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. */
1007 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1010 /** Number of meta pages - also hardcoded elsewhere */
1013 /** Meta page content.
1014 * A meta page is the start point for accessing a database snapshot.
1015 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1017 typedef struct MDB_meta {
1018 /** Stamp identifying this as an LMDB file. It must be set
1021 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1022 uint32_t mm_version;
1023 void *mm_address; /**< address for fixed mapping */
1024 size_t mm_mapsize; /**< size of mmap region */
1025 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1026 /** The size of pages used in this DB */
1027 #define mm_psize mm_dbs[FREE_DBI].md_pad
1028 /** Any persistent environment flags. @ref mdb_env */
1029 #define mm_flags mm_dbs[FREE_DBI].md_flags
1030 pgno_t mm_last_pg; /**< last used page in file */
1031 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1034 /** Buffer for a stack-allocated meta page.
1035 * The members define size and alignment, and silence type
1036 * aliasing warnings. They are not used directly; that could
1037 * mean incorrectly using several union members in parallel.
1039 typedef union MDB_metabuf {
1042 char mm_pad[PAGEHDRSZ];
1047 /** Auxiliary DB info.
1048 * The information here is mostly static/read-only. There is
1049 * only a single copy of this record in the environment.
1051 typedef struct MDB_dbx {
1052 MDB_val md_name; /**< name of the database */
1053 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1054 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1055 MDB_rel_func *md_rel; /**< user relocate function */
1056 void *md_relctx; /**< user-provided context for md_rel */
1059 /** A database transaction.
1060 * Every operation requires a transaction handle.
1063 MDB_txn *mt_parent; /**< parent of a nested txn */
1064 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1066 pgno_t mt_next_pgno; /**< next unallocated page */
1067 /** The ID of this transaction. IDs are integers incrementing from 1.
1068 * Only committed write transactions increment the ID. If a transaction
1069 * aborts, the ID may be re-used by the next writer.
1072 MDB_env *mt_env; /**< the DB environment */
1073 /** The list of pages that became unused during this transaction.
1075 MDB_IDL mt_free_pgs;
1076 /** The list of loose pages that became unused and may be reused
1077 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1079 MDB_page *mt_loose_pgs;
1080 /* #Number of loose pages (#mt_loose_pgs) */
1082 /** The sorted list of dirty pages we temporarily wrote to disk
1083 * because the dirty list was full. page numbers in here are
1084 * shifted left by 1, deleted slots have the LSB set.
1086 MDB_IDL mt_spill_pgs;
1088 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1089 MDB_ID2L dirty_list;
1090 /** For read txns: This thread/txn's reader table slot, or NULL. */
1093 /** Array of records for each DB known in the environment. */
1095 /** Array of MDB_db records for each known DB */
1097 /** Array of sequence numbers for each DB handle */
1098 unsigned int *mt_dbiseqs;
1099 /** @defgroup mt_dbflag Transaction DB Flags
1103 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1104 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1105 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1106 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1107 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1109 /** In write txns, array of cursors for each DB */
1110 MDB_cursor **mt_cursors;
1111 /** Array of flags for each DB */
1112 unsigned char *mt_dbflags;
1113 /** Number of DB records in use, or 0 when the txn is finished.
1114 * This number only ever increments until the txn finishes; we
1115 * don't decrement it when individual DB handles are closed.
1119 /** @defgroup mdb_txn Transaction Flags
1123 /** #mdb_txn_begin() flags */
1124 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1125 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1126 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1127 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1128 /* internal txn flags */
1129 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1130 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1131 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1132 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1133 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1134 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1135 /** most operations on the txn are currently illegal */
1136 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1138 unsigned int mt_flags; /**< @ref mdb_txn */
1139 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1140 * Includes ancestor txns' dirty pages not hidden by other txns'
1141 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1142 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1144 unsigned int mt_dirty_room;
1147 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1148 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1149 * raise this on a 64 bit machine.
1151 #define CURSOR_STACK 32
1155 /** Cursors are used for all DB operations.
1156 * A cursor holds a path of (page pointer, key index) from the DB
1157 * root to a position in the DB, plus other state. #MDB_DUPSORT
1158 * cursors include an xcursor to the current data item. Write txns
1159 * track their cursors and keep them up to date when data moves.
1160 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1161 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1164 /** Next cursor on this DB in this txn */
1165 MDB_cursor *mc_next;
1166 /** Backup of the original cursor if this cursor is a shadow */
1167 MDB_cursor *mc_backup;
1168 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1169 struct MDB_xcursor *mc_xcursor;
1170 /** The transaction that owns this cursor */
1172 /** The database handle this cursor operates on */
1174 /** The database record for this cursor */
1176 /** The database auxiliary record for this cursor */
1178 /** The @ref mt_dbflag for this database */
1179 unsigned char *mc_dbflag;
1180 unsigned short mc_snum; /**< number of pushed pages */
1181 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1182 /** @defgroup mdb_cursor Cursor Flags
1184 * Cursor state flags.
1187 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1188 #define C_EOF 0x02 /**< No more data */
1189 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1190 #define C_DEL 0x08 /**< last op was a cursor_del */
1191 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1192 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1194 unsigned int mc_flags; /**< @ref mdb_cursor */
1195 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1196 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1199 /** Context for sorted-dup records.
1200 * We could have gone to a fully recursive design, with arbitrarily
1201 * deep nesting of sub-databases. But for now we only handle these
1202 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1204 typedef struct MDB_xcursor {
1205 /** A sub-cursor for traversing the Dup DB */
1206 MDB_cursor mx_cursor;
1207 /** The database record for this Dup DB */
1209 /** The auxiliary DB record for this Dup DB */
1211 /** The @ref mt_dbflag for this Dup DB */
1212 unsigned char mx_dbflag;
1215 /** State of FreeDB old pages, stored in the MDB_env */
1216 typedef struct MDB_pgstate {
1217 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1218 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1221 /** The database environment. */
1223 HANDLE me_fd; /**< The main data file */
1224 HANDLE me_lfd; /**< The lock file */
1225 HANDLE me_mfd; /**< just for writing the meta pages */
1226 /** Failed to update the meta page. Probably an I/O error. */
1227 #define MDB_FATAL_ERROR 0x80000000U
1228 /** Some fields are initialized. */
1229 #define MDB_ENV_ACTIVE 0x20000000U
1230 /** me_txkey is set */
1231 #define MDB_ENV_TXKEY 0x10000000U
1232 /** fdatasync is unreliable */
1233 #define MDB_FSYNCONLY 0x08000000U
1234 uint32_t me_flags; /**< @ref mdb_env */
1235 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1236 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1237 unsigned int me_maxreaders; /**< size of the reader table */
1238 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1239 volatile int me_close_readers;
1240 MDB_dbi me_numdbs; /**< number of DBs opened */
1241 MDB_dbi me_maxdbs; /**< size of the DB table */
1242 MDB_PID_T me_pid; /**< process ID of this env */
1243 char *me_path; /**< path to the DB files */
1244 char *me_map; /**< the memory map of the data file */
1245 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1246 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1247 void *me_pbuf; /**< scratch area for DUPSORT put() */
1248 MDB_txn *me_txn; /**< current write transaction */
1249 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1250 size_t me_mapsize; /**< size of the data memory map */
1251 off_t me_size; /**< current file size */
1252 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1253 MDB_dbx *me_dbxs; /**< array of static DB info */
1254 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1255 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1256 pthread_key_t me_txkey; /**< thread-key for readers */
1257 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1258 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1259 # define me_pglast me_pgstate.mf_pglast
1260 # define me_pghead me_pgstate.mf_pghead
1261 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1262 /** IDL of pages that became unused in a write txn */
1263 MDB_IDL me_free_pgs;
1264 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1265 MDB_ID2L me_dirty_list;
1266 /** Max number of freelist items that can fit in a single overflow page */
1268 /** Max size of a node on a page */
1269 unsigned int me_nodemax;
1270 #if !(MDB_MAXKEYSIZE)
1271 unsigned int me_maxkey; /**< max size of a key */
1273 int me_live_reader; /**< have liveness lock in reader table */
1275 int me_pidquery; /**< Used in OpenProcess */
1277 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1278 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1279 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1281 mdb_mutex_t me_rmutex;
1282 mdb_mutex_t me_wmutex;
1284 void *me_userctx; /**< User-settable context */
1285 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1288 /** Nested transaction */
1289 typedef struct MDB_ntxn {
1290 MDB_txn mnt_txn; /**< the transaction */
1291 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1294 /** max number of pages to commit in one writev() call */
1295 #define MDB_COMMIT_PAGES 64
1296 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1297 #undef MDB_COMMIT_PAGES
1298 #define MDB_COMMIT_PAGES IOV_MAX
1301 /** max bytes to write in one call */
1302 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1304 /** Check \b txn and \b dbi arguments to a function */
1305 #define TXN_DBI_EXIST(txn, dbi, validity) \
1306 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1308 /** Check for misused \b dbi handles */
1309 #define TXN_DBI_CHANGED(txn, dbi) \
1310 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1312 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1313 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1314 static int mdb_page_touch(MDB_cursor *mc);
1316 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1317 "reset-tmp", "fail-begin", "fail-beginchild"}
1319 /* mdb_txn_end operation number, for logging */
1320 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1321 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1323 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1324 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1325 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1326 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1327 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1329 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1330 static int mdb_page_search_root(MDB_cursor *mc,
1331 MDB_val *key, int modify);
1332 #define MDB_PS_MODIFY 1
1333 #define MDB_PS_ROOTONLY 2
1334 #define MDB_PS_FIRST 4
1335 #define MDB_PS_LAST 8
1336 static int mdb_page_search(MDB_cursor *mc,
1337 MDB_val *key, int flags);
1338 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1340 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1341 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1342 pgno_t newpgno, unsigned int nflags);
1344 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1345 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1346 static int mdb_env_write_meta(MDB_txn *txn);
1347 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1348 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1350 static void mdb_env_close0(MDB_env *env, int excl);
1352 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1353 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1354 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1355 static void mdb_node_del(MDB_cursor *mc, int ksize);
1356 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1357 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1358 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1359 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1360 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1362 static int mdb_rebalance(MDB_cursor *mc);
1363 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1365 static void mdb_cursor_pop(MDB_cursor *mc);
1366 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1368 static int mdb_cursor_del0(MDB_cursor *mc);
1369 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1370 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1371 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1372 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1373 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1375 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1376 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1378 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1379 static void mdb_xcursor_init0(MDB_cursor *mc);
1380 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1382 static int mdb_drop0(MDB_cursor *mc, int subs);
1383 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1384 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1387 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1390 /** Compare two items pointing at size_t's of unknown alignment. */
1391 #ifdef MISALIGNED_OK
1392 # define mdb_cmp_clong mdb_cmp_long
1394 # define mdb_cmp_clong mdb_cmp_cint
1398 static SECURITY_DESCRIPTOR mdb_null_sd;
1399 static SECURITY_ATTRIBUTES mdb_all_sa;
1400 static int mdb_sec_inited;
1403 /** Return the library version info. */
1405 mdb_version(int *major, int *minor, int *patch)
1407 if (major) *major = MDB_VERSION_MAJOR;
1408 if (minor) *minor = MDB_VERSION_MINOR;
1409 if (patch) *patch = MDB_VERSION_PATCH;
1410 return MDB_VERSION_STRING;
1413 /** Table of descriptions for LMDB @ref errors */
1414 static char *const mdb_errstr[] = {
1415 "MDB_KEYEXIST: Key/data pair already exists",
1416 "MDB_NOTFOUND: No matching key/data pair found",
1417 "MDB_PAGE_NOTFOUND: Requested page not found",
1418 "MDB_CORRUPTED: Located page was wrong type",
1419 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1420 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1421 "MDB_INVALID: File is not an LMDB file",
1422 "MDB_MAP_FULL: Environment mapsize limit reached",
1423 "MDB_DBS_FULL: Environment maxdbs limit reached",
1424 "MDB_READERS_FULL: Environment maxreaders limit reached",
1425 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1426 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1427 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1428 "MDB_PAGE_FULL: Internal error - page has no more space",
1429 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1430 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1431 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1432 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1433 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1434 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1438 mdb_strerror(int err)
1441 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1442 * This works as long as no function between the call to mdb_strerror
1443 * and the actual use of the message uses more than 4K of stack.
1446 char buf[1024], *ptr = buf;
1450 return ("Successful return: 0");
1452 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1453 i = err - MDB_KEYEXIST;
1454 return mdb_errstr[i];
1458 /* These are the C-runtime error codes we use. The comment indicates
1459 * their numeric value, and the Win32 error they would correspond to
1460 * if the error actually came from a Win32 API. A major mess, we should
1461 * have used LMDB-specific error codes for everything.
1464 case ENOENT: /* 2, FILE_NOT_FOUND */
1465 case EIO: /* 5, ACCESS_DENIED */
1466 case ENOMEM: /* 12, INVALID_ACCESS */
1467 case EACCES: /* 13, INVALID_DATA */
1468 case EBUSY: /* 16, CURRENT_DIRECTORY */
1469 case EINVAL: /* 22, BAD_COMMAND */
1470 case ENOSPC: /* 28, OUT_OF_PAPER */
1471 return strerror(err);
1476 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1477 FORMAT_MESSAGE_IGNORE_INSERTS,
1478 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1481 return strerror(err);
1485 /** assert(3) variant in cursor context */
1486 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1487 /** assert(3) variant in transaction context */
1488 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1489 /** assert(3) variant in environment context */
1490 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1493 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1494 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1497 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1498 const char *func, const char *file, int line)
1501 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1502 file, line, expr_txt, func);
1503 if (env->me_assert_func)
1504 env->me_assert_func(env, buf);
1505 fprintf(stderr, "%s\n", buf);
1509 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1513 /** Return the page number of \b mp which may be sub-page, for debug output */
1515 mdb_dbg_pgno(MDB_page *mp)
1518 COPY_PGNO(ret, mp->mp_pgno);
1522 /** Display a key in hexadecimal and return the address of the result.
1523 * @param[in] key the key to display
1524 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1525 * @return The key in hexadecimal form.
1528 mdb_dkey(MDB_val *key, char *buf)
1531 unsigned char *c = key->mv_data;
1537 if (key->mv_size > DKBUF_MAXKEYSIZE)
1538 return "MDB_MAXKEYSIZE";
1539 /* may want to make this a dynamic check: if the key is mostly
1540 * printable characters, print it as-is instead of converting to hex.
1544 for (i=0; i<key->mv_size; i++)
1545 ptr += sprintf(ptr, "%02x", *c++);
1547 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1553 mdb_leafnode_type(MDB_node *n)
1555 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1556 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1557 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1560 /** Display all the keys in the page. */
1562 mdb_page_list(MDB_page *mp)
1564 pgno_t pgno = mdb_dbg_pgno(mp);
1565 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1567 unsigned int i, nkeys, nsize, total = 0;
1571 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1572 case P_BRANCH: type = "Branch page"; break;
1573 case P_LEAF: type = "Leaf page"; break;
1574 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1575 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1576 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1578 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1579 pgno, mp->mp_pages, state);
1582 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1583 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1586 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1590 nkeys = NUMKEYS(mp);
1591 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1593 for (i=0; i<nkeys; i++) {
1594 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1595 key.mv_size = nsize = mp->mp_pad;
1596 key.mv_data = LEAF2KEY(mp, i, nsize);
1598 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1601 node = NODEPTR(mp, i);
1602 key.mv_size = node->mn_ksize;
1603 key.mv_data = node->mn_data;
1604 nsize = NODESIZE + key.mv_size;
1605 if (IS_BRANCH(mp)) {
1606 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1610 if (F_ISSET(node->mn_flags, F_BIGDATA))
1611 nsize += sizeof(pgno_t);
1613 nsize += NODEDSZ(node);
1615 nsize += sizeof(indx_t);
1616 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1617 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1619 total = EVEN(total);
1621 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1622 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1626 mdb_cursor_chk(MDB_cursor *mc)
1632 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1633 for (i=0; i<mc->mc_top; i++) {
1635 node = NODEPTR(mp, mc->mc_ki[i]);
1636 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1639 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1645 /** Count all the pages in each DB and in the freelist
1646 * and make sure it matches the actual number of pages
1648 * All named DBs must be open for a correct count.
1650 static void mdb_audit(MDB_txn *txn)
1654 MDB_ID freecount, count;
1659 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1660 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1661 freecount += *(MDB_ID *)data.mv_data;
1662 mdb_tassert(txn, rc == MDB_NOTFOUND);
1665 for (i = 0; i<txn->mt_numdbs; i++) {
1667 if (!(txn->mt_dbflags[i] & DB_VALID))
1669 mdb_cursor_init(&mc, txn, i, &mx);
1670 if (txn->mt_dbs[i].md_root == P_INVALID)
1672 count += txn->mt_dbs[i].md_branch_pages +
1673 txn->mt_dbs[i].md_leaf_pages +
1674 txn->mt_dbs[i].md_overflow_pages;
1675 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1676 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1677 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1680 mp = mc.mc_pg[mc.mc_top];
1681 for (j=0; j<NUMKEYS(mp); j++) {
1682 MDB_node *leaf = NODEPTR(mp, j);
1683 if (leaf->mn_flags & F_SUBDATA) {
1685 memcpy(&db, NODEDATA(leaf), sizeof(db));
1686 count += db.md_branch_pages + db.md_leaf_pages +
1687 db.md_overflow_pages;
1691 mdb_tassert(txn, rc == MDB_NOTFOUND);
1694 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1695 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1696 txn->mt_txnid, freecount, count+NUM_METAS,
1697 freecount+count+NUM_METAS, txn->mt_next_pgno);
1703 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1705 return txn->mt_dbxs[dbi].md_cmp(a, b);
1709 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1711 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1712 #if UINT_MAX < SIZE_MAX
1713 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1714 dcmp = mdb_cmp_clong;
1719 /** Allocate memory for a page.
1720 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1723 mdb_page_malloc(MDB_txn *txn, unsigned num)
1725 MDB_env *env = txn->mt_env;
1726 MDB_page *ret = env->me_dpages;
1727 size_t psize = env->me_psize, sz = psize, off;
1728 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1729 * For a single page alloc, we init everything after the page header.
1730 * For multi-page, we init the final page; if the caller needed that
1731 * many pages they will be filling in at least up to the last page.
1735 VGMEMP_ALLOC(env, ret, sz);
1736 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1737 env->me_dpages = ret->mp_next;
1740 psize -= off = PAGEHDRSZ;
1745 if ((ret = malloc(sz)) != NULL) {
1746 VGMEMP_ALLOC(env, ret, sz);
1747 if (!(env->me_flags & MDB_NOMEMINIT)) {
1748 memset((char *)ret + off, 0, psize);
1752 txn->mt_flags |= MDB_TXN_ERROR;
1756 /** Free a single page.
1757 * Saves single pages to a list, for future reuse.
1758 * (This is not used for multi-page overflow pages.)
1761 mdb_page_free(MDB_env *env, MDB_page *mp)
1763 mp->mp_next = env->me_dpages;
1764 VGMEMP_FREE(env, mp);
1765 env->me_dpages = mp;
1768 /** Free a dirty page */
1770 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1772 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1773 mdb_page_free(env, dp);
1775 /* large pages just get freed directly */
1776 VGMEMP_FREE(env, dp);
1781 /** Return all dirty pages to dpage list */
1783 mdb_dlist_free(MDB_txn *txn)
1785 MDB_env *env = txn->mt_env;
1786 MDB_ID2L dl = txn->mt_u.dirty_list;
1787 unsigned i, n = dl[0].mid;
1789 for (i = 1; i <= n; i++) {
1790 mdb_dpage_free(env, dl[i].mptr);
1795 /** Loosen or free a single page.
1796 * Saves single pages to a list for future reuse
1797 * in this same txn. It has been pulled from the freeDB
1798 * and already resides on the dirty list, but has been
1799 * deleted. Use these pages first before pulling again
1802 * If the page wasn't dirtied in this txn, just add it
1803 * to this txn's free list.
1806 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1809 pgno_t pgno = mp->mp_pgno;
1810 MDB_txn *txn = mc->mc_txn;
1812 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1813 if (txn->mt_parent) {
1814 MDB_ID2 *dl = txn->mt_u.dirty_list;
1815 /* If txn has a parent, make sure the page is in our
1819 unsigned x = mdb_mid2l_search(dl, pgno);
1820 if (x <= dl[0].mid && dl[x].mid == pgno) {
1821 if (mp != dl[x].mptr) { /* bad cursor? */
1822 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1823 txn->mt_flags |= MDB_TXN_ERROR;
1824 return MDB_CORRUPTED;
1831 /* no parent txn, so it's just ours */
1836 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1838 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1839 txn->mt_loose_pgs = mp;
1840 txn->mt_loose_count++;
1841 mp->mp_flags |= P_LOOSE;
1843 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1851 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1852 * @param[in] mc A cursor handle for the current operation.
1853 * @param[in] pflags Flags of the pages to update:
1854 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1855 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1856 * @return 0 on success, non-zero on failure.
1859 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1861 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1862 MDB_txn *txn = mc->mc_txn;
1868 int rc = MDB_SUCCESS, level;
1870 /* Mark pages seen by cursors */
1871 if (mc->mc_flags & C_UNTRACK)
1872 mc = NULL; /* will find mc in mt_cursors */
1873 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1874 for (; mc; mc=mc->mc_next) {
1875 if (!(mc->mc_flags & C_INITIALIZED))
1877 for (m3 = mc;; m3 = &mx->mx_cursor) {
1879 for (j=0; j<m3->mc_snum; j++) {
1881 if ((mp->mp_flags & Mask) == pflags)
1882 mp->mp_flags ^= P_KEEP;
1884 mx = m3->mc_xcursor;
1885 /* Proceed to mx if it is at a sub-database */
1886 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1888 if (! (mp && (mp->mp_flags & P_LEAF)))
1890 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1891 if (!(leaf->mn_flags & F_SUBDATA))
1900 /* Mark dirty root pages */
1901 for (i=0; i<txn->mt_numdbs; i++) {
1902 if (txn->mt_dbflags[i] & DB_DIRTY) {
1903 pgno_t pgno = txn->mt_dbs[i].md_root;
1904 if (pgno == P_INVALID)
1906 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1908 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1909 dp->mp_flags ^= P_KEEP;
1917 static int mdb_page_flush(MDB_txn *txn, int keep);
1919 /** Spill pages from the dirty list back to disk.
1920 * This is intended to prevent running into #MDB_TXN_FULL situations,
1921 * but note that they may still occur in a few cases:
1922 * 1) our estimate of the txn size could be too small. Currently this
1923 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1924 * 2) child txns may run out of space if their parents dirtied a
1925 * lot of pages and never spilled them. TODO: we probably should do
1926 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1927 * the parent's dirty_room is below a given threshold.
1929 * Otherwise, if not using nested txns, it is expected that apps will
1930 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1931 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1932 * If the txn never references them again, they can be left alone.
1933 * If the txn only reads them, they can be used without any fuss.
1934 * If the txn writes them again, they can be dirtied immediately without
1935 * going thru all of the work of #mdb_page_touch(). Such references are
1936 * handled by #mdb_page_unspill().
1938 * Also note, we never spill DB root pages, nor pages of active cursors,
1939 * because we'll need these back again soon anyway. And in nested txns,
1940 * we can't spill a page in a child txn if it was already spilled in a
1941 * parent txn. That would alter the parent txns' data even though
1942 * the child hasn't committed yet, and we'd have no way to undo it if
1943 * the child aborted.
