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 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1306 static int mdb_page_search_root(MDB_cursor *mc,
1307 MDB_val *key, int modify);
1308 #define MDB_PS_MODIFY 1
1309 #define MDB_PS_ROOTONLY 2
1310 #define MDB_PS_FIRST 4
1311 #define MDB_PS_LAST 8
1312 static int mdb_page_search(MDB_cursor *mc,
1313 MDB_val *key, int flags);
1314 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1316 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1317 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1318 pgno_t newpgno, unsigned int nflags);
1320 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1321 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1322 static int mdb_env_write_meta(MDB_txn *txn);
1323 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1324 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1326 static void mdb_env_close0(MDB_env *env, int excl);
1328 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1329 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1330 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1331 static void mdb_node_del(MDB_cursor *mc, int ksize);
1332 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1333 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1334 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1335 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1336 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1338 static int mdb_rebalance(MDB_cursor *mc);
1339 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1341 static void mdb_cursor_pop(MDB_cursor *mc);
1342 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1344 static int mdb_cursor_del0(MDB_cursor *mc);
1345 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1346 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1347 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1348 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1349 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1351 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1352 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1354 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1355 static void mdb_xcursor_init0(MDB_cursor *mc);
1356 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1358 static int mdb_drop0(MDB_cursor *mc, int subs);
1359 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1360 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1363 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1366 /** Compare two items pointing at size_t's of unknown alignment. */
1367 #ifdef MISALIGNED_OK
1368 # define mdb_cmp_clong mdb_cmp_long
1370 # define mdb_cmp_clong mdb_cmp_cint
1374 static SECURITY_DESCRIPTOR mdb_null_sd;
1375 static SECURITY_ATTRIBUTES mdb_all_sa;
1376 static int mdb_sec_inited;
1379 /** Return the library version info. */
1381 mdb_version(int *major, int *minor, int *patch)
1383 if (major) *major = MDB_VERSION_MAJOR;
1384 if (minor) *minor = MDB_VERSION_MINOR;
1385 if (patch) *patch = MDB_VERSION_PATCH;
1386 return MDB_VERSION_STRING;
1389 /** Table of descriptions for LMDB @ref errors */
1390 static char *const mdb_errstr[] = {
1391 "MDB_KEYEXIST: Key/data pair already exists",
1392 "MDB_NOTFOUND: No matching key/data pair found",
1393 "MDB_PAGE_NOTFOUND: Requested page not found",
1394 "MDB_CORRUPTED: Located page was wrong type",
1395 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1396 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1397 "MDB_INVALID: File is not an LMDB file",
1398 "MDB_MAP_FULL: Environment mapsize limit reached",
1399 "MDB_DBS_FULL: Environment maxdbs limit reached",
1400 "MDB_READERS_FULL: Environment maxreaders limit reached",
1401 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1402 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1403 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1404 "MDB_PAGE_FULL: Internal error - page has no more space",
1405 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1406 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1407 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1408 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1409 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1410 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1414 mdb_strerror(int err)
1417 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1418 * This works as long as no function between the call to mdb_strerror
1419 * and the actual use of the message uses more than 4K of stack.
1422 char buf[1024], *ptr = buf;
1426 return ("Successful return: 0");
1428 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1429 i = err - MDB_KEYEXIST;
1430 return mdb_errstr[i];
1434 /* These are the C-runtime error codes we use. The comment indicates
1435 * their numeric value, and the Win32 error they would correspond to
1436 * if the error actually came from a Win32 API. A major mess, we should
1437 * have used LMDB-specific error codes for everything.
1440 case ENOENT: /* 2, FILE_NOT_FOUND */
1441 case EIO: /* 5, ACCESS_DENIED */
1442 case ENOMEM: /* 12, INVALID_ACCESS */
1443 case EACCES: /* 13, INVALID_DATA */
1444 case EBUSY: /* 16, CURRENT_DIRECTORY */
1445 case EINVAL: /* 22, BAD_COMMAND */
1446 case ENOSPC: /* 28, OUT_OF_PAPER */
1447 return strerror(err);
1452 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1453 FORMAT_MESSAGE_IGNORE_INSERTS,
1454 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1457 return strerror(err);
1461 /** assert(3) variant in cursor context */
1462 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1463 /** assert(3) variant in transaction context */
1464 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1465 /** assert(3) variant in environment context */
1466 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1469 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1470 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1473 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1474 const char *func, const char *file, int line)
1477 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1478 file, line, expr_txt, func);
1479 if (env->me_assert_func)
1480 env->me_assert_func(env, buf);
1481 fprintf(stderr, "%s\n", buf);
1485 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1489 /** Return the page number of \b mp which may be sub-page, for debug output */
1491 mdb_dbg_pgno(MDB_page *mp)
1494 COPY_PGNO(ret, mp->mp_pgno);
1498 /** Display a key in hexadecimal and return the address of the result.
1499 * @param[in] key the key to display
1500 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1501 * @return The key in hexadecimal form.
1504 mdb_dkey(MDB_val *key, char *buf)
1507 unsigned char *c = key->mv_data;
1513 if (key->mv_size > DKBUF_MAXKEYSIZE)
1514 return "MDB_MAXKEYSIZE";
1515 /* may want to make this a dynamic check: if the key is mostly
1516 * printable characters, print it as-is instead of converting to hex.
1520 for (i=0; i<key->mv_size; i++)
1521 ptr += sprintf(ptr, "%02x", *c++);
1523 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1529 mdb_leafnode_type(MDB_node *n)
1531 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1532 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1533 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1536 /** Display all the keys in the page. */
1538 mdb_page_list(MDB_page *mp)
1540 pgno_t pgno = mdb_dbg_pgno(mp);
1541 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1543 unsigned int i, nkeys, nsize, total = 0;
1547 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1548 case P_BRANCH: type = "Branch page"; break;
1549 case P_LEAF: type = "Leaf page"; break;
1550 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1551 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1552 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1554 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1555 pgno, mp->mp_pages, state);
1558 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1559 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1562 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1566 nkeys = NUMKEYS(mp);
1567 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1569 for (i=0; i<nkeys; i++) {
1570 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1571 key.mv_size = nsize = mp->mp_pad;
1572 key.mv_data = LEAF2KEY(mp, i, nsize);
1574 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1577 node = NODEPTR(mp, i);
1578 key.mv_size = node->mn_ksize;
1579 key.mv_data = node->mn_data;
1580 nsize = NODESIZE + key.mv_size;
1581 if (IS_BRANCH(mp)) {
1582 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1586 if (F_ISSET(node->mn_flags, F_BIGDATA))
1587 nsize += sizeof(pgno_t);
1589 nsize += NODEDSZ(node);
1591 nsize += sizeof(indx_t);
1592 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1593 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1595 total = EVEN(total);
1597 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1598 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1602 mdb_cursor_chk(MDB_cursor *mc)
1608 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1609 for (i=0; i<mc->mc_top; i++) {
1611 node = NODEPTR(mp, mc->mc_ki[i]);
1612 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1615 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1621 /** Count all the pages in each DB and in the freelist
1622 * and make sure it matches the actual number of pages
1624 * All named DBs must be open for a correct count.
1626 static void mdb_audit(MDB_txn *txn)
1630 MDB_ID freecount, count;
1635 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1636 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1637 freecount += *(MDB_ID *)data.mv_data;
1638 mdb_tassert(txn, rc == MDB_NOTFOUND);
1641 for (i = 0; i<txn->mt_numdbs; i++) {
1643 if (!(txn->mt_dbflags[i] & DB_VALID))
1645 mdb_cursor_init(&mc, txn, i, &mx);
1646 if (txn->mt_dbs[i].md_root == P_INVALID)
1648 count += txn->mt_dbs[i].md_branch_pages +
1649 txn->mt_dbs[i].md_leaf_pages +
1650 txn->mt_dbs[i].md_overflow_pages;
1651 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1652 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1653 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1656 mp = mc.mc_pg[mc.mc_top];
1657 for (j=0; j<NUMKEYS(mp); j++) {
1658 MDB_node *leaf = NODEPTR(mp, j);
1659 if (leaf->mn_flags & F_SUBDATA) {
1661 memcpy(&db, NODEDATA(leaf), sizeof(db));
1662 count += db.md_branch_pages + db.md_leaf_pages +
1663 db.md_overflow_pages;
1667 mdb_tassert(txn, rc == MDB_NOTFOUND);
1670 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1671 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1672 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1678 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1680 return txn->mt_dbxs[dbi].md_cmp(a, b);
1684 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1686 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1687 #if UINT_MAX < SIZE_MAX
1688 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1689 dcmp = mdb_cmp_clong;
1694 /** Allocate memory for a page.
1695 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1698 mdb_page_malloc(MDB_txn *txn, unsigned num)
1700 MDB_env *env = txn->mt_env;
1701 MDB_page *ret = env->me_dpages;
1702 size_t psize = env->me_psize, sz = psize, off;
1703 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1704 * For a single page alloc, we init everything after the page header.
1705 * For multi-page, we init the final page; if the caller needed that
1706 * many pages they will be filling in at least up to the last page.
1710 VGMEMP_ALLOC(env, ret, sz);
1711 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1712 env->me_dpages = ret->mp_next;
1715 psize -= off = PAGEHDRSZ;
1720 if ((ret = malloc(sz)) != NULL) {
1721 VGMEMP_ALLOC(env, ret, sz);
1722 if (!(env->me_flags & MDB_NOMEMINIT)) {
1723 memset((char *)ret + off, 0, psize);
1727 txn->mt_flags |= MDB_TXN_ERROR;
1731 /** Free a single page.
1732 * Saves single pages to a list, for future reuse.
1733 * (This is not used for multi-page overflow pages.)
1736 mdb_page_free(MDB_env *env, MDB_page *mp)
1738 mp->mp_next = env->me_dpages;
1739 VGMEMP_FREE(env, mp);
1740 env->me_dpages = mp;
1743 /** Free a dirty page */
1745 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1747 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1748 mdb_page_free(env, dp);
1750 /* large pages just get freed directly */
1751 VGMEMP_FREE(env, dp);
1756 /** Return all dirty pages to dpage list */
1758 mdb_dlist_free(MDB_txn *txn)
1760 MDB_env *env = txn->mt_env;
1761 MDB_ID2L dl = txn->mt_u.dirty_list;
1762 unsigned i, n = dl[0].mid;
1764 for (i = 1; i <= n; i++) {
1765 mdb_dpage_free(env, dl[i].mptr);
1770 /** Loosen or free a single page.
1771 * Saves single pages to a list for future reuse
1772 * in this same txn. It has been pulled from the freeDB
1773 * and already resides on the dirty list, but has been
1774 * deleted. Use these pages first before pulling again
1777 * If the page wasn't dirtied in this txn, just add it
1778 * to this txn's free list.
1781 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1784 pgno_t pgno = mp->mp_pgno;
1785 MDB_txn *txn = mc->mc_txn;
1787 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1788 if (txn->mt_parent) {
1789 MDB_ID2 *dl = txn->mt_u.dirty_list;
1790 /* If txn has a parent, make sure the page is in our
1794 unsigned x = mdb_mid2l_search(dl, pgno);
1795 if (x <= dl[0].mid && dl[x].mid == pgno) {
1796 if (mp != dl[x].mptr) { /* bad cursor? */
1797 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1798 txn->mt_flags |= MDB_TXN_ERROR;
1799 return MDB_CORRUPTED;
1806 /* no parent txn, so it's just ours */
1811 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1813 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1814 txn->mt_loose_pgs = mp;
1815 txn->mt_loose_count++;
1816 mp->mp_flags |= P_LOOSE;
1818 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1826 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1827 * @param[in] mc A cursor handle for the current operation.
1828 * @param[in] pflags Flags of the pages to update:
1829 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1830 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1831 * @return 0 on success, non-zero on failure.
1834 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1836 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1837 MDB_txn *txn = mc->mc_txn;
1843 int rc = MDB_SUCCESS, level;
1845 /* Mark pages seen by cursors */
1846 if (mc->mc_flags & C_UNTRACK)
1847 mc = NULL; /* will find mc in mt_cursors */
1848 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1849 for (; mc; mc=mc->mc_next) {
1850 if (!(mc->mc_flags & C_INITIALIZED))
1852 for (m3 = mc;; m3 = &mx->mx_cursor) {
1854 for (j=0; j<m3->mc_snum; j++) {
1856 if ((mp->mp_flags & Mask) == pflags)
1857 mp->mp_flags ^= P_KEEP;
1859 mx = m3->mc_xcursor;
1860 /* Proceed to mx if it is at a sub-database */
1861 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1863 if (! (mp && (mp->mp_flags & P_LEAF)))
1865 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1866 if (!(leaf->mn_flags & F_SUBDATA))
1875 /* Mark dirty root pages */
1876 for (i=0; i<txn->mt_numdbs; i++) {
1877 if (txn->mt_dbflags[i] & DB_DIRTY) {
1878 pgno_t pgno = txn->mt_dbs[i].md_root;
1879 if (pgno == P_INVALID)
1881 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1883 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1884 dp->mp_flags ^= P_KEEP;
1892 static int mdb_page_flush(MDB_txn *txn, int keep);
1894 /** Spill pages from the dirty list back to disk.
1895 * This is intended to prevent running into #MDB_TXN_FULL situations,
1896 * but note that they may still occur in a few cases:
1897 * 1) our estimate of the txn size could be too small. Currently this
1898 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1899 * 2) child txns may run out of space if their parents dirtied a
1900 * lot of pages and never spilled them. TODO: we probably should do
1901 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1902 * the parent's dirty_room is below a given threshold.
1904 * Otherwise, if not using nested txns, it is expected that apps will
1905 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1906 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1907 * If the txn never references them again, they can be left alone.
1908 * If the txn only reads them, they can be used without any fuss.
1909 * If the txn writes them again, they can be dirtied immediately without
1910 * going thru all of the work of #mdb_page_touch(). Such references are
1911 * handled by #mdb_page_unspill().
1913 * Also note, we never spill DB root pages, nor pages of active cursors,
1914 * because we'll need these back again soon anyway. And in nested txns,
1915 * we can't spill a page in a child txn if it was already spilled in a
1916 * parent txn. That would alter the parent txns' data even though
1917 * the child hasn't committed yet, and we'd have no way to undo it if
1918 * the child aborted.
1920 * @param[in] m0 cursor A cursor handle identifying the transaction and
1921 * database for which we are checking space.
1922 * @param[in] key For a put operation, the key being stored.
1923 * @param[in] data For a put operation, the data being stored.
1924 * @return 0 on success, non-zero on failure.
1927 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1929 MDB_txn *txn = m0->mc_txn;
1931 MDB_ID2L dl = txn->mt_u.dirty_list;
1932 unsigned int i, j, need;
1935 if (m0->mc_flags & C_SUB)
1938 /* Estimate how much space this op will take */
1939 i = m0->mc_db->md_depth;
1940 /* Named DBs also dirty the main DB */
1941 if (m0->mc_dbi > MAIN_DBI)
1942 i += txn->mt_dbs[MAIN_DBI].md_depth;
1943 /* For puts, roughly factor in the key+data size */
1945 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1946 i += i; /* double it for good measure */
1949 if (txn->mt_dirty_room > i)
1952 if (!txn->mt_spill_pgs) {
1953 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1954 if (!txn->mt_spill_pgs)
1957 /* purge deleted slots */
1958 MDB_IDL sl = txn->mt_spill_pgs;
1959 unsigned int num = sl[0];
1961 for (i=1; i<=num; i++) {
1968 /* Preserve pages which may soon be dirtied again */
1969 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1972 /* Less aggressive spill - we originally spilled the entire dirty list,
1973 * with a few exceptions for cursor pages and DB root pages. But this
1974 * turns out to be a lot of wasted effort because in a large txn many
1975 * of those pages will need to be used again. So now we spill only 1/8th
1976 * of the dirty pages. Testing revealed this to be a good tradeoff,
1977 * better than 1/2, 1/4, or 1/10.
1979 if (need < MDB_IDL_UM_MAX / 8)
1980 need = MDB_IDL_UM_MAX / 8;
1982 /* Save the page IDs of all the pages we're flushing */
1983 /* flush from the tail forward, this saves a lot of shifting later on. */
1984 for (i=dl[0].mid; i && need; i--) {
1985 MDB_ID pn = dl[i].mid << 1;
1987 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1989 /* Can't spill twice, make sure it's not already in a parent's
1992 if (txn->mt_parent) {
1994 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1995 if (tx2->mt_spill_pgs) {
1996 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1997 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1998 dp->mp_flags |= P_KEEP;
2006 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2010 mdb_midl_sort(txn->mt_spill_pgs);
2012 /* Flush the spilled part of dirty list */
2013 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2016 /* Reset any dirty pages we kept that page_flush didn't see */
2017 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2020 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2024 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2026 mdb_find_oldest(MDB_txn *txn)
2029 txnid_t mr, oldest = txn->mt_txnid - 1;
2030 if (txn->mt_env->me_txns) {
2031 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2032 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2043 /** Add a page to the txn's dirty list */
2045 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2048 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2050 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2051 insert = mdb_mid2l_append;
2053 insert = mdb_mid2l_insert;
2055 mid.mid = mp->mp_pgno;
2057 rc = insert(txn->mt_u.dirty_list, &mid);
2058 mdb_tassert(txn, rc == 0);
2059 txn->mt_dirty_room--;
2062 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2063 * me_pghead and mt_next_pgno.
2065 * If there are free pages available from older transactions, they
2066 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2067 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2068 * and move me_pglast to say which records were consumed. Only this
2069 * function can create me_pghead and move me_pglast/mt_next_pgno.
2070 * @param[in] mc cursor A cursor handle identifying the transaction and
2071 * database for which we are allocating.
2072 * @param[in] num the number of pages to allocate.
2073 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2074 * will always be satisfied by a single contiguous chunk of memory.
2075 * @return 0 on success, non-zero on failure.
2078 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2080 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2081 /* Get at most <Max_retries> more freeDB records once me_pghead
2082 * has enough pages. If not enough, use new pages from the map.
2083 * If <Paranoid> and mc is updating the freeDB, only get new
2084 * records if me_pghead is empty. Then the freelist cannot play
2085 * catch-up with itself by growing while trying to save it.
2087 enum { Paranoid = 1, Max_retries = 500 };
2089 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2091 int rc, retry = num * 60;
2092 MDB_txn *txn = mc->mc_txn;
2093 MDB_env *env = txn->mt_env;
2094 pgno_t pgno, *mop = env->me_pghead;
2095 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2097 txnid_t oldest = 0, last;
2102 /* If there are any loose pages, just use them */
2103 if (num == 1 && txn->mt_loose_pgs) {
2104 np = txn->mt_loose_pgs;
2105 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2106 txn->mt_loose_count--;
2107 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2115 /* If our dirty list is already full, we can't do anything */
2116 if (txn->mt_dirty_room == 0) {
2121 for (op = MDB_FIRST;; op = MDB_NEXT) {
2126 /* Seek a big enough contiguous page range. Prefer
2127 * pages at the tail, just truncating the list.
2133 if (mop[i-n2] == pgno+n2)
2140 if (op == MDB_FIRST) { /* 1st iteration */
2141 /* Prepare to fetch more and coalesce */
2142 last = env->me_pglast;
2143 oldest = env->me_pgoldest;
2144 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2147 key.mv_data = &last; /* will look up last+1 */
2148 key.mv_size = sizeof(last);
2150 if (Paranoid && mc->mc_dbi == FREE_DBI)
2153 if (Paranoid && retry < 0 && mop_len)
2157 /* Do not fetch more if the record will be too recent */
2158 if (oldest <= last) {
2160 oldest = mdb_find_oldest(txn);
2161 env->me_pgoldest = oldest;
2167 rc = mdb_cursor_get(&m2, &key, NULL, op);
2169 if (rc == MDB_NOTFOUND)
2173 last = *(txnid_t*)key.mv_data;
2174 if (oldest <= last) {
2176 oldest = mdb_find_oldest(txn);
2177 env->me_pgoldest = oldest;
2183 np = m2.mc_pg[m2.mc_top];
2184 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2185 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2188 idl = (MDB_ID *) data.mv_data;
2191 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2196 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2198 mop = env->me_pghead;
2200 env->me_pglast = last;
2202 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2203 last, txn->mt_dbs[FREE_DBI].md_root, i));
2205 DPRINTF(("IDL %"Z"u", idl[j]));
2207 /* Merge in descending sorted order */
2208 mdb_midl_xmerge(mop, idl);
2212 /* Use new pages from the map when nothing suitable in the freeDB */
2214 pgno = txn->mt_next_pgno;
2215 if (pgno + num >= env->me_maxpg) {
2216 DPUTS("DB size maxed out");
2222 if (env->me_flags & MDB_WRITEMAP) {
2223 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2225 if (!(np = mdb_page_malloc(txn, num))) {
2231 mop[0] = mop_len -= num;
2232 /* Move any stragglers down */
2233 for (j = i-num; j < mop_len; )
2234 mop[++j] = mop[++i];
2236 txn->mt_next_pgno = pgno + num;
2239 mdb_page_dirty(txn, np);
2245 txn->mt_flags |= MDB_TXN_ERROR;
2249 /** Copy the used portions of a non-overflow page.
2250 * @param[in] dst page to copy into
2251 * @param[in] src page to copy from
2252 * @param[in] psize size of a page
2255 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2257 enum { Align = sizeof(pgno_t) };
2258 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2260 /* If page isn't full, just copy the used portion. Adjust
2261 * alignment so memcpy may copy words instead of bytes.
2263 if ((unused &= -Align) && !IS_LEAF2(src)) {
2264 upper = (upper + PAGEBASE) & -Align;
2265 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2266 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2269 memcpy(dst, src, psize - unused);
2273 /** Pull a page off the txn's spill list, if present.
2274 * If a page being referenced was spilled to disk in this txn, bring
2275 * it back and make it dirty/writable again.
2276 * @param[in] txn the transaction handle.
2277 * @param[in] mp the page being referenced. It must not be dirty.
2278 * @param[out] ret the writable page, if any. ret is unchanged if
2279 * mp wasn't spilled.
2282 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2284 MDB_env *env = txn->mt_env;
2287 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2289 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2290 if (!tx2->mt_spill_pgs)
2292 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2293 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2296 if (txn->mt_dirty_room == 0)
2297 return MDB_TXN_FULL;
2298 if (IS_OVERFLOW(mp))
2302 if (env->me_flags & MDB_WRITEMAP) {
2305 np = mdb_page_malloc(txn, num);
2309 memcpy(np, mp, num * env->me_psize);
2311 mdb_page_copy(np, mp, env->me_psize);
2314 /* If in current txn, this page is no longer spilled.
