2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2015 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
82 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
83 /** fdatasync is broken on ext3/ext4fs on older kernels, see
84 * description in #mdb_env_open2 comments. You can safely
85 * define MDB_FDATASYNC_WORKS if this code will only be run
86 * on kernels 3.6 and newer.
88 #define BROKEN_FDATASYNC
101 #if defined(__sun) || defined(ANDROID)
102 /* Most platforms have posix_memalign, older may only have memalign */
103 #define HAVE_MEMALIGN 1
107 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
108 #include <netinet/in.h>
109 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
112 #if defined(__APPLE__) || defined (BSD)
113 # if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
114 # define MDB_USE_SYSV_SEM 1
116 # define MDB_FDATASYNC fsync
117 #elif defined(ANDROID)
118 # define MDB_FDATASYNC fsync
123 #ifdef MDB_USE_POSIX_SEM
124 # define MDB_USE_HASH 1
125 #include <semaphore.h>
126 #elif defined(MDB_USE_SYSV_SEM)
129 #ifdef _SEM_SEMUN_UNDEFINED
132 struct semid_ds *buf;
133 unsigned short *array;
135 #endif /* _SEM_SEMUN_UNDEFINED */
137 #define MDB_USE_POSIX_MUTEX 1
138 #endif /* MDB_USE_POSIX_SEM */
141 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
142 + defined(MDB_USE_POSIX_MUTEX) != 1
143 # error "Ambiguous shared-lock implementation"
147 #include <valgrind/memcheck.h>
148 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
149 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
150 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
151 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
152 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
154 #define VGMEMP_CREATE(h,r,z)
155 #define VGMEMP_ALLOC(h,a,s)
156 #define VGMEMP_FREE(h,a)
157 #define VGMEMP_DESTROY(h)
158 #define VGMEMP_DEFINED(a,s)
162 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
163 /* Solaris just defines one or the other */
164 # define LITTLE_ENDIAN 1234
165 # define BIG_ENDIAN 4321
166 # ifdef _LITTLE_ENDIAN
167 # define BYTE_ORDER LITTLE_ENDIAN
169 # define BYTE_ORDER BIG_ENDIAN
172 # define BYTE_ORDER __BYTE_ORDER
176 #ifndef LITTLE_ENDIAN
177 #define LITTLE_ENDIAN __LITTLE_ENDIAN
180 #define BIG_ENDIAN __BIG_ENDIAN
183 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
184 #define MISALIGNED_OK 1
190 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
191 # error "Unknown or unsupported endianness (BYTE_ORDER)"
192 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
193 # error "Two's complement, reasonably sized integer types, please"
197 /** Put infrequently used env functions in separate section */
199 # define ESECT __attribute__ ((section("__TEXT,text_env")))
201 # define ESECT __attribute__ ((section("text_env")))
207 /** @defgroup internal LMDB Internals
210 /** @defgroup compat Compatibility Macros
211 * A bunch of macros to minimize the amount of platform-specific ifdefs
212 * needed throughout the rest of the code. When the features this library
213 * needs are similar enough to POSIX to be hidden in a one-or-two line
214 * replacement, this macro approach is used.
218 /** Features under development */
223 /** Wrapper around __func__, which is a C99 feature */
224 #if __STDC_VERSION__ >= 199901L
225 # define mdb_func_ __func__
226 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
227 # define mdb_func_ __FUNCTION__
229 /* If a debug message says <mdb_unknown>(), update the #if statements above */
230 # define mdb_func_ "<mdb_unknown>"
233 /* Internal error codes, not exposed outside liblmdb */
234 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
236 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
237 #elif defined MDB_USE_SYSV_SEM
238 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
239 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
240 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
244 #define MDB_ROBUST_SUPPORTED 1
248 #define MDB_USE_HASH 1
249 #define MDB_PIDLOCK 0
250 #define THREAD_RET DWORD
251 #define pthread_t HANDLE
252 #define pthread_mutex_t HANDLE
253 #define pthread_cond_t HANDLE
254 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
255 #define pthread_key_t DWORD
256 #define pthread_self() GetCurrentThreadId()
257 #define pthread_key_create(x,y) \
258 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
259 #define pthread_key_delete(x) TlsFree(x)
260 #define pthread_getspecific(x) TlsGetValue(x)
261 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
262 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
263 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
264 #define pthread_cond_signal(x) SetEvent(*x)
265 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
266 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
267 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
268 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
269 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
270 #define mdb_mutex_consistent(mutex) 0
271 #define getpid() GetCurrentProcessId()
272 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
273 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
274 #define ErrCode() GetLastError()
275 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
276 #define close(fd) (CloseHandle(fd) ? 0 : -1)
277 #define munmap(ptr,len) UnmapViewOfFile(ptr)
278 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
279 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
281 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
285 #define THREAD_RET void *
286 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
287 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
288 #define Z "z" /**< printf format modifier for size_t */
290 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
291 #define MDB_PIDLOCK 1
293 #ifdef MDB_USE_POSIX_SEM
295 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
296 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
297 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
300 mdb_sem_wait(sem_t *sem)
303 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
307 #elif defined MDB_USE_SYSV_SEM
309 typedef struct mdb_mutex {
313 } mdb_mutex_t[1], *mdb_mutexref_t;
315 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
316 #define UNLOCK_MUTEX(mutex) do { \
317 struct sembuf sb = { 0, 1, SEM_UNDO }; \
318 sb.sem_num = (mutex)->semnum; \
319 *(mutex)->locked = 0; \
320 semop((mutex)->semid, &sb, 1); \
324 mdb_sem_wait(mdb_mutexref_t sem)
326 int rc, *locked = sem->locked;
327 struct sembuf sb = { 0, -1, SEM_UNDO };
328 sb.sem_num = sem->semnum;
330 if (!semop(sem->semid, &sb, 1)) {
331 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
335 } while ((rc = errno) == EINTR);
339 #define mdb_mutex_consistent(mutex) 0
341 #else /* MDB_USE_POSIX_MUTEX: */
342 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
343 * local variables keep it (mdb_mutexref_t).
345 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
346 * be the same, or an array[size 1] and a pointer.
349 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
351 /** Lock the reader or writer mutex.
352 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
354 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
355 /** Unlock the reader or writer mutex.
357 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
358 /** Mark mutex-protected data as repaired, after death of previous owner.
360 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
361 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
363 /** Get the error code for the last failed system function.
365 #define ErrCode() errno
367 /** An abstraction for a file handle.
368 * On POSIX systems file handles are small integers. On Windows
369 * they're opaque pointers.
373 /** A value for an invalid file handle.
374 * Mainly used to initialize file variables and signify that they are
377 #define INVALID_HANDLE_VALUE (-1)
379 /** Get the size of a memory page for the system.
380 * This is the basic size that the platform's memory manager uses, and is
381 * fundamental to the use of memory-mapped files.
383 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
386 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
388 #elif defined(MDB_USE_SYSV_SEM)
389 #define MNAME_LEN (sizeof(int))
391 #define MNAME_LEN (sizeof(pthread_mutex_t))
394 #ifdef MDB_USE_SYSV_SEM
395 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
397 #define SYSV_SEM_FLAG 0
402 #ifdef MDB_ROBUST_SUPPORTED
403 /** Lock mutex, handle any error, set rc = result.
404 * Return 0 on success, nonzero (not rc) on error.
406 #define LOCK_MUTEX(rc, env, mutex) \
407 (((rc) = LOCK_MUTEX0(mutex)) && \
408 ((rc) = mdb_mutex_failed(env, mutex, rc)))
409 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
411 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
412 #define mdb_mutex_failed(env, mutex, rc) (rc)
416 /** A flag for opening a file and requesting synchronous data writes.
417 * This is only used when writing a meta page. It's not strictly needed;
418 * we could just do a normal write and then immediately perform a flush.
419 * But if this flag is available it saves us an extra system call.
421 * @note If O_DSYNC is undefined but exists in /usr/include,
422 * preferably set some compiler flag to get the definition.
423 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
426 # define MDB_DSYNC O_DSYNC
430 /** Function for flushing the data of a file. Define this to fsync
431 * if fdatasync() is not supported.
433 #ifndef MDB_FDATASYNC
434 # define MDB_FDATASYNC fdatasync
438 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
449 /** A page number in the database.
450 * Note that 64 bit page numbers are overkill, since pages themselves
451 * already represent 12-13 bits of addressable memory, and the OS will
452 * always limit applications to a maximum of 63 bits of address space.
454 * @note In the #MDB_node structure, we only store 48 bits of this value,
455 * which thus limits us to only 60 bits of addressable data.
457 typedef MDB_ID pgno_t;
459 /** A transaction ID.
460 * See struct MDB_txn.mt_txnid for details.
462 typedef MDB_ID txnid_t;
464 /** @defgroup debug Debug Macros
468 /** Enable debug output. Needs variable argument macros (a C99 feature).
469 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
470 * read from and written to the database (used for free space management).
476 static int mdb_debug;
477 static txnid_t mdb_debug_start;
479 /** Print a debug message with printf formatting.
480 * Requires double parenthesis around 2 or more args.
482 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
483 # define DPRINTF0(fmt, ...) \
484 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
486 # define DPRINTF(args) ((void) 0)
488 /** Print a debug string.
489 * The string is printed literally, with no format processing.
491 #define DPUTS(arg) DPRINTF(("%s", arg))
492 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
494 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
497 /** @brief The maximum size of a database page.
499 * It is 32k or 64k, since value-PAGEBASE must fit in
500 * #MDB_page.%mp_upper.
502 * LMDB will use database pages < OS pages if needed.
503 * That causes more I/O in write transactions: The OS must
504 * know (read) the whole page before writing a partial page.
506 * Note that we don't currently support Huge pages. On Linux,
507 * regular data files cannot use Huge pages, and in general
508 * Huge pages aren't actually pageable. We rely on the OS
509 * demand-pager to read our data and page it out when memory
510 * pressure from other processes is high. So until OSs have
511 * actual paging support for Huge pages, they're not viable.
513 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
515 /** The minimum number of keys required in a database page.
516 * Setting this to a larger value will place a smaller bound on the
517 * maximum size of a data item. Data items larger than this size will
518 * be pushed into overflow pages instead of being stored directly in
519 * the B-tree node. This value used to default to 4. With a page size
520 * of 4096 bytes that meant that any item larger than 1024 bytes would
521 * go into an overflow page. That also meant that on average 2-3KB of
522 * each overflow page was wasted space. The value cannot be lower than
523 * 2 because then there would no longer be a tree structure. With this
524 * value, items larger than 2KB will go into overflow pages, and on
525 * average only 1KB will be wasted.
527 #define MDB_MINKEYS 2
529 /** A stamp that identifies a file as an LMDB file.
530 * There's nothing special about this value other than that it is easily
531 * recognizable, and it will reflect any byte order mismatches.
533 #define MDB_MAGIC 0xBEEFC0DE
535 /** The version number for a database's datafile format. */
536 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
537 /** The version number for a database's lockfile format. */
538 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
540 /** @brief The max size of a key we can write, or 0 for computed max.
542 * This macro should normally be left alone or set to 0.
543 * Note that a database with big keys or dupsort data cannot be
544 * reliably modified by a liblmdb which uses a smaller max.
545 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
547 * Other values are allowed, for backwards compat. However:
548 * A value bigger than the computed max can break if you do not
549 * know what you are doing, and liblmdb <= 0.9.10 can break when
550 * modifying a DB with keys/dupsort data bigger than its max.
552 * Data items in an #MDB_DUPSORT database are also limited to
553 * this size, since they're actually keys of a sub-DB. Keys and
554 * #MDB_DUPSORT data items must fit on a node in a regular page.
556 #ifndef MDB_MAXKEYSIZE
557 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
560 /** The maximum size of a key we can write to the environment. */
562 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
564 #define ENV_MAXKEY(env) ((env)->me_maxkey)
567 /** @brief The maximum size of a data item.
569 * We only store a 32 bit value for node sizes.
571 #define MAXDATASIZE 0xffffffffUL
574 /** Key size which fits in a #DKBUF.
577 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
580 * This is used for printing a hex dump of a key's contents.
582 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
583 /** Display a key in hex.
585 * Invoke a function to display a key in hex.
587 #define DKEY(x) mdb_dkey(x, kbuf)
593 /** An invalid page number.
594 * Mainly used to denote an empty tree.
596 #define P_INVALID (~(pgno_t)0)
598 /** Test if the flags \b f are set in a flag word \b w. */
599 #define F_ISSET(w, f) (((w) & (f)) == (f))
601 /** Round \b n up to an even number. */
602 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
604 /** Used for offsets within a single page.
605 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
608 typedef uint16_t indx_t;
610 /** Default size of memory map.
611 * This is certainly too small for any actual applications. Apps should always set
612 * the size explicitly using #mdb_env_set_mapsize().
614 #define DEFAULT_MAPSIZE 1048576
616 /** @defgroup readers Reader Lock Table
617 * Readers don't acquire any locks for their data access. Instead, they
618 * simply record their transaction ID in the reader table. The reader
619 * mutex is needed just to find an empty slot in the reader table. The
620 * slot's address is saved in thread-specific data so that subsequent read
621 * transactions started by the same thread need no further locking to proceed.
623 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
625 * No reader table is used if the database is on a read-only filesystem, or
626 * if #MDB_NOLOCK is set.
628 * Since the database uses multi-version concurrency control, readers don't
629 * actually need any locking. This table is used to keep track of which
630 * readers are using data from which old transactions, so that we'll know
631 * when a particular old transaction is no longer in use. Old transactions
632 * that have discarded any data pages can then have those pages reclaimed
633 * for use by a later write transaction.
635 * The lock table is constructed such that reader slots are aligned with the
636 * processor's cache line size. Any slot is only ever used by one thread.
637 * This alignment guarantees that there will be no contention or cache
638 * thrashing as threads update their own slot info, and also eliminates
639 * any need for locking when accessing a slot.
641 * A writer thread will scan every slot in the table to determine the oldest
642 * outstanding reader transaction. Any freed pages older than this will be
643 * reclaimed by the writer. The writer doesn't use any locks when scanning
644 * this table. This means that there's no guarantee that the writer will
645 * see the most up-to-date reader info, but that's not required for correct
646 * operation - all we need is to know the upper bound on the oldest reader,
647 * we don't care at all about the newest reader. So the only consequence of
648 * reading stale information here is that old pages might hang around a
649 * while longer before being reclaimed. That's actually good anyway, because
650 * the longer we delay reclaiming old pages, the more likely it is that a
651 * string of contiguous pages can be found after coalescing old pages from
652 * many old transactions together.
655 /** Number of slots in the reader table.
656 * This value was chosen somewhat arbitrarily. 126 readers plus a
657 * couple mutexes fit exactly into 8KB on my development machine.
658 * Applications should set the table size using #mdb_env_set_maxreaders().
660 #define DEFAULT_READERS 126
662 /** The size of a CPU cache line in bytes. We want our lock structures
663 * aligned to this size to avoid false cache line sharing in the
665 * This value works for most CPUs. For Itanium this should be 128.
671 /** The information we store in a single slot of the reader table.
672 * In addition to a transaction ID, we also record the process and
673 * thread ID that owns a slot, so that we can detect stale information,
674 * e.g. threads or processes that went away without cleaning up.
675 * @note We currently don't check for stale records. We simply re-init
676 * the table when we know that we're the only process opening the
679 typedef struct MDB_rxbody {
680 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
681 * Multiple readers that start at the same time will probably have the
682 * same ID here. Again, it's not important to exclude them from
683 * anything; all we need to know is which version of the DB they
684 * started from so we can avoid overwriting any data used in that
685 * particular version.
687 volatile txnid_t mrb_txnid;
688 /** The process ID of the process owning this reader txn. */
689 volatile MDB_PID_T mrb_pid;
690 /** The thread ID of the thread owning this txn. */
691 volatile MDB_THR_T mrb_tid;
694 /** The actual reader record, with cacheline padding. */
695 typedef struct MDB_reader {
698 /** shorthand for mrb_txnid */
699 #define mr_txnid mru.mrx.mrb_txnid
700 #define mr_pid mru.mrx.mrb_pid
701 #define mr_tid mru.mrx.mrb_tid
702 /** cache line alignment */
703 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
707 /** The header for the reader table.
708 * The table resides in a memory-mapped file. (This is a different file
709 * than is used for the main database.)
711 * For POSIX the actual mutexes reside in the shared memory of this
712 * mapped file. On Windows, mutexes are named objects allocated by the
713 * kernel; we store the mutex names in this mapped file so that other
714 * processes can grab them. This same approach is also used on
715 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
716 * process-shared POSIX mutexes. For these cases where a named object
717 * is used, the object name is derived from a 64 bit FNV hash of the
718 * environment pathname. As such, naming collisions are extremely
719 * unlikely. If a collision occurs, the results are unpredictable.
721 typedef struct MDB_txbody {
722 /** Stamp identifying this as an LMDB file. It must be set
725 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
727 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
728 char mtb_rmname[MNAME_LEN];
729 #elif defined(MDB_USE_SYSV_SEM)
733 /** Mutex protecting access to this table.
734 * This is the reader table lock used with LOCK_MUTEX().
736 mdb_mutex_t mtb_rmutex;
738 /** The ID of the last transaction committed to the database.
739 * This is recorded here only for convenience; the value can always
740 * be determined by reading the main database meta pages.
742 volatile txnid_t mtb_txnid;
743 /** The number of slots that have been used in the reader table.
744 * This always records the maximum count, it is not decremented
745 * when readers release their slots.
747 volatile unsigned mtb_numreaders;
750 /** The actual reader table definition. */
751 typedef struct MDB_txninfo {
754 #define mti_magic mt1.mtb.mtb_magic
755 #define mti_format mt1.mtb.mtb_format
756 #define mti_rmutex mt1.mtb.mtb_rmutex
757 #define mti_rmname mt1.mtb.mtb_rmname
758 #define mti_txnid mt1.mtb.mtb_txnid
759 #define mti_numreaders mt1.mtb.mtb_numreaders
760 #ifdef MDB_USE_SYSV_SEM
761 #define mti_semid mt1.mtb.mtb_semid
762 #define mti_rlocked mt1.mtb.mtb_rlocked
764 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
767 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
768 char mt2_wmname[MNAME_LEN];
769 #define mti_wmname mt2.mt2_wmname
770 #elif defined MDB_USE_SYSV_SEM
772 #define mti_wlocked mt2.mt2_wlocked
774 mdb_mutex_t mt2_wmutex;
775 #define mti_wmutex mt2.mt2_wmutex
777 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
779 MDB_reader mti_readers[1];
782 /** Lockfile format signature: version, features and field layout */
783 #define MDB_LOCK_FORMAT \
785 ((MDB_LOCK_VERSION) \
786 /* Flags which describe functionality */ \
787 + (SYSV_SEM_FLAG << 18) \
788 + (((MDB_PIDLOCK) != 0) << 16)))
791 /** Common header for all page types.
792 * Overflow records occupy a number of contiguous pages with no
793 * headers on any page after the first.
795 typedef struct MDB_page {
796 #define mp_pgno mp_p.p_pgno
797 #define mp_next mp_p.p_next
799 pgno_t p_pgno; /**< page number */
800 struct MDB_page *p_next; /**< for in-memory list of freed pages */
803 /** @defgroup mdb_page Page Flags
805 * Flags for the page headers.
808 #define P_BRANCH 0x01 /**< branch page */
809 #define P_LEAF 0x02 /**< leaf page */
810 #define P_OVERFLOW 0x04 /**< overflow page */
811 #define P_META 0x08 /**< meta page */
812 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
813 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
814 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
815 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
816 #define P_KEEP 0x8000 /**< leave this page alone during spill */
818 uint16_t mp_flags; /**< @ref mdb_page */
819 #define mp_lower mp_pb.pb.pb_lower
820 #define mp_upper mp_pb.pb.pb_upper
821 #define mp_pages mp_pb.pb_pages
824 indx_t pb_lower; /**< lower bound of free space */
825 indx_t pb_upper; /**< upper bound of free space */
827 uint32_t pb_pages; /**< number of overflow pages */
829 indx_t mp_ptrs[1]; /**< dynamic size */
832 /** Size of the page header, excluding dynamic data at the end */
833 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
835 /** Address of first usable data byte in a page, after the header */
836 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
838 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
839 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
841 /** Number of nodes on a page */
842 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
844 /** The amount of space remaining in the page */
845 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
847 /** The percentage of space used in the page, in tenths of a percent. */
848 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
849 ((env)->me_psize - PAGEHDRSZ))
850 /** The minimum page fill factor, in tenths of a percent.
851 * Pages emptier than this are candidates for merging.
853 #define FILL_THRESHOLD 250
855 /** Test if a page is a leaf page */
856 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
857 /** Test if a page is a LEAF2 page */
858 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
859 /** Test if a page is a branch page */
860 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
861 /** Test if a page is an overflow page */
862 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
863 /** Test if a page is a sub page */
864 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
866 /** The number of overflow pages needed to store the given size. */
867 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
869 /** Link in #MDB_txn.%mt_loose_pgs list */
870 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
872 /** Header for a single key/data pair within a page.
873 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
874 * We guarantee 2-byte alignment for 'MDB_node's.
876 typedef struct MDB_node {
877 /** lo and hi are used for data size on leaf nodes and for
878 * child pgno on branch nodes. On 64 bit platforms, flags
879 * is also used for pgno. (Branch nodes have no flags).
880 * They are in host byte order in case that lets some
881 * accesses be optimized into a 32-bit word access.
883 #if BYTE_ORDER == LITTLE_ENDIAN
884 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
886 unsigned short mn_hi, mn_lo;
888 /** @defgroup mdb_node Node Flags
890 * Flags for node headers.
893 #define F_BIGDATA 0x01 /**< data put on overflow page */
894 #define F_SUBDATA 0x02 /**< data is a sub-database */
895 #define F_DUPDATA 0x04 /**< data has duplicates */
897 /** valid flags for #mdb_node_add() */
898 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
901 unsigned short mn_flags; /**< @ref mdb_node */
902 unsigned short mn_ksize; /**< key size */
903 char mn_data[1]; /**< key and data are appended here */
906 /** Size of the node header, excluding dynamic data at the end */
907 #define NODESIZE offsetof(MDB_node, mn_data)
909 /** Bit position of top word in page number, for shifting mn_flags */
910 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
912 /** Size of a node in a branch page with a given key.
913 * This is just the node header plus the key, there is no data.
915 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
917 /** Size of a node in a leaf page with a given key and data.
918 * This is node header plus key plus data size.
920 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
922 /** Address of node \b i in page \b p */
923 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
925 /** Address of the key for the node */
926 #define NODEKEY(node) (void *)((node)->mn_data)
928 /** Address of the data for a node */
929 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
931 /** Get the page number pointed to by a branch node */
932 #define NODEPGNO(node) \
933 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
934 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
935 /** Set the page number in a branch node */
936 #define SETPGNO(node,pgno) do { \
937 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
938 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
940 /** Get the size of the data in a leaf node */
941 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
942 /** Set the size of the data for a leaf node */
943 #define SETDSZ(node,size) do { \
944 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
945 /** The size of a key in a node */
946 #define NODEKSZ(node) ((node)->mn_ksize)
948 /** Copy a page number from src to dst */
950 #define COPY_PGNO(dst,src) dst = src
952 #if SIZE_MAX > 4294967295UL
953 #define COPY_PGNO(dst,src) do { \
954 unsigned short *s, *d; \
955 s = (unsigned short *)&(src); \
956 d = (unsigned short *)&(dst); \
963 #define COPY_PGNO(dst,src) do { \
964 unsigned short *s, *d; \
965 s = (unsigned short *)&(src); \
966 d = (unsigned short *)&(dst); \
972 /** The address of a key in a LEAF2 page.
973 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
974 * There are no node headers, keys are stored contiguously.
976 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
978 /** Set the \b node's key into \b keyptr, if requested. */
979 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
980 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
982 /** Set the \b node's key into \b key. */
983 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
985 /** Information about a single database in the environment. */
986 typedef struct MDB_db {
987 uint32_t md_pad; /**< also ksize for LEAF2 pages */
988 uint16_t md_flags; /**< @ref mdb_dbi_open */
989 uint16_t md_depth; /**< depth of this tree */
990 pgno_t md_branch_pages; /**< number of internal pages */
991 pgno_t md_leaf_pages; /**< number of leaf pages */
992 pgno_t md_overflow_pages; /**< number of overflow pages */
993 size_t md_entries; /**< number of data items */
994 pgno_t md_root; /**< the root page of this tree */
997 /** mdb_dbi_open flags */
998 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
999 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1000 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1001 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1003 /** Handle for the DB used to track free pages. */
1005 /** Handle for the default DB. */
1008 /** Meta page content.
1009 * A meta page is the start point for accessing a database snapshot.
1010 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1012 typedef struct MDB_meta {
1013 /** Stamp identifying this as an LMDB file. It must be set
1016 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1017 uint32_t mm_version;
1018 void *mm_address; /**< address for fixed mapping */
1019 size_t mm_mapsize; /**< size of mmap region */
1020 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
1021 /** The size of pages used in this DB */
1022 #define mm_psize mm_dbs[0].md_pad
1023 /** Any persistent environment flags. @ref mdb_env */
1024 #define mm_flags mm_dbs[0].md_flags
1025 pgno_t mm_last_pg; /**< last used page in file */
1026 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1029 /** Buffer for a stack-allocated meta page.
1030 * The members define size and alignment, and silence type
1031 * aliasing warnings. They are not used directly; that could
1032 * mean incorrectly using several union members in parallel.
1034 typedef union MDB_metabuf {
1037 char mm_pad[PAGEHDRSZ];
1042 /** Auxiliary DB info.
1043 * The information here is mostly static/read-only. There is
1044 * only a single copy of this record in the environment.
1046 typedef struct MDB_dbx {
1047 MDB_val md_name; /**< name of the database */
1048 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1049 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1050 MDB_rel_func *md_rel; /**< user relocate function */
1051 void *md_relctx; /**< user-provided context for md_rel */
1054 /** A database transaction.
1055 * Every operation requires a transaction handle.
1058 MDB_txn *mt_parent; /**< parent of a nested txn */
1059 MDB_txn *mt_child; /**< nested txn under this txn */
1060 pgno_t mt_next_pgno; /**< next unallocated page */
1061 /** The ID of this transaction. IDs are integers incrementing from 1.
1062 * Only committed write transactions increment the ID. If a transaction
1063 * aborts, the ID may be re-used by the next writer.
1066 MDB_env *mt_env; /**< the DB environment */
1067 /** The list of pages that became unused during this transaction.