1945 * @param[in] m0 cursor A cursor handle identifying the transaction and
1946 * database for which we are checking space.
1947 * @param[in] key For a put operation, the key being stored.
1948 * @param[in] data For a put operation, the data being stored.
1949 * @return 0 on success, non-zero on failure.
1952 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1954 MDB_txn *txn = m0->mc_txn;
1956 MDB_ID2L dl = txn->mt_u.dirty_list;
1957 unsigned int i, j, need;
1960 if (m0->mc_flags & C_SUB)
1963 /* Estimate how much space this op will take */
1964 i = m0->mc_db->md_depth;
1965 /* Named DBs also dirty the main DB */
1966 if (m0->mc_dbi >= CORE_DBS)
1967 i += txn->mt_dbs[MAIN_DBI].md_depth;
1968 /* For puts, roughly factor in the key+data size */
1970 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1971 i += i; /* double it for good measure */
1974 if (txn->mt_dirty_room > i)
1977 if (!txn->mt_spill_pgs) {
1978 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1979 if (!txn->mt_spill_pgs)
1982 /* purge deleted slots */
1983 MDB_IDL sl = txn->mt_spill_pgs;
1984 unsigned int num = sl[0];
1986 for (i=1; i<=num; i++) {
1993 /* Preserve pages which may soon be dirtied again */
1994 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1997 /* Less aggressive spill - we originally spilled the entire dirty list,
1998 * with a few exceptions for cursor pages and DB root pages. But this
1999 * turns out to be a lot of wasted effort because in a large txn many
2000 * of those pages will need to be used again. So now we spill only 1/8th
2001 * of the dirty pages. Testing revealed this to be a good tradeoff,
2002 * better than 1/2, 1/4, or 1/10.
2004 if (need < MDB_IDL_UM_MAX / 8)
2005 need = MDB_IDL_UM_MAX / 8;
2007 /* Save the page IDs of all the pages we're flushing */
2008 /* flush from the tail forward, this saves a lot of shifting later on. */
2009 for (i=dl[0].mid; i && need; i--) {
2010 MDB_ID pn = dl[i].mid << 1;
2012 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2014 /* Can't spill twice, make sure it's not already in a parent's
2017 if (txn->mt_parent) {
2019 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2020 if (tx2->mt_spill_pgs) {
2021 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2022 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2023 dp->mp_flags |= P_KEEP;
2031 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2035 mdb_midl_sort(txn->mt_spill_pgs);
2037 /* Flush the spilled part of dirty list */
2038 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2041 /* Reset any dirty pages we kept that page_flush didn't see */
2042 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2045 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2049 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2051 mdb_find_oldest(MDB_txn *txn)
2054 txnid_t mr, oldest = txn->mt_txnid - 1;
2055 if (txn->mt_env->me_txns) {
2056 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2057 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2068 /** Add a page to the txn's dirty list */
2070 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2073 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2075 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2076 insert = mdb_mid2l_append;
2078 insert = mdb_mid2l_insert;
2080 mid.mid = mp->mp_pgno;
2082 rc = insert(txn->mt_u.dirty_list, &mid);
2083 mdb_tassert(txn, rc == 0);
2084 txn->mt_dirty_room--;
2087 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2088 * me_pghead and mt_next_pgno.
2090 * If there are free pages available from older transactions, they
2091 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2092 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2093 * and move me_pglast to say which records were consumed. Only this
2094 * function can create me_pghead and move me_pglast/mt_next_pgno.
2095 * @param[in] mc cursor A cursor handle identifying the transaction and
2096 * database for which we are allocating.
2097 * @param[in] num the number of pages to allocate.
2098 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2099 * will always be satisfied by a single contiguous chunk of memory.
2100 * @return 0 on success, non-zero on failure.
2103 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2105 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2106 /* Get at most <Max_retries> more freeDB records once me_pghead
2107 * has enough pages. If not enough, use new pages from the map.
2108 * If <Paranoid> and mc is updating the freeDB, only get new
2109 * records if me_pghead is empty. Then the freelist cannot play
2110 * catch-up with itself by growing while trying to save it.
2112 enum { Paranoid = 1, Max_retries = 500 };
2114 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2116 int rc, retry = num * 60;
2117 MDB_txn *txn = mc->mc_txn;
2118 MDB_env *env = txn->mt_env;
2119 pgno_t pgno, *mop = env->me_pghead;
2120 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2122 txnid_t oldest = 0, last;
2127 /* If there are any loose pages, just use them */
2128 if (num == 1 && txn->mt_loose_pgs) {
2129 np = txn->mt_loose_pgs;
2130 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2131 txn->mt_loose_count--;
2132 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2140 /* If our dirty list is already full, we can't do anything */
2141 if (txn->mt_dirty_room == 0) {
2146 for (op = MDB_FIRST;; op = MDB_NEXT) {
2151 /* Seek a big enough contiguous page range. Prefer
2152 * pages at the tail, just truncating the list.
2158 if (mop[i-n2] == pgno+n2)
2165 if (op == MDB_FIRST) { /* 1st iteration */
2166 /* Prepare to fetch more and coalesce */
2167 last = env->me_pglast;
2168 oldest = env->me_pgoldest;
2169 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2172 key.mv_data = &last; /* will look up last+1 */
2173 key.mv_size = sizeof(last);
2175 if (Paranoid && mc->mc_dbi == FREE_DBI)
2178 if (Paranoid && retry < 0 && mop_len)
2182 /* Do not fetch more if the record will be too recent */
2183 if (oldest <= last) {
2185 oldest = mdb_find_oldest(txn);
2186 env->me_pgoldest = oldest;
2192 rc = mdb_cursor_get(&m2, &key, NULL, op);
2194 if (rc == MDB_NOTFOUND)
2198 last = *(txnid_t*)key.mv_data;
2199 if (oldest <= last) {
2201 oldest = mdb_find_oldest(txn);
2202 env->me_pgoldest = oldest;
2208 np = m2.mc_pg[m2.mc_top];
2209 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2210 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2213 idl = (MDB_ID *) data.mv_data;
2216 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2221 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2223 mop = env->me_pghead;
2225 env->me_pglast = last;
2227 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2228 last, txn->mt_dbs[FREE_DBI].md_root, i));
2230 DPRINTF(("IDL %"Z"u", idl[j]));
2232 /* Merge in descending sorted order */
2233 mdb_midl_xmerge(mop, idl);
2237 /* Use new pages from the map when nothing suitable in the freeDB */
2239 pgno = txn->mt_next_pgno;
2240 if (pgno + num >= env->me_maxpg) {
2241 DPUTS("DB size maxed out");
2247 if (env->me_flags & MDB_WRITEMAP) {
2248 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2250 if (!(np = mdb_page_malloc(txn, num))) {
2256 mop[0] = mop_len -= num;
2257 /* Move any stragglers down */
2258 for (j = i-num; j < mop_len; )
2259 mop[++j] = mop[++i];
2261 txn->mt_next_pgno = pgno + num;
2264 mdb_page_dirty(txn, np);
2270 txn->mt_flags |= MDB_TXN_ERROR;
2274 /** Copy the used portions of a non-overflow page.
2275 * @param[in] dst page to copy into
2276 * @param[in] src page to copy from
2277 * @param[in] psize size of a page
2280 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2282 enum { Align = sizeof(pgno_t) };
2283 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2285 /* If page isn't full, just copy the used portion. Adjust
2286 * alignment so memcpy may copy words instead of bytes.
2288 if ((unused &= -Align) && !IS_LEAF2(src)) {
2289 upper = (upper + PAGEBASE) & -Align;
2290 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2291 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2294 memcpy(dst, src, psize - unused);
2298 /** Pull a page off the txn's spill list, if present.
2299 * If a page being referenced was spilled to disk in this txn, bring
2300 * it back and make it dirty/writable again.
2301 * @param[in] txn the transaction handle.
2302 * @param[in] mp the page being referenced. It must not be dirty.
2303 * @param[out] ret the writable page, if any. ret is unchanged if
2304 * mp wasn't spilled.
2307 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2309 MDB_env *env = txn->mt_env;
2312 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2314 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2315 if (!tx2->mt_spill_pgs)
2317 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2318 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2321 if (txn->mt_dirty_room == 0)
2322 return MDB_TXN_FULL;
2323 if (IS_OVERFLOW(mp))
2327 if (env->me_flags & MDB_WRITEMAP) {
2330 np = mdb_page_malloc(txn, num);
2334 memcpy(np, mp, num * env->me_psize);
2336 mdb_page_copy(np, mp, env->me_psize);
2339 /* If in current txn, this page is no longer spilled.
2340 * If it happens to be the last page, truncate the spill list.
2341 * Otherwise mark it as deleted by setting the LSB.
2343 if (x == txn->mt_spill_pgs[0])
2344 txn->mt_spill_pgs[0]--;
2346 txn->mt_spill_pgs[x] |= 1;
2347 } /* otherwise, if belonging to a parent txn, the
2348 * page remains spilled until child commits
2351 mdb_page_dirty(txn, np);
2352 np->mp_flags |= P_DIRTY;
2360 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2361 * @param[in] mc cursor pointing to the page to be touched
2362 * @return 0 on success, non-zero on failure.
2365 mdb_page_touch(MDB_cursor *mc)
2367 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2368 MDB_txn *txn = mc->mc_txn;
2369 MDB_cursor *m2, *m3;
2373 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2374 if (txn->mt_flags & MDB_TXN_SPILLS) {
2376 rc = mdb_page_unspill(txn, mp, &np);
2382 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2383 (rc = mdb_page_alloc(mc, 1, &np)))
2386 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2387 mp->mp_pgno, pgno));
2388 mdb_cassert(mc, mp->mp_pgno != pgno);
2389 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2390 /* Update the parent page, if any, to point to the new page */
2392 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2393 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2394 SETPGNO(node, pgno);
2396 mc->mc_db->md_root = pgno;
2398 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2399 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2401 /* If txn has a parent, make sure the page is in our
2405 unsigned x = mdb_mid2l_search(dl, pgno);
2406 if (x <= dl[0].mid && dl[x].mid == pgno) {
2407 if (mp != dl[x].mptr) { /* bad cursor? */
2408 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2409 txn->mt_flags |= MDB_TXN_ERROR;
2410 return MDB_CORRUPTED;
2415 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2417 np = mdb_page_malloc(txn, 1);
2422 rc = mdb_mid2l_insert(dl, &mid);
2423 mdb_cassert(mc, rc == 0);
2428 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2430 np->mp_flags |= P_DIRTY;
2433 /* Adjust cursors pointing to mp */
2434 mc->mc_pg[mc->mc_top] = np;
2435 m2 = txn->mt_cursors[mc->mc_dbi];
2436 if (mc->mc_flags & C_SUB) {
2437 for (; m2; m2=m2->mc_next) {
2438 m3 = &m2->mc_xcursor->mx_cursor;
2439 if (m3->mc_snum < mc->mc_snum) continue;
2440 if (m3->mc_pg[mc->mc_top] == mp)
2441 m3->mc_pg[mc->mc_top] = np;
2444 for (; m2; m2=m2->mc_next) {
2445 if (m2->mc_snum < mc->mc_snum) continue;
2446 if (m2->mc_pg[mc->mc_top] == mp) {
2447 m2->mc_pg[mc->mc_top] = np;
2448 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2450 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2452 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2453 if (!(leaf->mn_flags & F_SUBDATA))
2454 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2462 txn->mt_flags |= MDB_TXN_ERROR;
2467 mdb_env_sync(MDB_env *env, int force)
2470 if (env->me_flags & MDB_RDONLY)
2472 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2473 if (env->me_flags & MDB_WRITEMAP) {
2474 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2475 ? MS_ASYNC : MS_SYNC;
2476 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2479 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2483 #ifdef BROKEN_FDATASYNC
2484 if (env->me_flags & MDB_FSYNCONLY) {
2485 if (fsync(env->me_fd))
2489 if (MDB_FDATASYNC(env->me_fd))
2496 /** Back up parent txn's cursors, then grab the originals for tracking */
2498 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2500 MDB_cursor *mc, *bk;
2505 for (i = src->mt_numdbs; --i >= 0; ) {
2506 if ((mc = src->mt_cursors[i]) != NULL) {
2507 size = sizeof(MDB_cursor);
2509 size += sizeof(MDB_xcursor);
2510 for (; mc; mc = bk->mc_next) {
2516 mc->mc_db = &dst->mt_dbs[i];
2517 /* Kill pointers into src - and dst to reduce abuse: The
2518 * user may not use mc until dst ends. Otherwise we'd...
2520 mc->mc_txn = NULL; /* ...set this to dst */
2521 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2522 if ((mx = mc->mc_xcursor) != NULL) {
2523 *(MDB_xcursor *)(bk+1) = *mx;
2524 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2526 mc->mc_next = dst->mt_cursors[i];
2527 dst->mt_cursors[i] = mc;
2534 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2535 * @param[in] txn the transaction handle.
2536 * @param[in] merge true to keep changes to parent cursors, false to revert.
2537 * @return 0 on success, non-zero on failure.
2540 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2542 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2546 for (i = txn->mt_numdbs; --i >= 0; ) {
2547 for (mc = cursors[i]; mc; mc = next) {
2549 if ((bk = mc->mc_backup) != NULL) {
2551 /* Commit changes to parent txn */
2552 mc->mc_next = bk->mc_next;
2553 mc->mc_backup = bk->mc_backup;
2554 mc->mc_txn = bk->mc_txn;
2555 mc->mc_db = bk->mc_db;
2556 mc->mc_dbflag = bk->mc_dbflag;
2557 if ((mx = mc->mc_xcursor) != NULL)
2558 mx->mx_cursor.mc_txn = bk->mc_txn;
2560 /* Abort nested txn */
2562 if ((mx = mc->mc_xcursor) != NULL)
2563 *mx = *(MDB_xcursor *)(bk+1);
2567 /* Only malloced cursors are permanently tracked. */
2574 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2580 Pidset = F_SETLK, Pidcheck = F_GETLK
2584 /** Set or check a pid lock. Set returns 0 on success.
2585 * Check returns 0 if the process is certainly dead, nonzero if it may
2586 * be alive (the lock exists or an error happened so we do not know).
2588 * On Windows Pidset is a no-op, we merely check for the existence
2589 * of the process with the given pid. On POSIX we use a single byte
2590 * lock on the lockfile, set at an offset equal to the pid.
2593 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2595 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2598 if (op == Pidcheck) {
2599 h = OpenProcess(env->me_pidquery, FALSE, pid);
2600 /* No documented "no such process" code, but other program use this: */
2602 return ErrCode() != ERROR_INVALID_PARAMETER;
2603 /* A process exists until all handles to it close. Has it exited? */
2604 ret = WaitForSingleObject(h, 0) != 0;
2611 struct flock lock_info;
2612 memset(&lock_info, 0, sizeof(lock_info));
2613 lock_info.l_type = F_WRLCK;
2614 lock_info.l_whence = SEEK_SET;
2615 lock_info.l_start = pid;
2616 lock_info.l_len = 1;
2617 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2618 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2620 } else if ((rc = ErrCode()) == EINTR) {
2628 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2629 * @param[in] txn the transaction handle to initialize
2630 * @return 0 on success, non-zero on failure.
2633 mdb_txn_renew0(MDB_txn *txn)
2635 MDB_env *env = txn->mt_env;
2636 MDB_txninfo *ti = env->me_txns;
2638 unsigned int i, nr, flags = txn->mt_flags;
2640 int rc, new_notls = 0;
2642 if ((flags &= MDB_TXN_RDONLY) != 0) {
2644 meta = mdb_env_pick_meta(env);
2645 txn->mt_txnid = meta->mm_txnid;
2646 txn->mt_u.reader = NULL;
2648 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2649 pthread_getspecific(env->me_txkey);
2651 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2652 return MDB_BAD_RSLOT;
2654 MDB_PID_T pid = env->me_pid;
2655 MDB_THR_T tid = pthread_self();
2656 mdb_mutexref_t rmutex = env->me_rmutex;
2658 if (!env->me_live_reader) {
2659 rc = mdb_reader_pid(env, Pidset, pid);
2662 env->me_live_reader = 1;
2665 if (LOCK_MUTEX(rc, env, rmutex))
2667 nr = ti->mti_numreaders;
2668 for (i=0; i<nr; i++)
2669 if (ti->mti_readers[i].mr_pid == 0)
2671 if (i == env->me_maxreaders) {
2672 UNLOCK_MUTEX(rmutex);
2673 return MDB_READERS_FULL;
2675 r = &ti->mti_readers[i];
2676 /* Claim the reader slot, carefully since other code
2677 * uses the reader table un-mutexed: First reset the
2678 * slot, next publish it in mti_numreaders. After
2679 * that, it is safe for mdb_env_close() to touch it.
2680 * When it will be closed, we can finally claim it.
2683 r->mr_txnid = (txnid_t)-1;
2686 ti->mti_numreaders = ++nr;
2687 env->me_close_readers = nr;
2689 UNLOCK_MUTEX(rmutex);
2691 new_notls = (env->me_flags & MDB_NOTLS);
2692 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2697 do /* LY: Retry on a race, ITS#7970. */
2698 r->mr_txnid = ti->mti_txnid;
2699 while(r->mr_txnid != ti->mti_txnid);
2700 txn->mt_txnid = r->mr_txnid;
2701 txn->mt_u.reader = r;
2702 meta = env->me_metas[txn->mt_txnid & 1];
2706 /* Not yet touching txn == env->me_txn0, it may be active */
2708 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2710 txn->mt_txnid = ti->mti_txnid;
2711 meta = env->me_metas[txn->mt_txnid & 1];
2713 meta = mdb_env_pick_meta(env);
2714 txn->mt_txnid = meta->mm_txnid;
2718 if (txn->mt_txnid == mdb_debug_start)
2721 txn->mt_child = NULL;
2722 txn->mt_loose_pgs = NULL;
2723 txn->mt_loose_count = 0;
2724 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2725 txn->mt_u.dirty_list = env->me_dirty_list;
2726 txn->mt_u.dirty_list[0].mid = 0;
2727 txn->mt_free_pgs = env->me_free_pgs;
2728 txn->mt_free_pgs[0] = 0;
2729 txn->mt_spill_pgs = NULL;
2731 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2734 /* Copy the DB info and flags */
2735 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2737 /* Moved to here to avoid a data race in read TXNs */
2738 txn->mt_next_pgno = meta->mm_last_pg+1;
2740 txn->mt_flags = flags;
2743 txn->mt_numdbs = env->me_numdbs;
2744 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2745 x = env->me_dbflags[i];
2746 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2747 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2749 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2750 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2752 if (env->me_flags & MDB_FATAL_ERROR) {
2753 DPUTS("environment had fatal error, must shutdown!");
2755 } else if (env->me_maxpg < txn->mt_next_pgno) {
2756 rc = MDB_MAP_RESIZED;
2760 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2765 mdb_txn_renew(MDB_txn *txn)
2769 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2772 rc = mdb_txn_renew0(txn);
2773 if (rc == MDB_SUCCESS) {
2774 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2775 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2776 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2782 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2786 int rc, size, tsize;
2788 flags &= MDB_TXN_BEGIN_FLAGS;
2789 flags |= env->me_flags & MDB_WRITEMAP;
2791 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2795 /* Nested transactions: Max 1 child, write txns only, no writemap */
2796 flags |= parent->mt_flags;
2797 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2798 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2800 /* Child txns save MDB_pgstate and use own copy of cursors */
2801 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2802 size += tsize = sizeof(MDB_ntxn);
2803 } else if (flags & MDB_RDONLY) {
2804 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2805 size += tsize = sizeof(MDB_txn);
2807 /* Reuse preallocated write txn. However, do not touch it until
2808 * mdb_txn_renew0() succeeds, since it currently may be active.
2813 if ((txn = calloc(1, size)) == NULL) {
2814 DPRINTF(("calloc: %s", strerror(errno)));
2817 txn->mt_dbxs = env->me_dbxs; /* static */
2818 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2819 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2820 txn->mt_flags = flags;
2825 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2826 txn->mt_dbiseqs = parent->mt_dbiseqs;
2827 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2828 if (!txn->mt_u.dirty_list ||
2829 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2831 free(txn->mt_u.dirty_list);
2835 txn->mt_txnid = parent->mt_txnid;
2836 txn->mt_dirty_room = parent->mt_dirty_room;
2837 txn->mt_u.dirty_list[0].mid = 0;
2838 txn->mt_spill_pgs = NULL;
2839 txn->mt_next_pgno = parent->mt_next_pgno;
2840 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2841 parent->mt_child = txn;
2842 txn->mt_parent = parent;
2843 txn->mt_numdbs = parent->mt_numdbs;
2844 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2845 /* Copy parent's mt_dbflags, but clear DB_NEW */
2846 for (i=0; i<txn->mt_numdbs; i++)
2847 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2849 ntxn = (MDB_ntxn *)txn;
2850 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2851 if (env->me_pghead) {
2852 size = MDB_IDL_SIZEOF(env->me_pghead);
2853 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2855 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2860 rc = mdb_cursor_shadow(parent, txn);
2862 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2863 } else { /* MDB_RDONLY */
2864 txn->mt_dbiseqs = env->me_dbiseqs;
2866 rc = mdb_txn_renew0(txn);
2869 if (txn != env->me_txn0)
2872 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2874 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2875 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2876 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2883 mdb_txn_env(MDB_txn *txn)
2885 if(!txn) return NULL;
2890 mdb_txn_id(MDB_txn *txn)
2893 return txn->mt_txnid;
2896 /** Export or close DBI handles opened in this txn. */
2898 mdb_dbis_update(MDB_txn *txn, int keep)
2901 MDB_dbi n = txn->mt_numdbs;
2902 MDB_env *env = txn->mt_env;
2903 unsigned char *tdbflags = txn->mt_dbflags;
2905 for (i = n; --i >= CORE_DBS;) {
2906 if (tdbflags[i] & DB_NEW) {
2908 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2910 char *ptr = env->me_dbxs[i].md_name.mv_data;
2912 env->me_dbxs[i].md_name.mv_data = NULL;
2913 env->me_dbxs[i].md_name.mv_size = 0;
2914 env->me_dbflags[i] = 0;
2915 env->me_dbiseqs[i]++;
2921 if (keep && env->me_numdbs < n)
2925 /** End a transaction, except successful commit of a nested transaction.
2926 * May be called twice for readonly txns: First reset it, then abort.
