mc->mc_db->md_pad;
5716 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5720 mdb_cassert(mc, IS_LEAF(mp));
5721 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5723 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5724 mdb_xcursor_init1(mc, leaf);
5727 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5730 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5731 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5732 if (rc != MDB_SUCCESS)
5737 MDB_GET_KEY(leaf, key);
5741 /** Move the cursor to the previous data item. */
5743 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5749 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5751 mp = mc->mc_pg[mc->mc_top];
5753 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5754 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5755 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5756 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5757 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5758 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5759 if (rc == MDB_SUCCESS) {
5760 MDB_GET_KEY(leaf, key);
5761 mc->mc_flags &= ~C_EOF;
5767 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5768 if (op == MDB_PREV_DUP)
5769 return MDB_NOTFOUND;
5773 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5774 mdb_dbg_pgno(mp), (void *) mc));
5776 if (mc->mc_ki[mc->mc_top] == 0) {
5777 DPUTS("=====> move to prev sibling page");
5778 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5781 mp = mc->mc_pg[mc->mc_top];
5782 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5783 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5785 mc->mc_ki[mc->mc_top]--;
5787 mc->mc_flags &= ~C_EOF;
5789 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5790 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5793 key->mv_size = mc->mc_db->md_pad;
5794 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5798 mdb_cassert(mc, IS_LEAF(mp));
5799 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5801 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5802 mdb_xcursor_init1(mc, leaf);
5805 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5808 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5809 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5810 if (rc != MDB_SUCCESS)
5815 MDB_GET_KEY(leaf, key);
5819 /** Set the cursor on a specific data item. */
5821 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5822 MDB_cursor_op op, int *exactp)
5826 MDB_node *leaf = NULL;
5829 if (key->mv_size == 0)
5830 return MDB_BAD_VALSIZE;
5833 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5835 /* See if we're already on the right page */
5836 if (mc->mc_flags & C_INITIALIZED) {
5839 mp = mc->mc_pg[mc->mc_top];
5841 mc->mc_ki[mc->mc_top] = 0;
5842 return MDB_NOTFOUND;
5844 if (mp->mp_flags & P_LEAF2) {
5845 nodekey.mv_size = mc->mc_db->md_pad;
5846 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5848 leaf = NODEPTR(mp, 0);
5849 MDB_GET_KEY2(leaf, nodekey);
5851 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5853 /* Probably happens rarely, but first node on the page
5854 * was the one we wanted.
5856 mc->mc_ki[mc->mc_top] = 0;
5863 unsigned int nkeys = NUMKEYS(mp);
5865 if (mp->mp_flags & P_LEAF2) {
5866 nodekey.mv_data = LEAF2KEY(mp,
5867 nkeys-1, nodekey.mv_size);
5869 leaf = NODEPTR(mp, nkeys-1);
5870 MDB_GET_KEY2(leaf, nodekey);
5872 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5874 /* last node was the one we wanted */
5875 mc->mc_ki[mc->mc_top] = nkeys-1;
5881 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5882 /* This is definitely the right page, skip search_page */
5883 if (mp->mp_flags & P_LEAF2) {
5884 nodekey.mv_data = LEAF2KEY(mp,
5885 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5887 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5888 MDB_GET_KEY2(leaf, nodekey);
5890 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5892 /* current node was the one we wanted */
5902 /* If any parents have right-sibs, search.
5903 * Otherwise, there's nothing further.
5905 for (i=0; i<mc->mc_top; i++)
5907 NUMKEYS(mc->mc_pg[i])-1)
5909 if (i == mc->mc_top) {
5910 /* There are no other pages */
5911 mc->mc_ki[mc->mc_top] = nkeys;
5912 return MDB_NOTFOUND;
5916 /* There are no other pages */
5917 mc->mc_ki[mc->mc_top] = 0;
5918 if (op == MDB_SET_RANGE && !exactp) {
5922 return MDB_NOTFOUND;
5926 rc = mdb_page_search(mc, key, 0);
5927 if (rc != MDB_SUCCESS)
5930 mp = mc->mc_pg[mc->mc_top];
5931 mdb_cassert(mc, IS_LEAF(mp));
5934 leaf = mdb_node_search(mc, key, exactp);
5935 if (exactp != NULL && !*exactp) {
5936 /* MDB_SET specified and not an exact match. */
5937 return MDB_NOTFOUND;
5941 DPUTS("===> inexact leaf not found, goto sibling");
5942 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5943 mc->mc_flags |= C_EOF;
5944 return rc; /* no entries matched */
5946 mp = mc->mc_pg[mc->mc_top];
5947 mdb_cassert(mc, IS_LEAF(mp));
5948 leaf = NODEPTR(mp, 0);
5952 mc->mc_flags |= C_INITIALIZED;
5953 mc->mc_flags &= ~C_EOF;
5956 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5957 key->mv_size = mc->mc_db->md_pad;
5958 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5963 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5964 mdb_xcursor_init1(mc, leaf);
5967 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5968 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5969 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5972 if (op == MDB_GET_BOTH) {
5978 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5979 if (rc != MDB_SUCCESS)
5982 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5985 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5987 dcmp = mc->mc_dbx->md_dcmp;
5988 #if UINT_MAX < SIZE_MAX
5989 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5990 dcmp = mdb_cmp_clong;
5992 rc = dcmp(data, &olddata);
5994 if (op == MDB_GET_BOTH || rc > 0)
5995 return MDB_NOTFOUND;
6002 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6003 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6008 /* The key already matches in all other cases */
6009 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6010 MDB_GET_KEY(leaf, key);
6011 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6016 /** Move the cursor to the first item in the database. */
6018 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6024 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6026 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6027 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6028 if (rc != MDB_SUCCESS)
6031 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6033 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6034 mc->mc_flags |= C_INITIALIZED;
6035 mc->mc_flags &= ~C_EOF;
6037 mc->mc_ki[mc->mc_top] = 0;
6039 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6040 key->mv_size = mc->mc_db->md_pad;
6041 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6046 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6047 mdb_xcursor_init1(mc, leaf);
6048 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6052 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6056 MDB_GET_KEY(leaf, key);
6060 /** Move the cursor to the last item in the database. */
6062 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6068 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6070 if (!(mc->mc_flags & C_EOF)) {
6072 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6073 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6074 if (rc != MDB_SUCCESS)
6077 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6080 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6081 mc->mc_flags |= C_INITIALIZED|C_EOF;
6082 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6084 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6085 key->mv_size = mc->mc_db->md_pad;
6086 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6091 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6092 mdb_xcursor_init1(mc, leaf);
6093 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6097 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6102 MDB_GET_KEY(leaf, key);
6107 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6112 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6117 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6121 case MDB_GET_CURRENT:
6122 if (!(mc->mc_flags & C_INITIALIZED)) {
6125 MDB_page *mp = mc->mc_pg[mc->mc_top];
6126 int nkeys = NUMKEYS(mp);
6127 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6128 mc->mc_ki[mc->mc_top] = nkeys;
6134 key->mv_size = mc->mc_db->md_pad;
6135 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6137 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6138 MDB_GET_KEY(leaf, key);
6140 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6141 if (mc->mc_flags & C_DEL)
6142 mdb_xcursor_init1(mc, leaf);
6143 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6145 rc = mdb_node_read(mc->mc_txn, leaf, data);
6152 case MDB_GET_BOTH_RANGE:
6157 if (mc->mc_xcursor == NULL) {
6158 rc = MDB_INCOMPATIBLE;
6168 rc = mdb_cursor_set(mc, key, data, op,
6169 op == MDB_SET_RANGE ? NULL : &exact);
6172 case MDB_GET_MULTIPLE:
6173 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6177 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6178 rc = MDB_INCOMPATIBLE;
6182 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6183 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6186 case MDB_NEXT_MULTIPLE:
6191 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6192 rc = MDB_INCOMPATIBLE;
6195 if (!(mc->mc_flags & C_INITIALIZED))
6196 rc = mdb_cursor_first(mc, key, data);
6198 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6199 if (rc == MDB_SUCCESS) {
6200 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6203 mx = &mc->mc_xcursor->mx_cursor;
6204 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6206 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6207 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6215 case MDB_NEXT_NODUP:
6216 if (!(mc->mc_flags & C_INITIALIZED))
6217 rc = mdb_cursor_first(mc, key, data);
6219 rc = mdb_cursor_next(mc, key, data, op);
6223 case MDB_PREV_NODUP:
6224 if (!(mc->mc_flags & C_INITIALIZED)) {
6225 rc = mdb_cursor_last(mc, key, data);
6228 mc->mc_flags |= C_INITIALIZED;
6229 mc->mc_ki[mc->mc_top]++;
6231 rc = mdb_cursor_prev(mc, key, data, op);
6234 rc = mdb_cursor_first(mc, key, data);
6237 mfunc = mdb_cursor_first;
6239 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6243 if (mc->mc_xcursor == NULL) {
6244 rc = MDB_INCOMPATIBLE;
6248 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6249 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6250 MDB_GET_KEY(leaf, key);
6251 rc = mdb_node_read(mc->mc_txn, leaf, data);
6255 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6259 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6262 rc = mdb_cursor_last(mc, key, data);
6265 mfunc = mdb_cursor_last;
6268 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6273 if (mc->mc_flags & C_DEL)
6274 mc->mc_flags ^= C_DEL;
6279 /** Touch all the pages in the cursor stack. Set mc_top.
6280 * Makes sure all the pages are writable, before attempting a write operation.
6281 * @param[in] mc The cursor to operate on.