2315 * If it happens to be the last page, truncate the spill list.
2316 * Otherwise mark it as deleted by setting the LSB.
2318 if (x == txn->mt_spill_pgs[0])
2319 txn->mt_spill_pgs[0]--;
2321 txn->mt_spill_pgs[x] |= 1;
2322 } /* otherwise, if belonging to a parent txn, the
2323 * page remains spilled until child commits
2326 mdb_page_dirty(txn, np);
2327 np->mp_flags |= P_DIRTY;
2335 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2336 * @param[in] mc cursor pointing to the page to be touched
2337 * @return 0 on success, non-zero on failure.
2340 mdb_page_touch(MDB_cursor *mc)
2342 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2343 MDB_txn *txn = mc->mc_txn;
2344 MDB_cursor *m2, *m3;
2348 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2349 if (txn->mt_flags & MDB_TXN_SPILLS) {
2351 rc = mdb_page_unspill(txn, mp, &np);
2357 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2358 (rc = mdb_page_alloc(mc, 1, &np)))
2361 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2362 mp->mp_pgno, pgno));
2363 mdb_cassert(mc, mp->mp_pgno != pgno);
2364 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2365 /* Update the parent page, if any, to point to the new page */
2367 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2368 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2369 SETPGNO(node, pgno);
2371 mc->mc_db->md_root = pgno;
2373 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2374 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2376 /* If txn has a parent, make sure the page is in our
2380 unsigned x = mdb_mid2l_search(dl, pgno);
2381 if (x <= dl[0].mid && dl[x].mid == pgno) {
2382 if (mp != dl[x].mptr) { /* bad cursor? */
2383 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2384 txn->mt_flags |= MDB_TXN_ERROR;
2385 return MDB_CORRUPTED;
2390 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2392 np = mdb_page_malloc(txn, 1);
2397 rc = mdb_mid2l_insert(dl, &mid);
2398 mdb_cassert(mc, rc == 0);
2403 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2405 np->mp_flags |= P_DIRTY;
2408 /* Adjust cursors pointing to mp */
2409 mc->mc_pg[mc->mc_top] = np;
2410 m2 = txn->mt_cursors[mc->mc_dbi];
2411 if (mc->mc_flags & C_SUB) {
2412 for (; m2; m2=m2->mc_next) {
2413 m3 = &m2->mc_xcursor->mx_cursor;
2414 if (m3->mc_snum < mc->mc_snum) continue;
2415 if (m3->mc_pg[mc->mc_top] == mp)
2416 m3->mc_pg[mc->mc_top] = np;
2419 for (; m2; m2=m2->mc_next) {
2420 if (m2->mc_snum < mc->mc_snum) continue;
2421 if (m2->mc_pg[mc->mc_top] == mp) {
2422 m2->mc_pg[mc->mc_top] = np;
2423 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2425 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2427 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2428 if (!(leaf->mn_flags & F_SUBDATA))
2429 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2437 txn->mt_flags |= MDB_TXN_ERROR;
2442 mdb_env_sync(MDB_env *env, int force)
2445 if (env->me_flags & MDB_RDONLY)
2447 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2448 if (env->me_flags & MDB_WRITEMAP) {
2449 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2450 ? MS_ASYNC : MS_SYNC;
2451 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2454 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2458 #ifdef BROKEN_FDATASYNC
2459 if (env->me_flags & MDB_FSYNCONLY) {
2460 if (fsync(env->me_fd))
2464 if (MDB_FDATASYNC(env->me_fd))
2471 /** Back up parent txn's cursors, then grab the originals for tracking */
2473 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2475 MDB_cursor *mc, *bk;
2480 for (i = src->mt_numdbs; --i >= 0; ) {
2481 if ((mc = src->mt_cursors[i]) != NULL) {
2482 size = sizeof(MDB_cursor);
2484 size += sizeof(MDB_xcursor);
2485 for (; mc; mc = bk->mc_next) {
2491 mc->mc_db = &dst->mt_dbs[i];
2492 /* Kill pointers into src - and dst to reduce abuse: The
2493 * user may not use mc until dst ends. Otherwise we'd...
2495 mc->mc_txn = NULL; /* ...set this to dst */
2496 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2497 if ((mx = mc->mc_xcursor) != NULL) {
2498 *(MDB_xcursor *)(bk+1) = *mx;
2499 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2501 mc->mc_next = dst->mt_cursors[i];
2502 dst->mt_cursors[i] = mc;
2509 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2510 * @param[in] txn the transaction handle.
2511 * @param[in] merge true to keep changes to parent cursors, false to revert.
2512 * @return 0 on success, non-zero on failure.
2515 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2517 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2521 for (i = txn->mt_numdbs; --i >= 0; ) {
2522 for (mc = cursors[i]; mc; mc = next) {
2524 if ((bk = mc->mc_backup) != NULL) {
2526 /* Commit changes to parent txn */
2527 mc->mc_next = bk->mc_next;
2528 mc->mc_backup = bk->mc_backup;
2529 mc->mc_txn = bk->mc_txn;
2530 mc->mc_db = bk->mc_db;
2531 mc->mc_dbflag = bk->mc_dbflag;
2532 if ((mx = mc->mc_xcursor) != NULL)
2533 mx->mx_cursor.mc_txn = bk->mc_txn;
2535 /* Abort nested txn */
2537 if ((mx = mc->mc_xcursor) != NULL)
2538 *mx = *(MDB_xcursor *)(bk+1);
2542 /* Only malloced cursors are permanently tracked. */
2550 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2553 mdb_txn_reset0(MDB_txn *txn, const char *act);
2555 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2561 Pidset = F_SETLK, Pidcheck = F_GETLK
2565 /** Set or check a pid lock. Set returns 0 on success.
2566 * Check returns 0 if the process is certainly dead, nonzero if it may
2567 * be alive (the lock exists or an error happened so we do not know).
2569 * On Windows Pidset is a no-op, we merely check for the existence
2570 * of the process with the given pid. On POSIX we use a single byte
2571 * lock on the lockfile, set at an offset equal to the pid.
2574 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2576 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2579 if (op == Pidcheck) {
2580 h = OpenProcess(env->me_pidquery, FALSE, pid);
2581 /* No documented "no such process" code, but other program use this: */
2583 return ErrCode() != ERROR_INVALID_PARAMETER;
2584 /* A process exists until all handles to it close. Has it exited? */
2585 ret = WaitForSingleObject(h, 0) != 0;
2592 struct flock lock_info;
2593 memset(&lock_info, 0, sizeof(lock_info));
2594 lock_info.l_type = F_WRLCK;
2595 lock_info.l_whence = SEEK_SET;
2596 lock_info.l_start = pid;
2597 lock_info.l_len = 1;
2598 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2599 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2601 } else if ((rc = ErrCode()) == EINTR) {
2609 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2610 * @param[in] txn the transaction handle to initialize
2611 * @return 0 on success, non-zero on failure.
2614 mdb_txn_renew0(MDB_txn *txn)
2616 MDB_env *env = txn->mt_env;
2617 MDB_txninfo *ti = env->me_txns;
2619 unsigned int i, nr, flags = txn->mt_flags;
2621 int rc, new_notls = 0;
2623 if ((flags &= MDB_TXN_RDONLY) != 0) {
2625 meta = mdb_env_pick_meta(env);
2626 txn->mt_txnid = meta->mm_txnid;
2627 txn->mt_u.reader = NULL;
2629 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2630 pthread_getspecific(env->me_txkey);
2632 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2633 return MDB_BAD_RSLOT;
2635 MDB_PID_T pid = env->me_pid;
2636 MDB_THR_T tid = pthread_self();
2637 mdb_mutexref_t rmutex = env->me_rmutex;
2639 if (!env->me_live_reader) {
2640 rc = mdb_reader_pid(env, Pidset, pid);
2643 env->me_live_reader = 1;
2646 if (LOCK_MUTEX(rc, env, rmutex))
2648 nr = ti->mti_numreaders;
2649 for (i=0; i<nr; i++)
2650 if (ti->mti_readers[i].mr_pid == 0)
2652 if (i == env->me_maxreaders) {
2653 UNLOCK_MUTEX(rmutex);
2654 return MDB_READERS_FULL;
2656 r = &ti->mti_readers[i];
2657 /* Claim the reader slot, carefully since other code
2658 * uses the reader table un-mutexed: First reset the
2659 * slot, next publish it in mti_numreaders. After
2660 * that, it is safe for mdb_env_close() to touch it.
2661 * When it will be closed, we can finally claim it.
2664 r->mr_txnid = (txnid_t)-1;
2667 ti->mti_numreaders = ++nr;
2668 env->me_close_readers = nr;
2670 UNLOCK_MUTEX(rmutex);
2672 new_notls = (env->me_flags & MDB_NOTLS);
2673 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2678 do /* LY: Retry on a race, ITS#7970. */
2679 r->mr_txnid = ti->mti_txnid;
2680 while(r->mr_txnid != ti->mti_txnid);
2681 txn->mt_txnid = r->mr_txnid;
2682 txn->mt_u.reader = r;
2683 meta = env->me_metas[txn->mt_txnid & 1];
2685 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2687 /* Not yet touching txn == env->me_txn0, it may be active */
2689 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2691 txn->mt_txnid = ti->mti_txnid;
2692 meta = env->me_metas[txn->mt_txnid & 1];
2694 meta = mdb_env_pick_meta(env);
2695 txn->mt_txnid = meta->mm_txnid;
2699 if (txn->mt_txnid == mdb_debug_start)
2702 txn->mt_child = NULL;
2703 txn->mt_loose_pgs = NULL;
2704 txn->mt_loose_count = 0;
2705 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2706 txn->mt_u.dirty_list = env->me_dirty_list;
2707 txn->mt_u.dirty_list[0].mid = 0;
2708 txn->mt_free_pgs = env->me_free_pgs;
2709 txn->mt_free_pgs[0] = 0;
2710 txn->mt_spill_pgs = NULL;
2712 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2715 /* Copy the DB info and flags */
2716 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2718 /* Moved to here to avoid a data race in read TXNs */
2719 txn->mt_next_pgno = meta->mm_last_pg+1;
2721 txn->mt_flags = flags;
2724 txn->mt_numdbs = env->me_numdbs;
2725 for (i=2; i<txn->mt_numdbs; i++) {
2726 x = env->me_dbflags[i];
2727 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2728 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2730 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2731 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2733 if (env->me_maxpg < txn->mt_next_pgno) {
2734 mdb_txn_reset0(txn, "renew0-mapfail");
2736 txn->mt_u.reader->mr_pid = 0;
2737 txn->mt_u.reader = NULL;
2739 return MDB_MAP_RESIZED;
2746 mdb_txn_renew(MDB_txn *txn)
2750 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2753 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2754 DPUTS("environment had fatal error, must shutdown!");
2758 rc = mdb_txn_renew0(txn);
2759 if (rc == MDB_SUCCESS) {
2760 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2761 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2762 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2768 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2772 int rc, size, tsize;
2774 flags &= MDB_TXN_BEGIN_FLAGS;
2775 flags |= env->me_flags & MDB_WRITEMAP;
2777 if (env->me_flags & MDB_FATAL_ERROR) {
2778 DPUTS("environment had fatal error, must shutdown!");
2781 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2785 /* Nested transactions: Max 1 child, write txns only, no writemap */
2786 flags |= parent->mt_flags;
2787 if (parent->mt_child ||
2788 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2790 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2792 /* Child txns save MDB_pgstate and use own copy of cursors */
2793 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2794 size += tsize = sizeof(MDB_ntxn);
2795 } else if (flags & MDB_RDONLY) {
2796 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2797 size += tsize = sizeof(MDB_txn);
2799 /* Reuse preallocated write txn. However, do not touch it until
2800 * mdb_txn_renew0() succeeds, since it currently may be active.
2805 if ((txn = calloc(1, size)) == NULL) {
2806 DPRINTF(("calloc: %s", strerror(errno)));
2809 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2810 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2811 txn->mt_flags = flags;
2816 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2817 txn->mt_dbiseqs = parent->mt_dbiseqs;
2818 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2819 if (!txn->mt_u.dirty_list ||
2820 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2822 free(txn->mt_u.dirty_list);
2826 txn->mt_txnid = parent->mt_txnid;
2827 txn->mt_dirty_room = parent->mt_dirty_room;
2828 txn->mt_u.dirty_list[0].mid = 0;
2829 txn->mt_spill_pgs = NULL;
2830 txn->mt_next_pgno = parent->mt_next_pgno;
2831 parent->mt_child = txn;
2832 txn->mt_parent = parent;
2833 txn->mt_numdbs = parent->mt_numdbs;
2834 txn->mt_dbxs = parent->mt_dbxs;
2835 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2836 /* Copy parent's mt_dbflags, but clear DB_NEW */
2837 for (i=0; i<txn->mt_numdbs; i++)
2838 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2840 ntxn = (MDB_ntxn *)txn;
2841 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2842 if (env->me_pghead) {
2843 size = MDB_IDL_SIZEOF(env->me_pghead);
2844 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2846 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2851 rc = mdb_cursor_shadow(parent, txn);
2853 mdb_txn_reset0(txn, "beginchild-fail");
2854 } else { /* MDB_RDONLY */
2855 txn->mt_dbiseqs = env->me_dbiseqs;
2857 rc = mdb_txn_renew0(txn);
2860 if (txn != env->me_txn0)
2863 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2865 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2866 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2867 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2874 mdb_txn_env(MDB_txn *txn)
2876 if(!txn) return NULL;
2881 mdb_txn_id(MDB_txn *txn)
2884 return txn->mt_txnid;
2887 /** Export or close DBI handles opened in this txn. */
2889 mdb_dbis_update(MDB_txn *txn, int keep)
2892 MDB_dbi n = txn->mt_numdbs;
2893 MDB_env *env = txn->mt_env;
2894 unsigned char *tdbflags = txn->mt_dbflags;
2896 for (i = n; --i >= 2;) {
2897 if (tdbflags[i] & DB_NEW) {
2899 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2901 char *ptr = env->me_dbxs[i].md_name.mv_data;
2903 env->me_dbxs[i].md_name.mv_data = NULL;
2904 env->me_dbxs[i].md_name.mv_size = 0;
2905 env->me_dbflags[i] = 0;
2906 env->me_dbiseqs[i]++;
2912 if (keep && env->me_numdbs < n)
2916 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2917 * May be called twice for readonly txns: First reset it, then abort.
2918 * @param[in] txn the transaction handle to reset
2919 * @param[in] act why the transaction is being reset
2922 mdb_txn_reset0(MDB_txn *txn, const char *act)
2924 MDB_env *env = txn->mt_env;
2926 /* Close any DBI handles opened in this txn */
2927 mdb_dbis_update(txn, 0);
2929 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2930 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2931 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2933 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2934 if (txn->mt_u.reader) {
2935 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2936 if (!(env->me_flags & MDB_NOTLS))
2937 txn->mt_u.reader = NULL; /* txn does not own reader */
2939 txn->mt_numdbs = 0; /* close nothing if called again */
2940 txn->mt_dbxs = NULL; /* mark txn as reset */
2942 pgno_t *pghead = env->me_pghead;
2944 mdb_cursors_close(txn, 0);
2945 if (!(env->me_flags & MDB_WRITEMAP)) {
2946 mdb_dlist_free(txn);
2949 if (!txn->mt_parent) {
2950 mdb_midl_shrink(&txn->mt_free_pgs);
2951 env->me_free_pgs = txn->mt_free_pgs;
2953 env->me_pghead = NULL;
2957 /* The writer mutex was locked in mdb_txn_begin. */
2959 UNLOCK_MUTEX(env->me_wmutex);
2961 txn->mt_parent->mt_child = NULL;
2962 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2963 mdb_midl_free(txn->mt_free_pgs);
2964 mdb_midl_free(txn->mt_spill_pgs);
2965 free(txn->mt_u.dirty_list);
2968 mdb_midl_free(pghead);
2973 mdb_txn_reset(MDB_txn *txn)
2978 /* This call is only valid for read-only txns */
2979 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2982 mdb_txn_reset0(txn, "reset");
2986 mdb_txn_abort(MDB_txn *txn)
2992 mdb_txn_abort(txn->mt_child);
2994 mdb_txn_reset0(txn, "abort");
2995 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2996 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2997 txn->mt_u.reader->mr_pid = 0;
2999 if (txn != txn->mt_env->me_txn0)
3003 /** Save the freelist as of this transaction to the freeDB.
3004 * This changes the freelist. Keep trying until it stabilizes.
3007 mdb_freelist_save(MDB_txn *txn)
3009 /* env->me_pghead[] can grow and shrink during this call.
3010 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3011 * Page numbers cannot disappear from txn->mt_free_pgs[].
3014 MDB_env *env = txn->mt_env;
3015 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3016 txnid_t pglast = 0, head_id = 0;
3017 pgno_t freecnt = 0, *free_pgs, *mop;
3018 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3020 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3022 if (env->me_pghead) {
3023 /* Make sure first page of freeDB is touched and on freelist */
3024 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3025 if (rc && rc != MDB_NOTFOUND)
3029 if (!env->me_pghead && txn->mt_loose_pgs) {
3030 /* Put loose page numbers in mt_free_pgs, since
3031 * we may be unable to return them to me_pghead.
3033 MDB_page *mp = txn->mt_loose_pgs;
3034 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3036 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3037 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3038 txn->mt_loose_pgs = NULL;
3039 txn->mt_loose_count = 0;
3042 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3043 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3044 ? SSIZE_MAX : maxfree_1pg;
3047 /* Come back here after each Put() in case freelist changed */
3052 /* If using records from freeDB which we have not yet
3053 * deleted, delete them and any we reserved for me_pghead.
3055 while (pglast < env->me_pglast) {
3056 rc = mdb_cursor_first(&mc, &key, NULL);
3059 pglast = head_id = *(txnid_t *)key.mv_data;
3060 total_room = head_room = 0;
3061 mdb_tassert(txn, pglast <= env->me_pglast);
3062 rc = mdb_cursor_del(&mc, 0);
3067 /* Save the IDL of pages freed by this txn, to a single record */
3068 if (freecnt < txn->mt_free_pgs[0]) {
3070 /* Make sure last page of freeDB is touched and on freelist */
3071 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3072 if (rc && rc != MDB_NOTFOUND)
3075 free_pgs = txn->mt_free_pgs;
3076 /* Write to last page of freeDB */
3077 key.mv_size = sizeof(txn->mt_txnid);
3078 key.mv_data = &txn->mt_txnid;
3080 freecnt = free_pgs[0];
3081 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3082 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3085 /* Retry if mt_free_pgs[] grew during the Put() */
3086 free_pgs = txn->mt_free_pgs;
3087 } while (freecnt < free_pgs[0]);
3088 mdb_midl_sort(free_pgs);
3089 memcpy(data.mv_data, free_pgs, data.mv_size);
3092 unsigned int i = free_pgs[0];
3093 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3094 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3096 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3102 mop = env->me_pghead;
3103 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3105 /* Reserve records for me_pghead[]. Split it if multi-page,
3106 * to avoid searching freeDB for a page range. Use keys in
3107 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3109 if (total_room >= mop_len) {
3110 if (total_room == mop_len || --more < 0)
3112 } else if (head_room >= maxfree_1pg && head_id > 1) {
3113 /* Keep current record (overflow page), add a new one */
3117 /* (Re)write {key = head_id, IDL length = head_room} */
3118 total_room -= head_room;
3119 head_room = mop_len - total_room;
3120 if (head_room > maxfree_1pg && head_id > 1) {
3121 /* Overflow multi-page for part of me_pghead */
3122 head_room /= head_id; /* amortize page sizes */
3123 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3124 } else if (head_room < 0) {
3125 /* Rare case, not bothering to delete this record */
3128 key.mv_size = sizeof(head_id);
3129 key.mv_data = &head_id;
3130 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3131 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3134 /* IDL is initially empty, zero out at least the length */
3135 pgs = (pgno_t *)data.mv_data;
3136 j = head_room > clean_limit ? head_room : 0;
3140 total_room += head_room;
3143 /* Return loose page numbers to me_pghead, though usually none are
3144 * left at this point. The pages themselves remain in dirty_list.
3146 if (txn->mt_loose_pgs) {
3147 MDB_page *mp = txn->mt_loose_pgs;
3148 unsigned count = txn->mt_loose_count;
3150 /* Room for loose pages + temp IDL with same */
3151 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3153 mop = env->me_pghead;
3154 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3155 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3156 loose[ ++count ] = mp->mp_pgno;
3158 mdb_midl_sort(loose);
3159 mdb_midl_xmerge(mop, loose);
3160 txn->mt_loose_pgs = NULL;
3161 txn->mt_loose_count = 0;
3165 /* Fill in the reserved me_pghead records */
3171 rc = mdb_cursor_first(&mc, &key, &data);
3172 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3173 txnid_t id = *(txnid_t *)key.mv_data;
3174 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3177 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3179 if (len > mop_len) {
3181 data.mv_size = (len + 1) * sizeof(MDB_ID);
3183 data.mv_data = mop -= len;
3186 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3188 if (rc || !(mop_len -= len))
3195 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3196 * @param[in] txn the transaction that's being committed
3197 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3198 * @return 0 on success, non-zero on failure.