1069 MDB_IDL mt_free_pgs;
1070 /** The list of loose pages that became unused and may be reused
1071 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1073 MDB_page *mt_loose_pgs;
1074 /* #Number of loose pages (#mt_loose_pgs) */
1076 /** The sorted list of dirty pages we temporarily wrote to disk
1077 * because the dirty list was full. page numbers in here are
1078 * shifted left by 1, deleted slots have the LSB set.
1080 MDB_IDL mt_spill_pgs;
1082 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1083 MDB_ID2L dirty_list;
1084 /** For read txns: This thread/txn's reader table slot, or NULL. */
1087 /** Array of records for each DB known in the environment. */
1089 /** Array of MDB_db records for each known DB */
1091 /** Array of sequence numbers for each DB handle */
1092 unsigned int *mt_dbiseqs;
1093 /** @defgroup mt_dbflag Transaction DB Flags
1097 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1098 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1099 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1100 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1101 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1103 /** In write txns, array of cursors for each DB */
1104 MDB_cursor **mt_cursors;
1105 /** Array of flags for each DB */
1106 unsigned char *mt_dbflags;
1107 /** Number of DB records in use. This number only ever increments;
1108 * we don't decrement it when individual DB handles are closed.
1112 /** @defgroup mdb_txn Transaction Flags
1116 /** #mdb_txn_begin() flags */
1117 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1118 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1119 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1120 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1121 /* internal txn flags */
1122 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1123 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1124 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1125 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1127 unsigned int mt_flags; /**< @ref mdb_txn */
1128 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1129 * Includes ancestor txns' dirty pages not hidden by other txns'
1130 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1131 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1133 unsigned int mt_dirty_room;
1136 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1137 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1138 * raise this on a 64 bit machine.
1140 #define CURSOR_STACK 32
1144 /** Cursors are used for all DB operations.
1145 * A cursor holds a path of (page pointer, key index) from the DB
1146 * root to a position in the DB, plus other state. #MDB_DUPSORT
1147 * cursors include an xcursor to the current data item. Write txns
1148 * track their cursors and keep them up to date when data moves.
1149 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1150 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1153 /** Next cursor on this DB in this txn */
1154 MDB_cursor *mc_next;
1155 /** Backup of the original cursor if this cursor is a shadow */
1156 MDB_cursor *mc_backup;
1157 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1158 struct MDB_xcursor *mc_xcursor;
1159 /** The transaction that owns this cursor */
1161 /** The database handle this cursor operates on */
1163 /** The database record for this cursor */
1165 /** The database auxiliary record for this cursor */
1167 /** The @ref mt_dbflag for this database */
1168 unsigned char *mc_dbflag;
1169 unsigned short mc_snum; /**< number of pushed pages */
1170 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1171 /** @defgroup mdb_cursor Cursor Flags
1173 * Cursor state flags.
1176 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1177 #define C_EOF 0x02 /**< No more data */
1178 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1179 #define C_DEL 0x08 /**< last op was a cursor_del */
1180 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1181 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1183 unsigned int mc_flags; /**< @ref mdb_cursor */
1184 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1185 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1188 /** Context for sorted-dup records.
1189 * We could have gone to a fully recursive design, with arbitrarily
1190 * deep nesting of sub-databases. But for now we only handle these
1191 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1193 typedef struct MDB_xcursor {
1194 /** A sub-cursor for traversing the Dup DB */
1195 MDB_cursor mx_cursor;
1196 /** The database record for this Dup DB */
1198 /** The auxiliary DB record for this Dup DB */
1200 /** The @ref mt_dbflag for this Dup DB */
1201 unsigned char mx_dbflag;
1204 /** State of FreeDB old pages, stored in the MDB_env */
1205 typedef struct MDB_pgstate {
1206 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1207 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1210 /** The database environment. */
1212 HANDLE me_fd; /**< The main data file */
1213 HANDLE me_lfd; /**< The lock file */
1214 HANDLE me_mfd; /**< just for writing the meta pages */
1215 /** Failed to update the meta page. Probably an I/O error. */
1216 #define MDB_FATAL_ERROR 0x80000000U
1217 /** Some fields are initialized. */
1218 #define MDB_ENV_ACTIVE 0x20000000U
1219 /** me_txkey is set */
1220 #define MDB_ENV_TXKEY 0x10000000U
1221 /** fdatasync is unreliable */
1222 #define MDB_FSYNCONLY 0x08000000U
1223 uint32_t me_flags; /**< @ref mdb_env */
1224 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1225 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1226 unsigned int me_maxreaders; /**< size of the reader table */
1227 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1228 volatile int me_close_readers;
1229 MDB_dbi me_numdbs; /**< number of DBs opened */
1230 MDB_dbi me_maxdbs; /**< size of the DB table */
1231 MDB_PID_T me_pid; /**< process ID of this env */
1232 char *me_path; /**< path to the DB files */
1233 char *me_map; /**< the memory map of the data file */
1234 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1235 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1236 void *me_pbuf; /**< scratch area for DUPSORT put() */
1237 MDB_txn *me_txn; /**< current write transaction */
1238 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1239 size_t me_mapsize; /**< size of the data memory map */
1240 off_t me_size; /**< current file size */
1241 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1242 MDB_dbx *me_dbxs; /**< array of static DB info */
1243 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1244 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1245 pthread_key_t me_txkey; /**< thread-key for readers */
1246 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1247 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1248 # define me_pglast me_pgstate.mf_pglast
1249 # define me_pghead me_pgstate.mf_pghead
1250 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1251 /** IDL of pages that became unused in a write txn */
1252 MDB_IDL me_free_pgs;
1253 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1254 MDB_ID2L me_dirty_list;
1255 /** Max number of freelist items that can fit in a single overflow page */
1257 /** Max size of a node on a page */
1258 unsigned int me_nodemax;
1259 #if !(MDB_MAXKEYSIZE)
1260 unsigned int me_maxkey; /**< max size of a key */
1262 int me_live_reader; /**< have liveness lock in reader table */
1264 int me_pidquery; /**< Used in OpenProcess */
1266 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1267 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1268 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1270 mdb_mutex_t me_rmutex;
1271 mdb_mutex_t me_wmutex;
1273 void *me_userctx; /**< User-settable context */
1274 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1277 /** Nested transaction */
1278 typedef struct MDB_ntxn {
1279 MDB_txn mnt_txn; /**< the transaction */
1280 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1283 /** max number of pages to commit in one writev() call */
1284 #define MDB_COMMIT_PAGES 64
1285 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1286 #undef MDB_COMMIT_PAGES
1287 #define MDB_COMMIT_PAGES IOV_MAX
1290 /** max bytes to write in one call */
1291 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1293 /** Check \b txn and \b dbi arguments to a function */
1294 #define TXN_DBI_EXIST(txn, dbi, validity) \
1295 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1297 /** Check for misused \b dbi handles */
1298 #define TXN_DBI_CHANGED(txn, dbi) \
1299 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1301 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1302 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1303 static int mdb_page_touch(MDB_cursor *mc);
1305 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1306 "reset-tmp", "fail-begin", "fail-beginchild"}
1308 /* mdb_txn_end operation number, for logging */
1309 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1310 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1312 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1313 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1314 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1315 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1316 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1318 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1319 static int mdb_page_search_root(MDB_cursor *mc,
1320 MDB_val *key, int modify);
1321 #define MDB_PS_MODIFY 1
1322 #define MDB_PS_ROOTONLY 2
1323 #define MDB_PS_FIRST 4
1324 #define MDB_PS_LAST 8
1325 static int mdb_page_search(MDB_cursor *mc,
1326 MDB_val *key, int flags);
1327 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1329 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1330 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1331 pgno_t newpgno, unsigned int nflags);
1333 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1334 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1335 static int mdb_env_write_meta(MDB_txn *txn);
1336 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1337 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1339 static void mdb_env_close0(MDB_env *env, int excl);
1341 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1342 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1343 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1344 static void mdb_node_del(MDB_cursor *mc, int ksize);
1345 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1346 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1347 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1348 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1349 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1351 static int mdb_rebalance(MDB_cursor *mc);
1352 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1354 static void mdb_cursor_pop(MDB_cursor *mc);
1355 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1357 static int mdb_cursor_del0(MDB_cursor *mc);
1358 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1359 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1360 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1361 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1362 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1364 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1365 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1367 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1368 static void mdb_xcursor_init0(MDB_cursor *mc);
1369 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1371 static int mdb_drop0(MDB_cursor *mc, int subs);
1372 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1373 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1376 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1379 /** Compare two items pointing at size_t's of unknown alignment. */
1380 #ifdef MISALIGNED_OK
1381 # define mdb_cmp_clong mdb_cmp_long
1383 # define mdb_cmp_clong mdb_cmp_cint
1387 static SECURITY_DESCRIPTOR mdb_null_sd;
1388 static SECURITY_ATTRIBUTES mdb_all_sa;
1389 static int mdb_sec_inited;
1392 /** Return the library version info. */
1394 mdb_version(int *major, int *minor, int *patch)
1396 if (major) *major = MDB_VERSION_MAJOR;
1397 if (minor) *minor = MDB_VERSION_MINOR;
1398 if (patch) *patch = MDB_VERSION_PATCH;
1399 return MDB_VERSION_STRING;
1402 /** Table of descriptions for LMDB @ref errors */
1403 static char *const mdb_errstr[] = {
1404 "MDB_KEYEXIST: Key/data pair already exists",
1405 "MDB_NOTFOUND: No matching key/data pair found",
1406 "MDB_PAGE_NOTFOUND: Requested page not found",
1407 "MDB_CORRUPTED: Located page was wrong type",
1408 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1409 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1410 "MDB_INVALID: File is not an LMDB file",
1411 "MDB_MAP_FULL: Environment mapsize limit reached",
1412 "MDB_DBS_FULL: Environment maxdbs limit reached",
1413 "MDB_READERS_FULL: Environment maxreaders limit reached",
1414 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1415 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1416 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1417 "MDB_PAGE_FULL: Internal error - page has no more space",
1418 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1419 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1420 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1421 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1422 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1423 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1427 mdb_strerror(int err)
1430 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1431 * This works as long as no function between the call to mdb_strerror
1432 * and the actual use of the message uses more than 4K of stack.
1435 char buf[1024], *ptr = buf;
1439 return ("Successful return: 0");
1441 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1442 i = err - MDB_KEYEXIST;
1443 return mdb_errstr[i];
1447 /* These are the C-runtime error codes we use. The comment indicates
1448 * their numeric value, and the Win32 error they would correspond to
1449 * if the error actually came from a Win32 API. A major mess, we should
1450 * have used LMDB-specific error codes for everything.
1453 case ENOENT: /* 2, FILE_NOT_FOUND */
1454 case EIO: /* 5, ACCESS_DENIED */
1455 case ENOMEM: /* 12, INVALID_ACCESS */
1456 case EACCES: /* 13, INVALID_DATA */
1457 case EBUSY: /* 16, CURRENT_DIRECTORY */
1458 case EINVAL: /* 22, BAD_COMMAND */
1459 case ENOSPC: /* 28, OUT_OF_PAPER */
1460 return strerror(err);
1465 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1466 FORMAT_MESSAGE_IGNORE_INSERTS,
1467 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1470 return strerror(err);
1474 /** assert(3) variant in cursor context */
1475 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1476 /** assert(3) variant in transaction context */
1477 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1478 /** assert(3) variant in environment context */
1479 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1482 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1483 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1486 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1487 const char *func, const char *file, int line)
1490 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1491 file, line, expr_txt, func);
1492 if (env->me_assert_func)
1493 env->me_assert_func(env, buf);
1494 fprintf(stderr, "%s\n", buf);
1498 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1502 /** Return the page number of \b mp which may be sub-page, for debug output */
1504 mdb_dbg_pgno(MDB_page *mp)
1507 COPY_PGNO(ret, mp->mp_pgno);
1511 /** Display a key in hexadecimal and return the address of the result.
1512 * @param[in] key the key to display
1513 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1514 * @return The key in hexadecimal form.
1517 mdb_dkey(MDB_val *key, char *buf)
1520 unsigned char *c = key->mv_data;
1526 if (key->mv_size > DKBUF_MAXKEYSIZE)
1527 return "MDB_MAXKEYSIZE";
1528 /* may want to make this a dynamic check: if the key is mostly
1529 * printable characters, print it as-is instead of converting to hex.
1533 for (i=0; i<key->mv_size; i++)
1534 ptr += sprintf(ptr, "%02x", *c++);
1536 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1542 mdb_leafnode_type(MDB_node *n)
1544 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1545 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1546 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1549 /** Display all the keys in the page. */
1551 mdb_page_list(MDB_page *mp)
1553 pgno_t pgno = mdb_dbg_pgno(mp);
1554 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1556 unsigned int i, nkeys, nsize, total = 0;
1560 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1561 case P_BRANCH: type = "Branch page"; break;
1562 case P_LEAF: type = "Leaf page"; break;
1563 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1564 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1565 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1567 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1568 pgno, mp->mp_pages, state);
1571 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1572 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1575 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1579 nkeys = NUMKEYS(mp);
1580 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1582 for (i=0; i<nkeys; i++) {
1583 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1584 key.mv_size = nsize = mp->mp_pad;
1585 key.mv_data = LEAF2KEY(mp, i, nsize);
1587 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1590 node = NODEPTR(mp, i);
1591 key.mv_size = node->mn_ksize;
1592 key.mv_data = node->mn_data;
1593 nsize = NODESIZE + key.mv_size;
1594 if (IS_BRANCH(mp)) {
1595 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1599 if (F_ISSET(node->mn_flags, F_BIGDATA))
1600 nsize += sizeof(pgno_t);
1602 nsize += NODEDSZ(node);
1604 nsize += sizeof(indx_t);
1605 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1606 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1608 total = EVEN(total);
1610 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1611 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1615 mdb_cursor_chk(MDB_cursor *mc)
1621 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1622 for (i=0; i<mc->mc_top; i++) {
1624 node = NODEPTR(mp, mc->mc_ki[i]);
1625 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1628 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1634 /** Count all the pages in each DB and in the freelist
1635 * and make sure it matches the actual number of pages
1637 * All named DBs must be open for a correct count.
1639 static void mdb_audit(MDB_txn *txn)
1643 MDB_ID freecount, count;
1648 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1649 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1650 freecount += *(MDB_ID *)data.mv_data;
1651 mdb_tassert(txn, rc == MDB_NOTFOUND);
1654 for (i = 0; i<txn->mt_numdbs; i++) {
1656 if (!(txn->mt_dbflags[i] & DB_VALID))
1658 mdb_cursor_init(&mc, txn, i, &mx);
1659 if (txn->mt_dbs[i].md_root == P_INVALID)
1661 count += txn->mt_dbs[i].md_branch_pages +
1662 txn->mt_dbs[i].md_leaf_pages +
1663 txn->mt_dbs[i].md_overflow_pages;
1664 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1665 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1666 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1669 mp = mc.mc_pg[mc.mc_top];
1670 for (j=0; j<NUMKEYS(mp); j++) {
1671 MDB_node *leaf = NODEPTR(mp, j);
1672 if (leaf->mn_flags & F_SUBDATA) {
1674 memcpy(&db, NODEDATA(leaf), sizeof(db));
1675 count += db.md_branch_pages + db.md_leaf_pages +
1676 db.md_overflow_pages;
1680 mdb_tassert(txn, rc == MDB_NOTFOUND);
1683 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1684 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1685 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1691 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1693 return txn->mt_dbxs[dbi].md_cmp(a, b);
1697 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1699 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1700 #if UINT_MAX < SIZE_MAX
1701 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1702 dcmp = mdb_cmp_clong;
1707 /** Allocate memory for a page.
1708 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1711 mdb_page_malloc(MDB_txn *txn, unsigned num)
1713 MDB_env *env = txn->mt_env;
1714 MDB_page *ret = env->me_dpages;
1715 size_t psize = env->me_psize, sz = psize, off;
1716 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1717 * For a single page alloc, we init everything after the page header.
1718 * For multi-page, we init the final page; if the caller needed that
1719 * many pages they will be filling in at least up to the last page.
1723 VGMEMP_ALLOC(env, ret, sz);
1724 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1725 env->me_dpages = ret->mp_next;
1728 psize -= off = PAGEHDRSZ;
1733 if ((ret = malloc(sz)) != NULL) {
1734 VGMEMP_ALLOC(env, ret, sz);
1735 if (!(env->me_flags & MDB_NOMEMINIT)) {
1736 memset((char *)ret + off, 0, psize);
1740 txn->mt_flags |= MDB_TXN_ERROR;
1744 /** Free a single page.
1745 * Saves single pages to a list, for future reuse.
1746 * (This is not used for multi-page overflow pages.)
1749 mdb_page_free(MDB_env *env, MDB_page *mp)
1751 mp->mp_next = env->me_dpages;
1752 VGMEMP_FREE(env, mp);
1753 env->me_dpages = mp;
1756 /** Free a dirty page */
1758 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1760 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1761 mdb_page_free(env, dp);
1763 /* large pages just get freed directly */
1764 VGMEMP_FREE(env, dp);
1769 /** Return all dirty pages to dpage list */
1771 mdb_dlist_free(MDB_txn *txn)
1773 MDB_env *env = txn->mt_env;
1774 MDB_ID2L dl = txn->mt_u.dirty_list;
1775 unsigned i, n = dl[0].mid;
1777 for (i = 1; i <= n; i++) {
1778 mdb_dpage_free(env, dl[i].mptr);
1783 /** Loosen or free a single page.
1784 * Saves single pages to a list for future reuse
1785 * in this same txn. It has been pulled from the freeDB
1786 * and already resides on the dirty list, but has been
1787 * deleted. Use these pages first before pulling again
1790 * If the page wasn't dirtied in this txn, just add it
1791 * to this txn's free list.
1794 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1797 pgno_t pgno = mp->mp_pgno;
1798 MDB_txn *txn = mc->mc_txn;
1800 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1801 if (txn->mt_parent) {
1802 MDB_ID2 *dl = txn->mt_u.dirty_list;
1803 /* If txn has a parent, make sure the page is in our
1807 unsigned x = mdb_mid2l_search(dl, pgno);
1808 if (x <= dl[0].mid && dl[x].mid == pgno) {
1809 if (mp != dl[x].mptr) { /* bad cursor? */
1810 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1811 txn->mt_flags |= MDB_TXN_ERROR;
1812 return MDB_CORRUPTED;
1819 /* no parent txn, so it's just ours */
1824 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1826 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1827 txn->mt_loose_pgs = mp;
1828 txn->mt_loose_count++;
1829 mp->mp_flags |= P_LOOSE;
1831 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1839 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1840 * @param[in] mc A cursor handle for the current operation.
1841 * @param[in] pflags Flags of the pages to update:
1842 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1843 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1844 * @return 0 on success, non-zero on failure.
1847 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1849 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1850 MDB_txn *txn = mc->mc_txn;
1856 int rc = MDB_SUCCESS, level;
1858 /* Mark pages seen by cursors */
1859 if (mc->mc_flags & C_UNTRACK)
1860 mc = NULL; /* will find mc in mt_cursors */
1861 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1862 for (; mc; mc=mc->mc_next) {
1863 if (!(mc->mc_flags & C_INITIALIZED))
1865 for (m3 = mc;; m3 = &mx->mx_cursor) {
1867 for (j=0; j<m3->mc_snum; j++) {
1869 if ((mp->mp_flags & Mask) == pflags)
1870 mp->mp_flags ^= P_KEEP;
1872 mx = m3->mc_xcursor;
1873 /* Proceed to mx if it is at a sub-database */
1874 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1876 if (! (mp && (mp->mp_flags & P_LEAF)))
1878 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1879 if (!(leaf->mn_flags & F_SUBDATA))
1888 /* Mark dirty root pages */
1889 for (i=0; i<txn->mt_numdbs; i++) {
1890 if (txn->mt_dbflags[i] & DB_DIRTY) {
1891 pgno_t pgno = txn->mt_dbs[i].md_root;
1892 if (pgno == P_INVALID)
1894 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1896 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1897 dp->mp_flags ^= P_KEEP;
1905 static int mdb_page_flush(MDB_txn *txn, int keep);
1907 /** Spill pages from the dirty list back to disk.
1908 * This is intended to prevent running into #MDB_TXN_FULL situations,
1909 * but note that they may still occur in a few cases:
1910 * 1) our estimate of the txn size could be too small. Currently this
1911 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1912 * 2) child txns may run out of space if their parents dirtied a
1913 * lot of pages and never spilled them. TODO: we probably should do
1914 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1915 * the parent's dirty_room is below a given threshold.
1917 * Otherwise, if not using nested txns, it is expected that apps will
1918 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1919 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1920 * If the txn never references them again, they can be left alone.
1921 * If the txn only reads them, they can be used without any fuss.
1922 * If the txn writes them again, they can be dirtied immediately without
1923 * going thru all of the work of #mdb_page_touch(). Such references are
1924 * handled by #mdb_page_unspill().
1926 * Also note, we never spill DB root pages, nor pages of active cursors,
1927 * because we'll need these back again soon anyway. And in nested txns,
1928 * we can't spill a page in a child txn if it was already spilled in a
1929 * parent txn. That would alter the parent txns' data even though
1930 * the child hasn't committed yet, and we'd have no way to undo it if
1931 * the child aborted.
1933 * @param[in] m0 cursor A cursor handle identifying the transaction and
1934 * database for which we are checking space.
1935 * @param[in] key For a put operation, the key being stored.
1936 * @param[in] data For a put operation, the data being stored.
1937 * @return 0 on success, non-zero on failure.
1940 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1942 MDB_txn *txn = m0->mc_txn;
1944 MDB_ID2L dl = txn->mt_u.dirty_list;
1945 unsigned int i, j, need;
1948 if (m0->mc_flags & C_SUB)
1951 /* Estimate how much space this op will take */
1952 i = m0->mc_db->md_depth;
1953 /* Named DBs also dirty the main DB */
1954 if (m0->mc_dbi > MAIN_DBI)
1955 i += txn->mt_dbs[MAIN_DBI].md_depth;
1956 /* For puts, roughly factor in the key+data size */
1958 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1959 i += i; /* double it for good measure */
1962 if (txn->mt_dirty_room > i)
1965 if (!txn->mt_spill_pgs) {
1966 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1967 if (!txn->mt_spill_pgs)
1970 /* purge deleted slots */
1971 MDB_IDL sl = txn->mt_spill_pgs;
1972 unsigned int num = sl[0];
1974 for (i=1; i<=num; i++) {
1981 /* Preserve pages which may soon be dirtied again */
1982 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1985 /* Less aggressive spill - we originally spilled the entire dirty list,
1986 * with a few exceptions for cursor pages and DB root pages. But this
1987 * turns out to be a lot of wasted effort because in a large txn many
1988 * of those pages will need to be used again. So now we spill only 1/8th
1989 * of the dirty pages. Testing revealed this to be a good tradeoff,
1990 * better than 1/2, 1/4, or 1/10.
1992 if (need < MDB_IDL_UM_MAX / 8)
1993 need = MDB_IDL_UM_MAX / 8;
1995 /* Save the page IDs of all the pages we're flushing */
1996 /* flush from the tail forward, this saves a lot of shifting later on. */
1997 for (i=dl[0].mid; i && need; i--) {
1998 MDB_ID pn = dl[i].mid << 1;
2000 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2002 /* Can't spill twice, make sure it's not already in a parent's
2005 if (txn->mt_parent) {
2007 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2008 if (tx2->mt_spill_pgs) {
2009 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2010 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2011 dp->mp_flags |= P_KEEP;
2019 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2023 mdb_midl_sort(txn->mt_spill_pgs);
2025 /* Flush the spilled part of dirty list */
2026 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2029 /* Reset any dirty pages we kept that page_flush didn't see */
2030 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2033 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2037 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2039 mdb_find_oldest(MDB_txn *txn)
2042 txnid_t mr, oldest = txn->mt_txnid - 1;
2043 if (txn->mt_env->me_txns) {
2044 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2045 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2056 /** Add a page to the txn's dirty list */
2058 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2061 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2063 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2064 insert = mdb_mid2l_append;
2066 insert = mdb_mid2l_insert;
2068 mid.mid = mp->mp_pgno;
2070 rc = insert(txn->mt_u.dirty_list, &mid);
2071 mdb_tassert(txn, rc == 0);
2072 txn->mt_dirty_room--;
2075 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2076 * me_pghead and mt_next_pgno.
2078 * If there are free pages available from older transactions, they
2079 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2080 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2081 * and move me_pglast to say which records were consumed. Only this
2082 * function can create me_pghead and move me_pglast/mt_next_pgno.
2083 * @param[in] mc cursor A cursor handle identifying the transaction and
2084 * database for which we are allocating.
2085 * @param[in] num the number of pages to allocate.
2086 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2087 * will always be satisfied by a single contiguous chunk of memory.
2088 * @return 0 on success, non-zero on failure.
2091 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2093 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2094 /* Get at most <Max_retries> more freeDB records once me_pghead
2095 * has enough pages. If not enough, use new pages from the map.
2096 * If <Paranoid> and mc is updating the freeDB, only get new
2097 * records if me_pghead is empty. Then the freelist cannot play
2098 * catch-up with itself by growing while trying to save it.
2100 enum { Paranoid = 1, Max_retries = 500 };
2102 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2104 int rc, retry = num * 60;
2105 MDB_txn *txn = mc->mc_txn;
2106 MDB_env *env = txn->mt_env;
2107 pgno_t pgno, *mop = env->me_pghead;
2108 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2110 txnid_t oldest = 0, last;
2115 /* If there are any loose pages, just use them */
2116 if (num == 1 && txn->mt_loose_pgs) {
2117 np = txn->mt_loose_pgs;
2118 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2119 txn->mt_loose_count--;
2120 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2128 /* If our dirty list is already full, we can't do anything */
2129 if (txn->mt_dirty_room == 0) {
2134 for (op = MDB_FIRST;; op = MDB_NEXT) {
2139 /* Seek a big enough contiguous page range. Prefer
2140 * pages at the tail, just truncating the list.