6284 mdb_cursor_touch(MDB_cursor *mc)
6286 int rc = MDB_SUCCESS;
6288 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6291 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6293 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6294 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6297 *mc->mc_dbflag |= DB_DIRTY;
6302 rc = mdb_page_touch(mc);
6303 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6304 mc->mc_top = mc->mc_snum-1;
6309 /** Do not spill pages to disk if txn is getting full, may fail instead */
6310 #define MDB_NOSPILL 0x8000
6313 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6317 MDB_node *leaf = NULL;
6320 MDB_val xdata, *rdata, dkey, olddata;
6322 int do_sub = 0, insert_key, insert_data;
6323 unsigned int mcount = 0, dcount = 0, nospill;
6326 unsigned int nflags;
6329 if (mc == NULL || key == NULL)
6332 env = mc->mc_txn->mt_env;
6334 /* Check this first so counter will always be zero on any
6337 if (flags & MDB_MULTIPLE) {
6338 dcount = data[1].mv_size;
6339 data[1].mv_size = 0;
6340 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6341 return MDB_INCOMPATIBLE;
6344 nospill = flags & MDB_NOSPILL;
6345 flags &= ~MDB_NOSPILL;
6347 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6348 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6350 if (key->mv_size-1 >= ENV_MAXKEY(env))
6351 return MDB_BAD_VALSIZE;
6353 #if SIZE_MAX > MAXDATASIZE
6354 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6355 return MDB_BAD_VALSIZE;
6357 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6358 return MDB_BAD_VALSIZE;
6361 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6362 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6366 if (flags == MDB_CURRENT) {
6367 if (!(mc->mc_flags & C_INITIALIZED))
6370 } else if (mc->mc_db->md_root == P_INVALID) {
6371 /* new database, cursor has nothing to point to */
6374 mc->mc_flags &= ~C_INITIALIZED;
6379 if (flags & MDB_APPEND) {
6381 rc = mdb_cursor_last(mc, &k2, &d2);
6383 rc = mc->mc_dbx->md_cmp(key, &k2);
6386 mc->mc_ki[mc->mc_top]++;
6388 /* new key is <= last key */
6393 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6395 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6396 DPRINTF(("duplicate key [%s]", DKEY(key)));
6398 return MDB_KEYEXIST;
6400 if (rc && rc != MDB_NOTFOUND)
6404 if (mc->mc_flags & C_DEL)
6405 mc->mc_flags ^= C_DEL;
6407 /* Cursor is positioned, check for room in the dirty list */
6409 if (flags & MDB_MULTIPLE) {
6411 xdata.mv_size = data->mv_size * dcount;
6415 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6419 if (rc == MDB_NO_ROOT) {
6421 /* new database, write a root leaf page */
6422 DPUTS("allocating new root leaf page");
6423 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6426 mdb_cursor_push(mc, np);
6427 mc->mc_db->md_root = np->mp_pgno;
6428 mc->mc_db->md_depth++;
6429 *mc->mc_dbflag |= DB_DIRTY;
6430 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6432 np->mp_flags |= P_LEAF2;
6433 mc->mc_flags |= C_INITIALIZED;
6435 /* make sure all cursor pages are writable */
6436 rc2 = mdb_cursor_touch(mc);
6441 insert_key = insert_data = rc;
6443 /* The key does not exist */
6444 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6445 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6446 LEAFSIZE(key, data) > env->me_nodemax)
6448 /* Too big for a node, insert in sub-DB. Set up an empty
6449 * "old sub-page" for prep_subDB to expand to a full page.
6451 fp_flags = P_LEAF|P_DIRTY;
6453 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6454 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6455 olddata.mv_size = PAGEHDRSZ;
6459 /* there's only a key anyway, so this is a no-op */
6460 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6462 unsigned int ksize = mc->mc_db->md_pad;
6463 if (key->mv_size != ksize)
6464 return MDB_BAD_VALSIZE;
6465 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6466 memcpy(ptr, key->mv_data, ksize);
6468 /* if overwriting slot 0 of leaf, need to
6469 * update branch key if there is a parent page
6471 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6472 unsigned short top = mc->mc_top;
6474 /* slot 0 is always an empty key, find real slot */
6475 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6477 if (mc->mc_ki[mc->mc_top])
6478 rc2 = mdb_update_key(mc, key);
6489 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6490 olddata.mv_size = NODEDSZ(leaf);
6491 olddata.mv_data = NODEDATA(leaf);
6494 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6495 /* Prepare (sub-)page/sub-DB to accept the new item,
6496 * if needed. fp: old sub-page or a header faking
6497 * it. mp: new (sub-)page. offset: growth in page
6498 * size. xdata: node data with new page or DB.
6500 unsigned i, offset = 0;
6501 mp = fp = xdata.mv_data = env->me_pbuf;
6502 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6504 /* Was a single item before, must convert now */
6505 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6507 /* Just overwrite the current item */
6508 if (flags == MDB_CURRENT)
6510 dcmp = mc->mc_dbx->md_dcmp;
6511 #if UINT_MAX < SIZE_MAX
6512 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6513 dcmp = mdb_cmp_clong;
6515 /* does data match? */
6516 if (!dcmp(data, &olddata)) {
6517 if (flags & MDB_NODUPDATA)
6518 return MDB_KEYEXIST;
6523 /* Back up original data item */
6524 dkey.mv_size = olddata.mv_size;
6525 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6527 /* Make sub-page header for the dup items, with dummy body */
6528 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6529 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6530 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6531 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6532 fp->mp_flags |= P_LEAF2;
6533 fp->mp_pad = data->mv_size;
6534 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6536 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6537 (dkey.mv_size & 1) + (data->mv_size & 1);
6539 fp->mp_upper = xdata.mv_size - PAGEBASE;
6540 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6541 } else if (leaf->mn_flags & F_SUBDATA) {
6542 /* Data is on sub-DB, just store it */
6543 flags |= F_DUPDATA|F_SUBDATA;
6546 /* Data is on sub-page */
6547 fp = olddata.mv_data;
6550 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6551 offset = EVEN(NODESIZE + sizeof(indx_t) +
6555 offset = fp->mp_pad;
6556 if (SIZELEFT(fp) < offset) {
6557 offset *= 4; /* space for 4 more */
6560 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6562 fp->mp_flags |= P_DIRTY;
6563 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6564 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6568 xdata.mv_size = olddata.mv_size + offset;
6571 fp_flags = fp->mp_flags;
6572 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6573 /* Too big for a sub-page, convert to sub-DB */
6574 fp_flags &= ~P_SUBP;
6576 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6577 fp_flags |= P_LEAF2;
6578 dummy.md_pad = fp->mp_pad;
6579 dummy.md_flags = MDB_DUPFIXED;
6580 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6581 dummy.md_flags |= MDB_INTEGERKEY;
6587 dummy.md_branch_pages = 0;
6588 dummy.md_leaf_pages = 1;
6589 dummy.md_overflow_pages = 0;
6590 dummy.md_entries = NUMKEYS(fp);
6591 xdata.mv_size = sizeof(MDB_db);
6592 xdata.mv_data = &dummy;
6593 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6595 offset = env->me_psize - olddata.mv_size;
6596 flags |= F_DUPDATA|F_SUBDATA;
6597 dummy.md_root = mp->mp_pgno;
6600 mp->mp_flags = fp_flags | P_DIRTY;
6601 mp->mp_pad = fp->mp_pad;
6602 mp->mp_lower = fp->mp_lower;
6603 mp->mp_upper = fp->mp_upper + offset;
6604 if (fp_flags & P_LEAF2) {
6605 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6607 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6608 olddata.mv_size - fp->mp_upper - PAGEBASE);
6609 for (i=0; i<NUMKEYS(fp); i++)
6610 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6618 mdb_node_del(mc, 0);
6622 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6623 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6624 return MDB_INCOMPATIBLE;
6625 /* overflow page overwrites need special handling */
6626 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6629 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6631 memcpy(&pg, olddata.mv_data, sizeof(pg));
6632 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6634 ovpages = omp->mp_pages;
6636 /* Is the ov page large enough? */
6637 if (ovpages >= dpages) {
6638 if (!(omp->mp_flags & P_DIRTY) &&
6639 (level || (env->me_flags & MDB_WRITEMAP)))
6641 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6644 level = 0; /* dirty in this txn or clean */
6647 if (omp->mp_flags & P_DIRTY) {
6648 /* yes, overwrite it. Note in this case we don't
6649 * bother to try shrinking the page if the new data
6650 * is smaller than the overflow threshold.
6653 /* It is writable only in a parent txn */
6654 size_t sz = (size_t) env->me_psize * ovpages, off;
6655 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6661 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6662 mdb_cassert(mc, rc2 == 0);
6663 if (!(flags & MDB_RESERVE)) {
6664 /* Copy end of page, adjusting alignment so
6665 * compiler may copy words instead of bytes.
6667 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6668 memcpy((size_t *)((char *)np + off),
6669 (size_t *)((char *)omp + off), sz - off);
6672 memcpy(np, omp, sz); /* Copy beginning of page */
6675 SETDSZ(leaf, data->mv_size);
6676 if (F_ISSET(flags, MDB_RESERVE))
6677 data->mv_data = METADATA(omp);
6679 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6683 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6685 } else if (data->mv_size == olddata.mv_size) {
6686 /* same size, just replace it. Note that we could
6687 * also reuse this node if the new data is smaller,
6688 * but instead we opt to shrink the node in that case.
6690 if (F_ISSET(flags, MDB_RESERVE))
6691 data->mv_data = olddata.mv_data;
6692 else if (!(mc->mc_flags & C_SUB))
6693 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6695 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6700 mdb_node_del(mc, 0);
6706 nflags = flags & NODE_ADD_FLAGS;
6707 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6708 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6709 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6710 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6712 nflags |= MDB_SPLIT_REPLACE;
6713 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6715 /* There is room already in this leaf page. */
6716 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6717 if (rc == 0 && insert_key) {
6718 /* Adjust other cursors pointing to mp */
6719 MDB_cursor *m2, *m3;
6720 MDB_dbi dbi = mc->mc_dbi;
6721 unsigned i = mc->mc_top;
6722 MDB_page *mp = mc->mc_pg[i];
6724 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6725 if (mc->mc_flags & C_SUB)
6726 m3 = &m2->mc_xcursor->mx_cursor;
6729 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6730 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6737 if (rc == MDB_SUCCESS) {
6738 /* Now store the actual data in the child DB. Note that we're
6739 * storing the user data in the keys field, so there are strict
6740 * size limits on dupdata. The actual data fields of the child
6741 * DB are all zero size.
6749 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6750 if (flags & MDB_CURRENT) {
6751 xflags = MDB_CURRENT|MDB_NOSPILL;
6753 mdb_xcursor_init1(mc, leaf);
6754 xflags = (flags & MDB_NODUPDATA) ?
6755 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6757 /* converted, write the original data first */
6759 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6763 /* Adjust other cursors pointing to mp */
6765 unsigned i = mc->mc_top;
6766 MDB_page *mp = mc->mc_pg[i];
6768 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6769 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6770 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6771 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6772 mdb_xcursor_init1(m2, leaf);
6776 /* we've done our job */
6779 ecount = mc->mc_xcursor->mx_db.md_entries;
6780 if (flags & MDB_APPENDDUP)
6781 xflags |= MDB_APPEND;
6782 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6783 if (flags & F_SUBDATA) {
6784 void *db = NODEDATA(leaf);
6785 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6787 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6789 /* Increment count unless we just replaced an existing item. */
6791 mc->mc_db->md_entries++;
6793 /* Invalidate txn if we created an empty sub-DB */
6796 /* If we succeeded and the key didn't exist before,
6797 * make sure the cursor is marked valid.