3201 mdb_page_flush(MDB_txn *txn, int keep)
3203 MDB_env *env = txn->mt_env;
3204 MDB_ID2L dl = txn->mt_u.dirty_list;
3205 unsigned psize = env->me_psize, j;
3206 int i, pagecount = dl[0].mid, rc;
3207 size_t size = 0, pos = 0;
3209 MDB_page *dp = NULL;
3213 struct iovec iov[MDB_COMMIT_PAGES];
3214 ssize_t wpos = 0, wsize = 0, wres;
3215 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3221 if (env->me_flags & MDB_WRITEMAP) {
3222 /* Clear dirty flags */
3223 while (++i <= pagecount) {
3225 /* Don't flush this page yet */
3226 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3227 dp->mp_flags &= ~P_KEEP;
3231 dp->mp_flags &= ~P_DIRTY;
3236 /* Write the pages */
3238 if (++i <= pagecount) {
3240 /* Don't flush this page yet */
3241 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3242 dp->mp_flags &= ~P_KEEP;
3247 /* clear dirty flag */
3248 dp->mp_flags &= ~P_DIRTY;
3251 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3256 /* Windows actually supports scatter/gather I/O, but only on
3257 * unbuffered file handles. Since we're relying on the OS page
3258 * cache for all our data, that's self-defeating. So we just
3259 * write pages one at a time. We use the ov structure to set
3260 * the write offset, to at least save the overhead of a Seek
3263 DPRINTF(("committing page %"Z"u", pgno));
3264 memset(&ov, 0, sizeof(ov));
3265 ov.Offset = pos & 0xffffffff;
3266 ov.OffsetHigh = pos >> 16 >> 16;
3267 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3269 DPRINTF(("WriteFile: %d", rc));
3273 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3274 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3277 /* Write previous page(s) */
3278 #ifdef MDB_USE_PWRITEV
3279 wres = pwritev(env->me_fd, iov, n, wpos);
3282 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3285 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3289 DPRINTF(("lseek: %s", strerror(rc)));
3292 wres = writev(env->me_fd, iov, n);
3295 if (wres != wsize) {
3300 DPRINTF(("Write error: %s", strerror(rc)));
3302 rc = EIO; /* TODO: Use which error code? */
3303 DPUTS("short write, filesystem full?");
3314 DPRINTF(("committing page %"Z"u", pgno));
3315 next_pos = pos + size;
3316 iov[n].iov_len = size;
3317 iov[n].iov_base = (char *)dp;
3323 /* MIPS has cache coherency issues, this is a no-op everywhere else
3324 * Note: for any size >= on-chip cache size, entire on-chip cache is
3327 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3329 for (i = keep; ++i <= pagecount; ) {
3331 /* This is a page we skipped above */
3334 dl[j].mid = dp->mp_pgno;
3337 mdb_dpage_free(env, dp);
3342 txn->mt_dirty_room += i - j;
3348 mdb_txn_commit(MDB_txn *txn)
3357 if (txn->mt_child) {
3358 rc = mdb_txn_commit(txn->mt_child);
3365 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3366 mdb_dbis_update(txn, 1);
3367 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3372 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3373 DPUTS("error flag is set, can't commit");
3375 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3380 if (txn->mt_parent) {
3381 MDB_txn *parent = txn->mt_parent;
3385 unsigned x, y, len, ps_len;
3387 /* Append our free list to parent's */
3388 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3391 mdb_midl_free(txn->mt_free_pgs);
3392 /* Failures after this must either undo the changes
3393 * to the parent or set MDB_TXN_ERROR in the parent.
3396 parent->mt_next_pgno = txn->mt_next_pgno;
3397 parent->mt_flags = txn->mt_flags;
3399 /* Merge our cursors into parent's and close them */
3400 mdb_cursors_close(txn, 1);
3402 /* Update parent's DB table. */
3403 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3404 parent->mt_numdbs = txn->mt_numdbs;
3405 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3406 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3407 for (i=2; i<txn->mt_numdbs; i++) {
3408 /* preserve parent's DB_NEW status */
3409 x = parent->mt_dbflags[i] & DB_NEW;
3410 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3413 dst = parent->mt_u.dirty_list;
3414 src = txn->mt_u.dirty_list;
3415 /* Remove anything in our dirty list from parent's spill list */
3416 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3418 pspill[0] = (pgno_t)-1;
3419 /* Mark our dirty pages as deleted in parent spill list */
3420 for (i=0, len=src[0].mid; ++i <= len; ) {
3421 MDB_ID pn = src[i].mid << 1;
3422 while (pn > pspill[x])
3424 if (pn == pspill[x]) {
3429 /* Squash deleted pagenums if we deleted any */
3430 for (x=y; ++x <= ps_len; )
3431 if (!(pspill[x] & 1))
3432 pspill[++y] = pspill[x];
3436 /* Find len = length of merging our dirty list with parent's */
3438 dst[0].mid = 0; /* simplify loops */
3439 if (parent->mt_parent) {
3440 len = x + src[0].mid;
3441 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3442 for (i = x; y && i; y--) {
3443 pgno_t yp = src[y].mid;
3444 while (yp < dst[i].mid)
3446 if (yp == dst[i].mid) {
3451 } else { /* Simplify the above for single-ancestor case */
3452 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3454 /* Merge our dirty list with parent's */
3456 for (i = len; y; dst[i--] = src[y--]) {
3457 pgno_t yp = src[y].mid;
3458 while (yp < dst[x].mid)
3459 dst[i--] = dst[x--];
3460 if (yp == dst[x].mid)
3461 free(dst[x--].mptr);
3463 mdb_tassert(txn, i == x);
3465 free(txn->mt_u.dirty_list);
3466 parent->mt_dirty_room = txn->mt_dirty_room;
3467 if (txn->mt_spill_pgs) {
3468 if (parent->mt_spill_pgs) {
3469 /* TODO: Prevent failure here, so parent does not fail */
3470 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3472 parent->mt_flags |= MDB_TXN_ERROR;
3473 mdb_midl_free(txn->mt_spill_pgs);
3474 mdb_midl_sort(parent->mt_spill_pgs);
3476 parent->mt_spill_pgs = txn->mt_spill_pgs;
3480 /* Append our loose page list to parent's */
3481 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3483 *lp = txn->mt_loose_pgs;
3484 parent->mt_loose_count += txn->mt_loose_count;
3486 parent->mt_child = NULL;
3487 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3492 if (txn != env->me_txn) {
3493 DPUTS("attempt to commit unknown transaction");
3498 mdb_cursors_close(txn, 0);
3500 if (!txn->mt_u.dirty_list[0].mid &&
3501 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3504 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3505 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3507 /* Update DB root pointers */
3508 if (txn->mt_numdbs > 2) {
3512 data.mv_size = sizeof(MDB_db);
3514 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3515 for (i = 2; i < txn->mt_numdbs; i++) {
3516 if (txn->mt_dbflags[i] & DB_DIRTY) {
3517 if (TXN_DBI_CHANGED(txn, i)) {
3521 data.mv_data = &txn->mt_dbs[i];
3522 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3530 rc = mdb_freelist_save(txn);
3534 mdb_midl_free(env->me_pghead);
3535 env->me_pghead = NULL;
3536 mdb_midl_shrink(&txn->mt_free_pgs);
3537 env->me_free_pgs = txn->mt_free_pgs;
3543 if ((rc = mdb_page_flush(txn, 0)))
3545 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3546 (rc = mdb_env_sync(env, 0)))
3548 if ((rc = mdb_env_write_meta(txn)))
3551 /* Free P_LOOSE pages left behind in dirty_list */
3552 if (!(env->me_flags & MDB_WRITEMAP))
3553 mdb_dlist_free(txn);
3558 mdb_dbis_update(txn, 1);
3561 UNLOCK_MUTEX(env->me_wmutex);
3562 if (txn != env->me_txn0)
3572 /** Read the environment parameters of a DB environment before
3573 * mapping it into memory.
3574 * @param[in] env the environment handle
3575 * @param[out] meta address of where to store the meta information
3576 * @return 0 on success, non-zero on failure.
3579 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3585 enum { Size = sizeof(pbuf) };
3587 /* We don't know the page size yet, so use a minimum value.
3588 * Read both meta pages so we can use the latest one.
3591 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3595 memset(&ov, 0, sizeof(ov));
3597 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3598 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3601 rc = pread(env->me_fd, &pbuf, Size, off);
3604 if (rc == 0 && off == 0)
3606 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3607 DPRINTF(("read: %s", mdb_strerror(rc)));
3611 p = (MDB_page *)&pbuf;
3613 if (!F_ISSET(p->mp_flags, P_META)) {
3614 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3619 if (m->mm_magic != MDB_MAGIC) {
3620 DPUTS("meta has invalid magic");
3624 if (m->mm_version != MDB_DATA_VERSION) {
3625 DPRINTF(("database is version %u, expected version %u",
3626 m->mm_version, MDB_DATA_VERSION));
3627 return MDB_VERSION_MISMATCH;
3630 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3636 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3638 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3640 meta->mm_magic = MDB_MAGIC;
3641 meta->mm_version = MDB_DATA_VERSION;
3642 meta->mm_mapsize = env->me_mapsize;
3643 meta->mm_psize = env->me_psize;
3644 meta->mm_last_pg = 1;
3645 meta->mm_flags = env->me_flags & 0xffff;
3646 meta->mm_flags |= MDB_INTEGERKEY;
3647 meta->mm_dbs[0].md_root = P_INVALID;
3648 meta->mm_dbs[1].md_root = P_INVALID;
3651 /** Write the environment parameters of a freshly created DB environment.
3652 * @param[in] env the environment handle
3653 * @param[in] meta the #MDB_meta to write
3654 * @return 0 on success, non-zero on failure.
3657 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3665 memset(&ov, 0, sizeof(ov));
3666 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3668 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3671 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3672 len = pwrite(fd, ptr, size, pos); \
3673 if (len == -1 && ErrCode() == EINTR) continue; \
3674 rc = (len >= 0); break; } while(1)
3677 DPUTS("writing new meta page");
3679 psize = env->me_psize;
3681 p = calloc(2, psize);
3683 p->mp_flags = P_META;
3684 *(MDB_meta *)METADATA(p) = *meta;
3686 q = (MDB_page *)((char *)p + psize);
3688 q->mp_flags = P_META;
3689 *(MDB_meta *)METADATA(q) = *meta;
3691 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3694 else if ((unsigned) len == psize * 2)
3702 /** Update the environment info to commit a transaction.
3703 * @param[in] txn the transaction that's being committed
3704 * @return 0 on success, non-zero on failure.
3707 mdb_env_write_meta(MDB_txn *txn)
3710 MDB_meta meta, metab, *mp;
3714 int rc, len, toggle;
3723 toggle = txn->mt_txnid & 1;
3724 DPRINTF(("writing meta page %d for root page %"Z"u",
3725 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3728 flags = txn->mt_flags & env->me_flags;
3729 mp = env->me_metas[toggle];
3730 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3731 /* Persist any increases of mapsize config */
3732 if (mapsize < env->me_mapsize)
3733 mapsize = env->me_mapsize;
3735 if (flags & MDB_WRITEMAP) {
3736 mp->mm_mapsize = mapsize;
3737 mp->mm_dbs[0] = txn->mt_dbs[0];
3738 mp->mm_dbs[1] = txn->mt_dbs[1];
3739 mp->mm_last_pg = txn->mt_next_pgno - 1;
3740 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3741 !(defined(__i386__) || defined(__x86_64__))
3742 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3743 __sync_synchronize();
3745 mp->mm_txnid = txn->mt_txnid;
3746 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3747 unsigned meta_size = env->me_psize;
3748 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3751 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3752 if (meta_size < env->me_os_psize)
3753 meta_size += meta_size;
3758 if (MDB_MSYNC(ptr, meta_size, rc)) {
3765 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3766 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3768 meta.mm_mapsize = mapsize;
3769 meta.mm_dbs[0] = txn->mt_dbs[0];
3770 meta.mm_dbs[1] = txn->mt_dbs[1];
3771 meta.mm_last_pg = txn->mt_next_pgno - 1;
3772 meta.mm_txnid = txn->mt_txnid;
3774 off = offsetof(MDB_meta, mm_mapsize);
3775 ptr = (char *)&meta + off;
3776 len = sizeof(MDB_meta) - off;
3778 off += env->me_psize;
3781 /* Write to the SYNC fd */
3782 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3786 memset(&ov, 0, sizeof(ov));
3788 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3792 rc = pwrite(mfd, ptr, len, off);
3795 rc = rc < 0 ? ErrCode() : EIO;
3798 DPUTS("write failed, disk error?");
3799 /* On a failure, the pagecache still contains the new data.
3800 * Write some old data back, to prevent it from being used.
3801 * Use the non-SYNC fd; we know it will fail anyway.
3803 meta.mm_last_pg = metab.mm_last_pg;
3804 meta.mm_txnid = metab.mm_txnid;
3806 memset(&ov, 0, sizeof(ov));
3808 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3810 r2 = pwrite(env->me_fd, ptr, len, off);
3811 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3814 env->me_flags |= MDB_FATAL_ERROR;
3817 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3818 CACHEFLUSH(env->me_map + off, len, DCACHE);
3820 /* Memory ordering issues are irrelevant; since the entire writer
3821 * is wrapped by wmutex, all of these changes will become visible
3822 * after the wmutex is unlocked. Since the DB is multi-version,
3823 * readers will get consistent data regardless of how fresh or
3824 * how stale their view of these values is.
3827 env->me_txns->mti_txnid = txn->mt_txnid;
3832 /** Check both meta pages to see which one is newer.
3833 * @param[in] env the environment handle
3834 * @return newest #MDB_meta.
3837 mdb_env_pick_meta(const MDB_env *env)
3839 MDB_meta *const *metas = env->me_metas;
3840 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3844 mdb_env_create(MDB_env **env)
3848 e = calloc(1, sizeof(MDB_env));
3852 e->me_maxreaders = DEFAULT_READERS;
3853 e->me_maxdbs = e->me_numdbs = 2;
3854 e->me_fd = INVALID_HANDLE_VALUE;
3855 e->me_lfd = INVALID_HANDLE_VALUE;
3856 e->me_mfd = INVALID_HANDLE_VALUE;
3857 #ifdef MDB_USE_POSIX_SEM
3858 e->me_rmutex = SEM_FAILED;
3859 e->me_wmutex = SEM_FAILED;
3860 #elif defined MDB_USE_SYSV_SEM
3861 e->me_rmutex->semid = -1;
3862 e->me_wmutex->semid = -1;
3864 e->me_pid = getpid();
3865 GET_PAGESIZE(e->me_os_psize);
3866 VGMEMP_CREATE(e,0,0);
3872 mdb_env_map(MDB_env *env, void *addr)
3875 unsigned int flags = env->me_flags;
3879 LONG sizelo, sizehi;
3882 if (flags & MDB_RDONLY) {
3883 /* Don't set explicit map size, use whatever exists */
3888 msize = env->me_mapsize;
3889 sizelo = msize & 0xffffffff;
3890 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3892 /* Windows won't create mappings for zero length files.
3893 * and won't map more than the file size.
3894 * Just set the maxsize right now.
3896 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3897 || !SetEndOfFile(env->me_fd)
3898 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3902 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3903 PAGE_READWRITE : PAGE_READONLY,
3904 sizehi, sizelo, NULL);
3907 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3908 FILE_MAP_WRITE : FILE_MAP_READ,
3910 rc = env->me_map ? 0 : ErrCode();
3915 int prot = PROT_READ;
3916 if (flags & MDB_WRITEMAP) {
3918 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3921 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3923 if (env->me_map == MAP_FAILED) {
3928 if (flags & MDB_NORDAHEAD) {
3929 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3931 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3933 #ifdef POSIX_MADV_RANDOM
3934 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3935 #endif /* POSIX_MADV_RANDOM */
3936 #endif /* MADV_RANDOM */
3940 /* Can happen because the address argument to mmap() is just a
3941 * hint. mmap() can pick another, e.g. if the range is in use.
3942 * The MAP_FIXED flag would prevent that, but then mmap could
3943 * instead unmap existing pages to make room for the new map.
3945 if (addr && env->me_map != addr)
3946 return EBUSY; /* TODO: Make a new MDB_* error code? */
3948 p = (MDB_page *)env->me_map;
3949 env->me_metas[0] = METADATA(p);
3950 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3956 mdb_env_set_mapsize(MDB_env *env, size_t size)
3958 /* If env is already open, caller is responsible for making
3959 * sure there are no active txns.
3967 meta = mdb_env_pick_meta(env);
3969 size = meta->mm_mapsize;
3971 /* Silently round up to minimum if the size is too small */
3972 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3976 munmap(env->me_map, env->me_mapsize);
3977 env->me_mapsize = size;
3978 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3979 rc = mdb_env_map(env, old);
3983 env->me_mapsize = size;
3985 env->me_maxpg = env->me_mapsize / env->me_psize;
3990 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3994 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3999 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4001 if (env->me_map || readers < 1)
4003 env->me_maxreaders = readers;
4008 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4010 if (!env || !readers)
4012 *readers = env->me_maxreaders;
4017 mdb_fsize(HANDLE fd, size_t *size)
4020 LARGE_INTEGER fsize;
4022 if (!GetFileSizeEx(fd, &fsize))
4025 *size = fsize.QuadPart;
4037 #ifdef BROKEN_FDATASYNC
4038 #include <sys/utsname.h>
4039 #include <sys/vfs.h>
4042 /** Further setup required for opening an LMDB environment
4045 mdb_env_open2(MDB_env *env)
4047 unsigned int flags = env->me_flags;
4048 int i, newenv = 0, rc;
4052 /* See if we should use QueryLimited */
4054 if ((rc & 0xff) > 5)
4055 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4057 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4060 #ifdef BROKEN_FDATASYNC
4061 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4062 * https://lkml.org/lkml/2012/9/3/83
4063 * Kernels after 3.6-rc6 are known good.
4064 * https://lkml.org/lkml/2012/9/10/556
4065 * See if the DB is on ext3/ext4, then check for new enough kernel
4066 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4071 fstatfs(env->me_fd, &st);
4072 while (st.f_type == 0xEF53) {
4076 if (uts.release[0] < '3') {
4077 if (!strncmp(uts.release, "2.6.32.", 7)) {
4078 i = atoi(uts.release+7);
4080 break; /* 2.6.32.60 and newer is OK */
4081 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4082 i = atoi(uts.release+7);
4084 break; /* 2.6.34.15 and newer is OK */
4086 } else if (uts.release[0] == '3') {
4087 i = atoi(uts.release+2);
4089 break; /* 3.6 and newer is OK */
4091 i = atoi(uts.release+4);
4093 break; /* 3.5.4 and newer is OK */
4094 } else if (i == 2) {
4095 i = atoi(uts.release+4);
4097 break; /* 3.2.30 and newer is OK */
4099 } else { /* 4.x and newer is OK */
4102 env->me_flags |= MDB_FSYNCONLY;
4108 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4111 DPUTS("new mdbenv");
4113 env->me_psize = env->me_os_psize;
4114 if (env->me_psize > MAX_PAGESIZE)
4115 env->me_psize = MAX_PAGESIZE;
4116 memset(&meta, 0, sizeof(meta));
4117 mdb_env_init_meta0(env, &meta);
4118 meta.mm_mapsize = DEFAULT_MAPSIZE;
4120 env->me_psize = meta.mm_psize;
4123 /* Was a mapsize configured? */
4124 if (!env->me_mapsize) {
4125 env->me_mapsize = meta.mm_mapsize;
4128 /* Make sure mapsize >= committed data size. Even when using
4129 * mm_mapsize, which could be broken in old files (ITS#7789).
4131 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4132 if (env->me_mapsize < minsize)
4133 env->me_mapsize = minsize;
4135 meta.mm_mapsize = env->me_mapsize;
4137 if (newenv && !(flags & MDB_FIXEDMAP)) {
4138 /* mdb_env_map() may grow the datafile. Write the metapages
4139 * first, so the file will be valid if initialization fails.
4140 * Except with FIXEDMAP, since we do not yet know mm_address.
4141 * We could fill in mm_address later, but then a different
4142 * program might end up doing that - one with a memory layout
4143 * and map address which does not suit the main program.
4145 rc = mdb_env_init_meta(env, &meta);
4151 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4156 if (flags & MDB_FIXEDMAP)
4157 meta.mm_address = env->me_map;
4158 i = mdb_env_init_meta(env, &meta);
4159 if (i != MDB_SUCCESS) {
4164 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4165 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4167 #if !(MDB_MAXKEYSIZE)
4168 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4170 env->me_maxpg = env->me_mapsize / env->me_psize;
4174 MDB_meta *meta = mdb_env_pick_meta(env);
4175 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4177 DPRINTF(("opened database version %u, pagesize %u",
4178 meta->mm_version, env->me_psize));
4179 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4180 DPRINTF(("depth: %u", db->md_depth));
4181 DPRINTF(("entries: %"Z"u", db->md_entries));
4182 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4183 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4184 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4185 DPRINTF(("root: %"Z"u", db->md_root));
4193 /** Release a reader thread's slot in the reader lock table.
4194 * This function is called automatically when a thread exits.
4195 * @param[in] ptr This points to the slot in the reader lock table.
4198 mdb_env_reader_dest(void *ptr)
4200 MDB_reader *reader = ptr;
4206 /** Junk for arranging thread-specific callbacks on Windows. This is
4207 * necessarily platform and compiler-specific. Windows supports up
4208 * to 1088 keys. Let's assume nobody opens more than 64 environments
4209 * in a single process, for now. They can override this if needed.
4211 #ifndef MAX_TLS_KEYS
4212 #define MAX_TLS_KEYS 64
4214 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4215 static int mdb_tls_nkeys;
4217 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4221 case DLL_PROCESS_ATTACH: break;
4222 case DLL_THREAD_ATTACH: break;
4223 case DLL_THREAD_DETACH:
4224 for (i=0; i<mdb_tls_nkeys; i++) {
4225 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4227 mdb_env_reader_dest(r);
4231 case DLL_PROCESS_DETACH: break;
4236 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4238 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4242 /* Force some symbol references.
4243 * _tls_used forces the linker to create the TLS directory if not already done
4244 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4246 #pragma comment(linker, "/INCLUDE:_tls_used")
4247 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4248 #pragma const_seg(".CRT$XLB")
4249 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4250 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4253 #pragma comment(linker, "/INCLUDE:__tls_used")
4254 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4255 #pragma data_seg(".CRT$XLB")
4256 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4258 #endif /* WIN 32/64 */
4259 #endif /* !__GNUC__ */
4262 /** Downgrade the exclusive lock on the region back to shared */
4264 mdb_env_share_locks(MDB_env *env, int *excl)
4267 MDB_meta *meta = mdb_env_pick_meta(env);
4269 env->me_txns->mti_txnid = meta->mm_txnid;
4274 /* First acquire a shared lock. The Unlock will
4275 * then release the existing exclusive lock.
4277 memset(&ov, 0, sizeof(ov));
4278 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4281 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4287 struct flock lock_info;
4288 /* The shared lock replaces the existing lock */
4289 memset((void *)&lock_info, 0, sizeof(lock_info));
4290 lock_info.l_type = F_RDLCK;
4291 lock_info.l_whence = SEEK_SET;
4292 lock_info.l_start = 0;
4293 lock_info.l_len = 1;
4294 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4295 (rc = ErrCode()) == EINTR) ;
4296 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4303 /** Try to get exclusive lock, otherwise shared.
