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;
3790 memset(&ov, 0, sizeof(ov));
3792 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3797 rc = pwrite(mfd, ptr, len, off);
3800 rc = rc < 0 ? ErrCode() : EIO;
3805 DPUTS("write failed, disk error?");
3806 /* On a failure, the pagecache still contains the new data.
3807 * Write some old data back, to prevent it from being used.
3808 * Use the non-SYNC fd; we know it will fail anyway.
3810 meta.mm_last_pg = metab.mm_last_pg;
3811 meta.mm_txnid = metab.mm_txnid;
3813 memset(&ov, 0, sizeof(ov));
3815 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3817 r2 = pwrite(env->me_fd, ptr, len, off);
3818 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3821 env->me_flags |= MDB_FATAL_ERROR;
3824 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3825 CACHEFLUSH(env->me_map + off, len, DCACHE);
3827 /* Memory ordering issues are irrelevant; since the entire writer
3828 * is wrapped by wmutex, all of these changes will become visible
3829 * after the wmutex is unlocked. Since the DB is multi-version,
3830 * readers will get consistent data regardless of how fresh or
3831 * how stale their view of these values is.
3834 env->me_txns->mti_txnid = txn->mt_txnid;
3839 /** Check both meta pages to see which one is newer.
3840 * @param[in] env the environment handle
3841 * @return newest #MDB_meta.
3844 mdb_env_pick_meta(const MDB_env *env)
3846 MDB_meta *const *metas = env->me_metas;
3847 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3851 mdb_env_create(MDB_env **env)
3855 e = calloc(1, sizeof(MDB_env));
3859 e->me_maxreaders = DEFAULT_READERS;
3860 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3861 e->me_fd = INVALID_HANDLE_VALUE;
3862 e->me_lfd = INVALID_HANDLE_VALUE;
3863 e->me_mfd = INVALID_HANDLE_VALUE;
3864 #ifdef MDB_USE_POSIX_SEM
3865 e->me_rmutex = SEM_FAILED;
3866 e->me_wmutex = SEM_FAILED;
3867 #elif defined MDB_USE_SYSV_SEM
3868 e->me_rmutex->semid = -1;
3869 e->me_wmutex->semid = -1;
3871 e->me_pid = getpid();
3872 GET_PAGESIZE(e->me_os_psize);
3873 VGMEMP_CREATE(e,0,0);
3879 mdb_env_map(MDB_env *env, void *addr)
3882 unsigned int flags = env->me_flags;
3886 LONG sizelo, sizehi;
3889 if (flags & MDB_RDONLY) {
3890 /* Don't set explicit map size, use whatever exists */
3895 msize = env->me_mapsize;
3896 sizelo = msize & 0xffffffff;
3897 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3899 /* Windows won't create mappings for zero length files.
3900 * and won't map more than the file size.
3901 * Just set the maxsize right now.
3903 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3904 || !SetEndOfFile(env->me_fd)
3905 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3909 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3910 PAGE_READWRITE : PAGE_READONLY,
3911 sizehi, sizelo, NULL);
3914 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3915 FILE_MAP_WRITE : FILE_MAP_READ,
3917 rc = env->me_map ? 0 : ErrCode();
3922 int prot = PROT_READ;
3923 if (flags & MDB_WRITEMAP) {
3925 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3928 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3930 if (env->me_map == MAP_FAILED) {
3935 if (flags & MDB_NORDAHEAD) {
3936 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3938 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3940 #ifdef POSIX_MADV_RANDOM
3941 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3942 #endif /* POSIX_MADV_RANDOM */
3943 #endif /* MADV_RANDOM */
3947 /* Can happen because the address argument to mmap() is just a
3948 * hint. mmap() can pick another, e.g. if the range is in use.
3949 * The MAP_FIXED flag would prevent that, but then mmap could
3950 * instead unmap existing pages to make room for the new map.
3952 if (addr && env->me_map != addr)
3953 return EBUSY; /* TODO: Make a new MDB_* error code? */
3955 p = (MDB_page *)env->me_map;
3956 env->me_metas[0] = METADATA(p);
3957 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3963 mdb_env_set_mapsize(MDB_env *env, size_t size)
3965 /* If env is already open, caller is responsible for making
3966 * sure there are no active txns.
3974 meta = mdb_env_pick_meta(env);
3976 size = meta->mm_mapsize;
3978 /* Silently round up to minimum if the size is too small */
3979 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3983 munmap(env->me_map, env->me_mapsize);
3984 env->me_mapsize = size;
3985 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3986 rc = mdb_env_map(env, old);
3990 env->me_mapsize = size;
3992 env->me_maxpg = env->me_mapsize / env->me_psize;
3997 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4001 env->me_maxdbs = dbs + CORE_DBS;
4006 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4008 if (env->me_map || readers < 1)
4010 env->me_maxreaders = readers;
4015 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4017 if (!env || !readers)
4019 *readers = env->me_maxreaders;
4024 mdb_fsize(HANDLE fd, size_t *size)
4027 LARGE_INTEGER fsize;
4029 if (!GetFileSizeEx(fd, &fsize))
4032 *size = fsize.QuadPart;
4044 #ifdef BROKEN_FDATASYNC
4045 #include <sys/utsname.h>
4046 #include <sys/vfs.h>
4049 /** Further setup required for opening an LMDB environment
4052 mdb_env_open2(MDB_env *env)
4054 unsigned int flags = env->me_flags;
4055 int i, newenv = 0, rc;
4059 /* See if we should use QueryLimited */
4061 if ((rc & 0xff) > 5)
4062 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4064 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4067 #ifdef BROKEN_FDATASYNC
4068 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4069 * https://lkml.org/lkml/2012/9/3/83
4070 * Kernels after 3.6-rc6 are known good.
4071 * https://lkml.org/lkml/2012/9/10/556
4072 * See if the DB is on ext3/ext4, then check for new enough kernel
4073 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4078 fstatfs(env->me_fd, &st);
4079 while (st.f_type == 0xEF53) {
4083 if (uts.release[0] < '3') {
4084 if (!strncmp(uts.release, "2.6.32.", 7)) {
4085 i = atoi(uts.release+7);
4087 break; /* 2.6.32.60 and newer is OK */
4088 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4089 i = atoi(uts.release+7);
4091 break; /* 2.6.34.15 and newer is OK */
4093 } else if (uts.release[0] == '3') {
4094 i = atoi(uts.release+2);
4096 break; /* 3.6 and newer is OK */
4098 i = atoi(uts.release+4);
4100 break; /* 3.5.4 and newer is OK */
4101 } else if (i == 2) {
4102 i = atoi(uts.release+4);
4104 break; /* 3.2.30 and newer is OK */
4106 } else { /* 4.x and newer is OK */
4109 env->me_flags |= MDB_FSYNCONLY;
4115 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4118 DPUTS("new mdbenv");
4120 env->me_psize = env->me_os_psize;
4121 if (env->me_psize > MAX_PAGESIZE)
4122 env->me_psize = MAX_PAGESIZE;
4123 memset(&meta, 0, sizeof(meta));
4124 mdb_env_init_meta0(env, &meta);
4125 meta.mm_mapsize = DEFAULT_MAPSIZE;
4127 env->me_psize = meta.mm_psize;
4130 /* Was a mapsize configured? */
4131 if (!env->me_mapsize) {
4132 env->me_mapsize = meta.mm_mapsize;
4135 /* Make sure mapsize >= committed data size. Even when using
4136 * mm_mapsize, which could be broken in old files (ITS#7789).
4138 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4139 if (env->me_mapsize < minsize)
4140 env->me_mapsize = minsize;
4142 meta.mm_mapsize = env->me_mapsize;
4144 if (newenv && !(flags & MDB_FIXEDMAP)) {
4145 /* mdb_env_map() may grow the datafile. Write the metapages
4146 * first, so the file will be valid if initialization fails.
4147 * Except with FIXEDMAP, since we do not yet know mm_address.
4148 * We could fill in mm_address later, but then a different
4149 * program might end up doing that - one with a memory layout
4150 * and map address which does not suit the main program.
4152 rc = mdb_env_init_meta(env, &meta);
4158 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4163 if (flags & MDB_FIXEDMAP)
4164 meta.mm_address = env->me_map;
4165 i = mdb_env_init_meta(env, &meta);
4166 if (i != MDB_SUCCESS) {
4171 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4172 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4174 #if !(MDB_MAXKEYSIZE)
4175 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4177 env->me_maxpg = env->me_mapsize / env->me_psize;
4181 MDB_meta *meta = mdb_env_pick_meta(env);
4182 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4184 DPRINTF(("opened database version %u, pagesize %u",
4185 meta->mm_version, env->me_psize));
4186 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4187 DPRINTF(("depth: %u", db->md_depth));
4188 DPRINTF(("entries: %"Z"u", db->md_entries));
4189 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4190 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4191 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4192 DPRINTF(("root: %"Z"u", db->md_root));
4200 /** Release a reader thread's slot in the reader lock table.
4201 * This function is called automatically when a thread exits.
4202 * @param[in] ptr This points to the slot in the reader lock table.
4205 mdb_env_reader_dest(void *ptr)
4207 MDB_reader *reader = ptr;
4213 /** Junk for arranging thread-specific callbacks on Windows. This is
4214 * necessarily platform and compiler-specific. Windows supports up
4215 * to 1088 keys. Let's assume nobody opens more than 64 environments
4216 * in a single process, for now. They can override this if needed.
4218 #ifndef MAX_TLS_KEYS
4219 #define MAX_TLS_KEYS 64
4221 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4222 static int mdb_tls_nkeys;
4224 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4228 case DLL_PROCESS_ATTACH: break;
4229 case DLL_THREAD_ATTACH: break;
4230 case DLL_THREAD_DETACH:
4231 for (i=0; i<mdb_tls_nkeys; i++) {
4232 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4234 mdb_env_reader_dest(r);
4238 case DLL_PROCESS_DETACH: break;
4243 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4245 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4249 /* Force some symbol references.
4250 * _tls_used forces the linker to create the TLS directory if not already done
4251 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4253 #pragma comment(linker, "/INCLUDE:_tls_used")
4254 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4255 #pragma const_seg(".CRT$XLB")
4256 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4257 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4260 #pragma comment(linker, "/INCLUDE:__tls_used")
4261 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4262 #pragma data_seg(".CRT$XLB")
4263 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4265 #endif /* WIN 32/64 */
4266 #endif /* !__GNUC__ */
4269 /** Downgrade the exclusive lock on the region back to shared */
4271 mdb_env_share_locks(MDB_env *env, int *excl)
4274 MDB_meta *meta = mdb_env_pick_meta(env);
4276 env->me_txns->mti_txnid = meta->mm_txnid;
4281 /* First acquire a shared lock. The Unlock will
4282 * then release the existing exclusive lock.
4284 memset(&ov, 0, sizeof(ov));
4285 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4288 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4294 struct flock lock_info;
4295 /* The shared lock replaces the existing lock */
4296 memset((void *)&lock_info, 0, sizeof(lock_info));
4297 lock_info.l_type = F_RDLCK;
4298 lock_info.l_whence = SEEK_SET;
4299 lock_info.l_start = 0;
4300 lock_info.l_len = 1;
4301 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4302 (rc = ErrCode()) == EINTR) ;
4303 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4310 /** Try to get exclusive lock, otherwise shared.
4311 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4314 mdb_env_excl_lock(MDB_env *env, int *excl)
4318 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4322 memset(&ov, 0, sizeof(ov));
4323 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4330 struct flock lock_info;
4331 memset((void *)&lock_info, 0, sizeof(lock_info));
4332 lock_info.l_type = F_WRLCK;
4333 lock_info.l_whence = SEEK_SET;
4334 lock_info.l_start = 0;
4335 lock_info.l_len = 1;
4336 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4337 (rc = ErrCode()) == EINTR) ;
4341 # ifndef MDB_USE_POSIX_MUTEX
4342 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4345 lock_info.l_type = F_RDLCK;
4346 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4347 (rc = ErrCode()) == EINTR) ;
4357 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4359 * @(#) $Revision: 5.1 $
4360 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4361 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4363 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4367 * Please do not copyright this code. This code is in the public domain.
4369 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4370 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4371 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4372 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4373 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4374 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4375 * PERFORMANCE OF THIS SOFTWARE.
4378 * chongo <Landon Curt Noll> /\oo/\
4379 * http://www.isthe.com/chongo/
4381 * Share and Enjoy! :-)
4384 typedef unsigned long long mdb_hash_t;
4385 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4387 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4388 * @param[in] val value to hash
4389 * @param[in] hval initial value for hash
4390 * @return 64 bit hash
4392 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4393 * hval arg on the first call.
4396 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4398 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4399 unsigned char *end = s + val->mv_size;
4401 * FNV-1a hash each octet of the string
4404 /* xor the bottom with the current octet */
4405 hval ^= (mdb_hash_t)*s++;
4407 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4408 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4409 (hval << 7) + (hval << 8) + (hval << 40);
4411 /* return our new hash value */
4415 /** Hash the string and output the encoded hash.
4416 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4417 * very short name limits. We don't care about the encoding being reversible,
4418 * we just want to preserve as many bits of the input as possible in a
4419 * small printable string.
4420 * @param[in] str string to hash
4421 * @param[out] encbuf an array of 11 chars to hold the hash
4423 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4426 mdb_pack85(unsigned long l, char *out)
4430 for (i=0; i<5; i++) {
4431 *out++ = mdb_a85[l % 85];
4437 mdb_hash_enc(MDB_val *val, char *encbuf)
4439 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4441 mdb_pack85(h, encbuf);
4442 mdb_pack85(h>>32, encbuf+5);
4447 /** Open and/or initialize the lock region for the environment.
4448 * @param[in] env The LMDB environment.
4449 * @param[in] lpath The pathname of the file used for the lock region.
4450 * @param[in] mode The Unix permissions for the file, if we create it.
4451 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4452 * @return 0 on success, non-zero on failure.
4455 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4458 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4460 # define MDB_ERRCODE_ROFS EROFS
4461 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4462 # define MDB_CLOEXEC O_CLOEXEC
4465 # define MDB_CLOEXEC 0
4468 #ifdef MDB_USE_SYSV_SEM
4476 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4477 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4478 FILE_ATTRIBUTE_NORMAL, NULL);
4480 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4482 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4484 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4489 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4490 /* Lose record locks when exec*() */
4491 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4492 fcntl(env->me_lfd, F_SETFD, fdflags);
4495 if (!(env->me_flags & MDB_NOTLS)) {
4496 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4499 env->me_flags |= MDB_ENV_TXKEY;
4501 /* Windows TLS callbacks need help finding their TLS info. */
4502 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4506 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4510 /* Try to get exclusive lock. If we succeed, then
4511 * nobody is using the lock region and we should initialize it.
