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 2011-2016 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.
38 #if defined(__WIN64__)
39 #define _FILE_OFFSET_BITS 64
44 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
45 * as int64 which is wrong. MSVC doesn't define it at all, so just
49 #define MDB_THR_T DWORD
50 #include <sys/types.h>
53 # include <sys/param.h>
55 # define LITTLE_ENDIAN 1234
56 # define BIG_ENDIAN 4321
57 # define BYTE_ORDER LITTLE_ENDIAN
59 # define SSIZE_MAX INT_MAX
63 #include <sys/types.h>
65 #define MDB_PID_T pid_t
66 #define MDB_THR_T pthread_t
67 #include <sys/param.h>
70 #ifdef HAVE_SYS_FILE_H
76 #if defined(__mips) && defined(__linux)
77 /* MIPS has cache coherency issues, requires explicit cache control */
78 #include <asm/cachectl.h>
79 extern int cacheflush(char *addr, int nbytes, int cache);
80 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
82 #define CACHEFLUSH(addr, bytes, cache)
85 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
86 /** fdatasync is broken on ext3/ext4fs on older kernels, see
87 * description in #mdb_env_open2 comments. You can safely
88 * define MDB_FDATASYNC_WORKS if this code will only be run
89 * on kernels 3.6 and newer.
91 #define BROKEN_FDATASYNC
105 typedef SSIZE_T ssize_t;
110 #if defined(__sun) || defined(ANDROID)
111 /* Most platforms have posix_memalign, older may only have memalign */
112 #define HAVE_MEMALIGN 1
116 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
117 #include <netinet/in.h>
118 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
121 #if defined(__APPLE__) || defined (BSD)
122 # define MDB_USE_POSIX_SEM 1
123 # define MDB_FDATASYNC fsync
124 #elif defined(ANDROID)
125 # define MDB_FDATASYNC fsync
130 #ifdef MDB_USE_POSIX_SEM
131 # define MDB_USE_HASH 1
132 #include <semaphore.h>
134 #define MDB_USE_POSIX_MUTEX 1
138 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \
139 + defined(MDB_USE_POSIX_MUTEX) != 1
140 # error "Ambiguous shared-lock implementation"
144 #include <valgrind/memcheck.h>
145 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
146 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
147 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
148 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
149 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
151 #define VGMEMP_CREATE(h,r,z)
152 #define VGMEMP_ALLOC(h,a,s)
153 #define VGMEMP_FREE(h,a)
154 #define VGMEMP_DESTROY(h)
155 #define VGMEMP_DEFINED(a,s)
159 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
160 /* Solaris just defines one or the other */
161 # define LITTLE_ENDIAN 1234
162 # define BIG_ENDIAN 4321
163 # ifdef _LITTLE_ENDIAN
164 # define BYTE_ORDER LITTLE_ENDIAN
166 # define BYTE_ORDER BIG_ENDIAN
169 # define BYTE_ORDER __BYTE_ORDER
173 #ifndef LITTLE_ENDIAN
174 #define LITTLE_ENDIAN __LITTLE_ENDIAN
177 #define BIG_ENDIAN __BIG_ENDIAN
180 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
181 #define MISALIGNED_OK 1
187 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
188 # error "Unknown or unsupported endianness (BYTE_ORDER)"
189 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
190 # error "Two's complement, reasonably sized integer types, please"
194 /** Put infrequently used env functions in separate section */
196 # define ESECT __attribute__ ((section("__TEXT,text_env")))
198 # define ESECT __attribute__ ((section("text_env")))
205 #define CALL_CONV WINAPI
210 /** @defgroup internal LMDB Internals
213 /** @defgroup compat Compatibility Macros
214 * A bunch of macros to minimize the amount of platform-specific ifdefs
215 * needed throughout the rest of the code. When the features this library
216 * needs are similar enough to POSIX to be hidden in a one-or-two line
217 * replacement, this macro approach is used.
221 /** Features under development */
226 /** Wrapper around __func__, which is a C99 feature */
227 #if __STDC_VERSION__ >= 199901L
228 # define mdb_func_ __func__
229 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
230 # define mdb_func_ __FUNCTION__
232 /* If a debug message says <mdb_unknown>(), update the #if statements above */
233 # define mdb_func_ "<mdb_unknown>"
236 /* Internal error codes, not exposed outside liblmdb */
237 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
239 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
240 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
241 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
245 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
247 /** Some platforms define the EOWNERDEAD error code
248 * even though they don't support Robust Mutexes.
249 * Compile with -DMDB_USE_ROBUST=0, or use some other
250 * mechanism like -DMDB_USE_POSIX_SEM instead of
251 * -DMDB_USE_POSIX_MUTEX.
252 * (Posix semaphores are not robust.)
254 #ifndef MDB_USE_ROBUST
255 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
256 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
257 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
258 # define MDB_USE_ROBUST 0
260 # define MDB_USE_ROBUST 1
261 /* glibc < 2.12 only provided _np API */
262 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
263 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
264 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
265 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
266 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
269 #endif /* MDB_USE_ROBUST */
271 #if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
272 #define MDB_ROBUST_SUPPORTED 1
276 #define MDB_USE_HASH 1
277 #define MDB_PIDLOCK 0
278 #define THREAD_RET DWORD
279 #define pthread_t HANDLE
280 #define pthread_mutex_t HANDLE
281 #define pthread_cond_t HANDLE
282 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
283 #define pthread_key_t DWORD
284 #define pthread_self() GetCurrentThreadId()
285 #define pthread_key_create(x,y) \
286 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
287 #define pthread_key_delete(x) TlsFree(x)
288 #define pthread_getspecific(x) TlsGetValue(x)
289 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
290 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
291 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
292 #define pthread_cond_signal(x) SetEvent(*x)
293 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
294 #define THREAD_CREATE(thr,start,arg) \
295 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
296 #define THREAD_FINISH(thr) \
297 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
298 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
299 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
300 #define mdb_mutex_consistent(mutex) 0
301 #define getpid() GetCurrentProcessId()
302 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
303 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
304 #define ErrCode() GetLastError()
305 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
306 #define close(fd) (CloseHandle(fd) ? 0 : -1)
307 #define munmap(ptr,len) UnmapViewOfFile(ptr)
308 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
309 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
311 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
315 #define THREAD_RET void *
316 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
317 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
318 #define Z "z" /**< printf format modifier for size_t */
320 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
321 #define MDB_PIDLOCK 1
323 #ifdef MDB_USE_POSIX_SEM
325 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
326 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
327 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
330 mdb_sem_wait(sem_t *sem)
333 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
337 #else /* MDB_USE_POSIX_MUTEX: */
338 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
339 * local variables keep it (mdb_mutexref_t).
341 * When #mdb_mutexref_t is a pointer declaration and #mdb_mutex_t is
342 * not, then it is array[size 1] so it can be assigned to a pointer.
345 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
347 /** Lock the reader or writer mutex.
348 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
350 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
351 /** Unlock the reader or writer mutex.
353 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
354 /** Mark mutex-protected data as repaired, after death of previous owner.
356 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
357 #endif /* MDB_USE_POSIX_SEM */
359 /** Get the error code for the last failed system function.
361 #define ErrCode() errno
363 /** An abstraction for a file handle.
364 * On POSIX systems file handles are small integers. On Windows
365 * they're opaque pointers.
369 /** A value for an invalid file handle.
370 * Mainly used to initialize file variables and signify that they are
373 #define INVALID_HANDLE_VALUE (-1)
375 /** Get the size of a memory page for the system.
376 * This is the basic size that the platform's memory manager uses, and is
377 * fundamental to the use of memory-mapped files.
379 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
382 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
385 #define MNAME_LEN (sizeof(pthread_mutex_t))
390 #ifdef MDB_ROBUST_SUPPORTED
391 /** Lock mutex, handle any error, set rc = result.
392 * Return 0 on success, nonzero (not rc) on error.
394 #define LOCK_MUTEX(rc, env, mutex) \
395 (((rc) = LOCK_MUTEX0(mutex)) && \
396 ((rc) = mdb_mutex_failed(env, mutex, rc)))
397 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
399 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
400 #define mdb_mutex_failed(env, mutex, rc) (rc)
404 /** A flag for opening a file and requesting synchronous data writes.
405 * This is only used when writing a meta page. It's not strictly needed;
406 * we could just do a normal write and then immediately perform a flush.
407 * But if this flag is available it saves us an extra system call.
409 * @note If O_DSYNC is undefined but exists in /usr/include,
410 * preferably set some compiler flag to get the definition.
414 # define MDB_DSYNC O_DSYNC
416 # define MDB_DSYNC O_SYNC
421 /** Function for flushing the data of a file. Define this to fsync
422 * if fdatasync() is not supported.
424 #ifndef MDB_FDATASYNC
425 # define MDB_FDATASYNC fdatasync
429 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
440 /** A page number in the database.
441 * Note that 64 bit page numbers are overkill, since pages themselves
442 * already represent 12-13 bits of addressable memory, and the OS will
443 * always limit applications to a maximum of 63 bits of address space.
445 * @note In the #MDB_node structure, we only store 48 bits of this value,
446 * which thus limits us to only 60 bits of addressable data.
448 typedef MDB_ID pgno_t;
450 /** A transaction ID.
451 * See struct MDB_txn.mt_txnid for details.
453 typedef MDB_ID txnid_t;
455 /** @defgroup debug Debug Macros
459 /** Enable debug output. Needs variable argument macros (a C99 feature).
460 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
461 * read from and written to the database (used for free space management).
467 static int mdb_debug;
468 static txnid_t mdb_debug_start;
470 /** Print a debug message with printf formatting.
471 * Requires double parenthesis around 2 or more args.
473 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
474 # define DPRINTF0(fmt, ...) \
475 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
477 # define DPRINTF(args) ((void) 0)
479 /** Print a debug string.
480 * The string is printed literally, with no format processing.
482 #define DPUTS(arg) DPRINTF(("%s", arg))
483 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
485 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
488 /** @brief The maximum size of a database page.
490 * It is 32k or 64k, since value-PAGEBASE must fit in
491 * #MDB_page.%mp_upper.
493 * LMDB will use database pages < OS pages if needed.
494 * That causes more I/O in write transactions: The OS must
495 * know (read) the whole page before writing a partial page.
497 * Note that we don't currently support Huge pages. On Linux,
498 * regular data files cannot use Huge pages, and in general
499 * Huge pages aren't actually pageable. We rely on the OS
500 * demand-pager to read our data and page it out when memory
501 * pressure from other processes is high. So until OSs have
502 * actual paging support for Huge pages, they're not viable.
504 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
506 /** The minimum number of keys required in a database page.