2927 * @param[in] txn the transaction handle to end
2928 * @param[in] mode why and how to end the transaction
2931 mdb_txn_end(MDB_txn *txn, unsigned mode)
2933 MDB_env *env = txn->mt_env;
2935 static const char *const names[] = MDB_END_NAMES;
2938 /* Export or close DBI handles opened in this txn */
2939 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2941 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2942 names[mode & MDB_END_OPMASK],
2943 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2944 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2946 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2947 if (txn->mt_u.reader) {
2948 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2949 if (!(env->me_flags & MDB_NOTLS)) {
2950 txn->mt_u.reader = NULL; /* txn does not own reader */
2951 } else if (mode & MDB_END_SLOT) {
2952 txn->mt_u.reader->mr_pid = 0;
2953 txn->mt_u.reader = NULL;
2954 } /* else txn owns the slot until it does MDB_END_SLOT */
2956 txn->mt_numdbs = 0; /* prevent further DBI activity */
2957 txn->mt_flags |= MDB_TXN_FINISHED;
2959 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2960 pgno_t *pghead = env->me_pghead;
2962 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2963 mdb_cursors_close(txn, 0);
2964 if (!(env->me_flags & MDB_WRITEMAP)) {
2965 mdb_dlist_free(txn);
2969 txn->mt_flags = MDB_TXN_FINISHED;
2971 if (!txn->mt_parent) {
2972 mdb_midl_shrink(&txn->mt_free_pgs);
2973 env->me_free_pgs = txn->mt_free_pgs;
2975 env->me_pghead = NULL;
2979 mode = 0; /* txn == env->me_txn0, do not free() it */
2981 /* The writer mutex was locked in mdb_txn_begin. */
2983 UNLOCK_MUTEX(env->me_wmutex);
2985 txn->mt_parent->mt_child = NULL;
2986 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2987 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2988 mdb_midl_free(txn->mt_free_pgs);
2989 mdb_midl_free(txn->mt_spill_pgs);
2990 free(txn->mt_u.dirty_list);
2993 mdb_midl_free(pghead);
2996 if (mode & MDB_END_FREE)
3001 mdb_txn_reset(MDB_txn *txn)
3006 /* This call is only valid for read-only txns */
3007 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3010 mdb_txn_end(txn, MDB_END_RESET);
3014 mdb_txn_abort(MDB_txn *txn)
3020 mdb_txn_abort(txn->mt_child);
3022 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3025 /** Save the freelist as of this transaction to the freeDB.
3026 * This changes the freelist. Keep trying until it stabilizes.
3029 mdb_freelist_save(MDB_txn *txn)
3031 /* env->me_pghead[] can grow and shrink during this call.
3032 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3033 * Page numbers cannot disappear from txn->mt_free_pgs[].
3036 MDB_env *env = txn->mt_env;
3037 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3038 txnid_t pglast = 0, head_id = 0;
3039 pgno_t freecnt = 0, *free_pgs, *mop;
3040 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3042 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3044 if (env->me_pghead) {
3045 /* Make sure first page of freeDB is touched and on freelist */
3046 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3047 if (rc && rc != MDB_NOTFOUND)
3051 if (!env->me_pghead && txn->mt_loose_pgs) {
3052 /* Put loose page numbers in mt_free_pgs, since
3053 * we may be unable to return them to me_pghead.
3055 MDB_page *mp = txn->mt_loose_pgs;
3056 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3058 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3059 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3060 txn->mt_loose_pgs = NULL;
3061 txn->mt_loose_count = 0;
3064 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3065 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3066 ? SSIZE_MAX : maxfree_1pg;
3069 /* Come back here after each Put() in case freelist changed */
3074 /* If using records from freeDB which we have not yet
3075 * deleted, delete them and any we reserved for me_pghead.
3077 while (pglast < env->me_pglast) {
3078 rc = mdb_cursor_first(&mc, &key, NULL);
3081 pglast = head_id = *(txnid_t *)key.mv_data;
3082 total_room = head_room = 0;
3083 mdb_tassert(txn, pglast <= env->me_pglast);
3084 rc = mdb_cursor_del(&mc, 0);
3089 /* Save the IDL of pages freed by this txn, to a single record */
3090 if (freecnt < txn->mt_free_pgs[0]) {
3092 /* Make sure last page of freeDB is touched and on freelist */
3093 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3094 if (rc && rc != MDB_NOTFOUND)
3097 free_pgs = txn->mt_free_pgs;
3098 /* Write to last page of freeDB */
3099 key.mv_size = sizeof(txn->mt_txnid);
3100 key.mv_data = &txn->mt_txnid;
3102 freecnt = free_pgs[0];
3103 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3104 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3107 /* Retry if mt_free_pgs[] grew during the Put() */
3108 free_pgs = txn->mt_free_pgs;
3109 } while (freecnt < free_pgs[0]);
3110 mdb_midl_sort(free_pgs);
3111 memcpy(data.mv_data, free_pgs, data.mv_size);
3114 unsigned int i = free_pgs[0];
3115 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3116 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3118 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3124 mop = env->me_pghead;
3125 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3127 /* Reserve records for me_pghead[]. Split it if multi-page,
3128 * to avoid searching freeDB for a page range. Use keys in
3129 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3131 if (total_room >= mop_len) {
3132 if (total_room == mop_len || --more < 0)
3134 } else if (head_room >= maxfree_1pg && head_id > 1) {
3135 /* Keep current record (overflow page), add a new one */
3139 /* (Re)write {key = head_id, IDL length = head_room} */
3140 total_room -= head_room;
3141 head_room = mop_len - total_room;
3142 if (head_room > maxfree_1pg && head_id > 1) {
3143 /* Overflow multi-page for part of me_pghead */
3144 head_room /= head_id; /* amortize page sizes */
3145 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3146 } else if (head_room < 0) {
3147 /* Rare case, not bothering to delete this record */
3150 key.mv_size = sizeof(head_id);
3151 key.mv_data = &head_id;
3152 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3153 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3156 /* IDL is initially empty, zero out at least the length */
3157 pgs = (pgno_t *)data.mv_data;
3158 j = head_room > clean_limit ? head_room : 0;
3162 total_room += head_room;
3165 /* Return loose page numbers to me_pghead, though usually none are
3166 * left at this point. The pages themselves remain in dirty_list.
3168 if (txn->mt_loose_pgs) {
3169 MDB_page *mp = txn->mt_loose_pgs;
3170 unsigned count = txn->mt_loose_count;
3172 /* Room for loose pages + temp IDL with same */
3173 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3175 mop = env->me_pghead;
3176 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3177 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3178 loose[ ++count ] = mp->mp_pgno;
3180 mdb_midl_sort(loose);
3181 mdb_midl_xmerge(mop, loose);
3182 txn->mt_loose_pgs = NULL;
3183 txn->mt_loose_count = 0;
3187 /* Fill in the reserved me_pghead records */
3193 rc = mdb_cursor_first(&mc, &key, &data);
3194 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3195 txnid_t id = *(txnid_t *)key.mv_data;
3196 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3199 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3201 if (len > mop_len) {
3203 data.mv_size = (len + 1) * sizeof(MDB_ID);
3205 data.mv_data = mop -= len;
3208 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3210 if (rc || !(mop_len -= len))
3217 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3218 * @param[in] txn the transaction that's being committed
3219 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3220 * @return 0 on success, non-zero on failure.
3223 mdb_page_flush(MDB_txn *txn, int keep)
3225 MDB_env *env = txn->mt_env;
3226 MDB_ID2L dl = txn->mt_u.dirty_list;
3227 unsigned psize = env->me_psize, j;
3228 int i, pagecount = dl[0].mid, rc;
3229 size_t size = 0, pos = 0;
3231 MDB_page *dp = NULL;
3235 struct iovec iov[MDB_COMMIT_PAGES];
3236 ssize_t wpos = 0, wsize = 0, wres;
3237 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3243 if (env->me_flags & MDB_WRITEMAP) {
3244 /* Clear dirty flags */
3245 while (++i <= pagecount) {
3247 /* Don't flush this page yet */
3248 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3249 dp->mp_flags &= ~P_KEEP;
3253 dp->mp_flags &= ~P_DIRTY;
3258 /* Write the pages */
3260 if (++i <= pagecount) {
3262 /* Don't flush this page yet */
3263 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3264 dp->mp_flags &= ~P_KEEP;
3269 /* clear dirty flag */
3270 dp->mp_flags &= ~P_DIRTY;
3273 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3278 /* Windows actually supports scatter/gather I/O, but only on
3279 * unbuffered file handles. Since we're relying on the OS page
3280 * cache for all our data, that's self-defeating. So we just
3281 * write pages one at a time. We use the ov structure to set
3282 * the write offset, to at least save the overhead of a Seek
3285 DPRINTF(("committing page %"Z"u", pgno));
3286 memset(&ov, 0, sizeof(ov));
3287 ov.Offset = pos & 0xffffffff;
3288 ov.OffsetHigh = pos >> 16 >> 16;
3289 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3291 DPRINTF(("WriteFile: %d", rc));
3295 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3296 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3299 /* Write previous page(s) */
3300 #ifdef MDB_USE_PWRITEV
3301 wres = pwritev(env->me_fd, iov, n, wpos);
3304 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3307 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3311 DPRINTF(("lseek: %s", strerror(rc)));
3314 wres = writev(env->me_fd, iov, n);
3317 if (wres != wsize) {
3322 DPRINTF(("Write error: %s", strerror(rc)));
3324 rc = EIO; /* TODO: Use which error code? */
3325 DPUTS("short write, filesystem full?");
3336 DPRINTF(("committing page %"Z"u", pgno));
3337 next_pos = pos + size;
3338 iov[n].iov_len = size;
3339 iov[n].iov_base = (char *)dp;
3345 /* MIPS has cache coherency issues, this is a no-op everywhere else
3346 * Note: for any size >= on-chip cache size, entire on-chip cache is
3349 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3351 for (i = keep; ++i <= pagecount; ) {
3353 /* This is a page we skipped above */
3356 dl[j].mid = dp->mp_pgno;
3359 mdb_dpage_free(env, dp);
3364 txn->mt_dirty_room += i - j;
3370 mdb_txn_commit(MDB_txn *txn)
3373 unsigned int i, end_mode;
3379 /* mdb_txn_end() mode for a commit which writes nothing */
3380 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3382 if (txn->mt_child) {
3383 rc = mdb_txn_commit(txn->mt_child);
3390 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3394 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3395 DPUTS("txn has failed/finished, can't commit");
3397 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3402 if (txn->mt_parent) {
3403 MDB_txn *parent = txn->mt_parent;
3407 unsigned x, y, len, ps_len;
3409 /* Append our free list to parent's */
3410 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3413 mdb_midl_free(txn->mt_free_pgs);
3414 /* Failures after this must either undo the changes
3415 * to the parent or set MDB_TXN_ERROR in the parent.
3418 parent->mt_next_pgno = txn->mt_next_pgno;
3419 parent->mt_flags = txn->mt_flags;
3421 /* Merge our cursors into parent's and close them */
3422 mdb_cursors_close(txn, 1);
3424 /* Update parent's DB table. */
3425 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3426 parent->mt_numdbs = txn->mt_numdbs;
3427 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3428 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3429 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3430 /* preserve parent's DB_NEW status */
3431 x = parent->mt_dbflags[i] & DB_NEW;
3432 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3435 dst = parent->mt_u.dirty_list;
3436 src = txn->mt_u.dirty_list;
3437 /* Remove anything in our dirty list from parent's spill list */
3438 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3440 pspill[0] = (pgno_t)-1;
3441 /* Mark our dirty pages as deleted in parent spill list */
3442 for (i=0, len=src[0].mid; ++i <= len; ) {
3443 MDB_ID pn = src[i].mid << 1;
3444 while (pn > pspill[x])
3446 if (pn == pspill[x]) {
3451 /* Squash deleted pagenums if we deleted any */
3452 for (x=y; ++x <= ps_len; )
3453 if (!(pspill[x] & 1))
3454 pspill[++y] = pspill[x];
3458 /* Find len = length of merging our dirty list with parent's */
3460 dst[0].mid = 0; /* simplify loops */
3461 if (parent->mt_parent) {
3462 len = x + src[0].mid;
3463 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3464 for (i = x; y && i; y--) {
3465 pgno_t yp = src[y].mid;
3466 while (yp < dst[i].mid)
3468 if (yp == dst[i].mid) {
3473 } else { /* Simplify the above for single-ancestor case */
3474 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3476 /* Merge our dirty list with parent's */
3478 for (i = len; y; dst[i--] = src[y--]) {
3479 pgno_t yp = src[y].mid;
3480 while (yp < dst[x].mid)
3481 dst[i--] = dst[x--];
3482 if (yp == dst[x].mid)
3483 free(dst[x--].mptr);
3485 mdb_tassert(txn, i == x);
3487 free(txn->mt_u.dirty_list);
3488 parent->mt_dirty_room = txn->mt_dirty_room;
3489 if (txn->mt_spill_pgs) {
3490 if (parent->mt_spill_pgs) {
3491 /* TODO: Prevent failure here, so parent does not fail */
3492 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3494 parent->mt_flags |= MDB_TXN_ERROR;
3495 mdb_midl_free(txn->mt_spill_pgs);
3496 mdb_midl_sort(parent->mt_spill_pgs);
3498 parent->mt_spill_pgs = txn->mt_spill_pgs;
3502 /* Append our loose page list to parent's */
3503 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3505 *lp = txn->mt_loose_pgs;
3506 parent->mt_loose_count += txn->mt_loose_count;
3508 parent->mt_child = NULL;
3509 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3514 if (txn != env->me_txn) {
3515 DPUTS("attempt to commit unknown transaction");
3520 mdb_cursors_close(txn, 0);
3522 if (!txn->mt_u.dirty_list[0].mid &&
3523 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3526 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3527 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3529 /* Update DB root pointers */
3530 if (txn->mt_numdbs > CORE_DBS) {
3534 data.mv_size = sizeof(MDB_db);
3536 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3537 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3538 if (txn->mt_dbflags[i] & DB_DIRTY) {
3539 if (TXN_DBI_CHANGED(txn, i)) {
3543 data.mv_data = &txn->mt_dbs[i];
3544 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3552 rc = mdb_freelist_save(txn);
3556 mdb_midl_free(env->me_pghead);
3557 env->me_pghead = NULL;
3558 mdb_midl_shrink(&txn->mt_free_pgs);
3564 if ((rc = mdb_page_flush(txn, 0)))
3566 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3567 (rc = mdb_env_sync(env, 0)))
3569 if ((rc = mdb_env_write_meta(txn)))
3571 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3574 mdb_txn_end(txn, end_mode);
3582 /** Read the environment parameters of a DB environment before
3583 * mapping it into memory.
3584 * @param[in] env the environment handle
3585 * @param[out] meta address of where to store the meta information
3586 * @return 0 on success, non-zero on failure.
3589 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3595 enum { Size = sizeof(pbuf) };
3597 /* We don't know the page size yet, so use a minimum value.
3598 * Read both meta pages so we can use the latest one.
3601 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3605 memset(&ov, 0, sizeof(ov));
3607 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3608 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3611 rc = pread(env->me_fd, &pbuf, Size, off);
3614 if (rc == 0 && off == 0)
3616 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3617 DPRINTF(("read: %s", mdb_strerror(rc)));
3621 p = (MDB_page *)&pbuf;
3623 if (!F_ISSET(p->mp_flags, P_META)) {
3624 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3629 if (m->mm_magic != MDB_MAGIC) {
3630 DPUTS("meta has invalid magic");
3634 if (m->mm_version != MDB_DATA_VERSION) {
3635 DPRINTF(("database is version %u, expected version %u",
3636 m->mm_version, MDB_DATA_VERSION));
3637 return MDB_VERSION_MISMATCH;
3640 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3646 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3648 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3650 meta->mm_magic = MDB_MAGIC;
3651 meta->mm_version = MDB_DATA_VERSION;
3652 meta->mm_mapsize = env->me_mapsize;
3653 meta->mm_psize = env->me_psize;
3654 meta->mm_last_pg = NUM_METAS-1;
3655 meta->mm_flags = env->me_flags & 0xffff;
3656 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3657 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3658 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3661 /** Write the environment parameters of a freshly created DB environment.
3662 * @param[in] env the environment handle
3663 * @param[in] meta the #MDB_meta to write
3664 * @return 0 on success, non-zero on failure.
3667 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3675 memset(&ov, 0, sizeof(ov));
3676 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3678 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3681 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3682 len = pwrite(fd, ptr, size, pos); \
3683 if (len == -1 && ErrCode() == EINTR) continue; \
3684 rc = (len >= 0); break; } while(1)
3687 DPUTS("writing new meta page");
3689 psize = env->me_psize;
3691 p = calloc(NUM_METAS, psize);
3693 p->mp_flags = P_META;
3694 *(MDB_meta *)METADATA(p) = *meta;
3696 q = (MDB_page *)((char *)p + psize);
3698 q->mp_flags = P_META;
3699 *(MDB_meta *)METADATA(q) = *meta;
3701 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3704 else if ((unsigned) len == psize * NUM_METAS)
3712 /** Update the environment info to commit a transaction.
3713 * @param[in] txn the transaction that's being committed
3714 * @return 0 on success, non-zero on failure.
3717 mdb_env_write_meta(MDB_txn *txn)
3720 MDB_meta meta, metab, *mp;
3724 int rc, len, toggle;
3733 toggle = txn->mt_txnid & 1;
3734 DPRINTF(("writing meta page %d for root page %"Z"u",
3735 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3738 flags = txn->mt_flags & env->me_flags;
3739 mp = env->me_metas[toggle];
3740 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3741 /* Persist any increases of mapsize config */
3742 if (mapsize < env->me_mapsize)
3743 mapsize = env->me_mapsize;
3745 if (flags & MDB_WRITEMAP) {
3746 mp->mm_mapsize = mapsize;
3747 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3748 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3749 mp->mm_last_pg = txn->mt_next_pgno - 1;
3750 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3751 !(defined(__i386__) || defined(__x86_64__))
3752 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3753 __sync_synchronize();
3755 mp->mm_txnid = txn->mt_txnid;
3756 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3757 unsigned meta_size = env->me_psize;
3758 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3759 ptr = (char *)mp - PAGEHDRSZ;
3760 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3761 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3765 if (MDB_MSYNC(ptr, meta_size, rc)) {
3772 metab.mm_txnid = mp->mm_txnid;
3773 metab.mm_last_pg = mp->mm_last_pg;
3775 meta.mm_mapsize = mapsize;
3776 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3777 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3778 meta.mm_last_pg = txn->mt_next_pgno - 1;
3779 meta.mm_txnid = txn->mt_txnid;
3781 off = offsetof(MDB_meta, mm_mapsize);
3782 ptr = (char *)&meta + off;
3783 len = sizeof(MDB_meta) - off;
3784 off += (char *)mp - env->me_map;
3786 /* Write to the SYNC fd */
3787 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3791 memset(&ov, 0, sizeof(ov));
3793 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3797 rc = pwrite(mfd, ptr, len, off);
3800 rc = rc < 0 ? ErrCode() : EIO;
3803 DPUTS("write failed, disk error?");
3804 /* On a failure, the pagecache still contains the new data.
3805 * Write some old data back, to prevent it from being used.
3806 * Use the non-SYNC fd; we know it will fail anyway.
3808 meta.mm_last_pg = metab.mm_last_pg;
3809 meta.mm_txnid = metab.mm_txnid;
3811 memset(&ov, 0, sizeof(ov));
3813 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3815 r2 = pwrite(env->me_fd, ptr, len, off);
3816 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3819 env->me_flags |= MDB_FATAL_ERROR;
3822 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3823 CACHEFLUSH(env->me_map + off, len, DCACHE);
3825 /* Memory ordering issues are irrelevant; since the entire writer
3826 * is wrapped by wmutex, all of these changes will become visible
3827 * after the wmutex is unlocked. Since the DB is multi-version,
3828 * readers will get consistent data regardless of how fresh or
3829 * how stale their view of these values is.
3832 env->me_txns->mti_txnid = txn->mt_txnid;
3837 /** Check both meta pages to see which one is newer.
3838 * @param[in] env the environment handle
3839 * @return newest #MDB_meta.
3842 mdb_env_pick_meta(const MDB_env *env)
3844 MDB_meta *const *metas = env->me_metas;
3845 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3849 mdb_env_create(MDB_env **env)
3853 e = calloc(1, sizeof(MDB_env));
3857 e->me_maxreaders = DEFAULT_READERS;
3858 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3859 e->me_fd = INVALID_HANDLE_VALUE;
3860 e->me_lfd = INVALID_HANDLE_VALUE;
3861 e->me_mfd = INVALID_HANDLE_VALUE;
3862 #ifdef MDB_USE_POSIX_SEM
3863 e->me_rmutex = SEM_FAILED;
3864 e->me_wmutex = SEM_FAILED;
3865 #elif defined MDB_USE_SYSV_SEM
3866 e->me_rmutex->semid = -1;
3867 e->me_wmutex->semid = -1;
3869 e->me_pid = getpid();
3870 GET_PAGESIZE(e->me_os_psize);
3871 VGMEMP_CREATE(e,0,0);
3877 mdb_env_map(MDB_env *env, void *addr)
3880 unsigned int flags = env->me_flags;
3884 LONG sizelo, sizehi;
3887 if (flags & MDB_RDONLY) {
3888 /* Don't set explicit map size, use whatever exists */
3893 msize = env->me_mapsize;
3894 sizelo = msize & 0xffffffff;
3895 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3897 /* Windows won't create mappings for zero length files.
3898 * and won't map more than the file size.
3899 * Just set the maxsize right now.
3901 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3902 || !SetEndOfFile(env->me_fd)
3903 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3907 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3908 PAGE_READWRITE : PAGE_READONLY,
3909 sizehi, sizelo, NULL);
3912 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3913 FILE_MAP_WRITE : FILE_MAP_READ,
3915 rc = env->me_map ? 0 : ErrCode();
3920 int prot = PROT_READ;
3921 if (flags & MDB_WRITEMAP) {
3923 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3926 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3928 if (env->me_map == MAP_FAILED) {
3933 if (flags & MDB_NORDAHEAD) {
3934 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3936 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3938 #ifdef POSIX_MADV_RANDOM
3939 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3940 #endif /* POSIX_MADV_RANDOM */
3941 #endif /* MADV_RANDOM */
3945 /* Can happen because the address argument to mmap() is just a
3946 * hint. mmap() can pick another, e.g. if the range is in use.
3947 * The MAP_FIXED flag would prevent that, but then mmap could
3948 * instead unmap existing pages to make room for the new map.
3950 if (addr && env->me_map != addr)
3951 return EBUSY; /* TODO: Make a new MDB_* error code? */
3953 p = (MDB_page *)env->me_map;
3954 env->me_metas[0] = METADATA(p);
3955 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3961 mdb_env_set_mapsize(MDB_env *env, size_t size)
3963 /* If env is already open, caller is responsible for making
3964 * sure there are no active txns.
3972 meta = mdb_env_pick_meta(env);
3974 size = meta->mm_mapsize;
3976 /* Silently round up to minimum if the size is too small */
3977 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3981 munmap(env->me_map, env->me_mapsize);
3982 env->me_mapsize = size;
3983 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3984 rc = mdb_env_map(env, old);
3988 env->me_mapsize = size;
3990 env->me_maxpg = env->me_mapsize / env->me_psize;
3995 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3999 env->me_maxdbs = dbs + CORE_DBS;
4004 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4006 if (env->me_map || readers < 1)
4008 env->me_maxreaders = readers;
4013 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4015 if (!env || !readers)
4017 *readers = env->me_maxreaders;
4022 mdb_fsize(HANDLE fd, size_t *size)
4025 LARGE_INTEGER fsize;
4027 if (!GetFileSizeEx(fd, &fsize))
4030 *size = fsize.QuadPart;
4042 #ifdef BROKEN_FDATASYNC
4043 #include <sys/utsname.h>
4044 #include <sys/vfs.h>
4047 /** Further setup required for opening an LMDB environment
4050 mdb_env_open2(MDB_env *env)
4052 unsigned int flags = env->me_flags;
4053 int i, newenv = 0, rc;
4057 /* See if we should use QueryLimited */
4059 if ((rc & 0xff) > 5)
4060 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4062 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4065 #ifdef BROKEN_FDATASYNC
4066 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4067 * https://lkml.org/lkml/2012/9/3/83
4068 * Kernels after 3.6-rc6 are known good.