6799 mc->mc_flags |= C_INITIALIZED;
6801 if (flags & MDB_MULTIPLE) {
6804 /* let caller know how many succeeded, if any */
6805 data[1].mv_size = mcount;
6806 if (mcount < dcount) {
6807 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6808 insert_key = insert_data = 0;
6815 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6818 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6823 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6829 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6830 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6832 if (!(mc->mc_flags & C_INITIALIZED))
6835 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6836 return MDB_NOTFOUND;
6838 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6841 rc = mdb_cursor_touch(mc);
6845 mp = mc->mc_pg[mc->mc_top];
6848 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6850 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6851 if (flags & MDB_NODUPDATA) {
6852 /* mdb_cursor_del0() will subtract the final entry */
6853 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6855 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6856 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6858 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6861 /* If sub-DB still has entries, we're done */
6862 if (mc->mc_xcursor->mx_db.md_entries) {
6863 if (leaf->mn_flags & F_SUBDATA) {
6864 /* update subDB info */
6865 void *db = NODEDATA(leaf);
6866 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6869 /* shrink fake page */
6870 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6871 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6872 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6873 /* fix other sub-DB cursors pointed at this fake page */
6874 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6875 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6876 if (m2->mc_pg[mc->mc_top] == mp &&
6877 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6878 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6881 mc->mc_db->md_entries--;
6882 mc->mc_flags |= C_DEL;
6885 /* otherwise fall thru and delete the sub-DB */
6888 if (leaf->mn_flags & F_SUBDATA) {
6889 /* add all the child DB's pages to the free list */
6890 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6895 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6896 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6897 rc = MDB_INCOMPATIBLE;
6901 /* add overflow pages to free list */
6902 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6906 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6907 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6908 (rc = mdb_ovpage_free(mc, omp)))
6913 return mdb_cursor_del0(mc);
6916 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6920 /** Allocate and initialize new pages for a database.
6921 * @param[in] mc a cursor on the database being added to.
6922 * @param[in] flags flags defining what type of page is being allocated.
6923 * @param[in] num the number of pages to allocate. This is usually 1,
6924 * unless allocating overflow pages for a large record.
6925 * @param[out] mp Address of a page, or NULL on failure.
6926 * @return 0 on success, non-zero on failure.
6929 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6934 if ((rc = mdb_page_alloc(mc, num, &np)))
6936 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6937 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6938 np->mp_flags = flags | P_DIRTY;
6939 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6940 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6943 mc->mc_db->md_branch_pages++;
6944 else if (IS_LEAF(np))
6945 mc->mc_db->md_leaf_pages++;
6946 else if (IS_OVERFLOW(np)) {
6947 mc->mc_db->md_overflow_pages += num;
6955 /** Calculate the size of a leaf node.
6956 * The size depends on the environment's page size; if a data item
6957 * is too large it will be put onto an overflow page and the node
6958 * size will only include the key and not the data. Sizes are always
6959 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6960 * of the #MDB_node headers.
6961 * @param[in] env The environment handle.
6962 * @param[in] key The key for the node.
6963 * @param[in] data The data for the node.
6964 * @return The number of bytes needed to store the node.
6967 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6971 sz = LEAFSIZE(key, data);
6972 if (sz > env->me_nodemax) {
6973 /* put on overflow page */
6974 sz -= data->mv_size - sizeof(pgno_t);
6977 return EVEN(sz + sizeof(indx_t));
6980 /** Calculate the size of a branch node.
6981 * The size should depend on the environment's page size but since
6982 * we currently don't support spilling large keys onto overflow
6983 * pages, it's simply the size of the #MDB_node header plus the
6984 * size of the key. Sizes are always rounded up to an even number
6985 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6986 * @param[in] env The environment handle.
6987 * @param[in] key The key for the node.
6988 * @return The number of bytes needed to store the node.
6991 mdb_branch_size(MDB_env *env, MDB_val *key)
6996 if (sz > env->me_nodemax) {
6997 /* put on overflow page */
6998 /* not implemented */
6999 /* sz -= key->size - sizeof(pgno_t); */
7002 return sz + sizeof(indx_t);
7005 /** Add a node to the page pointed to by the cursor.
7006 * @param[in] mc The cursor for this operation.
7007 * @param[in] indx The index on the page where the new node should be added.
7008 * @param[in] key The key for the new node.
7009 * @param[in] data The data for the new node, if any.
7010 * @param[in] pgno The page number, if adding a branch node.
7011 * @param[in] flags Flags for the node.
7012 * @return 0 on success, non-zero on failure. Possible errors are:
7014 * <li>ENOMEM - failed to allocate overflow pages for the node.
7015 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7016 * should never happen since all callers already calculate the
7017 * page's free space before calling this function.
7021 mdb_node_add(MDB_cursor *mc, indx_t indx,
7022 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7025 size_t node_size = NODESIZE;
7029 MDB_page *mp = mc->mc_pg[mc->mc_top];
7030 MDB_page *ofp = NULL; /* overflow page */
7034 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7036 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7037 IS_LEAF(mp) ? "leaf" : "branch",
7038 IS_SUBP(mp) ? "sub-" : "",
7039 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7040 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7043 /* Move higher keys up one slot. */
7044 int ksize = mc->mc_db->md_pad, dif;
7045 char *ptr = LEAF2KEY(mp, indx, ksize);
7046 dif = NUMKEYS(mp) - indx;
7048 memmove(ptr+ksize, ptr, dif*ksize);
7049 /* insert new key */
7050 memcpy(ptr, key->mv_data, ksize);
7052 /* Just using these for counting */
7053 mp->mp_lower += sizeof(indx_t);
7054 mp->mp_upper -= ksize - sizeof(indx_t);
7058 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7060 node_size += key->mv_size;
7062 mdb_cassert(mc, key && data);
7063 if (F_ISSET(flags, F_BIGDATA)) {
7064 /* Data already on overflow page. */
7065 node_size += sizeof(pgno_t);
7066 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7067 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7069 /* Put data on overflow page. */
7070 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7071 data->mv_size, node_size+data->mv_size));
7072 node_size = EVEN(node_size + sizeof(pgno_t));
7073 if ((ssize_t)node_size > room)
7075 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7077 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7081 node_size += data->mv_size;
7084 node_size = EVEN(node_size);
7085 if ((ssize_t)node_size > room)
7089 /* Move higher pointers up one slot. */
7090 for (i = NUMKEYS(mp); i > indx; i--)
7091 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7093 /* Adjust free space offsets. */
7094 ofs = mp->mp_upper - node_size;
7095 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7096 mp->mp_ptrs[indx] = ofs;
7098 mp->mp_lower += sizeof(indx_t);
7100 /* Write the node data. */
7101 node = NODEPTR(mp, indx);
7102 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7103 node->mn_flags = flags;
7105 SETDSZ(node,data->mv_size);
7110 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7113 ndata = NODEDATA(node);
7115 if (F_ISSET(flags, F_BIGDATA))
7116 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7117 else if (F_ISSET(flags, MDB_RESERVE))
7118 data->mv_data = ndata;
7120 memcpy(ndata, data->mv_data, data->mv_size);
7122 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7123 ndata = METADATA(ofp);
7124 if (F_ISSET(flags, MDB_RESERVE))
7125 data->mv_data = ndata;
7127 memcpy(ndata, data->mv_data, data->mv_size);
7134 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7135 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7136 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7137 DPRINTF(("node size = %"Z"u", node_size));
7138 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7139 return MDB_PAGE_FULL;
7142 /** Delete the specified node from a page.
7143 * @param[in] mc Cursor pointing to the node to delete.
7144 * @param[in] ksize The size of a node. Only used if the page is
7145 * part of a #MDB_DUPFIXED database.
7148 mdb_node_del(MDB_cursor *mc, int ksize)
7150 MDB_page *mp = mc->mc_pg[mc->mc_top];
7151 indx_t indx = mc->mc_ki[mc->mc_top];
7153 indx_t i, j, numkeys, ptr;
7157 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7158 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7159 numkeys = NUMKEYS(mp);
7160 mdb_cassert(mc, indx < numkeys);
7163 int x = numkeys - 1 - indx;
7164 base = LEAF2KEY(mp, indx, ksize);
7166 memmove(base, base + ksize, x * ksize);
7167 mp->mp_lower -= sizeof(indx_t);
7168 mp->mp_upper += ksize - sizeof(indx_t);
7172 node = NODEPTR(mp, indx);
7173 sz = NODESIZE + node->mn_ksize;
7175 if (F_ISSET(node->mn_flags, F_BIGDATA))
7176 sz += sizeof(pgno_t);
7178 sz += NODEDSZ(node);
7182 ptr = mp->mp_ptrs[indx];
7183 for (i = j = 0; i < numkeys; i++) {
7185 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7186 if (mp->mp_ptrs[i] < ptr)
7187 mp->mp_ptrs[j] += sz;
7192 base = (char *)mp + mp->mp_upper + PAGEBASE;
7193 memmove(base + sz, base, ptr - mp->mp_upper);
7195 mp->mp_lower -= sizeof(indx_t);
7199 /** Compact the main page after deleting a node on a subpage.
7200 * @param[in] mp The main page to operate on.
7201 * @param[in] indx The index of the subpage on the main page.
7204 mdb_node_shrink(MDB_page *mp, indx_t indx)
7209 indx_t delta, nsize, len, ptr;
7212 node = NODEPTR(mp, indx);
7213 sp = (MDB_page *)NODEDATA(node);
7214 delta = SIZELEFT(sp);
7215 nsize = NODEDSZ(node) - delta;
7217 /* Prepare to shift upward, set len = length(subpage part to shift) */
7221 return; /* do not make the node uneven-sized */
7223 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7224 for (i = NUMKEYS(sp); --i >= 0; )
7225 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7228 sp->mp_upper = sp->mp_lower;
7229 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7230 SETDSZ(node, nsize);
7232 /* Shift <lower nodes...initial part of subpage> upward */
7233 base = (char *)mp + mp->mp_upper + PAGEBASE;
7234 memmove(base + delta, base, (char *)sp + len - base);
7236 ptr = mp->mp_ptrs[indx];
7237 for (i = NUMKEYS(mp); --i >= 0; ) {
7238 if (mp->mp_ptrs[i] <= ptr)
7239 mp->mp_ptrs[i] += delta;
7241 mp->mp_upper += delta;
7244 /** Initial setup of a sorted-dups cursor.
7245 * Sorted duplicates are implemented as a sub-database for the given key.
7246 * The duplicate data items are actually keys of the sub-database.
7247 * Operations on the duplicate data items are performed using a sub-cursor
7248 * initialized when the sub-database is first accessed. This function does
7249 * the preliminary setup of the sub-cursor, filling in the fields that
7250 * depend only on the parent DB.