4513 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4516 size = GetFileSize(env->me_lfd, NULL);
4518 size = lseek(env->me_lfd, 0, SEEK_END);
4519 if (size == -1) goto fail_errno;
4521 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4522 if (size < rsize && *excl > 0) {
4524 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4525 || !SetEndOfFile(env->me_lfd))
4528 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4532 size = rsize - sizeof(MDB_txninfo);
4533 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4538 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4540 if (!mh) goto fail_errno;
4541 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4543 if (!env->me_txns) goto fail_errno;
4545 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4547 if (m == MAP_FAILED) goto fail_errno;
4553 BY_HANDLE_FILE_INFORMATION stbuf;
4562 if (!mdb_sec_inited) {
4563 InitializeSecurityDescriptor(&mdb_null_sd,
4564 SECURITY_DESCRIPTOR_REVISION);
4565 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4566 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4567 mdb_all_sa.bInheritHandle = FALSE;
4568 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4571 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4572 idbuf.volume = stbuf.dwVolumeSerialNumber;
4573 idbuf.nhigh = stbuf.nFileIndexHigh;
4574 idbuf.nlow = stbuf.nFileIndexLow;
4575 val.mv_data = &idbuf;
4576 val.mv_size = sizeof(idbuf);
4577 mdb_hash_enc(&val, encbuf);
4578 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4579 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4580 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4581 if (!env->me_rmutex) goto fail_errno;
4582 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4583 if (!env->me_wmutex) goto fail_errno;
4584 #elif defined(MDB_USE_POSIX_SEM)
4593 #if defined(__NetBSD__)
4594 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4596 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4597 idbuf.dev = stbuf.st_dev;
4598 idbuf.ino = stbuf.st_ino;
4599 val.mv_data = &idbuf;
4600 val.mv_size = sizeof(idbuf);
4601 mdb_hash_enc(&val, encbuf);
4602 #ifdef MDB_SHORT_SEMNAMES
4603 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4605 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4606 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4607 /* Clean up after a previous run, if needed: Try to
4608 * remove both semaphores before doing anything else.
4610 sem_unlink(env->me_txns->mti_rmname);
4611 sem_unlink(env->me_txns->mti_wmname);
4612 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4613 O_CREAT|O_EXCL, mode, 1);
4614 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4615 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4616 O_CREAT|O_EXCL, mode, 1);
4617 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4618 #elif defined(MDB_USE_SYSV_SEM)
4619 unsigned short vals[2] = {1, 1};
4620 key_t key = ftok(lpath, 'M');
4623 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4627 if (semctl(semid, 0, SETALL, semu) < 0)
4629 env->me_txns->mti_semid = semid;
4630 #else /* MDB_USE_POSIX_MUTEX: */
4631 pthread_mutexattr_t mattr;
4633 if ((rc = pthread_mutexattr_init(&mattr))
4634 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4635 #ifdef MDB_ROBUST_SUPPORTED
4636 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4638 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4639 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4641 pthread_mutexattr_destroy(&mattr);
4642 #endif /* _WIN32 || ... */
4644 env->me_txns->mti_magic = MDB_MAGIC;
4645 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4646 env->me_txns->mti_txnid = 0;
4647 env->me_txns->mti_numreaders = 0;
4650 #ifdef MDB_USE_SYSV_SEM
4651 struct semid_ds buf;
4653 if (env->me_txns->mti_magic != MDB_MAGIC) {
4654 DPUTS("lock region has invalid magic");
4658 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4659 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4660 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4661 rc = MDB_VERSION_MISMATCH;
4665 if (rc && rc != EACCES && rc != EAGAIN) {
4669 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4670 if (!env->me_rmutex) goto fail_errno;
4671 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4672 if (!env->me_wmutex) goto fail_errno;
4673 #elif defined(MDB_USE_POSIX_SEM)
4674 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4675 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4676 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4677 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4678 #elif defined(MDB_USE_SYSV_SEM)
4679 semid = env->me_txns->mti_semid;
4681 /* check for read access */
4682 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4684 /* check for write access */
4685 if (semctl(semid, 0, IPC_SET, semu) < 0)
4689 #ifdef MDB_USE_SYSV_SEM
4690 env->me_rmutex->semid = semid;
4691 env->me_wmutex->semid = semid;
4692 env->me_rmutex->semnum = 0;
4693 env->me_wmutex->semnum = 1;
4694 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4695 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4706 /** The name of the lock file in the DB environment */
4707 #define LOCKNAME "/lock.mdb"
4708 /** The name of the data file in the DB environment */
4709 #define DATANAME "/data.mdb"
4710 /** The suffix of the lock file when no subdir is used */
4711 #define LOCKSUFF "-lock"
4712 /** Only a subset of the @ref mdb_env flags can be changed
4713 * at runtime. Changing other flags requires closing the
4714 * environment and re-opening it with the new flags.
4716 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4717 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4718 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4720 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4721 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4725 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4727 int oflags, rc, len, excl = -1;
4728 char *lpath, *dpath;
4730 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4734 if (flags & MDB_NOSUBDIR) {
4735 rc = len + sizeof(LOCKSUFF) + len + 1;
4737 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4742 if (flags & MDB_NOSUBDIR) {
4743 dpath = lpath + len + sizeof(LOCKSUFF);
4744 sprintf(lpath, "%s" LOCKSUFF, path);
4745 strcpy(dpath, path);
4747 dpath = lpath + len + sizeof(LOCKNAME);
4748 sprintf(lpath, "%s" LOCKNAME, path);
4749 sprintf(dpath, "%s" DATANAME, path);
4753 flags |= env->me_flags;
4754 if (flags & MDB_RDONLY) {
4755 /* silently ignore WRITEMAP when we're only getting read access */
4756 flags &= ~MDB_WRITEMAP;
4758 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4759 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4762 env->me_flags = flags |= MDB_ENV_ACTIVE;
4766 env->me_path = strdup(path);
4767 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4768 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4769 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4770 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4774 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4776 /* For RDONLY, get lockfile after we know datafile exists */
4777 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4778 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4784 if (F_ISSET(flags, MDB_RDONLY)) {
4785 oflags = GENERIC_READ;
4786 len = OPEN_EXISTING;
4788 oflags = GENERIC_READ|GENERIC_WRITE;
4791 mode = FILE_ATTRIBUTE_NORMAL;
4792 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4793 NULL, len, mode, NULL);
4795 if (F_ISSET(flags, MDB_RDONLY))
4798 oflags = O_RDWR | O_CREAT;
4800 env->me_fd = open(dpath, oflags, mode);
4802 if (env->me_fd == INVALID_HANDLE_VALUE) {
4807 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4808 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4813 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4814 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4815 env->me_mfd = env->me_fd;
4817 /* Synchronous fd for meta writes. Needed even with
4818 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4821 len = OPEN_EXISTING;
4822 env->me_mfd = CreateFile(dpath, oflags,
4823 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4824 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4827 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4829 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4834 DPRINTF(("opened dbenv %p", (void *) env));
4836 rc = mdb_env_share_locks(env, &excl);
4840 if (!(flags & MDB_RDONLY)) {
4842 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4843 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4844 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4845 (txn = calloc(1, size)))
4847 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4848 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4849 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4850 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4852 txn->mt_dbxs = env->me_dbxs;
4853 txn->mt_flags = MDB_TXN_FINISHED;
4863 mdb_env_close0(env, excl);
4869 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4871 mdb_env_close0(MDB_env *env, int excl)
4875 if (!(env->me_flags & MDB_ENV_ACTIVE))
4878 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4880 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4881 free(env->me_dbxs[i].md_name.mv_data);
4886 free(env->me_dbiseqs);
4887 free(env->me_dbflags);
4889 free(env->me_dirty_list);
4891 mdb_midl_free(env->me_free_pgs);
4893 if (env->me_flags & MDB_ENV_TXKEY) {
4894 pthread_key_delete(env->me_txkey);
4896 /* Delete our key from the global list */
4897 for (i=0; i<mdb_tls_nkeys; i++)
4898 if (mdb_tls_keys[i] == env->me_txkey) {
4899 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4907 munmap(env->me_map, env->me_mapsize);
4909 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4910 (void) close(env->me_mfd);
4911 if (env->me_fd != INVALID_HANDLE_VALUE)
4912 (void) close(env->me_fd);
4914 MDB_PID_T pid = env->me_pid;
4915 /* Clearing readers is done in this function because
4916 * me_txkey with its destructor must be disabled first.
4918 * We skip the the reader mutex, so we touch only
4919 * data owned by this process (me_close_readers and
4920 * our readers), and clear each reader atomically.
4922 for (i = env->me_close_readers; --i >= 0; )
4923 if (env->me_txns->mti_readers[i].mr_pid == pid)
4924 env->me_txns->mti_readers[i].mr_pid = 0;
4926 if (env->me_rmutex) {
4927 CloseHandle(env->me_rmutex);
4928 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4930 /* Windows automatically destroys the mutexes when
4931 * the last handle closes.
4933 #elif defined(MDB_USE_POSIX_SEM)
4934 if (env->me_rmutex != SEM_FAILED) {
4935 sem_close(env->me_rmutex);
4936 if (env->me_wmutex != SEM_FAILED)
4937 sem_close(env->me_wmutex);
4938 /* If we have the filelock: If we are the
4939 * only remaining user, clean up semaphores.
4942 mdb_env_excl_lock(env, &excl);
4944 sem_unlink(env->me_txns->mti_rmname);
4945 sem_unlink(env->me_txns->mti_wmname);
4948 #elif defined(MDB_USE_SYSV_SEM)
4949 if (env->me_rmutex->semid != -1) {
4950 /* If we have the filelock: If we are the
4951 * only remaining user, clean up semaphores.
4954 mdb_env_excl_lock(env, &excl);
4956 semctl(env->me_rmutex->semid, 0, IPC_RMID);
4959 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4961 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4964 /* Unlock the lockfile. Windows would have unlocked it
4965 * after closing anyway, but not necessarily at once.
4967 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4970 (void) close(env->me_lfd);
4973 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4977 mdb_env_close(MDB_env *env)
4984 VGMEMP_DESTROY(env);
4985 while ((dp = env->me_dpages) != NULL) {
4986 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4987 env->me_dpages = dp->mp_next;
4991 mdb_env_close0(env, 0);
4995 /** Compare two items pointing at aligned size_t's */
4997 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4999 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5000 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5003 /** Compare two items pointing at aligned unsigned int's.
5005 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5006 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5009 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5011 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5012 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5015 /** Compare two items pointing at unsigned ints of unknown alignment.
5016 * Nodes and keys are guaranteed to be 2-byte aligned.
5019 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5021 #if BYTE_ORDER == LITTLE_ENDIAN
5022 unsigned short *u, *c;
5025 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5026 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5029 } while(!x && u > (unsigned short *)a->mv_data);
5032 unsigned short *u, *c, *end;
5035 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5036 u = (unsigned short *)a->mv_data;
5037 c = (unsigned short *)b->mv_data;
5040 } while(!x && u < end);
5045 /** Compare two items lexically */
5047 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5054 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5060 diff = memcmp(a->mv_data, b->mv_data, len);
5061 return diff ? diff : len_diff<0 ? -1 : len_diff;
5064 /** Compare two items in reverse byte order */
5066 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5068 const unsigned char *p1, *p2, *p1_lim;
5072 p1_lim = (const unsigned char *)a->mv_data;
5073 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5074 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5076 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5082 while (p1 > p1_lim) {
5083 diff = *--p1 - *--p2;
5087 return len_diff<0 ? -1 : len_diff;
5090 /** Search for key within a page, using binary search.
5091 * Returns the smallest entry larger or equal to the key.
5092 * If exactp is non-null, stores whether the found entry was an exact match
5093 * in *exactp (1 or 0).
5094 * Updates the cursor index with the index of the found entry.
5095 * If no entry larger or equal to the key is found, returns NULL.
5098 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5100 unsigned int i = 0, nkeys;
5103 MDB_page *mp = mc->mc_pg[mc->mc_top];
5104 MDB_node *node = NULL;
5109 nkeys = NUMKEYS(mp);
5111 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5112 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5115 low = IS_LEAF(mp) ? 0 : 1;
5117 cmp = mc->mc_dbx->md_cmp;
5119 /* Branch pages have no data, so if using integer keys,
5120 * alignment is guaranteed. Use faster mdb_cmp_int.
5122 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5123 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5130 nodekey.mv_size = mc->mc_db->md_pad;
5131 node = NODEPTR(mp, 0); /* fake */
5132 while (low <= high) {
5133 i = (low + high) >> 1;
5134 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5135 rc = cmp(key, &nodekey);
5136 DPRINTF(("found leaf index %u [%s], rc = %i",
5137 i, DKEY(&nodekey), rc));
5146 while (low <= high) {
5147 i = (low + high) >> 1;
5149 node = NODEPTR(mp, i);
5150 nodekey.mv_size = NODEKSZ(node);
5151 nodekey.mv_data = NODEKEY(node);
5153 rc = cmp(key, &nodekey);
5156 DPRINTF(("found leaf index %u [%s], rc = %i",
5157 i, DKEY(&nodekey), rc));
5159 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5160 i, DKEY(&nodekey), NODEPGNO(node), rc));
5171 if (rc > 0) { /* Found entry is less than the key. */
5172 i++; /* Skip to get the smallest entry larger than key. */
5174 node = NODEPTR(mp, i);
5177 *exactp = (rc == 0 && nkeys > 0);
5178 /* store the key index */
5179 mc->mc_ki[mc->mc_top] = i;
5181 /* There is no entry larger or equal to the key. */
5184 /* nodeptr is fake for LEAF2 */
5190 mdb_cursor_adjust(MDB_cursor *mc, func)
5194 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5195 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5202 /** Pop a page off the top of the cursor's stack. */
5204 mdb_cursor_pop(MDB_cursor *mc)
5207 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5208 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5216 /** Push a page onto the top of the cursor's stack. */
5218 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5220 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5221 DDBI(mc), (void *) mc));
5223 if (mc->mc_snum >= CURSOR_STACK) {
5224 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5225 return MDB_CURSOR_FULL;
5228 mc->mc_top = mc->mc_snum++;
5229 mc->mc_pg[mc->mc_top] = mp;
5230 mc->mc_ki[mc->mc_top] = 0;
5235 /** Find the address of the page corresponding to a given page number.
5236 * @param[in] txn the transaction for this access.
5237 * @param[in] pgno the page number for the page to retrieve.
5238 * @param[out] ret address of a pointer where the page's address will be stored.
5239 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5240 * @return 0 on success, non-zero on failure.
5243 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5245 MDB_env *env = txn->mt_env;
5249 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5253 MDB_ID2L dl = tx2->mt_u.dirty_list;
5255 /* Spilled pages were dirtied in this txn and flushed
5256 * because the dirty list got full. Bring this page
5257 * back in from the map (but don't unspill it here,
5258 * leave that unless page_touch happens again).
5260 if (tx2->mt_spill_pgs) {
5261 MDB_ID pn = pgno << 1;
5262 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5263 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5264 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5269 unsigned x = mdb_mid2l_search(dl, pgno);
5270 if (x <= dl[0].mid && dl[x].mid == pgno) {
5276 } while ((tx2 = tx2->mt_parent) != NULL);
5279 if (pgno < txn->mt_next_pgno) {
5281 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5283 DPRINTF(("page %"Z"u not found", pgno));
5284 txn->mt_flags |= MDB_TXN_ERROR;
5285 return MDB_PAGE_NOTFOUND;
5295 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5296 * The cursor is at the root page, set up the rest of it.
5299 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5301 MDB_page *mp = mc->mc_pg[mc->mc_top];
5305 while (IS_BRANCH(mp)) {
5309 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5310 mdb_cassert(mc, NUMKEYS(mp) > 1);
5311 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5313 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5315 if (flags & MDB_PS_LAST)
5316 i = NUMKEYS(mp) - 1;
5319 node = mdb_node_search(mc, key, &exact);
5321 i = NUMKEYS(mp) - 1;
5323 i = mc->mc_ki[mc->mc_top];
5325 mdb_cassert(mc, i > 0);
5329 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5332 mdb_cassert(mc, i < NUMKEYS(mp));
5333 node = NODEPTR(mp, i);
5335 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5338 mc->mc_ki[mc->mc_top] = i;
5339 if ((rc = mdb_cursor_push(mc, mp)))
5342 if (flags & MDB_PS_MODIFY) {
5343 if ((rc = mdb_page_touch(mc)) != 0)
5345 mp = mc->mc_pg[mc->mc_top];
5350 DPRINTF(("internal error, index points to a %02X page!?",
5352 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5353 return MDB_CORRUPTED;
5356 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5357 key ? DKEY(key) : "null"));
5358 mc->mc_flags |= C_INITIALIZED;
5359 mc->mc_flags &= ~C_EOF;
5364 /** Search for the lowest key under the current branch page.