2146 if (mop[i-n2] == pgno+n2)
2153 if (op == MDB_FIRST) { /* 1st iteration */
2154 /* Prepare to fetch more and coalesce */
2155 last = env->me_pglast;
2156 oldest = env->me_pgoldest;
2157 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2160 key.mv_data = &last; /* will look up last+1 */
2161 key.mv_size = sizeof(last);
2163 if (Paranoid && mc->mc_dbi == FREE_DBI)
2166 if (Paranoid && retry < 0 && mop_len)
2170 /* Do not fetch more if the record will be too recent */
2171 if (oldest <= last) {
2173 oldest = mdb_find_oldest(txn);
2174 env->me_pgoldest = oldest;
2180 rc = mdb_cursor_get(&m2, &key, NULL, op);
2182 if (rc == MDB_NOTFOUND)
2186 last = *(txnid_t*)key.mv_data;
2187 if (oldest <= last) {
2189 oldest = mdb_find_oldest(txn);
2190 env->me_pgoldest = oldest;
2196 np = m2.mc_pg[m2.mc_top];
2197 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2198 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2201 idl = (MDB_ID *) data.mv_data;
2204 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2209 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2211 mop = env->me_pghead;
2213 env->me_pglast = last;
2215 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2216 last, txn->mt_dbs[FREE_DBI].md_root, i));
2218 DPRINTF(("IDL %"Z"u", idl[j]));
2220 /* Merge in descending sorted order */
2221 mdb_midl_xmerge(mop, idl);
2225 /* Use new pages from the map when nothing suitable in the freeDB */
2227 pgno = txn->mt_next_pgno;
2228 if (pgno + num >= env->me_maxpg) {
2229 DPUTS("DB size maxed out");
2235 if (env->me_flags & MDB_WRITEMAP) {
2236 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2238 if (!(np = mdb_page_malloc(txn, num))) {
2244 mop[0] = mop_len -= num;
2245 /* Move any stragglers down */
2246 for (j = i-num; j < mop_len; )
2247 mop[++j] = mop[++i];
2249 txn->mt_next_pgno = pgno + num;
2252 mdb_page_dirty(txn, np);
2258 txn->mt_flags |= MDB_TXN_ERROR;
2262 /** Copy the used portions of a non-overflow page.
2263 * @param[in] dst page to copy into
2264 * @param[in] src page to copy from
2265 * @param[in] psize size of a page
2268 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2270 enum { Align = sizeof(pgno_t) };
2271 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2273 /* If page isn't full, just copy the used portion. Adjust
2274 * alignment so memcpy may copy words instead of bytes.
2276 if ((unused &= -Align) && !IS_LEAF2(src)) {
2277 upper = (upper + PAGEBASE) & -Align;
2278 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2279 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2282 memcpy(dst, src, psize - unused);
2286 /** Pull a page off the txn's spill list, if present.
2287 * If a page being referenced was spilled to disk in this txn, bring
2288 * it back and make it dirty/writable again.
2289 * @param[in] txn the transaction handle.
2290 * @param[in] mp the page being referenced. It must not be dirty.
2291 * @param[out] ret the writable page, if any. ret is unchanged if
2292 * mp wasn't spilled.
2295 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2297 MDB_env *env = txn->mt_env;
2300 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2302 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2303 if (!tx2->mt_spill_pgs)
2305 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2306 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2309 if (txn->mt_dirty_room == 0)
2310 return MDB_TXN_FULL;
2311 if (IS_OVERFLOW(mp))
2315 if (env->me_flags & MDB_WRITEMAP) {
2318 np = mdb_page_malloc(txn, num);
2322 memcpy(np, mp, num * env->me_psize);
2324 mdb_page_copy(np, mp, env->me_psize);
2327 /* If in current txn, this page is no longer spilled.
2328 * If it happens to be the last page, truncate the spill list.
2329 * Otherwise mark it as deleted by setting the LSB.
2331 if (x == txn->mt_spill_pgs[0])
2332 txn->mt_spill_pgs[0]--;
2334 txn->mt_spill_pgs[x] |= 1;
2335 } /* otherwise, if belonging to a parent txn, the
2336 * page remains spilled until child commits
2339 mdb_page_dirty(txn, np);
2340 np->mp_flags |= P_DIRTY;
2348 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2349 * @param[in] mc cursor pointing to the page to be touched
2350 * @return 0 on success, non-zero on failure.
2353 mdb_page_touch(MDB_cursor *mc)
2355 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2356 MDB_txn *txn = mc->mc_txn;
2357 MDB_cursor *m2, *m3;
2361 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2362 if (txn->mt_flags & MDB_TXN_SPILLS) {
2364 rc = mdb_page_unspill(txn, mp, &np);
2370 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2371 (rc = mdb_page_alloc(mc, 1, &np)))
2374 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2375 mp->mp_pgno, pgno));
2376 mdb_cassert(mc, mp->mp_pgno != pgno);
2377 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2378 /* Update the parent page, if any, to point to the new page */
2380 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2381 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2382 SETPGNO(node, pgno);
2384 mc->mc_db->md_root = pgno;
2386 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2387 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2389 /* If txn has a parent, make sure the page is in our
2393 unsigned x = mdb_mid2l_search(dl, pgno);
2394 if (x <= dl[0].mid && dl[x].mid == pgno) {
2395 if (mp != dl[x].mptr) { /* bad cursor? */
2396 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2397 txn->mt_flags |= MDB_TXN_ERROR;
2398 return MDB_CORRUPTED;
2403 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2405 np = mdb_page_malloc(txn, 1);
2410 rc = mdb_mid2l_insert(dl, &mid);
2411 mdb_cassert(mc, rc == 0);
2416 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2418 np->mp_flags |= P_DIRTY;
2421 /* Adjust cursors pointing to mp */
2422 mc->mc_pg[mc->mc_top] = np;
2423 m2 = txn->mt_cursors[mc->mc_dbi];
2424 if (mc->mc_flags & C_SUB) {
2425 for (; m2; m2=m2->mc_next) {
2426 m3 = &m2->mc_xcursor->mx_cursor;
2427 if (m3->mc_snum < mc->mc_snum) continue;
2428 if (m3->mc_pg[mc->mc_top] == mp)
2429 m3->mc_pg[mc->mc_top] = np;
2432 for (; m2; m2=m2->mc_next) {
2433 if (m2->mc_snum < mc->mc_snum) continue;
2434 if (m2->mc_pg[mc->mc_top] == mp) {
2435 m2->mc_pg[mc->mc_top] = np;
2436 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2438 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2440 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2441 if (!(leaf->mn_flags & F_SUBDATA))
2442 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2450 txn->mt_flags |= MDB_TXN_ERROR;
2455 mdb_env_sync(MDB_env *env, int force)
2458 if (env->me_flags & MDB_RDONLY)
2460 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2461 if (env->me_flags & MDB_WRITEMAP) {
2462 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2463 ? MS_ASYNC : MS_SYNC;
2464 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2467 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2471 #ifdef BROKEN_FDATASYNC
2472 if (env->me_flags & MDB_FSYNCONLY) {
2473 if (fsync(env->me_fd))
2477 if (MDB_FDATASYNC(env->me_fd))
2484 /** Back up parent txn's cursors, then grab the originals for tracking */
2486 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2488 MDB_cursor *mc, *bk;
2493 for (i = src->mt_numdbs; --i >= 0; ) {
2494 if ((mc = src->mt_cursors[i]) != NULL) {
2495 size = sizeof(MDB_cursor);
2497 size += sizeof(MDB_xcursor);
2498 for (; mc; mc = bk->mc_next) {
2504 mc->mc_db = &dst->mt_dbs[i];
2505 /* Kill pointers into src - and dst to reduce abuse: The
2506 * user may not use mc until dst ends. Otherwise we'd...
2508 mc->mc_txn = NULL; /* ...set this to dst */
2509 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2510 if ((mx = mc->mc_xcursor) != NULL) {
2511 *(MDB_xcursor *)(bk+1) = *mx;
2512 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2514 mc->mc_next = dst->mt_cursors[i];
2515 dst->mt_cursors[i] = mc;
2522 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2523 * @param[in] txn the transaction handle.
2524 * @param[in] merge true to keep changes to parent cursors, false to revert.
2525 * @return 0 on success, non-zero on failure.
2528 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2530 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2534 for (i = txn->mt_numdbs; --i >= 0; ) {
2535 for (mc = cursors[i]; mc; mc = next) {
2537 if ((bk = mc->mc_backup) != NULL) {
2539 /* Commit changes to parent txn */
2540 mc->mc_next = bk->mc_next;
2541 mc->mc_backup = bk->mc_backup;
2542 mc->mc_txn = bk->mc_txn;
2543 mc->mc_db = bk->mc_db;
2544 mc->mc_dbflag = bk->mc_dbflag;
2545 if ((mx = mc->mc_xcursor) != NULL)
2546 mx->mx_cursor.mc_txn = bk->mc_txn;
2548 /* Abort nested txn */
2550 if ((mx = mc->mc_xcursor) != NULL)
2551 *mx = *(MDB_xcursor *)(bk+1);
2555 /* Only malloced cursors are permanently tracked. */
2562 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2568 Pidset = F_SETLK, Pidcheck = F_GETLK
2572 /** Set or check a pid lock. Set returns 0 on success.
2573 * Check returns 0 if the process is certainly dead, nonzero if it may
2574 * be alive (the lock exists or an error happened so we do not know).
2576 * On Windows Pidset is a no-op, we merely check for the existence
2577 * of the process with the given pid. On POSIX we use a single byte
2578 * lock on the lockfile, set at an offset equal to the pid.
2581 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2583 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2586 if (op == Pidcheck) {
2587 h = OpenProcess(env->me_pidquery, FALSE, pid);
2588 /* No documented "no such process" code, but other program use this: */
2590 return ErrCode() != ERROR_INVALID_PARAMETER;
2591 /* A process exists until all handles to it close. Has it exited? */
2592 ret = WaitForSingleObject(h, 0) != 0;
2599 struct flock lock_info;
2600 memset(&lock_info, 0, sizeof(lock_info));
2601 lock_info.l_type = F_WRLCK;
2602 lock_info.l_whence = SEEK_SET;
2603 lock_info.l_start = pid;
2604 lock_info.l_len = 1;
2605 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2606 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2608 } else if ((rc = ErrCode()) == EINTR) {
2616 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2617 * @param[in] txn the transaction handle to initialize
2618 * @return 0 on success, non-zero on failure.
2621 mdb_txn_renew0(MDB_txn *txn)
2623 MDB_env *env = txn->mt_env;
2624 MDB_txninfo *ti = env->me_txns;
2626 unsigned int i, nr, flags = txn->mt_flags;
2628 int rc, new_notls = 0;
2630 if ((flags &= MDB_TXN_RDONLY) != 0) {
2632 meta = mdb_env_pick_meta(env);
2633 txn->mt_txnid = meta->mm_txnid;
2634 txn->mt_u.reader = NULL;
2636 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2637 pthread_getspecific(env->me_txkey);
2639 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2640 return MDB_BAD_RSLOT;
2642 MDB_PID_T pid = env->me_pid;
2643 MDB_THR_T tid = pthread_self();
2644 mdb_mutexref_t rmutex = env->me_rmutex;
2646 if (!env->me_live_reader) {
2647 rc = mdb_reader_pid(env, Pidset, pid);
2650 env->me_live_reader = 1;
2653 if (LOCK_MUTEX(rc, env, rmutex))
2655 nr = ti->mti_numreaders;
2656 for (i=0; i<nr; i++)
2657 if (ti->mti_readers[i].mr_pid == 0)
2659 if (i == env->me_maxreaders) {
2660 UNLOCK_MUTEX(rmutex);
2661 return MDB_READERS_FULL;
2663 r = &ti->mti_readers[i];
2664 /* Claim the reader slot, carefully since other code
2665 * uses the reader table un-mutexed: First reset the
2666 * slot, next publish it in mti_numreaders. After
2667 * that, it is safe for mdb_env_close() to touch it.
2668 * When it will be closed, we can finally claim it.
2671 r->mr_txnid = (txnid_t)-1;
2674 ti->mti_numreaders = ++nr;
2675 env->me_close_readers = nr;
2677 UNLOCK_MUTEX(rmutex);
2679 new_notls = (env->me_flags & MDB_NOTLS);
2680 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2685 do /* LY: Retry on a race, ITS#7970. */
2686 r->mr_txnid = ti->mti_txnid;
2687 while(r->mr_txnid != ti->mti_txnid);
2688 txn->mt_txnid = r->mr_txnid;
2689 txn->mt_u.reader = r;
2690 meta = env->me_metas[txn->mt_txnid & 1];
2692 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2694 /* Not yet touching txn == env->me_txn0, it may be active */
2696 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2698 txn->mt_txnid = ti->mti_txnid;
2699 meta = env->me_metas[txn->mt_txnid & 1];
2701 meta = mdb_env_pick_meta(env);
2702 txn->mt_txnid = meta->mm_txnid;
2706 if (txn->mt_txnid == mdb_debug_start)
2709 txn->mt_child = NULL;
2710 txn->mt_loose_pgs = NULL;
2711 txn->mt_loose_count = 0;
2712 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2713 txn->mt_u.dirty_list = env->me_dirty_list;
2714 txn->mt_u.dirty_list[0].mid = 0;
2715 txn->mt_free_pgs = env->me_free_pgs;
2716 txn->mt_free_pgs[0] = 0;
2717 txn->mt_spill_pgs = NULL;
2719 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2722 /* Copy the DB info and flags */
2723 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2725 /* Moved to here to avoid a data race in read TXNs */
2726 txn->mt_next_pgno = meta->mm_last_pg+1;
2728 txn->mt_flags = flags;
2731 txn->mt_numdbs = env->me_numdbs;
2732 for (i=2; i<txn->mt_numdbs; i++) {
2733 x = env->me_dbflags[i];
2734 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2735 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2737 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2738 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2740 if (env->me_maxpg < txn->mt_next_pgno) {
2741 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2742 return MDB_MAP_RESIZED;
2749 mdb_txn_renew(MDB_txn *txn)
2753 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2756 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2757 DPUTS("environment had fatal error, must shutdown!");
2761 rc = mdb_txn_renew0(txn);
2762 if (rc == MDB_SUCCESS) {
2763 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2764 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2765 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2771 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2775 int rc, size, tsize;
2777 flags &= MDB_TXN_BEGIN_FLAGS;
2778 flags |= env->me_flags & MDB_WRITEMAP;
2780 if (env->me_flags & MDB_FATAL_ERROR) {
2781 DPUTS("environment had fatal error, must shutdown!");
2784 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2788 /* Nested transactions: Max 1 child, write txns only, no writemap */
2789 flags |= parent->mt_flags;
2790 if (parent->mt_child ||
2791 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2793 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2795 /* Child txns save MDB_pgstate and use own copy of cursors */
2796 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2797 size += tsize = sizeof(MDB_ntxn);
2798 } else if (flags & MDB_RDONLY) {
2799 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2800 size += tsize = sizeof(MDB_txn);
2802 /* Reuse preallocated write txn. However, do not touch it until
2803 * mdb_txn_renew0() succeeds, since it currently may be active.
2808 if ((txn = calloc(1, size)) == NULL) {
2809 DPRINTF(("calloc: %s", strerror(errno)));
2812 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2813 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2814 txn->mt_flags = flags;
2819 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2820 txn->mt_dbiseqs = parent->mt_dbiseqs;
2821 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2822 if (!txn->mt_u.dirty_list ||
2823 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2825 free(txn->mt_u.dirty_list);
2829 txn->mt_txnid = parent->mt_txnid;
2830 txn->mt_dirty_room = parent->mt_dirty_room;
2831 txn->mt_u.dirty_list[0].mid = 0;
2832 txn->mt_spill_pgs = NULL;
2833 txn->mt_next_pgno = parent->mt_next_pgno;
2834 parent->mt_child = txn;
2835 txn->mt_parent = parent;
2836 txn->mt_numdbs = parent->mt_numdbs;
2837 txn->mt_dbxs = parent->mt_dbxs;
2838 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2839 /* Copy parent's mt_dbflags, but clear DB_NEW */
2840 for (i=0; i<txn->mt_numdbs; i++)
2841 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2843 ntxn = (MDB_ntxn *)txn;
2844 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2845 if (env->me_pghead) {
2846 size = MDB_IDL_SIZEOF(env->me_pghead);
2847 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2849 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2854 rc = mdb_cursor_shadow(parent, txn);
2856 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2857 } else { /* MDB_RDONLY */
2858 txn->mt_dbiseqs = env->me_dbiseqs;
2860 rc = mdb_txn_renew0(txn);
2863 if (txn != env->me_txn0)
2866 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2868 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2869 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2870 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2877 mdb_txn_env(MDB_txn *txn)
2879 if(!txn) return NULL;
2884 mdb_txn_id(MDB_txn *txn)
2887 return txn->mt_txnid;
2890 /** Export or close DBI handles opened in this txn. */
2892 mdb_dbis_update(MDB_txn *txn, int keep)
2895 MDB_dbi n = txn->mt_numdbs;
2896 MDB_env *env = txn->mt_env;
2897 unsigned char *tdbflags = txn->mt_dbflags;
2899 for (i = n; --i >= 2;) {
2900 if (tdbflags[i] & DB_NEW) {
2902 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2904 char *ptr = env->me_dbxs[i].md_name.mv_data;
2906 env->me_dbxs[i].md_name.mv_data = NULL;
2907 env->me_dbxs[i].md_name.mv_size = 0;
2908 env->me_dbflags[i] = 0;
2909 env->me_dbiseqs[i]++;
2915 if (keep && env->me_numdbs < n)
2919 /** End a transaction, except successful commit of a nested transaction.
2920 * May be called twice for readonly txns: First reset it, then abort.
2921 * @param[in] txn the transaction handle to end
2922 * @param[in] mode why and how to end the transaction
2925 mdb_txn_end(MDB_txn *txn, unsigned mode)
2927 MDB_env *env = txn->mt_env;
2929 static const char *const names[] = MDB_END_NAMES;
2932 /* Export or close DBI handles opened in this txn */
2933 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2935 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2936 names[mode & MDB_END_OPMASK],
2937 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2938 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2940 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2941 if (txn->mt_u.reader) {
2942 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2943 if (!(env->me_flags & MDB_NOTLS)) {
2944 txn->mt_u.reader = NULL; /* txn does not own reader */
2945 } else if (mode & MDB_END_SLOT) {
2946 txn->mt_u.reader->mr_pid = 0;
2947 txn->mt_u.reader = NULL;
2948 } /* else txn owns the slot until it does MDB_END_SLOT */
2950 txn->mt_numdbs = 0; /* close nothing if called again */
2951 txn->mt_dbxs = NULL; /* mark txn as reset */
2953 pgno_t *pghead = env->me_pghead;
2955 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2956 mdb_cursors_close(txn, 0);
2957 if (!(env->me_flags & MDB_WRITEMAP)) {
2958 mdb_dlist_free(txn);
2961 if (!txn->mt_parent) {
2962 mdb_midl_shrink(&txn->mt_free_pgs);
2963 env->me_free_pgs = txn->mt_free_pgs;
2965 env->me_pghead = NULL;
2969 mode = 0; /* txn == env->me_txn0, do not free() it */
2971 /* The writer mutex was locked in mdb_txn_begin. */
2973 UNLOCK_MUTEX(env->me_wmutex);
2975 txn->mt_parent->mt_child = NULL;
2976 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2977 mdb_midl_free(txn->mt_free_pgs);
2978 mdb_midl_free(txn->mt_spill_pgs);
2979 free(txn->mt_u.dirty_list);
2982 mdb_midl_free(pghead);
2985 if (mode & MDB_END_FREE)
2990 mdb_txn_reset(MDB_txn *txn)
2995 /* This call is only valid for read-only txns */
2996 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2999 mdb_txn_end(txn, MDB_END_RESET);
3003 mdb_txn_abort(MDB_txn *txn)
3009 mdb_txn_abort(txn->mt_child);
3011 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3014 /** Save the freelist as of this transaction to the freeDB.
3015 * This changes the freelist. Keep trying until it stabilizes.
3018 mdb_freelist_save(MDB_txn *txn)
3020 /* env->me_pghead[] can grow and shrink during this call.
3021 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3022 * Page numbers cannot disappear from txn->mt_free_pgs[].
3025 MDB_env *env = txn->mt_env;
3026 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3027 txnid_t pglast = 0, head_id = 0;
3028 pgno_t freecnt = 0, *free_pgs, *mop;
3029 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3031 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3033 if (env->me_pghead) {
3034 /* Make sure first page of freeDB is touched and on freelist */
3035 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3036 if (rc && rc != MDB_NOTFOUND)
3040 if (!env->me_pghead && txn->mt_loose_pgs) {
3041 /* Put loose page numbers in mt_free_pgs, since
3042 * we may be unable to return them to me_pghead.
3044 MDB_page *mp = txn->mt_loose_pgs;
3045 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3047 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3048 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3049 txn->mt_loose_pgs = NULL;
3050 txn->mt_loose_count = 0;
3053 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3054 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3055 ? SSIZE_MAX : maxfree_1pg;
3058 /* Come back here after each Put() in case freelist changed */
3063 /* If using records from freeDB which we have not yet
3064 * deleted, delete them and any we reserved for me_pghead.
3066 while (pglast < env->me_pglast) {
3067 rc = mdb_cursor_first(&mc, &key, NULL);
3070 pglast = head_id = *(txnid_t *)key.mv_data;
3071 total_room = head_room = 0;
3072 mdb_tassert(txn, pglast <= env->me_pglast);
3073 rc = mdb_cursor_del(&mc, 0);
3078 /* Save the IDL of pages freed by this txn, to a single record */
3079 if (freecnt < txn->mt_free_pgs[0]) {
3081 /* Make sure last page of freeDB is touched and on freelist */
3082 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3083 if (rc && rc != MDB_NOTFOUND)
3086 free_pgs = txn->mt_free_pgs;
3087 /* Write to last page of freeDB */
3088 key.mv_size = sizeof(txn->mt_txnid);
3089 key.mv_data = &txn->mt_txnid;
3091 freecnt = free_pgs[0];
3092 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3093 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3096 /* Retry if mt_free_pgs[] grew during the Put() */
3097 free_pgs = txn->mt_free_pgs;
3098 } while (freecnt < free_pgs[0]);
3099 mdb_midl_sort(free_pgs);
3100 memcpy(data.mv_data, free_pgs, data.mv_size);
3103 unsigned int i = free_pgs[0];
3104 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3105 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3107 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3113 mop = env->me_pghead;
3114 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3116 /* Reserve records for me_pghead[]. Split it if multi-page,
3117 * to avoid searching freeDB for a page range. Use keys in
3118 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3120 if (total_room >= mop_len) {
3121 if (total_room == mop_len || --more < 0)
3123 } else if (head_room >= maxfree_1pg && head_id > 1) {
3124 /* Keep current record (overflow page), add a new one */
3128 /* (Re)write {key = head_id, IDL length = head_room} */
3129 total_room -= head_room;
3130 head_room = mop_len - total_room;
3131 if (head_room > maxfree_1pg && head_id > 1) {
3132 /* Overflow multi-page for part of me_pghead */
3133 head_room /= head_id; /* amortize page sizes */
3134 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3135 } else if (head_room < 0) {
3136 /* Rare case, not bothering to delete this record */
3139 key.mv_size = sizeof(head_id);
3140 key.mv_data = &head_id;
3141 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3142 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3145 /* IDL is initially empty, zero out at least the length */
3146 pgs = (pgno_t *)data.mv_data;
3147 j = head_room > clean_limit ? head_room : 0;
3151 total_room += head_room;
3154 /* Return loose page numbers to me_pghead, though usually none are
3155 * left at this point. The pages themselves remain in dirty_list.
3157 if (txn->mt_loose_pgs) {
3158 MDB_page *mp = txn->mt_loose_pgs;
3159 unsigned count = txn->mt_loose_count;
3161 /* Room for loose pages + temp IDL with same */
3162 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3164 mop = env->me_pghead;
3165 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3166 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3167 loose[ ++count ] = mp->mp_pgno;
3169 mdb_midl_sort(loose);
3170 mdb_midl_xmerge(mop, loose);
3171 txn->mt_loose_pgs = NULL;
3172 txn->mt_loose_count = 0;
3176 /* Fill in the reserved me_pghead records */
3182 rc = mdb_cursor_first(&mc, &key, &data);
3183 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3184 txnid_t id = *(txnid_t *)key.mv_data;
3185 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3188 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3190 if (len > mop_len) {
3192 data.mv_size = (len + 1) * sizeof(MDB_ID);
3194 data.mv_data = mop -= len;
3197 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3199 if (rc || !(mop_len -= len))
3206 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3207 * @param[in] txn the transaction that's being committed
3208 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3209 * @return 0 on success, non-zero on failure.
3212 mdb_page_flush(MDB_txn *txn, int keep)
3214 MDB_env *env = txn->mt_env;
3215 MDB_ID2L dl = txn->mt_u.dirty_list;
3216 unsigned psize = env->me_psize, j;
3217 int i, pagecount = dl[0].mid, rc;
3218 size_t size = 0, pos = 0;
3220 MDB_page *dp = NULL;
3224 struct iovec iov[MDB_COMMIT_PAGES];
3225 ssize_t wpos = 0, wsize = 0, wres;
3226 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3232 if (env->me_flags & MDB_WRITEMAP) {
3233 /* Clear dirty flags */
3234 while (++i <= pagecount) {
3236 /* Don't flush this page yet */
3237 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3238 dp->mp_flags &= ~P_KEEP;
3242 dp->mp_flags &= ~P_DIRTY;
3247 /* Write the pages */
3249 if (++i <= pagecount) {
3251 /* Don't flush this page yet */
3252 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3253 dp->mp_flags &= ~P_KEEP;
3258 /* clear dirty flag */
3259 dp->mp_flags &= ~P_DIRTY;
3262 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3267 /* Windows actually supports scatter/gather I/O, but only on
3268 * unbuffered file handles. Since we're relying on the OS page
3269 * cache for all our data, that's self-defeating. So we just
3270 * write pages one at a time. We use the ov structure to set
3271 * the write offset, to at least save the overhead of a Seek
3274 DPRINTF(("committing page %"Z"u", pgno));
3275 memset(&ov, 0, sizeof(ov));
3276 ov.Offset = pos & 0xffffffff;
3277 ov.OffsetHigh = pos >> 16 >> 16;
3278 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3280 DPRINTF(("WriteFile: %d", rc));
3284 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3285 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3288 /* Write previous page(s) */
3289 #ifdef MDB_USE_PWRITEV
3290 wres = pwritev(env->me_fd, iov, n, wpos);
3293 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3296 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3300 DPRINTF(("lseek: %s", strerror(rc)));
3303 wres = writev(env->me_fd, iov, n);
3306 if (wres != wsize) {
3311 DPRINTF(("Write error: %s", strerror(rc)));
3313 rc = EIO; /* TODO: Use which error code? */
3314 DPUTS("short write, filesystem full?");
3325 DPRINTF(("committing page %"Z"u", pgno));
3326 next_pos = pos + size;
3327 iov[n].iov_len = size;
3328 iov[n].iov_base = (char *)dp;
3334 /* MIPS has cache coherency issues, this is a no-op everywhere else
3335 * Note: for any size >= on-chip cache size, entire on-chip cache is
3338 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3340 for (i = keep; ++i <= pagecount; ) {
3342 /* This is a page we skipped above */
3345 dl[j].mid = dp->mp_pgno;
3348 mdb_dpage_free(env, dp);
3353 txn->mt_dirty_room += i - j;
3359 mdb_txn_commit(MDB_txn *txn)
3362 unsigned int i, end_mode;
3368 /* mdb_txn_end() mode for a commit which writes nothing */
3369 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3371 if (txn->mt_child) {
3372 rc = mdb_txn_commit(txn->mt_child);
3379 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3383 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3384 DPUTS("error flag is set, can't commit");
3386 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3391 if (txn->mt_parent) {
3392 MDB_txn *parent = txn->mt_parent;
3396 unsigned x, y, len, ps_len;
3398 /* Append our free list to parent's */
3399 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3402 mdb_midl_free(txn->mt_free_pgs);
3403 /* Failures after this must either undo the changes
3404 * to the parent or set MDB_TXN_ERROR in the parent.