7251 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7254 mdb_xcursor_init0(MDB_cursor *mc)
7256 MDB_xcursor *mx = mc->mc_xcursor;
7258 mx->mx_cursor.mc_xcursor = NULL;
7259 mx->mx_cursor.mc_txn = mc->mc_txn;
7260 mx->mx_cursor.mc_db = &mx->mx_db;
7261 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7262 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7263 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7264 mx->mx_cursor.mc_snum = 0;
7265 mx->mx_cursor.mc_top = 0;
7266 mx->mx_cursor.mc_flags = C_SUB;
7267 mx->mx_dbx.md_name.mv_size = 0;
7268 mx->mx_dbx.md_name.mv_data = NULL;
7269 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7270 mx->mx_dbx.md_dcmp = NULL;
7271 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7274 /** Final setup of a sorted-dups cursor.
7275 * Sets up the fields that depend on the data from the main cursor.
7276 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7277 * @param[in] node The data containing the #MDB_db record for the
7278 * sorted-dup database.
7281 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7283 MDB_xcursor *mx = mc->mc_xcursor;
7285 if (node->mn_flags & F_SUBDATA) {
7286 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7287 mx->mx_cursor.mc_pg[0] = 0;
7288 mx->mx_cursor.mc_snum = 0;
7289 mx->mx_cursor.mc_top = 0;
7290 mx->mx_cursor.mc_flags = C_SUB;
7292 MDB_page *fp = NODEDATA(node);
7293 mx->mx_db.md_pad = 0;
7294 mx->mx_db.md_flags = 0;
7295 mx->mx_db.md_depth = 1;
7296 mx->mx_db.md_branch_pages = 0;
7297 mx->mx_db.md_leaf_pages = 1;
7298 mx->mx_db.md_overflow_pages = 0;
7299 mx->mx_db.md_entries = NUMKEYS(fp);
7300 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7301 mx->mx_cursor.mc_snum = 1;
7302 mx->mx_cursor.mc_top = 0;
7303 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7304 mx->mx_cursor.mc_pg[0] = fp;
7305 mx->mx_cursor.mc_ki[0] = 0;
7306 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7307 mx->mx_db.md_flags = MDB_DUPFIXED;
7308 mx->mx_db.md_pad = fp->mp_pad;
7309 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7310 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7313 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7314 mx->mx_db.md_root));
7315 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7316 #if UINT_MAX < SIZE_MAX
7317 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7318 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7322 /** Initialize a cursor for a given transaction and database. */
7324 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7327 mc->mc_backup = NULL;
7330 mc->mc_db = &txn->mt_dbs[dbi];
7331 mc->mc_dbx = &txn->mt_dbxs[dbi];
7332 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7338 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7339 mdb_tassert(txn, mx != NULL);
7340 mc->mc_xcursor = mx;
7341 mdb_xcursor_init0(mc);
7343 mc->mc_xcursor = NULL;
7345 if (*mc->mc_dbflag & DB_STALE) {
7346 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7351 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7354 size_t size = sizeof(MDB_cursor);
7356 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7359 if (txn->mt_flags & MDB_TXN_BLOCKED)
7362 /* Allow read access to the freelist */
7363 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7366 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7367 size += sizeof(MDB_xcursor);
7369 if ((mc = malloc(size)) != NULL) {
7370 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7371 if (txn->mt_cursors) {
7372 mc->mc_next = txn->mt_cursors[dbi];
7373 txn->mt_cursors[dbi] = mc;
7374 mc->mc_flags |= C_UNTRACK;
7386 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7388 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7391 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7394 if (txn->mt_flags & MDB_TXN_BLOCKED)
7397 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7401 /* Return the count of duplicate data items for the current key */
7403 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7407 if (mc == NULL || countp == NULL)
7410 if (mc->mc_xcursor == NULL)
7411 return MDB_INCOMPATIBLE;
7413 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7416 if (!(mc->mc_flags & C_INITIALIZED))
7419 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7420 return MDB_NOTFOUND;
7422 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7423 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7426 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7429 *countp = mc->mc_xcursor->mx_db.md_entries;
7435 mdb_cursor_close(MDB_cursor *mc)
7437 if (mc && !mc->mc_backup) {
7438 /* remove from txn, if tracked */
7439 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7440 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7441 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7443 *prev = mc->mc_next;
7450 mdb_cursor_txn(MDB_cursor *mc)
7452 if (!mc) return NULL;
7457 mdb_cursor_dbi(MDB_cursor *mc)
7462 /** Replace the key for a branch node with a new key.
7463 * @param[in] mc Cursor pointing to the node to operate on.
7464 * @param[in] key The new key to use.
7465 * @return 0 on success, non-zero on failure.
7468 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7474 int delta, ksize, oksize;
7475 indx_t ptr, i, numkeys, indx;
7478 indx = mc->mc_ki[mc->mc_top];
7479 mp = mc->mc_pg[mc->mc_top];
7480 node = NODEPTR(mp, indx);
7481 ptr = mp->mp_ptrs[indx];
7485 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7486 k2.mv_data = NODEKEY(node);
7487 k2.mv_size = node->mn_ksize;
7488 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7490 mdb_dkey(&k2, kbuf2),
7496 /* Sizes must be 2-byte aligned. */
7497 ksize = EVEN(key->mv_size);
7498 oksize = EVEN(node->mn_ksize);
7499 delta = ksize - oksize;
7501 /* Shift node contents if EVEN(key length) changed. */
7503 if (delta > 0 && SIZELEFT(mp) < delta) {
7505 /* not enough space left, do a delete and split */
7506 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7507 pgno = NODEPGNO(node);
7508 mdb_node_del(mc, 0);
7509 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7512 numkeys = NUMKEYS(mp);
7513 for (i = 0; i < numkeys; i++) {
7514 if (mp->mp_ptrs[i] <= ptr)
7515 mp->mp_ptrs[i] -= delta;
7518 base = (char *)mp + mp->mp_upper + PAGEBASE;
7519 len = ptr - mp->mp_upper + NODESIZE;
7520 memmove(base - delta, base, len);
7521 mp->mp_upper -= delta;
7523 node = NODEPTR(mp, indx);
7526 /* But even if no shift was needed, update ksize */
7527 if (node->mn_ksize != key->mv_size)
7528 node->mn_ksize = key->mv_size;
7531 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7537 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7539 /** Move a node from csrc to cdst.
7542 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7549 unsigned short flags;
7553 /* Mark src and dst as dirty. */
7554 if ((rc = mdb_page_touch(csrc)) ||
7555 (rc = mdb_page_touch(cdst)))
7558 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7559 key.mv_size = csrc->mc_db->md_pad;
7560 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7562 data.mv_data = NULL;
7566 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7567 mdb_cassert(csrc, !((size_t)srcnode & 1));
7568 srcpg = NODEPGNO(srcnode);
7569 flags = srcnode->mn_flags;
7570 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7571 unsigned int snum = csrc->mc_snum;
7573 /* must find the lowest key below src */
7574 rc = mdb_page_search_lowest(csrc);
7577 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7578 key.mv_size = csrc->mc_db->md_pad;
7579 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7581 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7582 key.mv_size = NODEKSZ(s2);
7583 key.mv_data = NODEKEY(s2);
7585 csrc->mc_snum = snum--;
7586 csrc->mc_top = snum;
7588 key.mv_size = NODEKSZ(srcnode);
7589 key.mv_data = NODEKEY(srcnode);
7591 data.mv_size = NODEDSZ(srcnode);
7592 data.mv_data = NODEDATA(srcnode);
7594 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7595 unsigned int snum = cdst->mc_snum;
7598 /* must find the lowest key below dst */
7599 mdb_cursor_copy(cdst, &mn);
7600 rc = mdb_page_search_lowest(&mn);
7603 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7604 bkey.mv_size = mn.mc_db->md_pad;
7605 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7607 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7608 bkey.mv_size = NODEKSZ(s2);
7609 bkey.mv_data = NODEKEY(s2);
7611 mn.mc_snum = snum--;
7614 rc = mdb_update_key(&mn, &bkey);
7619 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7620 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7621 csrc->mc_ki[csrc->mc_top],
7623 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7624 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7626 /* Add the node to the destination page.
7628 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7629 if (rc != MDB_SUCCESS)
7632 /* Delete the node from the source page.
7634 mdb_node_del(csrc, key.mv_size);
7637 /* Adjust other cursors pointing to mp */
7638 MDB_cursor *m2, *m3;
7639 MDB_dbi dbi = csrc->mc_dbi;
7640 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7642 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7643 if (csrc->mc_flags & C_SUB)
7644 m3 = &m2->mc_xcursor->mx_cursor;
7647 if (m3 == csrc) continue;
7648 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7649 csrc->mc_ki[csrc->mc_top]) {
7650 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7651 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7656 /* Update the parent separators.
7658 if (csrc->mc_ki[csrc->mc_top] == 0) {
7659 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7660 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7661 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7663 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7664 key.mv_size = NODEKSZ(srcnode);
7665 key.mv_data = NODEKEY(srcnode);
7667 DPRINTF(("update separator for source page %"Z"u to [%s]",
7668 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7669 mdb_cursor_copy(csrc, &mn);
7672 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7675 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7677 indx_t ix = csrc->mc_ki[csrc->mc_top];
7678 nullkey.mv_size = 0;
7679 csrc->mc_ki[csrc->mc_top] = 0;
7680 rc = mdb_update_key(csrc, &nullkey);
7681 csrc->mc_ki[csrc->mc_top] = ix;
7682 mdb_cassert(csrc, rc == MDB_SUCCESS);
7686 if (cdst->mc_ki[cdst->mc_top] == 0) {
7687 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7688 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7689 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7691 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7692 key.mv_size = NODEKSZ(srcnode);
7693 key.mv_data = NODEKEY(srcnode);
7695 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7696 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7697 mdb_cursor_copy(cdst, &mn);
7700 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7703 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7705 indx_t ix = cdst->mc_ki[cdst->mc_top];
7706 nullkey.mv_size = 0;
7707 cdst->mc_ki[cdst->mc_top] = 0;
7708 rc = mdb_update_key(cdst, &nullkey);
7709 cdst->mc_ki[cdst->mc_top] = ix;
7710 mdb_cassert(cdst, rc == MDB_SUCCESS);
7717 /** Merge one page into another.
7718 * The nodes from the page pointed to by \b csrc will
7719 * be copied to the page pointed to by \b cdst and then
7720 * the \b csrc page will be freed.
7721 * @param[in] csrc Cursor pointing to the source page.
7722 * @param[in] cdst Cursor pointing to the destination page.
7723 * @return 0 on success, non-zero on failure.
7726 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7728 MDB_page *psrc, *pdst;
7735 psrc = csrc->mc_pg[csrc->mc_top];
7736 pdst = cdst->mc_pg[cdst->mc_top];
7738 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7740 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7741 mdb_cassert(csrc, cdst->mc_snum > 1);
7743 /* Mark dst as dirty. */
7744 if ((rc = mdb_page_touch(cdst)))
7747 /* Move all nodes from src to dst.