4304 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4307 mdb_env_excl_lock(MDB_env *env, int *excl)
4311 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4315 memset(&ov, 0, sizeof(ov));
4316 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4323 struct flock lock_info;
4324 memset((void *)&lock_info, 0, sizeof(lock_info));
4325 lock_info.l_type = F_WRLCK;
4326 lock_info.l_whence = SEEK_SET;
4327 lock_info.l_start = 0;
4328 lock_info.l_len = 1;
4329 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4330 (rc = ErrCode()) == EINTR) ;
4334 # ifndef MDB_USE_POSIX_MUTEX
4335 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4338 lock_info.l_type = F_RDLCK;
4339 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4340 (rc = ErrCode()) == EINTR) ;
4350 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4352 * @(#) $Revision: 5.1 $
4353 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4354 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4356 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4360 * Please do not copyright this code. This code is in the public domain.
4362 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4363 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4364 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4365 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4366 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4367 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4368 * PERFORMANCE OF THIS SOFTWARE.
4371 * chongo <Landon Curt Noll> /\oo/\
4372 * http://www.isthe.com/chongo/
4374 * Share and Enjoy! :-)
4377 typedef unsigned long long mdb_hash_t;
4378 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4380 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4381 * @param[in] val value to hash
4382 * @param[in] hval initial value for hash
4383 * @return 64 bit hash
4385 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4386 * hval arg on the first call.
4389 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4391 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4392 unsigned char *end = s + val->mv_size;
4394 * FNV-1a hash each octet of the string
4397 /* xor the bottom with the current octet */
4398 hval ^= (mdb_hash_t)*s++;
4400 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4401 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4402 (hval << 7) + (hval << 8) + (hval << 40);
4404 /* return our new hash value */
4408 /** Hash the string and output the encoded hash.
4409 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4410 * very short name limits. We don't care about the encoding being reversible,
4411 * we just want to preserve as many bits of the input as possible in a
4412 * small printable string.
4413 * @param[in] str string to hash
4414 * @param[out] encbuf an array of 11 chars to hold the hash
4416 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4419 mdb_pack85(unsigned long l, char *out)
4423 for (i=0; i<5; i++) {
4424 *out++ = mdb_a85[l % 85];
4430 mdb_hash_enc(MDB_val *val, char *encbuf)
4432 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4434 mdb_pack85(h, encbuf);
4435 mdb_pack85(h>>32, encbuf+5);
4440 /** Open and/or initialize the lock region for the environment.
4441 * @param[in] env The LMDB environment.
4442 * @param[in] lpath The pathname of the file used for the lock region.
4443 * @param[in] mode The Unix permissions for the file, if we create it.
4444 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4445 * @return 0 on success, non-zero on failure.
4448 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4451 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4453 # define MDB_ERRCODE_ROFS EROFS
4454 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4455 # define MDB_CLOEXEC O_CLOEXEC
4458 # define MDB_CLOEXEC 0
4461 #ifdef MDB_USE_SYSV_SEM
4469 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4470 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4471 FILE_ATTRIBUTE_NORMAL, NULL);
4473 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4475 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4477 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4482 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4483 /* Lose record locks when exec*() */
4484 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4485 fcntl(env->me_lfd, F_SETFD, fdflags);
4488 if (!(env->me_flags & MDB_NOTLS)) {
4489 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4492 env->me_flags |= MDB_ENV_TXKEY;
4494 /* Windows TLS callbacks need help finding their TLS info. */
4495 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4499 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4503 /* Try to get exclusive lock. If we succeed, then
4504 * nobody is using the lock region and we should initialize it.
4506 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4509 size = GetFileSize(env->me_lfd, NULL);
4511 size = lseek(env->me_lfd, 0, SEEK_END);
4512 if (size == -1) goto fail_errno;
4514 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4515 if (size < rsize && *excl > 0) {
4517 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4518 || !SetEndOfFile(env->me_lfd))
4521 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4525 size = rsize - sizeof(MDB_txninfo);
4526 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4531 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4533 if (!mh) goto fail_errno;
4534 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4536 if (!env->me_txns) goto fail_errno;
4538 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4540 if (m == MAP_FAILED) goto fail_errno;
4546 BY_HANDLE_FILE_INFORMATION stbuf;
4555 if (!mdb_sec_inited) {
4556 InitializeSecurityDescriptor(&mdb_null_sd,
4557 SECURITY_DESCRIPTOR_REVISION);
4558 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4559 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4560 mdb_all_sa.bInheritHandle = FALSE;
4561 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4564 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4565 idbuf.volume = stbuf.dwVolumeSerialNumber;
4566 idbuf.nhigh = stbuf.nFileIndexHigh;
4567 idbuf.nlow = stbuf.nFileIndexLow;
4568 val.mv_data = &idbuf;
4569 val.mv_size = sizeof(idbuf);
4570 mdb_hash_enc(&val, encbuf);
4571 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4572 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4573 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4574 if (!env->me_rmutex) goto fail_errno;
4575 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4576 if (!env->me_wmutex) goto fail_errno;
4577 #elif defined(MDB_USE_POSIX_SEM)
4586 #if defined(__NetBSD__)
4587 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4589 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4590 idbuf.dev = stbuf.st_dev;
4591 idbuf.ino = stbuf.st_ino;
4592 val.mv_data = &idbuf;
4593 val.mv_size = sizeof(idbuf);
4594 mdb_hash_enc(&val, encbuf);
4595 #ifdef MDB_SHORT_SEMNAMES
4596 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4598 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4599 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4600 /* Clean up after a previous run, if needed: Try to
4601 * remove both semaphores before doing anything else.
4603 sem_unlink(env->me_txns->mti_rmname);
4604 sem_unlink(env->me_txns->mti_wmname);
4605 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4606 O_CREAT|O_EXCL, mode, 1);
4607 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4608 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4609 O_CREAT|O_EXCL, mode, 1);
4610 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4611 #elif defined(MDB_USE_SYSV_SEM)
4612 unsigned short vals[2] = {1, 1};
4613 key_t key = ftok(lpath, 'M');
4616 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4620 if (semctl(semid, 0, SETALL, semu) < 0)
4622 env->me_txns->mti_semid = semid;
4623 #else /* MDB_USE_POSIX_MUTEX: */
4624 pthread_mutexattr_t mattr;
4626 if ((rc = pthread_mutexattr_init(&mattr))
4627 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4628 #ifdef MDB_ROBUST_SUPPORTED
4629 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4631 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4632 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4634 pthread_mutexattr_destroy(&mattr);
4635 #endif /* _WIN32 || ... */
4637 env->me_txns->mti_magic = MDB_MAGIC;
4638 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4639 env->me_txns->mti_txnid = 0;
4640 env->me_txns->mti_numreaders = 0;
4643 #ifdef MDB_USE_SYSV_SEM
4644 struct semid_ds buf;
4646 if (env->me_txns->mti_magic != MDB_MAGIC) {
4647 DPUTS("lock region has invalid magic");
4651 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4652 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4653 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4654 rc = MDB_VERSION_MISMATCH;
4658 if (rc && rc != EACCES && rc != EAGAIN) {
4662 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4663 if (!env->me_rmutex) goto fail_errno;
4664 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4665 if (!env->me_wmutex) goto fail_errno;
4666 #elif defined(MDB_USE_POSIX_SEM)
4667 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4668 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4669 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4670 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4671 #elif defined(MDB_USE_SYSV_SEM)
4672 semid = env->me_txns->mti_semid;
4674 /* check for read access */
4675 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4677 /* check for write access */
4678 if (semctl(semid, 0, IPC_SET, semu) < 0)
4682 #ifdef MDB_USE_SYSV_SEM
4683 env->me_rmutex->semid = semid;
4684 env->me_wmutex->semid = semid;
4685 env->me_rmutex->semnum = 0;
4686 env->me_wmutex->semnum = 1;
4687 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4688 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4699 /** The name of the lock file in the DB environment */
4700 #define LOCKNAME "/lock.mdb"
4701 /** The name of the data file in the DB environment */
4702 #define DATANAME "/data.mdb"
4703 /** The suffix of the lock file when no subdir is used */
4704 #define LOCKSUFF "-lock"
4705 /** Only a subset of the @ref mdb_env flags can be changed
4706 * at runtime. Changing other flags requires closing the
4707 * environment and re-opening it with the new flags.
4709 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4710 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4711 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4713 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4714 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4718 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4720 int oflags, rc, len, excl = -1;
4721 char *lpath, *dpath;
4723 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4727 if (flags & MDB_NOSUBDIR) {
4728 rc = len + sizeof(LOCKSUFF) + len + 1;
4730 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4735 if (flags & MDB_NOSUBDIR) {
4736 dpath = lpath + len + sizeof(LOCKSUFF);
4737 sprintf(lpath, "%s" LOCKSUFF, path);
4738 strcpy(dpath, path);
4740 dpath = lpath + len + sizeof(LOCKNAME);
4741 sprintf(lpath, "%s" LOCKNAME, path);
4742 sprintf(dpath, "%s" DATANAME, path);
4746 flags |= env->me_flags;
4747 if (flags & MDB_RDONLY) {
4748 /* silently ignore WRITEMAP when we're only getting read access */
4749 flags &= ~MDB_WRITEMAP;
4751 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4752 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4755 env->me_flags = flags |= MDB_ENV_ACTIVE;
4759 env->me_path = strdup(path);
4760 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4761 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4762 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4763 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4767 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4769 /* For RDONLY, get lockfile after we know datafile exists */
4770 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4771 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4777 if (F_ISSET(flags, MDB_RDONLY)) {
4778 oflags = GENERIC_READ;
4779 len = OPEN_EXISTING;
4781 oflags = GENERIC_READ|GENERIC_WRITE;
4784 mode = FILE_ATTRIBUTE_NORMAL;
4785 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4786 NULL, len, mode, NULL);
4788 if (F_ISSET(flags, MDB_RDONLY))
4791 oflags = O_RDWR | O_CREAT;
4793 env->me_fd = open(dpath, oflags, mode);
4795 if (env->me_fd == INVALID_HANDLE_VALUE) {
4800 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4801 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4806 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4807 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4808 env->me_mfd = env->me_fd;
4810 /* Synchronous fd for meta writes. Needed even with
4811 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4814 len = OPEN_EXISTING;
4815 env->me_mfd = CreateFile(dpath, oflags,
4816 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4817 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4820 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4822 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4827 DPRINTF(("opened dbenv %p", (void *) env));
4829 rc = mdb_env_share_locks(env, &excl);
4833 if (!(flags & MDB_RDONLY)) {
4835 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4836 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4837 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4838 (txn = calloc(1, size)))
4840 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4841 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4842 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4843 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4845 txn->mt_dbxs = env->me_dbxs;
4855 mdb_env_close0(env, excl);
4861 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4863 mdb_env_close0(MDB_env *env, int excl)
4867 if (!(env->me_flags & MDB_ENV_ACTIVE))
4870 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4872 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4873 free(env->me_dbxs[i].md_name.mv_data);
4878 free(env->me_dbiseqs);
4879 free(env->me_dbflags);
4881 free(env->me_dirty_list);
4883 mdb_midl_free(env->me_free_pgs);
4885 if (env->me_flags & MDB_ENV_TXKEY) {
4886 pthread_key_delete(env->me_txkey);
4888 /* Delete our key from the global list */
4889 for (i=0; i<mdb_tls_nkeys; i++)
4890 if (mdb_tls_keys[i] == env->me_txkey) {
4891 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4899 munmap(env->me_map, env->me_mapsize);
4901 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4902 (void) close(env->me_mfd);
4903 if (env->me_fd != INVALID_HANDLE_VALUE)
4904 (void) close(env->me_fd);
4906 MDB_PID_T pid = env->me_pid;
4907 /* Clearing readers is done in this function because
4908 * me_txkey with its destructor must be disabled first.
4910 * We skip the the reader mutex, so we touch only
4911 * data owned by this process (me_close_readers and
4912 * our readers), and clear each reader atomically.
4914 for (i = env->me_close_readers; --i >= 0; )
4915 if (env->me_txns->mti_readers[i].mr_pid == pid)
4916 env->me_txns->mti_readers[i].mr_pid = 0;
4918 if (env->me_rmutex) {
4919 CloseHandle(env->me_rmutex);
4920 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4922 /* Windows automatically destroys the mutexes when
4923 * the last handle closes.
4925 #elif defined(MDB_USE_POSIX_SEM)
4926 if (env->me_rmutex != SEM_FAILED) {
4927 sem_close(env->me_rmutex);
4928 if (env->me_wmutex != SEM_FAILED)
4929 sem_close(env->me_wmutex);
4930 /* If we have the filelock: If we are the
4931 * only remaining user, clean up semaphores.
4934 mdb_env_excl_lock(env, &excl);
4936 sem_unlink(env->me_txns->mti_rmname);
4937 sem_unlink(env->me_txns->mti_wmname);
4940 #elif defined(MDB_USE_SYSV_SEM)
4941 if (env->me_rmutex->semid != -1) {
4942 /* If we have the filelock: If we are the
4943 * only remaining user, clean up semaphores.
4946 mdb_env_excl_lock(env, &excl);
4948 semctl(env->me_rmutex->semid, 0, IPC_RMID);
4951 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4953 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4956 /* Unlock the lockfile. Windows would have unlocked it
4957 * after closing anyway, but not necessarily at once.
4959 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4962 (void) close(env->me_lfd);
4965 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4969 mdb_env_close(MDB_env *env)
4976 VGMEMP_DESTROY(env);
4977 while ((dp = env->me_dpages) != NULL) {
4978 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4979 env->me_dpages = dp->mp_next;
4983 mdb_env_close0(env, 0);
4987 /** Compare two items pointing at aligned size_t's */
4989 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4991 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4992 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4995 /** Compare two items pointing at aligned unsigned int's.
4997 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4998 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5001 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5003 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5004 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5007 /** Compare two items pointing at unsigned ints of unknown alignment.
5008 * Nodes and keys are guaranteed to be 2-byte aligned.
5011 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5013 #if BYTE_ORDER == LITTLE_ENDIAN
5014 unsigned short *u, *c;
5017 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5018 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5021 } while(!x && u > (unsigned short *)a->mv_data);
5024 unsigned short *u, *c, *end;
5027 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5028 u = (unsigned short *)a->mv_data;
5029 c = (unsigned short *)b->mv_data;
5032 } while(!x && u < end);
5037 /** Compare two items lexically */
5039 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5046 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5052 diff = memcmp(a->mv_data, b->mv_data, len);
5053 return diff ? diff : len_diff<0 ? -1 : len_diff;
5056 /** Compare two items in reverse byte order */
5058 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5060 const unsigned char *p1, *p2, *p1_lim;
5064 p1_lim = (const unsigned char *)a->mv_data;
5065 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5066 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5068 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5074 while (p1 > p1_lim) {
5075 diff = *--p1 - *--p2;
5079 return len_diff<0 ? -1 : len_diff;
5082 /** Search for key within a page, using binary search.
5083 * Returns the smallest entry larger or equal to the key.
5084 * If exactp is non-null, stores whether the found entry was an exact match
5085 * in *exactp (1 or 0).
5086 * Updates the cursor index with the index of the found entry.
5087 * If no entry larger or equal to the key is found, returns NULL.
5090 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5092 unsigned int i = 0, nkeys;
5095 MDB_page *mp = mc->mc_pg[mc->mc_top];
5096 MDB_node *node = NULL;
5101 nkeys = NUMKEYS(mp);
5103 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5104 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5107 low = IS_LEAF(mp) ? 0 : 1;
5109 cmp = mc->mc_dbx->md_cmp;
5111 /* Branch pages have no data, so if using integer keys,
5112 * alignment is guaranteed. Use faster mdb_cmp_int.
5114 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5115 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5122 nodekey.mv_size = mc->mc_db->md_pad;
5123 node = NODEPTR(mp, 0); /* fake */
5124 while (low <= high) {
5125 i = (low + high) >> 1;
5126 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5127 rc = cmp(key, &nodekey);
5128 DPRINTF(("found leaf index %u [%s], rc = %i",
5129 i, DKEY(&nodekey), rc));
5138 while (low <= high) {
5139 i = (low + high) >> 1;
5141 node = NODEPTR(mp, i);
5142 nodekey.mv_size = NODEKSZ(node);
5143 nodekey.mv_data = NODEKEY(node);
5145 rc = cmp(key, &nodekey);
5148 DPRINTF(("found leaf index %u [%s], rc = %i",
5149 i, DKEY(&nodekey), rc));
5151 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5152 i, DKEY(&nodekey), NODEPGNO(node), rc));
5163 if (rc > 0) { /* Found entry is less than the key. */
5164 i++; /* Skip to get the smallest entry larger than key. */
5166 node = NODEPTR(mp, i);
5169 *exactp = (rc == 0 && nkeys > 0);
5170 /* store the key index */
5171 mc->mc_ki[mc->mc_top] = i;
5173 /* There is no entry larger or equal to the key. */
5176 /* nodeptr is fake for LEAF2 */
5182 mdb_cursor_adjust(MDB_cursor *mc, func)
5186 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5187 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5194 /** Pop a page off the top of the cursor's stack. */
5196 mdb_cursor_pop(MDB_cursor *mc)
5199 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5200 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5208 /** Push a page onto the top of the cursor's stack. */
5210 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5212 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5213 DDBI(mc), (void *) mc));
5215 if (mc->mc_snum >= CURSOR_STACK) {
5216 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5217 return MDB_CURSOR_FULL;
5220 mc->mc_top = mc->mc_snum++;
5221 mc->mc_pg[mc->mc_top] = mp;
5222 mc->mc_ki[mc->mc_top] = 0;
5227 /** Find the address of the page corresponding to a given page number.
5228 * @param[in] txn the transaction for this access.
5229 * @param[in] pgno the page number for the page to retrieve.
5230 * @param[out] ret address of a pointer where the page's address will be stored.
5231 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5232 * @return 0 on success, non-zero on failure.
5235 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5237 MDB_env *env = txn->mt_env;
5241 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5245 MDB_ID2L dl = tx2->mt_u.dirty_list;
5247 /* Spilled pages were dirtied in this txn and flushed
5248 * because the dirty list got full. Bring this page
5249 * back in from the map (but don't unspill it here,
5250 * leave that unless page_touch happens again).
5252 if (tx2->mt_spill_pgs) {
5253 MDB_ID pn = pgno << 1;
5254 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5255 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5256 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5261 unsigned x = mdb_mid2l_search(dl, pgno);
5262 if (x <= dl[0].mid && dl[x].mid == pgno) {
5268 } while ((tx2 = tx2->mt_parent) != NULL);
5271 if (pgno < txn->mt_next_pgno) {
5273 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5275 DPRINTF(("page %"Z"u not found", pgno));
5276 txn->mt_flags |= MDB_TXN_ERROR;
5277 return MDB_PAGE_NOTFOUND;
5287 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5288 * The cursor is at the root page, set up the rest of it.
5291 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5293 MDB_page *mp = mc->mc_pg[mc->mc_top];
5297 while (IS_BRANCH(mp)) {
5301 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5302 mdb_cassert(mc, NUMKEYS(mp) > 1);
5303 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5305 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5307 if (flags & MDB_PS_LAST)
5308 i = NUMKEYS(mp) - 1;
5311 node = mdb_node_search(mc, key, &exact);
5313 i = NUMKEYS(mp) - 1;
5315 i = mc->mc_ki[mc->mc_top];
5317 mdb_cassert(mc, i > 0);
5321 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5324 mdb_cassert(mc, i < NUMKEYS(mp));
5325 node = NODEPTR(mp, i);
5327 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5330 mc->mc_ki[mc->mc_top] = i;
5331 if ((rc = mdb_cursor_push(mc, mp)))
5334 if (flags & MDB_PS_MODIFY) {
5335 if ((rc = mdb_page_touch(mc)) != 0)
5337 mp = mc->mc_pg[mc->mc_top];
5342 DPRINTF(("internal error, index points to a %02X page!?",
5344 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5345 return MDB_CORRUPTED;
5348 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5349 key ? DKEY(key) : "null"));
5350 mc->mc_flags |= C_INITIALIZED;
5351 mc->mc_flags &= ~C_EOF;
5356 /** Search for the lowest key under the current branch page.
5357 * This just bypasses a NUMKEYS check in the current page
5358 * before calling mdb_page_search_root(), because the callers
5359 * are all in situations where the current page is known to
5363 mdb_page_search_lowest(MDB_cursor *mc)
5365 MDB_page *mp = mc->mc_pg[mc->mc_top];
5366 MDB_node *node = NODEPTR(mp, 0);
5369 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5372 mc->mc_ki[mc->mc_top] = 0;
5373 if ((rc = mdb_cursor_push(mc, mp)))
5375 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5378 /** Search for the page a given key should be in.
5379 * Push it and its parent pages on the cursor stack.
5380 * @param[in,out] mc the cursor for this operation.
5381 * @param[in] key the key to search for, or NULL for first/last page.
5382 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5383 * are touched (updated with new page numbers).
5384 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5385 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5386 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5387 * @return 0 on success, non-zero on failure.
5390 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5395 /* Make sure the txn is still viable, then find the root from
5396 * the txn's db table and set it as the root of the cursor's stack.
5398 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5399 DPUTS("transaction has failed, must abort");
5402 /* Make sure we're using an up-to-date root */
5403 if (*mc->mc_dbflag & DB_STALE) {
5405 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5407 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5408 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5415 MDB_node *leaf = mdb_node_search(&mc2,
5416 &mc->mc_dbx->md_name, &exact);
5418 return MDB_NOTFOUND;
5419 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5420 return MDB_INCOMPATIBLE; /* not a named DB */
5421 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5424 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5426 /* The txn may not know this DBI, or another process may
5427 * have dropped and recreated the DB with other flags.
5429 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5430 return MDB_INCOMPATIBLE;
5431 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5433 *mc->mc_dbflag &= ~DB_STALE;
5435 root = mc->mc_db->md_root;
5437 if (root == P_INVALID) { /* Tree is empty. */
5438 DPUTS("tree is empty");
5439 return MDB_NOTFOUND;
5443 mdb_cassert(mc, root > 1);
5444 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5445 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5451 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5452 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5454 if (flags & MDB_PS_MODIFY) {
5455 if ((rc = mdb_page_touch(mc)))
5459 if (flags & MDB_PS_ROOTONLY)
5462 return mdb_page_search_root(mc, key, flags);
5466 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5468 MDB_txn *txn = mc->mc_txn;
5469 pgno_t pg = mp->mp_pgno;
5470 unsigned x = 0, ovpages = mp->mp_pages;
5471 MDB_env *env = txn->mt_env;
5472 MDB_IDL sl = txn->mt_spill_pgs;
5473 MDB_ID pn = pg << 1;
5476 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5477 /* If the page is dirty or on the spill list we just acquired it,
5478 * so we should give it back to our current free list, if any.