3407 parent->mt_next_pgno = txn->mt_next_pgno;
3408 parent->mt_flags = txn->mt_flags;
3410 /* Merge our cursors into parent's and close them */
3411 mdb_cursors_close(txn, 1);
3413 /* Update parent's DB table. */
3414 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3415 parent->mt_numdbs = txn->mt_numdbs;
3416 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3417 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3418 for (i=2; i<txn->mt_numdbs; i++) {
3419 /* preserve parent's DB_NEW status */
3420 x = parent->mt_dbflags[i] & DB_NEW;
3421 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3424 dst = parent->mt_u.dirty_list;
3425 src = txn->mt_u.dirty_list;
3426 /* Remove anything in our dirty list from parent's spill list */
3427 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3429 pspill[0] = (pgno_t)-1;
3430 /* Mark our dirty pages as deleted in parent spill list */
3431 for (i=0, len=src[0].mid; ++i <= len; ) {
3432 MDB_ID pn = src[i].mid << 1;
3433 while (pn > pspill[x])
3435 if (pn == pspill[x]) {
3440 /* Squash deleted pagenums if we deleted any */
3441 for (x=y; ++x <= ps_len; )
3442 if (!(pspill[x] & 1))
3443 pspill[++y] = pspill[x];
3447 /* Find len = length of merging our dirty list with parent's */
3449 dst[0].mid = 0; /* simplify loops */
3450 if (parent->mt_parent) {
3451 len = x + src[0].mid;
3452 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3453 for (i = x; y && i; y--) {
3454 pgno_t yp = src[y].mid;
3455 while (yp < dst[i].mid)
3457 if (yp == dst[i].mid) {
3462 } else { /* Simplify the above for single-ancestor case */
3463 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3465 /* Merge our dirty list with parent's */
3467 for (i = len; y; dst[i--] = src[y--]) {
3468 pgno_t yp = src[y].mid;
3469 while (yp < dst[x].mid)
3470 dst[i--] = dst[x--];
3471 if (yp == dst[x].mid)
3472 free(dst[x--].mptr);
3474 mdb_tassert(txn, i == x);
3476 free(txn->mt_u.dirty_list);
3477 parent->mt_dirty_room = txn->mt_dirty_room;
3478 if (txn->mt_spill_pgs) {
3479 if (parent->mt_spill_pgs) {
3480 /* TODO: Prevent failure here, so parent does not fail */
3481 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3483 parent->mt_flags |= MDB_TXN_ERROR;
3484 mdb_midl_free(txn->mt_spill_pgs);
3485 mdb_midl_sort(parent->mt_spill_pgs);
3487 parent->mt_spill_pgs = txn->mt_spill_pgs;
3491 /* Append our loose page list to parent's */
3492 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3494 *lp = txn->mt_loose_pgs;
3495 parent->mt_loose_count += txn->mt_loose_count;
3497 parent->mt_child = NULL;
3498 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3503 if (txn != env->me_txn) {
3504 DPUTS("attempt to commit unknown transaction");
3509 mdb_cursors_close(txn, 0);
3511 if (!txn->mt_u.dirty_list[0].mid &&
3512 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3515 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3516 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3518 /* Update DB root pointers */
3519 if (txn->mt_numdbs > 2) {
3523 data.mv_size = sizeof(MDB_db);
3525 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3526 for (i = 2; i < txn->mt_numdbs; i++) {
3527 if (txn->mt_dbflags[i] & DB_DIRTY) {
3528 if (TXN_DBI_CHANGED(txn, i)) {
3532 data.mv_data = &txn->mt_dbs[i];
3533 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3541 rc = mdb_freelist_save(txn);
3545 mdb_midl_free(env->me_pghead);
3546 env->me_pghead = NULL;
3547 mdb_midl_shrink(&txn->mt_free_pgs);
3553 if ((rc = mdb_page_flush(txn, 0)))
3555 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3556 (rc = mdb_env_sync(env, 0)))
3558 if ((rc = mdb_env_write_meta(txn)))
3560 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3563 mdb_txn_end(txn, end_mode);
3571 /** Read the environment parameters of a DB environment before
3572 * mapping it into memory.
3573 * @param[in] env the environment handle
3574 * @param[out] meta address of where to store the meta information
3575 * @return 0 on success, non-zero on failure.
3578 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3584 enum { Size = sizeof(pbuf) };
3586 /* We don't know the page size yet, so use a minimum value.
3587 * Read both meta pages so we can use the latest one.
3590 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3594 memset(&ov, 0, sizeof(ov));
3596 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3597 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3600 rc = pread(env->me_fd, &pbuf, Size, off);
3603 if (rc == 0 && off == 0)
3605 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3606 DPRINTF(("read: %s", mdb_strerror(rc)));
3610 p = (MDB_page *)&pbuf;
3612 if (!F_ISSET(p->mp_flags, P_META)) {
3613 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3618 if (m->mm_magic != MDB_MAGIC) {
3619 DPUTS("meta has invalid magic");
3623 if (m->mm_version != MDB_DATA_VERSION) {
3624 DPRINTF(("database is version %u, expected version %u",
3625 m->mm_version, MDB_DATA_VERSION));
3626 return MDB_VERSION_MISMATCH;
3629 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3635 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3637 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3639 meta->mm_magic = MDB_MAGIC;
3640 meta->mm_version = MDB_DATA_VERSION;
3641 meta->mm_mapsize = env->me_mapsize;
3642 meta->mm_psize = env->me_psize;
3643 meta->mm_last_pg = 1;
3644 meta->mm_flags = env->me_flags & 0xffff;
3645 meta->mm_flags |= MDB_INTEGERKEY;
3646 meta->mm_dbs[0].md_root = P_INVALID;
3647 meta->mm_dbs[1].md_root = P_INVALID;
3650 /** Write the environment parameters of a freshly created DB environment.
3651 * @param[in] env the environment handle
3652 * @param[in] meta the #MDB_meta to write
3653 * @return 0 on success, non-zero on failure.
3656 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3664 memset(&ov, 0, sizeof(ov));
3665 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3667 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3670 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3671 len = pwrite(fd, ptr, size, pos); \
3672 if (len == -1 && ErrCode() == EINTR) continue; \
3673 rc = (len >= 0); break; } while(1)
3676 DPUTS("writing new meta page");
3678 psize = env->me_psize;
3680 p = calloc(2, psize);
3682 p->mp_flags = P_META;
3683 *(MDB_meta *)METADATA(p) = *meta;
3685 q = (MDB_page *)((char *)p + psize);
3687 q->mp_flags = P_META;
3688 *(MDB_meta *)METADATA(q) = *meta;
3690 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3693 else if ((unsigned) len == psize * 2)
3701 /** Update the environment info to commit a transaction.
3702 * @param[in] txn the transaction that's being committed
3703 * @return 0 on success, non-zero on failure.
3706 mdb_env_write_meta(MDB_txn *txn)
3709 MDB_meta meta, metab, *mp;
3713 int rc, len, toggle;
3722 toggle = txn->mt_txnid & 1;
3723 DPRINTF(("writing meta page %d for root page %"Z"u",
3724 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3727 flags = txn->mt_flags & env->me_flags;
3728 mp = env->me_metas[toggle];
3729 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3730 /* Persist any increases of mapsize config */
3731 if (mapsize < env->me_mapsize)
3732 mapsize = env->me_mapsize;
3734 if (flags & MDB_WRITEMAP) {
3735 mp->mm_mapsize = mapsize;
3736 mp->mm_dbs[0] = txn->mt_dbs[0];
3737 mp->mm_dbs[1] = txn->mt_dbs[1];
3738 mp->mm_last_pg = txn->mt_next_pgno - 1;
3739 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3740 !(defined(__i386__) || defined(__x86_64__))
3741 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3742 __sync_synchronize();
3744 mp->mm_txnid = txn->mt_txnid;
3745 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3746 unsigned meta_size = env->me_psize;
3747 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3750 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3751 if (meta_size < env->me_os_psize)
3752 meta_size += meta_size;
3757 if (MDB_MSYNC(ptr, meta_size, rc)) {
3764 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3765 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3767 meta.mm_mapsize = mapsize;
3768 meta.mm_dbs[0] = txn->mt_dbs[0];
3769 meta.mm_dbs[1] = txn->mt_dbs[1];
3770 meta.mm_last_pg = txn->mt_next_pgno - 1;
3771 meta.mm_txnid = txn->mt_txnid;
3773 off = offsetof(MDB_meta, mm_mapsize);
3774 ptr = (char *)&meta + off;
3775 len = sizeof(MDB_meta) - off;
3777 off += env->me_psize;
3780 /* Write to the SYNC fd */
3781 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3785 memset(&ov, 0, sizeof(ov));
3787 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3791 rc = pwrite(mfd, ptr, len, off);
3794 rc = rc < 0 ? ErrCode() : EIO;
3797 DPUTS("write failed, disk error?");
3798 /* On a failure, the pagecache still contains the new data.
3799 * Write some old data back, to prevent it from being used.
3800 * Use the non-SYNC fd; we know it will fail anyway.
3802 meta.mm_last_pg = metab.mm_last_pg;
3803 meta.mm_txnid = metab.mm_txnid;
3805 memset(&ov, 0, sizeof(ov));
3807 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3809 r2 = pwrite(env->me_fd, ptr, len, off);
3810 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3813 env->me_flags |= MDB_FATAL_ERROR;
3816 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3817 CACHEFLUSH(env->me_map + off, len, DCACHE);
3819 /* Memory ordering issues are irrelevant; since the entire writer
3820 * is wrapped by wmutex, all of these changes will become visible
3821 * after the wmutex is unlocked. Since the DB is multi-version,
3822 * readers will get consistent data regardless of how fresh or
3823 * how stale their view of these values is.
3826 env->me_txns->mti_txnid = txn->mt_txnid;
3831 /** Check both meta pages to see which one is newer.
3832 * @param[in] env the environment handle
3833 * @return newest #MDB_meta.
3836 mdb_env_pick_meta(const MDB_env *env)
3838 MDB_meta *const *metas = env->me_metas;
3839 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3843 mdb_env_create(MDB_env **env)
3847 e = calloc(1, sizeof(MDB_env));
3851 e->me_maxreaders = DEFAULT_READERS;
3852 e->me_maxdbs = e->me_numdbs = 2;
3853 e->me_fd = INVALID_HANDLE_VALUE;
3854 e->me_lfd = INVALID_HANDLE_VALUE;
3855 e->me_mfd = INVALID_HANDLE_VALUE;
3856 #ifdef MDB_USE_POSIX_SEM
3857 e->me_rmutex = SEM_FAILED;
3858 e->me_wmutex = SEM_FAILED;
3859 #elif defined MDB_USE_SYSV_SEM
3860 e->me_rmutex->semid = -1;
3861 e->me_wmutex->semid = -1;
3863 e->me_pid = getpid();
3864 GET_PAGESIZE(e->me_os_psize);
3865 VGMEMP_CREATE(e,0,0);
3871 mdb_env_map(MDB_env *env, void *addr)
3874 unsigned int flags = env->me_flags;
3878 LONG sizelo, sizehi;
3881 if (flags & MDB_RDONLY) {
3882 /* Don't set explicit map size, use whatever exists */
3887 msize = env->me_mapsize;
3888 sizelo = msize & 0xffffffff;
3889 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3891 /* Windows won't create mappings for zero length files.
3892 * and won't map more than the file size.
3893 * Just set the maxsize right now.
3895 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3896 || !SetEndOfFile(env->me_fd)
3897 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3901 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3902 PAGE_READWRITE : PAGE_READONLY,
3903 sizehi, sizelo, NULL);
3906 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3907 FILE_MAP_WRITE : FILE_MAP_READ,
3909 rc = env->me_map ? 0 : ErrCode();
3914 int prot = PROT_READ;
3915 if (flags & MDB_WRITEMAP) {
3917 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3920 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3922 if (env->me_map == MAP_FAILED) {
3927 if (flags & MDB_NORDAHEAD) {
3928 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3930 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3932 #ifdef POSIX_MADV_RANDOM
3933 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3934 #endif /* POSIX_MADV_RANDOM */
3935 #endif /* MADV_RANDOM */
3939 /* Can happen because the address argument to mmap() is just a
3940 * hint. mmap() can pick another, e.g. if the range is in use.
3941 * The MAP_FIXED flag would prevent that, but then mmap could
3942 * instead unmap existing pages to make room for the new map.
3944 if (addr && env->me_map != addr)
3945 return EBUSY; /* TODO: Make a new MDB_* error code? */
3947 p = (MDB_page *)env->me_map;
3948 env->me_metas[0] = METADATA(p);
3949 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3955 mdb_env_set_mapsize(MDB_env *env, size_t size)
3957 /* If env is already open, caller is responsible for making
3958 * sure there are no active txns.
3966 meta = mdb_env_pick_meta(env);
3968 size = meta->mm_mapsize;
3970 /* Silently round up to minimum if the size is too small */
3971 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3975 munmap(env->me_map, env->me_mapsize);
3976 env->me_mapsize = size;
3977 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3978 rc = mdb_env_map(env, old);
3982 env->me_mapsize = size;
3984 env->me_maxpg = env->me_mapsize / env->me_psize;
3989 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3993 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3998 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4000 if (env->me_map || readers < 1)
4002 env->me_maxreaders = readers;
4007 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4009 if (!env || !readers)
4011 *readers = env->me_maxreaders;
4016 mdb_fsize(HANDLE fd, size_t *size)
4019 LARGE_INTEGER fsize;
4021 if (!GetFileSizeEx(fd, &fsize))
4024 *size = fsize.QuadPart;
4036 #ifdef BROKEN_FDATASYNC
4037 #include <sys/utsname.h>
4038 #include <sys/vfs.h>
4041 /** Further setup required for opening an LMDB environment
4044 mdb_env_open2(MDB_env *env)
4046 unsigned int flags = env->me_flags;
4047 int i, newenv = 0, rc;
4051 /* See if we should use QueryLimited */
4053 if ((rc & 0xff) > 5)
4054 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4056 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4059 #ifdef BROKEN_FDATASYNC
4060 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4061 * https://lkml.org/lkml/2012/9/3/83
4062 * Kernels after 3.6-rc6 are known good.
4063 * https://lkml.org/lkml/2012/9/10/556
4064 * See if the DB is on ext3/ext4, then check for new enough kernel
4065 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4070 fstatfs(env->me_fd, &st);
4071 while (st.f_type == 0xEF53) {
4075 if (uts.release[0] < '3') {
4076 if (!strncmp(uts.release, "2.6.32.", 7)) {
4077 i = atoi(uts.release+7);
4079 break; /* 2.6.32.60 and newer is OK */
4080 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4081 i = atoi(uts.release+7);
4083 break; /* 2.6.34.15 and newer is OK */
4085 } else if (uts.release[0] == '3') {
4086 i = atoi(uts.release+2);
4088 break; /* 3.6 and newer is OK */
4090 i = atoi(uts.release+4);
4092 break; /* 3.5.4 and newer is OK */
4093 } else if (i == 2) {
4094 i = atoi(uts.release+4);
4096 break; /* 3.2.30 and newer is OK */
4098 } else { /* 4.x and newer is OK */
4101 env->me_flags |= MDB_FSYNCONLY;
4107 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4110 DPUTS("new mdbenv");
4112 env->me_psize = env->me_os_psize;
4113 if (env->me_psize > MAX_PAGESIZE)
4114 env->me_psize = MAX_PAGESIZE;
4115 memset(&meta, 0, sizeof(meta));
4116 mdb_env_init_meta0(env, &meta);
4117 meta.mm_mapsize = DEFAULT_MAPSIZE;
4119 env->me_psize = meta.mm_psize;
4122 /* Was a mapsize configured? */
4123 if (!env->me_mapsize) {
4124 env->me_mapsize = meta.mm_mapsize;
4127 /* Make sure mapsize >= committed data size. Even when using
4128 * mm_mapsize, which could be broken in old files (ITS#7789).
4130 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4131 if (env->me_mapsize < minsize)
4132 env->me_mapsize = minsize;
4134 meta.mm_mapsize = env->me_mapsize;
4136 if (newenv && !(flags & MDB_FIXEDMAP)) {
4137 /* mdb_env_map() may grow the datafile. Write the metapages
4138 * first, so the file will be valid if initialization fails.
4139 * Except with FIXEDMAP, since we do not yet know mm_address.
4140 * We could fill in mm_address later, but then a different
4141 * program might end up doing that - one with a memory layout
4142 * and map address which does not suit the main program.
4144 rc = mdb_env_init_meta(env, &meta);
4150 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4155 if (flags & MDB_FIXEDMAP)
4156 meta.mm_address = env->me_map;
4157 i = mdb_env_init_meta(env, &meta);
4158 if (i != MDB_SUCCESS) {
4163 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4164 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4166 #if !(MDB_MAXKEYSIZE)
4167 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4169 env->me_maxpg = env->me_mapsize / env->me_psize;
4173 MDB_meta *meta = mdb_env_pick_meta(env);
4174 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4176 DPRINTF(("opened database version %u, pagesize %u",
4177 meta->mm_version, env->me_psize));
4178 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4179 DPRINTF(("depth: %u", db->md_depth));
4180 DPRINTF(("entries: %"Z"u", db->md_entries));
4181 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4182 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4183 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4184 DPRINTF(("root: %"Z"u", db->md_root));
4192 /** Release a reader thread's slot in the reader lock table.
4193 * This function is called automatically when a thread exits.
4194 * @param[in] ptr This points to the slot in the reader lock table.
4197 mdb_env_reader_dest(void *ptr)
4199 MDB_reader *reader = ptr;
4205 /** Junk for arranging thread-specific callbacks on Windows. This is
4206 * necessarily platform and compiler-specific. Windows supports up
4207 * to 1088 keys. Let's assume nobody opens more than 64 environments
4208 * in a single process, for now. They can override this if needed.
4210 #ifndef MAX_TLS_KEYS
4211 #define MAX_TLS_KEYS 64
4213 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4214 static int mdb_tls_nkeys;
4216 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4220 case DLL_PROCESS_ATTACH: break;
4221 case DLL_THREAD_ATTACH: break;
4222 case DLL_THREAD_DETACH:
4223 for (i=0; i<mdb_tls_nkeys; i++) {
4224 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4226 mdb_env_reader_dest(r);
4230 case DLL_PROCESS_DETACH: break;
4235 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4237 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4241 /* Force some symbol references.
4242 * _tls_used forces the linker to create the TLS directory if not already done
4243 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4245 #pragma comment(linker, "/INCLUDE:_tls_used")
4246 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4247 #pragma const_seg(".CRT$XLB")
4248 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4249 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4252 #pragma comment(linker, "/INCLUDE:__tls_used")
4253 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4254 #pragma data_seg(".CRT$XLB")
4255 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4257 #endif /* WIN 32/64 */
4258 #endif /* !__GNUC__ */
4261 /** Downgrade the exclusive lock on the region back to shared */
4263 mdb_env_share_locks(MDB_env *env, int *excl)
4266 MDB_meta *meta = mdb_env_pick_meta(env);
4268 env->me_txns->mti_txnid = meta->mm_txnid;
4273 /* First acquire a shared lock. The Unlock will
4274 * then release the existing exclusive lock.
4276 memset(&ov, 0, sizeof(ov));
4277 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4280 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4286 struct flock lock_info;
4287 /* The shared lock replaces the existing lock */
4288 memset((void *)&lock_info, 0, sizeof(lock_info));
4289 lock_info.l_type = F_RDLCK;
4290 lock_info.l_whence = SEEK_SET;
4291 lock_info.l_start = 0;
4292 lock_info.l_len = 1;
4293 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4294 (rc = ErrCode()) == EINTR) ;
4295 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4302 /** Try to get exclusive lock, otherwise shared.
4303 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4306 mdb_env_excl_lock(MDB_env *env, int *excl)
4310 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4314 memset(&ov, 0, sizeof(ov));
4315 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4322 struct flock lock_info;
4323 memset((void *)&lock_info, 0, sizeof(lock_info));
4324 lock_info.l_type = F_WRLCK;
4325 lock_info.l_whence = SEEK_SET;
4326 lock_info.l_start = 0;
4327 lock_info.l_len = 1;
4328 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4329 (rc = ErrCode()) == EINTR) ;
4333 # ifndef MDB_USE_POSIX_MUTEX
4334 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4337 lock_info.l_type = F_RDLCK;
4338 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4339 (rc = ErrCode()) == EINTR) ;
4349 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4351 * @(#) $Revision: 5.1 $
4352 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4353 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4355 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4359 * Please do not copyright this code. This code is in the public domain.
4361 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4362 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4363 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4364 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4365 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4366 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4367 * PERFORMANCE OF THIS SOFTWARE.
4370 * chongo <Landon Curt Noll> /\oo/\
4371 * http://www.isthe.com/chongo/
4373 * Share and Enjoy! :-)
4376 typedef unsigned long long mdb_hash_t;
4377 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4379 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4380 * @param[in] val value to hash
4381 * @param[in] hval initial value for hash
4382 * @return 64 bit hash
4384 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4385 * hval arg on the first call.
4388 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4390 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4391 unsigned char *end = s + val->mv_size;
4393 * FNV-1a hash each octet of the string
4396 /* xor the bottom with the current octet */
4397 hval ^= (mdb_hash_t)*s++;
4399 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4400 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4401 (hval << 7) + (hval << 8) + (hval << 40);
4403 /* return our new hash value */
4407 /** Hash the string and output the encoded hash.
4408 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4409 * very short name limits. We don't care about the encoding being reversible,
4410 * we just want to preserve as many bits of the input as possible in a
4411 * small printable string.
4412 * @param[in] str string to hash
4413 * @param[out] encbuf an array of 11 chars to hold the hash
4415 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4418 mdb_pack85(unsigned long l, char *out)
4422 for (i=0; i<5; i++) {
4423 *out++ = mdb_a85[l % 85];
4429 mdb_hash_enc(MDB_val *val, char *encbuf)
4431 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4433 mdb_pack85(h, encbuf);
4434 mdb_pack85(h>>32, encbuf+5);
4439 /** Open and/or initialize the lock region for the environment.
4440 * @param[in] env The LMDB environment.
4441 * @param[in] lpath The pathname of the file used for the lock region.
4442 * @param[in] mode The Unix permissions for the file, if we create it.
4443 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4444 * @return 0 on success, non-zero on failure.
4447 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4450 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4452 # define MDB_ERRCODE_ROFS EROFS
4453 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4454 # define MDB_CLOEXEC O_CLOEXEC
4457 # define MDB_CLOEXEC 0
4460 #ifdef MDB_USE_SYSV_SEM
4468 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4469 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4470 FILE_ATTRIBUTE_NORMAL, NULL);
4472 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4474 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4476 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4481 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4482 /* Lose record locks when exec*() */
4483 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4484 fcntl(env->me_lfd, F_SETFD, fdflags);
4487 if (!(env->me_flags & MDB_NOTLS)) {
4488 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4491 env->me_flags |= MDB_ENV_TXKEY;
4493 /* Windows TLS callbacks need help finding their TLS info. */
4494 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4498 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4502 /* Try to get exclusive lock. If we succeed, then
4503 * nobody is using the lock region and we should initialize it.
4505 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4508 size = GetFileSize(env->me_lfd, NULL);
4510 size = lseek(env->me_lfd, 0, SEEK_END);
4511 if (size == -1) goto fail_errno;
4513 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4514 if (size < rsize && *excl > 0) {
4516 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4517 || !SetEndOfFile(env->me_lfd))
4520 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4524 size = rsize - sizeof(MDB_txninfo);
4525 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4530 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4532 if (!mh) goto fail_errno;
4533 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4535 if (!env->me_txns) goto fail_errno;
4537 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4539 if (m == MAP_FAILED) goto fail_errno;
4545 BY_HANDLE_FILE_INFORMATION stbuf;
4554 if (!mdb_sec_inited) {
4555 InitializeSecurityDescriptor(&mdb_null_sd,
4556 SECURITY_DESCRIPTOR_REVISION);
4557 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4558 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4559 mdb_all_sa.bInheritHandle = FALSE;
4560 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4563 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4564 idbuf.volume = stbuf.dwVolumeSerialNumber;
4565 idbuf.nhigh = stbuf.nFileIndexHigh;
4566 idbuf.nlow = stbuf.nFileIndexLow;
4567 val.mv_data = &idbuf;
4568 val.mv_size = sizeof(idbuf);
4569 mdb_hash_enc(&val, encbuf);
4570 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4571 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4572 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4573 if (!env->me_rmutex) goto fail_errno;
4574 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4575 if (!env->me_wmutex) goto fail_errno;
4576 #elif defined(MDB_USE_POSIX_SEM)
4585 #if defined(__NetBSD__)
4586 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4588 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4589 idbuf.dev = stbuf.st_dev;
4590 idbuf.ino = stbuf.st_ino;
4591 val.mv_data = &idbuf;
4592 val.mv_size = sizeof(idbuf);
4593 mdb_hash_enc(&val, encbuf);
4594 #ifdef MDB_SHORT_SEMNAMES
4595 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4597 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4598 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4599 /* Clean up after a previous run, if needed: Try to
4600 * remove both semaphores before doing anything else.