4069 * https://lkml.org/lkml/2012/9/10/556
4070 * See if the DB is on ext3/ext4, then check for new enough kernel
4071 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4076 fstatfs(env->me_fd, &st);
4077 while (st.f_type == 0xEF53) {
4081 if (uts.release[0] < '3') {
4082 if (!strncmp(uts.release, "2.6.32.", 7)) {
4083 i = atoi(uts.release+7);
4085 break; /* 2.6.32.60 and newer is OK */
4086 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4087 i = atoi(uts.release+7);
4089 break; /* 2.6.34.15 and newer is OK */
4091 } else if (uts.release[0] == '3') {
4092 i = atoi(uts.release+2);
4094 break; /* 3.6 and newer is OK */
4096 i = atoi(uts.release+4);
4098 break; /* 3.5.4 and newer is OK */
4099 } else if (i == 2) {
4100 i = atoi(uts.release+4);
4102 break; /* 3.2.30 and newer is OK */
4104 } else { /* 4.x and newer is OK */
4107 env->me_flags |= MDB_FSYNCONLY;
4113 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4116 DPUTS("new mdbenv");
4118 env->me_psize = env->me_os_psize;
4119 if (env->me_psize > MAX_PAGESIZE)
4120 env->me_psize = MAX_PAGESIZE;
4121 memset(&meta, 0, sizeof(meta));
4122 mdb_env_init_meta0(env, &meta);
4123 meta.mm_mapsize = DEFAULT_MAPSIZE;
4125 env->me_psize = meta.mm_psize;
4128 /* Was a mapsize configured? */
4129 if (!env->me_mapsize) {
4130 env->me_mapsize = meta.mm_mapsize;
4133 /* Make sure mapsize >= committed data size. Even when using
4134 * mm_mapsize, which could be broken in old files (ITS#7789).
4136 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4137 if (env->me_mapsize < minsize)
4138 env->me_mapsize = minsize;
4140 meta.mm_mapsize = env->me_mapsize;
4142 if (newenv && !(flags & MDB_FIXEDMAP)) {
4143 /* mdb_env_map() may grow the datafile. Write the metapages
4144 * first, so the file will be valid if initialization fails.
4145 * Except with FIXEDMAP, since we do not yet know mm_address.
4146 * We could fill in mm_address later, but then a different
4147 * program might end up doing that - one with a memory layout
4148 * and map address which does not suit the main program.
4150 rc = mdb_env_init_meta(env, &meta);
4156 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4161 if (flags & MDB_FIXEDMAP)
4162 meta.mm_address = env->me_map;
4163 i = mdb_env_init_meta(env, &meta);
4164 if (i != MDB_SUCCESS) {
4169 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4170 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4172 #if !(MDB_MAXKEYSIZE)
4173 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4175 env->me_maxpg = env->me_mapsize / env->me_psize;
4179 MDB_meta *meta = mdb_env_pick_meta(env);
4180 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4182 DPRINTF(("opened database version %u, pagesize %u",
4183 meta->mm_version, env->me_psize));
4184 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4185 DPRINTF(("depth: %u", db->md_depth));
4186 DPRINTF(("entries: %"Z"u", db->md_entries));
4187 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4188 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4189 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4190 DPRINTF(("root: %"Z"u", db->md_root));
4198 /** Release a reader thread's slot in the reader lock table.
4199 * This function is called automatically when a thread exits.
4200 * @param[in] ptr This points to the slot in the reader lock table.
4203 mdb_env_reader_dest(void *ptr)
4205 MDB_reader *reader = ptr;
4211 /** Junk for arranging thread-specific callbacks on Windows. This is
4212 * necessarily platform and compiler-specific. Windows supports up
4213 * to 1088 keys. Let's assume nobody opens more than 64 environments
4214 * in a single process, for now. They can override this if needed.
4216 #ifndef MAX_TLS_KEYS
4217 #define MAX_TLS_KEYS 64
4219 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4220 static int mdb_tls_nkeys;
4222 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4226 case DLL_PROCESS_ATTACH: break;
4227 case DLL_THREAD_ATTACH: break;
4228 case DLL_THREAD_DETACH:
4229 for (i=0; i<mdb_tls_nkeys; i++) {
4230 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4232 mdb_env_reader_dest(r);
4236 case DLL_PROCESS_DETACH: break;
4241 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4243 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4247 /* Force some symbol references.
4248 * _tls_used forces the linker to create the TLS directory if not already done
4249 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4251 #pragma comment(linker, "/INCLUDE:_tls_used")
4252 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4253 #pragma const_seg(".CRT$XLB")
4254 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4255 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4258 #pragma comment(linker, "/INCLUDE:__tls_used")
4259 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4260 #pragma data_seg(".CRT$XLB")
4261 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4263 #endif /* WIN 32/64 */
4264 #endif /* !__GNUC__ */
4267 /** Downgrade the exclusive lock on the region back to shared */
4269 mdb_env_share_locks(MDB_env *env, int *excl)
4272 MDB_meta *meta = mdb_env_pick_meta(env);
4274 env->me_txns->mti_txnid = meta->mm_txnid;
4279 /* First acquire a shared lock. The Unlock will
4280 * then release the existing exclusive lock.
4282 memset(&ov, 0, sizeof(ov));
4283 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4286 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4292 struct flock lock_info;
4293 /* The shared lock replaces the existing lock */
4294 memset((void *)&lock_info, 0, sizeof(lock_info));
4295 lock_info.l_type = F_RDLCK;
4296 lock_info.l_whence = SEEK_SET;
4297 lock_info.l_start = 0;
4298 lock_info.l_len = 1;
4299 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4300 (rc = ErrCode()) == EINTR) ;
4301 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4308 /** Try to get exclusive lock, otherwise shared.
4309 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4312 mdb_env_excl_lock(MDB_env *env, int *excl)
4316 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4320 memset(&ov, 0, sizeof(ov));
4321 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4328 struct flock lock_info;
4329 memset((void *)&lock_info, 0, sizeof(lock_info));
4330 lock_info.l_type = F_WRLCK;
4331 lock_info.l_whence = SEEK_SET;
4332 lock_info.l_start = 0;
4333 lock_info.l_len = 1;
4334 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4335 (rc = ErrCode()) == EINTR) ;
4339 # ifndef MDB_USE_POSIX_MUTEX
4340 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4343 lock_info.l_type = F_RDLCK;
4344 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4345 (rc = ErrCode()) == EINTR) ;
4355 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4357 * @(#) $Revision: 5.1 $
4358 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4359 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4361 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4365 * Please do not copyright this code. This code is in the public domain.
4367 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4368 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4369 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4370 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4371 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4372 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4373 * PERFORMANCE OF THIS SOFTWARE.
4376 * chongo <Landon Curt Noll> /\oo/\
4377 * http://www.isthe.com/chongo/
4379 * Share and Enjoy! :-)
4382 typedef unsigned long long mdb_hash_t;
4383 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4385 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4386 * @param[in] val value to hash
4387 * @param[in] hval initial value for hash
4388 * @return 64 bit hash
4390 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4391 * hval arg on the first call.
4394 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4396 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4397 unsigned char *end = s + val->mv_size;
4399 * FNV-1a hash each octet of the string
4402 /* xor the bottom with the current octet */
4403 hval ^= (mdb_hash_t)*s++;
4405 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4406 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4407 (hval << 7) + (hval << 8) + (hval << 40);
4409 /* return our new hash value */
4413 /** Hash the string and output the encoded hash.
4414 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4415 * very short name limits. We don't care about the encoding being reversible,
4416 * we just want to preserve as many bits of the input as possible in a
4417 * small printable string.
4418 * @param[in] str string to hash
4419 * @param[out] encbuf an array of 11 chars to hold the hash
4421 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4424 mdb_pack85(unsigned long l, char *out)
4428 for (i=0; i<5; i++) {
4429 *out++ = mdb_a85[l % 85];
4435 mdb_hash_enc(MDB_val *val, char *encbuf)
4437 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4439 mdb_pack85(h, encbuf);
4440 mdb_pack85(h>>32, encbuf+5);
4445 /** Open and/or initialize the lock region for the environment.
4446 * @param[in] env The LMDB environment.
4447 * @param[in] lpath The pathname of the file used for the lock region.
4448 * @param[in] mode The Unix permissions for the file, if we create it.
4449 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4450 * @return 0 on success, non-zero on failure.
4453 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4456 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4458 # define MDB_ERRCODE_ROFS EROFS
4459 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4460 # define MDB_CLOEXEC O_CLOEXEC
4463 # define MDB_CLOEXEC 0
4466 #ifdef MDB_USE_SYSV_SEM
4474 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4475 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4476 FILE_ATTRIBUTE_NORMAL, NULL);
4478 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4480 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4482 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4487 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4488 /* Lose record locks when exec*() */
4489 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4490 fcntl(env->me_lfd, F_SETFD, fdflags);
4493 if (!(env->me_flags & MDB_NOTLS)) {
4494 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4497 env->me_flags |= MDB_ENV_TXKEY;
4499 /* Windows TLS callbacks need help finding their TLS info. */
4500 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4504 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4508 /* Try to get exclusive lock. If we succeed, then
4509 * nobody is using the lock region and we should initialize it.
4511 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4514 size = GetFileSize(env->me_lfd, NULL);
4516 size = lseek(env->me_lfd, 0, SEEK_END);
4517 if (size == -1) goto fail_errno;
4519 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4520 if (size < rsize && *excl > 0) {
4522 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4523 || !SetEndOfFile(env->me_lfd))
4526 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4530 size = rsize - sizeof(MDB_txninfo);
4531 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4536 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4538 if (!mh) goto fail_errno;
4539 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4541 if (!env->me_txns) goto fail_errno;
4543 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4545 if (m == MAP_FAILED) goto fail_errno;
4551 BY_HANDLE_FILE_INFORMATION stbuf;
4560 if (!mdb_sec_inited) {
4561 InitializeSecurityDescriptor(&mdb_null_sd,
4562 SECURITY_DESCRIPTOR_REVISION);
4563 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4564 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4565 mdb_all_sa.bInheritHandle = FALSE;
4566 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4569 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4570 idbuf.volume = stbuf.dwVolumeSerialNumber;
4571 idbuf.nhigh = stbuf.nFileIndexHigh;
4572 idbuf.nlow = stbuf.nFileIndexLow;
4573 val.mv_data = &idbuf;
4574 val.mv_size = sizeof(idbuf);
4575 mdb_hash_enc(&val, encbuf);
4576 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4577 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4578 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4579 if (!env->me_rmutex) goto fail_errno;
4580 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4581 if (!env->me_wmutex) goto fail_errno;
4582 #elif defined(MDB_USE_POSIX_SEM)
4591 #if defined(__NetBSD__)
4592 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4594 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4595 idbuf.dev = stbuf.st_dev;
4596 idbuf.ino = stbuf.st_ino;
4597 val.mv_data = &idbuf;
4598 val.mv_size = sizeof(idbuf);
4599 mdb_hash_enc(&val, encbuf);
4600 #ifdef MDB_SHORT_SEMNAMES
4601 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4603 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4604 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4605 /* Clean up after a previous run, if needed: Try to
4606 * remove both semaphores before doing anything else.
4608 sem_unlink(env->me_txns->mti_rmname);
4609 sem_unlink(env->me_txns->mti_wmname);
4610 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4611 O_CREAT|O_EXCL, mode, 1);
4612 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4613 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4614 O_CREAT|O_EXCL, mode, 1);
4615 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4616 #elif defined(MDB_USE_SYSV_SEM)
4617 unsigned short vals[2] = {1, 1};
4618 key_t key = ftok(lpath, 'M');
4621 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4625 if (semctl(semid, 0, SETALL, semu) < 0)
4627 env->me_txns->mti_semid = semid;
4628 #else /* MDB_USE_POSIX_MUTEX: */
4629 pthread_mutexattr_t mattr;
4631 if ((rc = pthread_mutexattr_init(&mattr))
4632 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4633 #ifdef MDB_ROBUST_SUPPORTED
4634 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4636 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4637 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4639 pthread_mutexattr_destroy(&mattr);
4640 #endif /* _WIN32 || ... */
4642 env->me_txns->mti_magic = MDB_MAGIC;
4643 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4644 env->me_txns->mti_txnid = 0;
4645 env->me_txns->mti_numreaders = 0;
4648 #ifdef MDB_USE_SYSV_SEM
4649 struct semid_ds buf;
4651 if (env->me_txns->mti_magic != MDB_MAGIC) {
4652 DPUTS("lock region has invalid magic");
4656 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4657 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4658 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4659 rc = MDB_VERSION_MISMATCH;
4663 if (rc && rc != EACCES && rc != EAGAIN) {
4667 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4668 if (!env->me_rmutex) goto fail_errno;
4669 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4670 if (!env->me_wmutex) goto fail_errno;
4671 #elif defined(MDB_USE_POSIX_SEM)
4672 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4673 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4674 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4675 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4676 #elif defined(MDB_USE_SYSV_SEM)
4677 semid = env->me_txns->mti_semid;
4679 /* check for read access */
4680 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4682 /* check for write access */
4683 if (semctl(semid, 0, IPC_SET, semu) < 0)
4687 #ifdef MDB_USE_SYSV_SEM
4688 env->me_rmutex->semid = semid;
4689 env->me_wmutex->semid = semid;
4690 env->me_rmutex->semnum = 0;
4691 env->me_wmutex->semnum = 1;
4692 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4693 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4704 /** The name of the lock file in the DB environment */
4705 #define LOCKNAME "/lock.mdb"
4706 /** The name of the data file in the DB environment */
4707 #define DATANAME "/data.mdb"
4708 /** The suffix of the lock file when no subdir is used */
4709 #define LOCKSUFF "-lock"
4710 /** Only a subset of the @ref mdb_env flags can be changed
4711 * at runtime. Changing other flags requires closing the
4712 * environment and re-opening it with the new flags.
4714 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4715 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4716 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4718 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4719 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4723 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4725 int oflags, rc, len, excl = -1;
4726 char *lpath, *dpath;
4728 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4732 if (flags & MDB_NOSUBDIR) {
4733 rc = len + sizeof(LOCKSUFF) + len + 1;
4735 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4740 if (flags & MDB_NOSUBDIR) {
4741 dpath = lpath + len + sizeof(LOCKSUFF);
4742 sprintf(lpath, "%s" LOCKSUFF, path);
4743 strcpy(dpath, path);
4745 dpath = lpath + len + sizeof(LOCKNAME);
4746 sprintf(lpath, "%s" LOCKNAME, path);
4747 sprintf(dpath, "%s" DATANAME, path);
4751 flags |= env->me_flags;
4752 if (flags & MDB_RDONLY) {
4753 /* silently ignore WRITEMAP when we're only getting read access */
4754 flags &= ~MDB_WRITEMAP;
4756 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4757 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4760 env->me_flags = flags |= MDB_ENV_ACTIVE;
4764 env->me_path = strdup(path);
4765 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4766 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4767 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4768 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4772 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4774 /* For RDONLY, get lockfile after we know datafile exists */
4775 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4776 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4782 if (F_ISSET(flags, MDB_RDONLY)) {
4783 oflags = GENERIC_READ;
4784 len = OPEN_EXISTING;
4786 oflags = GENERIC_READ|GENERIC_WRITE;
4789 mode = FILE_ATTRIBUTE_NORMAL;
4790 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4791 NULL, len, mode, NULL);
4793 if (F_ISSET(flags, MDB_RDONLY))
4796 oflags = O_RDWR | O_CREAT;
4798 env->me_fd = open(dpath, oflags, mode);
4800 if (env->me_fd == INVALID_HANDLE_VALUE) {
4805 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4806 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4811 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4812 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4813 env->me_mfd = env->me_fd;
4815 /* Synchronous fd for meta writes. Needed even with
4816 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4819 len = OPEN_EXISTING;
4820 env->me_mfd = CreateFile(dpath, oflags,
4821 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4822 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4825 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4827 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4832 DPRINTF(("opened dbenv %p", (void *) env));
4834 rc = mdb_env_share_locks(env, &excl);
4838 if (!(flags & MDB_RDONLY)) {
4840 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4841 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4842 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4843 (txn = calloc(1, size)))
4845 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4846 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4847 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4848 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4850 txn->mt_dbxs = env->me_dbxs;
4851 txn->mt_flags = MDB_TXN_FINISHED;
4861 mdb_env_close0(env, excl);
4867 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4869 mdb_env_close0(MDB_env *env, int excl)
4873 if (!(env->me_flags & MDB_ENV_ACTIVE))
4876 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4878 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4879 free(env->me_dbxs[i].md_name.mv_data);
4884 free(env->me_dbiseqs);
4885 free(env->me_dbflags);
4887 free(env->me_dirty_list);
4889 mdb_midl_free(env->me_free_pgs);
4891 if (env->me_flags & MDB_ENV_TXKEY) {
4892 pthread_key_delete(env->me_txkey);
4894 /* Delete our key from the global list */
4895 for (i=0; i<mdb_tls_nkeys; i++)
4896 if (mdb_tls_keys[i] == env->me_txkey) {
4897 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4905 munmap(env->me_map, env->me_mapsize);
4907 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4908 (void) close(env->me_mfd);
4909 if (env->me_fd != INVALID_HANDLE_VALUE)
4910 (void) close(env->me_fd);
4912 MDB_PID_T pid = env->me_pid;
4913 /* Clearing readers is done in this function because
4914 * me_txkey with its destructor must be disabled first.
4916 * We skip the the reader mutex, so we touch only
4917 * data owned by this process (me_close_readers and
4918 * our readers), and clear each reader atomically.
4920 for (i = env->me_close_readers; --i >= 0; )
4921 if (env->me_txns->mti_readers[i].mr_pid == pid)
4922 env->me_txns->mti_readers[i].mr_pid = 0;
4924 if (env->me_rmutex) {
4925 CloseHandle(env->me_rmutex);
4926 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4928 /* Windows automatically destroys the mutexes when
4929 * the last handle closes.
4931 #elif defined(MDB_USE_POSIX_SEM)
4932 if (env->me_rmutex != SEM_FAILED) {
4933 sem_close(env->me_rmutex);
4934 if (env->me_wmutex != SEM_FAILED)
4935 sem_close(env->me_wmutex);
4936 /* If we have the filelock: If we are the
4937 * only remaining user, clean up semaphores.
4940 mdb_env_excl_lock(env, &excl);
4942 sem_unlink(env->me_txns->mti_rmname);
4943 sem_unlink(env->me_txns->mti_wmname);
4946 #elif defined(MDB_USE_SYSV_SEM)
4947 if (env->me_rmutex->semid != -1) {
4948 /* If we have the filelock: If we are the
4949 * only remaining user, clean up semaphores.
4952 mdb_env_excl_lock(env, &excl);
4954 semctl(env->me_rmutex->semid, 0, IPC_RMID);
4957 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4959 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4962 /* Unlock the lockfile. Windows would have unlocked it
4963 * after closing anyway, but not necessarily at once.
4965 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4968 (void) close(env->me_lfd);
4971 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4975 mdb_env_close(MDB_env *env)
4982 VGMEMP_DESTROY(env);
4983 while ((dp = env->me_dpages) != NULL) {
4984 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4985 env->me_dpages = dp->mp_next;
4989 mdb_env_close0(env, 0);
4993 /** Compare two items pointing at aligned size_t's */
4995 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4997 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4998 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5001 /** Compare two items pointing at aligned unsigned int's.
5003 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5004 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5007 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5009 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5010 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5013 /** Compare two items pointing at unsigned ints of unknown alignment.
5014 * Nodes and keys are guaranteed to be 2-byte aligned.
5017 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5019 #if BYTE_ORDER == LITTLE_ENDIAN
5020 unsigned short *u, *c;
5023 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5024 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5027 } while(!x && u > (unsigned short *)a->mv_data);
5030 unsigned short *u, *c, *end;
5033 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5034 u = (unsigned short *)a->mv_data;
5035 c = (unsigned short *)b->mv_data;
5038 } while(!x && u < end);
5043 /** Compare two items lexically */
5045 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5052 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5058 diff = memcmp(a->mv_data, b->mv_data, len);
5059 return diff ? diff : len_diff<0 ? -1 : len_diff;
5062 /** Compare two items in reverse byte order */
5064 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5066 const unsigned char *p1, *p2, *p1_lim;
5070 p1_lim = (const unsigned char *)a->mv_data;
5071 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5072 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5074 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5080 while (p1 > p1_lim) {
5081 diff = *--p1 - *--p2;
5085 return len_diff<0 ? -1 : len_diff;
5088 /** Search for key within a page, using binary search.
5089 * Returns the smallest entry larger or equal to the key.
5090 * If exactp is non-null, stores whether the found entry was an exact match
5091 * in *exactp (1 or 0).
5092 * Updates the cursor index with the index of the found entry.
5093 * If no entry larger or equal to the key is found, returns NULL.
5096 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5098 unsigned int i = 0, nkeys;
5101 MDB_page *mp = mc->mc_pg[mc->mc_top];
5102 MDB_node *node = NULL;
5107 nkeys = NUMKEYS(mp);
5109 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5110 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5113 low = IS_LEAF(mp) ? 0 : 1;
5115 cmp = mc->mc_dbx->md_cmp;
5117 /* Branch pages have no data, so if using integer keys,
5118 * alignment is guaranteed. Use faster mdb_cmp_int.
5120 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5121 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5128 nodekey.mv_size = mc->mc_db->md_pad;
5129 node = NODEPTR(mp, 0); /* fake */
5130 while (low <= high) {
5131 i = (low + high) >> 1;
5132 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5133 rc = cmp(key, &nodekey);
5134 DPRINTF(("found leaf index %u [%s], rc = %i",
5135 i, DKEY(&nodekey), rc));
5144 while (low <= high) {
5145 i = (low + high) >> 1;
5147 node = NODEPTR(mp, i);
5148 nodekey.mv_size = NODEKSZ(node);
5149 nodekey.mv_data = NODEKEY(node);
5151 rc = cmp(key, &nodekey);
5154 DPRINTF(("found leaf index %u [%s], rc = %i",
5155 i, DKEY(&nodekey), rc));
5157 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5158 i, DKEY(&nodekey), NODEPGNO(node), rc));
5169 if (rc > 0) { /* Found entry is less than the key. */
5170 i++; /* Skip to get the smallest entry larger than key. */
5172 node = NODEPTR(mp, i);
5175 *exactp = (rc == 0 && nkeys > 0);
5176 /* store the key index */
5177 mc->mc_ki[mc->mc_top] = i;
5179 /* There is no entry larger or equal to the key. */
5182 /* nodeptr is fake for LEAF2 */
5188 mdb_cursor_adjust(MDB_cursor *mc, func)
5192 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5193 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5200 /** Pop a page off the top of the cursor's stack. */
5202 mdb_cursor_pop(MDB_cursor *mc)
5205 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5206 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5214 /** Push a page onto the top of the cursor's stack. */
5216 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5218 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5219 DDBI(mc), (void *) mc));
5221 if (mc->mc_snum >= CURSOR_STACK) {
5222 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5223 return MDB_CURSOR_FULL;
5226 mc->mc_top = mc->mc_snum++;
5227 mc->mc_pg[mc->mc_top] = mp;
5228 mc->mc_ki[mc->mc_top] = 0;
5233 /** Find the address of the page corresponding to a given page number.
5234 * @param[in] txn the transaction for this access.
5235 * @param[in] pgno the page number for the page to retrieve.
5236 * @param[out] ret address of a pointer where the page's address will be stored.
5237 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5238 * @return 0 on success, non-zero on failure.