7749 j = nkeys = NUMKEYS(pdst);
7750 if (IS_LEAF2(psrc)) {
7751 key.mv_size = csrc->mc_db->md_pad;
7752 key.mv_data = METADATA(psrc);
7753 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7754 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7755 if (rc != MDB_SUCCESS)
7757 key.mv_data = (char *)key.mv_data + key.mv_size;
7760 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7761 srcnode = NODEPTR(psrc, i);
7762 if (i == 0 && IS_BRANCH(psrc)) {
7765 mdb_cursor_copy(csrc, &mn);
7766 /* must find the lowest key below src */
7767 rc = mdb_page_search_lowest(&mn);
7770 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7771 key.mv_size = mn.mc_db->md_pad;
7772 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7774 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7775 key.mv_size = NODEKSZ(s2);
7776 key.mv_data = NODEKEY(s2);
7779 key.mv_size = srcnode->mn_ksize;
7780 key.mv_data = NODEKEY(srcnode);
7783 data.mv_size = NODEDSZ(srcnode);
7784 data.mv_data = NODEDATA(srcnode);
7785 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7786 if (rc != MDB_SUCCESS)
7791 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7792 pdst->mp_pgno, NUMKEYS(pdst),
7793 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7795 /* Unlink the src page from parent and add to free list.
7798 mdb_node_del(csrc, 0);
7799 if (csrc->mc_ki[csrc->mc_top] == 0) {
7801 rc = mdb_update_key(csrc, &key);
7809 psrc = csrc->mc_pg[csrc->mc_top];
7810 /* If not operating on FreeDB, allow this page to be reused
7811 * in this txn. Otherwise just add to free list.
7813 rc = mdb_page_loose(csrc, psrc);
7817 csrc->mc_db->md_leaf_pages--;
7819 csrc->mc_db->md_branch_pages--;
7821 /* Adjust other cursors pointing to mp */
7822 MDB_cursor *m2, *m3;
7823 MDB_dbi dbi = csrc->mc_dbi;
7825 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7826 if (csrc->mc_flags & C_SUB)
7827 m3 = &m2->mc_xcursor->mx_cursor;
7830 if (m3 == csrc) continue;
7831 if (m3->mc_snum < csrc->mc_snum) continue;
7832 if (m3->mc_pg[csrc->mc_top] == psrc) {
7833 m3->mc_pg[csrc->mc_top] = pdst;
7834 m3->mc_ki[csrc->mc_top] += nkeys;
7839 unsigned int snum = cdst->mc_snum;
7840 uint16_t depth = cdst->mc_db->md_depth;
7841 mdb_cursor_pop(cdst);
7842 rc = mdb_rebalance(cdst);
7843 /* Did the tree shrink? */
7844 if (depth > cdst->mc_db->md_depth)
7846 cdst->mc_snum = snum;
7847 cdst->mc_top = snum-1;
7852 /** Copy the contents of a cursor.
7853 * @param[in] csrc The cursor to copy from.
7854 * @param[out] cdst The cursor to copy to.
7857 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7861 cdst->mc_txn = csrc->mc_txn;
7862 cdst->mc_dbi = csrc->mc_dbi;
7863 cdst->mc_db = csrc->mc_db;
7864 cdst->mc_dbx = csrc->mc_dbx;
7865 cdst->mc_snum = csrc->mc_snum;
7866 cdst->mc_top = csrc->mc_top;
7867 cdst->mc_flags = csrc->mc_flags;
7869 for (i=0; i<csrc->mc_snum; i++) {
7870 cdst->mc_pg[i] = csrc->mc_pg[i];
7871 cdst->mc_ki[i] = csrc->mc_ki[i];
7875 /** Rebalance the tree after a delete operation.
7876 * @param[in] mc Cursor pointing to the page where rebalancing
7878 * @return 0 on success, non-zero on failure.
7881 mdb_rebalance(MDB_cursor *mc)
7885 unsigned int ptop, minkeys;
7889 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7890 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7891 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7892 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7893 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7895 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7896 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7897 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7898 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7902 if (mc->mc_snum < 2) {
7903 MDB_page *mp = mc->mc_pg[0];
7905 DPUTS("Can't rebalance a subpage, ignoring");
7908 if (NUMKEYS(mp) == 0) {
7909 DPUTS("tree is completely empty");
7910 mc->mc_db->md_root = P_INVALID;
7911 mc->mc_db->md_depth = 0;
7912 mc->mc_db->md_leaf_pages = 0;
7913 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7916 /* Adjust cursors pointing to mp */
7919 mc->mc_flags &= ~C_INITIALIZED;
7921 MDB_cursor *m2, *m3;
7922 MDB_dbi dbi = mc->mc_dbi;
7924 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7925 if (mc->mc_flags & C_SUB)
7926 m3 = &m2->mc_xcursor->mx_cursor;
7929 if (m3->mc_snum < mc->mc_snum) continue;
7930 if (m3->mc_pg[0] == mp) {
7933 m3->mc_flags &= ~C_INITIALIZED;
7937 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7939 DPUTS("collapsing root page!");
7940 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7943 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7944 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7947 mc->mc_db->md_depth--;
7948 mc->mc_db->md_branch_pages--;
7949 mc->mc_ki[0] = mc->mc_ki[1];
7950 for (i = 1; i<mc->mc_db->md_depth; i++) {
7951 mc->mc_pg[i] = mc->mc_pg[i+1];
7952 mc->mc_ki[i] = mc->mc_ki[i+1];
7955 /* Adjust other cursors pointing to mp */
7956 MDB_cursor *m2, *m3;
7957 MDB_dbi dbi = mc->mc_dbi;
7959 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7960 if (mc->mc_flags & C_SUB)
7961 m3 = &m2->mc_xcursor->mx_cursor;
7964 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7965 if (m3->mc_pg[0] == mp) {
7966 for (i=0; i<m3->mc_snum; i++) {
7967 m3->mc_pg[i] = m3->mc_pg[i+1];
7968 m3->mc_ki[i] = m3->mc_ki[i+1];
7976 DPUTS("root page doesn't need rebalancing");
7980 /* The parent (branch page) must have at least 2 pointers,
7981 * otherwise the tree is invalid.
7983 ptop = mc->mc_top-1;
7984 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7986 /* Leaf page fill factor is below the threshold.
7987 * Try to move keys from left or right neighbor, or
7988 * merge with a neighbor page.
7993 mdb_cursor_copy(mc, &mn);
7994 mn.mc_xcursor = NULL;
7996 oldki = mc->mc_ki[mc->mc_top];
7997 if (mc->mc_ki[ptop] == 0) {
7998 /* We're the leftmost leaf in our parent.
8000 DPUTS("reading right neighbor");
8002 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8003 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8006 mn.mc_ki[mn.mc_top] = 0;
8007 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8009 /* There is at least one neighbor to the left.
8011 DPUTS("reading left neighbor");
8013 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8014 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8017 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8018 mc->mc_ki[mc->mc_top] = 0;
8021 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8022 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8023 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8025 /* If the neighbor page is above threshold and has enough keys,
8026 * move one key from it. Otherwise we should try to merge them.
8027 * (A branch page must never have less than 2 keys.)
8029 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
8030 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8031 rc = mdb_node_move(&mn, mc);
8032 if (mc->mc_ki[ptop]) {
8036 if (mc->mc_ki[ptop] == 0) {
8037 rc = mdb_page_merge(&mn, mc);
8040 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8041 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8042 /* We want mdb_rebalance to find mn when doing fixups */
8043 if (mc->mc_flags & C_SUB) {
8044 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8045 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8046 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8048 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8049 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8051 rc = mdb_page_merge(mc, &mn);
8052 if (mc->mc_flags & C_SUB)
8053 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8055 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8056 mdb_cursor_copy(&mn, mc);
8058 mc->mc_flags &= ~C_EOF;
8060 mc->mc_ki[mc->mc_top] = oldki;
8064 /** Complete a delete operation started by #mdb_cursor_del(). */
8066 mdb_cursor_del0(MDB_cursor *mc)
8073 ki = mc->mc_ki[mc->mc_top];
8074 mdb_node_del(mc, mc->mc_db->md_pad);
8075 mc->mc_db->md_entries--;
8076 rc = mdb_rebalance(mc);
8078 if (rc == MDB_SUCCESS) {
8079 MDB_cursor *m2, *m3;
8080 MDB_dbi dbi = mc->mc_dbi;
8082 /* DB is totally empty now, just bail out.
8083 * Other cursors adjustments were already done
8084 * by mdb_rebalance and aren't needed here.
8089 mp = mc->mc_pg[mc->mc_top];
8090 nkeys = NUMKEYS(mp);
8092 /* if mc points past last node in page, find next sibling */
8093 if (mc->mc_ki[mc->mc_top] >= nkeys) {
8094 rc = mdb_cursor_sibling(mc, 1);
8095 if (rc == MDB_NOTFOUND) {
8096 mc->mc_flags |= C_EOF;
8101 /* Adjust other cursors pointing to mp */
8102 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8103 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8104 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8106 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8108 if (m3->mc_pg[mc->mc_top] == mp) {
8109 if (m3->mc_ki[mc->mc_top] >= ki) {
8110 m3->mc_flags |= C_DEL;
8111 if (m3->mc_ki[mc->mc_top] > ki)
8112 m3->mc_ki[mc->mc_top]--;
8113 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8114 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8116 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8117 rc = mdb_cursor_sibling(m3, 1);
8118 if (rc == MDB_NOTFOUND) {
8119 m3->mc_flags |= C_EOF;
8125 mc->mc_flags |= C_DEL;
8129 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8134 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8135 MDB_val *key, MDB_val *data)
8137 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8140 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8141 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8143 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8144 /* must ignore any data */
8148 return mdb_del0(txn, dbi, key, data, 0);
8152 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8153 MDB_val *key, MDB_val *data, unsigned flags)
8158 MDB_val rdata, *xdata;
8162 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8164 mdb_cursor_init(&mc, txn, dbi, &mx);
8173 flags |= MDB_NODUPDATA;
8175 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8177 /* let mdb_page_split know about this cursor if needed:
8178 * delete will trigger a rebalance; if it needs to move
8179 * a node from one page to another, it will have to
8180 * update the parent's separator key(s). If the new sepkey
8181 * is larger than the current one, the parent page may
8182 * run out of space, triggering a split. We need this
8183 * cursor to be consistent until the end of the rebalance.
8185 mc.mc_flags |= C_UNTRACK;
8186 mc.mc_next = txn->mt_cursors[dbi];
8187 txn->mt_cursors[dbi] = &mc;
8188 rc = mdb_cursor_del(&mc, flags);
8189 txn->mt_cursors[dbi] = mc.mc_next;
8194 /** Split a page and insert a new node.
8195 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8196 * The cursor will be updated to point to the actual page and index where
8197 * the node got inserted after the split.
8198 * @param[in] newkey The key for the newly inserted node.