4602 sem_unlink(env->me_txns->mti_rmname);
4603 sem_unlink(env->me_txns->mti_wmname);
4604 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4605 O_CREAT|O_EXCL, mode, 1);
4606 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4607 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4608 O_CREAT|O_EXCL, mode, 1);
4609 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4610 #elif defined(MDB_USE_SYSV_SEM)
4611 unsigned short vals[2] = {1, 1};
4612 key_t key = ftok(lpath, 'M');
4615 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4619 if (semctl(semid, 0, SETALL, semu) < 0)
4621 env->me_txns->mti_semid = semid;
4622 #else /* MDB_USE_POSIX_MUTEX: */
4623 pthread_mutexattr_t mattr;
4625 if ((rc = pthread_mutexattr_init(&mattr))
4626 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4627 #ifdef MDB_ROBUST_SUPPORTED
4628 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4630 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4631 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4633 pthread_mutexattr_destroy(&mattr);
4634 #endif /* _WIN32 || ... */
4636 env->me_txns->mti_magic = MDB_MAGIC;
4637 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4638 env->me_txns->mti_txnid = 0;
4639 env->me_txns->mti_numreaders = 0;
4642 #ifdef MDB_USE_SYSV_SEM
4643 struct semid_ds buf;
4645 if (env->me_txns->mti_magic != MDB_MAGIC) {
4646 DPUTS("lock region has invalid magic");
4650 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4651 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4652 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4653 rc = MDB_VERSION_MISMATCH;
4657 if (rc && rc != EACCES && rc != EAGAIN) {
4661 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4662 if (!env->me_rmutex) goto fail_errno;
4663 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4664 if (!env->me_wmutex) goto fail_errno;
4665 #elif defined(MDB_USE_POSIX_SEM)
4666 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4667 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4668 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4669 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4670 #elif defined(MDB_USE_SYSV_SEM)
4671 semid = env->me_txns->mti_semid;
4673 /* check for read access */
4674 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4676 /* check for write access */
4677 if (semctl(semid, 0, IPC_SET, semu) < 0)
4681 #ifdef MDB_USE_SYSV_SEM
4682 env->me_rmutex->semid = semid;
4683 env->me_wmutex->semid = semid;
4684 env->me_rmutex->semnum = 0;
4685 env->me_wmutex->semnum = 1;
4686 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4687 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4698 /** The name of the lock file in the DB environment */
4699 #define LOCKNAME "/lock.mdb"
4700 /** The name of the data file in the DB environment */
4701 #define DATANAME "/data.mdb"
4702 /** The suffix of the lock file when no subdir is used */
4703 #define LOCKSUFF "-lock"
4704 /** Only a subset of the @ref mdb_env flags can be changed
4705 * at runtime. Changing other flags requires closing the
4706 * environment and re-opening it with the new flags.
4708 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4709 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4710 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4712 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4713 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4717 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4719 int oflags, rc, len, excl = -1;
4720 char *lpath, *dpath;
4722 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4726 if (flags & MDB_NOSUBDIR) {
4727 rc = len + sizeof(LOCKSUFF) + len + 1;
4729 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4734 if (flags & MDB_NOSUBDIR) {
4735 dpath = lpath + len + sizeof(LOCKSUFF);
4736 sprintf(lpath, "%s" LOCKSUFF, path);
4737 strcpy(dpath, path);
4739 dpath = lpath + len + sizeof(LOCKNAME);
4740 sprintf(lpath, "%s" LOCKNAME, path);
4741 sprintf(dpath, "%s" DATANAME, path);
4745 flags |= env->me_flags;
4746 if (flags & MDB_RDONLY) {
4747 /* silently ignore WRITEMAP when we're only getting read access */
4748 flags &= ~MDB_WRITEMAP;
4750 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4751 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4754 env->me_flags = flags |= MDB_ENV_ACTIVE;
4758 env->me_path = strdup(path);
4759 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4760 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4761 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4762 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4766 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4768 /* For RDONLY, get lockfile after we know datafile exists */
4769 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4770 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4776 if (F_ISSET(flags, MDB_RDONLY)) {
4777 oflags = GENERIC_READ;
4778 len = OPEN_EXISTING;
4780 oflags = GENERIC_READ|GENERIC_WRITE;
4783 mode = FILE_ATTRIBUTE_NORMAL;
4784 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4785 NULL, len, mode, NULL);
4787 if (F_ISSET(flags, MDB_RDONLY))
4790 oflags = O_RDWR | O_CREAT;
4792 env->me_fd = open(dpath, oflags, mode);
4794 if (env->me_fd == INVALID_HANDLE_VALUE) {
4799 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4800 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4805 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4806 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4807 env->me_mfd = env->me_fd;
4809 /* Synchronous fd for meta writes. Needed even with
4810 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4813 len = OPEN_EXISTING;
4814 env->me_mfd = CreateFile(dpath, oflags,
4815 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4816 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4819 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4821 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4826 DPRINTF(("opened dbenv %p", (void *) env));
4828 rc = mdb_env_share_locks(env, &excl);
4832 if (!(flags & MDB_RDONLY)) {
4834 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4835 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4836 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4837 (txn = calloc(1, size)))
4839 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4840 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4841 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4842 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4844 txn->mt_dbxs = env->me_dbxs;
4854 mdb_env_close0(env, excl);
4860 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4862 mdb_env_close0(MDB_env *env, int excl)
4866 if (!(env->me_flags & MDB_ENV_ACTIVE))
4869 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4871 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4872 free(env->me_dbxs[i].md_name.mv_data);
4877 free(env->me_dbiseqs);
4878 free(env->me_dbflags);
4880 free(env->me_dirty_list);
4882 mdb_midl_free(env->me_free_pgs);
4884 if (env->me_flags & MDB_ENV_TXKEY) {
4885 pthread_key_delete(env->me_txkey);
4887 /* Delete our key from the global list */
4888 for (i=0; i<mdb_tls_nkeys; i++)
4889 if (mdb_tls_keys[i] == env->me_txkey) {
4890 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4898 munmap(env->me_map, env->me_mapsize);
4900 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4901 (void) close(env->me_mfd);
4902 if (env->me_fd != INVALID_HANDLE_VALUE)
4903 (void) close(env->me_fd);
4905 MDB_PID_T pid = env->me_pid;
4906 /* Clearing readers is done in this function because
4907 * me_txkey with its destructor must be disabled first.
4909 * We skip the the reader mutex, so we touch only
4910 * data owned by this process (me_close_readers and
4911 * our readers), and clear each reader atomically.
4913 for (i = env->me_close_readers; --i >= 0; )
4914 if (env->me_txns->mti_readers[i].mr_pid == pid)
4915 env->me_txns->mti_readers[i].mr_pid = 0;
4917 if (env->me_rmutex) {
4918 CloseHandle(env->me_rmutex);
4919 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4921 /* Windows automatically destroys the mutexes when
4922 * the last handle closes.
4924 #elif defined(MDB_USE_POSIX_SEM)
4925 if (env->me_rmutex != SEM_FAILED) {
4926 sem_close(env->me_rmutex);
4927 if (env->me_wmutex != SEM_FAILED)
4928 sem_close(env->me_wmutex);
4929 /* If we have the filelock: If we are the
4930 * only remaining user, clean up semaphores.
4933 mdb_env_excl_lock(env, &excl);
4935 sem_unlink(env->me_txns->mti_rmname);
4936 sem_unlink(env->me_txns->mti_wmname);
4939 #elif defined(MDB_USE_SYSV_SEM)
4940 if (env->me_rmutex->semid != -1) {
4941 /* If we have the filelock: If we are the
4942 * only remaining user, clean up semaphores.
4945 mdb_env_excl_lock(env, &excl);
4947 semctl(env->me_rmutex->semid, 0, IPC_RMID);
4950 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4952 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4955 /* Unlock the lockfile. Windows would have unlocked it
4956 * after closing anyway, but not necessarily at once.
4958 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4961 (void) close(env->me_lfd);
4964 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4968 mdb_env_close(MDB_env *env)
4975 VGMEMP_DESTROY(env);
4976 while ((dp = env->me_dpages) != NULL) {
4977 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4978 env->me_dpages = dp->mp_next;
4982 mdb_env_close0(env, 0);
4986 /** Compare two items pointing at aligned size_t's */
4988 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4990 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4991 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4994 /** Compare two items pointing at aligned unsigned int's.
4996 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4997 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5000 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5002 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5003 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5006 /** Compare two items pointing at unsigned ints of unknown alignment.
5007 * Nodes and keys are guaranteed to be 2-byte aligned.
5010 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5012 #if BYTE_ORDER == LITTLE_ENDIAN
5013 unsigned short *u, *c;
5016 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5017 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5020 } while(!x && u > (unsigned short *)a->mv_data);
5023 unsigned short *u, *c, *end;
5026 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5027 u = (unsigned short *)a->mv_data;
5028 c = (unsigned short *)b->mv_data;
5031 } while(!x && u < end);
5036 /** Compare two items lexically */
5038 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5045 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5051 diff = memcmp(a->mv_data, b->mv_data, len);
5052 return diff ? diff : len_diff<0 ? -1 : len_diff;
5055 /** Compare two items in reverse byte order */
5057 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5059 const unsigned char *p1, *p2, *p1_lim;
5063 p1_lim = (const unsigned char *)a->mv_data;
5064 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5065 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5067 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5073 while (p1 > p1_lim) {
5074 diff = *--p1 - *--p2;
5078 return len_diff<0 ? -1 : len_diff;
5081 /** Search for key within a page, using binary search.
5082 * Returns the smallest entry larger or equal to the key.
5083 * If exactp is non-null, stores whether the found entry was an exact match
5084 * in *exactp (1 or 0).
5085 * Updates the cursor index with the index of the found entry.
5086 * If no entry larger or equal to the key is found, returns NULL.
5089 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5091 unsigned int i = 0, nkeys;
5094 MDB_page *mp = mc->mc_pg[mc->mc_top];
5095 MDB_node *node = NULL;
5100 nkeys = NUMKEYS(mp);
5102 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5103 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5106 low = IS_LEAF(mp) ? 0 : 1;
5108 cmp = mc->mc_dbx->md_cmp;
5110 /* Branch pages have no data, so if using integer keys,
5111 * alignment is guaranteed. Use faster mdb_cmp_int.
5113 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5114 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5121 nodekey.mv_size = mc->mc_db->md_pad;
5122 node = NODEPTR(mp, 0); /* fake */
5123 while (low <= high) {
5124 i = (low + high) >> 1;
5125 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5126 rc = cmp(key, &nodekey);
5127 DPRINTF(("found leaf index %u [%s], rc = %i",
5128 i, DKEY(&nodekey), rc));
5137 while (low <= high) {
5138 i = (low + high) >> 1;
5140 node = NODEPTR(mp, i);
5141 nodekey.mv_size = NODEKSZ(node);
5142 nodekey.mv_data = NODEKEY(node);
5144 rc = cmp(key, &nodekey);
5147 DPRINTF(("found leaf index %u [%s], rc = %i",
5148 i, DKEY(&nodekey), rc));
5150 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5151 i, DKEY(&nodekey), NODEPGNO(node), rc));
5162 if (rc > 0) { /* Found entry is less than the key. */
5163 i++; /* Skip to get the smallest entry larger than key. */
5165 node = NODEPTR(mp, i);
5168 *exactp = (rc == 0 && nkeys > 0);
5169 /* store the key index */
5170 mc->mc_ki[mc->mc_top] = i;
5172 /* There is no entry larger or equal to the key. */
5175 /* nodeptr is fake for LEAF2 */
5181 mdb_cursor_adjust(MDB_cursor *mc, func)
5185 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5186 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5193 /** Pop a page off the top of the cursor's stack. */
5195 mdb_cursor_pop(MDB_cursor *mc)
5198 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5199 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5207 /** Push a page onto the top of the cursor's stack. */
5209 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5211 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5212 DDBI(mc), (void *) mc));
5214 if (mc->mc_snum >= CURSOR_STACK) {
5215 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5216 return MDB_CURSOR_FULL;
5219 mc->mc_top = mc->mc_snum++;
5220 mc->mc_pg[mc->mc_top] = mp;
5221 mc->mc_ki[mc->mc_top] = 0;
5226 /** Find the address of the page corresponding to a given page number.
5227 * @param[in] txn the transaction for this access.
5228 * @param[in] pgno the page number for the page to retrieve.
5229 * @param[out] ret address of a pointer where the page's address will be stored.
5230 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5231 * @return 0 on success, non-zero on failure.
5234 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5236 MDB_env *env = txn->mt_env;
5240 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5244 MDB_ID2L dl = tx2->mt_u.dirty_list;
5246 /* Spilled pages were dirtied in this txn and flushed
5247 * because the dirty list got full. Bring this page
5248 * back in from the map (but don't unspill it here,
5249 * leave that unless page_touch happens again).
5251 if (tx2->mt_spill_pgs) {
5252 MDB_ID pn = pgno << 1;
5253 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5254 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5255 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5260 unsigned x = mdb_mid2l_search(dl, pgno);
5261 if (x <= dl[0].mid && dl[x].mid == pgno) {
5267 } while ((tx2 = tx2->mt_parent) != NULL);
5270 if (pgno < txn->mt_next_pgno) {
5272 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5274 DPRINTF(("page %"Z"u not found", pgno));
5275 txn->mt_flags |= MDB_TXN_ERROR;
5276 return MDB_PAGE_NOTFOUND;
5286 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5287 * The cursor is at the root page, set up the rest of it.
5290 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5292 MDB_page *mp = mc->mc_pg[mc->mc_top];
5296 while (IS_BRANCH(mp)) {
5300 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5301 mdb_cassert(mc, NUMKEYS(mp) > 1);
5302 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5304 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5306 if (flags & MDB_PS_LAST)
5307 i = NUMKEYS(mp) - 1;
5310 node = mdb_node_search(mc, key, &exact);
5312 i = NUMKEYS(mp) - 1;
5314 i = mc->mc_ki[mc->mc_top];
5316 mdb_cassert(mc, i > 0);
5320 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5323 mdb_cassert(mc, i < NUMKEYS(mp));
5324 node = NODEPTR(mp, i);
5326 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5329 mc->mc_ki[mc->mc_top] = i;
5330 if ((rc = mdb_cursor_push(mc, mp)))
5333 if (flags & MDB_PS_MODIFY) {
5334 if ((rc = mdb_page_touch(mc)) != 0)
5336 mp = mc->mc_pg[mc->mc_top];
5341 DPRINTF(("internal error, index points to a %02X page!?",
5343 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5344 return MDB_CORRUPTED;
5347 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5348 key ? DKEY(key) : "null"));
5349 mc->mc_flags |= C_INITIALIZED;
5350 mc->mc_flags &= ~C_EOF;
5355 /** Search for the lowest key under the current branch page.
5356 * This just bypasses a NUMKEYS check in the current page
5357 * before calling mdb_page_search_root(), because the callers
5358 * are all in situations where the current page is known to
5362 mdb_page_search_lowest(MDB_cursor *mc)
5364 MDB_page *mp = mc->mc_pg[mc->mc_top];
5365 MDB_node *node = NODEPTR(mp, 0);
5368 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5371 mc->mc_ki[mc->mc_top] = 0;
5372 if ((rc = mdb_cursor_push(mc, mp)))
5374 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5377 /** Search for the page a given key should be in.
5378 * Push it and its parent pages on the cursor stack.
5379 * @param[in,out] mc the cursor for this operation.
5380 * @param[in] key the key to search for, or NULL for first/last page.
5381 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5382 * are touched (updated with new page numbers).
5383 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5384 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5385 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5386 * @return 0 on success, non-zero on failure.
5389 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5394 /* Make sure the txn is still viable, then find the root from
5395 * the txn's db table and set it as the root of the cursor's stack.
5397 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5398 DPUTS("transaction has failed, must abort");
5401 /* Make sure we're using an up-to-date root */
5402 if (*mc->mc_dbflag & DB_STALE) {
5404 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5406 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5407 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5414 MDB_node *leaf = mdb_node_search(&mc2,
5415 &mc->mc_dbx->md_name, &exact);
5417 return MDB_NOTFOUND;
5418 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5419 return MDB_INCOMPATIBLE; /* not a named DB */
5420 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5423 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5425 /* The txn may not know this DBI, or another process may
5426 * have dropped and recreated the DB with other flags.
5428 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5429 return MDB_INCOMPATIBLE;
5430 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5432 *mc->mc_dbflag &= ~DB_STALE;
5434 root = mc->mc_db->md_root;
5436 if (root == P_INVALID) { /* Tree is empty. */
5437 DPUTS("tree is empty");
5438 return MDB_NOTFOUND;
5442 mdb_cassert(mc, root > 1);
5443 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5444 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5450 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5451 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5453 if (flags & MDB_PS_MODIFY) {
5454 if ((rc = mdb_page_touch(mc)))
5458 if (flags & MDB_PS_ROOTONLY)
5461 return mdb_page_search_root(mc, key, flags);
5465 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5467 MDB_txn *txn = mc->mc_txn;
5468 pgno_t pg = mp->mp_pgno;
5469 unsigned x = 0, ovpages = mp->mp_pages;
5470 MDB_env *env = txn->mt_env;
5471 MDB_IDL sl = txn->mt_spill_pgs;
5472 MDB_ID pn = pg << 1;
5475 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5476 /* If the page is dirty or on the spill list we just acquired it,
5477 * so we should give it back to our current free list, if any.
5478 * Otherwise put it onto the list of pages we freed in this txn.
5480 * Won't create me_pghead: me_pglast must be inited along with it.
5481 * Unsupported in nested txns: They would need to hide the page
5482 * range in ancestor txns' dirty and spilled lists.
5484 if (env->me_pghead &&
5486 ((mp->mp_flags & P_DIRTY) ||
5487 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5491 MDB_ID2 *dl, ix, iy;
5492 rc = mdb_midl_need(&env->me_pghead, ovpages);
5495 if (!(mp->mp_flags & P_DIRTY)) {
5496 /* This page is no longer spilled */
5503 /* Remove from dirty list */
5504 dl = txn->mt_u.dirty_list;
5506 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5512 mdb_cassert(mc, x > 1);
5514 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5515 txn->mt_flags |= MDB_TXN_ERROR;
5516 return MDB_CORRUPTED;
5519 if (!(env->me_flags & MDB_WRITEMAP))
5520 mdb_dpage_free(env, mp);
5522 /* Insert in me_pghead */
5523 mop = env->me_pghead;
5524 j = mop[0] + ovpages;
5525 for (i = mop[0]; i && mop[i] < pg; i--)
5531 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5535 mc->mc_db->md_overflow_pages -= ovpages;
5539 /** Return the data associated with a given node.
5540 * @param[in] txn The transaction for this operation.
5541 * @param[in] leaf The node being read.
5542 * @param[out] data Updated to point to the node's data.
5543 * @return 0 on success, non-zero on failure.
5546 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5548 MDB_page *omp; /* overflow page */
5552 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5553 data->mv_size = NODEDSZ(leaf);
5554 data->mv_data = NODEDATA(leaf);
5558 /* Read overflow data.
5560 data->mv_size = NODEDSZ(leaf);
5561 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5562 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5563 DPRINTF(("read overflow page %"Z"u failed", pgno));
5566 data->mv_data = METADATA(omp);
5572 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5573 MDB_val *key, MDB_val *data)
5580 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5582 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5585 if (txn->mt_flags & MDB_TXN_ERROR)
5588 mdb_cursor_init(&mc, txn, dbi, &mx);
5589 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5592 /** Find a sibling for a page.
5593 * Replaces the page at the top of the cursor's stack with the
5594 * specified sibling, if one exists.
5595 * @param[in] mc The cursor for this operation.
5596 * @param[in] move_right Non-zero if the right sibling is requested,
5597 * otherwise the left sibling.
5598 * @return 0 on success, non-zero on failure.
5601 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5607 if (mc->mc_snum < 2) {
5608 return MDB_NOTFOUND; /* root has no siblings */
5612 DPRINTF(("parent page is page %"Z"u, index %u",
5613 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5615 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5616 : (mc->mc_ki[mc->mc_top] == 0)) {
5617 DPRINTF(("no more keys left, moving to %s sibling",
5618 move_right ? "right" : "left"));
5619 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5620 /* undo cursor_pop before returning */
5627 mc->mc_ki[mc->mc_top]++;
5629 mc->mc_ki[mc->mc_top]--;
5630 DPRINTF(("just moving to %s index key %u",
5631 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5633 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5635 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5636 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5637 /* mc will be inconsistent if caller does mc_snum++ as above */
5638 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5642 mdb_cursor_push(mc, mp);
5644 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5649 /** Move the cursor to the next data item. */
5651 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5657 if (mc->mc_flags & C_EOF) {
5658 return MDB_NOTFOUND;
5661 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5663 mp = mc->mc_pg[mc->mc_top];
5665 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5666 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5667 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5668 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5669 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5670 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5671 if (rc == MDB_SUCCESS)
5672 MDB_GET_KEY(leaf, key);
5677 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5678 if (op == MDB_NEXT_DUP)
5679 return MDB_NOTFOUND;
5683 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5684 mdb_dbg_pgno(mp), (void *) mc));
5685 if (mc->mc_flags & C_DEL)
5688 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5689 DPUTS("=====> move to next sibling page");
5690 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5691 mc->mc_flags |= C_EOF;
5694 mp = mc->mc_pg[mc->mc_top];
5695 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5697 mc->mc_ki[mc->mc_top]++;
5700 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5701 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5704 key->mv_size = mc->mc_db->md_pad;
5705 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5709 mdb_cassert(mc, IS_LEAF(mp));
5710 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5712 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5713 mdb_xcursor_init1(mc, leaf);
5716 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5719 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5720 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5721 if (rc != MDB_SUCCESS)
5726 MDB_GET_KEY(leaf, key);
5730 /** Move the cursor to the previous data item. */
5732 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5738 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5740 mp = mc->mc_pg[mc->mc_top];
5742 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5743 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5744 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5745 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5746 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5747 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5748 if (rc == MDB_SUCCESS) {
5749 MDB_GET_KEY(leaf, key);
5750 mc->mc_flags &= ~C_EOF;
5756 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5757 if (op == MDB_PREV_DUP)
5758 return MDB_NOTFOUND;
5762 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5763 mdb_dbg_pgno(mp), (void *) mc));
5765 if (mc->mc_ki[mc->mc_top] == 0) {
5766 DPUTS("=====> move to prev sibling page");
5767 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5770 mp = mc->mc_pg[mc->mc_top];
5771 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5772 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5774 mc->mc_ki[mc->mc_top]--;
5776 mc->mc_flags &= ~C_EOF;
5778 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5779 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5782 key->mv_size = mc->mc_db->md_pad;
5783 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5787 mdb_cassert(mc, IS_LEAF(mp));
5788 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5790 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5791 mdb_xcursor_init1(mc, leaf);
5794 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5797 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5798 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5799 if (rc != MDB_SUCCESS)
5804 MDB_GET_KEY(leaf, key);
5808 /** Set the cursor on a specific data item. */
5810 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5811 MDB_cursor_op op, int *exactp)
5815 MDB_node *leaf = NULL;
5818 if (key->mv_size == 0)
5819 return MDB_BAD_VALSIZE;
5822 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5824 /* See if we're already on the right page */
5825 if (mc->mc_flags & C_INITIALIZED) {
5828 mp = mc->mc_pg[mc->mc_top];
5830 mc->mc_ki[mc->mc_top] = 0;
5831 return MDB_NOTFOUND;
5833 if (mp->mp_flags & P_LEAF2) {
5834 nodekey.mv_size = mc->mc_db->md_pad;
5835 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5837 leaf = NODEPTR(mp, 0);
5838 MDB_GET_KEY2(leaf, nodekey);
5840 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5842 /* Probably happens rarely, but first node on the page
5843 * was the one we wanted.
5845 mc->mc_ki[mc->mc_top] = 0;
5852 unsigned int nkeys = NUMKEYS(mp);
5854 if (mp->mp_flags & P_LEAF2) {
5855 nodekey.mv_data = LEAF2KEY(mp,
5856 nkeys-1, nodekey.mv_size);
5858 leaf = NODEPTR(mp, nkeys-1);
5859 MDB_GET_KEY2(leaf, nodekey);
5861 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5863 /* last node was the one we wanted */
5864 mc->mc_ki[mc->mc_top] = nkeys-1;
5870 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5871 /* This is definitely the right page, skip search_page */
5872 if (mp->mp_flags & P_LEAF2) {
5873 nodekey.mv_data = LEAF2KEY(mp,
5874 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5876 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5877 MDB_GET_KEY2(leaf, nodekey);
5879 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5881 /* current node was the one we wanted */
5891 /* If any parents have right-sibs, search.
5892 * Otherwise, there's nothing further.