5241 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5243 MDB_env *env = txn->mt_env;
5247 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5251 MDB_ID2L dl = tx2->mt_u.dirty_list;
5253 /* Spilled pages were dirtied in this txn and flushed
5254 * because the dirty list got full. Bring this page
5255 * back in from the map (but don't unspill it here,
5256 * leave that unless page_touch happens again).
5258 if (tx2->mt_spill_pgs) {
5259 MDB_ID pn = pgno << 1;
5260 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5261 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5262 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5267 unsigned x = mdb_mid2l_search(dl, pgno);
5268 if (x <= dl[0].mid && dl[x].mid == pgno) {
5274 } while ((tx2 = tx2->mt_parent) != NULL);
5277 if (pgno < txn->mt_next_pgno) {
5279 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5281 DPRINTF(("page %"Z"u not found", pgno));
5282 txn->mt_flags |= MDB_TXN_ERROR;
5283 return MDB_PAGE_NOTFOUND;
5293 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5294 * The cursor is at the root page, set up the rest of it.
5297 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5299 MDB_page *mp = mc->mc_pg[mc->mc_top];
5303 while (IS_BRANCH(mp)) {
5307 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5308 mdb_cassert(mc, NUMKEYS(mp) > 1);
5309 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5311 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5313 if (flags & MDB_PS_LAST)
5314 i = NUMKEYS(mp) - 1;
5317 node = mdb_node_search(mc, key, &exact);
5319 i = NUMKEYS(mp) - 1;
5321 i = mc->mc_ki[mc->mc_top];
5323 mdb_cassert(mc, i > 0);
5327 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5330 mdb_cassert(mc, i < NUMKEYS(mp));
5331 node = NODEPTR(mp, i);
5333 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5336 mc->mc_ki[mc->mc_top] = i;
5337 if ((rc = mdb_cursor_push(mc, mp)))
5340 if (flags & MDB_PS_MODIFY) {
5341 if ((rc = mdb_page_touch(mc)) != 0)
5343 mp = mc->mc_pg[mc->mc_top];
5348 DPRINTF(("internal error, index points to a %02X page!?",
5350 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5351 return MDB_CORRUPTED;
5354 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5355 key ? DKEY(key) : "null"));
5356 mc->mc_flags |= C_INITIALIZED;
5357 mc->mc_flags &= ~C_EOF;
5362 /** Search for the lowest key under the current branch page.
5363 * This just bypasses a NUMKEYS check in the current page
5364 * before calling mdb_page_search_root(), because the callers
5365 * are all in situations where the current page is known to
5369 mdb_page_search_lowest(MDB_cursor *mc)
5371 MDB_page *mp = mc->mc_pg[mc->mc_top];
5372 MDB_node *node = NODEPTR(mp, 0);
5375 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5378 mc->mc_ki[mc->mc_top] = 0;
5379 if ((rc = mdb_cursor_push(mc, mp)))
5381 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5384 /** Search for the page a given key should be in.
5385 * Push it and its parent pages on the cursor stack.
5386 * @param[in,out] mc the cursor for this operation.
5387 * @param[in] key the key to search for, or NULL for first/last page.
5388 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5389 * are touched (updated with new page numbers).
5390 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5391 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5392 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5393 * @return 0 on success, non-zero on failure.
5396 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5401 /* Make sure the txn is still viable, then find the root from
5402 * the txn's db table and set it as the root of the cursor's stack.
5404 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5405 DPUTS("transaction may not be used now");
5408 /* Make sure we're using an up-to-date root */
5409 if (*mc->mc_dbflag & DB_STALE) {
5411 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5413 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5414 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5421 MDB_node *leaf = mdb_node_search(&mc2,
5422 &mc->mc_dbx->md_name, &exact);
5424 return MDB_NOTFOUND;
5425 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5426 return MDB_INCOMPATIBLE; /* not a named DB */
5427 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5430 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5432 /* The txn may not know this DBI, or another process may
5433 * have dropped and recreated the DB with other flags.
5435 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5436 return MDB_INCOMPATIBLE;
5437 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5439 *mc->mc_dbflag &= ~DB_STALE;
5441 root = mc->mc_db->md_root;
5443 if (root == P_INVALID) { /* Tree is empty. */
5444 DPUTS("tree is empty");
5445 return MDB_NOTFOUND;
5449 mdb_cassert(mc, root > 1);
5450 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5451 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5457 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5458 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5460 if (flags & MDB_PS_MODIFY) {
5461 if ((rc = mdb_page_touch(mc)))
5465 if (flags & MDB_PS_ROOTONLY)
5468 return mdb_page_search_root(mc, key, flags);
5472 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5474 MDB_txn *txn = mc->mc_txn;
5475 pgno_t pg = mp->mp_pgno;
5476 unsigned x = 0, ovpages = mp->mp_pages;
5477 MDB_env *env = txn->mt_env;
5478 MDB_IDL sl = txn->mt_spill_pgs;
5479 MDB_ID pn = pg << 1;
5482 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5483 /* If the page is dirty or on the spill list we just acquired it,
5484 * so we should give it back to our current free list, if any.
5485 * Otherwise put it onto the list of pages we freed in this txn.
5487 * Won't create me_pghead: me_pglast must be inited along with it.
5488 * Unsupported in nested txns: They would need to hide the page
5489 * range in ancestor txns' dirty and spilled lists.
5491 if (env->me_pghead &&
5493 ((mp->mp_flags & P_DIRTY) ||
5494 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5498 MDB_ID2 *dl, ix, iy;
5499 rc = mdb_midl_need(&env->me_pghead, ovpages);
5502 if (!(mp->mp_flags & P_DIRTY)) {
5503 /* This page is no longer spilled */
5510 /* Remove from dirty list */
5511 dl = txn->mt_u.dirty_list;
5513 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5519 mdb_cassert(mc, x > 1);
5521 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5522 txn->mt_flags |= MDB_TXN_ERROR;
5523 return MDB_CORRUPTED;
5526 if (!(env->me_flags & MDB_WRITEMAP))
5527 mdb_dpage_free(env, mp);
5529 /* Insert in me_pghead */
5530 mop = env->me_pghead;
5531 j = mop[0] + ovpages;
5532 for (i = mop[0]; i && mop[i] < pg; i--)
5538 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5542 mc->mc_db->md_overflow_pages -= ovpages;
5546 /** Return the data associated with a given node.
5547 * @param[in] txn The transaction for this operation.
5548 * @param[in] leaf The node being read.
5549 * @param[out] data Updated to point to the node's data.
5550 * @return 0 on success, non-zero on failure.
5553 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5555 MDB_page *omp; /* overflow page */
5559 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5560 data->mv_size = NODEDSZ(leaf);
5561 data->mv_data = NODEDATA(leaf);
5565 /* Read overflow data.
5567 data->mv_size = NODEDSZ(leaf);
5568 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5569 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5570 DPRINTF(("read overflow page %"Z"u failed", pgno));
5573 data->mv_data = METADATA(omp);
5579 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5580 MDB_val *key, MDB_val *data)
5587 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5589 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5592 if (txn->mt_flags & MDB_TXN_BLOCKED)
5595 mdb_cursor_init(&mc, txn, dbi, &mx);
5596 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5599 /** Find a sibling for a page.
5600 * Replaces the page at the top of the cursor's stack with the
5601 * specified sibling, if one exists.
5602 * @param[in] mc The cursor for this operation.
5603 * @param[in] move_right Non-zero if the right sibling is requested,
5604 * otherwise the left sibling.
5605 * @return 0 on success, non-zero on failure.
5608 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5614 if (mc->mc_snum < 2) {
5615 return MDB_NOTFOUND; /* root has no siblings */
5619 DPRINTF(("parent page is page %"Z"u, index %u",
5620 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5622 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5623 : (mc->mc_ki[mc->mc_top] == 0)) {
5624 DPRINTF(("no more keys left, moving to %s sibling",
5625 move_right ? "right" : "left"));
5626 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5627 /* undo cursor_pop before returning */
5634 mc->mc_ki[mc->mc_top]++;
5636 mc->mc_ki[mc->mc_top]--;
5637 DPRINTF(("just moving to %s index key %u",
5638 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5640 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5642 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5643 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5644 /* mc will be inconsistent if caller does mc_snum++ as above */
5645 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5649 mdb_cursor_push(mc, mp);
5651 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5656 /** Move the cursor to the next data item. */
5658 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5664 if (mc->mc_flags & C_EOF) {
5665 return MDB_NOTFOUND;
5668 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5670 mp = mc->mc_pg[mc->mc_top];
5672 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5673 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5674 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5675 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5676 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5677 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5678 if (rc == MDB_SUCCESS)
5679 MDB_GET_KEY(leaf, key);
5684 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5685 if (op == MDB_NEXT_DUP)
5686 return MDB_NOTFOUND;
5690 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5691 mdb_dbg_pgno(mp), (void *) mc));
5692 if (mc->mc_flags & C_DEL)
5695 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5696 DPUTS("=====> move to next sibling page");
5697 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5698 mc->mc_flags |= C_EOF;
5701 mp = mc->mc_pg[mc->mc_top];
5702 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5704 mc->mc_ki[mc->mc_top]++;
5707 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5708 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5711 key->mv_size = mc->mc_db->md_pad;
5712 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5716 mdb_cassert(mc, IS_LEAF(mp));
5717 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5719 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5720 mdb_xcursor_init1(mc, leaf);
5723 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5726 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5727 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5728 if (rc != MDB_SUCCESS)
5733 MDB_GET_KEY(leaf, key);
5737 /** Move the cursor to the previous data item. */
5739 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5745 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5747 mp = mc->mc_pg[mc->mc_top];
5749 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5750 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5751 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5752 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5753 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5754 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5755 if (rc == MDB_SUCCESS) {
5756 MDB_GET_KEY(leaf, key);
5757 mc->mc_flags &= ~C_EOF;
5763 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5764 if (op == MDB_PREV_DUP)
5765 return MDB_NOTFOUND;
5769 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5770 mdb_dbg_pgno(mp), (void *) mc));
5772 if (mc->mc_ki[mc->mc_top] == 0) {
5773 DPUTS("=====> move to prev sibling page");
5774 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5777 mp = mc->mc_pg[mc->mc_top];
5778 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5779 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5781 mc->mc_ki[mc->mc_top]--;
5783 mc->mc_flags &= ~C_EOF;
5785 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5786 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5789 key->mv_size = mc->mc_db->md_pad;
5790 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5794 mdb_cassert(mc, IS_LEAF(mp));
5795 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5797 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5798 mdb_xcursor_init1(mc, leaf);
5801 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5804 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5805 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5806 if (rc != MDB_SUCCESS)
5811 MDB_GET_KEY(leaf, key);
5815 /** Set the cursor on a specific data item. */
5817 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5818 MDB_cursor_op op, int *exactp)
5822 MDB_node *leaf = NULL;
5825 if (key->mv_size == 0)
5826 return MDB_BAD_VALSIZE;
5829 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5831 /* See if we're already on the right page */
5832 if (mc->mc_flags & C_INITIALIZED) {
5835 mp = mc->mc_pg[mc->mc_top];
5837 mc->mc_ki[mc->mc_top] = 0;
5838 return MDB_NOTFOUND;
5840 if (mp->mp_flags & P_LEAF2) {
5841 nodekey.mv_size = mc->mc_db->md_pad;
5842 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5844 leaf = NODEPTR(mp, 0);
5845 MDB_GET_KEY2(leaf, nodekey);
5847 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5849 /* Probably happens rarely, but first node on the page
5850 * was the one we wanted.
5852 mc->mc_ki[mc->mc_top] = 0;
5859 unsigned int nkeys = NUMKEYS(mp);
5861 if (mp->mp_flags & P_LEAF2) {
5862 nodekey.mv_data = LEAF2KEY(mp,
5863 nkeys-1, nodekey.mv_size);
5865 leaf = NODEPTR(mp, nkeys-1);
5866 MDB_GET_KEY2(leaf, nodekey);
5868 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5870 /* last node was the one we wanted */
5871 mc->mc_ki[mc->mc_top] = nkeys-1;
5877 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5878 /* This is definitely the right page, skip search_page */
5879 if (mp->mp_flags & P_LEAF2) {
5880 nodekey.mv_data = LEAF2KEY(mp,
5881 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5883 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5884 MDB_GET_KEY2(leaf, nodekey);
5886 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5888 /* current node was the one we wanted */
5898 /* If any parents have right-sibs, search.
5899 * Otherwise, there's nothing further.
5901 for (i=0; i<mc->mc_top; i++)
5903 NUMKEYS(mc->mc_pg[i])-1)
5905 if (i == mc->mc_top) {
5906 /* There are no other pages */
5907 mc->mc_ki[mc->mc_top] = nkeys;
5908 return MDB_NOTFOUND;
5912 /* There are no other pages */
5913 mc->mc_ki[mc->mc_top] = 0;
5914 if (op == MDB_SET_RANGE && !exactp) {
5918 return MDB_NOTFOUND;
5922 rc = mdb_page_search(mc, key, 0);
5923 if (rc != MDB_SUCCESS)
5926 mp = mc->mc_pg[mc->mc_top];
5927 mdb_cassert(mc, IS_LEAF(mp));
5930 leaf = mdb_node_search(mc, key, exactp);
5931 if (exactp != NULL && !*exactp) {
5932 /* MDB_SET specified and not an exact match. */
5933 return MDB_NOTFOUND;
5937 DPUTS("===> inexact leaf not found, goto sibling");
5938 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5939 mc->mc_flags |= C_EOF;
5940 return rc; /* no entries matched */
5942 mp = mc->mc_pg[mc->mc_top];
5943 mdb_cassert(mc, IS_LEAF(mp));
5944 leaf = NODEPTR(mp, 0);
5948 mc->mc_flags |= C_INITIALIZED;
5949 mc->mc_flags &= ~C_EOF;
5952 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5953 key->mv_size = mc->mc_db->md_pad;
5954 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5959 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5960 mdb_xcursor_init1(mc, leaf);
5963 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5964 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5965 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5968 if (op == MDB_GET_BOTH) {
5974 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5975 if (rc != MDB_SUCCESS)
5978 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5981 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5983 dcmp = mc->mc_dbx->md_dcmp;
5984 #if UINT_MAX < SIZE_MAX
5985 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5986 dcmp = mdb_cmp_clong;
5988 rc = dcmp(data, &olddata);
5990 if (op == MDB_GET_BOTH || rc > 0)
5991 return MDB_NOTFOUND;
5998 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5999 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6004 /* The key already matches in all other cases */
6005 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6006 MDB_GET_KEY(leaf, key);
6007 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6012 /** Move the cursor to the first item in the database. */
6014 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6020 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6022 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6023 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6024 if (rc != MDB_SUCCESS)
6027 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6029 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6030 mc->mc_flags |= C_INITIALIZED;
6031 mc->mc_flags &= ~C_EOF;
6033 mc->mc_ki[mc->mc_top] = 0;
6035 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6036 key->mv_size = mc->mc_db->md_pad;
6037 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6042 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6043 mdb_xcursor_init1(mc, leaf);
6044 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6048 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6052 MDB_GET_KEY(leaf, key);
6056 /** Move the cursor to the last item in the database. */
6058 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6064 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6066 if (!(mc->mc_flags & C_EOF)) {
6068 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6069 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6070 if (rc != MDB_SUCCESS)
6073 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6076 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6077 mc->mc_flags |= C_INITIALIZED|C_EOF;
6078 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6080 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6081 key->mv_size = mc->mc_db->md_pad;
6082 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6087 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6088 mdb_xcursor_init1(mc, leaf);
6089 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6093 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6098 MDB_GET_KEY(leaf, key);
6103 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6108 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6113 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6117 case MDB_GET_CURRENT:
6118 if (!(mc->mc_flags & C_INITIALIZED)) {
6121 MDB_page *mp = mc->mc_pg[mc->mc_top];
6122 int nkeys = NUMKEYS(mp);
6123 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6124 mc->mc_ki[mc->mc_top] = nkeys;
6130 key->mv_size = mc->mc_db->md_pad;
6131 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6133 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6134 MDB_GET_KEY(leaf, key);
6136 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6137 if (mc->mc_flags & C_DEL)
6138 mdb_xcursor_init1(mc, leaf);
6139 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6141 rc = mdb_node_read(mc->mc_txn, leaf, data);
6148 case MDB_GET_BOTH_RANGE:
6153 if (mc->mc_xcursor == NULL) {
6154 rc = MDB_INCOMPATIBLE;
6164 rc = mdb_cursor_set(mc, key, data, op,
6165 op == MDB_SET_RANGE ? NULL : &exact);
6168 case MDB_GET_MULTIPLE:
6169 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6173 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6174 rc = MDB_INCOMPATIBLE;
6178 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6179 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6182 case MDB_NEXT_MULTIPLE:
6187 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6188 rc = MDB_INCOMPATIBLE;
6191 if (!(mc->mc_flags & C_INITIALIZED))
6192 rc = mdb_cursor_first(mc, key, data);
6194 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6195 if (rc == MDB_SUCCESS) {
6196 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6199 mx = &mc->mc_xcursor->mx_cursor;
6200 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6202 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6203 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6211 case MDB_NEXT_NODUP:
6212 if (!(mc->mc_flags & C_INITIALIZED))
6213 rc = mdb_cursor_first(mc, key, data);
6215 rc = mdb_cursor_next(mc, key, data, op);
6219 case MDB_PREV_NODUP:
6220 if (!(mc->mc_flags & C_INITIALIZED)) {
6221 rc = mdb_cursor_last(mc, key, data);
6224 mc->mc_flags |= C_INITIALIZED;
6225 mc->mc_ki[mc->mc_top]++;
6227 rc = mdb_cursor_prev(mc, key, data, op);
6230 rc = mdb_cursor_first(mc, key, data);
6233 mfunc = mdb_cursor_first;
6235 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6239 if (mc->mc_xcursor == NULL) {
6240 rc = MDB_INCOMPATIBLE;
6244 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6245 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6246 MDB_GET_KEY(leaf, key);
6247 rc = mdb_node_read(mc->mc_txn, leaf, data);
6251 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6255 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6258 rc = mdb_cursor_last(mc, key, data);
6261 mfunc = mdb_cursor_last;
6264 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6269 if (mc->mc_flags & C_DEL)
6270 mc->mc_flags ^= C_DEL;
6275 /** Touch all the pages in the cursor stack. Set mc_top.
6276 * Makes sure all the pages are writable, before attempting a write operation.
6277 * @param[in] mc The cursor to operate on.
6280 mdb_cursor_touch(MDB_cursor *mc)
6282 int rc = MDB_SUCCESS;
6284 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6287 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6289 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6290 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6293 *mc->mc_dbflag |= DB_DIRTY;
6298 rc = mdb_page_touch(mc);
6299 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6300 mc->mc_top = mc->mc_snum-1;
6305 /** Do not spill pages to disk if txn is getting full, may fail instead */
6306 #define MDB_NOSPILL 0x8000
6309 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6313 MDB_node *leaf = NULL;
6316 MDB_val xdata, *rdata, dkey, olddata;
6318 int do_sub = 0, insert_key, insert_data;
6319 unsigned int mcount = 0, dcount = 0, nospill;
6322 unsigned int nflags;
6325 if (mc == NULL || key == NULL)
6328 env = mc->mc_txn->mt_env;
6330 /* Check this first so counter will always be zero on any
6333 if (flags & MDB_MULTIPLE) {
6334 dcount = data[1].mv_size;
6335 data[1].mv_size = 0;
6336 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6337 return MDB_INCOMPATIBLE;
6340 nospill = flags & MDB_NOSPILL;
6341 flags &= ~MDB_NOSPILL;
6343 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6344 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6346 if (key->mv_size-1 >= ENV_MAXKEY(env))
6347 return MDB_BAD_VALSIZE;
6349 #if SIZE_MAX > MAXDATASIZE
6350 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6351 return MDB_BAD_VALSIZE;
6353 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6354 return MDB_BAD_VALSIZE;
6357 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6358 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6362 if (flags == MDB_CURRENT) {
6363 if (!(mc->mc_flags & C_INITIALIZED))
6366 } else if (mc->mc_db->md_root == P_INVALID) {
6367 /* new database, cursor has nothing to point to */
6370 mc->mc_flags &= ~C_INITIALIZED;
6375 if (flags & MDB_APPEND) {
6377 rc = mdb_cursor_last(mc, &k2, &d2);
6379 rc = mc->mc_dbx->md_cmp(key, &k2);
6382 mc->mc_ki[mc->mc_top]++;
6384 /* new key is <= last key */
6389 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6391 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6392 DPRINTF(("duplicate key [%s]", DKEY(key)));
6394 return MDB_KEYEXIST;
6396 if (rc && rc != MDB_NOTFOUND)
6400 if (mc->mc_flags & C_DEL)
6401 mc->mc_flags ^= C_DEL;
6403 /* Cursor is positioned, check for room in the dirty list */
6405 if (flags & MDB_MULTIPLE) {
6407 xdata.mv_size = data->mv_size * dcount;
6411 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6415 if (rc == MDB_NO_ROOT) {
6417 /* new database, write a root leaf page */
6418 DPUTS("allocating new root leaf page");
6419 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6422 mdb_cursor_push(mc, np);
6423 mc->mc_db->md_root = np->mp_pgno;
6424 mc->mc_db->md_depth++;
6425 *mc->mc_dbflag |= DB_DIRTY;
6426 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6428 np->mp_flags |= P_LEAF2;
6429 mc->mc_flags |= C_INITIALIZED;
6431 /* make sure all cursor pages are writable */
6432 rc2 = mdb_cursor_touch(mc);
6437 insert_key = insert_data = rc;
6439 /* The key does not exist */
6440 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6441 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6442 LEAFSIZE(key, data) > env->me_nodemax)
6444 /* Too big for a node, insert in sub-DB. Set up an empty
6445 * "old sub-page" for prep_subDB to expand to a full page.
6447 fp_flags = P_LEAF|P_DIRTY;
6449 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6450 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6451 olddata.mv_size = PAGEHDRSZ;
6455 /* there's only a key anyway, so this is a no-op */
6456 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6458 unsigned int ksize = mc->mc_db->md_pad;
6459 if (key->mv_size != ksize)
6460 return MDB_BAD_VALSIZE;
6461 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6462 memcpy(ptr, key->mv_data, ksize);
6464 /* if overwriting slot 0 of leaf, need to
6465 * update branch key if there is a parent page
6467 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6468 unsigned short top = mc->mc_top;
6470 /* slot 0 is always an empty key, find real slot */
6471 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6473 if (mc->mc_ki[mc->mc_top])
6474 rc2 = mdb_update_key(mc, key);
6485 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6486 olddata.mv_size = NODEDSZ(leaf);
6487 olddata.mv_data = NODEDATA(leaf);
6490 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6491 /* Prepare (sub-)page/sub-DB to accept the new item,
6492 * if needed. fp: old sub-page or a header faking
6493 * it. mp: new (sub-)page. offset: growth in page
6494 * size. xdata: node data with new page or DB.