5479 * Otherwise put it onto the list of pages we freed in this txn.
5481 * Won't create me_pghead: me_pglast must be inited along with it.
5482 * Unsupported in nested txns: They would need to hide the page
5483 * range in ancestor txns' dirty and spilled lists.
5485 if (env->me_pghead &&
5487 ((mp->mp_flags & P_DIRTY) ||
5488 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5492 MDB_ID2 *dl, ix, iy;
5493 rc = mdb_midl_need(&env->me_pghead, ovpages);
5496 if (!(mp->mp_flags & P_DIRTY)) {
5497 /* This page is no longer spilled */
5504 /* Remove from dirty list */
5505 dl = txn->mt_u.dirty_list;
5507 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5513 mdb_cassert(mc, x > 1);
5515 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5516 txn->mt_flags |= MDB_TXN_ERROR;
5517 return MDB_CORRUPTED;
5520 if (!(env->me_flags & MDB_WRITEMAP))
5521 mdb_dpage_free(env, mp);
5523 /* Insert in me_pghead */
5524 mop = env->me_pghead;
5525 j = mop[0] + ovpages;
5526 for (i = mop[0]; i && mop[i] < pg; i--)
5532 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5536 mc->mc_db->md_overflow_pages -= ovpages;
5540 /** Return the data associated with a given node.
5541 * @param[in] txn The transaction for this operation.
5542 * @param[in] leaf The node being read.
5543 * @param[out] data Updated to point to the node's data.
5544 * @return 0 on success, non-zero on failure.
5547 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5549 MDB_page *omp; /* overflow page */
5553 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5554 data->mv_size = NODEDSZ(leaf);
5555 data->mv_data = NODEDATA(leaf);
5559 /* Read overflow data.
5561 data->mv_size = NODEDSZ(leaf);
5562 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5563 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5564 DPRINTF(("read overflow page %"Z"u failed", pgno));
5567 data->mv_data = METADATA(omp);
5573 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5574 MDB_val *key, MDB_val *data)
5581 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5583 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5586 if (txn->mt_flags & MDB_TXN_ERROR)
5589 mdb_cursor_init(&mc, txn, dbi, &mx);
5590 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5593 /** Find a sibling for a page.
5594 * Replaces the page at the top of the cursor's stack with the
5595 * specified sibling, if one exists.
5596 * @param[in] mc The cursor for this operation.
5597 * @param[in] move_right Non-zero if the right sibling is requested,
5598 * otherwise the left sibling.
5599 * @return 0 on success, non-zero on failure.
5602 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5608 if (mc->mc_snum < 2) {
5609 return MDB_NOTFOUND; /* root has no siblings */
5613 DPRINTF(("parent page is page %"Z"u, index %u",
5614 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5616 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5617 : (mc->mc_ki[mc->mc_top] == 0)) {
5618 DPRINTF(("no more keys left, moving to %s sibling",
5619 move_right ? "right" : "left"));
5620 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5621 /* undo cursor_pop before returning */
5628 mc->mc_ki[mc->mc_top]++;
5630 mc->mc_ki[mc->mc_top]--;
5631 DPRINTF(("just moving to %s index key %u",
5632 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5634 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5636 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5637 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5638 /* mc will be inconsistent if caller does mc_snum++ as above */
5639 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5643 mdb_cursor_push(mc, mp);
5645 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5650 /** Move the cursor to the next data item. */
5652 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5658 if (mc->mc_flags & C_EOF) {
5659 return MDB_NOTFOUND;
5662 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5664 mp = mc->mc_pg[mc->mc_top];
5666 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5667 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5668 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5669 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5670 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5671 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5672 if (rc == MDB_SUCCESS)
5673 MDB_GET_KEY(leaf, key);
5678 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5679 if (op == MDB_NEXT_DUP)
5680 return MDB_NOTFOUND;
5684 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5685 mdb_dbg_pgno(mp), (void *) mc));
5686 if (mc->mc_flags & C_DEL)
5689 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5690 DPUTS("=====> move to next sibling page");
5691 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5692 mc->mc_flags |= C_EOF;
5695 mp = mc->mc_pg[mc->mc_top];
5696 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5698 mc->mc_ki[mc->mc_top]++;
5701 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5702 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5705 key->mv_size = mc->mc_db->md_pad;
5706 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5710 mdb_cassert(mc, IS_LEAF(mp));
5711 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5713 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5714 mdb_xcursor_init1(mc, leaf);
5717 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5720 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5721 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5722 if (rc != MDB_SUCCESS)
5727 MDB_GET_KEY(leaf, key);
5731 /** Move the cursor to the previous data item. */
5733 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5739 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5741 mp = mc->mc_pg[mc->mc_top];
5743 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5744 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5745 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5746 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5747 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5748 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5749 if (rc == MDB_SUCCESS) {
5750 MDB_GET_KEY(leaf, key);
5751 mc->mc_flags &= ~C_EOF;
5757 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5758 if (op == MDB_PREV_DUP)
5759 return MDB_NOTFOUND;
5763 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5764 mdb_dbg_pgno(mp), (void *) mc));
5766 if (mc->mc_ki[mc->mc_top] == 0) {
5767 DPUTS("=====> move to prev sibling page");
5768 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5771 mp = mc->mc_pg[mc->mc_top];
5772 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5773 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5775 mc->mc_ki[mc->mc_top]--;
5777 mc->mc_flags &= ~C_EOF;
5779 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5780 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5783 key->mv_size = mc->mc_db->md_pad;
5784 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5788 mdb_cassert(mc, IS_LEAF(mp));
5789 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5791 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5792 mdb_xcursor_init1(mc, leaf);
5795 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5798 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5799 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5800 if (rc != MDB_SUCCESS)
5805 MDB_GET_KEY(leaf, key);
5809 /** Set the cursor on a specific data item. */
5811 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5812 MDB_cursor_op op, int *exactp)
5816 MDB_node *leaf = NULL;
5819 if (key->mv_size == 0)
5820 return MDB_BAD_VALSIZE;
5823 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5825 /* See if we're already on the right page */
5826 if (mc->mc_flags & C_INITIALIZED) {
5829 mp = mc->mc_pg[mc->mc_top];
5831 mc->mc_ki[mc->mc_top] = 0;
5832 return MDB_NOTFOUND;
5834 if (mp->mp_flags & P_LEAF2) {
5835 nodekey.mv_size = mc->mc_db->md_pad;
5836 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5838 leaf = NODEPTR(mp, 0);
5839 MDB_GET_KEY2(leaf, nodekey);
5841 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5843 /* Probably happens rarely, but first node on the page
5844 * was the one we wanted.
5846 mc->mc_ki[mc->mc_top] = 0;
5853 unsigned int nkeys = NUMKEYS(mp);
5855 if (mp->mp_flags & P_LEAF2) {
5856 nodekey.mv_data = LEAF2KEY(mp,
5857 nkeys-1, nodekey.mv_size);
5859 leaf = NODEPTR(mp, nkeys-1);
5860 MDB_GET_KEY2(leaf, nodekey);
5862 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5864 /* last node was the one we wanted */
5865 mc->mc_ki[mc->mc_top] = nkeys-1;
5871 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5872 /* This is definitely the right page, skip search_page */
5873 if (mp->mp_flags & P_LEAF2) {
5874 nodekey.mv_data = LEAF2KEY(mp,
5875 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5877 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5878 MDB_GET_KEY2(leaf, nodekey);
5880 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5882 /* current node was the one we wanted */
5892 /* If any parents have right-sibs, search.
5893 * Otherwise, there's nothing further.
5895 for (i=0; i<mc->mc_top; i++)
5897 NUMKEYS(mc->mc_pg[i])-1)
5899 if (i == mc->mc_top) {
5900 /* There are no other pages */
5901 mc->mc_ki[mc->mc_top] = nkeys;
5902 return MDB_NOTFOUND;
5906 /* There are no other pages */
5907 mc->mc_ki[mc->mc_top] = 0;
5908 if (op == MDB_SET_RANGE && !exactp) {
5912 return MDB_NOTFOUND;
5916 rc = mdb_page_search(mc, key, 0);
5917 if (rc != MDB_SUCCESS)
5920 mp = mc->mc_pg[mc->mc_top];
5921 mdb_cassert(mc, IS_LEAF(mp));
5924 leaf = mdb_node_search(mc, key, exactp);
5925 if (exactp != NULL && !*exactp) {
5926 /* MDB_SET specified and not an exact match. */
5927 return MDB_NOTFOUND;
5931 DPUTS("===> inexact leaf not found, goto sibling");
5932 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5933 mc->mc_flags |= C_EOF;
5934 return rc; /* no entries matched */
5936 mp = mc->mc_pg[mc->mc_top];
5937 mdb_cassert(mc, IS_LEAF(mp));
5938 leaf = NODEPTR(mp, 0);
5942 mc->mc_flags |= C_INITIALIZED;
5943 mc->mc_flags &= ~C_EOF;
5946 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5947 key->mv_size = mc->mc_db->md_pad;
5948 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5953 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5954 mdb_xcursor_init1(mc, leaf);
5957 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5958 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5959 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5962 if (op == MDB_GET_BOTH) {
5968 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5969 if (rc != MDB_SUCCESS)
5972 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5975 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5977 dcmp = mc->mc_dbx->md_dcmp;
5978 #if UINT_MAX < SIZE_MAX
5979 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5980 dcmp = mdb_cmp_clong;
5982 rc = dcmp(data, &olddata);
5984 if (op == MDB_GET_BOTH || rc > 0)
5985 return MDB_NOTFOUND;
5992 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5993 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5998 /* The key already matches in all other cases */
5999 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6000 MDB_GET_KEY(leaf, key);
6001 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6006 /** Move the cursor to the first item in the database. */
6008 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6014 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6016 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6017 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6018 if (rc != MDB_SUCCESS)
6021 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6023 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6024 mc->mc_flags |= C_INITIALIZED;
6025 mc->mc_flags &= ~C_EOF;
6027 mc->mc_ki[mc->mc_top] = 0;
6029 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6030 key->mv_size = mc->mc_db->md_pad;
6031 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6036 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6037 mdb_xcursor_init1(mc, leaf);
6038 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6042 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6046 MDB_GET_KEY(leaf, key);
6050 /** Move the cursor to the last item in the database. */
6052 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6058 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6060 if (!(mc->mc_flags & C_EOF)) {
6062 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6063 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6064 if (rc != MDB_SUCCESS)
6067 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6070 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6071 mc->mc_flags |= C_INITIALIZED|C_EOF;
6072 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6074 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6075 key->mv_size = mc->mc_db->md_pad;
6076 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6081 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6082 mdb_xcursor_init1(mc, leaf);
6083 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6087 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6092 MDB_GET_KEY(leaf, key);
6097 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6102 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6107 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6111 case MDB_GET_CURRENT:
6112 if (!(mc->mc_flags & C_INITIALIZED)) {
6115 MDB_page *mp = mc->mc_pg[mc->mc_top];
6116 int nkeys = NUMKEYS(mp);
6117 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6118 mc->mc_ki[mc->mc_top] = nkeys;
6124 key->mv_size = mc->mc_db->md_pad;
6125 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6127 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6128 MDB_GET_KEY(leaf, key);
6130 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6131 if (mc->mc_flags & C_DEL)
6132 mdb_xcursor_init1(mc, leaf);
6133 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6135 rc = mdb_node_read(mc->mc_txn, leaf, data);
6142 case MDB_GET_BOTH_RANGE:
6147 if (mc->mc_xcursor == NULL) {
6148 rc = MDB_INCOMPATIBLE;
6158 rc = mdb_cursor_set(mc, key, data, op,
6159 op == MDB_SET_RANGE ? NULL : &exact);
6162 case MDB_GET_MULTIPLE:
6163 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6167 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6168 rc = MDB_INCOMPATIBLE;
6172 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6173 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6176 case MDB_NEXT_MULTIPLE:
6181 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6182 rc = MDB_INCOMPATIBLE;
6185 if (!(mc->mc_flags & C_INITIALIZED))
6186 rc = mdb_cursor_first(mc, key, data);
6188 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6189 if (rc == MDB_SUCCESS) {
6190 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6193 mx = &mc->mc_xcursor->mx_cursor;
6194 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6196 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6197 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6205 case MDB_NEXT_NODUP:
6206 if (!(mc->mc_flags & C_INITIALIZED))
6207 rc = mdb_cursor_first(mc, key, data);
6209 rc = mdb_cursor_next(mc, key, data, op);
6213 case MDB_PREV_NODUP:
6214 if (!(mc->mc_flags & C_INITIALIZED)) {
6215 rc = mdb_cursor_last(mc, key, data);
6218 mc->mc_flags |= C_INITIALIZED;
6219 mc->mc_ki[mc->mc_top]++;
6221 rc = mdb_cursor_prev(mc, key, data, op);
6224 rc = mdb_cursor_first(mc, key, data);
6227 mfunc = mdb_cursor_first;
6229 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6233 if (mc->mc_xcursor == NULL) {
6234 rc = MDB_INCOMPATIBLE;
6238 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6239 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6240 MDB_GET_KEY(leaf, key);
6241 rc = mdb_node_read(mc->mc_txn, leaf, data);
6245 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6249 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6252 rc = mdb_cursor_last(mc, key, data);
6255 mfunc = mdb_cursor_last;
6258 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6263 if (mc->mc_flags & C_DEL)
6264 mc->mc_flags ^= C_DEL;
6269 /** Touch all the pages in the cursor stack. Set mc_top.
6270 * Makes sure all the pages are writable, before attempting a write operation.
6271 * @param[in] mc The cursor to operate on.
6274 mdb_cursor_touch(MDB_cursor *mc)
6276 int rc = MDB_SUCCESS;
6278 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6281 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6283 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6284 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6287 *mc->mc_dbflag |= DB_DIRTY;
6292 rc = mdb_page_touch(mc);
6293 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6294 mc->mc_top = mc->mc_snum-1;
6299 /** Do not spill pages to disk if txn is getting full, may fail instead */
6300 #define MDB_NOSPILL 0x8000
6303 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6307 MDB_node *leaf = NULL;
6310 MDB_val xdata, *rdata, dkey, olddata;
6312 int do_sub = 0, insert_key, insert_data;
6313 unsigned int mcount = 0, dcount = 0, nospill;
6316 unsigned int nflags;
6319 if (mc == NULL || key == NULL)
6322 env = mc->mc_txn->mt_env;
6324 /* Check this first so counter will always be zero on any
6327 if (flags & MDB_MULTIPLE) {
6328 dcount = data[1].mv_size;
6329 data[1].mv_size = 0;
6330 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6331 return MDB_INCOMPATIBLE;
6334 nospill = flags & MDB_NOSPILL;
6335 flags &= ~MDB_NOSPILL;
6337 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6338 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6340 if (key->mv_size-1 >= ENV_MAXKEY(env))
6341 return MDB_BAD_VALSIZE;
6343 #if SIZE_MAX > MAXDATASIZE
6344 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6345 return MDB_BAD_VALSIZE;
6347 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6348 return MDB_BAD_VALSIZE;
6351 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6352 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6356 if (flags == MDB_CURRENT) {
6357 if (!(mc->mc_flags & C_INITIALIZED))
6360 } else if (mc->mc_db->md_root == P_INVALID) {
6361 /* new database, cursor has nothing to point to */
6364 mc->mc_flags &= ~C_INITIALIZED;
6369 if (flags & MDB_APPEND) {
6371 rc = mdb_cursor_last(mc, &k2, &d2);
6373 rc = mc->mc_dbx->md_cmp(key, &k2);
6376 mc->mc_ki[mc->mc_top]++;
6378 /* new key is <= last key */
6383 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6385 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6386 DPRINTF(("duplicate key [%s]", DKEY(key)));
6388 return MDB_KEYEXIST;
6390 if (rc && rc != MDB_NOTFOUND)
6394 if (mc->mc_flags & C_DEL)
6395 mc->mc_flags ^= C_DEL;
6397 /* Cursor is positioned, check for room in the dirty list */
6399 if (flags & MDB_MULTIPLE) {
6401 xdata.mv_size = data->mv_size * dcount;
6405 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6409 if (rc == MDB_NO_ROOT) {
6411 /* new database, write a root leaf page */
6412 DPUTS("allocating new root leaf page");
6413 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6416 mdb_cursor_push(mc, np);
6417 mc->mc_db->md_root = np->mp_pgno;
6418 mc->mc_db->md_depth++;
6419 *mc->mc_dbflag |= DB_DIRTY;
6420 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6422 np->mp_flags |= P_LEAF2;
6423 mc->mc_flags |= C_INITIALIZED;
6425 /* make sure all cursor pages are writable */
6426 rc2 = mdb_cursor_touch(mc);
6431 insert_key = insert_data = rc;
6433 /* The key does not exist */
6434 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6435 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6436 LEAFSIZE(key, data) > env->me_nodemax)
6438 /* Too big for a node, insert in sub-DB. Set up an empty
6439 * "old sub-page" for prep_subDB to expand to a full page.
6441 fp_flags = P_LEAF|P_DIRTY;
6443 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6444 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6445 olddata.mv_size = PAGEHDRSZ;
6449 /* there's only a key anyway, so this is a no-op */
6450 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6452 unsigned int ksize = mc->mc_db->md_pad;
6453 if (key->mv_size != ksize)
6454 return MDB_BAD_VALSIZE;
6455 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6456 memcpy(ptr, key->mv_data, ksize);
6458 /* if overwriting slot 0 of leaf, need to
6459 * update branch key if there is a parent page
6461 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6462 unsigned short top = mc->mc_top;
6464 /* slot 0 is always an empty key, find real slot */
6465 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6467 if (mc->mc_ki[mc->mc_top])
6468 rc2 = mdb_update_key(mc, key);
6479 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6480 olddata.mv_size = NODEDSZ(leaf);
6481 olddata.mv_data = NODEDATA(leaf);
6484 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6485 /* Prepare (sub-)page/sub-DB to accept the new item,
6486 * if needed. fp: old sub-page or a header faking
6487 * it. mp: new (sub-)page. offset: growth in page
6488 * size. xdata: node data with new page or DB.
6490 unsigned i, offset = 0;
6491 mp = fp = xdata.mv_data = env->me_pbuf;
6492 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6494 /* Was a single item before, must convert now */
6495 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6497 /* Just overwrite the current item */
6498 if (flags == MDB_CURRENT)
6500 dcmp = mc->mc_dbx->md_dcmp;
6501 #if UINT_MAX < SIZE_MAX
6502 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6503 dcmp = mdb_cmp_clong;
6505 /* does data match? */
6506 if (!dcmp(data, &olddata)) {
6507 if (flags & MDB_NODUPDATA)
6508 return MDB_KEYEXIST;
6513 /* Back up original data item */
6514 dkey.mv_size = olddata.mv_size;
6515 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6517 /* Make sub-page header for the dup items, with dummy body */
6518 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6519 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6520 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6521 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6522 fp->mp_flags |= P_LEAF2;
6523 fp->mp_pad = data->mv_size;
6524 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6526 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6527 (dkey.mv_size & 1) + (data->mv_size & 1);
6529 fp->mp_upper = xdata.mv_size - PAGEBASE;
6530 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6531 } else if (leaf->mn_flags & F_SUBDATA) {
6532 /* Data is on sub-DB, just store it */
6533 flags |= F_DUPDATA|F_SUBDATA;
6536 /* Data is on sub-page */
6537 fp = olddata.mv_data;
6540 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6541 offset = EVEN(NODESIZE + sizeof(indx_t) +
6545 offset = fp->mp_pad;
6546 if (SIZELEFT(fp) < offset) {
6547 offset *= 4; /* space for 4 more */
6550 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6552 fp->mp_flags |= P_DIRTY;
6553 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6554 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6558 xdata.mv_size = olddata.mv_size + offset;
6561 fp_flags = fp->mp_flags;
6562 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6563 /* Too big for a sub-page, convert to sub-DB */
6564 fp_flags &= ~P_SUBP;
6566 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6567 fp_flags |= P_LEAF2;
6568 dummy.md_pad = fp->mp_pad;
6569 dummy.md_flags = MDB_DUPFIXED;
6570 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6571 dummy.md_flags |= MDB_INTEGERKEY;
6577 dummy.md_branch_pages = 0;
6578 dummy.md_leaf_pages = 1;
6579 dummy.md_overflow_pages = 0;
6580 dummy.md_entries = NUMKEYS(fp);
6581 xdata.mv_size = sizeof(MDB_db);
6582 xdata.mv_data = &dummy;
6583 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6585 offset = env->me_psize - olddata.mv_size;
6586 flags |= F_DUPDATA|F_SUBDATA;
6587 dummy.md_root = mp->mp_pgno;
6590 mp->mp_flags = fp_flags | P_DIRTY;
6591 mp->mp_pad = fp->mp_pad;
6592 mp->mp_lower = fp->mp_lower;
6593 mp->mp_upper = fp->mp_upper + offset;
6594 if (fp_flags & P_LEAF2) {
6595 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6597 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6598 olddata.mv_size - fp->mp_upper - PAGEBASE);
6599 for (i=0; i<NUMKEYS(fp); i++)
6600 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6608 mdb_node_del(mc, 0);
6612 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6613 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6614 return MDB_INCOMPATIBLE;
6615 /* overflow page overwrites need special handling */
6616 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6619 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6621 memcpy(&pg, olddata.mv_data, sizeof(pg));
6622 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6624 ovpages = omp->mp_pages;
6626 /* Is the ov page large enough? */
6627 if (ovpages >= dpages) {
6628 if (!(omp->mp_flags & P_DIRTY) &&
6629 (level || (env->me_flags & MDB_WRITEMAP)))
6631 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6634 level = 0; /* dirty in this txn or clean */
6637 if (omp->mp_flags & P_DIRTY) {
6638 /* yes, overwrite it. Note in this case we don't
6639 * bother to try shrinking the page if the new data
6640 * is smaller than the overflow threshold.