5894 for (i=0; i<mc->mc_top; i++)
5896 NUMKEYS(mc->mc_pg[i])-1)
5898 if (i == mc->mc_top) {
5899 /* There are no other pages */
5900 mc->mc_ki[mc->mc_top] = nkeys;
5901 return MDB_NOTFOUND;
5905 /* There are no other pages */
5906 mc->mc_ki[mc->mc_top] = 0;
5907 if (op == MDB_SET_RANGE && !exactp) {
5911 return MDB_NOTFOUND;
5915 rc = mdb_page_search(mc, key, 0);
5916 if (rc != MDB_SUCCESS)
5919 mp = mc->mc_pg[mc->mc_top];
5920 mdb_cassert(mc, IS_LEAF(mp));
5923 leaf = mdb_node_search(mc, key, exactp);
5924 if (exactp != NULL && !*exactp) {
5925 /* MDB_SET specified and not an exact match. */
5926 return MDB_NOTFOUND;
5930 DPUTS("===> inexact leaf not found, goto sibling");
5931 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5932 mc->mc_flags |= C_EOF;
5933 return rc; /* no entries matched */
5935 mp = mc->mc_pg[mc->mc_top];
5936 mdb_cassert(mc, IS_LEAF(mp));
5937 leaf = NODEPTR(mp, 0);
5941 mc->mc_flags |= C_INITIALIZED;
5942 mc->mc_flags &= ~C_EOF;
5945 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5946 key->mv_size = mc->mc_db->md_pad;
5947 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5952 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5953 mdb_xcursor_init1(mc, leaf);
5956 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5957 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5958 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5961 if (op == MDB_GET_BOTH) {
5967 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5968 if (rc != MDB_SUCCESS)
5971 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5974 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5976 dcmp = mc->mc_dbx->md_dcmp;
5977 #if UINT_MAX < SIZE_MAX
5978 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5979 dcmp = mdb_cmp_clong;
5981 rc = dcmp(data, &olddata);
5983 if (op == MDB_GET_BOTH || rc > 0)
5984 return MDB_NOTFOUND;
5991 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5992 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5997 /* The key already matches in all other cases */
5998 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5999 MDB_GET_KEY(leaf, key);
6000 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6005 /** Move the cursor to the first item in the database. */
6007 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6013 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6015 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6016 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6017 if (rc != MDB_SUCCESS)
6020 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6022 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6023 mc->mc_flags |= C_INITIALIZED;
6024 mc->mc_flags &= ~C_EOF;
6026 mc->mc_ki[mc->mc_top] = 0;
6028 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6029 key->mv_size = mc->mc_db->md_pad;
6030 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6035 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6036 mdb_xcursor_init1(mc, leaf);
6037 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6041 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6045 MDB_GET_KEY(leaf, key);
6049 /** Move the cursor to the last item in the database. */
6051 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6057 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6059 if (!(mc->mc_flags & C_EOF)) {
6061 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6062 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6063 if (rc != MDB_SUCCESS)
6066 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6069 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6070 mc->mc_flags |= C_INITIALIZED|C_EOF;
6071 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6073 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6074 key->mv_size = mc->mc_db->md_pad;
6075 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6080 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6081 mdb_xcursor_init1(mc, leaf);
6082 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6086 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6091 MDB_GET_KEY(leaf, key);
6096 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6101 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6106 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6110 case MDB_GET_CURRENT:
6111 if (!(mc->mc_flags & C_INITIALIZED)) {
6114 MDB_page *mp = mc->mc_pg[mc->mc_top];
6115 int nkeys = NUMKEYS(mp);
6116 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6117 mc->mc_ki[mc->mc_top] = nkeys;
6123 key->mv_size = mc->mc_db->md_pad;
6124 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6126 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6127 MDB_GET_KEY(leaf, key);
6129 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6130 if (mc->mc_flags & C_DEL)
6131 mdb_xcursor_init1(mc, leaf);
6132 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6134 rc = mdb_node_read(mc->mc_txn, leaf, data);
6141 case MDB_GET_BOTH_RANGE:
6146 if (mc->mc_xcursor == NULL) {
6147 rc = MDB_INCOMPATIBLE;
6157 rc = mdb_cursor_set(mc, key, data, op,
6158 op == MDB_SET_RANGE ? NULL : &exact);
6161 case MDB_GET_MULTIPLE:
6162 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6166 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6167 rc = MDB_INCOMPATIBLE;
6171 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6172 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6175 case MDB_NEXT_MULTIPLE:
6180 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6181 rc = MDB_INCOMPATIBLE;
6184 if (!(mc->mc_flags & C_INITIALIZED))
6185 rc = mdb_cursor_first(mc, key, data);
6187 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6188 if (rc == MDB_SUCCESS) {
6189 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6192 mx = &mc->mc_xcursor->mx_cursor;
6193 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6195 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6196 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6204 case MDB_NEXT_NODUP:
6205 if (!(mc->mc_flags & C_INITIALIZED))
6206 rc = mdb_cursor_first(mc, key, data);
6208 rc = mdb_cursor_next(mc, key, data, op);
6212 case MDB_PREV_NODUP:
6213 if (!(mc->mc_flags & C_INITIALIZED)) {
6214 rc = mdb_cursor_last(mc, key, data);
6217 mc->mc_flags |= C_INITIALIZED;
6218 mc->mc_ki[mc->mc_top]++;
6220 rc = mdb_cursor_prev(mc, key, data, op);
6223 rc = mdb_cursor_first(mc, key, data);
6226 mfunc = mdb_cursor_first;
6228 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6232 if (mc->mc_xcursor == NULL) {
6233 rc = MDB_INCOMPATIBLE;
6237 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6238 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6239 MDB_GET_KEY(leaf, key);
6240 rc = mdb_node_read(mc->mc_txn, leaf, data);
6244 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6248 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6251 rc = mdb_cursor_last(mc, key, data);
6254 mfunc = mdb_cursor_last;
6257 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6262 if (mc->mc_flags & C_DEL)
6263 mc->mc_flags ^= C_DEL;
6268 /** Touch all the pages in the cursor stack. Set mc_top.
6269 * Makes sure all the pages are writable, before attempting a write operation.
6270 * @param[in] mc The cursor to operate on.
6273 mdb_cursor_touch(MDB_cursor *mc)
6275 int rc = MDB_SUCCESS;
6277 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6280 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6282 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6283 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6286 *mc->mc_dbflag |= DB_DIRTY;
6291 rc = mdb_page_touch(mc);
6292 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6293 mc->mc_top = mc->mc_snum-1;
6298 /** Do not spill pages to disk if txn is getting full, may fail instead */
6299 #define MDB_NOSPILL 0x8000
6302 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6306 MDB_node *leaf = NULL;
6309 MDB_val xdata, *rdata, dkey, olddata;
6311 int do_sub = 0, insert_key, insert_data;
6312 unsigned int mcount = 0, dcount = 0, nospill;
6315 unsigned int nflags;
6318 if (mc == NULL || key == NULL)
6321 env = mc->mc_txn->mt_env;
6323 /* Check this first so counter will always be zero on any
6326 if (flags & MDB_MULTIPLE) {
6327 dcount = data[1].mv_size;
6328 data[1].mv_size = 0;
6329 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6330 return MDB_INCOMPATIBLE;
6333 nospill = flags & MDB_NOSPILL;
6334 flags &= ~MDB_NOSPILL;
6336 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6337 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6339 if (key->mv_size-1 >= ENV_MAXKEY(env))
6340 return MDB_BAD_VALSIZE;
6342 #if SIZE_MAX > MAXDATASIZE
6343 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6344 return MDB_BAD_VALSIZE;
6346 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6347 return MDB_BAD_VALSIZE;
6350 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6351 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6355 if (flags == MDB_CURRENT) {
6356 if (!(mc->mc_flags & C_INITIALIZED))
6359 } else if (mc->mc_db->md_root == P_INVALID) {
6360 /* new database, cursor has nothing to point to */
6363 mc->mc_flags &= ~C_INITIALIZED;
6368 if (flags & MDB_APPEND) {
6370 rc = mdb_cursor_last(mc, &k2, &d2);
6372 rc = mc->mc_dbx->md_cmp(key, &k2);
6375 mc->mc_ki[mc->mc_top]++;
6377 /* new key is <= last key */
6382 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6384 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6385 DPRINTF(("duplicate key [%s]", DKEY(key)));
6387 return MDB_KEYEXIST;
6389 if (rc && rc != MDB_NOTFOUND)
6393 if (mc->mc_flags & C_DEL)
6394 mc->mc_flags ^= C_DEL;
6396 /* Cursor is positioned, check for room in the dirty list */
6398 if (flags & MDB_MULTIPLE) {
6400 xdata.mv_size = data->mv_size * dcount;
6404 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6408 if (rc == MDB_NO_ROOT) {
6410 /* new database, write a root leaf page */
6411 DPUTS("allocating new root leaf page");
6412 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6415 mdb_cursor_push(mc, np);
6416 mc->mc_db->md_root = np->mp_pgno;
6417 mc->mc_db->md_depth++;
6418 *mc->mc_dbflag |= DB_DIRTY;
6419 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6421 np->mp_flags |= P_LEAF2;
6422 mc->mc_flags |= C_INITIALIZED;
6424 /* make sure all cursor pages are writable */
6425 rc2 = mdb_cursor_touch(mc);
6430 insert_key = insert_data = rc;
6432 /* The key does not exist */
6433 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6434 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6435 LEAFSIZE(key, data) > env->me_nodemax)
6437 /* Too big for a node, insert in sub-DB. Set up an empty
6438 * "old sub-page" for prep_subDB to expand to a full page.
6440 fp_flags = P_LEAF|P_DIRTY;
6442 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6443 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6444 olddata.mv_size = PAGEHDRSZ;
6448 /* there's only a key anyway, so this is a no-op */
6449 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6451 unsigned int ksize = mc->mc_db->md_pad;
6452 if (key->mv_size != ksize)
6453 return MDB_BAD_VALSIZE;
6454 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6455 memcpy(ptr, key->mv_data, ksize);
6457 /* if overwriting slot 0 of leaf, need to
6458 * update branch key if there is a parent page
6460 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6461 unsigned short top = mc->mc_top;
6463 /* slot 0 is always an empty key, find real slot */
6464 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6466 if (mc->mc_ki[mc->mc_top])
6467 rc2 = mdb_update_key(mc, key);
6478 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6479 olddata.mv_size = NODEDSZ(leaf);
6480 olddata.mv_data = NODEDATA(leaf);
6483 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6484 /* Prepare (sub-)page/sub-DB to accept the new item,
6485 * if needed. fp: old sub-page or a header faking
6486 * it. mp: new (sub-)page. offset: growth in page
6487 * size. xdata: node data with new page or DB.
6489 unsigned i, offset = 0;
6490 mp = fp = xdata.mv_data = env->me_pbuf;
6491 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6493 /* Was a single item before, must convert now */
6494 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6496 /* Just overwrite the current item */
6497 if (flags == MDB_CURRENT)
6499 dcmp = mc->mc_dbx->md_dcmp;
6500 #if UINT_MAX < SIZE_MAX
6501 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6502 dcmp = mdb_cmp_clong;
6504 /* does data match? */
6505 if (!dcmp(data, &olddata)) {
6506 if (flags & MDB_NODUPDATA)
6507 return MDB_KEYEXIST;
6512 /* Back up original data item */
6513 dkey.mv_size = olddata.mv_size;
6514 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6516 /* Make sub-page header for the dup items, with dummy body */
6517 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6518 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6519 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6520 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6521 fp->mp_flags |= P_LEAF2;
6522 fp->mp_pad = data->mv_size;
6523 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6525 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6526 (dkey.mv_size & 1) + (data->mv_size & 1);
6528 fp->mp_upper = xdata.mv_size - PAGEBASE;
6529 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6530 } else if (leaf->mn_flags & F_SUBDATA) {
6531 /* Data is on sub-DB, just store it */
6532 flags |= F_DUPDATA|F_SUBDATA;
6535 /* Data is on sub-page */
6536 fp = olddata.mv_data;
6539 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6540 offset = EVEN(NODESIZE + sizeof(indx_t) +
6544 offset = fp->mp_pad;
6545 if (SIZELEFT(fp) < offset) {
6546 offset *= 4; /* space for 4 more */
6549 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6551 fp->mp_flags |= P_DIRTY;
6552 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6553 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6557 xdata.mv_size = olddata.mv_size + offset;
6560 fp_flags = fp->mp_flags;
6561 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6562 /* Too big for a sub-page, convert to sub-DB */
6563 fp_flags &= ~P_SUBP;
6565 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6566 fp_flags |= P_LEAF2;
6567 dummy.md_pad = fp->mp_pad;
6568 dummy.md_flags = MDB_DUPFIXED;
6569 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6570 dummy.md_flags |= MDB_INTEGERKEY;
6576 dummy.md_branch_pages = 0;
6577 dummy.md_leaf_pages = 1;
6578 dummy.md_overflow_pages = 0;
6579 dummy.md_entries = NUMKEYS(fp);
6580 xdata.mv_size = sizeof(MDB_db);
6581 xdata.mv_data = &dummy;
6582 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6584 offset = env->me_psize - olddata.mv_size;
6585 flags |= F_DUPDATA|F_SUBDATA;
6586 dummy.md_root = mp->mp_pgno;
6589 mp->mp_flags = fp_flags | P_DIRTY;
6590 mp->mp_pad = fp->mp_pad;
6591 mp->mp_lower = fp->mp_lower;
6592 mp->mp_upper = fp->mp_upper + offset;
6593 if (fp_flags & P_LEAF2) {
6594 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6596 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6597 olddata.mv_size - fp->mp_upper - PAGEBASE);
6598 for (i=0; i<NUMKEYS(fp); i++)
6599 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6607 mdb_node_del(mc, 0);
6611 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6612 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6613 return MDB_INCOMPATIBLE;
6614 /* overflow page overwrites need special handling */
6615 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6618 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6620 memcpy(&pg, olddata.mv_data, sizeof(pg));
6621 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6623 ovpages = omp->mp_pages;
6625 /* Is the ov page large enough? */
6626 if (ovpages >= dpages) {
6627 if (!(omp->mp_flags & P_DIRTY) &&
6628 (level || (env->me_flags & MDB_WRITEMAP)))
6630 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6633 level = 0; /* dirty in this txn or clean */
6636 if (omp->mp_flags & P_DIRTY) {
6637 /* yes, overwrite it. Note in this case we don't
6638 * bother to try shrinking the page if the new data
6639 * is smaller than the overflow threshold.
6642 /* It is writable only in a parent txn */
6643 size_t sz = (size_t) env->me_psize * ovpages, off;
6644 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6650 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6651 mdb_cassert(mc, rc2 == 0);
6652 if (!(flags & MDB_RESERVE)) {
6653 /* Copy end of page, adjusting alignment so
6654 * compiler may copy words instead of bytes.
6656 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6657 memcpy((size_t *)((char *)np + off),
6658 (size_t *)((char *)omp + off), sz - off);
6661 memcpy(np, omp, sz); /* Copy beginning of page */
6664 SETDSZ(leaf, data->mv_size);
6665 if (F_ISSET(flags, MDB_RESERVE))
6666 data->mv_data = METADATA(omp);
6668 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6672 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6674 } else if (data->mv_size == olddata.mv_size) {
6675 /* same size, just replace it. Note that we could
6676 * also reuse this node if the new data is smaller,
6677 * but instead we opt to shrink the node in that case.
6679 if (F_ISSET(flags, MDB_RESERVE))
6680 data->mv_data = olddata.mv_data;
6681 else if (!(mc->mc_flags & C_SUB))
6682 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6684 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6689 mdb_node_del(mc, 0);
6695 nflags = flags & NODE_ADD_FLAGS;
6696 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6697 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6698 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6699 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6701 nflags |= MDB_SPLIT_REPLACE;
6702 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6704 /* There is room already in this leaf page. */
6705 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6706 if (rc == 0 && insert_key) {
6707 /* Adjust other cursors pointing to mp */
6708 MDB_cursor *m2, *m3;
6709 MDB_dbi dbi = mc->mc_dbi;
6710 unsigned i = mc->mc_top;
6711 MDB_page *mp = mc->mc_pg[i];
6713 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6714 if (mc->mc_flags & C_SUB)
6715 m3 = &m2->mc_xcursor->mx_cursor;
6718 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6719 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6726 if (rc == MDB_SUCCESS) {
6727 /* Now store the actual data in the child DB. Note that we're
6728 * storing the user data in the keys field, so there are strict
6729 * size limits on dupdata. The actual data fields of the child
6730 * DB are all zero size.
6738 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6739 if (flags & MDB_CURRENT) {
6740 xflags = MDB_CURRENT|MDB_NOSPILL;
6742 mdb_xcursor_init1(mc, leaf);
6743 xflags = (flags & MDB_NODUPDATA) ?
6744 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6746 /* converted, write the original data first */
6748 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6752 /* Adjust other cursors pointing to mp */
6754 unsigned i = mc->mc_top;
6755 MDB_page *mp = mc->mc_pg[i];
6757 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6758 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6759 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6760 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6761 mdb_xcursor_init1(m2, leaf);
6765 /* we've done our job */
6768 ecount = mc->mc_xcursor->mx_db.md_entries;
6769 if (flags & MDB_APPENDDUP)
6770 xflags |= MDB_APPEND;
6771 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6772 if (flags & F_SUBDATA) {
6773 void *db = NODEDATA(leaf);
6774 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6776 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6778 /* Increment count unless we just replaced an existing item. */
6780 mc->mc_db->md_entries++;
6782 /* Invalidate txn if we created an empty sub-DB */
6785 /* If we succeeded and the key didn't exist before,
6786 * make sure the cursor is marked valid.
6788 mc->mc_flags |= C_INITIALIZED;
6790 if (flags & MDB_MULTIPLE) {
6793 /* let caller know how many succeeded, if any */
6794 data[1].mv_size = mcount;
6795 if (mcount < dcount) {
6796 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6797 insert_key = insert_data = 0;
6804 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6807 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6812 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6818 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6819 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6821 if (!(mc->mc_flags & C_INITIALIZED))
6824 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6825 return MDB_NOTFOUND;
6827 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6830 rc = mdb_cursor_touch(mc);
6834 mp = mc->mc_pg[mc->mc_top];
6837 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6839 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6840 if (flags & MDB_NODUPDATA) {
6841 /* mdb_cursor_del0() will subtract the final entry */
6842 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6844 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6845 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6847 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6850 /* If sub-DB still has entries, we're done */
6851 if (mc->mc_xcursor->mx_db.md_entries) {
6852 if (leaf->mn_flags & F_SUBDATA) {
6853 /* update subDB info */
6854 void *db = NODEDATA(leaf);
6855 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6858 /* shrink fake page */
6859 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6860 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6861 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6862 /* fix other sub-DB cursors pointed at this fake page */
6863 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6864 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6865 if (m2->mc_pg[mc->mc_top] == mp &&
6866 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6867 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6870 mc->mc_db->md_entries--;
6871 mc->mc_flags |= C_DEL;
6874 /* otherwise fall thru and delete the sub-DB */
6877 if (leaf->mn_flags & F_SUBDATA) {
6878 /* add all the child DB's pages to the free list */
6879 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6884 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6885 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6886 rc = MDB_INCOMPATIBLE;
6890 /* add overflow pages to free list */
6891 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6895 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6896 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6897 (rc = mdb_ovpage_free(mc, omp)))
6902 return mdb_cursor_del0(mc);
6905 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6909 /** Allocate and initialize new pages for a database.
6910 * @param[in] mc a cursor on the database being added to.
6911 * @param[in] flags flags defining what type of page is being allocated.
6912 * @param[in] num the number of pages to allocate. This is usually 1,
6913 * unless allocating overflow pages for a large record.
6914 * @param[out] mp Address of a page, or NULL on failure.
6915 * @return 0 on success, non-zero on failure.
6918 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6923 if ((rc = mdb_page_alloc(mc, num, &np)))
6925 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6926 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6927 np->mp_flags = flags | P_DIRTY;
6928 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6929 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6932 mc->mc_db->md_branch_pages++;
6933 else if (IS_LEAF(np))
6934 mc->mc_db->md_leaf_pages++;
6935 else if (IS_OVERFLOW(np)) {
6936 mc->mc_db->md_overflow_pages += num;
6944 /** Calculate the size of a leaf node.
6945 * The size depends on the environment's page size; if a data item
6946 * is too large it will be put onto an overflow page and the node
6947 * size will only include the key and not the data. Sizes are always
6948 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6949 * of the #MDB_node headers.
6950 * @param[in] env The environment handle.
6951 * @param[in] key The key for the node.
6952 * @param[in] data The data for the node.
6953 * @return The number of bytes needed to store the node.
6956 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6960 sz = LEAFSIZE(key, data);
6961 if (sz > env->me_nodemax) {
6962 /* put on overflow page */
6963 sz -= data->mv_size - sizeof(pgno_t);
6966 return EVEN(sz + sizeof(indx_t));
6969 /** Calculate the size of a branch node.
6970 * The size should depend on the environment's page size but since
6971 * we currently don't support spilling large keys onto overflow
6972 * pages, it's simply the size of the #MDB_node header plus the
6973 * size of the key. Sizes are always rounded up to an even number
6974 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6975 * @param[in] env The environment handle.
6976 * @param[in] key The key for the node.
6977 * @return The number of bytes needed to store the node.
6980 mdb_branch_size(MDB_env *env, MDB_val *key)
6985 if (sz > env->me_nodemax) {
6986 /* put on overflow page */
6987 /* not implemented */
6988 /* sz -= key->size - sizeof(pgno_t); */
6991 return sz + sizeof(indx_t);
6994 /** Add a node to the page pointed to by the cursor.
6995 * @param[in] mc The cursor for this operation.
6996 * @param[in] indx The index on the page where the new node should be added.
6997 * @param[in] key The key for the new node.
6998 * @param[in] data The data for the new node, if any.
6999 * @param[in] pgno The page number, if adding a branch node.
7000 * @param[in] flags Flags for the node.
7001 * @return 0 on success, non-zero on failure. Possible errors are:
7003 * <li>ENOMEM - failed to allocate overflow pages for the node.
7004 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7005 * should never happen since all callers already calculate the
7006 * page's free space before calling this function.
7010 mdb_node_add(MDB_cursor *mc, indx_t indx,
7011 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7014 size_t node_size = NODESIZE;
7018 MDB_page *mp = mc->mc_pg[mc->mc_top];
7019 MDB_page *ofp = NULL; /* overflow page */
7023 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7025 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7026 IS_LEAF(mp) ? "leaf" : "branch",
7027 IS_SUBP(mp) ? "sub-" : "",
7028 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7029 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7032 /* Move higher keys up one slot. */
7033 int ksize = mc->mc_db->md_pad, dif;
7034 char *ptr = LEAF2KEY(mp, indx, ksize);
7035 dif = NUMKEYS(mp) - indx;
7037 memmove(ptr+ksize, ptr, dif*ksize);
7038 /* insert new key */
7039 memcpy(ptr, key->mv_data, ksize);
7041 /* Just using these for counting */
7042 mp->mp_lower += sizeof(indx_t);
7043 mp->mp_upper -= ksize - sizeof(indx_t);
7047 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7049 node_size += key->mv_size;
7051 mdb_cassert(mc, key && data);
7052 if (F_ISSET(flags, F_BIGDATA)) {
7053 /* Data already on overflow page. */
7054 node_size += sizeof(pgno_t);
7055 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7056 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7058 /* Put data on overflow page. */
7059 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7060 data->mv_size, node_size+data->mv_size));
7061 node_size = EVEN(node_size + sizeof(pgno_t));
7062 if ((ssize_t)node_size > room)
7064 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7066 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7070 node_size += data->mv_size;
7073 node_size = EVEN(node_size);
7074 if ((ssize_t)node_size > room)
7078 /* Move higher pointers up one slot. */
7079 for (i = NUMKEYS(mp); i > indx; i--)
7080 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7082 /* Adjust free space offsets. */
7083 ofs = mp->mp_upper - node_size;
7084 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7085 mp->mp_ptrs[indx] = ofs;
7087 mp->mp_lower += sizeof(indx_t);
7089 /* Write the node data. */
7090 node = NODEPTR(mp, indx);
7091 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7092 node->mn_flags = flags;
7094 SETDSZ(node,data->mv_size);
7099 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7102 ndata = NODEDATA(node);
7104 if (F_ISSET(flags, F_BIGDATA))
7105 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7106 else if (F_ISSET(flags, MDB_RESERVE))
7107 data->mv_data = ndata;
7109 memcpy(ndata, data->mv_data, data->mv_size);
7111 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7112 ndata = METADATA(ofp);
7113 if (F_ISSET(flags, MDB_RESERVE))
7114 data->mv_data = ndata;
7116 memcpy(ndata, data->mv_data, data->mv_size);
7123 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7124 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7125 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7126 DPRINTF(("node size = %"Z"u", node_size));
7127 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7128 return MDB_PAGE_FULL;
7131 /** Delete the specified node from a page.
7132 * @param[in] mc Cursor pointing to the node to delete.
7133 * @param[in] ksize The size of a node. Only used if the page is
7134 * part of a #MDB_DUPFIXED database.
7137 mdb_node_del(MDB_cursor *mc, int ksize)
7139 MDB_page *mp = mc->mc_pg[mc->mc_top];
7140 indx_t indx = mc->mc_ki[mc->mc_top];
7142 indx_t i, j, numkeys, ptr;
7146 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7147 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7148 numkeys = NUMKEYS(mp);
7149 mdb_cassert(mc, indx < numkeys);
7152 int x = numkeys - 1 - indx;
7153 base = LEAF2KEY(mp, indx, ksize);
7155 memmove(base, base + ksize, x * ksize);
7156 mp->mp_lower -= sizeof(indx_t);
7157 mp->mp_upper += ksize - sizeof(indx_t);
7161 node = NODEPTR(mp, indx);
7162 sz = NODESIZE + node->mn_ksize;
7164 if (F_ISSET(node->mn_flags, F_BIGDATA))
7165 sz += sizeof(pgno_t);
7167 sz += NODEDSZ(node);
7171 ptr = mp->mp_ptrs[indx];
7172 for (i = j = 0; i < numkeys; i++) {
7174 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7175 if (mp->mp_ptrs[i] < ptr)
7176 mp->mp_ptrs[j] += sz;
7181 base = (char *)mp + mp->mp_upper + PAGEBASE;
7182 memmove(base + sz, base, ptr - mp->mp_upper);
7184 mp->mp_lower -= sizeof(indx_t);
7188 /** Compact the main page after deleting a node on a subpage.
7189 * @param[in] mp The main page to operate on.
7190 * @param[in] indx The index of the subpage on the main page.
7193 mdb_node_shrink(MDB_page *mp, indx_t indx)
7198 indx_t delta, nsize, len, ptr;
7201 node = NODEPTR(mp, indx);
7202 sp = (MDB_page *)NODEDATA(node);
7203 delta = SIZELEFT(sp);
7204 nsize = NODEDSZ(node) - delta;
7206 /* Prepare to shift upward, set len = length(subpage part to shift) */
7210 return; /* do not make the node uneven-sized */
7212 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7213 for (i = NUMKEYS(sp); --i >= 0; )
7214 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7217 sp->mp_upper = sp->mp_lower;
7218 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7219 SETDSZ(node, nsize);
7221 /* Shift <lower nodes...initial part of subpage> upward */
7222 base = (char *)mp + mp->mp_upper + PAGEBASE;
7223 memmove(base + delta, base, (char *)sp + len - base);
7225 ptr = mp->mp_ptrs[indx];
7226 for (i = NUMKEYS(mp); --i >= 0; ) {
7227 if (mp->mp_ptrs[i] <= ptr)
7228 mp->mp_ptrs[i] += delta;
7230 mp->mp_upper += delta;
7233 /** Initial setup of a sorted-dups cursor.
7234 * Sorted duplicates are implemented as a sub-database for the given key.
7235 * The duplicate data items are actually keys of the sub-database.
7236 * Operations on the duplicate data items are performed using a sub-cursor
7237 * initialized when the sub-database is first accessed. This function does
7238 * the preliminary setup of the sub-cursor, filling in the fields that
7239 * depend only on the parent DB.
7240 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7243 mdb_xcursor_init0(MDB_cursor *mc)
7245 MDB_xcursor *mx = mc->mc_xcursor;
7247 mx->mx_cursor.mc_xcursor = NULL;
7248 mx->mx_cursor.mc_txn = mc->mc_txn;
7249 mx->mx_cursor.mc_db = &mx->mx_db;
7250 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7251 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7252 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7253 mx->mx_cursor.mc_snum = 0;
7254 mx->mx_cursor.mc_top = 0;
7255 mx->mx_cursor.mc_flags = C_SUB;
7256 mx->mx_dbx.md_name.mv_size = 0;
7257 mx->mx_dbx.md_name.mv_data = NULL;
7258 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7259 mx->mx_dbx.md_dcmp = NULL;
7260 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7263 /** Final setup of a sorted-dups cursor.