507 * Setting this to a larger value will place a smaller bound on the
508 * maximum size of a data item. Data items larger than this size will
509 * be pushed into overflow pages instead of being stored directly in
510 * the B-tree node. This value used to default to 4. With a page size
511 * of 4096 bytes that meant that any item larger than 1024 bytes would
512 * go into an overflow page. That also meant that on average 2-3KB of
513 * each overflow page was wasted space. The value cannot be lower than
514 * 2 because then there would no longer be a tree structure. With this
515 * value, items larger than 2KB will go into overflow pages, and on
516 * average only 1KB will be wasted.
518 #define MDB_MINKEYS 2
520 /** A stamp that identifies a file as an LMDB file.
521 * There's nothing special about this value other than that it is easily
522 * recognizable, and it will reflect any byte order mismatches.
524 #define MDB_MAGIC 0xBEEFC0DE
526 /** The version number for a database's datafile format. */
527 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
528 /** The version number for a database's lockfile format. */
529 #define MDB_LOCK_VERSION 1
531 /** @brief The max size of a key we can write, or 0 for computed max.
533 * This macro should normally be left alone or set to 0.
534 * Note that a database with big keys or dupsort data cannot be
535 * reliably modified by a liblmdb which uses a smaller max.
536 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
538 * Other values are allowed, for backwards compat. However:
539 * A value bigger than the computed max can break if you do not
540 * know what you are doing, and liblmdb <= 0.9.10 can break when
541 * modifying a DB with keys/dupsort data bigger than its max.
543 * Data items in an #MDB_DUPSORT database are also limited to
544 * this size, since they're actually keys of a sub-DB. Keys and
545 * #MDB_DUPSORT data items must fit on a node in a regular page.
547 #ifndef MDB_MAXKEYSIZE
548 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
551 /** The maximum size of a key we can write to the environment. */
553 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
555 #define ENV_MAXKEY(env) ((env)->me_maxkey)
558 /** @brief The maximum size of a data item.
560 * We only store a 32 bit value for node sizes.
562 #define MAXDATASIZE 0xffffffffUL
565 /** Key size which fits in a #DKBUF.
568 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
571 * This is used for printing a hex dump of a key's contents.
573 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
574 /** Display a key in hex.
576 * Invoke a function to display a key in hex.
578 #define DKEY(x) mdb_dkey(x, kbuf)
584 /** An invalid page number.
585 * Mainly used to denote an empty tree.
587 #define P_INVALID (~(pgno_t)0)
589 /** Test if the flags \b f are set in a flag word \b w. */
590 #define F_ISSET(w, f) (((w) & (f)) == (f))
592 /** Round \b n up to an even number. */
593 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
595 /** Used for offsets within a single page.
596 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
599 typedef uint16_t indx_t;
601 /** Default size of memory map.
602 * This is certainly too small for any actual applications. Apps should always set
603 * the size explicitly using #mdb_env_set_mapsize().
605 #define DEFAULT_MAPSIZE 1048576
607 /** @defgroup readers Reader Lock Table
608 * Readers don't acquire any locks for their data access. Instead, they
609 * simply record their transaction ID in the reader table. The reader
610 * mutex is needed just to find an empty slot in the reader table. The
611 * slot's address is saved in thread-specific data so that subsequent read
612 * transactions started by the same thread need no further locking to proceed.
614 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
616 * No reader table is used if the database is on a read-only filesystem, or
617 * if #MDB_NOLOCK is set.
619 * Since the database uses multi-version concurrency control, readers don't
620 * actually need any locking. This table is used to keep track of which
621 * readers are using data from which old transactions, so that we'll know
622 * when a particular old transaction is no longer in use. Old transactions
623 * that have discarded any data pages can then have those pages reclaimed
624 * for use by a later write transaction.
626 * The lock table is constructed such that reader slots are aligned with the
627 * processor's cache line size. Any slot is only ever used by one thread.
628 * This alignment guarantees that there will be no contention or cache
629 * thrashing as threads update their own slot info, and also eliminates
630 * any need for locking when accessing a slot.
632 * A writer thread will scan every slot in the table to determine the oldest
633 * outstanding reader transaction. Any freed pages older than this will be
634 * reclaimed by the writer. The writer doesn't use any locks when scanning
635 * this table. This means that there's no guarantee that the writer will
636 * see the most up-to-date reader info, but that's not required for correct
637 * operation - all we need is to know the upper bound on the oldest reader,
638 * we don't care at all about the newest reader. So the only consequence of
639 * reading stale information here is that old pages might hang around a
640 * while longer before being reclaimed. That's actually good anyway, because
641 * the longer we delay reclaiming old pages, the more likely it is that a
642 * string of contiguous pages can be found after coalescing old pages from
643 * many old transactions together.
646 /** Number of slots in the reader table.
647 * This value was chosen somewhat arbitrarily. 126 readers plus a
648 * couple mutexes fit exactly into 8KB on my development machine.
649 * Applications should set the table size using #mdb_env_set_maxreaders().
651 #define DEFAULT_READERS 126
653 /** The size of a CPU cache line in bytes. We want our lock structures
654 * aligned to this size to avoid false cache line sharing in the
656 * This value works for most CPUs. For Itanium this should be 128.
662 /** The information we store in a single slot of the reader table.
663 * In addition to a transaction ID, we also record the process and
664 * thread ID that owns a slot, so that we can detect stale information,
665 * e.g. threads or processes that went away without cleaning up.
666 * @note We currently don't check for stale records. We simply re-init
667 * the table when we know that we're the only process opening the
670 typedef struct MDB_rxbody {
671 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
672 * Multiple readers that start at the same time will probably have the
673 * same ID here. Again, it's not important to exclude them from
674 * anything; all we need to know is which version of the DB they
675 * started from so we can avoid overwriting any data used in that
676 * particular version.
678 volatile txnid_t mrb_txnid;
679 /** The process ID of the process owning this reader txn. */
680 volatile MDB_PID_T mrb_pid;
681 /** The thread ID of the thread owning this txn. */
682 volatile MDB_THR_T mrb_tid;
685 /** The actual reader record, with cacheline padding. */
686 typedef struct MDB_reader {
689 /** shorthand for mrb_txnid */
690 #define mr_txnid mru.mrx.mrb_txnid
691 #define mr_pid mru.mrx.mrb_pid
692 #define mr_tid mru.mrx.mrb_tid
693 /** cache line alignment */
694 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
698 /** The header for the reader table.
699 * The table resides in a memory-mapped file. (This is a different file
700 * than is used for the main database.)
702 * For POSIX the actual mutexes reside in the shared memory of this
703 * mapped file. On Windows, mutexes are named objects allocated by the
704 * kernel; we store the mutex names in this mapped file so that other
705 * processes can grab them. This same approach is also used on
706 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
707 * process-shared POSIX mutexes. For these cases where a named object
708 * is used, the object name is derived from a 64 bit FNV hash of the
709 * environment pathname. As such, naming collisions are extremely
710 * unlikely. If a collision occurs, the results are unpredictable.
712 typedef struct MDB_txbody {
713 /** Stamp identifying this as an LMDB file. It must be set
716 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
718 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
719 char mtb_rmname[MNAME_LEN];
721 /** Mutex protecting access to this table.
722 * This is the reader table lock used with LOCK_MUTEX().
724 mdb_mutex_t mtb_rmutex;
726 /** The ID of the last transaction committed to the database.
727 * This is recorded here only for convenience; the value can always
728 * be determined by reading the main database meta pages.
730 volatile txnid_t mtb_txnid;
731 /** The number of slots that have been used in the reader table.
732 * This always records the maximum count, it is not decremented
733 * when readers release their slots.
735 volatile unsigned mtb_numreaders;
738 /** The actual reader table definition. */
739 typedef struct MDB_txninfo {
742 #define mti_magic mt1.mtb.mtb_magic
743 #define mti_format mt1.mtb.mtb_format
744 #define mti_rmutex mt1.mtb.mtb_rmutex
745 #define mti_rmname mt1.mtb.mtb_rmname
746 #define mti_txnid mt1.mtb.mtb_txnid
747 #define mti_numreaders mt1.mtb.mtb_numreaders
748 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
751 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
752 char mt2_wmname[MNAME_LEN];
753 #define mti_wmname mt2.mt2_wmname
755 mdb_mutex_t mt2_wmutex;
756 #define mti_wmutex mt2.mt2_wmutex
758 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
760 MDB_reader mti_readers[1];
763 /** Lockfile format signature: version, features and field layout */
764 #define MDB_LOCK_FORMAT \
766 ((MDB_LOCK_VERSION) \
767 /* Flags which describe functionality */ \
768 + (((MDB_PIDLOCK) != 0) << 16)))
771 /** Common header for all page types.
772 * Overflow records occupy a number of contiguous pages with no
773 * headers on any page after the first.
775 typedef struct MDB_page {
776 #define mp_pgno mp_p.p_pgno
777 #define mp_next mp_p.p_next
779 pgno_t p_pgno; /**< page number */
780 struct MDB_page *p_next; /**< for in-memory list of freed pages */
783 /** @defgroup mdb_page Page Flags
785 * Flags for the page headers.
788 #define P_BRANCH 0x01 /**< branch page */
789 #define P_LEAF 0x02 /**< leaf page */
790 #define P_OVERFLOW 0x04 /**< overflow page */
791 #define P_META 0x08 /**< meta page */
792 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
793 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
794 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
795 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
796 #define P_KEEP 0x8000 /**< leave this page alone during spill */
798 uint16_t mp_flags; /**< @ref mdb_page */
799 #define mp_lower mp_pb.pb.pb_lower
800 #define mp_upper mp_pb.pb.pb_upper
801 #define mp_pages mp_pb.pb_pages
804 indx_t pb_lower; /**< lower bound of free space */
805 indx_t pb_upper; /**< upper bound of free space */
807 uint32_t pb_pages; /**< number of overflow pages */
809 indx_t mp_ptrs[1]; /**< dynamic size */
812 /** Size of the page header, excluding dynamic data at the end */
813 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
815 /** Address of first usable data byte in a page, after the header */
816 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
818 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
819 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
821 /** Number of nodes on a page */
822 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
824 /** The amount of space remaining in the page */
825 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
827 /** The percentage of space used in the page, in tenths of a percent. */
828 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
829 ((env)->me_psize - PAGEHDRSZ))
830 /** The minimum page fill factor, in tenths of a percent.
831 * Pages emptier than this are candidates for merging.
833 #define FILL_THRESHOLD 250
835 /** Test if a page is a leaf page */
836 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
837 /** Test if a page is a LEAF2 page */
838 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
839 /** Test if a page is a branch page */
840 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
841 /** Test if a page is an overflow page */
842 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
843 /** Test if a page is a sub page */
844 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
846 /** The number of overflow pages needed to store the given size. */
847 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
849 /** Link in #MDB_txn.%mt_loose_pgs list */
850 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
852 /** Header for a single key/data pair within a page.