6496 unsigned i, offset = 0;
6497 mp = fp = xdata.mv_data = env->me_pbuf;
6498 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6500 /* Was a single item before, must convert now */
6501 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6503 /* Just overwrite the current item */
6504 if (flags == MDB_CURRENT)
6506 dcmp = mc->mc_dbx->md_dcmp;
6507 #if UINT_MAX < SIZE_MAX
6508 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6509 dcmp = mdb_cmp_clong;
6511 /* does data match? */
6512 if (!dcmp(data, &olddata)) {
6513 if (flags & MDB_NODUPDATA)
6514 return MDB_KEYEXIST;
6519 /* Back up original data item */
6520 dkey.mv_size = olddata.mv_size;
6521 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6523 /* Make sub-page header for the dup items, with dummy body */
6524 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6525 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6526 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6527 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6528 fp->mp_flags |= P_LEAF2;
6529 fp->mp_pad = data->mv_size;
6530 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6532 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6533 (dkey.mv_size & 1) + (data->mv_size & 1);
6535 fp->mp_upper = xdata.mv_size - PAGEBASE;
6536 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6537 } else if (leaf->mn_flags & F_SUBDATA) {
6538 /* Data is on sub-DB, just store it */
6539 flags |= F_DUPDATA|F_SUBDATA;
6542 /* Data is on sub-page */
6543 fp = olddata.mv_data;
6546 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6547 offset = EVEN(NODESIZE + sizeof(indx_t) +
6551 offset = fp->mp_pad;
6552 if (SIZELEFT(fp) < offset) {
6553 offset *= 4; /* space for 4 more */
6556 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6558 fp->mp_flags |= P_DIRTY;
6559 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6560 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6564 xdata.mv_size = olddata.mv_size + offset;
6567 fp_flags = fp->mp_flags;
6568 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6569 /* Too big for a sub-page, convert to sub-DB */
6570 fp_flags &= ~P_SUBP;
6572 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6573 fp_flags |= P_LEAF2;
6574 dummy.md_pad = fp->mp_pad;
6575 dummy.md_flags = MDB_DUPFIXED;
6576 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6577 dummy.md_flags |= MDB_INTEGERKEY;
6583 dummy.md_branch_pages = 0;
6584 dummy.md_leaf_pages = 1;
6585 dummy.md_overflow_pages = 0;
6586 dummy.md_entries = NUMKEYS(fp);
6587 xdata.mv_size = sizeof(MDB_db);
6588 xdata.mv_data = &dummy;
6589 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6591 offset = env->me_psize - olddata.mv_size;
6592 flags |= F_DUPDATA|F_SUBDATA;
6593 dummy.md_root = mp->mp_pgno;
6596 mp->mp_flags = fp_flags | P_DIRTY;
6597 mp->mp_pad = fp->mp_pad;
6598 mp->mp_lower = fp->mp_lower;
6599 mp->mp_upper = fp->mp_upper + offset;
6600 if (fp_flags & P_LEAF2) {
6601 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6603 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6604 olddata.mv_size - fp->mp_upper - PAGEBASE);
6605 for (i=0; i<NUMKEYS(fp); i++)
6606 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6614 mdb_node_del(mc, 0);
6618 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6619 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6620 return MDB_INCOMPATIBLE;
6621 /* overflow page overwrites need special handling */
6622 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6625 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6627 memcpy(&pg, olddata.mv_data, sizeof(pg));
6628 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6630 ovpages = omp->mp_pages;
6632 /* Is the ov page large enough? */
6633 if (ovpages >= dpages) {
6634 if (!(omp->mp_flags & P_DIRTY) &&
6635 (level || (env->me_flags & MDB_WRITEMAP)))
6637 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6640 level = 0; /* dirty in this txn or clean */
6643 if (omp->mp_flags & P_DIRTY) {
6644 /* yes, overwrite it. Note in this case we don't
6645 * bother to try shrinking the page if the new data
6646 * is smaller than the overflow threshold.
6649 /* It is writable only in a parent txn */
6650 size_t sz = (size_t) env->me_psize * ovpages, off;
6651 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6657 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6658 mdb_cassert(mc, rc2 == 0);
6659 if (!(flags & MDB_RESERVE)) {
6660 /* Copy end of page, adjusting alignment so
6661 * compiler may copy words instead of bytes.
6663 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6664 memcpy((size_t *)((char *)np + off),
6665 (size_t *)((char *)omp + off), sz - off);
6668 memcpy(np, omp, sz); /* Copy beginning of page */
6671 SETDSZ(leaf, data->mv_size);
6672 if (F_ISSET(flags, MDB_RESERVE))
6673 data->mv_data = METADATA(omp);
6675 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6679 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6681 } else if (data->mv_size == olddata.mv_size) {
6682 /* same size, just replace it. Note that we could
6683 * also reuse this node if the new data is smaller,
6684 * but instead we opt to shrink the node in that case.
6686 if (F_ISSET(flags, MDB_RESERVE))
6687 data->mv_data = olddata.mv_data;
6688 else if (!(mc->mc_flags & C_SUB))
6689 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6691 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6696 mdb_node_del(mc, 0);
6702 nflags = flags & NODE_ADD_FLAGS;
6703 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6704 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6705 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6706 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6708 nflags |= MDB_SPLIT_REPLACE;
6709 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6711 /* There is room already in this leaf page. */
6712 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6713 if (rc == 0 && insert_key) {
6714 /* Adjust other cursors pointing to mp */
6715 MDB_cursor *m2, *m3;
6716 MDB_dbi dbi = mc->mc_dbi;
6717 unsigned i = mc->mc_top;
6718 MDB_page *mp = mc->mc_pg[i];
6720 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6721 if (mc->mc_flags & C_SUB)
6722 m3 = &m2->mc_xcursor->mx_cursor;
6725 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6726 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6733 if (rc == MDB_SUCCESS) {
6734 /* Now store the actual data in the child DB. Note that we're
6735 * storing the user data in the keys field, so there are strict
6736 * size limits on dupdata. The actual data fields of the child
6737 * DB are all zero size.
6745 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6746 if (flags & MDB_CURRENT) {
6747 xflags = MDB_CURRENT|MDB_NOSPILL;
6749 mdb_xcursor_init1(mc, leaf);
6750 xflags = (flags & MDB_NODUPDATA) ?
6751 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6753 /* converted, write the original data first */
6755 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6759 /* Adjust other cursors pointing to mp */
6761 unsigned i = mc->mc_top;
6762 MDB_page *mp = mc->mc_pg[i];
6764 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6765 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6766 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6767 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6768 mdb_xcursor_init1(m2, leaf);
6772 /* we've done our job */
6775 ecount = mc->mc_xcursor->mx_db.md_entries;
6776 if (flags & MDB_APPENDDUP)
6777 xflags |= MDB_APPEND;
6778 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6779 if (flags & F_SUBDATA) {
6780 void *db = NODEDATA(leaf);
6781 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6783 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6785 /* Increment count unless we just replaced an existing item. */
6787 mc->mc_db->md_entries++;
6789 /* Invalidate txn if we created an empty sub-DB */
6792 /* If we succeeded and the key didn't exist before,
6793 * make sure the cursor is marked valid.
6795 mc->mc_flags |= C_INITIALIZED;
6797 if (flags & MDB_MULTIPLE) {
6800 /* let caller know how many succeeded, if any */
6801 data[1].mv_size = mcount;
6802 if (mcount < dcount) {
6803 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6804 insert_key = insert_data = 0;
6811 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6814 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6819 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6825 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6826 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6828 if (!(mc->mc_flags & C_INITIALIZED))
6831 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6832 return MDB_NOTFOUND;
6834 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6837 rc = mdb_cursor_touch(mc);
6841 mp = mc->mc_pg[mc->mc_top];
6844 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6846 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6847 if (flags & MDB_NODUPDATA) {
6848 /* mdb_cursor_del0() will subtract the final entry */
6849 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6851 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6852 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6854 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6857 /* If sub-DB still has entries, we're done */
6858 if (mc->mc_xcursor->mx_db.md_entries) {
6859 if (leaf->mn_flags & F_SUBDATA) {
6860 /* update subDB info */
6861 void *db = NODEDATA(leaf);
6862 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6865 /* shrink fake page */
6866 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6867 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6868 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6869 /* fix other sub-DB cursors pointed at this fake page */
6870 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6871 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6872 if (m2->mc_pg[mc->mc_top] == mp &&
6873 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6874 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6877 mc->mc_db->md_entries--;
6878 mc->mc_flags |= C_DEL;
6881 /* otherwise fall thru and delete the sub-DB */
6884 if (leaf->mn_flags & F_SUBDATA) {
6885 /* add all the child DB's pages to the free list */
6886 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6891 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6892 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6893 rc = MDB_INCOMPATIBLE;
6897 /* add overflow pages to free list */
6898 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6902 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6903 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6904 (rc = mdb_ovpage_free(mc, omp)))
6909 return mdb_cursor_del0(mc);
6912 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6916 /** Allocate and initialize new pages for a database.
6917 * @param[in] mc a cursor on the database being added to.
6918 * @param[in] flags flags defining what type of page is being allocated.
6919 * @param[in] num the number of pages to allocate. This is usually 1,
6920 * unless allocating overflow pages for a large record.
6921 * @param[out] mp Address of a page, or NULL on failure.
6922 * @return 0 on success, non-zero on failure.
6925 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6930 if ((rc = mdb_page_alloc(mc, num, &np)))
6932 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6933 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6934 np->mp_flags = flags | P_DIRTY;
6935 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6936 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6939 mc->mc_db->md_branch_pages++;
6940 else if (IS_LEAF(np))
6941 mc->mc_db->md_leaf_pages++;
6942 else if (IS_OVERFLOW(np)) {
6943 mc->mc_db->md_overflow_pages += num;
6951 /** Calculate the size of a leaf node.
6952 * The size depends on the environment's page size; if a data item
6953 * is too large it will be put onto an overflow page and the node
6954 * size will only include the key and not the data. Sizes are always
6955 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6956 * of the #MDB_node headers.
6957 * @param[in] env The environment handle.
6958 * @param[in] key The key for the node.
6959 * @param[in] data The data for the node.
6960 * @return The number of bytes needed to store the node.
6963 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6967 sz = LEAFSIZE(key, data);
6968 if (sz > env->me_nodemax) {
6969 /* put on overflow page */
6970 sz -= data->mv_size - sizeof(pgno_t);
6973 return EVEN(sz + sizeof(indx_t));
6976 /** Calculate the size of a branch node.
6977 * The size should depend on the environment's page size but since
6978 * we currently don't support spilling large keys onto overflow
6979 * pages, it's simply the size of the #MDB_node header plus the
6980 * size of the key. Sizes are always rounded up to an even number
6981 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6982 * @param[in] env The environment handle.
6983 * @param[in] key The key for the node.
6984 * @return The number of bytes needed to store the node.
6987 mdb_branch_size(MDB_env *env, MDB_val *key)
6992 if (sz > env->me_nodemax) {
6993 /* put on overflow page */
6994 /* not implemented */
6995 /* sz -= key->size - sizeof(pgno_t); */
6998 return sz + sizeof(indx_t);
7001 /** Add a node to the page pointed to by the cursor.
7002 * @param[in] mc The cursor for this operation.
7003 * @param[in] indx The index on the page where the new node should be added.
7004 * @param[in] key The key for the new node.
7005 * @param[in] data The data for the new node, if any.
7006 * @param[in] pgno The page number, if adding a branch node.
7007 * @param[in] flags Flags for the node.
7008 * @return 0 on success, non-zero on failure. Possible errors are:
7010 * <li>ENOMEM - failed to allocate overflow pages for the node.
7011 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7012 * should never happen since all callers already calculate the
7013 * page's free space before calling this function.
7017 mdb_node_add(MDB_cursor *mc, indx_t indx,
7018 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7021 size_t node_size = NODESIZE;
7025 MDB_page *mp = mc->mc_pg[mc->mc_top];
7026 MDB_page *ofp = NULL; /* overflow page */
7030 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7032 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7033 IS_LEAF(mp) ? "leaf" : "branch",
7034 IS_SUBP(mp) ? "sub-" : "",
7035 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7036 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7039 /* Move higher keys up one slot. */
7040 int ksize = mc->mc_db->md_pad, dif;
7041 char *ptr = LEAF2KEY(mp, indx, ksize);
7042 dif = NUMKEYS(mp) - indx;
7044 memmove(ptr+ksize, ptr, dif*ksize);
7045 /* insert new key */
7046 memcpy(ptr, key->mv_data, ksize);
7048 /* Just using these for counting */
7049 mp->mp_lower += sizeof(indx_t);
7050 mp->mp_upper -= ksize - sizeof(indx_t);
7054 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7056 node_size += key->mv_size;
7058 mdb_cassert(mc, key && data);
7059 if (F_ISSET(flags, F_BIGDATA)) {
7060 /* Data already on overflow page. */
7061 node_size += sizeof(pgno_t);
7062 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7063 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7065 /* Put data on overflow page. */
7066 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7067 data->mv_size, node_size+data->mv_size));
7068 node_size = EVEN(node_size + sizeof(pgno_t));
7069 if ((ssize_t)node_size > room)
7071 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7073 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7077 node_size += data->mv_size;
7080 node_size = EVEN(node_size);
7081 if ((ssize_t)node_size > room)
7085 /* Move higher pointers up one slot. */
7086 for (i = NUMKEYS(mp); i > indx; i--)
7087 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7089 /* Adjust free space offsets. */
7090 ofs = mp->mp_upper - node_size;
7091 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7092 mp->mp_ptrs[indx] = ofs;
7094 mp->mp_lower += sizeof(indx_t);
7096 /* Write the node data. */
7097 node = NODEPTR(mp, indx);
7098 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7099 node->mn_flags = flags;
7101 SETDSZ(node,data->mv_size);
7106 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7109 ndata = NODEDATA(node);
7111 if (F_ISSET(flags, F_BIGDATA))
7112 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7113 else if (F_ISSET(flags, MDB_RESERVE))
7114 data->mv_data = ndata;
7116 memcpy(ndata, data->mv_data, data->mv_size);
7118 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7119 ndata = METADATA(ofp);
7120 if (F_ISSET(flags, MDB_RESERVE))
7121 data->mv_data = ndata;
7123 memcpy(ndata, data->mv_data, data->mv_size);
7130 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7131 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7132 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7133 DPRINTF(("node size = %"Z"u", node_size));
7134 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7135 return MDB_PAGE_FULL;
7138 /** Delete the specified node from a page.
7139 * @param[in] mc Cursor pointing to the node to delete.
7140 * @param[in] ksize The size of a node. Only used if the page is
7141 * part of a #MDB_DUPFIXED database.
7144 mdb_node_del(MDB_cursor *mc, int ksize)
7146 MDB_page *mp = mc->mc_pg[mc->mc_top];
7147 indx_t indx = mc->mc_ki[mc->mc_top];
7149 indx_t i, j, numkeys, ptr;
7153 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7154 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7155 numkeys = NUMKEYS(mp);
7156 mdb_cassert(mc, indx < numkeys);
7159 int x = numkeys - 1 - indx;
7160 base = LEAF2KEY(mp, indx, ksize);
7162 memmove(base, base + ksize, x * ksize);
7163 mp->mp_lower -= sizeof(indx_t);
7164 mp->mp_upper += ksize - sizeof(indx_t);
7168 node = NODEPTR(mp, indx);
7169 sz = NODESIZE + node->mn_ksize;
7171 if (F_ISSET(node->mn_flags, F_BIGDATA))
7172 sz += sizeof(pgno_t);
7174 sz += NODEDSZ(node);
7178 ptr = mp->mp_ptrs[indx];
7179 for (i = j = 0; i < numkeys; i++) {
7181 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7182 if (mp->mp_ptrs[i] < ptr)
7183 mp->mp_ptrs[j] += sz;
7188 base = (char *)mp + mp->mp_upper + PAGEBASE;
7189 memmove(base + sz, base, ptr - mp->mp_upper);
7191 mp->mp_lower -= sizeof(indx_t);
7195 /** Compact the main page after deleting a node on a subpage.
7196 * @param[in] mp The main page to operate on.
7197 * @param[in] indx The index of the subpage on the main page.
7200 mdb_node_shrink(MDB_page *mp, indx_t indx)
7205 indx_t delta, nsize, len, ptr;
7208 node = NODEPTR(mp, indx);
7209 sp = (MDB_page *)NODEDATA(node);
7210 delta = SIZELEFT(sp);
7211 nsize = NODEDSZ(node) - delta;
7213 /* Prepare to shift upward, set len = length(subpage part to shift) */
7217 return; /* do not make the node uneven-sized */
7219 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7220 for (i = NUMKEYS(sp); --i >= 0; )
7221 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7224 sp->mp_upper = sp->mp_lower;
7225 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7226 SETDSZ(node, nsize);
7228 /* Shift <lower nodes...initial part of subpage> upward */
7229 base = (char *)mp + mp->mp_upper + PAGEBASE;
7230 memmove(base + delta, base, (char *)sp + len - base);
7232 ptr = mp->mp_ptrs[indx];
7233 for (i = NUMKEYS(mp); --i >= 0; ) {
7234 if (mp->mp_ptrs[i] <= ptr)
7235 mp->mp_ptrs[i] += delta;
7237 mp->mp_upper += delta;
7240 /** Initial setup of a sorted-dups cursor.
7241 * Sorted duplicates are implemented as a sub-database for the given key.
7242 * The duplicate data items are actually keys of the sub-database.
7243 * Operations on the duplicate data items are performed using a sub-cursor
7244 * initialized when the sub-database is first accessed. This function does
7245 * the preliminary setup of the sub-cursor, filling in the fields that
7246 * depend only on the parent DB.
7247 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7250 mdb_xcursor_init0(MDB_cursor *mc)
7252 MDB_xcursor *mx = mc->mc_xcursor;
7254 mx->mx_cursor.mc_xcursor = NULL;
7255 mx->mx_cursor.mc_txn = mc->mc_txn;
7256 mx->mx_cursor.mc_db = &mx->mx_db;
7257 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7258 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7259 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7260 mx->mx_cursor.mc_snum = 0;
7261 mx->mx_cursor.mc_top = 0;
7262 mx->mx_cursor.mc_flags = C_SUB;
7263 mx->mx_dbx.md_name.mv_size = 0;
7264 mx->mx_dbx.md_name.mv_data = NULL;
7265 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7266 mx->mx_dbx.md_dcmp = NULL;
7267 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7270 /** Final setup of a sorted-dups cursor.
7271 * Sets up the fields that depend on the data from the main cursor.
7272 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7273 * @param[in] node The data containing the #MDB_db record for the
7274 * sorted-dup database.
7277 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7279 MDB_xcursor *mx = mc->mc_xcursor;
7281 if (node->mn_flags & F_SUBDATA) {
7282 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7283 mx->mx_cursor.mc_pg[0] = 0;
7284 mx->mx_cursor.mc_snum = 0;
7285 mx->mx_cursor.mc_top = 0;
7286 mx->mx_cursor.mc_flags = C_SUB;
7288 MDB_page *fp = NODEDATA(node);
7289 mx->mx_db.md_pad = 0;
7290 mx->mx_db.md_flags = 0;
7291 mx->mx_db.md_depth = 1;
7292 mx->mx_db.md_branch_pages = 0;
7293 mx->mx_db.md_leaf_pages = 1;
7294 mx->mx_db.md_overflow_pages = 0;
7295 mx->mx_db.md_entries = NUMKEYS(fp);
7296 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7297 mx->mx_cursor.mc_snum = 1;
7298 mx->mx_cursor.mc_top = 0;
7299 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7300 mx->mx_cursor.mc_pg[0] = fp;
7301 mx->mx_cursor.mc_ki[0] = 0;
7302 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7303 mx->mx_db.md_flags = MDB_DUPFIXED;
7304 mx->mx_db.md_pad = fp->mp_pad;
7305 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7306 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7309 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7310 mx->mx_db.md_root));
7311 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7312 #if UINT_MAX < SIZE_MAX
7313 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7314 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7318 /** Initialize a cursor for a given transaction and database. */
7320 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7323 mc->mc_backup = NULL;
7326 mc->mc_db = &txn->mt_dbs[dbi];
7327 mc->mc_dbx = &txn->mt_dbxs[dbi];
7328 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7334 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7335 mdb_tassert(txn, mx != NULL);
7336 mc->mc_xcursor = mx;
7337 mdb_xcursor_init0(mc);
7339 mc->mc_xcursor = NULL;
7341 if (*mc->mc_dbflag & DB_STALE) {
7342 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7347 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7350 size_t size = sizeof(MDB_cursor);
7352 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7355 if (txn->mt_flags & MDB_TXN_BLOCKED)
7358 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7361 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7362 size += sizeof(MDB_xcursor);
7364 if ((mc = malloc(size)) != NULL) {
7365 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7366 if (txn->mt_cursors) {
7367 mc->mc_next = txn->mt_cursors[dbi];
7368 txn->mt_cursors[dbi] = mc;
7369 mc->mc_flags |= C_UNTRACK;
7381 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7383 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7386 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7389 if (txn->mt_flags & MDB_TXN_BLOCKED)
7392 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7396 /* Return the count of duplicate data items for the current key */
7398 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7402 if (mc == NULL || countp == NULL)
7405 if (mc->mc_xcursor == NULL)
7406 return MDB_INCOMPATIBLE;
7408 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7411 if (!(mc->mc_flags & C_INITIALIZED))
7414 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7415 return MDB_NOTFOUND;
7417 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7418 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7421 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7424 *countp = mc->mc_xcursor->mx_db.md_entries;
7430 mdb_cursor_close(MDB_cursor *mc)
7432 if (mc && !mc->mc_backup) {
7433 /* remove from txn, if tracked */
7434 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7435 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7436 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7438 *prev = mc->mc_next;
7445 mdb_cursor_txn(MDB_cursor *mc)
7447 if (!mc) return NULL;
7452 mdb_cursor_dbi(MDB_cursor *mc)
7457 /** Replace the key for a branch node with a new key.
7458 * @param[in] mc Cursor pointing to the node to operate on.
7459 * @param[in] key The new key to use.
7460 * @return 0 on success, non-zero on failure.
7463 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7469 int delta, ksize, oksize;
7470 indx_t ptr, i, numkeys, indx;
7473 indx = mc->mc_ki[mc->mc_top];
7474 mp = mc->mc_pg[mc->mc_top];
7475 node = NODEPTR(mp, indx);
7476 ptr = mp->mp_ptrs[indx];
7480 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7481 k2.mv_data = NODEKEY(node);
7482 k2.mv_size = node->mn_ksize;
7483 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7485 mdb_dkey(&k2, kbuf2),
7491 /* Sizes must be 2-byte aligned. */
7492 ksize = EVEN(key->mv_size);
7493 oksize = EVEN(node->mn_ksize);
7494 delta = ksize - oksize;
7496 /* Shift node contents if EVEN(key length) changed. */
7498 if (delta > 0 && SIZELEFT(mp) < delta) {
7500 /* not enough space left, do a delete and split */
7501 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7502 pgno = NODEPGNO(node);
7503 mdb_node_del(mc, 0);
7504 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7507 numkeys = NUMKEYS(mp);
7508 for (i = 0; i < numkeys; i++) {
7509 if (mp->mp_ptrs[i] <= ptr)
7510 mp->mp_ptrs[i] -= delta;
7513 base = (char *)mp + mp->mp_upper + PAGEBASE;
7514 len = ptr - mp->mp_upper + NODESIZE;
7515 memmove(base - delta, base, len);
7516 mp->mp_upper -= delta;
7518 node = NODEPTR(mp, indx);
7521 /* But even if no shift was needed, update ksize */
7522 if (node->mn_ksize != key->mv_size)
7523 node->mn_ksize = key->mv_size;
7526 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7532 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7534 /** Move a node from csrc to cdst.