7264 * Sets up the fields that depend on the data from the main cursor.
7265 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7266 * @param[in] node The data containing the #MDB_db record for the
7267 * sorted-dup database.
7270 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7272 MDB_xcursor *mx = mc->mc_xcursor;
7274 if (node->mn_flags & F_SUBDATA) {
7275 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7276 mx->mx_cursor.mc_pg[0] = 0;
7277 mx->mx_cursor.mc_snum = 0;
7278 mx->mx_cursor.mc_top = 0;
7279 mx->mx_cursor.mc_flags = C_SUB;
7281 MDB_page *fp = NODEDATA(node);
7282 mx->mx_db.md_pad = 0;
7283 mx->mx_db.md_flags = 0;
7284 mx->mx_db.md_depth = 1;
7285 mx->mx_db.md_branch_pages = 0;
7286 mx->mx_db.md_leaf_pages = 1;
7287 mx->mx_db.md_overflow_pages = 0;
7288 mx->mx_db.md_entries = NUMKEYS(fp);
7289 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7290 mx->mx_cursor.mc_snum = 1;
7291 mx->mx_cursor.mc_top = 0;
7292 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7293 mx->mx_cursor.mc_pg[0] = fp;
7294 mx->mx_cursor.mc_ki[0] = 0;
7295 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7296 mx->mx_db.md_flags = MDB_DUPFIXED;
7297 mx->mx_db.md_pad = fp->mp_pad;
7298 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7299 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7302 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7303 mx->mx_db.md_root));
7304 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7305 #if UINT_MAX < SIZE_MAX
7306 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7307 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7311 /** Initialize a cursor for a given transaction and database. */
7313 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7316 mc->mc_backup = NULL;
7319 mc->mc_db = &txn->mt_dbs[dbi];
7320 mc->mc_dbx = &txn->mt_dbxs[dbi];
7321 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7327 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7328 mdb_tassert(txn, mx != NULL);
7329 mc->mc_xcursor = mx;
7330 mdb_xcursor_init0(mc);
7332 mc->mc_xcursor = NULL;
7334 if (*mc->mc_dbflag & DB_STALE) {
7335 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7340 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7343 size_t size = sizeof(MDB_cursor);
7345 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7348 if (txn->mt_flags & MDB_TXN_ERROR)
7351 /* Allow read access to the freelist */
7352 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7355 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7356 size += sizeof(MDB_xcursor);
7358 if ((mc = malloc(size)) != NULL) {
7359 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7360 if (txn->mt_cursors) {
7361 mc->mc_next = txn->mt_cursors[dbi];
7362 txn->mt_cursors[dbi] = mc;
7363 mc->mc_flags |= C_UNTRACK;
7375 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7377 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7380 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7383 if (txn->mt_flags & MDB_TXN_ERROR)
7386 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7390 /* Return the count of duplicate data items for the current key */
7392 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7396 if (mc == NULL || countp == NULL)
7399 if (mc->mc_xcursor == NULL)
7400 return MDB_INCOMPATIBLE;
7402 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7405 if (!(mc->mc_flags & C_INITIALIZED))
7408 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7409 return MDB_NOTFOUND;
7411 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7412 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7415 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7418 *countp = mc->mc_xcursor->mx_db.md_entries;
7424 mdb_cursor_close(MDB_cursor *mc)
7426 if (mc && !mc->mc_backup) {
7427 /* remove from txn, if tracked */
7428 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7429 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7430 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7432 *prev = mc->mc_next;
7439 mdb_cursor_txn(MDB_cursor *mc)
7441 if (!mc) return NULL;
7446 mdb_cursor_dbi(MDB_cursor *mc)
7451 /** Replace the key for a branch node with a new key.
7452 * @param[in] mc Cursor pointing to the node to operate on.
7453 * @param[in] key The new key to use.
7454 * @return 0 on success, non-zero on failure.
7457 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7463 int delta, ksize, oksize;
7464 indx_t ptr, i, numkeys, indx;
7467 indx = mc->mc_ki[mc->mc_top];
7468 mp = mc->mc_pg[mc->mc_top];
7469 node = NODEPTR(mp, indx);
7470 ptr = mp->mp_ptrs[indx];
7474 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7475 k2.mv_data = NODEKEY(node);
7476 k2.mv_size = node->mn_ksize;
7477 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7479 mdb_dkey(&k2, kbuf2),
7485 /* Sizes must be 2-byte aligned. */
7486 ksize = EVEN(key->mv_size);
7487 oksize = EVEN(node->mn_ksize);
7488 delta = ksize - oksize;
7490 /* Shift node contents if EVEN(key length) changed. */
7492 if (delta > 0 && SIZELEFT(mp) < delta) {
7494 /* not enough space left, do a delete and split */
7495 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7496 pgno = NODEPGNO(node);
7497 mdb_node_del(mc, 0);
7498 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7501 numkeys = NUMKEYS(mp);
7502 for (i = 0; i < numkeys; i++) {
7503 if (mp->mp_ptrs[i] <= ptr)
7504 mp->mp_ptrs[i] -= delta;
7507 base = (char *)mp + mp->mp_upper + PAGEBASE;
7508 len = ptr - mp->mp_upper + NODESIZE;
7509 memmove(base - delta, base, len);
7510 mp->mp_upper -= delta;
7512 node = NODEPTR(mp, indx);
7515 /* But even if no shift was needed, update ksize */
7516 if (node->mn_ksize != key->mv_size)
7517 node->mn_ksize = key->mv_size;
7520 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7526 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7528 /** Move a node from csrc to cdst.
7531 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7538 unsigned short flags;
7542 /* Mark src and dst as dirty. */
7543 if ((rc = mdb_page_touch(csrc)) ||
7544 (rc = mdb_page_touch(cdst)))
7547 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7548 key.mv_size = csrc->mc_db->md_pad;
7549 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7551 data.mv_data = NULL;
7555 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7556 mdb_cassert(csrc, !((size_t)srcnode & 1));
7557 srcpg = NODEPGNO(srcnode);
7558 flags = srcnode->mn_flags;
7559 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7560 unsigned int snum = csrc->mc_snum;
7562 /* must find the lowest key below src */
7563 rc = mdb_page_search_lowest(csrc);
7566 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7567 key.mv_size = csrc->mc_db->md_pad;
7568 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7570 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7571 key.mv_size = NODEKSZ(s2);
7572 key.mv_data = NODEKEY(s2);
7574 csrc->mc_snum = snum--;
7575 csrc->mc_top = snum;
7577 key.mv_size = NODEKSZ(srcnode);
7578 key.mv_data = NODEKEY(srcnode);
7580 data.mv_size = NODEDSZ(srcnode);
7581 data.mv_data = NODEDATA(srcnode);
7583 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7584 unsigned int snum = cdst->mc_snum;
7587 /* must find the lowest key below dst */
7588 mdb_cursor_copy(cdst, &mn);
7589 rc = mdb_page_search_lowest(&mn);
7592 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7593 bkey.mv_size = mn.mc_db->md_pad;
7594 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7596 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7597 bkey.mv_size = NODEKSZ(s2);
7598 bkey.mv_data = NODEKEY(s2);
7600 mn.mc_snum = snum--;
7603 rc = mdb_update_key(&mn, &bkey);
7608 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7609 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7610 csrc->mc_ki[csrc->mc_top],
7612 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7613 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7615 /* Add the node to the destination page.
7617 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7618 if (rc != MDB_SUCCESS)
7621 /* Delete the node from the source page.
7623 mdb_node_del(csrc, key.mv_size);
7626 /* Adjust other cursors pointing to mp */
7627 MDB_cursor *m2, *m3;
7628 MDB_dbi dbi = csrc->mc_dbi;
7629 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7631 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7632 if (csrc->mc_flags & C_SUB)
7633 m3 = &m2->mc_xcursor->mx_cursor;
7636 if (m3 == csrc) continue;
7637 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7638 csrc->mc_ki[csrc->mc_top]) {
7639 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7640 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7645 /* Update the parent separators.
7647 if (csrc->mc_ki[csrc->mc_top] == 0) {
7648 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7649 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7650 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7652 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7653 key.mv_size = NODEKSZ(srcnode);
7654 key.mv_data = NODEKEY(srcnode);
7656 DPRINTF(("update separator for source page %"Z"u to [%s]",
7657 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7658 mdb_cursor_copy(csrc, &mn);
7661 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7664 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7666 indx_t ix = csrc->mc_ki[csrc->mc_top];
7667 nullkey.mv_size = 0;
7668 csrc->mc_ki[csrc->mc_top] = 0;
7669 rc = mdb_update_key(csrc, &nullkey);
7670 csrc->mc_ki[csrc->mc_top] = ix;
7671 mdb_cassert(csrc, rc == MDB_SUCCESS);
7675 if (cdst->mc_ki[cdst->mc_top] == 0) {
7676 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7677 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7678 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7680 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7681 key.mv_size = NODEKSZ(srcnode);
7682 key.mv_data = NODEKEY(srcnode);
7684 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7685 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7686 mdb_cursor_copy(cdst, &mn);
7689 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7692 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7694 indx_t ix = cdst->mc_ki[cdst->mc_top];
7695 nullkey.mv_size = 0;
7696 cdst->mc_ki[cdst->mc_top] = 0;
7697 rc = mdb_update_key(cdst, &nullkey);
7698 cdst->mc_ki[cdst->mc_top] = ix;
7699 mdb_cassert(cdst, rc == MDB_SUCCESS);
7706 /** Merge one page into another.
7707 * The nodes from the page pointed to by \b csrc will
7708 * be copied to the page pointed to by \b cdst and then
7709 * the \b csrc page will be freed.
7710 * @param[in] csrc Cursor pointing to the source page.
7711 * @param[in] cdst Cursor pointing to the destination page.
7712 * @return 0 on success, non-zero on failure.
7715 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7717 MDB_page *psrc, *pdst;
7724 psrc = csrc->mc_pg[csrc->mc_top];
7725 pdst = cdst->mc_pg[cdst->mc_top];
7727 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7729 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7730 mdb_cassert(csrc, cdst->mc_snum > 1);
7732 /* Mark dst as dirty. */
7733 if ((rc = mdb_page_touch(cdst)))
7736 /* Move all nodes from src to dst.
7738 j = nkeys = NUMKEYS(pdst);
7739 if (IS_LEAF2(psrc)) {
7740 key.mv_size = csrc->mc_db->md_pad;
7741 key.mv_data = METADATA(psrc);
7742 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7743 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7744 if (rc != MDB_SUCCESS)
7746 key.mv_data = (char *)key.mv_data + key.mv_size;
7749 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7750 srcnode = NODEPTR(psrc, i);
7751 if (i == 0 && IS_BRANCH(psrc)) {
7754 mdb_cursor_copy(csrc, &mn);
7755 /* must find the lowest key below src */
7756 rc = mdb_page_search_lowest(&mn);
7759 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7760 key.mv_size = mn.mc_db->md_pad;
7761 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7763 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7764 key.mv_size = NODEKSZ(s2);
7765 key.mv_data = NODEKEY(s2);
7768 key.mv_size = srcnode->mn_ksize;
7769 key.mv_data = NODEKEY(srcnode);
7772 data.mv_size = NODEDSZ(srcnode);
7773 data.mv_data = NODEDATA(srcnode);
7774 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7775 if (rc != MDB_SUCCESS)
7780 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7781 pdst->mp_pgno, NUMKEYS(pdst),
7782 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7784 /* Unlink the src page from parent and add to free list.
7787 mdb_node_del(csrc, 0);
7788 if (csrc->mc_ki[csrc->mc_top] == 0) {
7790 rc = mdb_update_key(csrc, &key);
7798 psrc = csrc->mc_pg[csrc->mc_top];
7799 /* If not operating on FreeDB, allow this page to be reused
7800 * in this txn. Otherwise just add to free list.
7802 rc = mdb_page_loose(csrc, psrc);
7806 csrc->mc_db->md_leaf_pages--;
7808 csrc->mc_db->md_branch_pages--;
7810 /* Adjust other cursors pointing to mp */
7811 MDB_cursor *m2, *m3;
7812 MDB_dbi dbi = csrc->mc_dbi;
7814 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7815 if (csrc->mc_flags & C_SUB)
7816 m3 = &m2->mc_xcursor->mx_cursor;
7819 if (m3 == csrc) continue;
7820 if (m3->mc_snum < csrc->mc_snum) continue;
7821 if (m3->mc_pg[csrc->mc_top] == psrc) {
7822 m3->mc_pg[csrc->mc_top] = pdst;
7823 m3->mc_ki[csrc->mc_top] += nkeys;
7828 unsigned int snum = cdst->mc_snum;
7829 uint16_t depth = cdst->mc_db->md_depth;
7830 mdb_cursor_pop(cdst);
7831 rc = mdb_rebalance(cdst);
7832 /* Did the tree shrink? */
7833 if (depth > cdst->mc_db->md_depth)
7835 cdst->mc_snum = snum;
7836 cdst->mc_top = snum-1;
7841 /** Copy the contents of a cursor.
7842 * @param[in] csrc The cursor to copy from.
7843 * @param[out] cdst The cursor to copy to.
7846 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7850 cdst->mc_txn = csrc->mc_txn;
7851 cdst->mc_dbi = csrc->mc_dbi;
7852 cdst->mc_db = csrc->mc_db;
7853 cdst->mc_dbx = csrc->mc_dbx;
7854 cdst->mc_snum = csrc->mc_snum;
7855 cdst->mc_top = csrc->mc_top;
7856 cdst->mc_flags = csrc->mc_flags;
7858 for (i=0; i<csrc->mc_snum; i++) {
7859 cdst->mc_pg[i] = csrc->mc_pg[i];
7860 cdst->mc_ki[i] = csrc->mc_ki[i];
7864 /** Rebalance the tree after a delete operation.
7865 * @param[in] mc Cursor pointing to the page where rebalancing
7867 * @return 0 on success, non-zero on failure.
7870 mdb_rebalance(MDB_cursor *mc)
7874 unsigned int ptop, minkeys;
7878 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7879 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7880 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7881 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7882 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7884 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7885 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7886 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7887 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7891 if (mc->mc_snum < 2) {
7892 MDB_page *mp = mc->mc_pg[0];
7894 DPUTS("Can't rebalance a subpage, ignoring");
7897 if (NUMKEYS(mp) == 0) {
7898 DPUTS("tree is completely empty");
7899 mc->mc_db->md_root = P_INVALID;
7900 mc->mc_db->md_depth = 0;
7901 mc->mc_db->md_leaf_pages = 0;
7902 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7905 /* Adjust cursors pointing to mp */
7908 mc->mc_flags &= ~C_INITIALIZED;
7910 MDB_cursor *m2, *m3;
7911 MDB_dbi dbi = mc->mc_dbi;
7913 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7914 if (mc->mc_flags & C_SUB)
7915 m3 = &m2->mc_xcursor->mx_cursor;
7918 if (m3->mc_snum < mc->mc_snum) continue;
7919 if (m3->mc_pg[0] == mp) {
7922 m3->mc_flags &= ~C_INITIALIZED;
7926 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7928 DPUTS("collapsing root page!");
7929 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7932 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7933 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7936 mc->mc_db->md_depth--;
7937 mc->mc_db->md_branch_pages--;
7938 mc->mc_ki[0] = mc->mc_ki[1];
7939 for (i = 1; i<mc->mc_db->md_depth; i++) {
7940 mc->mc_pg[i] = mc->mc_pg[i+1];
7941 mc->mc_ki[i] = mc->mc_ki[i+1];
7944 /* Adjust other cursors pointing to mp */
7945 MDB_cursor *m2, *m3;
7946 MDB_dbi dbi = mc->mc_dbi;
7948 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7949 if (mc->mc_flags & C_SUB)
7950 m3 = &m2->mc_xcursor->mx_cursor;
7953 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7954 if (m3->mc_pg[0] == mp) {
7955 for (i=0; i<m3->mc_snum; i++) {
7956 m3->mc_pg[i] = m3->mc_pg[i+1];
7957 m3->mc_ki[i] = m3->mc_ki[i+1];
7965 DPUTS("root page doesn't need rebalancing");
7969 /* The parent (branch page) must have at least 2 pointers,
7970 * otherwise the tree is invalid.
7972 ptop = mc->mc_top-1;
7973 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7975 /* Leaf page fill factor is below the threshold.
7976 * Try to move keys from left or right neighbor, or
7977 * merge with a neighbor page.
7982 mdb_cursor_copy(mc, &mn);
7983 mn.mc_xcursor = NULL;
7985 oldki = mc->mc_ki[mc->mc_top];
7986 if (mc->mc_ki[ptop] == 0) {
7987 /* We're the leftmost leaf in our parent.
7989 DPUTS("reading right neighbor");
7991 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7992 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7995 mn.mc_ki[mn.mc_top] = 0;
7996 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7998 /* There is at least one neighbor to the left.
8000 DPUTS("reading left neighbor");
8002 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8003 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8006 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8007 mc->mc_ki[mc->mc_top] = 0;
8010 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8011 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8012 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8014 /* If the neighbor page is above threshold and has enough keys,
8015 * move one key from it. Otherwise we should try to merge them.
8016 * (A branch page must never have less than 2 keys.)
8018 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
8019 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8020 rc = mdb_node_move(&mn, mc);
8021 if (mc->mc_ki[ptop]) {
8025 if (mc->mc_ki[ptop] == 0) {
8026 rc = mdb_page_merge(&mn, mc);
8029 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8030 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8031 /* We want mdb_rebalance to find mn when doing fixups */
8032 if (mc->mc_flags & C_SUB) {
8033 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8034 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8035 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8037 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8038 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8040 rc = mdb_page_merge(mc, &mn);
8041 if (mc->mc_flags & C_SUB)
8042 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8044 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8045 mdb_cursor_copy(&mn, mc);
8047 mc->mc_flags &= ~C_EOF;
8049 mc->mc_ki[mc->mc_top] = oldki;
8053 /** Complete a delete operation started by #mdb_cursor_del(). */
8055 mdb_cursor_del0(MDB_cursor *mc)
8062 ki = mc->mc_ki[mc->mc_top];
8063 mdb_node_del(mc, mc->mc_db->md_pad);
8064 mc->mc_db->md_entries--;
8065 rc = mdb_rebalance(mc);
8067 if (rc == MDB_SUCCESS) {
8068 MDB_cursor *m2, *m3;
8069 MDB_dbi dbi = mc->mc_dbi;
8071 /* DB is totally empty now, just bail out.
8072 * Other cursors adjustments were already done
8073 * by mdb_rebalance and aren't needed here.
8078 mp = mc->mc_pg[mc->mc_top];
8079 nkeys = NUMKEYS(mp);
8081 /* if mc points past last node in page, find next sibling */
8082 if (mc->mc_ki[mc->mc_top] >= nkeys) {
8083 rc = mdb_cursor_sibling(mc, 1);
8084 if (rc == MDB_NOTFOUND) {
8085 mc->mc_flags |= C_EOF;
8090 /* Adjust other cursors pointing to mp */
8091 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8092 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8093 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8095 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8097 if (m3->mc_pg[mc->mc_top] == mp) {
8098 if (m3->mc_ki[mc->mc_top] >= ki) {
8099 m3->mc_flags |= C_DEL;
8100 if (m3->mc_ki[mc->mc_top] > ki)
8101 m3->mc_ki[mc->mc_top]--;
8102 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8103 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8105 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8106 rc = mdb_cursor_sibling(m3, 1);
8107 if (rc == MDB_NOTFOUND) {
8108 m3->mc_flags |= C_EOF;
8114 mc->mc_flags |= C_DEL;
8118 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8123 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8124 MDB_val *key, MDB_val *data)
8126 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8129 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
8130 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8132 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8133 /* must ignore any data */
8137 return mdb_del0(txn, dbi, key, data, 0);
8141 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8142 MDB_val *key, MDB_val *data, unsigned flags)
8147 MDB_val rdata, *xdata;
8151 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8153 mdb_cursor_init(&mc, txn, dbi, &mx);
8162 flags |= MDB_NODUPDATA;
8164 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8166 /* let mdb_page_split know about this cursor if needed:
8167 * delete will trigger a rebalance; if it needs to move
8168 * a node from one page to another, it will have to
8169 * update the parent's separator key(s). If the new sepkey
8170 * is larger than the current one, the parent page may
8171 * run out of space, triggering a split. We need this
8172 * cursor to be consistent until the end of the rebalance.
8174 mc.mc_flags |= C_UNTRACK;
8175 mc.mc_next = txn->mt_cursors[dbi];
8176 txn->mt_cursors[dbi] = &mc;
8177 rc = mdb_cursor_del(&mc, flags);
8178 txn->mt_cursors[dbi] = mc.mc_next;
8183 /** Split a page and insert a new node.
8184 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8185 * The cursor will be updated to point to the actual page and index where
8186 * the node got inserted after the split.
8187 * @param[in] newkey The key for the newly inserted node.
8188 * @param[in] newdata The data for the newly inserted node.
8189 * @param[in] newpgno The page number, if the new node is a branch node.
8190 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8191 * @return 0 on success, non-zero on failure.
8194 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8195 unsigned int nflags)
8198 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8201 int i, j, split_indx, nkeys, pmax;
8202 MDB_env *env = mc->mc_txn->mt_env;
8204 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8205 MDB_page *copy = NULL;
8206 MDB_page *mp, *rp, *pp;
8211 mp = mc->mc_pg[mc->mc_top];
8212 newindx = mc->mc_ki[mc->mc_top];
8213 nkeys = NUMKEYS(mp);
8215 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8216 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8217 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8219 /* Create a right sibling. */
8220 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8222 rp->mp_pad = mp->mp_pad;
8223 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8225 if (mc->mc_snum < 2) {
8226 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8228 /* shift current top to make room for new parent */
8229 mc->mc_pg[1] = mc->mc_pg[0];
8230 mc->mc_ki[1] = mc->mc_ki[0];
8233 mc->mc_db->md_root = pp->mp_pgno;
8234 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8235 mc->mc_db->md_depth++;
8238 /* Add left (implicit) pointer. */
8239 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8240 /* undo the pre-push */
8241 mc->mc_pg[0] = mc->mc_pg[1];
8242 mc->mc_ki[0] = mc->mc_ki[1];
8243 mc->mc_db->md_root = mp->mp_pgno;
8244 mc->mc_db->md_depth--;
8251 ptop = mc->mc_top-1;
8252 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8255 mc->mc_flags |= C_SPLITTING;
8256 mdb_cursor_copy(mc, &mn);
8257 mn.mc_pg[mn.mc_top] = rp;
8258 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8260 if (nflags & MDB_APPEND) {
8261 mn.mc_ki[mn.mc_top] = 0;
8263 split_indx = newindx;
8267 split_indx = (nkeys+1) / 2;
8272 unsigned int lsize, rsize, ksize;
8273 /* Move half of the keys to the right sibling */
8274 x = mc->mc_ki[mc->mc_top] - split_indx;
8275 ksize = mc->mc_db->md_pad;
8276 split = LEAF2KEY(mp, split_indx, ksize);
8277 rsize = (nkeys - split_indx) * ksize;
8278 lsize = (nkeys - split_indx) * sizeof(indx_t);
8279 mp->mp_lower -= lsize;
8280 rp->mp_lower += lsize;
8281 mp->mp_upper += rsize - lsize;
8282 rp->mp_upper -= rsize - lsize;
8283 sepkey.mv_size = ksize;
8284 if (newindx == split_indx) {
8285 sepkey.mv_data = newkey->mv_data;
8287 sepkey.mv_data = split;
8290 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8291 memcpy(rp->mp_ptrs, split, rsize);
8292 sepkey.mv_data = rp->mp_ptrs;
8293 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8294 memcpy(ins, newkey->mv_data, ksize);
8295 mp->mp_lower += sizeof(indx_t);
8296 mp->mp_upper -= ksize - sizeof(indx_t);
8299 memcpy(rp->mp_ptrs, split, x * ksize);
8300 ins = LEAF2KEY(rp, x, ksize);
8301 memcpy(ins, newkey->mv_data, ksize);
8302 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8303 rp->mp_lower += sizeof(indx_t);
8304 rp->mp_upper -= ksize - sizeof(indx_t);
8305 mc->mc_ki[mc->mc_top] = x;
8306 mc->mc_pg[mc->mc_top] = rp;
8309 int psize, nsize, k;
8310 /* Maximum free space in an empty page */
8311 pmax = env->me_psize - PAGEHDRSZ;
8313 nsize = mdb_leaf_size(env, newkey, newdata);
8315 nsize = mdb_branch_size(env, newkey);
8316 nsize = EVEN(nsize);
8318 /* grab a page to hold a temporary copy */
8319 copy = mdb_page_malloc(mc->mc_txn, 1);
8324 copy->mp_pgno = mp->mp_pgno;
8325 copy->mp_flags = mp->mp_flags;
8326 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8327 copy->mp_upper = env->me_psize - PAGEBASE;
8329 /* prepare to insert */
8330 for (i=0, j=0; i<nkeys; i++) {
8332 copy->mp_ptrs[j++] = 0;
8334 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8337 /* When items are relatively large the split point needs
8338 * to be checked, because being off-by-one will make the
8339 * difference between success or failure in mdb_node_add.
8341 * It's also relevant if a page happens to be laid out
8342 * such that one half of its nodes are all "small" and
8343 * the other half of its nodes are "large." If the new
8344 * item is also "large" and falls on the half with
8345 * "large" nodes, it also may not fit.
8347 * As a final tweak, if the new item goes on the last
8348 * spot on the page (and thus, onto the new page), bias
8349 * the split so the new page is emptier than the old page.
8350 * This yields better packing during sequential inserts.
8352 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8353 /* Find split point */
8355 if (newindx <= split_indx || newindx >= nkeys) {
8357 k = newindx >= nkeys ? nkeys : split_indx+2;
8362 for (; i!=k; i+=j) {
8367 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8368 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8370 if (F_ISSET(node->mn_flags, F_BIGDATA))
8371 psize += sizeof(pgno_t);
8373 psize += NODEDSZ(node);
8375 psize = EVEN(psize);
8377 if (psize > pmax || i == k-j) {
8378 split_indx = i + (j<0);
8383 if (split_indx == newindx) {
8384 sepkey.mv_size = newkey->mv_size;
8385 sepkey.mv_data = newkey->mv_data;
8387 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8388 sepkey.mv_size = node->mn_ksize;
8389 sepkey.mv_data = NODEKEY(node);
8394 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8396 /* Copy separator key to the parent.