6643 /* It is writable only in a parent txn */
6644 size_t sz = (size_t) env->me_psize * ovpages, off;
6645 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6651 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6652 mdb_cassert(mc, rc2 == 0);
6653 if (!(flags & MDB_RESERVE)) {
6654 /* Copy end of page, adjusting alignment so
6655 * compiler may copy words instead of bytes.
6657 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6658 memcpy((size_t *)((char *)np + off),
6659 (size_t *)((char *)omp + off), sz - off);
6662 memcpy(np, omp, sz); /* Copy beginning of page */
6665 SETDSZ(leaf, data->mv_size);
6666 if (F_ISSET(flags, MDB_RESERVE))
6667 data->mv_data = METADATA(omp);
6669 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6673 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6675 } else if (data->mv_size == olddata.mv_size) {
6676 /* same size, just replace it. Note that we could
6677 * also reuse this node if the new data is smaller,
6678 * but instead we opt to shrink the node in that case.
6680 if (F_ISSET(flags, MDB_RESERVE))
6681 data->mv_data = olddata.mv_data;
6682 else if (!(mc->mc_flags & C_SUB))
6683 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6685 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6690 mdb_node_del(mc, 0);
6696 nflags = flags & NODE_ADD_FLAGS;
6697 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6698 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6699 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6700 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6702 nflags |= MDB_SPLIT_REPLACE;
6703 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6705 /* There is room already in this leaf page. */
6706 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6707 if (rc == 0 && insert_key) {
6708 /* Adjust other cursors pointing to mp */
6709 MDB_cursor *m2, *m3;
6710 MDB_dbi dbi = mc->mc_dbi;
6711 unsigned i = mc->mc_top;
6712 MDB_page *mp = mc->mc_pg[i];
6714 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6715 if (mc->mc_flags & C_SUB)
6716 m3 = &m2->mc_xcursor->mx_cursor;
6719 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6720 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6727 if (rc == MDB_SUCCESS) {
6728 /* Now store the actual data in the child DB. Note that we're
6729 * storing the user data in the keys field, so there are strict
6730 * size limits on dupdata. The actual data fields of the child
6731 * DB are all zero size.
6739 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6740 if (flags & MDB_CURRENT) {
6741 xflags = MDB_CURRENT|MDB_NOSPILL;
6743 mdb_xcursor_init1(mc, leaf);
6744 xflags = (flags & MDB_NODUPDATA) ?
6745 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6747 /* converted, write the original data first */
6749 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6753 /* Adjust other cursors pointing to mp */
6755 unsigned i = mc->mc_top;
6756 MDB_page *mp = mc->mc_pg[i];
6758 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6759 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6760 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6761 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6762 mdb_xcursor_init1(m2, leaf);
6766 /* we've done our job */
6769 ecount = mc->mc_xcursor->mx_db.md_entries;
6770 if (flags & MDB_APPENDDUP)
6771 xflags |= MDB_APPEND;
6772 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6773 if (flags & F_SUBDATA) {
6774 void *db = NODEDATA(leaf);
6775 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6777 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6779 /* Increment count unless we just replaced an existing item. */
6781 mc->mc_db->md_entries++;
6783 /* Invalidate txn if we created an empty sub-DB */
6786 /* If we succeeded and the key didn't exist before,
6787 * make sure the cursor is marked valid.
6789 mc->mc_flags |= C_INITIALIZED;
6791 if (flags & MDB_MULTIPLE) {
6794 /* let caller know how many succeeded, if any */
6795 data[1].mv_size = mcount;
6796 if (mcount < dcount) {
6797 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6798 insert_key = insert_data = 0;
6805 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6808 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6813 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6819 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6820 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6822 if (!(mc->mc_flags & C_INITIALIZED))
6825 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6826 return MDB_NOTFOUND;
6828 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6831 rc = mdb_cursor_touch(mc);
6835 mp = mc->mc_pg[mc->mc_top];
6838 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6840 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6841 if (flags & MDB_NODUPDATA) {
6842 /* mdb_cursor_del0() will subtract the final entry */
6843 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6845 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6846 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6848 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6851 /* If sub-DB still has entries, we're done */
6852 if (mc->mc_xcursor->mx_db.md_entries) {
6853 if (leaf->mn_flags & F_SUBDATA) {
6854 /* update subDB info */
6855 void *db = NODEDATA(leaf);
6856 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6859 /* shrink fake page */
6860 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6861 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6862 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6863 /* fix other sub-DB cursors pointed at this fake page */
6864 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6865 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6866 if (m2->mc_pg[mc->mc_top] == mp &&
6867 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6868 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6871 mc->mc_db->md_entries--;
6872 mc->mc_flags |= C_DEL;
6875 /* otherwise fall thru and delete the sub-DB */
6878 if (leaf->mn_flags & F_SUBDATA) {
6879 /* add all the child DB's pages to the free list */
6880 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6885 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6886 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6887 rc = MDB_INCOMPATIBLE;
6891 /* add overflow pages to free list */
6892 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6896 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6897 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6898 (rc = mdb_ovpage_free(mc, omp)))
6903 return mdb_cursor_del0(mc);
6906 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6910 /** Allocate and initialize new pages for a database.
6911 * @param[in] mc a cursor on the database being added to.
6912 * @param[in] flags flags defining what type of page is being allocated.
6913 * @param[in] num the number of pages to allocate. This is usually 1,
6914 * unless allocating overflow pages for a large record.
6915 * @param[out] mp Address of a page, or NULL on failure.
6916 * @return 0 on success, non-zero on failure.
6919 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6924 if ((rc = mdb_page_alloc(mc, num, &np)))
6926 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6927 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6928 np->mp_flags = flags | P_DIRTY;
6929 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6930 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6933 mc->mc_db->md_branch_pages++;
6934 else if (IS_LEAF(np))
6935 mc->mc_db->md_leaf_pages++;
6936 else if (IS_OVERFLOW(np)) {
6937 mc->mc_db->md_overflow_pages += num;
6945 /** Calculate the size of a leaf node.
6946 * The size depends on the environment's page size; if a data item
6947 * is too large it will be put onto an overflow page and the node
6948 * size will only include the key and not the data. Sizes are always
6949 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6950 * of the #MDB_node headers.
6951 * @param[in] env The environment handle.
6952 * @param[in] key The key for the node.
6953 * @param[in] data The data for the node.
6954 * @return The number of bytes needed to store the node.
6957 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6961 sz = LEAFSIZE(key, data);
6962 if (sz > env->me_nodemax) {
6963 /* put on overflow page */
6964 sz -= data->mv_size - sizeof(pgno_t);
6967 return EVEN(sz + sizeof(indx_t));
6970 /** Calculate the size of a branch node.
6971 * The size should depend on the environment's page size but since
6972 * we currently don't support spilling large keys onto overflow
6973 * pages, it's simply the size of the #MDB_node header plus the
6974 * size of the key. Sizes are always rounded up to an even number
6975 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6976 * @param[in] env The environment handle.
6977 * @param[in] key The key for the node.
6978 * @return The number of bytes needed to store the node.
6981 mdb_branch_size(MDB_env *env, MDB_val *key)
6986 if (sz > env->me_nodemax) {
6987 /* put on overflow page */
6988 /* not implemented */
6989 /* sz -= key->size - sizeof(pgno_t); */
6992 return sz + sizeof(indx_t);
6995 /** Add a node to the page pointed to by the cursor.
6996 * @param[in] mc The cursor for this operation.
6997 * @param[in] indx The index on the page where the new node should be added.
6998 * @param[in] key The key for the new node.
6999 * @param[in] data The data for the new node, if any.
7000 * @param[in] pgno The page number, if adding a branch node.
7001 * @param[in] flags Flags for the node.
7002 * @return 0 on success, non-zero on failure. Possible errors are:
7004 * <li>ENOMEM - failed to allocate overflow pages for the node.
7005 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7006 * should never happen since all callers already calculate the
7007 * page's free space before calling this function.
7011 mdb_node_add(MDB_cursor *mc, indx_t indx,
7012 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7015 size_t node_size = NODESIZE;
7019 MDB_page *mp = mc->mc_pg[mc->mc_top];
7020 MDB_page *ofp = NULL; /* overflow page */
7024 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7026 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7027 IS_LEAF(mp) ? "leaf" : "branch",
7028 IS_SUBP(mp) ? "sub-" : "",
7029 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7030 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7033 /* Move higher keys up one slot. */
7034 int ksize = mc->mc_db->md_pad, dif;
7035 char *ptr = LEAF2KEY(mp, indx, ksize);
7036 dif = NUMKEYS(mp) - indx;
7038 memmove(ptr+ksize, ptr, dif*ksize);
7039 /* insert new key */
7040 memcpy(ptr, key->mv_data, ksize);
7042 /* Just using these for counting */
7043 mp->mp_lower += sizeof(indx_t);
7044 mp->mp_upper -= ksize - sizeof(indx_t);
7048 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7050 node_size += key->mv_size;
7052 mdb_cassert(mc, key && data);
7053 if (F_ISSET(flags, F_BIGDATA)) {
7054 /* Data already on overflow page. */
7055 node_size += sizeof(pgno_t);
7056 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7057 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7059 /* Put data on overflow page. */
7060 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7061 data->mv_size, node_size+data->mv_size));
7062 node_size = EVEN(node_size + sizeof(pgno_t));
7063 if ((ssize_t)node_size > room)
7065 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7067 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7071 node_size += data->mv_size;
7074 node_size = EVEN(node_size);
7075 if ((ssize_t)node_size > room)
7079 /* Move higher pointers up one slot. */
7080 for (i = NUMKEYS(mp); i > indx; i--)
7081 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7083 /* Adjust free space offsets. */
7084 ofs = mp->mp_upper - node_size;
7085 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7086 mp->mp_ptrs[indx] = ofs;
7088 mp->mp_lower += sizeof(indx_t);
7090 /* Write the node data. */
7091 node = NODEPTR(mp, indx);
7092 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7093 node->mn_flags = flags;
7095 SETDSZ(node,data->mv_size);
7100 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7103 ndata = NODEDATA(node);
7105 if (F_ISSET(flags, F_BIGDATA))
7106 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7107 else if (F_ISSET(flags, MDB_RESERVE))
7108 data->mv_data = ndata;
7110 memcpy(ndata, data->mv_data, data->mv_size);
7112 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7113 ndata = METADATA(ofp);
7114 if (F_ISSET(flags, MDB_RESERVE))
7115 data->mv_data = ndata;
7117 memcpy(ndata, data->mv_data, data->mv_size);
7124 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7125 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7126 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7127 DPRINTF(("node size = %"Z"u", node_size));
7128 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7129 return MDB_PAGE_FULL;
7132 /** Delete the specified node from a page.
7133 * @param[in] mc Cursor pointing to the node to delete.
7134 * @param[in] ksize The size of a node. Only used if the page is
7135 * part of a #MDB_DUPFIXED database.
7138 mdb_node_del(MDB_cursor *mc, int ksize)
7140 MDB_page *mp = mc->mc_pg[mc->mc_top];
7141 indx_t indx = mc->mc_ki[mc->mc_top];
7143 indx_t i, j, numkeys, ptr;
7147 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7148 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7149 numkeys = NUMKEYS(mp);
7150 mdb_cassert(mc, indx < numkeys);
7153 int x = numkeys - 1 - indx;
7154 base = LEAF2KEY(mp, indx, ksize);
7156 memmove(base, base + ksize, x * ksize);
7157 mp->mp_lower -= sizeof(indx_t);
7158 mp->mp_upper += ksize - sizeof(indx_t);
7162 node = NODEPTR(mp, indx);
7163 sz = NODESIZE + node->mn_ksize;
7165 if (F_ISSET(node->mn_flags, F_BIGDATA))
7166 sz += sizeof(pgno_t);
7168 sz += NODEDSZ(node);
7172 ptr = mp->mp_ptrs[indx];
7173 for (i = j = 0; i < numkeys; i++) {
7175 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7176 if (mp->mp_ptrs[i] < ptr)
7177 mp->mp_ptrs[j] += sz;
7182 base = (char *)mp + mp->mp_upper + PAGEBASE;
7183 memmove(base + sz, base, ptr - mp->mp_upper);
7185 mp->mp_lower -= sizeof(indx_t);
7189 /** Compact the main page after deleting a node on a subpage.
7190 * @param[in] mp The main page to operate on.
7191 * @param[in] indx The index of the subpage on the main page.
7194 mdb_node_shrink(MDB_page *mp, indx_t indx)
7199 indx_t delta, nsize, len, ptr;
7202 node = NODEPTR(mp, indx);
7203 sp = (MDB_page *)NODEDATA(node);
7204 delta = SIZELEFT(sp);
7205 nsize = NODEDSZ(node) - delta;
7207 /* Prepare to shift upward, set len = length(subpage part to shift) */
7211 return; /* do not make the node uneven-sized */
7213 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7214 for (i = NUMKEYS(sp); --i >= 0; )
7215 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7218 sp->mp_upper = sp->mp_lower;
7219 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7220 SETDSZ(node, nsize);
7222 /* Shift <lower nodes...initial part of subpage> upward */
7223 base = (char *)mp + mp->mp_upper + PAGEBASE;
7224 memmove(base + delta, base, (char *)sp + len - base);
7226 ptr = mp->mp_ptrs[indx];
7227 for (i = NUMKEYS(mp); --i >= 0; ) {
7228 if (mp->mp_ptrs[i] <= ptr)
7229 mp->mp_ptrs[i] += delta;
7231 mp->mp_upper += delta;
7234 /** Initial setup of a sorted-dups cursor.
7235 * Sorted duplicates are implemented as a sub-database for the given key.
7236 * The duplicate data items are actually keys of the sub-database.
7237 * Operations on the duplicate data items are performed using a sub-cursor
7238 * initialized when the sub-database is first accessed. This function does
7239 * the preliminary setup of the sub-cursor, filling in the fields that
7240 * depend only on the parent DB.
7241 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7244 mdb_xcursor_init0(MDB_cursor *mc)
7246 MDB_xcursor *mx = mc->mc_xcursor;
7248 mx->mx_cursor.mc_xcursor = NULL;
7249 mx->mx_cursor.mc_txn = mc->mc_txn;
7250 mx->mx_cursor.mc_db = &mx->mx_db;
7251 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7252 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7253 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7254 mx->mx_cursor.mc_snum = 0;
7255 mx->mx_cursor.mc_top = 0;
7256 mx->mx_cursor.mc_flags = C_SUB;
7257 mx->mx_dbx.md_name.mv_size = 0;
7258 mx->mx_dbx.md_name.mv_data = NULL;
7259 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7260 mx->mx_dbx.md_dcmp = NULL;
7261 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7264 /** Final setup of a sorted-dups cursor.
7265 * Sets up the fields that depend on the data from the main cursor.
7266 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7267 * @param[in] node The data containing the #MDB_db record for the
7268 * sorted-dup database.
7271 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7273 MDB_xcursor *mx = mc->mc_xcursor;
7275 if (node->mn_flags & F_SUBDATA) {
7276 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7277 mx->mx_cursor.mc_pg[0] = 0;
7278 mx->mx_cursor.mc_snum = 0;
7279 mx->mx_cursor.mc_top = 0;
7280 mx->mx_cursor.mc_flags = C_SUB;
7282 MDB_page *fp = NODEDATA(node);
7283 mx->mx_db.md_pad = 0;
7284 mx->mx_db.md_flags = 0;
7285 mx->mx_db.md_depth = 1;
7286 mx->mx_db.md_branch_pages = 0;
7287 mx->mx_db.md_leaf_pages = 1;
7288 mx->mx_db.md_overflow_pages = 0;
7289 mx->mx_db.md_entries = NUMKEYS(fp);
7290 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7291 mx->mx_cursor.mc_snum = 1;
7292 mx->mx_cursor.mc_top = 0;
7293 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7294 mx->mx_cursor.mc_pg[0] = fp;
7295 mx->mx_cursor.mc_ki[0] = 0;
7296 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7297 mx->mx_db.md_flags = MDB_DUPFIXED;
7298 mx->mx_db.md_pad = fp->mp_pad;
7299 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7300 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7303 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7304 mx->mx_db.md_root));
7305 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7306 #if UINT_MAX < SIZE_MAX
7307 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7308 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7312 /** Initialize a cursor for a given transaction and database. */
7314 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7317 mc->mc_backup = NULL;
7320 mc->mc_db = &txn->mt_dbs[dbi];
7321 mc->mc_dbx = &txn->mt_dbxs[dbi];
7322 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7328 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7329 mdb_tassert(txn, mx != NULL);
7330 mc->mc_xcursor = mx;
7331 mdb_xcursor_init0(mc);
7333 mc->mc_xcursor = NULL;
7335 if (*mc->mc_dbflag & DB_STALE) {
7336 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7341 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7344 size_t size = sizeof(MDB_cursor);
7346 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7349 if (txn->mt_flags & MDB_TXN_ERROR)
7352 /* Allow read access to the freelist */
7353 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7356 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7357 size += sizeof(MDB_xcursor);
7359 if ((mc = malloc(size)) != NULL) {
7360 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7361 if (txn->mt_cursors) {
7362 mc->mc_next = txn->mt_cursors[dbi];
7363 txn->mt_cursors[dbi] = mc;
7364 mc->mc_flags |= C_UNTRACK;
7376 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7378 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7381 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7384 if (txn->mt_flags & MDB_TXN_ERROR)
7387 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7391 /* Return the count of duplicate data items for the current key */
7393 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7397 if (mc == NULL || countp == NULL)
7400 if (mc->mc_xcursor == NULL)
7401 return MDB_INCOMPATIBLE;
7403 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7406 if (!(mc->mc_flags & C_INITIALIZED))
7409 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7410 return MDB_NOTFOUND;
7412 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7413 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7416 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7419 *countp = mc->mc_xcursor->mx_db.md_entries;
7425 mdb_cursor_close(MDB_cursor *mc)
7427 if (mc && !mc->mc_backup) {
7428 /* remove from txn, if tracked */
7429 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7430 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7431 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7433 *prev = mc->mc_next;
7440 mdb_cursor_txn(MDB_cursor *mc)
7442 if (!mc) return NULL;
7447 mdb_cursor_dbi(MDB_cursor *mc)
7452 /** Replace the key for a branch node with a new key.
7453 * @param[in] mc Cursor pointing to the node to operate on.
7454 * @param[in] key The new key to use.
7455 * @return 0 on success, non-zero on failure.
7458 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7464 int delta, ksize, oksize;
7465 indx_t ptr, i, numkeys, indx;
7468 indx = mc->mc_ki[mc->mc_top];
7469 mp = mc->mc_pg[mc->mc_top];
7470 node = NODEPTR(mp, indx);
7471 ptr = mp->mp_ptrs[indx];
7475 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7476 k2.mv_data = NODEKEY(node);
7477 k2.mv_size = node->mn_ksize;
7478 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7480 mdb_dkey(&k2, kbuf2),
7486 /* Sizes must be 2-byte aligned. */
7487 ksize = EVEN(key->mv_size);
7488 oksize = EVEN(node->mn_ksize);
7489 delta = ksize - oksize;
7491 /* Shift node contents if EVEN(key length) changed. */
7493 if (delta > 0 && SIZELEFT(mp) < delta) {
7495 /* not enough space left, do a delete and split */
7496 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7497 pgno = NODEPGNO(node);
7498 mdb_node_del(mc, 0);
7499 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7502 numkeys = NUMKEYS(mp);
7503 for (i = 0; i < numkeys; i++) {
7504 if (mp->mp_ptrs[i] <= ptr)
7505 mp->mp_ptrs[i] -= delta;
7508 base = (char *)mp + mp->mp_upper + PAGEBASE;
7509 len = ptr - mp->mp_upper + NODESIZE;
7510 memmove(base - delta, base, len);
7511 mp->mp_upper -= delta;
7513 node = NODEPTR(mp, indx);
7516 /* But even if no shift was needed, update ksize */
7517 if (node->mn_ksize != key->mv_size)
7518 node->mn_ksize = key->mv_size;
7521 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7527 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7529 /** Move a node from csrc to cdst.
7532 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7539 unsigned short flags;
7543 /* Mark src and dst as dirty. */
7544 if ((rc = mdb_page_touch(csrc)) ||
7545 (rc = mdb_page_touch(cdst)))
7548 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7549 key.mv_size = csrc->mc_db->md_pad;
7550 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7552 data.mv_data = NULL;
7556 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7557 mdb_cassert(csrc, !((size_t)srcnode & 1));
7558 srcpg = NODEPGNO(srcnode);
7559 flags = srcnode->mn_flags;
7560 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7561 unsigned int snum = csrc->mc_snum;
7563 /* must find the lowest key below src */
7564 rc = mdb_page_search_lowest(csrc);
7567 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7568 key.mv_size = csrc->mc_db->md_pad;
7569 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7571 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7572 key.mv_size = NODEKSZ(s2);
7573 key.mv_data = NODEKEY(s2);
7575 csrc->mc_snum = snum--;
7576 csrc->mc_top = snum;
7578 key.mv_size = NODEKSZ(srcnode);
7579 key.mv_data = NODEKEY(srcnode);
7581 data.mv_size = NODEDSZ(srcnode);
7582 data.mv_data = NODEDATA(srcnode);
7584 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7585 unsigned int snum = cdst->mc_snum;
7588 /* must find the lowest key below dst */
7589 mdb_cursor_copy(cdst, &mn);
7590 rc = mdb_page_search_lowest(&mn);
7593 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7594 bkey.mv_size = mn.mc_db->md_pad;
7595 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7597 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7598 bkey.mv_size = NODEKSZ(s2);
7599 bkey.mv_data = NODEKEY(s2);
7601 mn.mc_snum = snum--;
7604 rc = mdb_update_key(&mn, &bkey);
7609 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7610 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7611 csrc->mc_ki[csrc->mc_top],
7613 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7614 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7616 /* Add the node to the destination page.
7618 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7619 if (rc != MDB_SUCCESS)
7622 /* Delete the node from the source page.
7624 mdb_node_del(csrc, key.mv_size);
7627 /* Adjust other cursors pointing to mp */
7628 MDB_cursor *m2, *m3;
7629 MDB_dbi dbi = csrc->mc_dbi;
7630 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7632 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7633 if (csrc->mc_flags & C_SUB)
7634 m3 = &m2->mc_xcursor->mx_cursor;
7637 if (m3 == csrc) continue;
7638 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7639 csrc->mc_ki[csrc->mc_top]) {
7640 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7641 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7646 /* Update the parent separators.