853 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
854 * We guarantee 2-byte alignment for 'MDB_node's.
856 typedef struct MDB_node {
857 /** lo and hi are used for data size on leaf nodes and for
858 * child pgno on branch nodes. On 64 bit platforms, flags
859 * is also used for pgno. (Branch nodes have no flags).
860 * They are in host byte order in case that lets some
861 * accesses be optimized into a 32-bit word access.
863 #if BYTE_ORDER == LITTLE_ENDIAN
864 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
866 unsigned short mn_hi, mn_lo;
868 /** @defgroup mdb_node Node Flags
870 * Flags for node headers.
873 #define F_BIGDATA 0x01 /**< data put on overflow page */
874 #define F_SUBDATA 0x02 /**< data is a sub-database */
875 #define F_DUPDATA 0x04 /**< data has duplicates */
877 /** valid flags for #mdb_node_add() */
878 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
881 unsigned short mn_flags; /**< @ref mdb_node */
882 unsigned short mn_ksize; /**< key size */
883 char mn_data[1]; /**< key and data are appended here */
886 /** Size of the node header, excluding dynamic data at the end */
887 #define NODESIZE offsetof(MDB_node, mn_data)
889 /** Bit position of top word in page number, for shifting mn_flags */
890 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
892 /** Size of a node in a branch page with a given key.
893 * This is just the node header plus the key, there is no data.
895 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
897 /** Size of a node in a leaf page with a given key and data.
898 * This is node header plus key plus data size.
900 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
902 /** Address of node \b i in page \b p */
903 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
905 /** Address of the key for the node */
906 #define NODEKEY(node) (void *)((node)->mn_data)
908 /** Address of the data for a node */
909 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
911 /** Get the page number pointed to by a branch node */
912 #define NODEPGNO(node) \
913 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
914 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
915 /** Set the page number in a branch node */
916 #define SETPGNO(node,pgno) do { \
917 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
918 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
920 /** Get the size of the data in a leaf node */
921 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
922 /** Set the size of the data for a leaf node */
923 #define SETDSZ(node,size) do { \
924 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
925 /** The size of a key in a node */
926 #define NODEKSZ(node) ((node)->mn_ksize)
928 /** Copy a page number from src to dst */
930 #define COPY_PGNO(dst,src) dst = src
932 #if SIZE_MAX > 4294967295UL
933 #define COPY_PGNO(dst,src) do { \
934 unsigned short *s, *d; \
935 s = (unsigned short *)&(src); \
936 d = (unsigned short *)&(dst); \
943 #define COPY_PGNO(dst,src) do { \
944 unsigned short *s, *d; \
945 s = (unsigned short *)&(src); \
946 d = (unsigned short *)&(dst); \
952 /** The address of a key in a LEAF2 page.
953 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
954 * There are no node headers, keys are stored contiguously.
956 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
958 /** Set the \b node's key into \b keyptr, if requested. */
959 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
960 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
962 /** Set the \b node's key into \b key. */
963 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
965 /** Information about a single database in the environment. */
966 typedef struct MDB_db {
967 uint32_t md_pad; /**< also ksize for LEAF2 pages */
968 uint16_t md_flags; /**< @ref mdb_dbi_open */
969 uint16_t md_depth; /**< depth of this tree */
970 pgno_t md_branch_pages; /**< number of internal pages */
971 pgno_t md_leaf_pages; /**< number of leaf pages */
972 pgno_t md_overflow_pages; /**< number of overflow pages */
973 size_t md_entries; /**< number of data items */
974 pgno_t md_root; /**< the root page of this tree */
977 /** mdb_dbi_open flags */
978 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
979 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
980 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
981 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
983 /** Handle for the DB used to track free pages. */
985 /** Handle for the default DB. */
987 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
990 /** Number of meta pages - also hardcoded elsewhere */
993 /** Meta page content.
994 * A meta page is the start point for accessing a database snapshot.
995 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
997 typedef struct MDB_meta {
998 /** Stamp identifying this as an LMDB file. It must be set
1001 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1002 uint32_t mm_version;
1003 void *mm_address; /**< address for fixed mapping */
1004 size_t mm_mapsize; /**< size of mmap region */
1005 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1006 /** The size of pages used in this DB */
1007 #define mm_psize mm_dbs[FREE_DBI].md_pad
1008 /** Any persistent environment flags. @ref mdb_env */
1009 #define mm_flags mm_dbs[FREE_DBI].md_flags
1010 pgno_t mm_last_pg; /**< last used page in file */
1011 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1014 /** Buffer for a stack-allocated meta page.
1015 * The members define size and alignment, and silence type
1016 * aliasing warnings. They are not used directly; that could
1017 * mean incorrectly using several union members in parallel.
1019 typedef union MDB_metabuf {
1022 char mm_pad[PAGEHDRSZ];
1027 /** Auxiliary DB info.
1028 * The information here is mostly static/read-only. There is
1029 * only a single copy of this record in the environment.
1031 typedef struct MDB_dbx {
1032 MDB_val md_name; /**< name of the database */
1033 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1034 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1035 MDB_rel_func *md_rel; /**< user relocate function */
1036 void *md_relctx; /**< user-provided context for md_rel */
1039 /** A database transaction.
1040 * Every operation requires a transaction handle.
1043 MDB_txn *mt_parent; /**< parent of a nested txn */
1044 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1046 pgno_t mt_next_pgno; /**< next unallocated page */
1047 /** The ID of this transaction. IDs are integers incrementing from 1.
1048 * Only committed write transactions increment the ID. If a transaction
1049 * aborts, the ID may be re-used by the next writer.
1052 MDB_env *mt_env; /**< the DB environment */
1053 /** The list of pages that became unused during this transaction.
1055 MDB_IDL mt_free_pgs;
1056 /** The list of loose pages that became unused and may be reused
1057 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1059 MDB_page *mt_loose_pgs;
1060 /* #Number of loose pages (#mt_loose_pgs) */
1062 /** The sorted list of dirty pages we temporarily wrote to disk
1063 * because the dirty list was full. page numbers in here are
1064 * shifted left by 1, deleted slots have the LSB set.
1066 MDB_IDL mt_spill_pgs;
1068 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1069 MDB_ID2L dirty_list;
1070 /** For read txns: This thread/txn's reader table slot, or NULL. */
1073 /** Array of records for each DB known in the environment. */
1075 /** Array of MDB_db records for each known DB */
1077 /** Array of sequence numbers for each DB handle */
1078 unsigned int *mt_dbiseqs;
1079 /** @defgroup mt_dbflag Transaction DB Flags
1083 #define DB_DIRTY 0x01 /**< DB was written in this txn */
1084 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1085 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1086 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1087 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1088 #define DB_DUPDATA 0x20 /**< DB is #MDB_DUPSORT data */
1090 /** In write txns, array of cursors for each DB */
1091 MDB_cursor **mt_cursors;
1092 /** Array of flags for each DB */
1093 unsigned char *mt_dbflags;
1094 /** Number of DB records in use, or 0 when the txn is finished.
1095 * This number only ever increments until the txn finishes; we
1096 * don't decrement it when individual DB handles are closed.
1100 /** @defgroup mdb_txn Transaction Flags
1104 /** #mdb_txn_begin() flags */
1105 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1106 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1107 /* internal txn flags */
1108 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1109 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1110 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1111 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1112 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1113 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1114 /** most operations on the txn are currently illegal */
1115 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1117 unsigned int mt_flags; /**< @ref mdb_txn */
1118 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1119 * Includes ancestor txns' dirty pages not hidden by other txns'
1120 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1121 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1123 unsigned int mt_dirty_room;
1126 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1127 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1128 * raise this on a 64 bit machine.
1130 #define CURSOR_STACK 32
1134 /** Cursors are used for all DB operations.
1135 * A cursor holds a path of (page pointer, key index) from the DB
1136 * root to a position in the DB, plus other state. #MDB_DUPSORT
1137 * cursors include an xcursor to the current data item. Write txns
1138 * track their cursors and keep them up to date when data moves.
1139 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1140 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1143 /** Next cursor on this DB in this txn */
1144 MDB_cursor *mc_next;
1145 /** Backup of the original cursor if this cursor is a shadow */
1146 MDB_cursor *mc_backup;
1147 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1148 struct MDB_xcursor *mc_xcursor;
1149 /** The transaction that owns this cursor */
1151 /** The database handle this cursor operates on */
1153 /** The database record for this cursor */
1155 /** The database auxiliary record for this cursor */
1157 /** The @ref mt_dbflag for this database */
1158 unsigned char *mc_dbflag;
1159 unsigned short mc_snum; /**< number of pushed pages */
1160 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1161 /** @defgroup mdb_cursor Cursor Flags
1163 * Cursor state flags.
1166 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1167 #define C_EOF 0x02 /**< No more data */
1168 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1169 #define C_DEL 0x08 /**< last op was a cursor_del */
1170 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1172 unsigned int mc_flags; /**< @ref mdb_cursor */
1173 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1174 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1177 /** Context for sorted-dup records.
1178 * We could have gone to a fully recursive design, with arbitrarily
1179 * deep nesting of sub-databases. But for now we only handle these
1180 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1182 typedef struct MDB_xcursor {
1183 /** A sub-cursor for traversing the Dup DB */
1184 MDB_cursor mx_cursor;
1185 /** The database record for this Dup DB */
1187 /** The auxiliary DB record for this Dup DB */
1189 /** The @ref mt_dbflag for this Dup DB */
1190 unsigned char mx_dbflag;
1193 /** Check if there is an inited xcursor, so #XCURSOR_REFRESH() is proper */
1194 #define XCURSOR_INITED(mc) \
1195 ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
1197 /** Update sub-page pointer, if any, in \b mc->mc_xcursor. Needed
1198 * when the node which contains the sub-page may have moved. Called
1199 * with \b mp = mc->mc_pg[mc->mc_top], \b ki = mc->mc_ki[mc->mc_top].