5365 * This just bypasses a NUMKEYS check in the current page
5366 * before calling mdb_page_search_root(), because the callers
5367 * are all in situations where the current page is known to
5371 mdb_page_search_lowest(MDB_cursor *mc)
5373 MDB_page *mp = mc->mc_pg[mc->mc_top];
5374 MDB_node *node = NODEPTR(mp, 0);
5377 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5380 mc->mc_ki[mc->mc_top] = 0;
5381 if ((rc = mdb_cursor_push(mc, mp)))
5383 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5386 /** Search for the page a given key should be in.
5387 * Push it and its parent pages on the cursor stack.
5388 * @param[in,out] mc the cursor for this operation.
5389 * @param[in] key the key to search for, or NULL for first/last page.
5390 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5391 * are touched (updated with new page numbers).
5392 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5393 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5394 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5395 * @return 0 on success, non-zero on failure.
5398 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5403 /* Make sure the txn is still viable, then find the root from
5404 * the txn's db table and set it as the root of the cursor's stack.
5406 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5407 DPUTS("transaction may not be used now");
5410 /* Make sure we're using an up-to-date root */
5411 if (*mc->mc_dbflag & DB_STALE) {
5413 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5415 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5416 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5423 MDB_node *leaf = mdb_node_search(&mc2,
5424 &mc->mc_dbx->md_name, &exact);
5426 return MDB_NOTFOUND;
5427 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5428 return MDB_INCOMPATIBLE; /* not a named DB */
5429 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5432 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5434 /* The txn may not know this DBI, or another process may
5435 * have dropped and recreated the DB with other flags.
5437 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5438 return MDB_INCOMPATIBLE;
5439 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5441 *mc->mc_dbflag &= ~DB_STALE;
5443 root = mc->mc_db->md_root;
5445 if (root == P_INVALID) { /* Tree is empty. */
5446 DPUTS("tree is empty");
5447 return MDB_NOTFOUND;
5451 mdb_cassert(mc, root > 1);
5452 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5453 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5459 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5460 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5462 if (flags & MDB_PS_MODIFY) {
5463 if ((rc = mdb_page_touch(mc)))
5467 if (flags & MDB_PS_ROOTONLY)
5470 return mdb_page_search_root(mc, key, flags);
5474 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5476 MDB_txn *txn = mc->mc_txn;
5477 pgno_t pg = mp->mp_pgno;
5478 unsigned x = 0, ovpages = mp->mp_pages;
5479 MDB_env *env = txn->mt_env;
5480 MDB_IDL sl = txn->mt_spill_pgs;
5481 MDB_ID pn = pg << 1;
5484 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5485 /* If the page is dirty or on the spill list we just acquired it,
5486 * so we should give it back to our current free list, if any.
5487 * Otherwise put it onto the list of pages we freed in this txn.
5489 * Won't create me_pghead: me_pglast must be inited along with it.
5490 * Unsupported in nested txns: They would need to hide the page
5491 * range in ancestor txns' dirty and spilled lists.
5493 if (env->me_pghead &&
5495 ((mp->mp_flags & P_DIRTY) ||
5496 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5500 MDB_ID2 *dl, ix, iy;
5501 rc = mdb_midl_need(&env->me_pghead, ovpages);
5504 if (!(mp->mp_flags & P_DIRTY)) {
5505 /* This page is no longer spilled */
5512 /* Remove from dirty list */
5513 dl = txn->mt_u.dirty_list;
5515 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5521 mdb_cassert(mc, x > 1);
5523 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5524 txn->mt_flags |= MDB_TXN_ERROR;
5525 return MDB_CORRUPTED;
5528 if (!(env->me_flags & MDB_WRITEMAP))
5529 mdb_dpage_free(env, mp);
5531 /* Insert in me_pghead */
5532 mop = env->me_pghead;
5533 j = mop[0] + ovpages;
5534 for (i = mop[0]; i && mop[i] < pg; i--)
5540 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5544 mc->mc_db->md_overflow_pages -= ovpages;
5548 /** Return the data associated with a given node.
5549 * @param[in] txn The transaction for this operation.
5550 * @param[in] leaf The node being read.
5551 * @param[out] data Updated to point to the node's data.
5552 * @return 0 on success, non-zero on failure.
5555 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5557 MDB_page *omp; /* overflow page */
5561 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5562 data->mv_size = NODEDSZ(leaf);
5563 data->mv_data = NODEDATA(leaf);
5567 /* Read overflow data.
5569 data->mv_size = NODEDSZ(leaf);
5570 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5571 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5572 DPRINTF(("read overflow page %"Z"u failed", pgno));
5575 data->mv_data = METADATA(omp);
5581 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5582 MDB_val *key, MDB_val *data)
5589 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5591 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5594 if (txn->mt_flags & MDB_TXN_BLOCKED)
5597 mdb_cursor_init(&mc, txn, dbi, &mx);
5598 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5601 /** Find a sibling for a page.
5602 * Replaces the page at the top of the cursor's stack with the
5603 * specified sibling, if one exists.
5604 * @param[in] mc The cursor for this operation.
5605 * @param[in] move_right Non-zero if the right sibling is requested,
5606 * otherwise the left sibling.
5607 * @return 0 on success, non-zero on failure.
5610 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5616 if (mc->mc_snum < 2) {
5617 return MDB_NOTFOUND; /* root has no siblings */
5621 DPRINTF(("parent page is page %"Z"u, index %u",
5622 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5624 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5625 : (mc->mc_ki[mc->mc_top] == 0)) {
5626 DPRINTF(("no more keys left, moving to %s sibling",
5627 move_right ? "right" : "left"));
5628 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5629 /* undo cursor_pop before returning */
5636 mc->mc_ki[mc->mc_top]++;
5638 mc->mc_ki[mc->mc_top]--;
5639 DPRINTF(("just moving to %s index key %u",
5640 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5642 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5644 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5645 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5646 /* mc will be inconsistent if caller does mc_snum++ as above */
5647 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5651 mdb_cursor_push(mc, mp);
5653 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5658 /** Move the cursor to the next data item. */
5660 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5666 if (mc->mc_flags & C_EOF) {
5667 return MDB_NOTFOUND;
5670 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5672 mp = mc->mc_pg[mc->mc_top];
5674 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5675 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5676 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5677 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5678 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5679 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5680 if (rc == MDB_SUCCESS)
5681 MDB_GET_KEY(leaf, key);
5686 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5687 if (op == MDB_NEXT_DUP)
5688 return MDB_NOTFOUND;
5692 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5693 mdb_dbg_pgno(mp), (void *) mc));
5694 if (mc->mc_flags & C_DEL)
5697 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5698 DPUTS("=====> move to next sibling page");
5699 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5700 mc->mc_flags |= C_EOF;
5703 mp = mc->mc_pg[mc->mc_top];
5704 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5706 mc->mc_ki[mc->mc_top]++;
5709 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5710 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5713 key->mv_size = mc->mc_db->md_pad;
5714 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5718 mdb_cassert(mc, IS_LEAF(mp));
5719 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5721 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5722 mdb_xcursor_init1(mc, leaf);
5725 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5728 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5729 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5730 if (rc != MDB_SUCCESS)
5735 MDB_GET_KEY(leaf, key);
5739 /** Move the cursor to the previous data item. */
5741 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5747 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5749 mp = mc->mc_pg[mc->mc_top];
5751 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5752 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5753 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5754 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5755 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5756 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5757 if (rc == MDB_SUCCESS) {
5758 MDB_GET_KEY(leaf, key);
5759 mc->mc_flags &= ~C_EOF;
5765 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5766 if (op == MDB_PREV_DUP)
5767 return MDB_NOTFOUND;
5771 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5772 mdb_dbg_pgno(mp), (void *) mc));
5774 if (mc->mc_ki[mc->mc_top] == 0) {
5775 DPUTS("=====> move to prev sibling page");
5776 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5779 mp = mc->mc_pg[mc->mc_top];
5780 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5781 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5783 mc->mc_ki[mc->mc_top]--;
5785 mc->mc_flags &= ~C_EOF;
5787 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5788 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5791 key->mv_size = mc->mc_db->md_pad;
5792 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5796 mdb_cassert(mc, IS_LEAF(mp));
5797 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5799 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5800 mdb_xcursor_init1(mc, leaf);
5803 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5806 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5807 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5808 if (rc != MDB_SUCCESS)
5813 MDB_GET_KEY(leaf, key);
5817 /** Set the cursor on a specific data item. */
5819 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5820 MDB_cursor_op op, int *exactp)
5824 MDB_node *leaf = NULL;
5827 if (key->mv_size == 0)
5828 return MDB_BAD_VALSIZE;
5831 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5833 /* See if we're already on the right page */
5834 if (mc->mc_flags & C_INITIALIZED) {
5837 mp = mc->mc_pg[mc->mc_top];
5839 mc->mc_ki[mc->mc_top] = 0;
5840 return MDB_NOTFOUND;
5842 if (mp->mp_flags & P_LEAF2) {
5843 nodekey.mv_size = mc->mc_db->md_pad;
5844 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5846 leaf = NODEPTR(mp, 0);
5847 MDB_GET_KEY2(leaf, nodekey);
5849 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5851 /* Probably happens rarely, but first node on the page
5852 * was the one we wanted.
5854 mc->mc_ki[mc->mc_top] = 0;
5861 unsigned int nkeys = NUMKEYS(mp);
5863 if (mp->mp_flags & P_LEAF2) {
5864 nodekey.mv_data = LEAF2KEY(mp,
5865 nkeys-1, nodekey.mv_size);
5867 leaf = NODEPTR(mp, nkeys-1);
5868 MDB_GET_KEY2(leaf, nodekey);
5870 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5872 /* last node was the one we wanted */
5873 mc->mc_ki[mc->mc_top] = nkeys-1;
5879 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5880 /* This is definitely the right page, skip search_page */
5881 if (mp->mp_flags & P_LEAF2) {
5882 nodekey.mv_data = LEAF2KEY(mp,
5883 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5885 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5886 MDB_GET_KEY2(leaf, nodekey);
5888 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5890 /* current node was the one we wanted */
5900 /* If any parents have right-sibs, search.
5901 * Otherwise, there's nothing further.
5903 for (i=0; i<mc->mc_top; i++)
5905 NUMKEYS(mc->mc_pg[i])-1)
5907 if (i == mc->mc_top) {
5908 /* There are no other pages */
5909 mc->mc_ki[mc->mc_top] = nkeys;
5910 return MDB_NOTFOUND;
5914 /* There are no other pages */
5915 mc->mc_ki[mc->mc_top] = 0;
5916 if (op == MDB_SET_RANGE && !exactp) {
5920 return MDB_NOTFOUND;
5924 rc = mdb_page_search(mc, key, 0);
5925 if (rc != MDB_SUCCESS)
5928 mp = mc->mc_pg[mc->mc_top];
5929 mdb_cassert(mc, IS_LEAF(mp));
5932 leaf = mdb_node_search(mc, key, exactp);
5933 if (exactp != NULL && !*exactp) {
5934 /* MDB_SET specified and not an exact match. */
5935 return MDB_NOTFOUND;
5939 DPUTS("===> inexact leaf not found, goto sibling");
5940 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5941 mc->mc_flags |= C_EOF;
5942 return rc; /* no entries matched */
5944 mp = mc->mc_pg[mc->mc_top];
5945 mdb_cassert(mc, IS_LEAF(mp));
5946 leaf = NODEPTR(mp, 0);
5950 mc->mc_flags |= C_INITIALIZED;
5951 mc->mc_flags &= ~C_EOF;
5954 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5955 key->mv_size = mc->mc_db->md_pad;
5956 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5961 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5962 mdb_xcursor_init1(mc, leaf);
5965 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5966 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5967 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5970 if (op == MDB_GET_BOTH) {
5976 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5977 if (rc != MDB_SUCCESS)
5980 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5983 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5985 dcmp = mc->mc_dbx->md_dcmp;
5986 #if UINT_MAX < SIZE_MAX
5987 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5988 dcmp = mdb_cmp_clong;
5990 rc = dcmp(data, &olddata);
5992 if (op == MDB_GET_BOTH || rc > 0)
5993 return MDB_NOTFOUND;
6000 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6001 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6006 /* The key already matches in all other cases */
6007 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6008 MDB_GET_KEY(leaf, key);
6009 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6014 /** Move the cursor to the first item in the database. */
6016 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6022 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6024 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6025 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6026 if (rc != MDB_SUCCESS)
6029 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6031 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6032 mc->mc_flags |= C_INITIALIZED;
6033 mc->mc_flags &= ~C_EOF;
6035 mc->mc_ki[mc->mc_top] = 0;
6037 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6038 key->mv_size = mc->mc_db->md_pad;
6039 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6044 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6045 mdb_xcursor_init1(mc, leaf);
6046 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6050 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6054 MDB_GET_KEY(leaf, key);
6058 /** Move the cursor to the last item in the database. */
6060 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6066 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6068 if (!(mc->mc_flags & C_EOF)) {
6070 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6071 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6072 if (rc != MDB_SUCCESS)
6075 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6078 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6079 mc->mc_flags |= C_INITIALIZED|C_EOF;
6080 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6082 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6083 key->mv_size = mc->mc_db->md_pad;
6084 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6089 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6090 mdb_xcursor_init1(mc, leaf);
6091 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6095 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6100 MDB_GET_KEY(leaf, key);
6105 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6110 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6115 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6119 case MDB_GET_CURRENT:
6120 if (!(mc->mc_flags & C_INITIALIZED)) {
6123 MDB_page *mp = mc->mc_pg[mc->mc_top];
6124 int nkeys = NUMKEYS(mp);
6125 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6126 mc->mc_ki[mc->mc_top] = nkeys;
6132 key->mv_size = mc->mc_db->md_pad;
6133 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6135 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6136 MDB_GET_KEY(leaf, key);
6138 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6139 if (mc->mc_flags & C_DEL)
6140 mdb_xcursor_init1(mc, leaf);
6141 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6143 rc = mdb_node_read(mc->mc_txn, leaf, data);
6150 case MDB_GET_BOTH_RANGE:
6155 if (mc->mc_xcursor == NULL) {
6156 rc = MDB_INCOMPATIBLE;
6166 rc = mdb_cursor_set(mc, key, data, op,
6167 op == MDB_SET_RANGE ? NULL : &exact);
6170 case MDB_GET_MULTIPLE:
6171 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6175 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6176 rc = MDB_INCOMPATIBLE;
6180 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6181 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6184 case MDB_NEXT_MULTIPLE:
6189 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6190 rc = MDB_INCOMPATIBLE;
6193 if (!(mc->mc_flags & C_INITIALIZED))
6194 rc = mdb_cursor_first(mc, key, data);
6196 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6197 if (rc == MDB_SUCCESS) {
6198 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6201 mx = &mc->mc_xcursor->mx_cursor;
6202 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6204 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6205 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6213 case MDB_NEXT_NODUP:
6214 if (!(mc->mc_flags & C_INITIALIZED))
6215 rc = mdb_cursor_first(mc, key, data);
6217 rc = mdb_cursor_next(mc, key, data, op);
6221 case MDB_PREV_NODUP:
6222 if (!(mc->mc_flags & C_INITIALIZED)) {
6223 rc = mdb_cursor_last(mc, key, data);
6226 mc->mc_flags |= C_INITIALIZED;
6227 mc->mc_ki[mc->mc_top]++;
6229 rc = mdb_cursor_prev(mc, key, data, op);
6232 rc = mdb_cursor_first(mc, key, data);
6235 mfunc = mdb_cursor_first;
6237 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6241 if (mc->mc_xcursor == NULL) {
6242 rc = MDB_INCOMPATIBLE;
6246 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6247 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6248 MDB_GET_KEY(leaf, key);
6249 rc = mdb_node_read(mc->mc_txn, leaf, data);
6253 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6257 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6260 rc = mdb_cursor_last(mc, key, data);
6263 mfunc = mdb_cursor_last;
6266 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6271 if (mc->mc_flags & C_DEL)
6272 mc->mc_flags ^= C_DEL;
6277 /** Touch all the pages in the cursor stack. Set mc_top.