7648 if (csrc->mc_ki[csrc->mc_top] == 0) {
7649 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7650 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7651 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7653 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7654 key.mv_size = NODEKSZ(srcnode);
7655 key.mv_data = NODEKEY(srcnode);
7657 DPRINTF(("update separator for source page %"Z"u to [%s]",
7658 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7659 mdb_cursor_copy(csrc, &mn);
7662 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7665 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7667 indx_t ix = csrc->mc_ki[csrc->mc_top];
7668 nullkey.mv_size = 0;
7669 csrc->mc_ki[csrc->mc_top] = 0;
7670 rc = mdb_update_key(csrc, &nullkey);
7671 csrc->mc_ki[csrc->mc_top] = ix;
7672 mdb_cassert(csrc, rc == MDB_SUCCESS);
7676 if (cdst->mc_ki[cdst->mc_top] == 0) {
7677 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7678 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7679 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7681 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7682 key.mv_size = NODEKSZ(srcnode);
7683 key.mv_data = NODEKEY(srcnode);
7685 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7686 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7687 mdb_cursor_copy(cdst, &mn);
7690 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7693 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7695 indx_t ix = cdst->mc_ki[cdst->mc_top];
7696 nullkey.mv_size = 0;
7697 cdst->mc_ki[cdst->mc_top] = 0;
7698 rc = mdb_update_key(cdst, &nullkey);
7699 cdst->mc_ki[cdst->mc_top] = ix;
7700 mdb_cassert(cdst, rc == MDB_SUCCESS);
7707 /** Merge one page into another.
7708 * The nodes from the page pointed to by \b csrc will
7709 * be copied to the page pointed to by \b cdst and then
7710 * the \b csrc page will be freed.
7711 * @param[in] csrc Cursor pointing to the source page.
7712 * @param[in] cdst Cursor pointing to the destination page.
7713 * @return 0 on success, non-zero on failure.
7716 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7718 MDB_page *psrc, *pdst;
7725 psrc = csrc->mc_pg[csrc->mc_top];
7726 pdst = cdst->mc_pg[cdst->mc_top];
7728 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7730 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7731 mdb_cassert(csrc, cdst->mc_snum > 1);
7733 /* Mark dst as dirty. */
7734 if ((rc = mdb_page_touch(cdst)))
7737 /* Move all nodes from src to dst.
7739 j = nkeys = NUMKEYS(pdst);
7740 if (IS_LEAF2(psrc)) {
7741 key.mv_size = csrc->mc_db->md_pad;
7742 key.mv_data = METADATA(psrc);
7743 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7744 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7745 if (rc != MDB_SUCCESS)
7747 key.mv_data = (char *)key.mv_data + key.mv_size;
7750 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7751 srcnode = NODEPTR(psrc, i);
7752 if (i == 0 && IS_BRANCH(psrc)) {
7755 mdb_cursor_copy(csrc, &mn);
7756 /* must find the lowest key below src */
7757 rc = mdb_page_search_lowest(&mn);
7760 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7761 key.mv_size = mn.mc_db->md_pad;
7762 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7764 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7765 key.mv_size = NODEKSZ(s2);
7766 key.mv_data = NODEKEY(s2);
7769 key.mv_size = srcnode->mn_ksize;
7770 key.mv_data = NODEKEY(srcnode);
7773 data.mv_size = NODEDSZ(srcnode);
7774 data.mv_data = NODEDATA(srcnode);
7775 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7776 if (rc != MDB_SUCCESS)
7781 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7782 pdst->mp_pgno, NUMKEYS(pdst),
7783 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7785 /* Unlink the src page from parent and add to free list.
7788 mdb_node_del(csrc, 0);
7789 if (csrc->mc_ki[csrc->mc_top] == 0) {
7791 rc = mdb_update_key(csrc, &key);
7799 psrc = csrc->mc_pg[csrc->mc_top];
7800 /* If not operating on FreeDB, allow this page to be reused
7801 * in this txn. Otherwise just add to free list.
7803 rc = mdb_page_loose(csrc, psrc);
7807 csrc->mc_db->md_leaf_pages--;
7809 csrc->mc_db->md_branch_pages--;
7811 /* Adjust other cursors pointing to mp */
7812 MDB_cursor *m2, *m3;
7813 MDB_dbi dbi = csrc->mc_dbi;
7815 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7816 if (csrc->mc_flags & C_SUB)
7817 m3 = &m2->mc_xcursor->mx_cursor;
7820 if (m3 == csrc) continue;
7821 if (m3->mc_snum < csrc->mc_snum) continue;
7822 if (m3->mc_pg[csrc->mc_top] == psrc) {
7823 m3->mc_pg[csrc->mc_top] = pdst;
7824 m3->mc_ki[csrc->mc_top] += nkeys;
7829 unsigned int snum = cdst->mc_snum;
7830 uint16_t depth = cdst->mc_db->md_depth;
7831 mdb_cursor_pop(cdst);
7832 rc = mdb_rebalance(cdst);
7833 /* Did the tree shrink? */
7834 if (depth > cdst->mc_db->md_depth)
7836 cdst->mc_snum = snum;
7837 cdst->mc_top = snum-1;
7842 /** Copy the contents of a cursor.
7843 * @param[in] csrc The cursor to copy from.
7844 * @param[out] cdst The cursor to copy to.
7847 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7851 cdst->mc_txn = csrc->mc_txn;
7852 cdst->mc_dbi = csrc->mc_dbi;
7853 cdst->mc_db = csrc->mc_db;
7854 cdst->mc_dbx = csrc->mc_dbx;
7855 cdst->mc_snum = csrc->mc_snum;
7856 cdst->mc_top = csrc->mc_top;
7857 cdst->mc_flags = csrc->mc_flags;
7859 for (i=0; i<csrc->mc_snum; i++) {
7860 cdst->mc_pg[i] = csrc->mc_pg[i];
7861 cdst->mc_ki[i] = csrc->mc_ki[i];
7865 /** Rebalance the tree after a delete operation.
7866 * @param[in] mc Cursor pointing to the page where rebalancing
7868 * @return 0 on success, non-zero on failure.
7871 mdb_rebalance(MDB_cursor *mc)
7875 unsigned int ptop, minkeys;
7879 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7880 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7881 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7882 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7883 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7885 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7886 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7887 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7888 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7892 if (mc->mc_snum < 2) {
7893 MDB_page *mp = mc->mc_pg[0];
7895 DPUTS("Can't rebalance a subpage, ignoring");
7898 if (NUMKEYS(mp) == 0) {
7899 DPUTS("tree is completely empty");
7900 mc->mc_db->md_root = P_INVALID;
7901 mc->mc_db->md_depth = 0;
7902 mc->mc_db->md_leaf_pages = 0;
7903 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7906 /* Adjust cursors pointing to mp */
7909 mc->mc_flags &= ~C_INITIALIZED;
7911 MDB_cursor *m2, *m3;
7912 MDB_dbi dbi = mc->mc_dbi;
7914 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7915 if (mc->mc_flags & C_SUB)
7916 m3 = &m2->mc_xcursor->mx_cursor;
7919 if (m3->mc_snum < mc->mc_snum) continue;
7920 if (m3->mc_pg[0] == mp) {
7923 m3->mc_flags &= ~C_INITIALIZED;
7927 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7929 DPUTS("collapsing root page!");
7930 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7933 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7934 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7937 mc->mc_db->md_depth--;
7938 mc->mc_db->md_branch_pages--;
7939 mc->mc_ki[0] = mc->mc_ki[1];
7940 for (i = 1; i<mc->mc_db->md_depth; i++) {
7941 mc->mc_pg[i] = mc->mc_pg[i+1];
7942 mc->mc_ki[i] = mc->mc_ki[i+1];
7945 /* Adjust other cursors pointing to mp */
7946 MDB_cursor *m2, *m3;
7947 MDB_dbi dbi = mc->mc_dbi;
7949 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7950 if (mc->mc_flags & C_SUB)
7951 m3 = &m2->mc_xcursor->mx_cursor;
7954 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7955 if (m3->mc_pg[0] == mp) {
7956 for (i=0; i<m3->mc_snum; i++) {
7957 m3->mc_pg[i] = m3->mc_pg[i+1];
7958 m3->mc_ki[i] = m3->mc_ki[i+1];
7966 DPUTS("root page doesn't need rebalancing");
7970 /* The parent (branch page) must have at least 2 pointers,
7971 * otherwise the tree is invalid.
7973 ptop = mc->mc_top-1;
7974 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7976 /* Leaf page fill factor is below the threshold.
7977 * Try to move keys from left or right neighbor, or
7978 * merge with a neighbor page.
7983 mdb_cursor_copy(mc, &mn);
7984 mn.mc_xcursor = NULL;
7986 oldki = mc->mc_ki[mc->mc_top];
7987 if (mc->mc_ki[ptop] == 0) {
7988 /* We're the leftmost leaf in our parent.
7990 DPUTS("reading right neighbor");
7992 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7993 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7996 mn.mc_ki[mn.mc_top] = 0;
7997 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7999 /* There is at least one neighbor to the left.
8001 DPUTS("reading left neighbor");
8003 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8004 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8007 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8008 mc->mc_ki[mc->mc_top] = 0;
8011 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8012 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8013 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8015 /* If the neighbor page is above threshold and has enough keys,
8016 * move one key from it. Otherwise we should try to merge them.
8017 * (A branch page must never have less than 2 keys.)
8019 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
8020 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8021 rc = mdb_node_move(&mn, mc);
8022 if (mc->mc_ki[ptop]) {
8026 if (mc->mc_ki[ptop] == 0) {
8027 rc = mdb_page_merge(&mn, mc);
8030 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8031 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8032 /* We want mdb_rebalance to find mn when doing fixups */
8033 if (mc->mc_flags & C_SUB) {
8034 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8035 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8036 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8038 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8039 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8041 rc = mdb_page_merge(mc, &mn);
8042 if (mc->mc_flags & C_SUB)
8043 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8045 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8046 mdb_cursor_copy(&mn, mc);
8048 mc->mc_flags &= ~C_EOF;
8050 mc->mc_ki[mc->mc_top] = oldki;
8054 /** Complete a delete operation started by #mdb_cursor_del(). */
8056 mdb_cursor_del0(MDB_cursor *mc)
8063 ki = mc->mc_ki[mc->mc_top];
8064 mdb_node_del(mc, mc->mc_db->md_pad);
8065 mc->mc_db->md_entries--;
8066 rc = mdb_rebalance(mc);
8068 if (rc == MDB_SUCCESS) {
8069 MDB_cursor *m2, *m3;
8070 MDB_dbi dbi = mc->mc_dbi;
8072 /* DB is totally empty now, just bail out.
8073 * Other cursors adjustments were already done
8074 * by mdb_rebalance and aren't needed here.
8079 mp = mc->mc_pg[mc->mc_top];
8080 nkeys = NUMKEYS(mp);
8082 /* if mc points past last node in page, find next sibling */
8083 if (mc->mc_ki[mc->mc_top] >= nkeys) {
8084 rc = mdb_cursor_sibling(mc, 1);
8085 if (rc == MDB_NOTFOUND) {
8086 mc->mc_flags |= C_EOF;
8091 /* Adjust other cursors pointing to mp */
8092 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8093 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8094 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8096 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8098 if (m3->mc_pg[mc->mc_top] == mp) {
8099 if (m3->mc_ki[mc->mc_top] >= ki) {
8100 m3->mc_flags |= C_DEL;
8101 if (m3->mc_ki[mc->mc_top] > ki)
8102 m3->mc_ki[mc->mc_top]--;
8103 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8104 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8106 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8107 rc = mdb_cursor_sibling(m3, 1);
8108 if (rc == MDB_NOTFOUND) {
8109 m3->mc_flags |= C_EOF;
8115 mc->mc_flags |= C_DEL;
8119 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8124 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8125 MDB_val *key, MDB_val *data)
8127 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8130 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
8131 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8133 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8134 /* must ignore any data */
8138 return mdb_del0(txn, dbi, key, data, 0);
8142 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8143 MDB_val *key, MDB_val *data, unsigned flags)
8148 MDB_val rdata, *xdata;
8152 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8154 mdb_cursor_init(&mc, txn, dbi, &mx);
8163 flags |= MDB_NODUPDATA;
8165 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8167 /* let mdb_page_split know about this cursor if needed:
8168 * delete will trigger a rebalance; if it needs to move
8169 * a node from one page to another, it will have to
8170 * update the parent's separator key(s). If the new sepkey
8171 * is larger than the current one, the parent page may
8172 * run out of space, triggering a split. We need this
8173 * cursor to be consistent until the end of the rebalance.
8175 mc.mc_flags |= C_UNTRACK;
8176 mc.mc_next = txn->mt_cursors[dbi];
8177 txn->mt_cursors[dbi] = &mc;
8178 rc = mdb_cursor_del(&mc, flags);
8179 txn->mt_cursors[dbi] = mc.mc_next;
8184 /** Split a page and insert a new node.
8185 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8186 * The cursor will be updated to point to the actual page and index where
8187 * the node got inserted after the split.
8188 * @param[in] newkey The key for the newly inserted node.
8189 * @param[in] newdata The data for the newly inserted node.
8190 * @param[in] newpgno The page number, if the new node is a branch node.
8191 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8192 * @return 0 on success, non-zero on failure.
8195 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8196 unsigned int nflags)
8199 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8202 int i, j, split_indx, nkeys, pmax;
8203 MDB_env *env = mc->mc_txn->mt_env;
8205 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8206 MDB_page *copy = NULL;
8207 MDB_page *mp, *rp, *pp;
8212 mp = mc->mc_pg[mc->mc_top];
8213 newindx = mc->mc_ki[mc->mc_top];
8214 nkeys = NUMKEYS(mp);
8216 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8217 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8218 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8220 /* Create a right sibling. */
8221 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8223 rp->mp_pad = mp->mp_pad;
8224 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8226 if (mc->mc_snum < 2) {
8227 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8229 /* shift current top to make room for new parent */
8230 mc->mc_pg[1] = mc->mc_pg[0];
8231 mc->mc_ki[1] = mc->mc_ki[0];
8234 mc->mc_db->md_root = pp->mp_pgno;
8235 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8236 mc->mc_db->md_depth++;
8239 /* Add left (implicit) pointer. */
8240 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8241 /* undo the pre-push */
8242 mc->mc_pg[0] = mc->mc_pg[1];
8243 mc->mc_ki[0] = mc->mc_ki[1];
8244 mc->mc_db->md_root = mp->mp_pgno;
8245 mc->mc_db->md_depth--;
8252 ptop = mc->mc_top-1;
8253 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8256 mc->mc_flags |= C_SPLITTING;
8257 mdb_cursor_copy(mc, &mn);
8258 mn.mc_pg[mn.mc_top] = rp;
8259 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8261 if (nflags & MDB_APPEND) {
8262 mn.mc_ki[mn.mc_top] = 0;
8264 split_indx = newindx;
8268 split_indx = (nkeys+1) / 2;
8273 unsigned int lsize, rsize, ksize;
8274 /* Move half of the keys to the right sibling */
8275 x = mc->mc_ki[mc->mc_top] - split_indx;
8276 ksize = mc->mc_db->md_pad;
8277 split = LEAF2KEY(mp, split_indx, ksize);
8278 rsize = (nkeys - split_indx) * ksize;
8279 lsize = (nkeys - split_indx) * sizeof(indx_t);
8280 mp->mp_lower -= lsize;
8281 rp->mp_lower += lsize;
8282 mp->mp_upper += rsize - lsize;
8283 rp->mp_upper -= rsize - lsize;
8284 sepkey.mv_size = ksize;
8285 if (newindx == split_indx) {
8286 sepkey.mv_data = newkey->mv_data;
8288 sepkey.mv_data = split;
8291 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8292 memcpy(rp->mp_ptrs, split, rsize);
8293 sepkey.mv_data = rp->mp_ptrs;
8294 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8295 memcpy(ins, newkey->mv_data, ksize);
8296 mp->mp_lower += sizeof(indx_t);
8297 mp->mp_upper -= ksize - sizeof(indx_t);
8300 memcpy(rp->mp_ptrs, split, x * ksize);
8301 ins = LEAF2KEY(rp, x, ksize);
8302 memcpy(ins, newkey->mv_data, ksize);
8303 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8304 rp->mp_lower += sizeof(indx_t);
8305 rp->mp_upper -= ksize - sizeof(indx_t);
8306 mc->mc_ki[mc->mc_top] = x;
8307 mc->mc_pg[mc->mc_top] = rp;
8310 int psize, nsize, k;
8311 /* Maximum free space in an empty page */
8312 pmax = env->me_psize - PAGEHDRSZ;
8314 nsize = mdb_leaf_size(env, newkey, newdata);
8316 nsize = mdb_branch_size(env, newkey);
8317 nsize = EVEN(nsize);
8319 /* grab a page to hold a temporary copy */
8320 copy = mdb_page_malloc(mc->mc_txn, 1);
8325 copy->mp_pgno = mp->mp_pgno;
8326 copy->mp_flags = mp->mp_flags;
8327 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8328 copy->mp_upper = env->me_psize - PAGEBASE;
8330 /* prepare to insert */
8331 for (i=0, j=0; i<nkeys; i++) {
8333 copy->mp_ptrs[j++] = 0;
8335 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8338 /* When items are relatively large the split point needs
8339 * to be checked, because being off-by-one will make the
8340 * difference between success or failure in mdb_node_add.
8342 * It's also relevant if a page happens to be laid out
8343 * such that one half of its nodes are all "small" and
8344 * the other half of its nodes are "large." If the new
8345 * item is also "large" and falls on the half with
8346 * "large" nodes, it also may not fit.
8348 * As a final tweak, if the new item goes on the last
8349 * spot on the page (and thus, onto the new page), bias
8350 * the split so the new page is emptier than the old page.
8351 * This yields better packing during sequential inserts.
8353 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8354 /* Find split point */
8356 if (newindx <= split_indx || newindx >= nkeys) {
8358 k = newindx >= nkeys ? nkeys : split_indx+2;
8363 for (; i!=k; i+=j) {
8368 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8369 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8371 if (F_ISSET(node->mn_flags, F_BIGDATA))
8372 psize += sizeof(pgno_t);
8374 psize += NODEDSZ(node);
8376 psize = EVEN(psize);
8378 if (psize > pmax || i == k-j) {
8379 split_indx = i + (j<0);
8384 if (split_indx == newindx) {
8385 sepkey.mv_size = newkey->mv_size;
8386 sepkey.mv_data = newkey->mv_data;
8388 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8389 sepkey.mv_size = node->mn_ksize;
8390 sepkey.mv_data = NODEKEY(node);
8395 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8397 /* Copy separator key to the parent.
8399 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8403 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8408 if (mn.mc_snum == mc->mc_snum) {
8409 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8410 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8411 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8412 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8417 /* Right page might now have changed parent.
8418 * Check if left page also changed parent.
8420 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8421 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8422 for (i=0; i<ptop; i++) {
8423 mc->mc_pg[i] = mn.mc_pg[i];
8424 mc->mc_ki[i] = mn.mc_ki[i];
8426 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8427 if (mn.mc_ki[ptop]) {
8428 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8430 /* find right page's left sibling */
8431 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8432 mdb_cursor_sibling(mc, 0);
8437 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8440 mc->mc_flags ^= C_SPLITTING;
8441 if (rc != MDB_SUCCESS) {
8444 if (nflags & MDB_APPEND) {
8445 mc->mc_pg[mc->mc_top] = rp;
8446 mc->mc_ki[mc->mc_top] = 0;
8447 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8450 for (i=0; i<mc->mc_top; i++)
8451 mc->mc_ki[i] = mn.mc_ki[i];
8452 } else if (!IS_LEAF2(mp)) {
8454 mc->mc_pg[mc->mc_top] = rp;
8459 rkey.mv_data = newkey->mv_data;
8460 rkey.mv_size = newkey->mv_size;
8466 /* Update index for the new key. */
8467 mc->mc_ki[mc->mc_top] = j;
8469 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8470 rkey.mv_data = NODEKEY(node);
8471 rkey.mv_size = node->mn_ksize;
8473 xdata.mv_data = NODEDATA(node);
8474 xdata.mv_size = NODEDSZ(node);
8477 pgno = NODEPGNO(node);
8478 flags = node->mn_flags;
8481 if (!IS_LEAF(mp) && j == 0) {
8482 /* First branch index doesn't need key data. */
8486 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8492 mc->mc_pg[mc->mc_top] = copy;
8497 } while (i != split_indx);
8499 nkeys = NUMKEYS(copy);
8500 for (i=0; i<nkeys; i++)
8501 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8502 mp->mp_lower = copy->mp_lower;
8503 mp->mp_upper = copy->mp_upper;
8504 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8505 env->me_psize - copy->mp_upper - PAGEBASE);
8507 /* reset back to original page */
8508 if (newindx < split_indx) {
8509 mc->mc_pg[mc->mc_top] = mp;
8510 if (nflags & MDB_RESERVE) {
8511 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8512 if (!(node->mn_flags & F_BIGDATA))
8513 newdata->mv_data = NODEDATA(node);
8516 mc->mc_pg[mc->mc_top] = rp;
8518 /* Make sure mc_ki is still valid.