1201 #define XCURSOR_REFRESH(mc, mp, ki) do { \
1202 MDB_page *xr_pg = (mp); \
1203 MDB_node *xr_node = NODEPTR(xr_pg, ki); \
1204 if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \
1205 (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \
1208 /** State of FreeDB old pages, stored in the MDB_env */
1209 typedef struct MDB_pgstate {
1210 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1211 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1214 /** The database environment. */
1216 HANDLE me_fd; /**< The main data file */
1217 HANDLE me_lfd; /**< The lock file */
1218 HANDLE me_mfd; /**< just for writing the meta pages */
1219 /** Failed to update the meta page. Probably an I/O error. */
1220 #define MDB_FATAL_ERROR 0x80000000U
1221 /** Some fields are initialized. */
1222 #define MDB_ENV_ACTIVE 0x20000000U
1223 /** me_txkey is set */
1224 #define MDB_ENV_TXKEY 0x10000000U
1225 /** fdatasync is unreliable */
1226 #define MDB_FSYNCONLY 0x08000000U
1227 uint32_t me_flags; /**< @ref mdb_env */
1228 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1229 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1230 unsigned int me_maxreaders; /**< size of the reader table */
1231 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1232 volatile int me_close_readers;
1233 MDB_dbi me_numdbs; /**< number of DBs opened */
1234 MDB_dbi me_maxdbs; /**< size of the DB table */
1235 MDB_PID_T me_pid; /**< process ID of this env */
1236 char *me_path; /**< path to the DB files */
1237 char *me_map; /**< the memory map of the data file */
1238 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1239 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1240 void *me_pbuf; /**< scratch area for DUPSORT put() */
1241 MDB_txn *me_txn; /**< current write transaction */
1242 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1243 size_t me_mapsize; /**< size of the data memory map */
1244 off_t me_size; /**< current file size */
1245 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1246 MDB_dbx *me_dbxs; /**< array of static DB info */
1247 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1248 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1249 pthread_key_t me_txkey; /**< thread-key for readers */
1250 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1251 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1252 # define me_pglast me_pgstate.mf_pglast
1253 # define me_pghead me_pgstate.mf_pghead
1254 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1255 /** IDL of pages that became unused in a write txn */
1256 MDB_IDL me_free_pgs;
1257 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1258 MDB_ID2L me_dirty_list;
1259 /** Max number of freelist items that can fit in a single overflow page */
1261 /** Max size of a node on a page */
1262 unsigned int me_nodemax;
1263 #if !(MDB_MAXKEYSIZE)
1264 unsigned int me_maxkey; /**< max size of a key */
1266 int me_live_reader; /**< have liveness lock in reader table */
1268 int me_pidquery; /**< Used in OpenProcess */
1270 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1271 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1272 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1274 mdb_mutex_t me_rmutex;
1275 mdb_mutex_t me_wmutex;
1277 void *me_userctx; /**< User-settable context */
1278 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1281 /** Nested transaction */
1282 typedef struct MDB_ntxn {
1283 MDB_txn mnt_txn; /**< the transaction */
1284 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1287 /** max number of pages to commit in one writev() call */
1288 #define MDB_COMMIT_PAGES 64
1289 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1290 #undef MDB_COMMIT_PAGES
1291 #define MDB_COMMIT_PAGES IOV_MAX
1294 /** max bytes to write in one call */
1295 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1297 /** Check \b txn and \b dbi arguments to a function */
1298 #define TXN_DBI_EXIST(txn, dbi, validity) \
1299 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1301 /** Check for misused \b dbi handles */
1302 #define TXN_DBI_CHANGED(txn, dbi) \
1303 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1305 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1306 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1307 static int mdb_page_touch(MDB_cursor *mc);
1309 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1310 "reset-tmp", "fail-begin", "fail-beginchild"}
1312 /* mdb_txn_end operation number, for logging */
1313 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1314 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1316 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1317 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1318 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1319 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1320 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1322 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1323 static int mdb_page_search_root(MDB_cursor *mc,
1324 MDB_val *key, int modify);
1325 #define MDB_PS_MODIFY 1
1326 #define MDB_PS_ROOTONLY 2
1327 #define MDB_PS_FIRST 4
1328 #define MDB_PS_LAST 8
1329 static int mdb_page_search(MDB_cursor *mc,
1330 MDB_val *key, int flags);
1331 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1333 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1334 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1335 pgno_t newpgno, unsigned int nflags);
1337 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1338 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1339 static int mdb_env_write_meta(MDB_txn *txn);
1340 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1341 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1343 static void mdb_env_close0(MDB_env *env, int excl);
1345 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1346 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1347 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1348 static void mdb_node_del(MDB_cursor *mc, int ksize);
1349 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1350 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1351 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1352 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1353 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1355 static int mdb_rebalance(MDB_cursor *mc);
1356 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1358 static void mdb_cursor_pop(MDB_cursor *mc);
1359 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1361 static int mdb_cursor_del0(MDB_cursor *mc);
1362 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1363 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1364 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1365 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1366 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1368 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1369 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1371 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1372 static void mdb_xcursor_init0(MDB_cursor *mc);
1373 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1374 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1376 static int mdb_drop0(MDB_cursor *mc, int subs);
1377 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1378 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1381 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1384 /** Compare two items pointing at size_t's of unknown alignment. */
1385 #ifdef MISALIGNED_OK
1386 # define mdb_cmp_clong mdb_cmp_long
1388 # define mdb_cmp_clong mdb_cmp_cint
1392 static SECURITY_DESCRIPTOR mdb_null_sd;
1393 static SECURITY_ATTRIBUTES mdb_all_sa;
1394 static int mdb_sec_inited;
1396 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize);
1399 /** Return the library version info. */
1401 mdb_version(int *major, int *minor, int *patch)
1403 if (major) *major = MDB_VERSION_MAJOR;
1404 if (minor) *minor = MDB_VERSION_MINOR;
1405 if (patch) *patch = MDB_VERSION_PATCH;
1406 return MDB_VERSION_STRING;
1409 /** Table of descriptions for LMDB @ref errors */
1410 static char *const mdb_errstr[] = {
1411 "MDB_KEYEXIST: Key/data pair already exists",
1412 "MDB_NOTFOUND: No matching key/data pair found",
1413 "MDB_PAGE_NOTFOUND: Requested page not found",
1414 "MDB_CORRUPTED: Located page was wrong type",
1415 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1416 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1417 "MDB_INVALID: File is not an LMDB file",
1418 "MDB_MAP_FULL: Environment mapsize limit reached",
1419 "MDB_DBS_FULL: Environment maxdbs limit reached",
1420 "MDB_READERS_FULL: Environment maxreaders limit reached",
1421 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1422 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1423 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1424 "MDB_PAGE_FULL: Internal error - page has no more space",
1425 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1426 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1427 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1428 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1429 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1430 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1434 mdb_strerror(int err)
1437 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1438 * This works as long as no function between the call to mdb_strerror
1439 * and the actual use of the message uses more than 4K of stack.
1441 #define MSGSIZE 1024
1442 #define PADSIZE 4096
1443 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1447 return ("Successful return: 0");
1449 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1450 i = err - MDB_KEYEXIST;
1451 return mdb_errstr[i];
1455 /* These are the C-runtime error codes we use. The comment indicates
1456 * their numeric value, and the Win32 error they would correspond to
1457 * if the error actually came from a Win32 API. A major mess, we should
1458 * have used LMDB-specific error codes for everything.
1461 case ENOENT: /* 2, FILE_NOT_FOUND */
1462 case EIO: /* 5, ACCESS_DENIED */
1463 case ENOMEM: /* 12, INVALID_ACCESS */
1464 case EACCES: /* 13, INVALID_DATA */
1465 case EBUSY: /* 16, CURRENT_DIRECTORY */
1466 case EINVAL: /* 22, BAD_COMMAND */
1467 case ENOSPC: /* 28, OUT_OF_PAPER */
1468 return strerror(err);
1473 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1474 FORMAT_MESSAGE_IGNORE_INSERTS,
1475 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1478 return strerror(err);
1482 /** assert(3) variant in cursor context */
1483 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1484 /** assert(3) variant in transaction context */
1485 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1486 /** assert(3) variant in environment context */
1487 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1490 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1491 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1494 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1495 const char *func, const char *file, int line)
1498 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1499 file, line, expr_txt, func);
1500 if (env->me_assert_func)
1501 env->me_assert_func(env, buf);
1502 fprintf(stderr, "%s\n", buf);
1506 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1510 /** Return the page number of \b mp which may be sub-page, for debug output */
1512 mdb_dbg_pgno(MDB_page *mp)
1515 COPY_PGNO(ret, mp->mp_pgno);
1519 /** Display a key in hexadecimal and return the address of the result.
1520 * @param[in] key the key to display
1521 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1522 * @return The key in hexadecimal form.
1525 mdb_dkey(MDB_val *key, char *buf)
1528 unsigned char *c = key->mv_data;
1534 if (key->mv_size > DKBUF_MAXKEYSIZE)
1535 return "MDB_MAXKEYSIZE";
1536 /* may want to make this a dynamic check: if the key is mostly
1537 * printable characters, print it as-is instead of converting to hex.
1541 for (i=0; i<key->mv_size; i++)
1542 ptr += sprintf(ptr, "%02x", *c++);
1544 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1550 mdb_leafnode_type(MDB_node *n)
1552 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1553 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1554 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1557 /** Display all the keys in the page. */
1559 mdb_page_list(MDB_page *mp)
1561 pgno_t pgno = mdb_dbg_pgno(mp);
1562 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1564 unsigned int i, nkeys, nsize, total = 0;
1568 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1569 case P_BRANCH: type = "Branch page"; break;
1570 case P_LEAF: type = "Leaf page"; break;
1571 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1572 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1573 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1575 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1576 pgno, mp->mp_pages, state);
1579 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1580 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1583 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1587 nkeys = NUMKEYS(mp);
1588 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1590 for (i=0; i<nkeys; i++) {
1591 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1592 key.mv_size = nsize = mp->mp_pad;
1593 key.mv_data = LEAF2KEY(mp, i, nsize);
1595 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1598 node = NODEPTR(mp, i);
1599 key.mv_size = node->mn_ksize;
1600 key.mv_data = node->mn_data;
1601 nsize = NODESIZE + key.mv_size;
1602 if (IS_BRANCH(mp)) {
1603 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1607 if (F_ISSET(node->mn_flags, F_BIGDATA))
1608 nsize += sizeof(pgno_t);
1610 nsize += NODEDSZ(node);
1612 nsize += sizeof(indx_t);
1613 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1614 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1616 total = EVEN(total);
1618 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1619 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1623 mdb_cursor_chk(MDB_cursor *mc)
1629 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1630 for (i=0; i<mc->mc_top; i++) {
1632 node = NODEPTR(mp, mc->mc_ki[i]);
1633 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1636 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1638 if (XCURSOR_INITED(mc)) {
1639 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1640 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1641 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1649 /** Count all the pages in each DB and in the freelist
1650 * and make sure it matches the actual number of pages
1652 * All named DBs must be open for a correct count.