6278 * Makes sure all the pages are writable, before attempting a write operation.
6279 * @param[in] mc The cursor to operate on.
6282 mdb_cursor_touch(MDB_cursor *mc)
6284 int rc = MDB_SUCCESS;
6286 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6289 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6291 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6292 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6295 *mc->mc_dbflag |= DB_DIRTY;
6300 rc = mdb_page_touch(mc);
6301 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6302 mc->mc_top = mc->mc_snum-1;
6307 /** Do not spill pages to disk if txn is getting full, may fail instead */
6308 #define MDB_NOSPILL 0x8000
6311 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6315 MDB_node *leaf = NULL;
6318 MDB_val xdata, *rdata, dkey, olddata;
6320 int do_sub = 0, insert_key, insert_data;
6321 unsigned int mcount = 0, dcount = 0, nospill;
6324 unsigned int nflags;
6327 if (mc == NULL || key == NULL)
6330 env = mc->mc_txn->mt_env;
6332 /* Check this first so counter will always be zero on any
6335 if (flags & MDB_MULTIPLE) {
6336 dcount = data[1].mv_size;
6337 data[1].mv_size = 0;
6338 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6339 return MDB_INCOMPATIBLE;
6342 nospill = flags & MDB_NOSPILL;
6343 flags &= ~MDB_NOSPILL;
6345 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6346 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6348 if (key->mv_size-1 >= ENV_MAXKEY(env))
6349 return MDB_BAD_VALSIZE;
6351 #if SIZE_MAX > MAXDATASIZE
6352 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6353 return MDB_BAD_VALSIZE;
6355 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6356 return MDB_BAD_VALSIZE;
6359 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6360 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6364 if (flags == MDB_CURRENT) {
6365 if (!(mc->mc_flags & C_INITIALIZED))
6368 } else if (mc->mc_db->md_root == P_INVALID) {
6369 /* new database, cursor has nothing to point to */
6372 mc->mc_flags &= ~C_INITIALIZED;
6377 if (flags & MDB_APPEND) {
6379 rc = mdb_cursor_last(mc, &k2, &d2);
6381 rc = mc->mc_dbx->md_cmp(key, &k2);
6384 mc->mc_ki[mc->mc_top]++;
6386 /* new key is <= last key */
6391 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6393 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6394 DPRINTF(("duplicate key [%s]", DKEY(key)));
6396 return MDB_KEYEXIST;
6398 if (rc && rc != MDB_NOTFOUND)
6402 if (mc->mc_flags & C_DEL)
6403 mc->mc_flags ^= C_DEL;
6405 /* Cursor is positioned, check for room in the dirty list */
6407 if (flags & MDB_MULTIPLE) {
6409 xdata.mv_size = data->mv_size * dcount;
6413 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6417 if (rc == MDB_NO_ROOT) {
6419 /* new database, write a root leaf page */
6420 DPUTS("allocating new root leaf page");
6421 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6424 mdb_cursor_push(mc, np);
6425 mc->mc_db->md_root = np->mp_pgno;
6426 mc->mc_db->md_depth++;
6427 *mc->mc_dbflag |= DB_DIRTY;
6428 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6430 np->mp_flags |= P_LEAF2;
6431 mc->mc_flags |= C_INITIALIZED;
6433 /* make sure all cursor pages are writable */
6434 rc2 = mdb_cursor_touch(mc);
6439 insert_key = insert_data = rc;
6441 /* The key does not exist */
6442 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6443 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6444 LEAFSIZE(key, data) > env->me_nodemax)
6446 /* Too big for a node, insert in sub-DB. Set up an empty
6447 * "old sub-page" for prep_subDB to expand to a full page.
6449 fp_flags = P_LEAF|P_DIRTY;
6451 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6452 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6453 olddata.mv_size = PAGEHDRSZ;
6457 /* there's only a key anyway, so this is a no-op */
6458 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6460 unsigned int ksize = mc->mc_db->md_pad;
6461 if (key->mv_size != ksize)
6462 return MDB_BAD_VALSIZE;
6463 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6464 memcpy(ptr, key->mv_data, ksize);
6466 /* if overwriting slot 0 of leaf, need to
6467 * update branch key if there is a parent page
6469 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6470 unsigned short top = mc->mc_top;
6472 /* slot 0 is always an empty key, find real slot */
6473 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6475 if (mc->mc_ki[mc->mc_top])
6476 rc2 = mdb_update_key(mc, key);
6487 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6488 olddata.mv_size = NODEDSZ(leaf);
6489 olddata.mv_data = NODEDATA(leaf);
6492 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6493 /* Prepare (sub-)page/sub-DB to accept the new item,
6494 * if needed. fp: old sub-page or a header faking
6495 * it. mp: new (sub-)page. offset: growth in page
6496 * size. xdata: node data with new page or DB.
6498 unsigned i, offset = 0;
6499 mp = fp = xdata.mv_data = env->me_pbuf;
6500 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6502 /* Was a single item before, must convert now */
6503 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6505 /* Just overwrite the current item */
6506 if (flags == MDB_CURRENT)
6508 dcmp = mc->mc_dbx->md_dcmp;
6509 #if UINT_MAX < SIZE_MAX
6510 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6511 dcmp = mdb_cmp_clong;
6513 /* does data match? */
6514 if (!dcmp(data, &olddata)) {
6515 if (flags & MDB_NODUPDATA)
6516 return MDB_KEYEXIST;
6521 /* Back up original data item */
6522 dkey.mv_size = olddata.mv_size;
6523 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6525 /* Make sub-page header for the dup items, with dummy body */
6526 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6527 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6528 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6529 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6530 fp->mp_flags |= P_LEAF2;
6531 fp->mp_pad = data->mv_size;
6532 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6534 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6535 (dkey.mv_size & 1) + (data->mv_size & 1);
6537 fp->mp_upper = xdata.mv_size - PAGEBASE;
6538 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6539 } else if (leaf->mn_flags & F_SUBDATA) {
6540 /* Data is on sub-DB, just store it */
6541 flags |= F_DUPDATA|F_SUBDATA;
6544 /* Data is on sub-page */
6545 fp = olddata.mv_data;
6548 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6549 offset = EVEN(NODESIZE + sizeof(indx_t) +
6553 offset = fp->mp_pad;
6554 if (SIZELEFT(fp) < offset) {
6555 offset *= 4; /* space for 4 more */
6558 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6560 fp->mp_flags |= P_DIRTY;
6561 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6562 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6566 xdata.mv_size = olddata.mv_size + offset;
6569 fp_flags = fp->mp_flags;
6570 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6571 /* Too big for a sub-page, convert to sub-DB */
6572 fp_flags &= ~P_SUBP;
6574 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6575 fp_flags |= P_LEAF2;
6576 dummy.md_pad = fp->mp_pad;
6577 dummy.md_flags = MDB_DUPFIXED;
6578 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6579 dummy.md_flags |= MDB_INTEGERKEY;
6585 dummy.md_branch_pages = 0;
6586 dummy.md_leaf_pages = 1;
6587 dummy.md_overflow_pages = 0;
6588 dummy.md_entries = NUMKEYS(fp);
6589 xdata.mv_size = sizeof(MDB_db);
6590 xdata.mv_data = &dummy;
6591 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6593 offset = env->me_psize - olddata.mv_size;
6594 flags |= F_DUPDATA|F_SUBDATA;
6595 dummy.md_root = mp->mp_pgno;
6598 mp->mp_flags = fp_flags | P_DIRTY;
6599 mp->mp_pad = fp->mp_pad;
6600 mp->mp_lower = fp->mp_lower;
6601 mp->mp_upper = fp->mp_upper + offset;
6602 if (fp_flags & P_LEAF2) {
6603 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6605 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6606 olddata.mv_size - fp->mp_upper - PAGEBASE);
6607 for (i=0; i<NUMKEYS(fp); i++)
6608 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6616 mdb_node_del(mc, 0);
6620 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6621 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6622 return MDB_INCOMPATIBLE;
6623 /* overflow page overwrites need special handling */
6624 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6627 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6629 memcpy(&pg, olddata.mv_data, sizeof(pg));
6630 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6632 ovpages = omp->mp_pages;
6634 /* Is the ov page large enough? */
6635 if (ovpages >= dpages) {
6636 if (!(omp->mp_flags & P_DIRTY) &&
6637 (level || (env->me_flags & MDB_WRITEMAP)))
6639 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6642 level = 0; /* dirty in this txn or clean */
6645 if (omp->mp_flags & P_DIRTY) {
6646 /* yes, overwrite it. Note in this case we don't
6647 * bother to try shrinking the page if the new data
6648 * is smaller than the overflow threshold.
6651 /* It is writable only in a parent txn */
6652 size_t sz = (size_t) env->me_psize * ovpages, off;
6653 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6659 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6660 mdb_cassert(mc, rc2 == 0);
6661 if (!(flags & MDB_RESERVE)) {
6662 /* Copy end of page, adjusting alignment so
6663 * compiler may copy words instead of bytes.
6665 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6666 memcpy((size_t *)((char *)np + off),
6667 (size_t *)((char *)omp + off), sz - off);
6670 memcpy(np, omp, sz); /* Copy beginning of page */
6673 SETDSZ(leaf, data->mv_size);
6674 if (F_ISSET(flags, MDB_RESERVE))
6675 data->mv_data = METADATA(omp);
6677 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6681 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6683 } else if (data->mv_size == olddata.mv_size) {
6684 /* same size, just replace it. Note that we could
6685 * also reuse this node if the new data is smaller,
6686 * but instead we opt to shrink the node in that case.
6688 if (F_ISSET(flags, MDB_RESERVE))
6689 data->mv_data = olddata.mv_data;
6690 else if (!(mc->mc_flags & C_SUB))
6691 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6693 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6698 mdb_node_del(mc, 0);
6704 nflags = flags & NODE_ADD_FLAGS;
6705 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6706 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6707 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6708 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6710 nflags |= MDB_SPLIT_REPLACE;
6711 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6713 /* There is room already in this leaf page. */
6714 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6715 if (rc == 0 && insert_key) {
6716 /* Adjust other cursors pointing to mp */
6717 MDB_cursor *m2, *m3;
6718 MDB_dbi dbi = mc->mc_dbi;
6719 unsigned i = mc->mc_top;
6720 MDB_page *mp = mc->mc_pg[i];
6722 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6723 if (mc->mc_flags & C_SUB)
6724 m3 = &m2->mc_xcursor->mx_cursor;
6727 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6728 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6735 if (rc == MDB_SUCCESS) {
6736 /* Now store the actual data in the child DB. Note that we're
6737 * storing the user data in the keys field, so there are strict
6738 * size limits on dupdata. The actual data fields of the child
6739 * DB are all zero size.
6747 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6748 if (flags & MDB_CURRENT) {
6749 xflags = MDB_CURRENT|MDB_NOSPILL;
6751 mdb_xcursor_init1(mc, leaf);
6752 xflags = (flags & MDB_NODUPDATA) ?
6753 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6755 /* converted, write the original data first */
6757 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6761 /* Adjust other cursors pointing to mp */
6763 unsigned i = mc->mc_top;
6764 MDB_page *mp = mc->mc_pg[i];
6766 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6767 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6768 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6769 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6770 mdb_xcursor_init1(m2, leaf);
6774 /* we've done our job */
6777 ecount = mc->mc_xcursor->mx_db.md_entries;
6778 if (flags & MDB_APPENDDUP)
6779 xflags |= MDB_APPEND;
6780 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6781 if (flags & F_SUBDATA) {
6782 void *db = NODEDATA(leaf);
6783 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6785 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6787 /* Increment count unless we just replaced an existing item. */
6789 mc->mc_db->md_entries++;
6791 /* Invalidate txn if we created an empty sub-DB */
6794 /* If we succeeded and the key didn't exist before,
6795 * make sure the cursor is marked valid.
6797 mc->mc_flags |= C_INITIALIZED;
6799 if (flags & MDB_MULTIPLE) {
6802 /* let caller know how many succeeded, if any */
6803 data[1].mv_size = mcount;
6804 if (mcount < dcount) {
6805 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6806 insert_key = insert_data = 0;
6813 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6816 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6821 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6827 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6828 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6830 if (!(mc->mc_flags & C_INITIALIZED))
6833 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6834 return MDB_NOTFOUND;
6836 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6839 rc = mdb_cursor_touch(mc);
6843 mp = mc->mc_pg[mc->mc_top];
6846 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6848 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6849 if (flags & MDB_NODUPDATA) {
6850 /* mdb_cursor_del0() will subtract the final entry */
6851 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6853 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6854 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6856 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6859 /* If sub-DB still has entries, we're done */
6860 if (mc->mc_xcursor->mx_db.md_entries) {
6861 if (leaf->mn_flags & F_SUBDATA) {
6862 /* update subDB info */
6863 void *db = NODEDATA(leaf);
6864 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6867 /* shrink fake page */
6868 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6869 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6870 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6871 /* fix other sub-DB cursors pointed at this fake page */
6872 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6873 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6874 if (m2->mc_pg[mc->mc_top] == mp &&
6875 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6876 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6879 mc->mc_db->md_entries--;
6880 mc->mc_flags |= C_DEL;
6883 /* otherwise fall thru and delete the sub-DB */
6886 if (leaf->mn_flags & F_SUBDATA) {
6887 /* add all the child DB's pages to the free list */
6888 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6893 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6894 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6895 rc = MDB_INCOMPATIBLE;
6899 /* add overflow pages to free list */
6900 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6904 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6905 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6906 (rc = mdb_ovpage_free(mc, omp)))
6911 return mdb_cursor_del0(mc);
6914 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6918 /** Allocate and initialize new pages for a database.
6919 * @param[in] mc a cursor on the database being added to.
6920 * @param[in] flags flags defining what type of page is being allocated.
6921 * @param[in] num the number of pages to allocate. This is usually 1,
6922 * unless allocating overflow pages for a large record.
6923 * @param[out] mp Address of a page, or NULL on failure.
6924 * @return 0 on success, non-zero on failure.
6927 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6932 if ((rc = mdb_page_alloc(mc, num, &np)))
6934 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6935 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6936 np->mp_flags = flags | P_DIRTY;
6937 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6938 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6941 mc->mc_db->md_branch_pages++;
6942 else if (IS_LEAF(np))
6943 mc->mc_db->md_leaf_pages++;
6944 else if (IS_OVERFLOW(np)) {
6945 mc->mc_db->md_overflow_pages += num;
6953 /** Calculate the size of a leaf node.