7537 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7544 unsigned short flags;
7548 /* Mark src and dst as dirty. */
7549 if ((rc = mdb_page_touch(csrc)) ||
7550 (rc = mdb_page_touch(cdst)))
7553 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7554 key.mv_size = csrc->mc_db->md_pad;
7555 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7557 data.mv_data = NULL;
7561 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7562 mdb_cassert(csrc, !((size_t)srcnode & 1));
7563 srcpg = NODEPGNO(srcnode);
7564 flags = srcnode->mn_flags;
7565 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7566 unsigned int snum = csrc->mc_snum;
7568 /* must find the lowest key below src */
7569 rc = mdb_page_search_lowest(csrc);
7572 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7573 key.mv_size = csrc->mc_db->md_pad;
7574 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7576 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7577 key.mv_size = NODEKSZ(s2);
7578 key.mv_data = NODEKEY(s2);
7580 csrc->mc_snum = snum--;
7581 csrc->mc_top = snum;
7583 key.mv_size = NODEKSZ(srcnode);
7584 key.mv_data = NODEKEY(srcnode);
7586 data.mv_size = NODEDSZ(srcnode);
7587 data.mv_data = NODEDATA(srcnode);
7589 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7590 unsigned int snum = cdst->mc_snum;
7593 /* must find the lowest key below dst */
7594 mdb_cursor_copy(cdst, &mn);
7595 rc = mdb_page_search_lowest(&mn);
7598 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7599 bkey.mv_size = mn.mc_db->md_pad;
7600 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7602 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7603 bkey.mv_size = NODEKSZ(s2);
7604 bkey.mv_data = NODEKEY(s2);
7606 mn.mc_snum = snum--;
7609 rc = mdb_update_key(&mn, &bkey);
7614 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7615 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7616 csrc->mc_ki[csrc->mc_top],
7618 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7619 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7621 /* Add the node to the destination page.
7623 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7624 if (rc != MDB_SUCCESS)
7627 /* Delete the node from the source page.
7629 mdb_node_del(csrc, key.mv_size);
7632 /* Adjust other cursors pointing to mp */
7633 MDB_cursor *m2, *m3;
7634 MDB_dbi dbi = csrc->mc_dbi;
7635 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7637 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7638 if (csrc->mc_flags & C_SUB)
7639 m3 = &m2->mc_xcursor->mx_cursor;
7642 if (m3 == csrc) continue;
7643 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7644 csrc->mc_ki[csrc->mc_top]) {
7645 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7646 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7651 /* Update the parent separators.
7653 if (csrc->mc_ki[csrc->mc_top] == 0) {
7654 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7655 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7656 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7658 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7659 key.mv_size = NODEKSZ(srcnode);
7660 key.mv_data = NODEKEY(srcnode);
7662 DPRINTF(("update separator for source page %"Z"u to [%s]",
7663 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7664 mdb_cursor_copy(csrc, &mn);
7667 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7670 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7672 indx_t ix = csrc->mc_ki[csrc->mc_top];
7673 nullkey.mv_size = 0;
7674 csrc->mc_ki[csrc->mc_top] = 0;
7675 rc = mdb_update_key(csrc, &nullkey);
7676 csrc->mc_ki[csrc->mc_top] = ix;
7677 mdb_cassert(csrc, rc == MDB_SUCCESS);
7681 if (cdst->mc_ki[cdst->mc_top] == 0) {
7682 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7683 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7684 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7686 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7687 key.mv_size = NODEKSZ(srcnode);
7688 key.mv_data = NODEKEY(srcnode);
7690 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7691 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7692 mdb_cursor_copy(cdst, &mn);
7695 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7698 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7700 indx_t ix = cdst->mc_ki[cdst->mc_top];
7701 nullkey.mv_size = 0;
7702 cdst->mc_ki[cdst->mc_top] = 0;
7703 rc = mdb_update_key(cdst, &nullkey);
7704 cdst->mc_ki[cdst->mc_top] = ix;
7705 mdb_cassert(cdst, rc == MDB_SUCCESS);
7712 /** Merge one page into another.
7713 * The nodes from the page pointed to by \b csrc will
7714 * be copied to the page pointed to by \b cdst and then
7715 * the \b csrc page will be freed.
7716 * @param[in] csrc Cursor pointing to the source page.
7717 * @param[in] cdst Cursor pointing to the destination page.
7718 * @return 0 on success, non-zero on failure.
7721 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7723 MDB_page *psrc, *pdst;
7730 psrc = csrc->mc_pg[csrc->mc_top];
7731 pdst = cdst->mc_pg[cdst->mc_top];
7733 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7735 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7736 mdb_cassert(csrc, cdst->mc_snum > 1);
7738 /* Mark dst as dirty. */
7739 if ((rc = mdb_page_touch(cdst)))
7742 /* Move all nodes from src to dst.
7744 j = nkeys = NUMKEYS(pdst);
7745 if (IS_LEAF2(psrc)) {
7746 key.mv_size = csrc->mc_db->md_pad;
7747 key.mv_data = METADATA(psrc);
7748 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7749 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7750 if (rc != MDB_SUCCESS)
7752 key.mv_data = (char *)key.mv_data + key.mv_size;
7755 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7756 srcnode = NODEPTR(psrc, i);
7757 if (i == 0 && IS_BRANCH(psrc)) {
7760 mdb_cursor_copy(csrc, &mn);
7761 /* must find the lowest key below src */
7762 rc = mdb_page_search_lowest(&mn);
7765 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7766 key.mv_size = mn.mc_db->md_pad;
7767 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7769 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7770 key.mv_size = NODEKSZ(s2);
7771 key.mv_data = NODEKEY(s2);
7774 key.mv_size = srcnode->mn_ksize;
7775 key.mv_data = NODEKEY(srcnode);
7778 data.mv_size = NODEDSZ(srcnode);
7779 data.mv_data = NODEDATA(srcnode);
7780 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7781 if (rc != MDB_SUCCESS)
7786 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7787 pdst->mp_pgno, NUMKEYS(pdst),
7788 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7790 /* Unlink the src page from parent and add to free list.
7793 mdb_node_del(csrc, 0);
7794 if (csrc->mc_ki[csrc->mc_top] == 0) {
7796 rc = mdb_update_key(csrc, &key);
7804 psrc = csrc->mc_pg[csrc->mc_top];
7805 /* If not operating on FreeDB, allow this page to be reused
7806 * in this txn. Otherwise just add to free list.
7808 rc = mdb_page_loose(csrc, psrc);
7812 csrc->mc_db->md_leaf_pages--;
7814 csrc->mc_db->md_branch_pages--;
7816 /* Adjust other cursors pointing to mp */
7817 MDB_cursor *m2, *m3;
7818 MDB_dbi dbi = csrc->mc_dbi;
7820 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7821 if (csrc->mc_flags & C_SUB)
7822 m3 = &m2->mc_xcursor->mx_cursor;
7825 if (m3 == csrc) continue;
7826 if (m3->mc_snum < csrc->mc_snum) continue;
7827 if (m3->mc_pg[csrc->mc_top] == psrc) {
7828 m3->mc_pg[csrc->mc_top] = pdst;
7829 m3->mc_ki[csrc->mc_top] += nkeys;
7834 unsigned int snum = cdst->mc_snum;
7835 uint16_t depth = cdst->mc_db->md_depth;
7836 mdb_cursor_pop(cdst);
7837 rc = mdb_rebalance(cdst);
7838 /* Did the tree shrink? */
7839 if (depth > cdst->mc_db->md_depth)
7841 cdst->mc_snum = snum;
7842 cdst->mc_top = snum-1;
7847 /** Copy the contents of a cursor.
7848 * @param[in] csrc The cursor to copy from.
7849 * @param[out] cdst The cursor to copy to.
7852 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7856 cdst->mc_txn = csrc->mc_txn;
7857 cdst->mc_dbi = csrc->mc_dbi;
7858 cdst->mc_db = csrc->mc_db;
7859 cdst->mc_dbx = csrc->mc_dbx;
7860 cdst->mc_snum = csrc->mc_snum;
7861 cdst->mc_top = csrc->mc_top;
7862 cdst->mc_flags = csrc->mc_flags;
7864 for (i=0; i<csrc->mc_snum; i++) {
7865 cdst->mc_pg[i] = csrc->mc_pg[i];
7866 cdst->mc_ki[i] = csrc->mc_ki[i];
7870 /** Rebalance the tree after a delete operation.
7871 * @param[in] mc Cursor pointing to the page where rebalancing
7873 * @return 0 on success, non-zero on failure.
7876 mdb_rebalance(MDB_cursor *mc)
7880 unsigned int ptop, minkeys;
7884 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7885 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7886 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7887 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7888 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7890 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7891 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7892 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7893 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7897 if (mc->mc_snum < 2) {
7898 MDB_page *mp = mc->mc_pg[0];
7900 DPUTS("Can't rebalance a subpage, ignoring");
7903 if (NUMKEYS(mp) == 0) {
7904 DPUTS("tree is completely empty");
7905 mc->mc_db->md_root = P_INVALID;
7906 mc->mc_db->md_depth = 0;
7907 mc->mc_db->md_leaf_pages = 0;
7908 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7911 /* Adjust cursors pointing to mp */
7914 mc->mc_flags &= ~C_INITIALIZED;
7916 MDB_cursor *m2, *m3;
7917 MDB_dbi dbi = mc->mc_dbi;
7919 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7920 if (mc->mc_flags & C_SUB)
7921 m3 = &m2->mc_xcursor->mx_cursor;
7924 if (m3->mc_snum < mc->mc_snum) continue;
7925 if (m3->mc_pg[0] == mp) {
7928 m3->mc_flags &= ~C_INITIALIZED;
7932 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7934 DPUTS("collapsing root page!");
7935 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7938 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7939 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7942 mc->mc_db->md_depth--;
7943 mc->mc_db->md_branch_pages--;
7944 mc->mc_ki[0] = mc->mc_ki[1];
7945 for (i = 1; i<mc->mc_db->md_depth; i++) {
7946 mc->mc_pg[i] = mc->mc_pg[i+1];
7947 mc->mc_ki[i] = mc->mc_ki[i+1];
7950 /* Adjust other cursors pointing to mp */
7951 MDB_cursor *m2, *m3;
7952 MDB_dbi dbi = mc->mc_dbi;
7954 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7955 if (mc->mc_flags & C_SUB)
7956 m3 = &m2->mc_xcursor->mx_cursor;
7959 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7960 if (m3->mc_pg[0] == mp) {
7961 for (i=0; i<m3->mc_snum; i++) {
7962 m3->mc_pg[i] = m3->mc_pg[i+1];
7963 m3->mc_ki[i] = m3->mc_ki[i+1];
7971 DPUTS("root page doesn't need rebalancing");
7975 /* The parent (branch page) must have at least 2 pointers,
7976 * otherwise the tree is invalid.
7978 ptop = mc->mc_top-1;
7979 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7981 /* Leaf page fill factor is below the threshold.
7982 * Try to move keys from left or right neighbor, or
7983 * merge with a neighbor page.
7988 mdb_cursor_copy(mc, &mn);
7989 mn.mc_xcursor = NULL;
7991 oldki = mc->mc_ki[mc->mc_top];
7992 if (mc->mc_ki[ptop] == 0) {
7993 /* We're the leftmost leaf in our parent.
7995 DPUTS("reading right neighbor");
7997 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7998 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8001 mn.mc_ki[mn.mc_top] = 0;
8002 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8004 /* There is at least one neighbor to the left.
8006 DPUTS("reading left neighbor");
8008 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8009 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8012 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8013 mc->mc_ki[mc->mc_top] = 0;
8016 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8017 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8018 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8020 /* If the neighbor page is above threshold and has enough keys,
8021 * move one key from it. Otherwise we should try to merge them.
8022 * (A branch page must never have less than 2 keys.)
8024 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
8025 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8026 rc = mdb_node_move(&mn, mc);
8027 if (mc->mc_ki[ptop]) {
8031 if (mc->mc_ki[ptop] == 0) {
8032 rc = mdb_page_merge(&mn, mc);
8035 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8036 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8037 /* We want mdb_rebalance to find mn when doing fixups */
8038 if (mc->mc_flags & C_SUB) {
8039 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8040 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8041 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8043 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8044 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8046 rc = mdb_page_merge(mc, &mn);
8047 if (mc->mc_flags & C_SUB)
8048 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8050 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8051 mdb_cursor_copy(&mn, mc);
8053 mc->mc_flags &= ~C_EOF;
8055 mc->mc_ki[mc->mc_top] = oldki;
8059 /** Complete a delete operation started by #mdb_cursor_del(). */
8061 mdb_cursor_del0(MDB_cursor *mc)
8068 ki = mc->mc_ki[mc->mc_top];
8069 mdb_node_del(mc, mc->mc_db->md_pad);
8070 mc->mc_db->md_entries--;
8071 rc = mdb_rebalance(mc);
8073 if (rc == MDB_SUCCESS) {
8074 MDB_cursor *m2, *m3;
8075 MDB_dbi dbi = mc->mc_dbi;
8077 /* DB is totally empty now, just bail out.
8078 * Other cursors adjustments were already done
8079 * by mdb_rebalance and aren't needed here.
8084 mp = mc->mc_pg[mc->mc_top];
8085 nkeys = NUMKEYS(mp);
8087 /* if mc points past last node in page, find next sibling */
8088 if (mc->mc_ki[mc->mc_top] >= nkeys) {
8089 rc = mdb_cursor_sibling(mc, 1);
8090 if (rc == MDB_NOTFOUND) {
8091 mc->mc_flags |= C_EOF;
8096 /* Adjust other cursors pointing to mp */
8097 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8098 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8099 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8101 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8103 if (m3->mc_pg[mc->mc_top] == mp) {
8104 if (m3->mc_ki[mc->mc_top] >= ki) {
8105 m3->mc_flags |= C_DEL;
8106 if (m3->mc_ki[mc->mc_top] > ki)
8107 m3->mc_ki[mc->mc_top]--;
8108 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8109 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8111 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8112 rc = mdb_cursor_sibling(m3, 1);
8113 if (rc == MDB_NOTFOUND) {
8114 m3->mc_flags |= C_EOF;
8120 mc->mc_flags |= C_DEL;
8124 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8129 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8130 MDB_val *key, MDB_val *data)
8132 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8135 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8136 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8138 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8139 /* must ignore any data */
8143 return mdb_del0(txn, dbi, key, data, 0);
8147 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8148 MDB_val *key, MDB_val *data, unsigned flags)
8153 MDB_val rdata, *xdata;
8157 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8159 mdb_cursor_init(&mc, txn, dbi, &mx);
8168 flags |= MDB_NODUPDATA;
8170 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8172 /* let mdb_page_split know about this cursor if needed:
8173 * delete will trigger a rebalance; if it needs to move
8174 * a node from one page to another, it will have to
8175 * update the parent's separator key(s). If the new sepkey
8176 * is larger than the current one, the parent page may
8177 * run out of space, triggering a split. We need this
8178 * cursor to be consistent until the end of the rebalance.
8180 mc.mc_flags |= C_UNTRACK;
8181 mc.mc_next = txn->mt_cursors[dbi];
8182 txn->mt_cursors[dbi] = &mc;
8183 rc = mdb_cursor_del(&mc, flags);
8184 txn->mt_cursors[dbi] = mc.mc_next;
8189 /** Split a page and insert a new node.
8190 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8191 * The cursor will be updated to point to the actual page and index where
8192 * the node got inserted after the split.
8193 * @param[in] newkey The key for the newly inserted node.
8194 * @param[in] newdata The data for the newly inserted node.
8195 * @param[in] newpgno The page number, if the new node is a branch node.
8196 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8197 * @return 0 on success, non-zero on failure.
8200 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8201 unsigned int nflags)
8204 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8207 int i, j, split_indx, nkeys, pmax;
8208 MDB_env *env = mc->mc_txn->mt_env;
8210 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8211 MDB_page *copy = NULL;
8212 MDB_page *mp, *rp, *pp;
8217 mp = mc->mc_pg[mc->mc_top];
8218 newindx = mc->mc_ki[mc->mc_top];
8219 nkeys = NUMKEYS(mp);
8221 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8222 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8223 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8225 /* Create a right sibling. */
8226 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8228 rp->mp_pad = mp->mp_pad;
8229 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8231 if (mc->mc_snum < 2) {
8232 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8234 /* shift current top to make room for new parent */
8235 mc->mc_pg[1] = mc->mc_pg[0];
8236 mc->mc_ki[1] = mc->mc_ki[0];
8239 mc->mc_db->md_root = pp->mp_pgno;
8240 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8241 mc->mc_db->md_depth++;
8244 /* Add left (implicit) pointer. */
8245 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8246 /* undo the pre-push */
8247 mc->mc_pg[0] = mc->mc_pg[1];
8248 mc->mc_ki[0] = mc->mc_ki[1];
8249 mc->mc_db->md_root = mp->mp_pgno;
8250 mc->mc_db->md_depth--;
8257 ptop = mc->mc_top-1;
8258 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8261 mc->mc_flags |= C_SPLITTING;
8262 mdb_cursor_copy(mc, &mn);
8263 mn.mc_pg[mn.mc_top] = rp;
8264 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8266 if (nflags & MDB_APPEND) {
8267 mn.mc_ki[mn.mc_top] = 0;
8269 split_indx = newindx;
8273 split_indx = (nkeys+1) / 2;
8278 unsigned int lsize, rsize, ksize;
8279 /* Move half of the keys to the right sibling */
8280 x = mc->mc_ki[mc->mc_top] - split_indx;
8281 ksize = mc->mc_db->md_pad;
8282 split = LEAF2KEY(mp, split_indx, ksize);
8283 rsize = (nkeys - split_indx) * ksize;
8284 lsize = (nkeys - split_indx) * sizeof(indx_t);
8285 mp->mp_lower -= lsize;
8286 rp->mp_lower += lsize;
8287 mp->mp_upper += rsize - lsize;
8288 rp->mp_upper -= rsize - lsize;
8289 sepkey.mv_size = ksize;
8290 if (newindx == split_indx) {
8291 sepkey.mv_data = newkey->mv_data;
8293 sepkey.mv_data = split;
8296 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8297 memcpy(rp->mp_ptrs, split, rsize);
8298 sepkey.mv_data = rp->mp_ptrs;
8299 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8300 memcpy(ins, newkey->mv_data, ksize);
8301 mp->mp_lower += sizeof(indx_t);
8302 mp->mp_upper -= ksize - sizeof(indx_t);
8305 memcpy(rp->mp_ptrs, split, x * ksize);
8306 ins = LEAF2KEY(rp, x, ksize);
8307 memcpy(ins, newkey->mv_data, ksize);
8308 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8309 rp->mp_lower += sizeof(indx_t);
8310 rp->mp_upper -= ksize - sizeof(indx_t);
8311 mc->mc_ki[mc->mc_top] = x;
8312 mc->mc_pg[mc->mc_top] = rp;
8315 int psize, nsize, k;
8316 /* Maximum free space in an empty page */
8317 pmax = env->me_psize - PAGEHDRSZ;
8319 nsize = mdb_leaf_size(env, newkey, newdata);
8321 nsize = mdb_branch_size(env, newkey);
8322 nsize = EVEN(nsize);
8324 /* grab a page to hold a temporary copy */
8325 copy = mdb_page_malloc(mc->mc_txn, 1);
8330 copy->mp_pgno = mp->mp_pgno;
8331 copy->mp_flags = mp->mp_flags;
8332 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8333 copy->mp_upper = env->me_psize - PAGEBASE;
8335 /* prepare to insert */
8336 for (i=0, j=0; i<nkeys; i++) {
8338 copy->mp_ptrs[j++] = 0;
8340 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8343 /* When items are relatively large the split point needs
8344 * to be checked, because being off-by-one will make the
8345 * difference between success or failure in mdb_node_add.
8347 * It's also relevant if a page happens to be laid out
8348 * such that one half of its nodes are all "small" and
8349 * the other half of its nodes are "large." If the new
8350 * item is also "large" and falls on the half with
8351 * "large" nodes, it also may not fit.
8353 * As a final tweak, if the new item goes on the last
8354 * spot on the page (and thus, onto the new page), bias
8355 * the split so the new page is emptier than the old page.
8356 * This yields better packing during sequential inserts.
8358 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8359 /* Find split point */
8361 if (newindx <= split_indx || newindx >= nkeys) {
8363 k = newindx >= nkeys ? nkeys : split_indx+2;
8368 for (; i!=k; i+=j) {
8373 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8374 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8376 if (F_ISSET(node->mn_flags, F_BIGDATA))
8377 psize += sizeof(pgno_t);
8379 psize += NODEDSZ(node);
8381 psize = EVEN(psize);
8383 if (psize > pmax || i == k-j) {
8384 split_indx = i + (j<0);
8389 if (split_indx == newindx) {
8390 sepkey.mv_size = newkey->mv_size;
8391 sepkey.mv_data = newkey->mv_data;
8393 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8394 sepkey.mv_size = node->mn_ksize;
8395 sepkey.mv_data = NODEKEY(node);
8400 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8402 /* Copy separator key to the parent.
8404 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8408 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8413 if (mn.mc_snum == mc->mc_snum) {
8414 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8415 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8416 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8417 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8422 /* Right page might now have changed parent.
8423 * Check if left page also changed parent.
8425 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8426 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8427 for (i=0; i<ptop; i++) {
8428 mc->mc_pg[i] = mn.mc_pg[i];
8429 mc->mc_ki[i] = mn.mc_ki[i];
8431 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8432 if (mn.mc_ki[ptop]) {
8433 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8435 /* find right page's left sibling */
8436 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8437 mdb_cursor_sibling(mc, 0);
8442 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8445 mc->mc_flags ^= C_SPLITTING;
8446 if (rc != MDB_SUCCESS) {
8449 if (nflags & MDB_APPEND) {
8450 mc->mc_pg[mc->mc_top] = rp;
8451 mc->mc_ki[mc->mc_top] = 0;
8452 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8455 for (i=0; i<mc->mc_top; i++)
8456 mc->mc_ki[i] = mn.mc_ki[i];
8457 } else if (!IS_LEAF2(mp)) {
8459 mc->mc_pg[mc->mc_top] = rp;
8464 rkey.mv_data = newkey->mv_data;
8465 rkey.mv_size = newkey->mv_size;
8471 /* Update index for the new key. */
8472 mc->mc_ki[mc->mc_top] = j;
8474 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8475 rkey.mv_data = NODEKEY(node);
8476 rkey.mv_size = node->mn_ksize;
8478 xdata.mv_data = NODEDATA(node);
8479 xdata.mv_size = NODEDSZ(node);
8482 pgno = NODEPGNO(node);
8483 flags = node->mn_flags;
8486 if (!IS_LEAF(mp) && j == 0) {
8487 /* First branch index doesn't need key data. */
8491 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8497 mc->mc_pg[mc->mc_top] = copy;
8502 } while (i != split_indx);
8504 nkeys = NUMKEYS(copy);
8505 for (i=0; i<nkeys; i++)
8506 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8507 mp->mp_lower = copy->mp_lower;
8508 mp->mp_upper = copy->mp_upper;
8509 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8510 env->me_psize - copy->mp_upper - PAGEBASE);
8512 /* reset back to original page */
8513 if (newindx < split_indx) {
8514 mc->mc_pg[mc->mc_top] = mp;
8515 if (nflags & MDB_RESERVE) {
8516 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8517 if (!(node->mn_flags & F_BIGDATA))
8518 newdata->mv_data = NODEDATA(node);
8521 mc->mc_pg[mc->mc_top] = rp;
8523 /* Make sure mc_ki is still valid.