1654 static void mdb_audit(MDB_txn *txn)
1658 MDB_ID freecount, count;
1663 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1664 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1665 freecount += *(MDB_ID *)data.mv_data;
1666 mdb_tassert(txn, rc == MDB_NOTFOUND);
1669 for (i = 0; i<txn->mt_numdbs; i++) {
1671 if (!(txn->mt_dbflags[i] & DB_VALID))
1673 mdb_cursor_init(&mc, txn, i, &mx);
1674 if (txn->mt_dbs[i].md_root == P_INVALID)
1676 count += txn->mt_dbs[i].md_branch_pages +
1677 txn->mt_dbs[i].md_leaf_pages +
1678 txn->mt_dbs[i].md_overflow_pages;
1679 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1680 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1681 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1684 mp = mc.mc_pg[mc.mc_top];
1685 for (j=0; j<NUMKEYS(mp); j++) {
1686 MDB_node *leaf = NODEPTR(mp, j);
1687 if (leaf->mn_flags & F_SUBDATA) {
1689 memcpy(&db, NODEDATA(leaf), sizeof(db));
1690 count += db.md_branch_pages + db.md_leaf_pages +
1691 db.md_overflow_pages;
1695 mdb_tassert(txn, rc == MDB_NOTFOUND);
1698 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1699 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1700 txn->mt_txnid, freecount, count+NUM_METAS,
1701 freecount+count+NUM_METAS, txn->mt_next_pgno);
1707 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1709 return txn->mt_dbxs[dbi].md_cmp(a, b);
1713 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1715 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1716 #if UINT_MAX < SIZE_MAX
1717 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1718 dcmp = mdb_cmp_clong;
1723 /** Allocate memory for a page.
1724 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1727 mdb_page_malloc(MDB_txn *txn, unsigned num)
1729 MDB_env *env = txn->mt_env;
1730 MDB_page *ret = env->me_dpages;
1731 size_t psize = env->me_psize, sz = psize, off;
1732 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1733 * For a single page alloc, we init everything after the page header.
1734 * For multi-page, we init the final page; if the caller needed that
1735 * many pages they will be filling in at least up to the last page.
1739 VGMEMP_ALLOC(env, ret, sz);
1740 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1741 env->me_dpages = ret->mp_next;
1744 psize -= off = PAGEHDRSZ;
1749 if ((ret = malloc(sz)) != NULL) {
1750 VGMEMP_ALLOC(env, ret, sz);
1751 if (!(env->me_flags & MDB_NOMEMINIT)) {
1752 memset((char *)ret + off, 0, psize);
1756 txn->mt_flags |= MDB_TXN_ERROR;
1760 /** Free a single page.
1761 * Saves single pages to a list, for future reuse.
1762 * (This is not used for multi-page overflow pages.)
1765 mdb_page_free(MDB_env *env, MDB_page *mp)
1767 mp->mp_next = env->me_dpages;
1768 VGMEMP_FREE(env, mp);
1769 env->me_dpages = mp;
1772 /** Free a dirty page */
1774 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1776 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1777 mdb_page_free(env, dp);
1779 /* large pages just get freed directly */
1780 VGMEMP_FREE(env, dp);
1785 /** Return all dirty pages to dpage list */
1787 mdb_dlist_free(MDB_txn *txn)
1789 MDB_env *env = txn->mt_env;
1790 MDB_ID2L dl = txn->mt_u.dirty_list;
1791 unsigned i, n = dl[0].mid;
1793 for (i = 1; i <= n; i++) {
1794 mdb_dpage_free(env, dl[i].mptr);
1799 /** Loosen or free a single page.
1800 * Saves single pages to a list for future reuse
1801 * in this same txn. It has been pulled from the freeDB
1802 * and already resides on the dirty list, but has been
1803 * deleted. Use these pages first before pulling again
1806 * If the page wasn't dirtied in this txn, just add it
1807 * to this txn's free list.
1810 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1813 pgno_t pgno = mp->mp_pgno;
1814 MDB_txn *txn = mc->mc_txn;
1816 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1817 if (txn->mt_parent) {
1818 MDB_ID2 *dl = txn->mt_u.dirty_list;
1819 /* If txn has a parent, make sure the page is in our
1823 unsigned x = mdb_mid2l_search(dl, pgno);
1824 if (x <= dl[0].mid && dl[x].mid == pgno) {
1825 if (mp != dl[x].mptr) { /* bad cursor? */
1826 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1827 txn->mt_flags |= MDB_TXN_ERROR;
1828 return MDB_CORRUPTED;
1835 /* no parent txn, so it's just ours */
1840 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1842 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1843 txn->mt_loose_pgs = mp;
1844 txn->mt_loose_count++;
1845 mp->mp_flags |= P_LOOSE;
1847 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1855 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1856 * @param[in] mc A cursor handle for the current operation.
1857 * @param[in] pflags Flags of the pages to update:
1858 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1859 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1860 * @return 0 on success, non-zero on failure.
1863 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1865 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1866 MDB_txn *txn = mc->mc_txn;
1867 MDB_cursor *m3, *m0 = mc;
1872 int rc = MDB_SUCCESS, level;
1874 /* Mark pages seen by cursors */
1875 if (mc->mc_flags & C_UNTRACK)
1876 mc = NULL; /* will find mc in mt_cursors */
1877 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1878 for (; mc; mc=mc->mc_next) {
1879 if (!(mc->mc_flags & C_INITIALIZED))
1881 for (m3 = mc;; m3 = &mx->mx_cursor) {
1883 for (j=0; j<m3->mc_snum; j++) {
1885 if ((mp->mp_flags & Mask) == pflags)
1886 mp->mp_flags ^= P_KEEP;
1888 mx = m3->mc_xcursor;
1889 /* Proceed to mx if it is at a sub-database */
1890 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1892 if (! (mp && (mp->mp_flags & P_LEAF)))
1894 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1895 if (!(leaf->mn_flags & F_SUBDATA))
1904 /* Mark dirty root pages */
1905 for (i=0; i<txn->mt_numdbs; i++) {
1906 if (txn->mt_dbflags[i] & DB_DIRTY) {
1907 pgno_t pgno = txn->mt_dbs[i].md_root;
1908 if (pgno == P_INVALID)
1910 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
1912 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1913 dp->mp_flags ^= P_KEEP;
1921 static int mdb_page_flush(MDB_txn *txn, int keep);
1923 /** Spill pages from the dirty list back to disk.
1924 * This is intended to prevent running into #MDB_TXN_FULL situations,
1925 * but note that they may still occur in a few cases:
1926 * 1) our estimate of the txn size could be too small. Currently this
1927 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1928 * 2) child txns may run out of space if their parents dirtied a
1929 * lot of pages and never spilled them. TODO: we probably should do
1930 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1931 * the parent's dirty_room is below a given threshold.
1933 * Otherwise, if not using nested txns, it is expected that apps will
1934 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1935 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1936 * If the txn never references them again, they can be left alone.
1937 * If the txn only reads them, they can be used without any fuss.
1938 * If the txn writes them again, they can be dirtied immediately without
1939 * going thru all of the work of #mdb_page_touch(). Such references are
1940 * handled by #mdb_page_unspill().
1942 * Also note, we never spill DB root pages, nor pages of active cursors,
1943 * because we'll need these back again soon anyway. And in nested txns,
1944 * we can't spill a page in a child txn if it was already spilled in a
1945 * parent txn. That would alter the parent txns' data even though
1946 * the child hasn't committed yet, and we'd have no way to undo it if
1947 * the child aborted.
1949 * @param[in] m0 cursor A cursor handle identifying the transaction and
1950 * database for which we are checking space.
1951 * @param[in] key For a put operation, the key being stored.
1952 * @param[in] data For a put operation, the data being stored.
1953 * @return 0 on success, non-zero on failure.
1956 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1958 MDB_txn *txn = m0->mc_txn;
1960 MDB_ID2L dl = txn->mt_u.dirty_list;
1961 unsigned int i, j, need;
1964 if (m0->mc_flags & C_SUB)
1967 /* Estimate how much space this op will take */
1968 i = m0->mc_db->md_depth;
1969 /* Named DBs also dirty the main DB */
1970 if (m0->mc_dbi >= CORE_DBS)
1971 i += txn->mt_dbs[MAIN_DBI].md_depth;
1972 /* For puts, roughly factor in the key+data size */
1974 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1975 i += i; /* double it for good measure */
1978 if (txn->mt_dirty_room > i)
1981 if (!txn->mt_spill_pgs) {
1982 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1983 if (!txn->mt_spill_pgs)
1986 /* purge deleted slots */
1987 MDB_IDL sl = txn->mt_spill_pgs;
1988 unsigned int num = sl[0];
1990 for (i=1; i<=num; i++) {
1997 /* Preserve pages which may soon be dirtied again */
1998 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2001 /* Less aggressive spill - we originally spilled the entire dirty list,
2002 * with a few exceptions for cursor pages and DB root pages. But this
2003 * turns out to be a lot of wasted effort because in a large txn many
2004 * of those pages will need to be used again. So now we spill only 1/8th
2005 * of the dirty pages. Testing revealed this to be a good tradeoff,
2006 * better than 1/2, 1/4, or 1/10.
2008 if (need < MDB_IDL_UM_MAX / 8)
2009 need = MDB_IDL_UM_MAX / 8;
2011 /* Save the page IDs of all the pages we're flushing */
2012 /* flush from the tail forward, this saves a lot of shifting later on. */
2013 for (i=dl[0].mid; i && need; i--) {
2014 MDB_ID pn = dl[i].mid << 1;
2016 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2018 /* Can't spill twice, make sure it's not already in a parent's
2021 if (txn->mt_parent) {
2023 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2024 if (tx2->mt_spill_pgs) {
2025 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2026 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2027 dp->mp_flags |= P_KEEP;
2035 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2039 mdb_midl_sort(txn->mt_spill_pgs);
2041 /* Flush the spilled part of dirty list */
2042 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2045 /* Reset any dirty pages we kept that page_flush didn't see */
2046 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2049 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2053 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2055 mdb_find_oldest(MDB_txn *txn)
2058 txnid_t mr, oldest = txn->mt_txnid - 1;
2059 if (txn->mt_env->me_txns) {
2060 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2061 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2072 /** Add a page to the txn's dirty list */
2074 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2077 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2079 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2080 insert = mdb_mid2l_append;
2082 insert = mdb_mid2l_insert;
2084 mid.mid = mp->mp_pgno;
2086 rc = insert(txn->mt_u.dirty_list, &mid);
2087 mdb_tassert(txn, rc == 0);
2088 txn->mt_dirty_room--;
2091 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2092 * me_pghead and mt_next_pgno.
2094 * If there are free pages available from older transactions, they
2095 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2096 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2097 * and move me_pglast to say which records were consumed. Only this
2098 * function can create me_pghead and move me_pglast/mt_next_pgno.
2099 * @param[in] mc cursor A cursor handle identifying the transaction and
2100 * database for which we are allocating.
2101 * @param[in] num the number of pages to allocate.
2102 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2103 * will always be satisfied by a single contiguous chunk of memory.
2104 * @return 0 on success, non-zero on failure.
2107 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2109 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2110 /* Get at most <Max_retries> more freeDB records once me_pghead
2111 * has enough pages. If not enough, use new pages from the map.
2112 * If <Paranoid> and mc is updating the freeDB, only get new
2113 * records if me_pghead is empty. Then the freelist cannot play
2114 * catch-up with itself by growing while trying to save it.