6954 * The size depends on the environment's page size; if a data item
6955 * is too large it will be put onto an overflow page and the node
6956 * size will only include the key and not the data. Sizes are always
6957 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6958 * of the #MDB_node headers.
6959 * @param[in] env The environment handle.
6960 * @param[in] key The key for the node.
6961 * @param[in] data The data for the node.
6962 * @return The number of bytes needed to store the node.
6965 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6969 sz = LEAFSIZE(key, data);
6970 if (sz > env->me_nodemax) {
6971 /* put on overflow page */
6972 sz -= data->mv_size - sizeof(pgno_t);
6975 return EVEN(sz + sizeof(indx_t));
6978 /** Calculate the size of a branch node.
6979 * The size should depend on the environment's page size but since
6980 * we currently don't support spilling large keys onto overflow
6981 * pages, it's simply the size of the #MDB_node header plus the
6982 * size of the key. Sizes are always rounded up to an even number
6983 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6984 * @param[in] env The environment handle.
6985 * @param[in] key The key for the node.
6986 * @return The number of bytes needed to store the node.
6989 mdb_branch_size(MDB_env *env, MDB_val *key)
6994 if (sz > env->me_nodemax) {
6995 /* put on overflow page */
6996 /* not implemented */
6997 /* sz -= key->size - sizeof(pgno_t); */
7000 return sz + sizeof(indx_t);
7003 /** Add a node to the page pointed to by the cursor.
7004 * @param[in] mc The cursor for this operation.
7005 * @param[in] indx The index on the page where the new node should be added.
7006 * @param[in] key The key for the new node.
7007 * @param[in] data The data for the new node, if any.
7008 * @param[in] pgno The page number, if adding a branch node.
7009 * @param[in] flags Flags for the node.
7010 * @return 0 on success, non-zero on failure. Possible errors are:
7012 * <li>ENOMEM - failed to allocate overflow pages for the node.
7013 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7014 * should never happen since all callers already calculate the
7015 * page's free space before calling this function.
7019 mdb_node_add(MDB_cursor *mc, indx_t indx,
7020 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7023 size_t node_size = NODESIZE;
7027 MDB_page *mp = mc->mc_pg[mc->mc_top];
7028 MDB_page *ofp = NULL; /* overflow page */
7032 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7034 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7035 IS_LEAF(mp) ? "leaf" : "branch",
7036 IS_SUBP(mp) ? "sub-" : "",
7037 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7038 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7041 /* Move higher keys up one slot. */
7042 int ksize = mc->mc_db->md_pad, dif;
7043 char *ptr = LEAF2KEY(mp, indx, ksize);
7044 dif = NUMKEYS(mp) - indx;
7046 memmove(ptr+ksize, ptr, dif*ksize);
7047 /* insert new key */
7048 memcpy(ptr, key->mv_data, ksize);
7050 /* Just using these for counting */
7051 mp->mp_lower += sizeof(indx_t);
7052 mp->mp_upper -= ksize - sizeof(indx_t);
7056 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7058 node_size += key->mv_size;
7060 mdb_cassert(mc, key && data);
7061 if (F_ISSET(flags, F_BIGDATA)) {
7062 /* Data already on overflow page. */
7063 node_size += sizeof(pgno_t);
7064 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7065 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7067 /* Put data on overflow page. */
7068 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7069 data->mv_size, node_size+data->mv_size));
7070 node_size = EVEN(node_size + sizeof(pgno_t));
7071 if ((ssize_t)node_size > room)
7073 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7075 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7079 node_size += data->mv_size;
7082 node_size = EVEN(node_size);
7083 if ((ssize_t)node_size > room)
7087 /* Move higher pointers up one slot. */
7088 for (i = NUMKEYS(mp); i > indx; i--)
7089 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7091 /* Adjust free space offsets. */
7092 ofs = mp->mp_upper - node_size;
7093 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7094 mp->mp_ptrs[indx] = ofs;
7096 mp->mp_lower += sizeof(indx_t);
7098 /* Write the node data. */
7099 node = NODEPTR(mp, indx);
7100 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7101 node->mn_flags = flags;
7103 SETDSZ(node,data->mv_size);
7108 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7111 ndata = NODEDATA(node);
7113 if (F_ISSET(flags, F_BIGDATA))
7114 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7115 else if (F_ISSET(flags, MDB_RESERVE))
7116 data->mv_data = ndata;
7118 memcpy(ndata, data->mv_data, data->mv_size);
7120 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7121 ndata = METADATA(ofp);
7122 if (F_ISSET(flags, MDB_RESERVE))
7123 data->mv_data = ndata;
7125 memcpy(ndata, data->mv_data, data->mv_size);
7132 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7133 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7134 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7135 DPRINTF(("node size = %"Z"u", node_size));
7136 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7137 return MDB_PAGE_FULL;
7140 /** Delete the specified node from a page.
7141 * @param[in] mc Cursor pointing to the node to delete.
7142 * @param[in] ksize The size of a node. Only used if the page is
7143 * part of a #MDB_DUPFIXED database.
7146 mdb_node_del(MDB_cursor *mc, int ksize)
7148 MDB_page *mp = mc->mc_pg[mc->mc_top];
7149 indx_t indx = mc->mc_ki[mc->mc_top];
7151 indx_t i, j, numkeys, ptr;
7155 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7156 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7157 numkeys = NUMKEYS(mp);
7158 mdb_cassert(mc, indx < numkeys);
7161 int x = numkeys - 1 - indx;
7162 base = LEAF2KEY(mp, indx, ksize);
7164 memmove(base, base + ksize, x * ksize);
7165 mp->mp_lower -= sizeof(indx_t);
7166 mp->mp_upper += ksize - sizeof(indx_t);
7170 node = NODEPTR(mp, indx);
7171 sz = NODESIZE + node->mn_ksize;
7173 if (F_ISSET(node->mn_flags, F_BIGDATA))
7174 sz += sizeof(pgno_t);
7176 sz += NODEDSZ(node);
7180 ptr = mp->mp_ptrs[indx];
7181 for (i = j = 0; i < numkeys; i++) {
7183 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7184 if (mp->mp_ptrs[i] < ptr)
7185 mp->mp_ptrs[j] += sz;
7190 base = (char *)mp + mp->mp_upper + PAGEBASE;
7191 memmove(base + sz, base, ptr - mp->mp_upper);
7193 mp->mp_lower -= sizeof(indx_t);
7197 /** Compact the main page after deleting a node on a subpage.
7198 * @param[in] mp The main page to operate on.
7199 * @param[in] indx The index of the subpage on the main page.
7202 mdb_node_shrink(MDB_page *mp, indx_t indx)
7207 indx_t delta, nsize, len, ptr;
7210 node = NODEPTR(mp, indx);
7211 sp = (MDB_page *)NODEDATA(node);
7212 delta = SIZELEFT(sp);
7213 nsize = NODEDSZ(node) - delta;
7215 /* Prepare to shift upward, set len = length(subpage part to shift) */
7219 return; /* do not make the node uneven-sized */
7221 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7222 for (i = NUMKEYS(sp); --i >= 0; )
7223 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7226 sp->mp_upper = sp->mp_lower;
7227 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7228 SETDSZ(node, nsize);
7230 /* Shift <lower nodes...initial part of subpage> upward */
7231 base = (char *)mp + mp->mp_upper + PAGEBASE;
7232 memmove(base + delta, base, (char *)sp + len - base);
7234 ptr = mp->mp_ptrs[indx];
7235 for (i = NUMKEYS(mp); --i >= 0; ) {
7236 if (mp->mp_ptrs[i] <= ptr)
7237 mp->mp_ptrs[i] += delta;
7239 mp->mp_upper += delta;
7242 /** Initial setup of a sorted-dups cursor.
7243 * Sorted duplicates are implemented as a sub-database for the given key.
7244 * The duplicate data items are actually keys of the sub-database.
7245 * Operations on the duplicate data items are performed using a sub-cursor
7246 * initialized when the sub-database is first accessed. This function does
7247 * the preliminary setup of the sub-cursor, filling in the fields that
7248 * depend only on the parent DB.
7249 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7252 mdb_xcursor_init0(MDB_cursor *mc)
7254 MDB_xcursor *mx = mc->mc_xcursor;
7256 mx->mx_cursor.mc_xcursor = NULL;
7257 mx->mx_cursor.mc_txn = mc->mc_txn;
7258 mx->mx_cursor.mc_db = &mx->mx_db;
7259 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7260 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7261 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7262 mx->mx_cursor.mc_snum = 0;
7263 mx->mx_cursor.mc_top = 0;
7264 mx->mx_cursor.mc_flags = C_SUB;
7265 mx->mx_dbx.md_name.mv_size = 0;
7266 mx->mx_dbx.md_name.mv_data = NULL;
7267 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7268 mx->mx_dbx.md_dcmp = NULL;
7269 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7272 /** Final setup of a sorted-dups cursor.
7273 * Sets up the fields that depend on the data from the main cursor.
7274 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7275 * @param[in] node The data containing the #MDB_db record for the
7276 * sorted-dup database.
7279 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7281 MDB_xcursor *mx = mc->mc_xcursor;
7283 if (node->mn_flags & F_SUBDATA) {
7284 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7285 mx->mx_cursor.mc_pg[0] = 0;
7286 mx->mx_cursor.mc_snum = 0;
7287 mx->mx_cursor.mc_top = 0;
7288 mx->mx_cursor.mc_flags = C_SUB;
7290 MDB_page *fp = NODEDATA(node);
7291 mx->mx_db.md_pad = 0;
7292 mx->mx_db.md_flags = 0;
7293 mx->mx_db.md_depth = 1;
7294 mx->mx_db.md_branch_pages = 0;
7295 mx->mx_db.md_leaf_pages = 1;
7296 mx->mx_db.md_overflow_pages = 0;
7297 mx->mx_db.md_entries = NUMKEYS(fp);
7298 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7299 mx->mx_cursor.mc_snum = 1;
7300 mx->mx_cursor.mc_top = 0;
7301 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7302 mx->mx_cursor.mc_pg[0] = fp;
7303 mx->mx_cursor.mc_ki[0] = 0;
7304 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7305 mx->mx_db.md_flags = MDB_DUPFIXED;
7306 mx->mx_db.md_pad = fp->mp_pad;
7307 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7308 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7311 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7312 mx->mx_db.md_root));
7313 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7314 #if UINT_MAX < SIZE_MAX
7315 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7316 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7320 /** Initialize a cursor for a given transaction and database. */
7322 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7325 mc->mc_backup = NULL;
7328 mc->mc_db = &txn->mt_dbs[dbi];
7329 mc->mc_dbx = &txn->mt_dbxs[dbi];
7330 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7336 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7337 mdb_tassert(txn, mx != NULL);
7338 mc->mc_xcursor = mx;
7339 mdb_xcursor_init0(mc);
7341 mc->mc_xcursor = NULL;
7343 if (*mc->mc_dbflag & DB_STALE) {
7344 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7349 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7352 size_t size = sizeof(MDB_cursor);
7354 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7357 if (txn->mt_flags & MDB_TXN_BLOCKED)
7360 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7363 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7364 size += sizeof(MDB_xcursor);
7366 if ((mc = malloc(size)) != NULL) {
7367 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7368 if (txn->mt_cursors) {
7369 mc->mc_next = txn->mt_cursors[dbi];
7370 txn->mt_cursors[dbi] = mc;
7371 mc->mc_flags |= C_UNTRACK;
7383 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7385 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7388 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7391 if (txn->mt_flags & MDB_TXN_BLOCKED)
7394 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7398 /* Return the count of duplicate data items for the current key */
7400 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7404 if (mc == NULL || countp == NULL)
7407 if (mc->mc_xcursor == NULL)
7408 return MDB_INCOMPATIBLE;
7410 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7413 if (!(mc->mc_flags & C_INITIALIZED))
7416 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7417 return MDB_NOTFOUND;
7419 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7420 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7423 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7426 *countp = mc->mc_xcursor->mx_db.md_entries;
7432 mdb_cursor_close(MDB_cursor *mc)
7434 if (mc && !mc->mc_backup) {
7435 /* remove from txn, if tracked */
7436 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7437 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7438 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7440 *prev = mc->mc_next;
7447 mdb_cursor_txn(MDB_cursor *mc)
7449 if (!mc) return NULL;
7454 mdb_cursor_dbi(MDB_cursor *mc)
7459 /** Replace the key for a branch node with a new key.
7460 * @param[in] mc Cursor pointing to the node to operate on.
7461 * @param[in] key The new key to use.
7462 * @return 0 on success, non-zero on failure.
7465 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7471 int delta, ksize, oksize;
7472 indx_t ptr, i, numkeys, indx;
7475 indx = mc->mc_ki[mc->mc_top];
7476 mp = mc->mc_pg[mc->mc_top];
7477 node = NODEPTR(mp, indx);
7478 ptr = mp->mp_ptrs[indx];
7482 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7483 k2.mv_data = NODEKEY(node);
7484 k2.mv_size = node->mn_ksize;
7485 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7487 mdb_dkey(&k2, kbuf2),
7493 /* Sizes must be 2-byte aligned. */
7494 ksize = EVEN(key->mv_size);
7495 oksize = EVEN(node->mn_ksize);
7496 delta = ksize - oksize;
7498 /* Shift node contents if EVEN(key length) changed. */
7500 if (delta > 0 && SIZELEFT(mp) < delta) {
7502 /* not enough space left, do a delete and split */
7503 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7504 pgno = NODEPGNO(node);
7505 mdb_node_del(mc, 0);
7506 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7509 numkeys = NUMKEYS(mp);
7510 for (i = 0; i < numkeys; i++) {
7511 if (mp->mp_ptrs[i] <= ptr)
7512 mp->mp_ptrs[i] -= delta;
7515 base = (char *)mp + mp->mp_upper + PAGEBASE;
7516 len = ptr - mp->mp_upper + NODESIZE;
7517 memmove(base - delta, base, len);
7518 mp->mp_upper -= delta;
7520 node = NODEPTR(mp, indx);
7523 /* But even if no shift was needed, update ksize */
7524 if (node->mn_ksize != key->mv_size)
7525 node->mn_ksize = key->mv_size;
7528 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7534 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7536 /** Move a node from csrc to cdst.
7539 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7546 unsigned short flags;
7550 /* Mark src and dst as dirty. */
7551 if ((rc = mdb_page_touch(csrc)) ||
7552 (rc = mdb_page_touch(cdst)))
7555 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7556 key.mv_size = csrc->mc_db->md_pad;
7557 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7559 data.mv_data = NULL;
7563 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7564 mdb_cassert(csrc, !((size_t)srcnode & 1));
7565 srcpg = NODEPGNO(srcnode);
7566 flags = srcnode->mn_flags;
7567 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7568 unsigned int snum = csrc->mc_snum;
7570 /* must find the lowest key below src */
7571 rc = mdb_page_search_lowest(csrc);
7574 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7575 key.mv_size = csrc->mc_db->md_pad;
7576 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7578 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7579 key.mv_size = NODEKSZ(s2);
7580 key.mv_data = NODEKEY(s2);
7582 csrc->mc_snum = snum--;
7583 csrc->mc_top = snum;
7585 key.mv_size = NODEKSZ(srcnode);
7586 key.mv_data = NODEKEY(srcnode);
7588 data.mv_size = NODEDSZ(srcnode);
7589 data.mv_data = NODEDATA(srcnode);
7591 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7592 unsigned int snum = cdst->mc_snum;
7595 /* must find the lowest key below dst */
7596 mdb_cursor_copy(cdst, &mn);
7597 rc = mdb_page_search_lowest(&mn);
7600 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7601 bkey.mv_size = mn.mc_db->md_pad;
7602 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7604 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7605 bkey.mv_size = NODEKSZ(s2);
7606 bkey.mv_data = NODEKEY(s2);
7608 mn.mc_snum = snum--;
7611 rc = mdb_update_key(&mn, &bkey);
7616 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7617 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7618 csrc->mc_ki[csrc->mc_top],
7620 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7621 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7623 /* Add the node to the destination page.