8199 * @param[in] newdata The data for the newly inserted node.
8200 * @param[in] newpgno The page number, if the new node is a branch node.
8201 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8202 * @return 0 on success, non-zero on failure.
8205 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8206 unsigned int nflags)
8209 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8212 int i, j, split_indx, nkeys, pmax;
8213 MDB_env *env = mc->mc_txn->mt_env;
8215 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8216 MDB_page *copy = NULL;
8217 MDB_page *mp, *rp, *pp;
8222 mp = mc->mc_pg[mc->mc_top];
8223 newindx = mc->mc_ki[mc->mc_top];
8224 nkeys = NUMKEYS(mp);
8226 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8227 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8228 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8230 /* Create a right sibling. */
8231 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8233 rp->mp_pad = mp->mp_pad;
8234 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8236 if (mc->mc_snum < 2) {
8237 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8239 /* shift current top to make room for new parent */
8240 mc->mc_pg[1] = mc->mc_pg[0];
8241 mc->mc_ki[1] = mc->mc_ki[0];
8244 mc->mc_db->md_root = pp->mp_pgno;
8245 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8246 mc->mc_db->md_depth++;
8249 /* Add left (implicit) pointer. */
8250 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8251 /* undo the pre-push */
8252 mc->mc_pg[0] = mc->mc_pg[1];
8253 mc->mc_ki[0] = mc->mc_ki[1];
8254 mc->mc_db->md_root = mp->mp_pgno;
8255 mc->mc_db->md_depth--;
8262 ptop = mc->mc_top-1;
8263 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8266 mc->mc_flags |= C_SPLITTING;
8267 mdb_cursor_copy(mc, &mn);
8268 mn.mc_pg[mn.mc_top] = rp;
8269 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8271 if (nflags & MDB_APPEND) {
8272 mn.mc_ki[mn.mc_top] = 0;
8274 split_indx = newindx;
8278 split_indx = (nkeys+1) / 2;
8283 unsigned int lsize, rsize, ksize;
8284 /* Move half of the keys to the right sibling */
8285 x = mc->mc_ki[mc->mc_top] - split_indx;
8286 ksize = mc->mc_db->md_pad;
8287 split = LEAF2KEY(mp, split_indx, ksize);
8288 rsize = (nkeys - split_indx) * ksize;
8289 lsize = (nkeys - split_indx) * sizeof(indx_t);
8290 mp->mp_lower -= lsize;
8291 rp->mp_lower += lsize;
8292 mp->mp_upper += rsize - lsize;
8293 rp->mp_upper -= rsize - lsize;
8294 sepkey.mv_size = ksize;
8295 if (newindx == split_indx) {
8296 sepkey.mv_data = newkey->mv_data;
8298 sepkey.mv_data = split;
8301 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8302 memcpy(rp->mp_ptrs, split, rsize);
8303 sepkey.mv_data = rp->mp_ptrs;
8304 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8305 memcpy(ins, newkey->mv_data, ksize);
8306 mp->mp_lower += sizeof(indx_t);
8307 mp->mp_upper -= ksize - sizeof(indx_t);
8310 memcpy(rp->mp_ptrs, split, x * ksize);
8311 ins = LEAF2KEY(rp, x, ksize);
8312 memcpy(ins, newkey->mv_data, ksize);
8313 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8314 rp->mp_lower += sizeof(indx_t);
8315 rp->mp_upper -= ksize - sizeof(indx_t);
8316 mc->mc_ki[mc->mc_top] = x;
8317 mc->mc_pg[mc->mc_top] = rp;
8320 int psize, nsize, k;
8321 /* Maximum free space in an empty page */
8322 pmax = env->me_psize - PAGEHDRSZ;
8324 nsize = mdb_leaf_size(env, newkey, newdata);
8326 nsize = mdb_branch_size(env, newkey);
8327 nsize = EVEN(nsize);
8329 /* grab a page to hold a temporary copy */
8330 copy = mdb_page_malloc(mc->mc_txn, 1);
8335 copy->mp_pgno = mp->mp_pgno;
8336 copy->mp_flags = mp->mp_flags;
8337 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8338 copy->mp_upper = env->me_psize - PAGEBASE;
8340 /* prepare to insert */
8341 for (i=0, j=0; i<nkeys; i++) {
8343 copy->mp_ptrs[j++] = 0;
8345 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8348 /* When items are relatively large the split point needs
8349 * to be checked, because being off-by-one will make the
8350 * difference between success or failure in mdb_node_add.
8352 * It's also relevant if a page happens to be laid out
8353 * such that one half of its nodes are all "small" and
8354 * the other half of its nodes are "large." If the new
8355 * item is also "large" and falls on the half with
8356 * "large" nodes, it also may not fit.
8358 * As a final tweak, if the new item goes on the last
8359 * spot on the page (and thus, onto the new page), bias
8360 * the split so the new page is emptier than the old page.
8361 * This yields better packing during sequential inserts.
8363 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8364 /* Find split point */
8366 if (newindx <= split_indx || newindx >= nkeys) {
8368 k = newindx >= nkeys ? nkeys : split_indx+2;
8373 for (; i!=k; i+=j) {
8378 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8379 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8381 if (F_ISSET(node->mn_flags, F_BIGDATA))
8382 psize += sizeof(pgno_t);
8384 psize += NODEDSZ(node);
8386 psize = EVEN(psize);
8388 if (psize > pmax || i == k-j) {
8389 split_indx = i + (j<0);
8394 if (split_indx == newindx) {
8395 sepkey.mv_size = newkey->mv_size;
8396 sepkey.mv_data = newkey->mv_data;
8398 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8399 sepkey.mv_size = node->mn_ksize;
8400 sepkey.mv_data = NODEKEY(node);
8405 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8407 /* Copy separator key to the parent.
8409 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8413 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8418 if (mn.mc_snum == mc->mc_snum) {
8419 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8420 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8421 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8422 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8427 /* Right page might now have changed parent.
8428 * Check if left page also changed parent.
8430 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8431 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8432 for (i=0; i<ptop; i++) {
8433 mc->mc_pg[i] = mn.mc_pg[i];
8434 mc->mc_ki[i] = mn.mc_ki[i];
8436 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8437 if (mn.mc_ki[ptop]) {
8438 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8440 /* find right page's left sibling */
8441 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8442 mdb_cursor_sibling(mc, 0);
8447 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8450 mc->mc_flags ^= C_SPLITTING;
8451 if (rc != MDB_SUCCESS) {
8454 if (nflags & MDB_APPEND) {
8455 mc->mc_pg[mc->mc_top] = rp;
8456 mc->mc_ki[mc->mc_top] = 0;
8457 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8460 for (i=0; i<mc->mc_top; i++)
8461 mc->mc_ki[i] = mn.mc_ki[i];
8462 } else if (!IS_LEAF2(mp)) {
8464 mc->mc_pg[mc->mc_top] = rp;
8469 rkey.mv_data = newkey->mv_data;
8470 rkey.mv_size = newkey->mv_size;
8476 /* Update index for the new key. */
8477 mc->mc_ki[mc->mc_top] = j;
8479 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8480 rkey.mv_data = NODEKEY(node);
8481 rkey.mv_size = node->mn_ksize;
8483 xdata.mv_data = NODEDATA(node);
8484 xdata.mv_size = NODEDSZ(node);
8487 pgno = NODEPGNO(node);
8488 flags = node->mn_flags;
8491 if (!IS_LEAF(mp) && j == 0) {
8492 /* First branch index doesn't need key data. */
8496 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8502 mc->mc_pg[mc->mc_top] = copy;
8507 } while (i != split_indx);
8509 nkeys = NUMKEYS(copy);
8510 for (i=0; i<nkeys; i++)
8511 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8512 mp->mp_lower = copy->mp_lower;
8513 mp->mp_upper = copy->mp_upper;
8514 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8515 env->me_psize - copy->mp_upper - PAGEBASE);
8517 /* reset back to original page */
8518 if (newindx < split_indx) {
8519 mc->mc_pg[mc->mc_top] = mp;
8520 if (nflags & MDB_RESERVE) {
8521 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8522 if (!(node->mn_flags & F_BIGDATA))
8523 newdata->mv_data = NODEDATA(node);
8526 mc->mc_pg[mc->mc_top] = rp;
8528 /* Make sure mc_ki is still valid.