2116 enum { Paranoid = 1, Max_retries = 500 };
2118 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2120 int rc, retry = num * 60;
2121 MDB_txn *txn = mc->mc_txn;
2122 MDB_env *env = txn->mt_env;
2123 pgno_t pgno, *mop = env->me_pghead;
2124 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2126 txnid_t oldest = 0, last;
2131 /* If there are any loose pages, just use them */
2132 if (num == 1 && txn->mt_loose_pgs) {
2133 np = txn->mt_loose_pgs;
2134 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2135 txn->mt_loose_count--;
2136 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2144 /* If our dirty list is already full, we can't do anything */
2145 if (txn->mt_dirty_room == 0) {
2150 for (op = MDB_FIRST;; op = MDB_NEXT) {
2155 /* Seek a big enough contiguous page range. Prefer
2156 * pages at the tail, just truncating the list.
2162 if (mop[i-n2] == pgno+n2)
2169 if (op == MDB_FIRST) { /* 1st iteration */
2170 /* Prepare to fetch more and coalesce */
2171 last = env->me_pglast;
2172 oldest = env->me_pgoldest;
2173 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2176 key.mv_data = &last; /* will look up last+1 */
2177 key.mv_size = sizeof(last);
2179 if (Paranoid && mc->mc_dbi == FREE_DBI)
2182 if (Paranoid && retry < 0 && mop_len)
2186 /* Do not fetch more if the record will be too recent */
2187 if (oldest <= last) {
2189 oldest = mdb_find_oldest(txn);
2190 env->me_pgoldest = oldest;
2196 rc = mdb_cursor_get(&m2, &key, NULL, op);
2198 if (rc == MDB_NOTFOUND)
2202 last = *(txnid_t*)key.mv_data;
2203 if (oldest <= last) {
2205 oldest = mdb_find_oldest(txn);
2206 env->me_pgoldest = oldest;
2212 np = m2.mc_pg[m2.mc_top];
2213 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2214 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2217 idl = (MDB_ID *) data.mv_data;
2220 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2225 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2227 mop = env->me_pghead;
2229 env->me_pglast = last;
2231 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2232 last, txn->mt_dbs[FREE_DBI].md_root, i));
2234 DPRINTF(("IDL %"Z"u", idl[j]));
2236 /* Merge in descending sorted order */
2237 mdb_midl_xmerge(mop, idl);
2241 /* Use new pages from the map when nothing suitable in the freeDB */
2243 pgno = txn->mt_next_pgno;
2244 if (pgno + num >= env->me_maxpg) {
2245 DPUTS("DB size maxed out");
2251 if (env->me_flags & MDB_WRITEMAP) {
2252 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2254 if (!(np = mdb_page_malloc(txn, num))) {
2260 mop[0] = mop_len -= num;
2261 /* Move any stragglers down */
2262 for (j = i-num; j < mop_len; )
2263 mop[++j] = mop[++i];
2265 txn->mt_next_pgno = pgno + num;
2268 mdb_page_dirty(txn, np);
2274 txn->mt_flags |= MDB_TXN_ERROR;
2278 /** Copy the used portions of a non-overflow page.
2279 * @param[in] dst page to copy into
2280 * @param[in] src page to copy from
2281 * @param[in] psize size of a page
2284 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2286 enum { Align = sizeof(pgno_t) };
2287 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2289 /* If page isn't full, just copy the used portion. Adjust
2290 * alignment so memcpy may copy words instead of bytes.
2292 if ((unused &= -Align) && !IS_LEAF2(src)) {
2293 upper = (upper + PAGEBASE) & -Align;
2294 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2295 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2298 memcpy(dst, src, psize - unused);
2302 /** Pull a page off the txn's spill list, if present.
2303 * If a page being referenced was spilled to disk in this txn, bring
2304 * it back and make it dirty/writable again.
2305 * @param[in] txn the transaction handle.
2306 * @param[in] mp the page being referenced. It must not be dirty.
2307 * @param[out] ret the writable page, if any. ret is unchanged if
2308 * mp wasn't spilled.
2311 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2313 MDB_env *env = txn->mt_env;
2316 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2318 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2319 if (!tx2->mt_spill_pgs)
2321 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2322 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2325 if (txn->mt_dirty_room == 0)
2326 return MDB_TXN_FULL;
2327 if (IS_OVERFLOW(mp))
2331 if (env->me_flags & MDB_WRITEMAP) {
2334 np = mdb_page_malloc(txn, num);
2338 memcpy(np, mp, num * env->me_psize);
2340 mdb_page_copy(np, mp, env->me_psize);
2343 /* If in current txn, this page is no longer spilled.
2344 * If it happens to be the last page, truncate the spill list.
2345 * Otherwise mark it as deleted by setting the LSB.
2347 if (x == txn->mt_spill_pgs[0])
2348 txn->mt_spill_pgs[0]--;
2350 txn->mt_spill_pgs[x] |= 1;
2351 } /* otherwise, if belonging to a parent txn, the
2352 * page remains spilled until child commits
2355 mdb_page_dirty(txn, np);
2356 np->mp_flags |= P_DIRTY;
2364 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2365 * @param[in] mc cursor pointing to the page to be touched
2366 * @return 0 on success, non-zero on failure.
2369 mdb_page_touch(MDB_cursor *mc)
2371 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2372 MDB_txn *txn = mc->mc_txn;
2373 MDB_cursor *m2, *m3;
2377 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2378 if (txn->mt_flags & MDB_TXN_SPILLS) {
2380 rc = mdb_page_unspill(txn, mp, &np);
2386 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2387 (rc = mdb_page_alloc(mc, 1, &np)))
2390 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2391 mp->mp_pgno, pgno));
2392 mdb_cassert(mc, mp->mp_pgno != pgno);
2393 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2394 /* Update the parent page, if any, to point to the new page */
2396 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2397 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2398 SETPGNO(node, pgno);
2400 mc->mc_db->md_root = pgno;
2402 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2403 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2405 /* If txn has a parent, make sure the page is in our
2409 unsigned x = mdb_mid2l_search(dl, pgno);
2410 if (x <= dl[0].mid && dl[x].mid == pgno) {
2411 if (mp != dl[x].mptr) { /* bad cursor? */
2412 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2413 txn->mt_flags |= MDB_TXN_ERROR;
2414 return MDB_CORRUPTED;
2419 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2421 np = mdb_page_malloc(txn, 1);
2426 rc = mdb_mid2l_insert(dl, &mid);
2427 mdb_cassert(mc, rc == 0);
2432 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2434 np->mp_flags |= P_DIRTY;
2437 /* Adjust cursors pointing to mp */
2438 mc->mc_pg[mc->mc_top] = np;
2439 m2 = txn->mt_cursors[mc->mc_dbi];
2440 if (mc->mc_flags & C_SUB) {
2441 for (; m2; m2=m2->mc_next) {
2442 m3 = &m2->mc_xcursor->mx_cursor;
2443 if (m3->mc_snum < mc->mc_snum) continue;
2444 if (m3->mc_pg[mc->mc_top] == mp)
2445 m3->mc_pg[mc->mc_top] = np;
2448 for (; m2; m2=m2->mc_next) {
2449 if (m2->mc_snum < mc->mc_snum) continue;
2450 if (m2 == mc) continue;
2451 if (m2->mc_pg[mc->mc_top] == mp) {
2452 m2->mc_pg[mc->mc_top] = np;
2453 if (XCURSOR_INITED(m2) && IS_LEAF(np))
2454 XCURSOR_REFRESH(m2, np, m2->mc_ki[mc->mc_top]);
2461 txn->mt_flags |= MDB_TXN_ERROR;
2466 mdb_env_sync(MDB_env *env, int force)
2469 if (env->me_flags & MDB_RDONLY)
2471 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2472 if (env->me_flags & MDB_WRITEMAP) {
2473 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2474 ? MS_ASYNC : MS_SYNC;
2475 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2478 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2482 #ifdef BROKEN_FDATASYNC
2483 if (env->me_flags & MDB_FSYNCONLY) {
2484 if (fsync(env->me_fd))
2488 if (MDB_FDATASYNC(env->me_fd))
2495 /** Back up parent txn's cursors, then grab the originals for tracking */
2497 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2499 MDB_cursor *mc, *bk;
2504 for (i = src->mt_numdbs; --i >= 0; ) {
2505 if ((mc = src->mt_cursors[i]) != NULL) {
2506 size = sizeof(MDB_cursor);
2508 size += sizeof(MDB_xcursor);
2509 for (; mc; mc = bk->mc_next) {
2515 mc->mc_db = &dst->mt_dbs[i];
2516 /* Kill pointers into src to reduce abuse: The
2517 * user may not use mc until dst ends. But we need a valid
2518 * txn pointer here for cursor fixups to keep working.
2521 mc->mc_dbflag = &dst->mt_dbflags[i];
2522 if ((mx = mc->mc_xcursor) != NULL) {
2523 *(MDB_xcursor *)(bk+1) = *mx;
2524 mx->mx_cursor.mc_txn = dst;
2526 mc->mc_next = dst->mt_cursors[i];
2527 dst->mt_cursors[i] = mc;
2534 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2535 * @param[in] txn the transaction handle.
2536 * @param[in] merge true to keep changes to parent cursors, false to revert.
2537 * @return 0 on success, non-zero on failure.
2540 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2542 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2546 for (i = txn->mt_numdbs; --i >= 0; ) {
2547 for (mc = cursors[i]; mc; mc = next) {
2549 if ((bk = mc->mc_backup) != NULL) {
2551 /* Commit changes to parent txn */
2552 mc->mc_next = bk->mc_next;
2553 mc->mc_backup = bk->mc_backup;
2554 mc->mc_txn = bk->mc_txn;
2555 mc->mc_db = bk->mc_db;
2556 mc->mc_dbflag = bk->mc_dbflag;
2557 if ((mx = mc->mc_xcursor) != NULL)
2558 mx->mx_cursor.mc_txn = bk->mc_txn;
2560 /* Abort nested txn */
2562 if ((mx = mc->mc_xcursor) != NULL)
2563 *mx = *(MDB_xcursor *)(bk+1);
2567 /* Only malloced cursors are permanently tracked. */
2574 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2580 Pidset = F_SETLK, Pidcheck = F_GETLK
2584 /** Set or check a pid lock. Set returns 0 on success.
2585 * Check returns 0 if the process is certainly dead, nonzero if it may
2586 * be alive (the lock exists or an error happened so we do not know).
2588 * On Windows Pidset is a no-op, we merely check for the existence
2589 * of the process with the given pid. On POSIX we use a single byte
2590 * lock on the lockfile, set at an offset equal to the pid.