7625 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7626 if (rc != MDB_SUCCESS)
7629 /* Delete the node from the source page.
7631 mdb_node_del(csrc, key.mv_size);
7634 /* Adjust other cursors pointing to mp */
7635 MDB_cursor *m2, *m3;
7636 MDB_dbi dbi = csrc->mc_dbi;
7637 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7639 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7640 if (csrc->mc_flags & C_SUB)
7641 m3 = &m2->mc_xcursor->mx_cursor;
7644 if (m3 == csrc) continue;
7645 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7646 csrc->mc_ki[csrc->mc_top]) {
7647 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7648 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7653 /* Update the parent separators.
7655 if (csrc->mc_ki[csrc->mc_top] == 0) {
7656 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7657 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7658 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7660 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7661 key.mv_size = NODEKSZ(srcnode);
7662 key.mv_data = NODEKEY(srcnode);
7664 DPRINTF(("update separator for source page %"Z"u to [%s]",
7665 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7666 mdb_cursor_copy(csrc, &mn);
7669 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7672 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7674 indx_t ix = csrc->mc_ki[csrc->mc_top];
7675 nullkey.mv_size = 0;
7676 csrc->mc_ki[csrc->mc_top] = 0;
7677 rc = mdb_update_key(csrc, &nullkey);
7678 csrc->mc_ki[csrc->mc_top] = ix;
7679 mdb_cassert(csrc, rc == MDB_SUCCESS);
7683 if (cdst->mc_ki[cdst->mc_top] == 0) {
7684 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7685 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7686 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7688 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7689 key.mv_size = NODEKSZ(srcnode);
7690 key.mv_data = NODEKEY(srcnode);
7692 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7693 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7694 mdb_cursor_copy(cdst, &mn);
7697 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7700 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7702 indx_t ix = cdst->mc_ki[cdst->mc_top];
7703 nullkey.mv_size = 0;
7704 cdst->mc_ki[cdst->mc_top] = 0;
7705 rc = mdb_update_key(cdst, &nullkey);
7706 cdst->mc_ki[cdst->mc_top] = ix;
7707 mdb_cassert(cdst, rc == MDB_SUCCESS);
7714 /** Merge one page into another.
7715 * The nodes from the page pointed to by \b csrc will
7716 * be copied to the page pointed to by \b cdst and then
7717 * the \b csrc page will be freed.
7718 * @param[in] csrc Cursor pointing to the source page.
7719 * @param[in] cdst Cursor pointing to the destination page.
7720 * @return 0 on success, non-zero on failure.
7723 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7725 MDB_page *psrc, *pdst;
7732 psrc = csrc->mc_pg[csrc->mc_top];
7733 pdst = cdst->mc_pg[cdst->mc_top];
7735 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7737 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7738 mdb_cassert(csrc, cdst->mc_snum > 1);
7740 /* Mark dst as dirty. */
7741 if ((rc = mdb_page_touch(cdst)))
7744 /* Move all nodes from src to dst.
7746 j = nkeys = NUMKEYS(pdst);
7747 if (IS_LEAF2(psrc)) {
7748 key.mv_size = csrc->mc_db->md_pad;
7749 key.mv_data = METADATA(psrc);
7750 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7751 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7752 if (rc != MDB_SUCCESS)
7754 key.mv_data = (char *)key.mv_data + key.mv_size;
7757 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7758 srcnode = NODEPTR(psrc, i);
7759 if (i == 0 && IS_BRANCH(psrc)) {
7762 mdb_cursor_copy(csrc, &mn);
7763 /* must find the lowest key below src */
7764 rc = mdb_page_search_lowest(&mn);
7767 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7768 key.mv_size = mn.mc_db->md_pad;
7769 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7771 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7772 key.mv_size = NODEKSZ(s2);
7773 key.mv_data = NODEKEY(s2);
7776 key.mv_size = srcnode->mn_ksize;
7777 key.mv_data = NODEKEY(srcnode);
7780 data.mv_size = NODEDSZ(srcnode);
7781 data.mv_data = NODEDATA(srcnode);
7782 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7783 if (rc != MDB_SUCCESS)
7788 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7789 pdst->mp_pgno, NUMKEYS(pdst),
7790 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7792 /* Unlink the src page from parent and add to free list.
7795 mdb_node_del(csrc, 0);
7796 if (csrc->mc_ki[csrc->mc_top] == 0) {
7798 rc = mdb_update_key(csrc, &key);
7806 psrc = csrc->mc_pg[csrc->mc_top];
7807 /* If not operating on FreeDB, allow this page to be reused
7808 * in this txn. Otherwise just add to free list.
7810 rc = mdb_page_loose(csrc, psrc);
7814 csrc->mc_db->md_leaf_pages--;
7816 csrc->mc_db->md_branch_pages--;
7818 /* Adjust other cursors pointing to mp */
7819 MDB_cursor *m2, *m3;
7820 MDB_dbi dbi = csrc->mc_dbi;
7822 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7823 if (csrc->mc_flags & C_SUB)
7824 m3 = &m2->mc_xcursor->mx_cursor;
7827 if (m3 == csrc) continue;
7828 if (m3->mc_snum < csrc->mc_snum) continue;
7829 if (m3->mc_pg[csrc->mc_top] == psrc) {
7830 m3->mc_pg[csrc->mc_top] = pdst;
7831 m3->mc_ki[csrc->mc_top] += nkeys;
7836 unsigned int snum = cdst->mc_snum;
7837 uint16_t depth = cdst->mc_db->md_depth;
7838 mdb_cursor_pop(cdst);
7839 rc = mdb_rebalance(cdst);
7840 /* Did the tree shrink? */
7841 if (depth > cdst->mc_db->md_depth)
7843 cdst->mc_snum = snum;
7844 cdst->mc_top = snum-1;
7849 /** Copy the contents of a cursor.
7850 * @param[in] csrc The cursor to copy from.
7851 * @param[out] cdst The cursor to copy to.
7854 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7858 cdst->mc_txn = csrc->mc_txn;
7859 cdst->mc_dbi = csrc->mc_dbi;
7860 cdst->mc_db = csrc->mc_db;
7861 cdst->mc_dbx = csrc->mc_dbx;
7862 cdst->mc_snum = csrc->mc_snum;
7863 cdst->mc_top = csrc->mc_top;
7864 cdst->mc_flags = csrc->mc_flags;
7866 for (i=0; i<csrc->mc_snum; i++) {
7867 cdst->mc_pg[i] = csrc->mc_pg[i];
7868 cdst->mc_ki[i] = csrc->mc_ki[i];
7872 /** Rebalance the tree after a delete operation.
7873 * @param[in] mc Cursor pointing to the page where rebalancing
7875 * @return 0 on success, non-zero on failure.
7878 mdb_rebalance(MDB_cursor *mc)
7882 unsigned int ptop, minkeys;
7886 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7887 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7888 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7889 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7890 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7892 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7893 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7894 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7895 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7899 if (mc->mc_snum < 2) {
7900 MDB_page *mp = mc->mc_pg[0];
7902 DPUTS("Can't rebalance a subpage, ignoring");
7905 if (NUMKEYS(mp) == 0) {
7906 DPUTS("tree is completely empty");
7907 mc->mc_db->md_root = P_INVALID;
7908 mc->mc_db->md_depth = 0;
7909 mc->mc_db->md_leaf_pages = 0;
7910 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7913 /* Adjust cursors pointing to mp */
7916 mc->mc_flags &= ~C_INITIALIZED;
7918 MDB_cursor *m2, *m3;
7919 MDB_dbi dbi = mc->mc_dbi;
7921 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7922 if (mc->mc_flags & C_SUB)
7923 m3 = &m2->mc_xcursor->mx_cursor;
7926 if (m3->mc_snum < mc->mc_snum) continue;
7927 if (m3->mc_pg[0] == mp) {
7930 m3->mc_flags &= ~C_INITIALIZED;
7934 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7936 DPUTS("collapsing root page!");
7937 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7940 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7941 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7944 mc->mc_db->md_depth--;
7945 mc->mc_db->md_branch_pages--;
7946 mc->mc_ki[0] = mc->mc_ki[1];
7947 for (i = 1; i<mc->mc_db->md_depth; i++) {
7948 mc->mc_pg[i] = mc->mc_pg[i+1];
7949 mc->mc_ki[i] = mc->mc_ki[i+1];
7952 /* Adjust other cursors pointing to mp */
7953 MDB_cursor *m2, *m3;
7954 MDB_dbi dbi = mc->mc_dbi;
7956 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7957 if (mc->mc_flags & C_SUB)
7958 m3 = &m2->mc_xcursor->mx_cursor;
7961 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7962 if (m3->mc_pg[0] == mp) {
7963 for (i=0; i<m3->mc_snum; i++) {
7964 m3->mc_pg[i] = m3->mc_pg[i+1];
7965 m3->mc_ki[i] = m3->mc_ki[i+1];
7973 DPUTS("root page doesn't need rebalancing");
7977 /* The parent (branch page) must have at least 2 pointers,
7978 * otherwise the tree is invalid.
7980 ptop = mc->mc_top-1;
7981 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7983 /* Leaf page fill factor is below the threshold.
7984 * Try to move keys from left or right neighbor, or
7985 * merge with a neighbor page.
7990 mdb_cursor_copy(mc, &mn);
7991 mn.mc_xcursor = NULL;
7993 oldki = mc->mc_ki[mc->mc_top];
7994 if (mc->mc_ki[ptop] == 0) {
7995 /* We're the leftmost leaf in our parent.
7997 DPUTS("reading right neighbor");
7999 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8000 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8003 mn.mc_ki[mn.mc_top] = 0;
8004 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8006 /* There is at least one neighbor to the left.
8008 DPUTS("reading left neighbor");
8010 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8011 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8014 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8015 mc->mc_ki[mc->mc_top] = 0;
8018 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8019 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8020 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8022 /* If the neighbor page is above threshold and has enough keys,
8023 * move one key from it. Otherwise we should try to merge them.
8024 * (A branch page must never have less than 2 keys.)
8026 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
8027 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8028 rc = mdb_node_move(&mn, mc);
8029 if (mc->mc_ki[ptop]) {
8033 if (mc->mc_ki[ptop] == 0) {
8034 rc = mdb_page_merge(&mn, mc);
8037 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8038 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8039 /* We want mdb_rebalance to find mn when doing fixups */
8040 if (mc->mc_flags & C_SUB) {
8041 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8042 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8043 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8045 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8046 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8048 rc = mdb_page_merge(mc, &mn);
8049 if (mc->mc_flags & C_SUB)
8050 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8052 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8053 mdb_cursor_copy(&mn, mc);
8055 mc->mc_flags &= ~C_EOF;
8057 mc->mc_ki[mc->mc_top] = oldki;
8061 /** Complete a delete operation started by #mdb_cursor_del(). */
8063 mdb_cursor_del0(MDB_cursor *mc)
8070 ki = mc->mc_ki[mc->mc_top];
8071 mdb_node_del(mc, mc->mc_db->md_pad);
8072 mc->mc_db->md_entries--;
8073 rc = mdb_rebalance(mc);
8075 if (rc == MDB_SUCCESS) {
8076 MDB_cursor *m2, *m3;
8077 MDB_dbi dbi = mc->mc_dbi;
8079 /* DB is totally empty now, just bail out.
8080 * Other cursors adjustments were already done
8081 * by mdb_rebalance and aren't needed here.
8086 mp = mc->mc_pg[mc->mc_top];
8087 nkeys = NUMKEYS(mp);
8089 /* if mc points past last node in page, find next sibling */
8090 if (mc->mc_ki[mc->mc_top] >= nkeys) {
8091 rc = mdb_cursor_sibling(mc, 1);
8092 if (rc == MDB_NOTFOUND) {
8093 mc->mc_flags |= C_EOF;
8098 /* Adjust other cursors pointing to mp */
8099 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8100 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8101 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8103 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8105 if (m3->mc_pg[mc->mc_top] == mp) {
8106 if (m3->mc_ki[mc->mc_top] >= ki) {
8107 m3->mc_flags |= C_DEL;
8108 if (m3->mc_ki[mc->mc_top] > ki)
8109 m3->mc_ki[mc->mc_top]--;
8110 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8111 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8113 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8114 rc = mdb_cursor_sibling(m3, 1);
8115 if (rc == MDB_NOTFOUND) {
8116 m3->mc_flags |= C_EOF;
8122 mc->mc_flags |= C_DEL;
8126 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8131 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8132 MDB_val *key, MDB_val *data)
8134 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8137 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8138 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8140 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8141 /* must ignore any data */
8145 return mdb_del0(txn, dbi, key, data, 0);
8149 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8150 MDB_val *key, MDB_val *data, unsigned flags)
8155 MDB_val rdata, *xdata;
8159 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8161 mdb_cursor_init(&mc, txn, dbi, &mx);
8170 flags |= MDB_NODUPDATA;
8172 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8174 /* let mdb_page_split know about this cursor if needed:
8175 * delete will trigger a rebalance; if it needs to move
8176 * a node from one page to another, it will have to
8177 * update the parent's separator key(s). If the new sepkey
8178 * is larger than the current one, the parent page may
8179 * run out of space, triggering a split. We need this
8180 * cursor to be consistent until the end of the rebalance.
8182 mc.mc_flags |= C_UNTRACK;
8183 mc.mc_next = txn->mt_cursors[dbi];
8184 txn->mt_cursors[dbi] = &mc;
8185 rc = mdb_cursor_del(&mc, flags);
8186 txn->mt_cursors[dbi] = mc.mc_next;
8191 /** Split a page and insert a new node.
8192 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8193 * The cursor will be updated to point to the actual page and index where
8194 * the node got inserted after the split.
8195 * @param[in] newkey The key for the newly inserted node.
8196 * @param[in] newdata The data for the newly inserted node.
8197 * @param[in] newpgno The page number, if the new node is a branch node.
8198 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8199 * @return 0 on success, non-zero on failure.