8530 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8531 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8532 for (i=0; i<=ptop; i++) {
8533 mc->mc_pg[i] = mn.mc_pg[i];
8534 mc->mc_ki[i] = mn.mc_ki[i];
8541 /* Adjust other cursors pointing to mp */
8542 MDB_cursor *m2, *m3;
8543 MDB_dbi dbi = mc->mc_dbi;
8544 int fixup = NUMKEYS(mp);
8546 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8547 if (mc->mc_flags & C_SUB)
8548 m3 = &m2->mc_xcursor->mx_cursor;
8553 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8555 if (m3->mc_flags & C_SPLITTING)
8560 for (k=m3->mc_top; k>=0; k--) {
8561 m3->mc_ki[k+1] = m3->mc_ki[k];
8562 m3->mc_pg[k+1] = m3->mc_pg[k];
8564 if (m3->mc_ki[0] >= split_indx) {
8569 m3->mc_pg[0] = mc->mc_pg[0];
8573 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8574 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8575 m3->mc_ki[mc->mc_top]++;
8576 if (m3->mc_ki[mc->mc_top] >= fixup) {
8577 m3->mc_pg[mc->mc_top] = rp;
8578 m3->mc_ki[mc->mc_top] -= fixup;
8579 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8581 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8582 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8587 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8590 if (copy) /* tmp page */
8591 mdb_page_free(env, copy);
8593 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8598 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8599 MDB_val *key, MDB_val *data, unsigned int flags)
8604 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8607 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8610 mdb_cursor_init(&mc, txn, dbi, &mx);
8611 return mdb_cursor_put(&mc, key, data, flags);
8615 #define MDB_WBUF (1024*1024)
8618 /** State needed for a compacting copy. */
8619 typedef struct mdb_copy {
8620 pthread_mutex_t mc_mutex;
8621 pthread_cond_t mc_cond;
8628 pgno_t mc_next_pgno;
8631 volatile int mc_new;
8636 /** Dedicated writer thread for compacting copy. */
8637 static THREAD_RET ESECT
8638 mdb_env_copythr(void *arg)
8642 int toggle = 0, wsize, rc;
8645 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8648 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8651 pthread_mutex_lock(&my->mc_mutex);
8653 pthread_cond_signal(&my->mc_cond);
8656 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8657 if (my->mc_new < 0) {
8662 wsize = my->mc_wlen[toggle];
8663 ptr = my->mc_wbuf[toggle];
8666 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8670 } else if (len > 0) {
8684 /* If there's an overflow page tail, write it too */
8685 if (my->mc_olen[toggle]) {
8686 wsize = my->mc_olen[toggle];
8687 ptr = my->mc_over[toggle];
8688 my->mc_olen[toggle] = 0;
8691 my->mc_wlen[toggle] = 0;
8693 pthread_cond_signal(&my->mc_cond);
8695 pthread_cond_signal(&my->mc_cond);
8696 pthread_mutex_unlock(&my->mc_mutex);
8697 return (THREAD_RET)0;
8701 /** Tell the writer thread there's a buffer ready to write */
8703 mdb_env_cthr_toggle(mdb_copy *my, int st)
8705 int toggle = my->mc_toggle ^ 1;
8706 pthread_mutex_lock(&my->mc_mutex);
8707 if (my->mc_status) {
8708 pthread_mutex_unlock(&my->mc_mutex);
8709 return my->mc_status;
8711 while (my->mc_new == 1)
8712 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8714 my->mc_toggle = toggle;
8715 pthread_cond_signal(&my->mc_cond);
8716 pthread_mutex_unlock(&my->mc_mutex);
8720 /** Depth-first tree traversal for compacting copy. */
8722 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8725 MDB_txn *txn = my->mc_txn;
8727 MDB_page *mo, *mp, *leaf;
8732 /* Empty DB, nothing to do */
8733 if (*pg == P_INVALID)
8740 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8743 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8747 /* Make cursor pages writable */
8748 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8752 for (i=0; i<mc.mc_top; i++) {
8753 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8754 mc.mc_pg[i] = (MDB_page *)ptr;
8755 ptr += my->mc_env->me_psize;
8758 /* This is writable space for a leaf page. Usually not needed. */
8759 leaf = (MDB_page *)ptr;
8761 toggle = my->mc_toggle;
8762 while (mc.mc_snum > 0) {
8764 mp = mc.mc_pg[mc.mc_top];
8768 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8769 for (i=0; i<n; i++) {
8770 ni = NODEPTR(mp, i);
8771 if (ni->mn_flags & F_BIGDATA) {
8775 /* Need writable leaf */
8777 mc.mc_pg[mc.mc_top] = leaf;
8778 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8780 ni = NODEPTR(mp, i);
8783 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8784 rc = mdb_page_get(txn, pg, &omp, NULL);
8787 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8788 rc = mdb_env_cthr_toggle(my, 1);
8791 toggle = my->mc_toggle;
8793 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8794 memcpy(mo, omp, my->mc_env->me_psize);
8795 mo->mp_pgno = my->mc_next_pgno;
8796 my->mc_next_pgno += omp->mp_pages;
8797 my->mc_wlen[toggle] += my->mc_env->me_psize;
8798 if (omp->mp_pages > 1) {
8799 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8800 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8801 rc = mdb_env_cthr_toggle(my, 1);
8804 toggle = my->mc_toggle;
8806 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8807 } else if (ni->mn_flags & F_SUBDATA) {
8810 /* Need writable leaf */
8812 mc.mc_pg[mc.mc_top] = leaf;
8813 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8815 ni = NODEPTR(mp, i);
8818 memcpy(&db, NODEDATA(ni), sizeof(db));
8819 my->mc_toggle = toggle;
8820 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8823 toggle = my->mc_toggle;
8824 memcpy(NODEDATA(ni), &db, sizeof(db));
8829 mc.mc_ki[mc.mc_top]++;
8830 if (mc.mc_ki[mc.mc_top] < n) {
8833 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8835 rc = mdb_page_get(txn, pg, &mp, NULL);
8840 mc.mc_ki[mc.mc_top] = 0;
8841 if (IS_BRANCH(mp)) {
8842 /* Whenever we advance to a sibling branch page,
8843 * we must proceed all the way down to its first leaf.
8845 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8848 mc.mc_pg[mc.mc_top] = mp;
8852 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8853 rc = mdb_env_cthr_toggle(my, 1);
8856 toggle = my->mc_toggle;
8858 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8859 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8860 mo->mp_pgno = my->mc_next_pgno++;
8861 my->mc_wlen[toggle] += my->mc_env->me_psize;
8863 /* Update parent if there is one */
8864 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8865 SETPGNO(ni, mo->mp_pgno);
8866 mdb_cursor_pop(&mc);
8868 /* Otherwise we're done */
8878 /** Copy environment with compaction. */
8880 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8885 MDB_txn *txn = NULL;
8890 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8891 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8892 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8893 if (my.mc_wbuf[0] == NULL)
8896 pthread_mutex_init(&my.mc_mutex, NULL);
8897 pthread_cond_init(&my.mc_cond, NULL);
8898 #ifdef HAVE_MEMALIGN
8899 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8900 if (my.mc_wbuf[0] == NULL)
8903 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8908 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8909 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8914 my.mc_next_pgno = 2;
8920 THREAD_CREATE(thr, mdb_env_copythr, &my);
8922 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8926 mp = (MDB_page *)my.mc_wbuf[0];
8927 memset(mp, 0, 2*env->me_psize);
8929 mp->mp_flags = P_META;
8930 mm = (MDB_meta *)METADATA(mp);
8931 mdb_env_init_meta0(env, mm);
8932 mm->mm_address = env->me_metas[0]->mm_address;
8934 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8936 mp->mp_flags = P_META;
8937 *(MDB_meta *)METADATA(mp) = *mm;
8938 mm = (MDB_meta *)METADATA(mp);
8940 /* Count the number of free pages, subtract from lastpg to find
8941 * number of active pages
8944 MDB_ID freecount = 0;
8947 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8948 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8949 freecount += *(MDB_ID *)data.mv_data;
8950 freecount += txn->mt_dbs[0].md_branch_pages +
8951 txn->mt_dbs[0].md_leaf_pages +
8952 txn->mt_dbs[0].md_overflow_pages;
8954 /* Set metapage 1 */
8955 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8956 mm->mm_dbs[1] = txn->mt_dbs[1];
8957 if (mm->mm_last_pg > 1) {
8958 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8961 mm->mm_dbs[1].md_root = P_INVALID;
8964 my.mc_wlen[0] = env->me_psize * 2;
8966 pthread_mutex_lock(&my.mc_mutex);
8968 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8969 pthread_mutex_unlock(&my.mc_mutex);
8970 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8971 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8972 rc = mdb_env_cthr_toggle(&my, 1);
8973 mdb_env_cthr_toggle(&my, -1);
8974 pthread_mutex_lock(&my.mc_mutex);
8976 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8977 pthread_mutex_unlock(&my.mc_mutex);
8982 CloseHandle(my.mc_cond);
8983 CloseHandle(my.mc_mutex);
8984 _aligned_free(my.mc_wbuf[0]);
8986 pthread_cond_destroy(&my.mc_cond);
8987 pthread_mutex_destroy(&my.mc_mutex);
8988 free(my.mc_wbuf[0]);
8993 /** Copy environment as-is. */
8995 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8997 MDB_txn *txn = NULL;
8998 mdb_mutexref_t wmutex = NULL;
9004 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9008 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9011 /* Do the lock/unlock of the reader mutex before starting the
9012 * write txn. Otherwise other read txns could block writers.
9014 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9019 /* We must start the actual read txn after blocking writers */
9020 mdb_txn_end(txn, MDB_END_RESET_TMP);
9022 /* Temporarily block writers until we snapshot the meta pages */
9023 wmutex = env->me_wmutex;
9024 if (LOCK_MUTEX(rc, env, wmutex))
9027 rc = mdb_txn_renew0(txn);
9029 UNLOCK_MUTEX(wmutex);
9034 wsize = env->me_psize * 2;
9038 DO_WRITE(rc, fd, ptr, w2, len);
9042 } else if (len > 0) {
9048 /* Non-blocking or async handles are not supported */
9054 UNLOCK_MUTEX(wmutex);
9059 w2 = txn->mt_next_pgno * env->me_psize;
9062 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9069 if (wsize > MAX_WRITE)
9073 DO_WRITE(rc, fd, ptr, w2, len);
9077 } else if (len > 0) {
9094 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9096 if (flags & MDB_CP_COMPACT)
9097 return mdb_env_copyfd1(env, fd);
9099 return mdb_env_copyfd0(env, fd);
9103 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9105 return mdb_env_copyfd2(env, fd, 0);
9109 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9113 HANDLE newfd = INVALID_HANDLE_VALUE;
9115 if (env->me_flags & MDB_NOSUBDIR) {
9116 lpath = (char *)path;
9119 len += sizeof(DATANAME);
9120 lpath = malloc(len);
9123 sprintf(lpath, "%s" DATANAME, path);
9126 /* The destination path must exist, but the destination file must not.
9127 * We don't want the OS to cache the writes, since the source data is
9128 * already in the OS cache.
9131 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9132 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9134 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9136 if (newfd == INVALID_HANDLE_VALUE) {
9141 if (env->me_psize >= env->me_os_psize) {
9143 /* Set O_DIRECT if the file system supports it */
9144 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9145 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9147 #ifdef F_NOCACHE /* __APPLE__ */
9148 rc = fcntl(newfd, F_NOCACHE, 1);
9156 rc = mdb_env_copyfd2(env, newfd, flags);
9159 if (!(env->me_flags & MDB_NOSUBDIR))
9161 if (newfd != INVALID_HANDLE_VALUE)
9162 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9169 mdb_env_copy(MDB_env *env, const char *path)
9171 return mdb_env_copy2(env, path, 0);
9175 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9177 if (flag & ~CHANGEABLE)
9180 env->me_flags |= flag;
9182 env->me_flags &= ~flag;
9187 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9192 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9197 mdb_env_set_userctx(MDB_env *env, void *ctx)
9201 env->me_userctx = ctx;
9206 mdb_env_get_userctx(MDB_env *env)
9208 return env ? env->me_userctx : NULL;
9212 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9217 env->me_assert_func = func;
9223 mdb_env_get_path(MDB_env *env, const char **arg)
9228 *arg = env->me_path;
9233 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9242 /** Common code for #mdb_stat() and #mdb_env_stat().
9243 * @param[in] env the environment to operate in.
9244 * @param[in] db the #MDB_db record containing the stats to return.
9245 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9246 * @return 0, this function always succeeds.
9249 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9251 arg->ms_psize = env->me_psize;
9252 arg->ms_depth = db->md_depth;
9253 arg->ms_branch_pages = db->md_branch_pages;
9254 arg->ms_leaf_pages = db->md_leaf_pages;
9255 arg->ms_overflow_pages = db->md_overflow_pages;
9256 arg->ms_entries = db->md_entries;
9262 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9266 if (env == NULL || arg == NULL)
9269 meta = mdb_env_pick_meta(env);
9271 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9275 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9279 if (env == NULL || arg == NULL)
9282 meta = mdb_env_pick_meta(env);
9283 arg->me_mapaddr = meta->mm_address;
9284 arg->me_last_pgno = meta->mm_last_pg;
9285 arg->me_last_txnid = meta->mm_txnid;
9287 arg->me_mapsize = env->me_mapsize;
9288 arg->me_maxreaders = env->me_maxreaders;
9289 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9293 /** Set the default comparison functions for a database.