2593 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2595 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2598 if (op == Pidcheck) {
2599 h = OpenProcess(env->me_pidquery, FALSE, pid);
2600 /* No documented "no such process" code, but other program use this: */
2602 return ErrCode() != ERROR_INVALID_PARAMETER;
2603 /* A process exists until all handles to it close. Has it exited? */
2604 ret = WaitForSingleObject(h, 0) != 0;
2611 struct flock lock_info;
2612 memset(&lock_info, 0, sizeof(lock_info));
2613 lock_info.l_type = F_WRLCK;
2614 lock_info.l_whence = SEEK_SET;
2615 lock_info.l_start = pid;
2616 lock_info.l_len = 1;
2617 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2618 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2620 } else if ((rc = ErrCode()) == EINTR) {
2628 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2629 * @param[in] txn the transaction handle to initialize
2630 * @return 0 on success, non-zero on failure.
2633 mdb_txn_renew0(MDB_txn *txn)
2635 MDB_env *env = txn->mt_env;
2636 MDB_txninfo *ti = env->me_txns;
2638 unsigned int i, nr, flags = txn->mt_flags;
2640 int rc, new_notls = 0;
2642 if ((flags &= MDB_TXN_RDONLY) != 0) {
2644 meta = mdb_env_pick_meta(env);
2645 txn->mt_txnid = meta->mm_txnid;
2646 txn->mt_u.reader = NULL;
2648 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2649 pthread_getspecific(env->me_txkey);
2651 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2652 return MDB_BAD_RSLOT;
2654 MDB_PID_T pid = env->me_pid;
2655 MDB_THR_T tid = pthread_self();
2656 mdb_mutexref_t rmutex = env->me_rmutex;
2658 if (!env->me_live_reader) {
2659 rc = mdb_reader_pid(env, Pidset, pid);
2662 env->me_live_reader = 1;
2665 if (LOCK_MUTEX(rc, env, rmutex))
2667 nr = ti->mti_numreaders;
2668 for (i=0; i<nr; i++)
2669 if (ti->mti_readers[i].mr_pid == 0)
2671 if (i == env->me_maxreaders) {
2672 UNLOCK_MUTEX(rmutex);
2673 return MDB_READERS_FULL;
2675 r = &ti->mti_readers[i];
2676 /* Claim the reader slot, carefully since other code
2677 * uses the reader table un-mutexed: First reset the
2678 * slot, next publish it in mti_numreaders. After
2679 * that, it is safe for mdb_env_close() to touch it.
2680 * When it will be closed, we can finally claim it.
2683 r->mr_txnid = (txnid_t)-1;
2686 ti->mti_numreaders = ++nr;
2687 env->me_close_readers = nr;
2689 UNLOCK_MUTEX(rmutex);
2691 new_notls = (env->me_flags & MDB_NOTLS);
2692 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2697 do /* LY: Retry on a race, ITS#7970. */
2698 r->mr_txnid = ti->mti_txnid;
2699 while(r->mr_txnid != ti->mti_txnid);
2700 txn->mt_txnid = r->mr_txnid;
2701 txn->mt_u.reader = r;
2702 meta = env->me_metas[txn->mt_txnid & 1];
2706 /* Not yet touching txn == env->me_txn0, it may be active */
2708 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2710 txn->mt_txnid = ti->mti_txnid;
2711 meta = env->me_metas[txn->mt_txnid & 1];
2713 meta = mdb_env_pick_meta(env);
2714 txn->mt_txnid = meta->mm_txnid;
2718 if (txn->mt_txnid == mdb_debug_start)
2721 txn->mt_child = NULL;
2722 txn->mt_loose_pgs = NULL;
2723 txn->mt_loose_count = 0;
2724 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2725 txn->mt_u.dirty_list = env->me_dirty_list;
2726 txn->mt_u.dirty_list[0].mid = 0;
2727 txn->mt_free_pgs = env->me_free_pgs;
2728 txn->mt_free_pgs[0] = 0;
2729 txn->mt_spill_pgs = NULL;
2731 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2734 /* Copy the DB info and flags */
2735 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2737 /* Moved to here to avoid a data race in read TXNs */
2738 txn->mt_next_pgno = meta->mm_last_pg+1;
2740 txn->mt_flags = flags;
2743 txn->mt_numdbs = env->me_numdbs;
2744 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2745 x = env->me_dbflags[i];
2746 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2747 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2749 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2750 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2752 if (env->me_flags & MDB_FATAL_ERROR) {
2753 DPUTS("environment had fatal error, must shutdown!");
2755 } else if (env->me_maxpg < txn->mt_next_pgno) {
2756 rc = MDB_MAP_RESIZED;
2760 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2765 mdb_txn_renew(MDB_txn *txn)
2769 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2772 rc = mdb_txn_renew0(txn);
2773 if (rc == MDB_SUCCESS) {
2774 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2775 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2776 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2782 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2786 int rc, size, tsize;
2788 flags &= MDB_TXN_BEGIN_FLAGS;
2789 flags |= env->me_flags & MDB_WRITEMAP;
2791 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2795 /* Nested transactions: Max 1 child, write txns only, no writemap */
2796 flags |= parent->mt_flags;
2797 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2798 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2800 /* Child txns save MDB_pgstate and use own copy of cursors */
2801 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2802 size += tsize = sizeof(MDB_ntxn);
2803 } else if (flags & MDB_RDONLY) {
2804 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2805 size += tsize = sizeof(MDB_txn);
2807 /* Reuse preallocated write txn. However, do not touch it until
2808 * mdb_txn_renew0() succeeds, since it currently may be active.
2813 if ((txn = calloc(1, size)) == NULL) {
2814 DPRINTF(("calloc: %s", strerror(errno)));
2817 txn->mt_dbxs = env->me_dbxs; /* static */
2818 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2819 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2820 txn->mt_flags = flags;
2825 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2826 txn->mt_dbiseqs = parent->mt_dbiseqs;
2827 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2828 if (!txn->mt_u.dirty_list ||
2829 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2831 free(txn->mt_u.dirty_list);
2835 txn->mt_txnid = parent->mt_txnid;
2836 txn->mt_dirty_room = parent->mt_dirty_room;
2837 txn->mt_u.dirty_list[0].mid = 0;
2838 txn->mt_spill_pgs = NULL;
2839 txn->mt_next_pgno = parent->mt_next_pgno;
2840 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2841 parent->mt_child = txn;
2842 txn->mt_parent = parent;
2843 txn->mt_numdbs = parent->mt_numdbs;
2844 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2845 /* Copy parent's mt_dbflags, but clear DB_NEW */
2846 for (i=0; i<txn->mt_numdbs; i++)
2847 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2849 ntxn = (MDB_ntxn *)txn;
2850 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2851 if (env->me_pghead) {
2852 size = MDB_IDL_SIZEOF(env->me_pghead);
2853 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2855 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2860 rc = mdb_cursor_shadow(parent, txn);
2862 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2863 } else { /* MDB_RDONLY */
2864 txn->mt_dbiseqs = env->me_dbiseqs;
2866 rc = mdb_txn_renew0(txn);
2869 if (txn != env->me_txn0)
2872 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2874 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2875 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2876 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2883 mdb_txn_env(MDB_txn *txn)
2885 if(!txn) return NULL;
2890 mdb_txn_id(MDB_txn *txn)
2893 return txn->mt_txnid;
2896 /** Export or close DBI handles opened in this txn. */
2898 mdb_dbis_update(MDB_txn *txn, int keep)
2901 MDB_dbi n = txn->mt_numdbs;
2902 MDB_env *env = txn->mt_env;
2903 unsigned char *tdbflags = txn->mt_dbflags;
2905 for (i = n; --i >= CORE_DBS;) {
2906 if (tdbflags[i] & DB_NEW) {
2908 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2910 char *ptr = env->me_dbxs[i].md_name.mv_data;
2912 env->me_dbxs[i].md_name.mv_data = NULL;
2913 env->me_dbxs[i].md_name.mv_size = 0;
2914 env->me_dbflags[i] = 0;
2915 env->me_dbiseqs[i]++;
2921 if (keep && env->me_numdbs < n)
2925 /** End a transaction, except successful commit of a nested transaction.
2926 * May be called twice for readonly txns: First reset it, then abort.
2927 * @param[in] txn the transaction handle to end
2928 * @param[in] mode why and how to end the transaction
2931 mdb_txn_end(MDB_txn *txn, unsigned mode)
2933 MDB_env *env = txn->mt_env;
2935 static const char *const names[] = MDB_END_NAMES;
2938 /* Export or close DBI handles opened in this txn */
2939 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2941 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2942 names[mode & MDB_END_OPMASK],
2943 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2944 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2946 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2947 if (txn->mt_u.reader) {
2948 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2949 if (!(env->me_flags & MDB_NOTLS)) {
2950 txn->mt_u.reader = NULL; /* txn does not own reader */
2951 } else if (mode & MDB_END_SLOT) {
2952 txn->mt_u.reader->mr_pid = 0;
2953 txn->mt_u.reader = NULL;
2954 } /* else txn owns the slot until it does MDB_END_SLOT */
2956 txn->mt_numdbs = 0; /* prevent further DBI activity */
2957 txn->mt_flags |= MDB_TXN_FINISHED;
2959 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2960 pgno_t *pghead = env->me_pghead;
2962 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2963 mdb_cursors_close(txn, 0);
2964 if (!(env->me_flags & MDB_WRITEMAP)) {
2965 mdb_dlist_free(txn);
2969 txn->mt_flags = MDB_TXN_FINISHED;
2971 if (!txn->mt_parent) {
2972 mdb_midl_shrink(&txn->mt_free_pgs);
2973 env->me_free_pgs = txn->mt_free_pgs;
2975 env->me_pghead = NULL;
2979 mode = 0; /* txn == env->me_txn0, do not free() it */
2981 /* The writer mutex was locked in mdb_txn_begin. */
2983 UNLOCK_MUTEX(env->me_wmutex);
2985 txn->mt_parent->mt_child = NULL;
2986 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2987 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2988 mdb_midl_free(txn->mt_free_pgs);
2989 mdb_midl_free(txn->mt_spill_pgs);
2990 free(txn->mt_u.dirty_list);
2993 mdb_midl_free(pghead);
2996 if (mode & MDB_END_FREE)
3001 mdb_txn_reset(MDB_txn *txn)
3006 /* This call is only valid for read-only txns */
3007 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3010 mdb_txn_end(txn, MDB_END_RESET);
3014 mdb_txn_abort(MDB_txn *txn)
3020 mdb_txn_abort(txn->mt_child);
3022 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3025 /** Save the freelist as of this transaction to the freeDB.
3026 * This changes the freelist. Keep trying until it stabilizes.
3029 mdb_freelist_save(MDB_txn *txn)
3031 /* env->me_pghead[] can grow and shrink during this call.
3032 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3033 * Page numbers cannot disappear from txn->mt_free_pgs[].