8520 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8521 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8522 for (i=0; i<=ptop; i++) {
8523 mc->mc_pg[i] = mn.mc_pg[i];
8524 mc->mc_ki[i] = mn.mc_ki[i];
8531 /* Adjust other cursors pointing to mp */
8532 MDB_cursor *m2, *m3;
8533 MDB_dbi dbi = mc->mc_dbi;
8534 int fixup = NUMKEYS(mp);
8536 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8537 if (mc->mc_flags & C_SUB)
8538 m3 = &m2->mc_xcursor->mx_cursor;
8543 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8545 if (m3->mc_flags & C_SPLITTING)
8550 for (k=m3->mc_top; k>=0; k--) {
8551 m3->mc_ki[k+1] = m3->mc_ki[k];
8552 m3->mc_pg[k+1] = m3->mc_pg[k];
8554 if (m3->mc_ki[0] >= split_indx) {
8559 m3->mc_pg[0] = mc->mc_pg[0];
8563 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8564 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8565 m3->mc_ki[mc->mc_top]++;
8566 if (m3->mc_ki[mc->mc_top] >= fixup) {
8567 m3->mc_pg[mc->mc_top] = rp;
8568 m3->mc_ki[mc->mc_top] -= fixup;
8569 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8571 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8572 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8577 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8580 if (copy) /* tmp page */
8581 mdb_page_free(env, copy);
8583 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8588 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8589 MDB_val *key, MDB_val *data, unsigned int flags)
8594 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8597 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8600 mdb_cursor_init(&mc, txn, dbi, &mx);
8601 return mdb_cursor_put(&mc, key, data, flags);
8605 #define MDB_WBUF (1024*1024)
8608 /** State needed for a compacting copy. */
8609 typedef struct mdb_copy {
8610 pthread_mutex_t mc_mutex;
8611 pthread_cond_t mc_cond;
8618 pgno_t mc_next_pgno;
8621 volatile int mc_new;
8626 /** Dedicated writer thread for compacting copy. */
8627 static THREAD_RET ESECT
8628 mdb_env_copythr(void *arg)
8632 int toggle = 0, wsize, rc;
8635 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8638 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8641 pthread_mutex_lock(&my->mc_mutex);
8643 pthread_cond_signal(&my->mc_cond);
8646 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8647 if (my->mc_new < 0) {
8652 wsize = my->mc_wlen[toggle];
8653 ptr = my->mc_wbuf[toggle];
8656 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8660 } else if (len > 0) {
8674 /* If there's an overflow page tail, write it too */
8675 if (my->mc_olen[toggle]) {
8676 wsize = my->mc_olen[toggle];
8677 ptr = my->mc_over[toggle];
8678 my->mc_olen[toggle] = 0;
8681 my->mc_wlen[toggle] = 0;
8683 pthread_cond_signal(&my->mc_cond);
8685 pthread_cond_signal(&my->mc_cond);
8686 pthread_mutex_unlock(&my->mc_mutex);
8687 return (THREAD_RET)0;
8691 /** Tell the writer thread there's a buffer ready to write */
8693 mdb_env_cthr_toggle(mdb_copy *my, int st)
8695 int toggle = my->mc_toggle ^ 1;
8696 pthread_mutex_lock(&my->mc_mutex);
8697 if (my->mc_status) {
8698 pthread_mutex_unlock(&my->mc_mutex);
8699 return my->mc_status;
8701 while (my->mc_new == 1)
8702 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8704 my->mc_toggle = toggle;
8705 pthread_cond_signal(&my->mc_cond);
8706 pthread_mutex_unlock(&my->mc_mutex);
8710 /** Depth-first tree traversal for compacting copy. */
8712 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8715 MDB_txn *txn = my->mc_txn;
8717 MDB_page *mo, *mp, *leaf;
8722 /* Empty DB, nothing to do */
8723 if (*pg == P_INVALID)
8730 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8733 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8737 /* Make cursor pages writable */
8738 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8742 for (i=0; i<mc.mc_top; i++) {
8743 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8744 mc.mc_pg[i] = (MDB_page *)ptr;
8745 ptr += my->mc_env->me_psize;
8748 /* This is writable space for a leaf page. Usually not needed. */
8749 leaf = (MDB_page *)ptr;
8751 toggle = my->mc_toggle;
8752 while (mc.mc_snum > 0) {
8754 mp = mc.mc_pg[mc.mc_top];
8758 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8759 for (i=0; i<n; i++) {
8760 ni = NODEPTR(mp, i);
8761 if (ni->mn_flags & F_BIGDATA) {
8765 /* Need writable leaf */
8767 mc.mc_pg[mc.mc_top] = leaf;
8768 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8770 ni = NODEPTR(mp, i);
8773 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8774 rc = mdb_page_get(txn, pg, &omp, NULL);
8777 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8778 rc = mdb_env_cthr_toggle(my, 1);
8781 toggle = my->mc_toggle;
8783 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8784 memcpy(mo, omp, my->mc_env->me_psize);
8785 mo->mp_pgno = my->mc_next_pgno;
8786 my->mc_next_pgno += omp->mp_pages;
8787 my->mc_wlen[toggle] += my->mc_env->me_psize;
8788 if (omp->mp_pages > 1) {
8789 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8790 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8791 rc = mdb_env_cthr_toggle(my, 1);
8794 toggle = my->mc_toggle;
8796 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8797 } else if (ni->mn_flags & F_SUBDATA) {
8800 /* Need writable leaf */
8802 mc.mc_pg[mc.mc_top] = leaf;
8803 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8805 ni = NODEPTR(mp, i);
8808 memcpy(&db, NODEDATA(ni), sizeof(db));
8809 my->mc_toggle = toggle;
8810 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8813 toggle = my->mc_toggle;
8814 memcpy(NODEDATA(ni), &db, sizeof(db));
8819 mc.mc_ki[mc.mc_top]++;
8820 if (mc.mc_ki[mc.mc_top] < n) {
8823 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8825 rc = mdb_page_get(txn, pg, &mp, NULL);
8830 mc.mc_ki[mc.mc_top] = 0;
8831 if (IS_BRANCH(mp)) {
8832 /* Whenever we advance to a sibling branch page,
8833 * we must proceed all the way down to its first leaf.
8835 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8838 mc.mc_pg[mc.mc_top] = mp;
8842 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8843 rc = mdb_env_cthr_toggle(my, 1);
8846 toggle = my->mc_toggle;
8848 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8849 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8850 mo->mp_pgno = my->mc_next_pgno++;
8851 my->mc_wlen[toggle] += my->mc_env->me_psize;
8853 /* Update parent if there is one */
8854 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8855 SETPGNO(ni, mo->mp_pgno);
8856 mdb_cursor_pop(&mc);
8858 /* Otherwise we're done */
8868 /** Copy environment with compaction. */
8870 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8875 MDB_txn *txn = NULL;
8880 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8881 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8882 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8883 if (my.mc_wbuf[0] == NULL)
8886 pthread_mutex_init(&my.mc_mutex, NULL);
8887 pthread_cond_init(&my.mc_cond, NULL);
8888 #ifdef HAVE_MEMALIGN
8889 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8890 if (my.mc_wbuf[0] == NULL)
8893 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8898 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8899 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8904 my.mc_next_pgno = 2;
8910 THREAD_CREATE(thr, mdb_env_copythr, &my);
8912 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8916 mp = (MDB_page *)my.mc_wbuf[0];
8917 memset(mp, 0, 2*env->me_psize);
8919 mp->mp_flags = P_META;
8920 mm = (MDB_meta *)METADATA(mp);
8921 mdb_env_init_meta0(env, mm);
8922 mm->mm_address = env->me_metas[0]->mm_address;
8924 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8926 mp->mp_flags = P_META;
8927 *(MDB_meta *)METADATA(mp) = *mm;
8928 mm = (MDB_meta *)METADATA(mp);
8930 /* Count the number of free pages, subtract from lastpg to find
8931 * number of active pages
8934 MDB_ID freecount = 0;
8937 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8938 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8939 freecount += *(MDB_ID *)data.mv_data;
8940 freecount += txn->mt_dbs[0].md_branch_pages +
8941 txn->mt_dbs[0].md_leaf_pages +
8942 txn->mt_dbs[0].md_overflow_pages;
8944 /* Set metapage 1 */
8945 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8946 mm->mm_dbs[1] = txn->mt_dbs[1];
8947 if (mm->mm_last_pg > 1) {
8948 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8951 mm->mm_dbs[1].md_root = P_INVALID;
8954 my.mc_wlen[0] = env->me_psize * 2;
8956 pthread_mutex_lock(&my.mc_mutex);
8958 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8959 pthread_mutex_unlock(&my.mc_mutex);
8960 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8961 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8962 rc = mdb_env_cthr_toggle(&my, 1);
8963 mdb_env_cthr_toggle(&my, -1);
8964 pthread_mutex_lock(&my.mc_mutex);
8966 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8967 pthread_mutex_unlock(&my.mc_mutex);
8972 CloseHandle(my.mc_cond);
8973 CloseHandle(my.mc_mutex);
8974 _aligned_free(my.mc_wbuf[0]);
8976 pthread_cond_destroy(&my.mc_cond);
8977 pthread_mutex_destroy(&my.mc_mutex);
8978 free(my.mc_wbuf[0]);
8983 /** Copy environment as-is. */
8985 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8987 MDB_txn *txn = NULL;
8988 mdb_mutexref_t wmutex = NULL;
8994 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8998 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9001 /* Do the lock/unlock of the reader mutex before starting the
9002 * write txn. Otherwise other read txns could block writers.
9004 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9009 /* We must start the actual read txn after blocking writers */
9010 mdb_txn_reset0(txn, "reset-stage1");
9012 /* Temporarily block writers until we snapshot the meta pages */
9013 wmutex = env->me_wmutex;
9014 if (LOCK_MUTEX(rc, env, wmutex))
9017 rc = mdb_txn_renew0(txn);
9019 UNLOCK_MUTEX(wmutex);
9024 wsize = env->me_psize * 2;
9028 DO_WRITE(rc, fd, ptr, w2, len);
9032 } else if (len > 0) {
9038 /* Non-blocking or async handles are not supported */
9044 UNLOCK_MUTEX(wmutex);
9049 w2 = txn->mt_next_pgno * env->me_psize;
9052 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9059 if (wsize > MAX_WRITE)
9063 DO_WRITE(rc, fd, ptr, w2, len);
9067 } else if (len > 0) {
9084 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9086 if (flags & MDB_CP_COMPACT)
9087 return mdb_env_copyfd1(env, fd);
9089 return mdb_env_copyfd0(env, fd);
9093 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9095 return mdb_env_copyfd2(env, fd, 0);
9099 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9103 HANDLE newfd = INVALID_HANDLE_VALUE;
9105 if (env->me_flags & MDB_NOSUBDIR) {
9106 lpath = (char *)path;
9109 len += sizeof(DATANAME);
9110 lpath = malloc(len);
9113 sprintf(lpath, "%s" DATANAME, path);
9116 /* The destination path must exist, but the destination file must not.
9117 * We don't want the OS to cache the writes, since the source data is
9118 * already in the OS cache.
9121 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9122 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9124 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9126 if (newfd == INVALID_HANDLE_VALUE) {
9131 if (env->me_psize >= env->me_os_psize) {
9133 /* Set O_DIRECT if the file system supports it */
9134 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9135 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9137 #ifdef F_NOCACHE /* __APPLE__ */
9138 rc = fcntl(newfd, F_NOCACHE, 1);
9146 rc = mdb_env_copyfd2(env, newfd, flags);
9149 if (!(env->me_flags & MDB_NOSUBDIR))
9151 if (newfd != INVALID_HANDLE_VALUE)
9152 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9159 mdb_env_copy(MDB_env *env, const char *path)
9161 return mdb_env_copy2(env, path, 0);
9165 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9167 if (flag & ~CHANGEABLE)
9170 env->me_flags |= flag;
9172 env->me_flags &= ~flag;
9177 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9182 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9187 mdb_env_set_userctx(MDB_env *env, void *ctx)
9191 env->me_userctx = ctx;
9196 mdb_env_get_userctx(MDB_env *env)
9198 return env ? env->me_userctx : NULL;
9202 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9207 env->me_assert_func = func;
9213 mdb_env_get_path(MDB_env *env, const char **arg)
9218 *arg = env->me_path;
9223 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9232 /** Common code for #mdb_stat() and #mdb_env_stat().
9233 * @param[in] env the environment to operate in.
9234 * @param[in] db the #MDB_db record containing the stats to return.
9235 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9236 * @return 0, this function always succeeds.
9239 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9241 arg->ms_psize = env->me_psize;
9242 arg->ms_depth = db->md_depth;
9243 arg->ms_branch_pages = db->md_branch_pages;
9244 arg->ms_leaf_pages = db->md_leaf_pages;
9245 arg->ms_overflow_pages = db->md_overflow_pages;
9246 arg->ms_entries = db->md_entries;
9252 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9256 if (env == NULL || arg == NULL)
9259 meta = mdb_env_pick_meta(env);
9261 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9265 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9269 if (env == NULL || arg == NULL)
9272 meta = mdb_env_pick_meta(env);
9273 arg->me_mapaddr = meta->mm_address;
9274 arg->me_last_pgno = meta->mm_last_pg;
9275 arg->me_last_txnid = meta->mm_txnid;
9277 arg->me_mapsize = env->me_mapsize;
9278 arg->me_maxreaders = env->me_maxreaders;
9279 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9283 /** Set the default comparison functions for a database.
9284 * Called immediately after a database is opened to set the defaults.
9285 * The user can then override them with #mdb_set_compare() or
9286 * #mdb_set_dupsort().
9287 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9288 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9291 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9293 uint16_t f = txn->mt_dbs[dbi].md_flags;
9295 txn->mt_dbxs[dbi].md_cmp =
9296 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9297 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9299 txn->mt_dbxs[dbi].md_dcmp =
9300 !(f & MDB_DUPSORT) ? 0 :
9301 ((f & MDB_INTEGERDUP)
9302 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9303 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9306 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9312 int rc, dbflag, exact;
9313 unsigned int unused = 0, seq;
9316 if (flags & ~VALID_FLAGS)
9318 if (txn->mt_flags & MDB_TXN_ERROR)
9324 if (flags & PERSISTENT_FLAGS) {
9325 uint16_t f2 = flags & PERSISTENT_FLAGS;
9326 /* make sure flag changes get committed */
9327 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9328 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9329 txn->mt_flags |= MDB_TXN_DIRTY;
9332 mdb_default_cmp(txn, MAIN_DBI);
9336 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9337 mdb_default_cmp(txn, MAIN_DBI);
9340 /* Is the DB already open? */
9342 for (i=2; i<txn->mt_numdbs; i++) {
9343 if (!txn->mt_dbxs[i].md_name.mv_size) {
9344 /* Remember this free slot */
9345 if (!unused) unused = i;
9348 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9349 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9355 /* If no free slot and max hit, fail */
9356 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9357 return MDB_DBS_FULL;
9359 /* Cannot mix named databases with some mainDB flags */
9360 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9361 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9363 /* Find the DB info */
9364 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9367 key.mv_data = (void *)name;
9368 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9369 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9370 if (rc == MDB_SUCCESS) {
9371 /* make sure this is actually a DB */
9372 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9373 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9374 return MDB_INCOMPATIBLE;
9375 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9376 /* Create if requested */
9377 data.mv_size = sizeof(MDB_db);
9378 data.mv_data = &dummy;
9379 memset(&dummy, 0, sizeof(dummy));
9380 dummy.md_root = P_INVALID;
9381 dummy.md_flags = flags & PERSISTENT_FLAGS;
9382 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9386 /* OK, got info, add to table */
9387 if (rc == MDB_SUCCESS) {
9388 unsigned int slot = unused ? unused : txn->mt_numdbs;
9389 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9390 txn->mt_dbxs[slot].md_name.mv_size = len;
9391 txn->mt_dbxs[slot].md_rel = NULL;
9392 txn->mt_dbflags[slot] = dbflag;
9393 /* txn-> and env-> are the same in read txns, use
9394 * tmp variable to avoid undefined assignment
9396 seq = ++txn->mt_env->me_dbiseqs[slot];
9397 txn->mt_dbiseqs[slot] = seq;
9399 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9401 mdb_default_cmp(txn, slot);
9411 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9413 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9416 if (txn->mt_flags & MDB_TXN_ERROR)
9419 if (txn->mt_dbflags[dbi] & DB_STALE) {
9422 /* Stale, must read the DB's root. cursor_init does it for us. */
9423 mdb_cursor_init(&mc, txn, dbi, &mx);
9425 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9428 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9431 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9433 ptr = env->me_dbxs[dbi].md_name.mv_data;
9434 /* If there was no name, this was already closed */
9436 env->me_dbxs[dbi].md_name.mv_data = NULL;
9437 env->me_dbxs[dbi].md_name.mv_size = 0;
9438 env->me_dbflags[dbi] = 0;
9439 env->me_dbiseqs[dbi]++;
9444 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9446 /* We could return the flags for the FREE_DBI too but what's the point? */
9447 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9449 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9453 /** Add all the DB's pages to the free list.
9454 * @param[in] mc Cursor on the DB to free.
9455 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9456 * @return 0 on success, non-zero on failure.
9459 mdb_drop0(MDB_cursor *mc, int subs)
9463 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9464 if (rc == MDB_SUCCESS) {
9465 MDB_txn *txn = mc->mc_txn;
9470 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9471 * This also avoids any P_LEAF2 pages, which have no nodes.
9473 if (mc->mc_flags & C_SUB)
9476 mdb_cursor_copy(mc, &mx);
9477 while (mc->mc_snum > 0) {
9478 MDB_page *mp = mc->mc_pg[mc->mc_top];
9479 unsigned n = NUMKEYS(mp);
9481 for (i=0; i<n; i++) {
9482 ni = NODEPTR(mp, i);
9483 if (ni->mn_flags & F_BIGDATA) {
9486 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9487 rc = mdb_page_get(txn, pg, &omp, NULL);
9490 mdb_cassert(mc, IS_OVERFLOW(omp));
9491 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9495 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9496 mdb_xcursor_init1(mc, ni);
9497 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9503 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9505 for (i=0; i<n; i++) {
9507 ni = NODEPTR(mp, i);
9510 mdb_midl_xappend(txn->mt_free_pgs, pg);
9515 mc->mc_ki[mc->mc_top] = i;
9516 rc = mdb_cursor_sibling(mc, 1);
9518 if (rc != MDB_NOTFOUND)
9520 /* no more siblings, go back to beginning
9521 * of previous level.
9525 for (i=1; i<mc->mc_snum; i++) {
9527 mc->mc_pg[i] = mx.mc_pg[i];
9532 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9535 txn->mt_flags |= MDB_TXN_ERROR;
9536 } else if (rc == MDB_NOTFOUND) {
9542 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9544 MDB_cursor *mc, *m2;
9547 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9550 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9553 if (TXN_DBI_CHANGED(txn, dbi))
9556 rc = mdb_cursor_open(txn, dbi, &mc);
9560 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9561 /* Invalidate the dropped DB's cursors */
9562 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9563 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9567 /* Can't delete the main DB */
9568 if (del && dbi > MAIN_DBI) {
9569 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9571 txn->mt_dbflags[dbi] = DB_STALE;
9572 mdb_dbi_close(txn->mt_env, dbi);
9574 txn->mt_flags |= MDB_TXN_ERROR;
9577 /* reset the DB record, mark it dirty */
9578 txn->mt_dbflags[dbi] |= DB_DIRTY;
9579 txn->mt_dbs[dbi].md_depth = 0;
9580 txn->mt_dbs[dbi].md_branch_pages = 0;
9581 txn->mt_dbs[dbi].md_leaf_pages = 0;
9582 txn->mt_dbs[dbi].md_overflow_pages = 0;
9583 txn->mt_dbs[dbi].md_entries = 0;
9584 txn->mt_dbs[dbi].md_root = P_INVALID;
9586 txn->mt_flags |= MDB_TXN_DIRTY;
9589 mdb_cursor_close(mc);
9593 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9595 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9598 txn->mt_dbxs[dbi].md_cmp = cmp;
9602 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9604 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9607 txn->mt_dbxs[dbi].md_dcmp = cmp;
9611 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9613 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9616 txn->mt_dbxs[dbi].md_rel = rel;
9620 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9622 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9625 txn->mt_dbxs[dbi].md_relctx = ctx;
9630 mdb_env_get_maxkeysize(MDB_env *env)
9632 return ENV_MAXKEY(env);
9636 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9638 unsigned int i, rdrs;
9641 int rc = 0, first = 1;
9645 if (!env->me_txns) {
9646 return func("(no reader locks)\n", ctx);
9648 rdrs = env->me_txns->mti_numreaders;
9649 mr = env->me_txns->mti_readers;
9650 for (i=0; i<rdrs; i++) {
9652 txnid_t txnid = mr[i].mr_txnid;
9653 sprintf(buf, txnid == (txnid_t)-1 ?
9654 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9655 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9658 rc = func(" pid thread txnid\n", ctx);
9662 rc = func(buf, ctx);
9668 rc = func("(no active readers)\n", ctx);
9673 /** Insert pid into list if not already present.
9674 * return -1 if already present.
9677 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9679 /* binary search of pid in list */
9681 unsigned cursor = 1;
9683 unsigned n = ids[0];
9686 unsigned pivot = n >> 1;
9687 cursor = base + pivot + 1;
9688 val = pid - ids[cursor];
9693 } else if ( val > 0 ) {
9698 /* found, so it's a duplicate */
9707 for (n = ids[0]; n > cursor; n--)
9714 mdb_reader_check(MDB_env *env, int *dead)
9720 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9723 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9725 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9727 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9728 unsigned int i, j, rdrs;
9730 MDB_PID_T *pids, pid;
9731 int rc = MDB_SUCCESS, count = 0;
9733 rdrs = env->me_txns->mti_numreaders;
9734 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9738 mr = env->me_txns->mti_readers;
9739 for (i=0; i<rdrs; i++) {
9741 if (pid && pid != env->me_pid) {
9742 if (mdb_pid_insert(pids, pid) == 0) {
9743 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9744 /* Stale reader found */
9747 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9748 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9750 rdrs = 0; /* the above checked all readers */
9752 /* Recheck, a new process may have reused pid */
9753 if (mdb_reader_pid(env, Pidcheck, pid))
9758 if (mr[j].mr_pid == pid) {
9759 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9760 (unsigned) pid, mr[j].mr_txnid));
9765 UNLOCK_MUTEX(rmutex);
9776 #ifdef MDB_ROBUST_SUPPORTED
9777 /** Handle #LOCK_MUTEX0() failure.
9778 * Try to repair the lock file if the mutex owner died.
9779 * @param[in] env the environment handle
9780 * @param[in] mutex LOCK_MUTEX0() mutex
9781 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9782 * @return 0 on success with the mutex locked, or an error code on failure.
9785 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9790 if (rc == MDB_OWNERDEAD) {
9791 /* We own the mutex. Clean up after dead previous owner. */
9793 rlocked = (mutex == env->me_rmutex);
9795 /* Keep mti_txnid updated, otherwise next writer can
9796 * overwrite data which latest meta page refers to.
9798 meta = mdb_env_pick_meta(env);
9799 env->me_txns->mti_txnid = meta->mm_txnid;
9800 /* env is hosed if the dead thread was ours */
9802 env->me_flags |= MDB_FATAL_ERROR;
9807 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9808 (rc ? "this process' env is hosed" : "recovering")));
9809 rc2 = mdb_reader_check0(env, rlocked, NULL);
9811 rc2 = mdb_mutex_consistent(mutex);
9812 if (rc || (rc = rc2)) {
9813 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9814 UNLOCK_MUTEX(mutex);
9820 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9825 #endif /* MDB_ROBUST_SUPPORTED */
projects / openldap / blob