8202 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8203 unsigned int nflags)
8206 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8209 int i, j, split_indx, nkeys, pmax;
8210 MDB_env *env = mc->mc_txn->mt_env;
8212 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8213 MDB_page *copy = NULL;
8214 MDB_page *mp, *rp, *pp;
8219 mp = mc->mc_pg[mc->mc_top];
8220 newindx = mc->mc_ki[mc->mc_top];
8221 nkeys = NUMKEYS(mp);
8223 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8224 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8225 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8227 /* Create a right sibling. */
8228 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8230 rp->mp_pad = mp->mp_pad;
8231 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8233 if (mc->mc_snum < 2) {
8234 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8236 /* shift current top to make room for new parent */
8237 mc->mc_pg[1] = mc->mc_pg[0];
8238 mc->mc_ki[1] = mc->mc_ki[0];
8241 mc->mc_db->md_root = pp->mp_pgno;
8242 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8243 mc->mc_db->md_depth++;
8246 /* Add left (implicit) pointer. */
8247 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8248 /* undo the pre-push */
8249 mc->mc_pg[0] = mc->mc_pg[1];
8250 mc->mc_ki[0] = mc->mc_ki[1];
8251 mc->mc_db->md_root = mp->mp_pgno;
8252 mc->mc_db->md_depth--;
8259 ptop = mc->mc_top-1;
8260 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8263 mc->mc_flags |= C_SPLITTING;
8264 mdb_cursor_copy(mc, &mn);
8265 mn.mc_pg[mn.mc_top] = rp;
8266 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8268 if (nflags & MDB_APPEND) {
8269 mn.mc_ki[mn.mc_top] = 0;
8271 split_indx = newindx;
8275 split_indx = (nkeys+1) / 2;
8280 unsigned int lsize, rsize, ksize;
8281 /* Move half of the keys to the right sibling */
8282 x = mc->mc_ki[mc->mc_top] - split_indx;
8283 ksize = mc->mc_db->md_pad;
8284 split = LEAF2KEY(mp, split_indx, ksize);
8285 rsize = (nkeys - split_indx) * ksize;
8286 lsize = (nkeys - split_indx) * sizeof(indx_t);
8287 mp->mp_lower -= lsize;
8288 rp->mp_lower += lsize;
8289 mp->mp_upper += rsize - lsize;
8290 rp->mp_upper -= rsize - lsize;
8291 sepkey.mv_size = ksize;
8292 if (newindx == split_indx) {
8293 sepkey.mv_data = newkey->mv_data;
8295 sepkey.mv_data = split;
8298 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8299 memcpy(rp->mp_ptrs, split, rsize);
8300 sepkey.mv_data = rp->mp_ptrs;
8301 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8302 memcpy(ins, newkey->mv_data, ksize);
8303 mp->mp_lower += sizeof(indx_t);
8304 mp->mp_upper -= ksize - sizeof(indx_t);
8307 memcpy(rp->mp_ptrs, split, x * ksize);
8308 ins = LEAF2KEY(rp, x, ksize);
8309 memcpy(ins, newkey->mv_data, ksize);
8310 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8311 rp->mp_lower += sizeof(indx_t);
8312 rp->mp_upper -= ksize - sizeof(indx_t);
8313 mc->mc_ki[mc->mc_top] = x;
8314 mc->mc_pg[mc->mc_top] = rp;
8317 int psize, nsize, k;
8318 /* Maximum free space in an empty page */
8319 pmax = env->me_psize - PAGEHDRSZ;
8321 nsize = mdb_leaf_size(env, newkey, newdata);
8323 nsize = mdb_branch_size(env, newkey);
8324 nsize = EVEN(nsize);
8326 /* grab a page to hold a temporary copy */
8327 copy = mdb_page_malloc(mc->mc_txn, 1);
8332 copy->mp_pgno = mp->mp_pgno;
8333 copy->mp_flags = mp->mp_flags;
8334 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8335 copy->mp_upper = env->me_psize - PAGEBASE;
8337 /* prepare to insert */
8338 for (i=0, j=0; i<nkeys; i++) {
8340 copy->mp_ptrs[j++] = 0;
8342 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8345 /* When items are relatively large the split point needs
8346 * to be checked, because being off-by-one will make the
8347 * difference between success or failure in mdb_node_add.
8349 * It's also relevant if a page happens to be laid out
8350 * such that one half of its nodes are all "small" and
8351 * the other half of its nodes are "large." If the new
8352 * item is also "large" and falls on the half with
8353 * "large" nodes, it also may not fit.
8355 * As a final tweak, if the new item goes on the last
8356 * spot on the page (and thus, onto the new page), bias
8357 * the split so the new page is emptier than the old page.
8358 * This yields better packing during sequential inserts.
8360 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8361 /* Find split point */
8363 if (newindx <= split_indx || newindx >= nkeys) {
8365 k = newindx >= nkeys ? nkeys : split_indx+2;
8370 for (; i!=k; i+=j) {
8375 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8376 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8378 if (F_ISSET(node->mn_flags, F_BIGDATA))
8379 psize += sizeof(pgno_t);
8381 psize += NODEDSZ(node);
8383 psize = EVEN(psize);
8385 if (psize > pmax || i == k-j) {
8386 split_indx = i + (j<0);
8391 if (split_indx == newindx) {
8392 sepkey.mv_size = newkey->mv_size;
8393 sepkey.mv_data = newkey->mv_data;
8395 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8396 sepkey.mv_size = node->mn_ksize;
8397 sepkey.mv_data = NODEKEY(node);
8402 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8404 /* Copy separator key to the parent.
8406 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8410 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8415 if (mn.mc_snum == mc->mc_snum) {
8416 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8417 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8418 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8419 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8424 /* Right page might now have changed parent.
8425 * Check if left page also changed parent.
8427 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8428 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8429 for (i=0; i<ptop; i++) {
8430 mc->mc_pg[i] = mn.mc_pg[i];
8431 mc->mc_ki[i] = mn.mc_ki[i];
8433 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8434 if (mn.mc_ki[ptop]) {
8435 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8437 /* find right page's left sibling */
8438 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8439 mdb_cursor_sibling(mc, 0);
8444 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8447 mc->mc_flags ^= C_SPLITTING;
8448 if (rc != MDB_SUCCESS) {
8451 if (nflags & MDB_APPEND) {
8452 mc->mc_pg[mc->mc_top] = rp;
8453 mc->mc_ki[mc->mc_top] = 0;
8454 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8457 for (i=0; i<mc->mc_top; i++)
8458 mc->mc_ki[i] = mn.mc_ki[i];
8459 } else if (!IS_LEAF2(mp)) {
8461 mc->mc_pg[mc->mc_top] = rp;
8466 rkey.mv_data = newkey->mv_data;
8467 rkey.mv_size = newkey->mv_size;
8473 /* Update index for the new key. */
8474 mc->mc_ki[mc->mc_top] = j;
8476 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8477 rkey.mv_data = NODEKEY(node);
8478 rkey.mv_size = node->mn_ksize;
8480 xdata.mv_data = NODEDATA(node);
8481 xdata.mv_size = NODEDSZ(node);
8484 pgno = NODEPGNO(node);
8485 flags = node->mn_flags;
8488 if (!IS_LEAF(mp) && j == 0) {
8489 /* First branch index doesn't need key data. */
8493 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8499 mc->mc_pg[mc->mc_top] = copy;
8504 } while (i != split_indx);
8506 nkeys = NUMKEYS(copy);
8507 for (i=0; i<nkeys; i++)
8508 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8509 mp->mp_lower = copy->mp_lower;
8510 mp->mp_upper = copy->mp_upper;
8511 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8512 env->me_psize - copy->mp_upper - PAGEBASE);
8514 /* reset back to original page */
8515 if (newindx < split_indx) {
8516 mc->mc_pg[mc->mc_top] = mp;
8517 if (nflags & MDB_RESERVE) {
8518 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8519 if (!(node->mn_flags & F_BIGDATA))
8520 newdata->mv_data = NODEDATA(node);
8523 mc->mc_pg[mc->mc_top] = rp;
8525 /* Make sure mc_ki is still valid.
8527 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8528 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8529 for (i=0; i<=ptop; i++) {
8530 mc->mc_pg[i] = mn.mc_pg[i];
8531 mc->mc_ki[i] = mn.mc_ki[i];
8538 /* Adjust other cursors pointing to mp */
8539 MDB_cursor *m2, *m3;
8540 MDB_dbi dbi = mc->mc_dbi;
8541 int fixup = NUMKEYS(mp);
8543 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8544 if (mc->mc_flags & C_SUB)
8545 m3 = &m2->mc_xcursor->mx_cursor;
8550 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8552 if (m3->mc_flags & C_SPLITTING)
8557 for (k=m3->mc_top; k>=0; k--) {
8558 m3->mc_ki[k+1] = m3->mc_ki[k];
8559 m3->mc_pg[k+1] = m3->mc_pg[k];
8561 if (m3->mc_ki[0] >= split_indx) {
8566 m3->mc_pg[0] = mc->mc_pg[0];
8570 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8571 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8572 m3->mc_ki[mc->mc_top]++;
8573 if (m3->mc_ki[mc->mc_top] >= fixup) {
8574 m3->mc_pg[mc->mc_top] = rp;
8575 m3->mc_ki[mc->mc_top] -= fixup;
8576 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8578 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8579 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8584 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8587 if (copy) /* tmp page */
8588 mdb_page_free(env, copy);
8590 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8595 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8596 MDB_val *key, MDB_val *data, unsigned int flags)
8601 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8604 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8607 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8608 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8610 mdb_cursor_init(&mc, txn, dbi, &mx);
8611 return mdb_cursor_put(&mc, key, data, flags);
8615 #define MDB_WBUF (1024*1024)
8618 /** State needed for a compacting copy. */
8619 typedef struct mdb_copy {
8620 pthread_mutex_t mc_mutex;
8621 pthread_cond_t mc_cond;
8628 pgno_t mc_next_pgno;
8631 volatile int mc_new;
8636 /** Dedicated writer thread for compacting copy. */
8637 static THREAD_RET ESECT
8638 mdb_env_copythr(void *arg)
8642 int toggle = 0, wsize, rc;
8645 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8648 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8651 pthread_mutex_lock(&my->mc_mutex);
8653 pthread_cond_signal(&my->mc_cond);
8656 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8657 if (my->mc_new < 0) {
8662 wsize = my->mc_wlen[toggle];
8663 ptr = my->mc_wbuf[toggle];
8666 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8670 } else if (len > 0) {
8684 /* If there's an overflow page tail, write it too */
8685 if (my->mc_olen[toggle]) {
8686 wsize = my->mc_olen[toggle];
8687 ptr = my->mc_over[toggle];
8688 my->mc_olen[toggle] = 0;
8691 my->mc_wlen[toggle] = 0;
8693 pthread_cond_signal(&my->mc_cond);
8695 pthread_cond_signal(&my->mc_cond);
8696 pthread_mutex_unlock(&my->mc_mutex);
8697 return (THREAD_RET)0;
8701 /** Tell the writer thread there's a buffer ready to write */
8703 mdb_env_cthr_toggle(mdb_copy *my, int st)
8705 int toggle = my->mc_toggle ^ 1;
8706 pthread_mutex_lock(&my->mc_mutex);
8707 if (my->mc_status) {
8708 pthread_mutex_unlock(&my->mc_mutex);
8709 return my->mc_status;
8711 while (my->mc_new == 1)
8712 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8714 my->mc_toggle = toggle;
8715 pthread_cond_signal(&my->mc_cond);
8716 pthread_mutex_unlock(&my->mc_mutex);
8720 /** Depth-first tree traversal for compacting copy. */
8722 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8725 MDB_txn *txn = my->mc_txn;
8727 MDB_page *mo, *mp, *leaf;
8732 /* Empty DB, nothing to do */
8733 if (*pg == P_INVALID)
8740 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8743 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8747 /* Make cursor pages writable */
8748 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8752 for (i=0; i<mc.mc_top; i++) {
8753 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8754 mc.mc_pg[i] = (MDB_page *)ptr;
8755 ptr += my->mc_env->me_psize;
8758 /* This is writable space for a leaf page. Usually not needed. */
8759 leaf = (MDB_page *)ptr;
8761 toggle = my->mc_toggle;
8762 while (mc.mc_snum > 0) {
8764 mp = mc.mc_pg[mc.mc_top];
8768 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8769 for (i=0; i<n; i++) {
8770 ni = NODEPTR(mp, i);
8771 if (ni->mn_flags & F_BIGDATA) {
8775 /* Need writable leaf */
8777 mc.mc_pg[mc.mc_top] = leaf;
8778 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8780 ni = NODEPTR(mp, i);
8783 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8784 rc = mdb_page_get(txn, pg, &omp, NULL);
8787 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8788 rc = mdb_env_cthr_toggle(my, 1);
8791 toggle = my->mc_toggle;
8793 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8794 memcpy(mo, omp, my->mc_env->me_psize);
8795 mo->mp_pgno = my->mc_next_pgno;
8796 my->mc_next_pgno += omp->mp_pages;
8797 my->mc_wlen[toggle] += my->mc_env->me_psize;
8798 if (omp->mp_pages > 1) {
8799 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8800 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8801 rc = mdb_env_cthr_toggle(my, 1);
8804 toggle = my->mc_toggle;
8806 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8807 } else if (ni->mn_flags & F_SUBDATA) {
8810 /* Need writable leaf */
8812 mc.mc_pg[mc.mc_top] = leaf;
8813 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8815 ni = NODEPTR(mp, i);
8818 memcpy(&db, NODEDATA(ni), sizeof(db));
8819 my->mc_toggle = toggle;
8820 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8823 toggle = my->mc_toggle;
8824 memcpy(NODEDATA(ni), &db, sizeof(db));
8829 mc.mc_ki[mc.mc_top]++;
8830 if (mc.mc_ki[mc.mc_top] < n) {
8833 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8835 rc = mdb_page_get(txn, pg, &mp, NULL);
8840 mc.mc_ki[mc.mc_top] = 0;
8841 if (IS_BRANCH(mp)) {
8842 /* Whenever we advance to a sibling branch page,
8843 * we must proceed all the way down to its first leaf.
8845 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8848 mc.mc_pg[mc.mc_top] = mp;
8852 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8853 rc = mdb_env_cthr_toggle(my, 1);
8856 toggle = my->mc_toggle;
8858 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8859 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8860 mo->mp_pgno = my->mc_next_pgno++;
8861 my->mc_wlen[toggle] += my->mc_env->me_psize;
8863 /* Update parent if there is one */
8864 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8865 SETPGNO(ni, mo->mp_pgno);
8866 mdb_cursor_pop(&mc);
8868 /* Otherwise we're done */
8878 /** Copy environment with compaction. */
8880 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8885 MDB_txn *txn = NULL;
8890 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8891 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8892 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8893 if (my.mc_wbuf[0] == NULL)
8896 pthread_mutex_init(&my.mc_mutex, NULL);
8897 pthread_cond_init(&my.mc_cond, NULL);
8898 #ifdef HAVE_MEMALIGN
8899 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8900 if (my.mc_wbuf[0] == NULL)
8903 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8908 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8909 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8914 my.mc_next_pgno = NUM_METAS;
8920 THREAD_CREATE(thr, mdb_env_copythr, &my);
8922 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8926 mp = (MDB_page *)my.mc_wbuf[0];
8927 memset(mp, 0, NUM_METAS * env->me_psize);
8929 mp->mp_flags = P_META;
8930 mm = (MDB_meta *)METADATA(mp);
8931 mdb_env_init_meta0(env, mm);
8932 mm->mm_address = env->me_metas[0]->mm_address;
8934 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8936 mp->mp_flags = P_META;
8937 *(MDB_meta *)METADATA(mp) = *mm;
8938 mm = (MDB_meta *)METADATA(mp);
8940 /* Count the number of free pages, subtract from lastpg to find
8941 * number of active pages
8944 MDB_ID freecount = 0;
8947 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8948 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8949 freecount += *(MDB_ID *)data.mv_data;
8950 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
8951 txn->mt_dbs[FREE_DBI].md_leaf_pages +
8952 txn->mt_dbs[FREE_DBI].md_overflow_pages;
8954 /* Set metapage 1 */
8955 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8956 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
8957 if (mm->mm_last_pg > NUM_METAS-1) {
8958 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
8961 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
8964 my.mc_wlen[0] = env->me_psize * NUM_METAS;
8966 pthread_mutex_lock(&my.mc_mutex);
8968 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8969 pthread_mutex_unlock(&my.mc_mutex);
8970 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
8971 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8972 rc = mdb_env_cthr_toggle(&my, 1);
8973 mdb_env_cthr_toggle(&my, -1);
8974 pthread_mutex_lock(&my.mc_mutex);
8976 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8977 pthread_mutex_unlock(&my.mc_mutex);
8982 CloseHandle(my.mc_cond);
8983 CloseHandle(my.mc_mutex);
8984 _aligned_free(my.mc_wbuf[0]);
8986 pthread_cond_destroy(&my.mc_cond);
8987 pthread_mutex_destroy(&my.mc_mutex);
8988 free(my.mc_wbuf[0]);
8993 /** Copy environment as-is. */
8995 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8997 MDB_txn *txn = NULL;
8998 mdb_mutexref_t wmutex = NULL;
9004 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9008 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9011 /* Do the lock/unlock of the reader mutex before starting the
9012 * write txn. Otherwise other read txns could block writers.