8525 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8526 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8527 for (i=0; i<=ptop; i++) {
8528 mc->mc_pg[i] = mn.mc_pg[i];
8529 mc->mc_ki[i] = mn.mc_ki[i];
8536 /* Adjust other cursors pointing to mp */
8537 MDB_cursor *m2, *m3;
8538 MDB_dbi dbi = mc->mc_dbi;
8539 int fixup = NUMKEYS(mp);
8541 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8542 if (mc->mc_flags & C_SUB)
8543 m3 = &m2->mc_xcursor->mx_cursor;
8548 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8550 if (m3->mc_flags & C_SPLITTING)
8555 for (k=m3->mc_top; k>=0; k--) {
8556 m3->mc_ki[k+1] = m3->mc_ki[k];
8557 m3->mc_pg[k+1] = m3->mc_pg[k];
8559 if (m3->mc_ki[0] >= split_indx) {
8564 m3->mc_pg[0] = mc->mc_pg[0];
8568 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8569 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8570 m3->mc_ki[mc->mc_top]++;
8571 if (m3->mc_ki[mc->mc_top] >= fixup) {
8572 m3->mc_pg[mc->mc_top] = rp;
8573 m3->mc_ki[mc->mc_top] -= fixup;
8574 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8576 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8577 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8582 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8585 if (copy) /* tmp page */
8586 mdb_page_free(env, copy);
8588 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8593 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8594 MDB_val *key, MDB_val *data, unsigned int flags)
8599 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8602 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8605 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8606 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8608 mdb_cursor_init(&mc, txn, dbi, &mx);
8609 return mdb_cursor_put(&mc, key, data, flags);
8613 #define MDB_WBUF (1024*1024)
8616 /** State needed for a compacting copy. */
8617 typedef struct mdb_copy {
8618 pthread_mutex_t mc_mutex;
8619 pthread_cond_t mc_cond;
8626 pgno_t mc_next_pgno;
8629 volatile int mc_new;
8634 /** Dedicated writer thread for compacting copy. */
8635 static THREAD_RET ESECT
8636 mdb_env_copythr(void *arg)
8640 int toggle = 0, wsize, rc;
8643 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8646 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8649 pthread_mutex_lock(&my->mc_mutex);
8651 pthread_cond_signal(&my->mc_cond);
8654 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8655 if (my->mc_new < 0) {
8660 wsize = my->mc_wlen[toggle];
8661 ptr = my->mc_wbuf[toggle];
8664 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8668 } else if (len > 0) {
8682 /* If there's an overflow page tail, write it too */
8683 if (my->mc_olen[toggle]) {
8684 wsize = my->mc_olen[toggle];
8685 ptr = my->mc_over[toggle];
8686 my->mc_olen[toggle] = 0;
8689 my->mc_wlen[toggle] = 0;
8691 pthread_cond_signal(&my->mc_cond);
8693 pthread_cond_signal(&my->mc_cond);
8694 pthread_mutex_unlock(&my->mc_mutex);
8695 return (THREAD_RET)0;
8699 /** Tell the writer thread there's a buffer ready to write */
8701 mdb_env_cthr_toggle(mdb_copy *my, int st)
8703 int toggle = my->mc_toggle ^ 1;
8704 pthread_mutex_lock(&my->mc_mutex);
8705 if (my->mc_status) {
8706 pthread_mutex_unlock(&my->mc_mutex);
8707 return my->mc_status;
8709 while (my->mc_new == 1)
8710 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8712 my->mc_toggle = toggle;
8713 pthread_cond_signal(&my->mc_cond);
8714 pthread_mutex_unlock(&my->mc_mutex);
8718 /** Depth-first tree traversal for compacting copy. */
8720 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8723 MDB_txn *txn = my->mc_txn;
8725 MDB_page *mo, *mp, *leaf;
8730 /* Empty DB, nothing to do */
8731 if (*pg == P_INVALID)
8738 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8741 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8745 /* Make cursor pages writable */
8746 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8750 for (i=0; i<mc.mc_top; i++) {
8751 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8752 mc.mc_pg[i] = (MDB_page *)ptr;
8753 ptr += my->mc_env->me_psize;
8756 /* This is writable space for a leaf page. Usually not needed. */
8757 leaf = (MDB_page *)ptr;
8759 toggle = my->mc_toggle;
8760 while (mc.mc_snum > 0) {
8762 mp = mc.mc_pg[mc.mc_top];
8766 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8767 for (i=0; i<n; i++) {
8768 ni = NODEPTR(mp, i);
8769 if (ni->mn_flags & F_BIGDATA) {
8773 /* Need writable leaf */
8775 mc.mc_pg[mc.mc_top] = leaf;
8776 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8778 ni = NODEPTR(mp, i);
8781 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8782 rc = mdb_page_get(txn, pg, &omp, NULL);
8785 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8786 rc = mdb_env_cthr_toggle(my, 1);
8789 toggle = my->mc_toggle;
8791 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8792 memcpy(mo, omp, my->mc_env->me_psize);
8793 mo->mp_pgno = my->mc_next_pgno;
8794 my->mc_next_pgno += omp->mp_pages;
8795 my->mc_wlen[toggle] += my->mc_env->me_psize;
8796 if (omp->mp_pages > 1) {
8797 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8798 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8799 rc = mdb_env_cthr_toggle(my, 1);
8802 toggle = my->mc_toggle;
8804 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8805 } else if (ni->mn_flags & F_SUBDATA) {
8808 /* Need writable leaf */
8810 mc.mc_pg[mc.mc_top] = leaf;
8811 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8813 ni = NODEPTR(mp, i);
8816 memcpy(&db, NODEDATA(ni), sizeof(db));
8817 my->mc_toggle = toggle;
8818 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8821 toggle = my->mc_toggle;
8822 memcpy(NODEDATA(ni), &db, sizeof(db));
8827 mc.mc_ki[mc.mc_top]++;
8828 if (mc.mc_ki[mc.mc_top] < n) {
8831 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8833 rc = mdb_page_get(txn, pg, &mp, NULL);
8838 mc.mc_ki[mc.mc_top] = 0;
8839 if (IS_BRANCH(mp)) {
8840 /* Whenever we advance to a sibling branch page,
8841 * we must proceed all the way down to its first leaf.
8843 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8846 mc.mc_pg[mc.mc_top] = mp;
8850 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8851 rc = mdb_env_cthr_toggle(my, 1);
8854 toggle = my->mc_toggle;
8856 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8857 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8858 mo->mp_pgno = my->mc_next_pgno++;
8859 my->mc_wlen[toggle] += my->mc_env->me_psize;
8861 /* Update parent if there is one */
8862 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8863 SETPGNO(ni, mo->mp_pgno);
8864 mdb_cursor_pop(&mc);
8866 /* Otherwise we're done */
8876 /** Copy environment with compaction. */
8878 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8883 MDB_txn *txn = NULL;
8888 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8889 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8890 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8891 if (my.mc_wbuf[0] == NULL)
8894 pthread_mutex_init(&my.mc_mutex, NULL);
8895 pthread_cond_init(&my.mc_cond, NULL);
8896 #ifdef HAVE_MEMALIGN
8897 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8898 if (my.mc_wbuf[0] == NULL)
8901 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8906 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8907 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8912 my.mc_next_pgno = NUM_METAS;
8918 THREAD_CREATE(thr, mdb_env_copythr, &my);
8920 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8924 mp = (MDB_page *)my.mc_wbuf[0];
8925 memset(mp, 0, NUM_METAS * env->me_psize);
8927 mp->mp_flags = P_META;
8928 mm = (MDB_meta *)METADATA(mp);
8929 mdb_env_init_meta0(env, mm);
8930 mm->mm_address = env->me_metas[0]->mm_address;
8932 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8934 mp->mp_flags = P_META;
8935 *(MDB_meta *)METADATA(mp) = *mm;
8936 mm = (MDB_meta *)METADATA(mp);
8938 /* Count the number of free pages, subtract from lastpg to find
8939 * number of active pages
8942 MDB_ID freecount = 0;
8945 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8946 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8947 freecount += *(MDB_ID *)data.mv_data;
8948 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
8949 txn->mt_dbs[FREE_DBI].md_leaf_pages +
8950 txn->mt_dbs[FREE_DBI].md_overflow_pages;
8952 /* Set metapage 1 */
8953 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8954 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
8955 if (mm->mm_last_pg > NUM_METAS-1) {
8956 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
8959 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
8962 my.mc_wlen[0] = env->me_psize * NUM_METAS;
8964 pthread_mutex_lock(&my.mc_mutex);
8966 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8967 pthread_mutex_unlock(&my.mc_mutex);
8968 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
8969 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8970 rc = mdb_env_cthr_toggle(&my, 1);
8971 mdb_env_cthr_toggle(&my, -1);
8972 pthread_mutex_lock(&my.mc_mutex);
8974 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8975 pthread_mutex_unlock(&my.mc_mutex);
8980 CloseHandle(my.mc_cond);
8981 CloseHandle(my.mc_mutex);
8982 _aligned_free(my.mc_wbuf[0]);
8984 pthread_cond_destroy(&my.mc_cond);
8985 pthread_mutex_destroy(&my.mc_mutex);
8986 free(my.mc_wbuf[0]);
8991 /** Copy environment as-is. */
8993 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8995 MDB_txn *txn = NULL;
8996 mdb_mutexref_t wmutex = NULL;
9002 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9006 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9009 /* Do the lock/unlock of the reader mutex before starting the
9010 * write txn. Otherwise other read txns could block writers.
9012 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9017 /* We must start the actual read txn after blocking writers */
9018 mdb_txn_end(txn, MDB_END_RESET_TMP);
9020 /* Temporarily block writers until we snapshot the meta pages */
9021 wmutex = env->me_wmutex;
9022 if (LOCK_MUTEX(rc, env, wmutex))
9025 rc = mdb_txn_renew0(txn);
9027 UNLOCK_MUTEX(wmutex);
9032 wsize = env->me_psize * NUM_METAS;
9036 DO_WRITE(rc, fd, ptr, w2, len);
9040 } else if (len > 0) {
9046 /* Non-blocking or async handles are not supported */
9052 UNLOCK_MUTEX(wmutex);
9057 w2 = txn->mt_next_pgno * env->me_psize;
9060 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9067 if (wsize > MAX_WRITE)
9071 DO_WRITE(rc, fd, ptr, w2, len);
9075 } else if (len > 0) {
9092 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9094 if (flags & MDB_CP_COMPACT)
9095 return mdb_env_copyfd1(env, fd);
9097 return mdb_env_copyfd0(env, fd);
9101 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9103 return mdb_env_copyfd2(env, fd, 0);
9107 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9111 HANDLE newfd = INVALID_HANDLE_VALUE;
9113 if (env->me_flags & MDB_NOSUBDIR) {
9114 lpath = (char *)path;
9117 len += sizeof(DATANAME);
9118 lpath = malloc(len);
9121 sprintf(lpath, "%s" DATANAME, path);
9124 /* The destination path must exist, but the destination file must not.
9125 * We don't want the OS to cache the writes, since the source data is
9126 * already in the OS cache.
9129 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9130 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9132 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9134 if (newfd == INVALID_HANDLE_VALUE) {
9139 if (env->me_psize >= env->me_os_psize) {
9141 /* Set O_DIRECT if the file system supports it */
9142 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9143 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9145 #ifdef F_NOCACHE /* __APPLE__ */
9146 rc = fcntl(newfd, F_NOCACHE, 1);
9154 rc = mdb_env_copyfd2(env, newfd, flags);
9157 if (!(env->me_flags & MDB_NOSUBDIR))
9159 if (newfd != INVALID_HANDLE_VALUE)
9160 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9167 mdb_env_copy(MDB_env *env, const char *path)
9169 return mdb_env_copy2(env, path, 0);
9173 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9175 if (flag & ~CHANGEABLE)
9178 env->me_flags |= flag;
9180 env->me_flags &= ~flag;
9185 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9190 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9195 mdb_env_set_userctx(MDB_env *env, void *ctx)
9199 env->me_userctx = ctx;
9204 mdb_env_get_userctx(MDB_env *env)
9206 return env ? env->me_userctx : NULL;
9210 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9215 env->me_assert_func = func;
9221 mdb_env_get_path(MDB_env *env, const char **arg)
9226 *arg = env->me_path;
9231 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9240 /** Common code for #mdb_stat() and #mdb_env_stat().
9241 * @param[in] env the environment to operate in.
9242 * @param[in] db the #MDB_db record containing the stats to return.
9243 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9244 * @return 0, this function always succeeds.
9247 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9249 arg->ms_psize = env->me_psize;
9250 arg->ms_depth = db->md_depth;
9251 arg->ms_branch_pages = db->md_branch_pages;
9252 arg->ms_leaf_pages = db->md_leaf_pages;
9253 arg->ms_overflow_pages = db->md_overflow_pages;
9254 arg->ms_entries = db->md_entries;
9260 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9264 if (env == NULL || arg == NULL)
9267 meta = mdb_env_pick_meta(env);
9269 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9273 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9277 if (env == NULL || arg == NULL)
9280 meta = mdb_env_pick_meta(env);
9281 arg->me_mapaddr = meta->mm_address;
9282 arg->me_last_pgno = meta->mm_last_pg;
9283 arg->me_last_txnid = meta->mm_txnid;
9285 arg->me_mapsize = env->me_mapsize;
9286 arg->me_maxreaders = env->me_maxreaders;
9287 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9291 /** Set the default comparison functions for a database.
9292 * Called immediately after a database is opened to set the defaults.
9293 * The user can then override them with #mdb_set_compare() or
9294 * #mdb_set_dupsort().
9295 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9296 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9299 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9301 uint16_t f = txn->mt_dbs[dbi].md_flags;
9303 txn->mt_dbxs[dbi].md_cmp =
9304 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9305 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9307 txn->mt_dbxs[dbi].md_dcmp =
9308 !(f & MDB_DUPSORT) ? 0 :
9309 ((f & MDB_INTEGERDUP)
9310 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9311 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9314 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9320 int rc, dbflag, exact;
9321 unsigned int unused = 0, seq;
9324 if (flags & ~VALID_FLAGS)
9326 if (txn->mt_flags & MDB_TXN_BLOCKED)
9332 if (flags & PERSISTENT_FLAGS) {
9333 uint16_t f2 = flags & PERSISTENT_FLAGS;
9334 /* make sure flag changes get committed */
9335 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9336 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9337 txn->mt_flags |= MDB_TXN_DIRTY;
9340 mdb_default_cmp(txn, MAIN_DBI);
9344 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9345 mdb_default_cmp(txn, MAIN_DBI);
9348 /* Is the DB already open? */
9350 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9351 if (!txn->mt_dbxs[i].md_name.mv_size) {
9352 /* Remember this free slot */
9353 if (!unused) unused = i;
9356 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9357 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9363 /* If no free slot and max hit, fail */
9364 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9365 return MDB_DBS_FULL;
9367 /* Cannot mix named databases with some mainDB flags */
9368 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9369 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9371 /* Find the DB info */
9372 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9375 key.mv_data = (void *)name;
9376 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9377 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9378 if (rc == MDB_SUCCESS) {
9379 /* make sure this is actually a DB */
9380 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9381 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9382 return MDB_INCOMPATIBLE;
9383 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9384 /* Create if requested */
9385 data.mv_size = sizeof(MDB_db);
9386 data.mv_data = &dummy;
9387 memset(&dummy, 0, sizeof(dummy));
9388 dummy.md_root = P_INVALID;
9389 dummy.md_flags = flags & PERSISTENT_FLAGS;
9390 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9394 /* OK, got info, add to table */
9395 if (rc == MDB_SUCCESS) {
9396 unsigned int slot = unused ? unused : txn->mt_numdbs;
9397 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9398 txn->mt_dbxs[slot].md_name.mv_size = len;
9399 txn->mt_dbxs[slot].md_rel = NULL;
9400 txn->mt_dbflags[slot] = dbflag;
9401 /* txn-> and env-> are the same in read txns, use
9402 * tmp variable to avoid undefined assignment
9404 seq = ++txn->mt_env->me_dbiseqs[slot];
9405 txn->mt_dbiseqs[slot] = seq;
9407 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9409 mdb_default_cmp(txn, slot);
9419 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9421 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9424 if (txn->mt_flags & MDB_TXN_BLOCKED)
9427 if (txn->mt_dbflags[dbi] & DB_STALE) {
9430 /* Stale, must read the DB's root. cursor_init does it for us. */
9431 mdb_cursor_init(&mc, txn, dbi, &mx);
9433 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9436 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9439 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9441 ptr = env->me_dbxs[dbi].md_name.mv_data;
9442 /* If there was no name, this was already closed */
9444 env->me_dbxs[dbi].md_name.mv_data = NULL;
9445 env->me_dbxs[dbi].md_name.mv_size = 0;
9446 env->me_dbflags[dbi] = 0;
9447 env->me_dbiseqs[dbi]++;
9452 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9454 /* We could return the flags for the FREE_DBI too but what's the point? */
9455 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9457 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9461 /** Add all the DB's pages to the free list.
9462 * @param[in] mc Cursor on the DB to free.
9463 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9464 * @return 0 on success, non-zero on failure.
9467 mdb_drop0(MDB_cursor *mc, int subs)
9471 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9472 if (rc == MDB_SUCCESS) {
9473 MDB_txn *txn = mc->mc_txn;
9478 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9479 * This also avoids any P_LEAF2 pages, which have no nodes.
9481 if (mc->mc_flags & C_SUB)
9484 mdb_cursor_copy(mc, &mx);
9485 while (mc->mc_snum > 0) {
9486 MDB_page *mp = mc->mc_pg[mc->mc_top];
9487 unsigned n = NUMKEYS(mp);
9489 for (i=0; i<n; i++) {
9490 ni = NODEPTR(mp, i);
9491 if (ni->mn_flags & F_BIGDATA) {
9494 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9495 rc = mdb_page_get(txn, pg, &omp, NULL);
9498 mdb_cassert(mc, IS_OVERFLOW(omp));
9499 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9503 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9504 mdb_xcursor_init1(mc, ni);
9505 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9511 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9513 for (i=0; i<n; i++) {
9515 ni = NODEPTR(mp, i);
9518 mdb_midl_xappend(txn->mt_free_pgs, pg);
9523 mc->mc_ki[mc->mc_top] = i;
9524 rc = mdb_cursor_sibling(mc, 1);
9526 if (rc != MDB_NOTFOUND)
9528 /* no more siblings, go back to beginning
9529 * of previous level.
9533 for (i=1; i<mc->mc_snum; i++) {
9535 mc->mc_pg[i] = mx.mc_pg[i];
9540 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9543 txn->mt_flags |= MDB_TXN_ERROR;
9544 } else if (rc == MDB_NOTFOUND) {
9550 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9552 MDB_cursor *mc, *m2;
9555 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9558 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9561 if (TXN_DBI_CHANGED(txn, dbi))
9564 rc = mdb_cursor_open(txn, dbi, &mc);
9568 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9569 /* Invalidate the dropped DB's cursors */
9570 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9571 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9575 /* Can't delete the main DB */
9576 if (del && dbi >= CORE_DBS) {
9577 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9579 txn->mt_dbflags[dbi] = DB_STALE;
9580 mdb_dbi_close(txn->mt_env, dbi);
9582 txn->mt_flags |= MDB_TXN_ERROR;
9585 /* reset the DB record, mark it dirty */
9586 txn->mt_dbflags[dbi] |= DB_DIRTY;
9587 txn->mt_dbs[dbi].md_depth = 0;
9588 txn->mt_dbs[dbi].md_branch_pages = 0;
9589 txn->mt_dbs[dbi].md_leaf_pages = 0;
9590 txn->mt_dbs[dbi].md_overflow_pages = 0;
9591 txn->mt_dbs[dbi].md_entries = 0;
9592 txn->mt_dbs[dbi].md_root = P_INVALID;
9594 txn->mt_flags |= MDB_TXN_DIRTY;
9597 mdb_cursor_close(mc);
9601 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9603 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9606 txn->mt_dbxs[dbi].md_cmp = cmp;
9610 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9612 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9615 txn->mt_dbxs[dbi].md_dcmp = cmp;
9619 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9621 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9624 txn->mt_dbxs[dbi].md_rel = rel;
9628 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9630 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9633 txn->mt_dbxs[dbi].md_relctx = ctx;
9638 mdb_env_get_maxkeysize(MDB_env *env)
9640 return ENV_MAXKEY(env);
9644 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9646 unsigned int i, rdrs;
9649 int rc = 0, first = 1;
9653 if (!env->me_txns) {
9654 return func("(no reader locks)\n", ctx);
9656 rdrs = env->me_txns->mti_numreaders;
9657 mr = env->me_txns->mti_readers;
9658 for (i=0; i<rdrs; i++) {
9660 txnid_t txnid = mr[i].mr_txnid;
9661 sprintf(buf, txnid == (txnid_t)-1 ?
9662 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9663 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9666 rc = func(" pid thread txnid\n", ctx);
9670 rc = func(buf, ctx);
9676 rc = func("(no active readers)\n", ctx);
9681 /** Insert pid into list if not already present.
9682 * return -1 if already present.
9685 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9687 /* binary search of pid in list */
9689 unsigned cursor = 1;
9691 unsigned n = ids[0];
9694 unsigned pivot = n >> 1;
9695 cursor = base + pivot + 1;
9696 val = pid - ids[cursor];
9701 } else if ( val > 0 ) {
9706 /* found, so it's a duplicate */
9715 for (n = ids[0]; n > cursor; n--)
9722 mdb_reader_check(MDB_env *env, int *dead)
9728 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9731 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9733 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9735 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9736 unsigned int i, j, rdrs;
9738 MDB_PID_T *pids, pid;
9739 int rc = MDB_SUCCESS, count = 0;
9741 rdrs = env->me_txns->mti_numreaders;
9742 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9746 mr = env->me_txns->mti_readers;
9747 for (i=0; i<rdrs; i++) {
9749 if (pid && pid != env->me_pid) {
9750 if (mdb_pid_insert(pids, pid) == 0) {
9751 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9752 /* Stale reader found */
9755 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9756 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9758 rdrs = 0; /* the above checked all readers */
9760 /* Recheck, a new process may have reused pid */
9761 if (mdb_reader_pid(env, Pidcheck, pid))
9766 if (mr[j].mr_pid == pid) {
9767 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9768 (unsigned) pid, mr[j].mr_txnid));
9773 UNLOCK_MUTEX(rmutex);
9784 #ifdef MDB_ROBUST_SUPPORTED
9785 /** Handle #LOCK_MUTEX0() failure.
9786 * Try to repair the lock file if the mutex owner died.
9787 * @param[in] env the environment handle
9788 * @param[in] mutex LOCK_MUTEX0() mutex
9789 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9790 * @return 0 on success with the mutex locked, or an error code on failure.
9793 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9798 if (rc == MDB_OWNERDEAD) {
9799 /* We own the mutex. Clean up after dead previous owner. */
9801 rlocked = (mutex == env->me_rmutex);
9803 /* Keep mti_txnid updated, otherwise next writer can
9804 * overwrite data which latest meta page refers to.
9806 meta = mdb_env_pick_meta(env);
9807 env->me_txns->mti_txnid = meta->mm_txnid;
9808 /* env is hosed if the dead thread was ours */
9810 env->me_flags |= MDB_FATAL_ERROR;
9815 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9816 (rc ? "this process' env is hosed" : "recovering")));
9817 rc2 = mdb_reader_check0(env, rlocked, NULL);
9819 rc2 = mdb_mutex_consistent(mutex);
9820 if (rc || (rc = rc2)) {
9821 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9822 UNLOCK_MUTEX(mutex);
9828 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9833 #endif /* MDB_ROBUST_SUPPORTED */