8398 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8402 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8407 if (mn.mc_snum == mc->mc_snum) {
8408 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8409 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8410 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8411 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8416 /* Right page might now have changed parent.
8417 * Check if left page also changed parent.
8419 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8420 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8421 for (i=0; i<ptop; i++) {
8422 mc->mc_pg[i] = mn.mc_pg[i];
8423 mc->mc_ki[i] = mn.mc_ki[i];
8425 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8426 if (mn.mc_ki[ptop]) {
8427 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8429 /* find right page's left sibling */
8430 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8431 mdb_cursor_sibling(mc, 0);
8436 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8439 mc->mc_flags ^= C_SPLITTING;
8440 if (rc != MDB_SUCCESS) {
8443 if (nflags & MDB_APPEND) {
8444 mc->mc_pg[mc->mc_top] = rp;
8445 mc->mc_ki[mc->mc_top] = 0;
8446 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8449 for (i=0; i<mc->mc_top; i++)
8450 mc->mc_ki[i] = mn.mc_ki[i];
8451 } else if (!IS_LEAF2(mp)) {
8453 mc->mc_pg[mc->mc_top] = rp;
8458 rkey.mv_data = newkey->mv_data;
8459 rkey.mv_size = newkey->mv_size;
8465 /* Update index for the new key. */
8466 mc->mc_ki[mc->mc_top] = j;
8468 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8469 rkey.mv_data = NODEKEY(node);
8470 rkey.mv_size = node->mn_ksize;
8472 xdata.mv_data = NODEDATA(node);
8473 xdata.mv_size = NODEDSZ(node);
8476 pgno = NODEPGNO(node);
8477 flags = node->mn_flags;
8480 if (!IS_LEAF(mp) && j == 0) {
8481 /* First branch index doesn't need key data. */
8485 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8491 mc->mc_pg[mc->mc_top] = copy;
8496 } while (i != split_indx);
8498 nkeys = NUMKEYS(copy);
8499 for (i=0; i<nkeys; i++)
8500 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8501 mp->mp_lower = copy->mp_lower;
8502 mp->mp_upper = copy->mp_upper;
8503 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8504 env->me_psize - copy->mp_upper - PAGEBASE);
8506 /* reset back to original page */
8507 if (newindx < split_indx) {
8508 mc->mc_pg[mc->mc_top] = mp;
8509 if (nflags & MDB_RESERVE) {
8510 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8511 if (!(node->mn_flags & F_BIGDATA))
8512 newdata->mv_data = NODEDATA(node);
8515 mc->mc_pg[mc->mc_top] = rp;
8517 /* Make sure mc_ki is still valid.
8519 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8520 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8521 for (i=0; i<=ptop; i++) {
8522 mc->mc_pg[i] = mn.mc_pg[i];
8523 mc->mc_ki[i] = mn.mc_ki[i];
8530 /* Adjust other cursors pointing to mp */
8531 MDB_cursor *m2, *m3;
8532 MDB_dbi dbi = mc->mc_dbi;
8533 int fixup = NUMKEYS(mp);
8535 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8536 if (mc->mc_flags & C_SUB)
8537 m3 = &m2->mc_xcursor->mx_cursor;
8542 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8544 if (m3->mc_flags & C_SPLITTING)
8549 for (k=m3->mc_top; k>=0; k--) {
8550 m3->mc_ki[k+1] = m3->mc_ki[k];
8551 m3->mc_pg[k+1] = m3->mc_pg[k];
8553 if (m3->mc_ki[0] >= split_indx) {
8558 m3->mc_pg[0] = mc->mc_pg[0];
8562 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8563 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8564 m3->mc_ki[mc->mc_top]++;
8565 if (m3->mc_ki[mc->mc_top] >= fixup) {
8566 m3->mc_pg[mc->mc_top] = rp;
8567 m3->mc_ki[mc->mc_top] -= fixup;
8568 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8570 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8571 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8576 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8579 if (copy) /* tmp page */
8580 mdb_page_free(env, copy);
8582 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8587 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8588 MDB_val *key, MDB_val *data, unsigned int flags)
8593 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8596 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8599 mdb_cursor_init(&mc, txn, dbi, &mx);
8600 return mdb_cursor_put(&mc, key, data, flags);
8604 #define MDB_WBUF (1024*1024)
8607 /** State needed for a compacting copy. */
8608 typedef struct mdb_copy {
8609 pthread_mutex_t mc_mutex;
8610 pthread_cond_t mc_cond;
8617 pgno_t mc_next_pgno;
8620 volatile int mc_new;
8625 /** Dedicated writer thread for compacting copy. */
8626 static THREAD_RET ESECT
8627 mdb_env_copythr(void *arg)
8631 int toggle = 0, wsize, rc;
8634 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8637 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8640 pthread_mutex_lock(&my->mc_mutex);
8642 pthread_cond_signal(&my->mc_cond);
8645 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8646 if (my->mc_new < 0) {
8651 wsize = my->mc_wlen[toggle];
8652 ptr = my->mc_wbuf[toggle];
8655 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8659 } else if (len > 0) {
8673 /* If there's an overflow page tail, write it too */
8674 if (my->mc_olen[toggle]) {
8675 wsize = my->mc_olen[toggle];
8676 ptr = my->mc_over[toggle];
8677 my->mc_olen[toggle] = 0;
8680 my->mc_wlen[toggle] = 0;
8682 pthread_cond_signal(&my->mc_cond);
8684 pthread_cond_signal(&my->mc_cond);
8685 pthread_mutex_unlock(&my->mc_mutex);
8686 return (THREAD_RET)0;
8690 /** Tell the writer thread there's a buffer ready to write */
8692 mdb_env_cthr_toggle(mdb_copy *my, int st)
8694 int toggle = my->mc_toggle ^ 1;
8695 pthread_mutex_lock(&my->mc_mutex);
8696 if (my->mc_status) {
8697 pthread_mutex_unlock(&my->mc_mutex);
8698 return my->mc_status;
8700 while (my->mc_new == 1)
8701 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8703 my->mc_toggle = toggle;
8704 pthread_cond_signal(&my->mc_cond);
8705 pthread_mutex_unlock(&my->mc_mutex);
8709 /** Depth-first tree traversal for compacting copy. */
8711 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8714 MDB_txn *txn = my->mc_txn;
8716 MDB_page *mo, *mp, *leaf;
8721 /* Empty DB, nothing to do */
8722 if (*pg == P_INVALID)
8729 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8732 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8736 /* Make cursor pages writable */
8737 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8741 for (i=0; i<mc.mc_top; i++) {
8742 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8743 mc.mc_pg[i] = (MDB_page *)ptr;
8744 ptr += my->mc_env->me_psize;
8747 /* This is writable space for a leaf page. Usually not needed. */
8748 leaf = (MDB_page *)ptr;
8750 toggle = my->mc_toggle;
8751 while (mc.mc_snum > 0) {
8753 mp = mc.mc_pg[mc.mc_top];
8757 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8758 for (i=0; i<n; i++) {
8759 ni = NODEPTR(mp, i);
8760 if (ni->mn_flags & F_BIGDATA) {
8764 /* Need writable leaf */
8766 mc.mc_pg[mc.mc_top] = leaf;
8767 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8769 ni = NODEPTR(mp, i);
8772 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8773 rc = mdb_page_get(txn, pg, &omp, NULL);
8776 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8777 rc = mdb_env_cthr_toggle(my, 1);
8780 toggle = my->mc_toggle;
8782 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8783 memcpy(mo, omp, my->mc_env->me_psize);
8784 mo->mp_pgno = my->mc_next_pgno;
8785 my->mc_next_pgno += omp->mp_pages;
8786 my->mc_wlen[toggle] += my->mc_env->me_psize;
8787 if (omp->mp_pages > 1) {
8788 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8789 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8790 rc = mdb_env_cthr_toggle(my, 1);
8793 toggle = my->mc_toggle;
8795 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8796 } else if (ni->mn_flags & F_SUBDATA) {
8799 /* Need writable leaf */
8801 mc.mc_pg[mc.mc_top] = leaf;
8802 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8804 ni = NODEPTR(mp, i);
8807 memcpy(&db, NODEDATA(ni), sizeof(db));
8808 my->mc_toggle = toggle;
8809 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8812 toggle = my->mc_toggle;
8813 memcpy(NODEDATA(ni), &db, sizeof(db));
8818 mc.mc_ki[mc.mc_top]++;
8819 if (mc.mc_ki[mc.mc_top] < n) {
8822 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8824 rc = mdb_page_get(txn, pg, &mp, NULL);
8829 mc.mc_ki[mc.mc_top] = 0;
8830 if (IS_BRANCH(mp)) {
8831 /* Whenever we advance to a sibling branch page,
8832 * we must proceed all the way down to its first leaf.
8834 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8837 mc.mc_pg[mc.mc_top] = mp;
8841 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8842 rc = mdb_env_cthr_toggle(my, 1);
8845 toggle = my->mc_toggle;
8847 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8848 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8849 mo->mp_pgno = my->mc_next_pgno++;
8850 my->mc_wlen[toggle] += my->mc_env->me_psize;
8852 /* Update parent if there is one */
8853 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8854 SETPGNO(ni, mo->mp_pgno);
8855 mdb_cursor_pop(&mc);
8857 /* Otherwise we're done */
8867 /** Copy environment with compaction. */
8869 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8874 MDB_txn *txn = NULL;
8879 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8880 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8881 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8882 if (my.mc_wbuf[0] == NULL)
8885 pthread_mutex_init(&my.mc_mutex, NULL);
8886 pthread_cond_init(&my.mc_cond, NULL);
8887 #ifdef HAVE_MEMALIGN
8888 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8889 if (my.mc_wbuf[0] == NULL)
8892 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8897 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8898 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8903 my.mc_next_pgno = 2;
8909 THREAD_CREATE(thr, mdb_env_copythr, &my);
8911 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8915 mp = (MDB_page *)my.mc_wbuf[0];
8916 memset(mp, 0, 2*env->me_psize);
8918 mp->mp_flags = P_META;
8919 mm = (MDB_meta *)METADATA(mp);
8920 mdb_env_init_meta0(env, mm);
8921 mm->mm_address = env->me_metas[0]->mm_address;
8923 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8925 mp->mp_flags = P_META;
8926 *(MDB_meta *)METADATA(mp) = *mm;
8927 mm = (MDB_meta *)METADATA(mp);
8929 /* Count the number of free pages, subtract from lastpg to find
8930 * number of active pages
8933 MDB_ID freecount = 0;
8936 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8937 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8938 freecount += *(MDB_ID *)data.mv_data;
8939 freecount += txn->mt_dbs[0].md_branch_pages +
8940 txn->mt_dbs[0].md_leaf_pages +
8941 txn->mt_dbs[0].md_overflow_pages;
8943 /* Set metapage 1 */
8944 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8945 mm->mm_dbs[1] = txn->mt_dbs[1];
8946 if (mm->mm_last_pg > 1) {
8947 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8950 mm->mm_dbs[1].md_root = P_INVALID;
8953 my.mc_wlen[0] = env->me_psize * 2;
8955 pthread_mutex_lock(&my.mc_mutex);
8957 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8958 pthread_mutex_unlock(&my.mc_mutex);
8959 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8960 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8961 rc = mdb_env_cthr_toggle(&my, 1);
8962 mdb_env_cthr_toggle(&my, -1);
8963 pthread_mutex_lock(&my.mc_mutex);
8965 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8966 pthread_mutex_unlock(&my.mc_mutex);
8971 CloseHandle(my.mc_cond);
8972 CloseHandle(my.mc_mutex);
8973 _aligned_free(my.mc_wbuf[0]);
8975 pthread_cond_destroy(&my.mc_cond);
8976 pthread_mutex_destroy(&my.mc_mutex);
8977 free(my.mc_wbuf[0]);
8982 /** Copy environment as-is. */
8984 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8986 MDB_txn *txn = NULL;
8987 mdb_mutexref_t wmutex = NULL;
8993 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8997 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9000 /* Do the lock/unlock of the reader mutex before starting the
9001 * write txn. Otherwise other read txns could block writers.
9003 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9008 /* We must start the actual read txn after blocking writers */
9009 mdb_txn_end(txn, MDB_END_RESET_TMP);
9011 /* Temporarily block writers until we snapshot the meta pages */
9012 wmutex = env->me_wmutex;
9013 if (LOCK_MUTEX(rc, env, wmutex))
9016 rc = mdb_txn_renew0(txn);
9018 UNLOCK_MUTEX(wmutex);
9023 wsize = env->me_psize * 2;
9027 DO_WRITE(rc, fd, ptr, w2, len);
9031 } else if (len > 0) {
9037 /* Non-blocking or async handles are not supported */
9043 UNLOCK_MUTEX(wmutex);
9048 w2 = txn->mt_next_pgno * env->me_psize;
9051 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9058 if (wsize > MAX_WRITE)
9062 DO_WRITE(rc, fd, ptr, w2, len);
9066 } else if (len > 0) {
9083 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9085 if (flags & MDB_CP_COMPACT)
9086 return mdb_env_copyfd1(env, fd);
9088 return mdb_env_copyfd0(env, fd);
9092 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9094 return mdb_env_copyfd2(env, fd, 0);
9098 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9102 HANDLE newfd = INVALID_HANDLE_VALUE;
9104 if (env->me_flags & MDB_NOSUBDIR) {
9105 lpath = (char *)path;
9108 len += sizeof(DATANAME);
9109 lpath = malloc(len);
9112 sprintf(lpath, "%s" DATANAME, path);
9115 /* The destination path must exist, but the destination file must not.
9116 * We don't want the OS to cache the writes, since the source data is
9117 * already in the OS cache.
9120 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9121 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9123 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9125 if (newfd == INVALID_HANDLE_VALUE) {
9130 if (env->me_psize >= env->me_os_psize) {
9132 /* Set O_DIRECT if the file system supports it */
9133 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9134 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9136 #ifdef F_NOCACHE /* __APPLE__ */
9137 rc = fcntl(newfd, F_NOCACHE, 1);
9145 rc = mdb_env_copyfd2(env, newfd, flags);
9148 if (!(env->me_flags & MDB_NOSUBDIR))
9150 if (newfd != INVALID_HANDLE_VALUE)
9151 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9158 mdb_env_copy(MDB_env *env, const char *path)
9160 return mdb_env_copy2(env, path, 0);
9164 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9166 if (flag & ~CHANGEABLE)
9169 env->me_flags |= flag;
9171 env->me_flags &= ~flag;
9176 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9181 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9186 mdb_env_set_userctx(MDB_env *env, void *ctx)
9190 env->me_userctx = ctx;
9195 mdb_env_get_userctx(MDB_env *env)
9197 return env ? env->me_userctx : NULL;
9201 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9206 env->me_assert_func = func;
9212 mdb_env_get_path(MDB_env *env, const char **arg)
9217 *arg = env->me_path;
9222 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9231 /** Common code for #mdb_stat() and #mdb_env_stat().
9232 * @param[in] env the environment to operate in.
9233 * @param[in] db the #MDB_db record containing the stats to return.
9234 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9235 * @return 0, this function always succeeds.
9238 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9240 arg->ms_psize = env->me_psize;
9241 arg->ms_depth = db->md_depth;
9242 arg->ms_branch_pages = db->md_branch_pages;
9243 arg->ms_leaf_pages = db->md_leaf_pages;
9244 arg->ms_overflow_pages = db->md_overflow_pages;
9245 arg->ms_entries = db->md_entries;
9251 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9255 if (env == NULL || arg == NULL)
9258 meta = mdb_env_pick_meta(env);
9260 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9264 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9268 if (env == NULL || arg == NULL)
9271 meta = mdb_env_pick_meta(env);
9272 arg->me_mapaddr = meta->mm_address;
9273 arg->me_last_pgno = meta->mm_last_pg;
9274 arg->me_last_txnid = meta->mm_txnid;
9276 arg->me_mapsize = env->me_mapsize;
9277 arg->me_maxreaders = env->me_maxreaders;
9278 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9282 /** Set the default comparison functions for a database.
9283 * Called immediately after a database is opened to set the defaults.
9284 * The user can then override them with #mdb_set_compare() or
9285 * #mdb_set_dupsort().
9286 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9287 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9290 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9292 uint16_t f = txn->mt_dbs[dbi].md_flags;
9294 txn->mt_dbxs[dbi].md_cmp =
9295 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9296 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9298 txn->mt_dbxs[dbi].md_dcmp =
9299 !(f & MDB_DUPSORT) ? 0 :
9300 ((f & MDB_INTEGERDUP)
9301 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9302 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9305 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9311 int rc, dbflag, exact;
9312 unsigned int unused = 0, seq;
9315 if (flags & ~VALID_FLAGS)
9317 if (txn->mt_flags & MDB_TXN_ERROR)
9323 if (flags & PERSISTENT_FLAGS) {
9324 uint16_t f2 = flags & PERSISTENT_FLAGS;
9325 /* make sure flag changes get committed */
9326 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9327 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9328 txn->mt_flags |= MDB_TXN_DIRTY;
9331 mdb_default_cmp(txn, MAIN_DBI);
9335 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9336 mdb_default_cmp(txn, MAIN_DBI);
9339 /* Is the DB already open? */
9341 for (i=2; i<txn->mt_numdbs; i++) {
9342 if (!txn->mt_dbxs[i].md_name.mv_size) {
9343 /* Remember this free slot */
9344 if (!unused) unused = i;
9347 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9348 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9354 /* If no free slot and max hit, fail */
9355 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9356 return MDB_DBS_FULL;
9358 /* Cannot mix named databases with some mainDB flags */
9359 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9360 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9362 /* Find the DB info */
9363 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9366 key.mv_data = (void *)name;
9367 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9368 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9369 if (rc == MDB_SUCCESS) {
9370 /* make sure this is actually a DB */
9371 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9372 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9373 return MDB_INCOMPATIBLE;
9374 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9375 /* Create if requested */
9376 data.mv_size = sizeof(MDB_db);
9377 data.mv_data = &dummy;
9378 memset(&dummy, 0, sizeof(dummy));
9379 dummy.md_root = P_INVALID;
9380 dummy.md_flags = flags & PERSISTENT_FLAGS;
9381 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9385 /* OK, got info, add to table */
9386 if (rc == MDB_SUCCESS) {
9387 unsigned int slot = unused ? unused : txn->mt_numdbs;
9388 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9389 txn->mt_dbxs[slot].md_name.mv_size = len;
9390 txn->mt_dbxs[slot].md_rel = NULL;
9391 txn->mt_dbflags[slot] = dbflag;
9392 /* txn-> and env-> are the same in read txns, use
9393 * tmp variable to avoid undefined assignment
9395 seq = ++txn->mt_env->me_dbiseqs[slot];
9396 txn->mt_dbiseqs[slot] = seq;
9398 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9400 mdb_default_cmp(txn, slot);
9410 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9412 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9415 if (txn->mt_flags & MDB_TXN_ERROR)
9418 if (txn->mt_dbflags[dbi] & DB_STALE) {
9421 /* Stale, must read the DB's root. cursor_init does it for us. */
9422 mdb_cursor_init(&mc, txn, dbi, &mx);
9424 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9427 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9430 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9432 ptr = env->me_dbxs[dbi].md_name.mv_data;
9433 /* If there was no name, this was already closed */
9435 env->me_dbxs[dbi].md_name.mv_data = NULL;
9436 env->me_dbxs[dbi].md_name.mv_size = 0;
9437 env->me_dbflags[dbi] = 0;
9438 env->me_dbiseqs[dbi]++;
9443 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9445 /* We could return the flags for the FREE_DBI too but what's the point? */
9446 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9448 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9452 /** Add all the DB's pages to the free list.
9453 * @param[in] mc Cursor on the DB to free.
9454 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9455 * @return 0 on success, non-zero on failure.
9458 mdb_drop0(MDB_cursor *mc, int subs)
9462 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9463 if (rc == MDB_SUCCESS) {
9464 MDB_txn *txn = mc->mc_txn;
9469 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9470 * This also avoids any P_LEAF2 pages, which have no nodes.
9472 if (mc->mc_flags & C_SUB)
9475 mdb_cursor_copy(mc, &mx);
9476 while (mc->mc_snum > 0) {
9477 MDB_page *mp = mc->mc_pg[mc->mc_top];
9478 unsigned n = NUMKEYS(mp);
9480 for (i=0; i<n; i++) {
9481 ni = NODEPTR(mp, i);
9482 if (ni->mn_flags & F_BIGDATA) {
9485 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9486 rc = mdb_page_get(txn, pg, &omp, NULL);
9489 mdb_cassert(mc, IS_OVERFLOW(omp));
9490 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9494 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9495 mdb_xcursor_init1(mc, ni);
9496 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9502 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9504 for (i=0; i<n; i++) {
9506 ni = NODEPTR(mp, i);
9509 mdb_midl_xappend(txn->mt_free_pgs, pg);
9514 mc->mc_ki[mc->mc_top] = i;
9515 rc = mdb_cursor_sibling(mc, 1);
9517 if (rc != MDB_NOTFOUND)
9519 /* no more siblings, go back to beginning
9520 * of previous level.
9524 for (i=1; i<mc->mc_snum; i++) {
9526 mc->mc_pg[i] = mx.mc_pg[i];
9531 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9534 txn->mt_flags |= MDB_TXN_ERROR;
9535 } else if (rc == MDB_NOTFOUND) {
9541 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9543 MDB_cursor *mc, *m2;
9546 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9549 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9552 if (TXN_DBI_CHANGED(txn, dbi))
9555 rc = mdb_cursor_open(txn, dbi, &mc);
9559 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9560 /* Invalidate the dropped DB's cursors */
9561 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9562 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9566 /* Can't delete the main DB */
9567 if (del && dbi > MAIN_DBI) {
9568 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9570 txn->mt_dbflags[dbi] = DB_STALE;
9571 mdb_dbi_close(txn->mt_env, dbi);
9573 txn->mt_flags |= MDB_TXN_ERROR;
9576 /* reset the DB record, mark it dirty */
9577 txn->mt_dbflags[dbi] |= DB_DIRTY;
9578 txn->mt_dbs[dbi].md_depth = 0;
9579 txn->mt_dbs[dbi].md_branch_pages = 0;
9580 txn->mt_dbs[dbi].md_leaf_pages = 0;
9581 txn->mt_dbs[dbi].md_overflow_pages = 0;
9582 txn->mt_dbs[dbi].md_entries = 0;
9583 txn->mt_dbs[dbi].md_root = P_INVALID;
9585 txn->mt_flags |= MDB_TXN_DIRTY;
9588 mdb_cursor_close(mc);
9592 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9594 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9597 txn->mt_dbxs[dbi].md_cmp = cmp;
9601 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9603 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9606 txn->mt_dbxs[dbi].md_dcmp = cmp;
9610 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9612 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9615 txn->mt_dbxs[dbi].md_rel = rel;
9619 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9621 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9624 txn->mt_dbxs[dbi].md_relctx = ctx;
9629 mdb_env_get_maxkeysize(MDB_env *env)
9631 return ENV_MAXKEY(env);
9635 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9637 unsigned int i, rdrs;
9640 int rc = 0, first = 1;
9644 if (!env->me_txns) {
9645 return func("(no reader locks)\n", ctx);
9647 rdrs = env->me_txns->mti_numreaders;
9648 mr = env->me_txns->mti_readers;
9649 for (i=0; i<rdrs; i++) {
9651 txnid_t txnid = mr[i].mr_txnid;
9652 sprintf(buf, txnid == (txnid_t)-1 ?
9653 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9654 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9657 rc = func(" pid thread txnid\n", ctx);
9661 rc = func(buf, ctx);
9667 rc = func("(no active readers)\n", ctx);
9672 /** Insert pid into list if not already present.
9673 * return -1 if already present.
9676 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9678 /* binary search of pid in list */
9680 unsigned cursor = 1;
9682 unsigned n = ids[0];
9685 unsigned pivot = n >> 1;
9686 cursor = base + pivot + 1;
9687 val = pid - ids[cursor];
9692 } else if ( val > 0 ) {
9697 /* found, so it's a duplicate */
9706 for (n = ids[0]; n > cursor; n--)
9713 mdb_reader_check(MDB_env *env, int *dead)
9719 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9722 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9724 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9726 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9727 unsigned int i, j, rdrs;
9729 MDB_PID_T *pids, pid;
9730 int rc = MDB_SUCCESS, count = 0;
9732 rdrs = env->me_txns->mti_numreaders;
9733 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9737 mr = env->me_txns->mti_readers;
9738 for (i=0; i<rdrs; i++) {
9740 if (pid && pid != env->me_pid) {
9741 if (mdb_pid_insert(pids, pid) == 0) {
9742 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9743 /* Stale reader found */
9746 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9747 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9749 rdrs = 0; /* the above checked all readers */
9751 /* Recheck, a new process may have reused pid */
9752 if (mdb_reader_pid(env, Pidcheck, pid))
9757 if (mr[j].mr_pid == pid) {
9758 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9759 (unsigned) pid, mr[j].mr_txnid));
9764 UNLOCK_MUTEX(rmutex);
9775 #ifdef MDB_ROBUST_SUPPORTED
9776 /** Handle #LOCK_MUTEX0() failure.
9777 * Try to repair the lock file if the mutex owner died.
9778 * @param[in] env the environment handle
9779 * @param[in] mutex LOCK_MUTEX0() mutex
9780 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9781 * @return 0 on success with the mutex locked, or an error code on failure.
9784 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9789 if (rc == MDB_OWNERDEAD) {
9790 /* We own the mutex. Clean up after dead previous owner. */
9792 rlocked = (mutex == env->me_rmutex);
9794 /* Keep mti_txnid updated, otherwise next writer can
9795 * overwrite data which latest meta page refers to.
9797 meta = mdb_env_pick_meta(env);
9798 env->me_txns->mti_txnid = meta->mm_txnid;
9799 /* env is hosed if the dead thread was ours */
9801 env->me_flags |= MDB_FATAL_ERROR;
9806 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9807 (rc ? "this process' env is hosed" : "recovering")));
9808 rc2 = mdb_reader_check0(env, rlocked, NULL);
9810 rc2 = mdb_mutex_consistent(mutex);
9811 if (rc || (rc = rc2)) {
9812 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9813 UNLOCK_MUTEX(mutex);
9819 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9824 #endif /* MDB_ROBUST_SUPPORTED */
projects / openldap / blob