9014 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9019 /* We must start the actual read txn after blocking writers */
9020 mdb_txn_end(txn, MDB_END_RESET_TMP);
9022 /* Temporarily block writers until we snapshot the meta pages */
9023 wmutex = env->me_wmutex;
9024 if (LOCK_MUTEX(rc, env, wmutex))
9027 rc = mdb_txn_renew0(txn);
9029 UNLOCK_MUTEX(wmutex);
9034 wsize = env->me_psize * NUM_METAS;
9038 DO_WRITE(rc, fd, ptr, w2, len);
9042 } else if (len > 0) {
9048 /* Non-blocking or async handles are not supported */
9054 UNLOCK_MUTEX(wmutex);
9059 w2 = txn->mt_next_pgno * env->me_psize;
9062 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9069 if (wsize > MAX_WRITE)
9073 DO_WRITE(rc, fd, ptr, w2, len);
9077 } else if (len > 0) {
9094 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9096 if (flags & MDB_CP_COMPACT)
9097 return mdb_env_copyfd1(env, fd);
9099 return mdb_env_copyfd0(env, fd);
9103 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9105 return mdb_env_copyfd2(env, fd, 0);
9109 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9113 HANDLE newfd = INVALID_HANDLE_VALUE;
9115 if (env->me_flags & MDB_NOSUBDIR) {
9116 lpath = (char *)path;
9119 len += sizeof(DATANAME);
9120 lpath = malloc(len);
9123 sprintf(lpath, "%s" DATANAME, path);
9126 /* The destination path must exist, but the destination file must not.
9127 * We don't want the OS to cache the writes, since the source data is
9128 * already in the OS cache.
9131 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9132 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9134 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9136 if (newfd == INVALID_HANDLE_VALUE) {
9141 if (env->me_psize >= env->me_os_psize) {
9143 /* Set O_DIRECT if the file system supports it */
9144 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9145 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9147 #ifdef F_NOCACHE /* __APPLE__ */
9148 rc = fcntl(newfd, F_NOCACHE, 1);
9156 rc = mdb_env_copyfd2(env, newfd, flags);
9159 if (!(env->me_flags & MDB_NOSUBDIR))
9161 if (newfd != INVALID_HANDLE_VALUE)
9162 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9169 mdb_env_copy(MDB_env *env, const char *path)
9171 return mdb_env_copy2(env, path, 0);
9175 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9177 if (flag & ~CHANGEABLE)
9180 env->me_flags |= flag;
9182 env->me_flags &= ~flag;
9187 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9192 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9197 mdb_env_set_userctx(MDB_env *env, void *ctx)
9201 env->me_userctx = ctx;
9206 mdb_env_get_userctx(MDB_env *env)
9208 return env ? env->me_userctx : NULL;
9212 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9217 env->me_assert_func = func;
9223 mdb_env_get_path(MDB_env *env, const char **arg)
9228 *arg = env->me_path;
9233 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9242 /** Common code for #mdb_stat() and #mdb_env_stat().
9243 * @param[in] env the environment to operate in.
9244 * @param[in] db the #MDB_db record containing the stats to return.
9245 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9246 * @return 0, this function always succeeds.
9249 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9251 arg->ms_psize = env->me_psize;
9252 arg->ms_depth = db->md_depth;
9253 arg->ms_branch_pages = db->md_branch_pages;
9254 arg->ms_leaf_pages = db->md_leaf_pages;
9255 arg->ms_overflow_pages = db->md_overflow_pages;
9256 arg->ms_entries = db->md_entries;
9262 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9266 if (env == NULL || arg == NULL)
9269 meta = mdb_env_pick_meta(env);
9271 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9275 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9279 if (env == NULL || arg == NULL)
9282 meta = mdb_env_pick_meta(env);
9283 arg->me_mapaddr = meta->mm_address;
9284 arg->me_last_pgno = meta->mm_last_pg;
9285 arg->me_last_txnid = meta->mm_txnid;
9287 arg->me_mapsize = env->me_mapsize;
9288 arg->me_maxreaders = env->me_maxreaders;
9289 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9293 /** Set the default comparison functions for a database.
9294 * Called immediately after a database is opened to set the defaults.
9295 * The user can then override them with #mdb_set_compare() or
9296 * #mdb_set_dupsort().
9297 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9298 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9301 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9303 uint16_t f = txn->mt_dbs[dbi].md_flags;
9305 txn->mt_dbxs[dbi].md_cmp =
9306 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9307 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9309 txn->mt_dbxs[dbi].md_dcmp =
9310 !(f & MDB_DUPSORT) ? 0 :
9311 ((f & MDB_INTEGERDUP)
9312 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9313 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9316 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9322 int rc, dbflag, exact;
9323 unsigned int unused = 0, seq;
9326 if (flags & ~VALID_FLAGS)
9328 if (txn->mt_flags & MDB_TXN_BLOCKED)
9334 if (flags & PERSISTENT_FLAGS) {
9335 uint16_t f2 = flags & PERSISTENT_FLAGS;
9336 /* make sure flag changes get committed */
9337 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9338 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9339 txn->mt_flags |= MDB_TXN_DIRTY;
9342 mdb_default_cmp(txn, MAIN_DBI);
9346 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9347 mdb_default_cmp(txn, MAIN_DBI);
9350 /* Is the DB already open? */
9352 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9353 if (!txn->mt_dbxs[i].md_name.mv_size) {
9354 /* Remember this free slot */
9355 if (!unused) unused = i;
9358 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9359 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9365 /* If no free slot and max hit, fail */
9366 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9367 return MDB_DBS_FULL;
9369 /* Cannot mix named databases with some mainDB flags */
9370 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9371 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9373 /* Find the DB info */
9374 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9377 key.mv_data = (void *)name;
9378 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9379 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9380 if (rc == MDB_SUCCESS) {
9381 /* make sure this is actually a DB */
9382 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9383 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9384 return MDB_INCOMPATIBLE;
9385 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9386 /* Create if requested */
9387 data.mv_size = sizeof(MDB_db);
9388 data.mv_data = &dummy;
9389 memset(&dummy, 0, sizeof(dummy));
9390 dummy.md_root = P_INVALID;
9391 dummy.md_flags = flags & PERSISTENT_FLAGS;
9392 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9396 /* OK, got info, add to table */
9397 if (rc == MDB_SUCCESS) {
9398 unsigned int slot = unused ? unused : txn->mt_numdbs;
9399 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9400 txn->mt_dbxs[slot].md_name.mv_size = len;
9401 txn->mt_dbxs[slot].md_rel = NULL;
9402 txn->mt_dbflags[slot] = dbflag;
9403 /* txn-> and env-> are the same in read txns, use
9404 * tmp variable to avoid undefined assignment
9406 seq = ++txn->mt_env->me_dbiseqs[slot];
9407 txn->mt_dbiseqs[slot] = seq;
9409 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9411 mdb_default_cmp(txn, slot);
9421 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9423 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9426 if (txn->mt_flags & MDB_TXN_BLOCKED)
9429 if (txn->mt_dbflags[dbi] & DB_STALE) {
9432 /* Stale, must read the DB's root. cursor_init does it for us. */
9433 mdb_cursor_init(&mc, txn, dbi, &mx);
9435 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9438 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9441 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9443 ptr = env->me_dbxs[dbi].md_name.mv_data;
9444 /* If there was no name, this was already closed */
9446 env->me_dbxs[dbi].md_name.mv_data = NULL;
9447 env->me_dbxs[dbi].md_name.mv_size = 0;
9448 env->me_dbflags[dbi] = 0;
9449 env->me_dbiseqs[dbi]++;
9454 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9456 /* We could return the flags for the FREE_DBI too but what's the point? */
9457 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9459 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9463 /** Add all the DB's pages to the free list.
9464 * @param[in] mc Cursor on the DB to free.
9465 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9466 * @return 0 on success, non-zero on failure.
9469 mdb_drop0(MDB_cursor *mc, int subs)
9473 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9474 if (rc == MDB_SUCCESS) {
9475 MDB_txn *txn = mc->mc_txn;
9480 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9481 * This also avoids any P_LEAF2 pages, which have no nodes.
9483 if (mc->mc_flags & C_SUB)
9486 mdb_cursor_copy(mc, &mx);
9487 while (mc->mc_snum > 0) {
9488 MDB_page *mp = mc->mc_pg[mc->mc_top];
9489 unsigned n = NUMKEYS(mp);
9491 for (i=0; i<n; i++) {
9492 ni = NODEPTR(mp, i);
9493 if (ni->mn_flags & F_BIGDATA) {
9496 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9497 rc = mdb_page_get(txn, pg, &omp, NULL);
9500 mdb_cassert(mc, IS_OVERFLOW(omp));
9501 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9505 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9506 mdb_xcursor_init1(mc, ni);
9507 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9513 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9515 for (i=0; i<n; i++) {
9517 ni = NODEPTR(mp, i);
9520 mdb_midl_xappend(txn->mt_free_pgs, pg);
9525 mc->mc_ki[mc->mc_top] = i;
9526 rc = mdb_cursor_sibling(mc, 1);
9528 if (rc != MDB_NOTFOUND)
9530 /* no more siblings, go back to beginning
9531 * of previous level.
9535 for (i=1; i<mc->mc_snum; i++) {
9537 mc->mc_pg[i] = mx.mc_pg[i];
9542 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9545 txn->mt_flags |= MDB_TXN_ERROR;
9546 } else if (rc == MDB_NOTFOUND) {
9552 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9554 MDB_cursor *mc, *m2;
9557 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9560 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9563 if (TXN_DBI_CHANGED(txn, dbi))
9566 rc = mdb_cursor_open(txn, dbi, &mc);
9570 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9571 /* Invalidate the dropped DB's cursors */
9572 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9573 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9577 /* Can't delete the main DB */
9578 if (del && dbi >= CORE_DBS) {
9579 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9581 txn->mt_dbflags[dbi] = DB_STALE;
9582 mdb_dbi_close(txn->mt_env, dbi);
9584 txn->mt_flags |= MDB_TXN_ERROR;
9587 /* reset the DB record, mark it dirty */
9588 txn->mt_dbflags[dbi] |= DB_DIRTY;
9589 txn->mt_dbs[dbi].md_depth = 0;
9590 txn->mt_dbs[dbi].md_branch_pages = 0;
9591 txn->mt_dbs[dbi].md_leaf_pages = 0;
9592 txn->mt_dbs[dbi].md_overflow_pages = 0;
9593 txn->mt_dbs[dbi].md_entries = 0;
9594 txn->mt_dbs[dbi].md_root = P_INVALID;
9596 txn->mt_flags |= MDB_TXN_DIRTY;
9599 mdb_cursor_close(mc);
9603 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9605 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9608 txn->mt_dbxs[dbi].md_cmp = cmp;
9612 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9614 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9617 txn->mt_dbxs[dbi].md_dcmp = cmp;
9621 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9623 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9626 txn->mt_dbxs[dbi].md_rel = rel;
9630 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9632 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9635 txn->mt_dbxs[dbi].md_relctx = ctx;
9640 mdb_env_get_maxkeysize(MDB_env *env)
9642 return ENV_MAXKEY(env);
9646 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9648 unsigned int i, rdrs;
9651 int rc = 0, first = 1;
9655 if (!env->me_txns) {
9656 return func("(no reader locks)\n", ctx);
9658 rdrs = env->me_txns->mti_numreaders;
9659 mr = env->me_txns->mti_readers;
9660 for (i=0; i<rdrs; i++) {
9662 txnid_t txnid = mr[i].mr_txnid;
9663 sprintf(buf, txnid == (txnid_t)-1 ?
9664 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9665 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9668 rc = func(" pid thread txnid\n", ctx);
9672 rc = func(buf, ctx);
9678 rc = func("(no active readers)\n", ctx);
9683 /** Insert pid into list if not already present.
9684 * return -1 if already present.
9687 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9689 /* binary search of pid in list */
9691 unsigned cursor = 1;
9693 unsigned n = ids[0];
9696 unsigned pivot = n >> 1;
9697 cursor = base + pivot + 1;
9698 val = pid - ids[cursor];
9703 } else if ( val > 0 ) {
9708 /* found, so it's a duplicate */
9717 for (n = ids[0]; n > cursor; n--)
9724 mdb_reader_check(MDB_env *env, int *dead)
9730 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9733 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9735 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9737 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9738 unsigned int i, j, rdrs;
9740 MDB_PID_T *pids, pid;
9741 int rc = MDB_SUCCESS, count = 0;
9743 rdrs = env->me_txns->mti_numreaders;
9744 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9748 mr = env->me_txns->mti_readers;
9749 for (i=0; i<rdrs; i++) {
9751 if (pid && pid != env->me_pid) {
9752 if (mdb_pid_insert(pids, pid) == 0) {
9753 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9754 /* Stale reader found */
9757 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9758 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9760 rdrs = 0; /* the above checked all readers */
9762 /* Recheck, a new process may have reused pid */
9763 if (mdb_reader_pid(env, Pidcheck, pid))
9768 if (mr[j].mr_pid == pid) {
9769 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9770 (unsigned) pid, mr[j].mr_txnid));
9775 UNLOCK_MUTEX(rmutex);
9786 #ifdef MDB_ROBUST_SUPPORTED
9787 /** Handle #LOCK_MUTEX0() failure.
9788 * Try to repair the lock file if the mutex owner died.
9789 * @param[in] env the environment handle
9790 * @param[in] mutex LOCK_MUTEX0() mutex
9791 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9792 * @return 0 on success with the mutex locked, or an error code on failure.
9795 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9800 if (rc == MDB_OWNERDEAD) {
9801 /* We own the mutex. Clean up after dead previous owner. */
9803 rlocked = (mutex == env->me_rmutex);
9805 /* Keep mti_txnid updated, otherwise next writer can
9806 * overwrite data which latest meta page refers to.
9808 meta = mdb_env_pick_meta(env);
9809 env->me_txns->mti_txnid = meta->mm_txnid;
9810 /* env is hosed if the dead thread was ours */
9812 env->me_flags |= MDB_FATAL_ERROR;
9817 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9818 (rc ? "this process' env is hosed" : "recovering")));
9819 rc2 = mdb_reader_check0(env, rlocked, NULL);
9821 rc2 = mdb_mutex_consistent(mutex);
9822 if (rc || (rc = rc2)) {
9823 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9824 UNLOCK_MUTEX(mutex);
9830 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9835 #endif /* MDB_ROBUST_SUPPORTED */
projects / platform / upstream / lmdb.git / blob