9294 * Called immediately after a database is opened to set the defaults.
9295 * The user can then override them with #mdb_set_compare() or
9296 * #mdb_set_dupsort().
9297 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9298 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9301 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9303 uint16_t f = txn->mt_dbs[dbi].md_flags;
9305 txn->mt_dbxs[dbi].md_cmp =
9306 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9307 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9309 txn->mt_dbxs[dbi].md_dcmp =
9310 !(f & MDB_DUPSORT) ? 0 :
9311 ((f & MDB_INTEGERDUP)
9312 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9313 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9316 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9322 int rc, dbflag, exact;
9323 unsigned int unused = 0, seq;
9326 if (flags & ~VALID_FLAGS)
9328 if (txn->mt_flags & MDB_TXN_BLOCKED)
9334 if (flags & PERSISTENT_FLAGS) {
9335 uint16_t f2 = flags & PERSISTENT_FLAGS;
9336 /* make sure flag changes get committed */
9337 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9338 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9339 txn->mt_flags |= MDB_TXN_DIRTY;
9342 mdb_default_cmp(txn, MAIN_DBI);
9346 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9347 mdb_default_cmp(txn, MAIN_DBI);
9350 /* Is the DB already open? */
9352 for (i=2; i<txn->mt_numdbs; i++) {
9353 if (!txn->mt_dbxs[i].md_name.mv_size) {
9354 /* Remember this free slot */
9355 if (!unused) unused = i;
9358 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9359 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9365 /* If no free slot and max hit, fail */
9366 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9367 return MDB_DBS_FULL;
9369 /* Cannot mix named databases with some mainDB flags */
9370 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9371 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9373 /* Find the DB info */
9374 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9377 key.mv_data = (void *)name;
9378 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9379 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9380 if (rc == MDB_SUCCESS) {
9381 /* make sure this is actually a DB */
9382 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9383 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9384 return MDB_INCOMPATIBLE;
9385 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9386 /* Create if requested */
9387 data.mv_size = sizeof(MDB_db);
9388 data.mv_data = &dummy;
9389 memset(&dummy, 0, sizeof(dummy));
9390 dummy.md_root = P_INVALID;
9391 dummy.md_flags = flags & PERSISTENT_FLAGS;
9392 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9396 /* OK, got info, add to table */
9397 if (rc == MDB_SUCCESS) {
9398 unsigned int slot = unused ? unused : txn->mt_numdbs;
9399 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9400 txn->mt_dbxs[slot].md_name.mv_size = len;
9401 txn->mt_dbxs[slot].md_rel = NULL;
9402 txn->mt_dbflags[slot] = dbflag;
9403 /* txn-> and env-> are the same in read txns, use
9404 * tmp variable to avoid undefined assignment
9406 seq = ++txn->mt_env->me_dbiseqs[slot];
9407 txn->mt_dbiseqs[slot] = seq;
9409 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9411 mdb_default_cmp(txn, slot);
9421 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9423 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9426 if (txn->mt_flags & MDB_TXN_BLOCKED)
9429 if (txn->mt_dbflags[dbi] & DB_STALE) {
9432 /* Stale, must read the DB's root. cursor_init does it for us. */
9433 mdb_cursor_init(&mc, txn, dbi, &mx);
9435 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9438 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9441 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9443 ptr = env->me_dbxs[dbi].md_name.mv_data;
9444 /* If there was no name, this was already closed */
9446 env->me_dbxs[dbi].md_name.mv_data = NULL;
9447 env->me_dbxs[dbi].md_name.mv_size = 0;
9448 env->me_dbflags[dbi] = 0;
9449 env->me_dbiseqs[dbi]++;
9454 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9456 /* We could return the flags for the FREE_DBI too but what's the point? */
9457 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9459 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9463 /** Add all the DB's pages to the free list.
9464 * @param[in] mc Cursor on the DB to free.
9465 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9466 * @return 0 on success, non-zero on failure.
9469 mdb_drop0(MDB_cursor *mc, int subs)
9473 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9474 if (rc == MDB_SUCCESS) {
9475 MDB_txn *txn = mc->mc_txn;
9480 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9481 * This also avoids any P_LEAF2 pages, which have no nodes.
9483 if (mc->mc_flags & C_SUB)
9486 mdb_cursor_copy(mc, &mx);
9487 while (mc->mc_snum > 0) {
9488 MDB_page *mp = mc->mc_pg[mc->mc_top];
9489 unsigned n = NUMKEYS(mp);
9491 for (i=0; i<n; i++) {
9492 ni = NODEPTR(mp, i);
9493 if (ni->mn_flags & F_BIGDATA) {
9496 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9497 rc = mdb_page_get(txn, pg, &omp, NULL);
9500 mdb_cassert(mc, IS_OVERFLOW(omp));
9501 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9505 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9506 mdb_xcursor_init1(mc, ni);
9507 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9513 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9515 for (i=0; i<n; i++) {
9517 ni = NODEPTR(mp, i);
9520 mdb_midl_xappend(txn->mt_free_pgs, pg);
9525 mc->mc_ki[mc->mc_top] = i;
9526 rc = mdb_cursor_sibling(mc, 1);
9528 if (rc != MDB_NOTFOUND)
9530 /* no more siblings, go back to beginning
9531 * of previous level.
9535 for (i=1; i<mc->mc_snum; i++) {
9537 mc->mc_pg[i] = mx.mc_pg[i];
9542 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9545 txn->mt_flags |= MDB_TXN_ERROR;
9546 } else if (rc == MDB_NOTFOUND) {
9552 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9554 MDB_cursor *mc, *m2;
9557 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9560 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9563 if (TXN_DBI_CHANGED(txn, dbi))
9566 rc = mdb_cursor_open(txn, dbi, &mc);
9570 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9571 /* Invalidate the dropped DB's cursors */
9572 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9573 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9577 /* Can't delete the main DB */
9578 if (del && dbi > MAIN_DBI) {
9579 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9581 txn->mt_dbflags[dbi] = DB_STALE;
9582 mdb_dbi_close(txn->mt_env, dbi);
9584 txn->mt_flags |= MDB_TXN_ERROR;
9587 /* reset the DB record, mark it dirty */
9588 txn->mt_dbflags[dbi] |= DB_DIRTY;
9589 txn->mt_dbs[dbi].md_depth = 0;
9590 txn->mt_dbs[dbi].md_branch_pages = 0;
9591 txn->mt_dbs[dbi].md_leaf_pages = 0;
9592 txn->mt_dbs[dbi].md_overflow_pages = 0;
9593 txn->mt_dbs[dbi].md_entries = 0;
9594 txn->mt_dbs[dbi].md_root = P_INVALID;
9596 txn->mt_flags |= MDB_TXN_DIRTY;
9599 mdb_cursor_close(mc);
9603 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9605 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9608 txn->mt_dbxs[dbi].md_cmp = cmp;
9612 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9614 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9617 txn->mt_dbxs[dbi].md_dcmp = cmp;
9621 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9623 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9626 txn->mt_dbxs[dbi].md_rel = rel;
9630 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9632 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9635 txn->mt_dbxs[dbi].md_relctx = ctx;
9640 mdb_env_get_maxkeysize(MDB_env *env)
9642 return ENV_MAXKEY(env);
9646 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9648 unsigned int i, rdrs;
9651 int rc = 0, first = 1;
9655 if (!env->me_txns) {
9656 return func("(no reader locks)\n", ctx);
9658 rdrs = env->me_txns->mti_numreaders;
9659 mr = env->me_txns->mti_readers;
9660 for (i=0; i<rdrs; i++) {
9662 txnid_t txnid = mr[i].mr_txnid;
9663 sprintf(buf, txnid == (txnid_t)-1 ?
9664 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9665 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9668 rc = func(" pid thread txnid\n", ctx);
9672 rc = func(buf, ctx);
9678 rc = func("(no active readers)\n", ctx);
9683 /** Insert pid into list if not already present.
9684 * return -1 if already present.
9687 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9689 /* binary search of pid in list */
9691 unsigned cursor = 1;
9693 unsigned n = ids[0];
9696 unsigned pivot = n >> 1;
9697 cursor = base + pivot + 1;
9698 val = pid - ids[cursor];
9703 } else if ( val > 0 ) {
9708 /* found, so it's a duplicate */
9717 for (n = ids[0]; n > cursor; n--)
9724 mdb_reader_check(MDB_env *env, int *dead)
9730 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9733 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9735 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9737 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9738 unsigned int i, j, rdrs;
9740 MDB_PID_T *pids, pid;
9741 int rc = MDB_SUCCESS, count = 0;
9743 rdrs = env->me_txns->mti_numreaders;
9744 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9748 mr = env->me_txns->mti_readers;
9749 for (i=0; i<rdrs; i++) {
9751 if (pid && pid != env->me_pid) {
9752 if (mdb_pid_insert(pids, pid) == 0) {
9753 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9754 /* Stale reader found */
9757 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9758 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9760 rdrs = 0; /* the above checked all readers */
9762 /* Recheck, a new process may have reused pid */
9763 if (mdb_reader_pid(env, Pidcheck, pid))
9768 if (mr[j].mr_pid == pid) {
9769 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9770 (unsigned) pid, mr[j].mr_txnid));
9775 UNLOCK_MUTEX(rmutex);
9786 #ifdef MDB_ROBUST_SUPPORTED
9787 /** Handle #LOCK_MUTEX0() failure.
9788 * Try to repair the lock file if the mutex owner died.
9789 * @param[in] env the environment handle
9790 * @param[in] mutex LOCK_MUTEX0() mutex
9791 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9792 * @return 0 on success with the mutex locked, or an error code on failure.
9795 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9800 if (rc == MDB_OWNERDEAD) {
9801 /* We own the mutex. Clean up after dead previous owner. */
9803 rlocked = (mutex == env->me_rmutex);
9805 /* Keep mti_txnid updated, otherwise next writer can
9806 * overwrite data which latest meta page refers to.
9808 meta = mdb_env_pick_meta(env);
9809 env->me_txns->mti_txnid = meta->mm_txnid;
9810 /* env is hosed if the dead thread was ours */
9812 env->me_flags |= MDB_FATAL_ERROR;
9817 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9818 (rc ? "this process' env is hosed" : "recovering")));
9819 rc2 = mdb_reader_check0(env, rlocked, NULL);
9821 rc2 = mdb_mutex_consistent(mutex);
9822 if (rc || (rc = rc2)) {
9823 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9824 UNLOCK_MUTEX(mutex);
9830 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9835 #endif /* MDB_ROBUST_SUPPORTED */