3036 MDB_env *env = txn->mt_env;
3037 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3038 txnid_t pglast = 0, head_id = 0;
3039 pgno_t freecnt = 0, *free_pgs, *mop;
3040 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3042 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3044 if (env->me_pghead) {
3045 /* Make sure first page of freeDB is touched and on freelist */
3046 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3047 if (rc && rc != MDB_NOTFOUND)
3051 if (!env->me_pghead && txn->mt_loose_pgs) {
3052 /* Put loose page numbers in mt_free_pgs, since
3053 * we may be unable to return them to me_pghead.
3055 MDB_page *mp = txn->mt_loose_pgs;
3056 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3058 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3059 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3060 txn->mt_loose_pgs = NULL;
3061 txn->mt_loose_count = 0;
3064 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3065 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3066 ? SSIZE_MAX : maxfree_1pg;
3069 /* Come back here after each Put() in case freelist changed */
3074 /* If using records from freeDB which we have not yet
3075 * deleted, delete them and any we reserved for me_pghead.
3077 while (pglast < env->me_pglast) {
3078 rc = mdb_cursor_first(&mc, &key, NULL);
3081 pglast = head_id = *(txnid_t *)key.mv_data;
3082 total_room = head_room = 0;
3083 mdb_tassert(txn, pglast <= env->me_pglast);
3084 rc = mdb_cursor_del(&mc, 0);
3089 /* Save the IDL of pages freed by this txn, to a single record */
3090 if (freecnt < txn->mt_free_pgs[0]) {
3092 /* Make sure last page of freeDB is touched and on freelist */
3093 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3094 if (rc && rc != MDB_NOTFOUND)
3097 free_pgs = txn->mt_free_pgs;
3098 /* Write to last page of freeDB */
3099 key.mv_size = sizeof(txn->mt_txnid);
3100 key.mv_data = &txn->mt_txnid;
3102 freecnt = free_pgs[0];
3103 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3104 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3107 /* Retry if mt_free_pgs[] grew during the Put() */
3108 free_pgs = txn->mt_free_pgs;
3109 } while (freecnt < free_pgs[0]);
3110 mdb_midl_sort(free_pgs);
3111 memcpy(data.mv_data, free_pgs, data.mv_size);
3114 unsigned int i = free_pgs[0];
3115 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3116 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3118 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3124 mop = env->me_pghead;
3125 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3127 /* Reserve records for me_pghead[]. Split it if multi-page,
3128 * to avoid searching freeDB for a page range. Use keys in
3129 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3131 if (total_room >= mop_len) {
3132 if (total_room == mop_len || --more < 0)
3134 } else if (head_room >= maxfree_1pg && head_id > 1) {
3135 /* Keep current record (overflow page), add a new one */
3139 /* (Re)write {key = head_id, IDL length = head_room} */
3140 total_room -= head_room;
3141 head_room = mop_len - total_room;
3142 if (head_room > maxfree_1pg && head_id > 1) {
3143 /* Overflow multi-page for part of me_pghead */
3144 head_room /= head_id; /* amortize page sizes */
3145 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3146 } else if (head_room < 0) {
3147 /* Rare case, not bothering to delete this record */
3150 key.mv_size = sizeof(head_id);
3151 key.mv_data = &head_id;
3152 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3153 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3156 /* IDL is initially empty, zero out at least the length */
3157 pgs = (pgno_t *)data.mv_data;
3158 j = head_room > clean_limit ? head_room : 0;
3162 total_room += head_room;
3165 /* Return loose page numbers to me_pghead, though usually none are
3166 * left at this point. The pages themselves remain in dirty_list.
3168 if (txn->mt_loose_pgs) {
3169 MDB_page *mp = txn->mt_loose_pgs;
3170 unsigned count = txn->mt_loose_count;
3172 /* Room for loose pages + temp IDL with same */
3173 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3175 mop = env->me_pghead;
3176 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3177 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3178 loose[ ++count ] = mp->mp_pgno;
3180 mdb_midl_sort(loose);
3181 mdb_midl_xmerge(mop, loose);
3182 txn->mt_loose_pgs = NULL;
3183 txn->mt_loose_count = 0;
3187 /* Fill in the reserved me_pghead records */
3193 rc = mdb_cursor_first(&mc, &key, &data);
3194 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3195 txnid_t id = *(txnid_t *)key.mv_data;
3196 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3199 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3201 if (len > mop_len) {
3203 data.mv_size = (len + 1) * sizeof(MDB_ID);
3205 data.mv_data = mop -= len;
3208 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3210 if (rc || !(mop_len -= len))
3217 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3218 * @param[in] txn the transaction that's being committed
3219 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3220 * @return 0 on success, non-zero on failure.
3223 mdb_page_flush(MDB_txn *txn, int keep)
3225 MDB_env *env = txn->mt_env;
3226 MDB_ID2L dl = txn->mt_u.dirty_list;
3227 unsigned psize = env->me_psize, j;
3228 int i, pagecount = dl[0].mid, rc;
3229 size_t size = 0, pos = 0;
3231 MDB_page *dp = NULL;
3235 struct iovec iov[MDB_COMMIT_PAGES];
3236 ssize_t wpos = 0, wsize = 0, wres;
3237 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3243 if (env->me_flags & MDB_WRITEMAP) {
3244 /* Clear dirty flags */
3245 while (++i <= pagecount) {
3247 /* Don't flush this page yet */
3248 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3249 dp->mp_flags &= ~P_KEEP;
3253 dp->mp_flags &= ~P_DIRTY;
3258 /* Write the pages */
3260 if (++i <= pagecount) {
3262 /* Don't flush this page yet */
3263 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3264 dp->mp_flags &= ~P_KEEP;
3269 /* clear dirty flag */
3270 dp->mp_flags &= ~P_DIRTY;
3273 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3278 /* Windows actually supports scatter/gather I/O, but only on
3279 * unbuffered file handles. Since we're relying on the OS page
3280 * cache for all our data, that's self-defeating. So we just
3281 * write pages one at a time. We use the ov structure to set
3282 * the write offset, to at least save the overhead of a Seek
3285 DPRINTF(("committing page %"Z"u", pgno));
3286 memset(&ov, 0, sizeof(ov));
3287 ov.Offset = pos & 0xffffffff;
3288 ov.OffsetHigh = pos >> 16 >> 16;
3289 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3291 DPRINTF(("WriteFile: %d", rc));
3295 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3296 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3299 /* Write previous page(s) */
3300 #ifdef MDB_USE_PWRITEV
3301 wres = pwritev(env->me_fd, iov, n, wpos);
3304 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3307 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3311 DPRINTF(("lseek: %s", strerror(rc)));
3314 wres = writev(env->me_fd, iov, n);
3317 if (wres != wsize) {
3322 DPRINTF(("Write error: %s", strerror(rc)));
3324 rc = EIO; /* TODO: Use which error code? */
3325 DPUTS("short write, filesystem full?");
3336 DPRINTF(("committing page %"Z"u", pgno));
3337 next_pos = pos + size;
3338 iov[n].iov_len = size;
3339 iov[n].iov_base = (char *)dp;
3345 /* MIPS has cache coherency issues, this is a no-op everywhere else
3346 * Note: for any size >= on-chip cache size, entire on-chip cache is
3349 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3351 for (i = keep; ++i <= pagecount; ) {
3353 /* This is a page we skipped above */
3356 dl[j].mid = dp->mp_pgno;
3359 mdb_dpage_free(env, dp);
3364 txn->mt_dirty_room += i - j;
3370 mdb_txn_commit(MDB_txn *txn)
3373 unsigned int i, end_mode;
3379 /* mdb_txn_end() mode for a commit which writes nothing */
3380 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3382 if (txn->mt_child) {
3383 rc = mdb_txn_commit(txn->mt_child);
3390 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3394 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3395 DPUTS("txn has failed/finished, can't commit");
3397 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3402 if (txn->mt_parent) {
3403 MDB_txn *parent = txn->mt_parent;
3407 unsigned x, y, len, ps_len;
3409 /* Append our free list to parent's */
3410 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3413 mdb_midl_free(txn->mt_free_pgs);
3414 /* Failures after this must either undo the changes
3415 * to the parent or set MDB_TXN_ERROR in the parent.
3418 parent->mt_next_pgno = txn->mt_next_pgno;
3419 parent->mt_flags = txn->mt_flags;
3421 /* Merge our cursors into parent's and close them */
3422 mdb_cursors_close(txn, 1);
3424 /* Update parent's DB table. */
3425 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3426 parent->mt_numdbs = txn->mt_numdbs;
3427 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3428 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3429 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3430 /* preserve parent's DB_NEW status */
3431 x = parent->mt_dbflags[i] & DB_NEW;
3432 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3435 dst = parent->mt_u.dirty_list;
3436 src = txn->mt_u.dirty_list;
3437 /* Remove anything in our dirty list from parent's spill list */
3438 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3440 pspill[0] = (pgno_t)-1;
3441 /* Mark our dirty pages as deleted in parent spill list */
3442 for (i=0, len=src[0].mid; ++i <= len; ) {
3443 MDB_ID pn = src[i].mid << 1;
3444 while (pn > pspill[x])
3446 if (pn == pspill[x]) {
3451 /* Squash deleted pagenums if we deleted any */
3452 for (x=y; ++x <= ps_len; )
3453 if (!(pspill[x] & 1))
3454 pspill[++y] = pspill[x];
3458 /* Remove anything in our spill list from parent's dirty list */
3459 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3460 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3461 MDB_ID pn = txn->mt_spill_pgs[i];
3463 continue; /* deleted spillpg */
3465 y = mdb_mid2l_search(dst, pn);
3466 if (y <= dst[0].mid && dst[y].mid == pn) {
3468 while (y < dst[0].mid) {
3477 /* Find len = length of merging our dirty list with parent's */
3479 dst[0].mid = 0; /* simplify loops */
3480 if (parent->mt_parent) {
3481 len = x + src[0].mid;
3482 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3483 for (i = x; y && i; y--) {
3484 pgno_t yp = src[y].mid;
3485 while (yp < dst[i].mid)
3487 if (yp == dst[i].mid) {
3492 } else { /* Simplify the above for single-ancestor case */
3493 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3495 /* Merge our dirty list with parent's */
3497 for (i = len; y; dst[i--] = src[y--]) {
3498 pgno_t yp = src[y].mid;
3499 while (yp < dst[x].mid)
3500 dst[i--] = dst[x--];
3501 if (yp == dst[x].mid)
3502 free(dst[x--].mptr);
3504 mdb_tassert(txn, i == x);
3506 free(txn->mt_u.dirty_list);
3507 parent->mt_dirty_room = txn->mt_dirty_room;
3508 if (txn->mt_spill_pgs) {
3509 if (parent->mt_spill_pgs) {
3510 /* TODO: Prevent failure here, so parent does not fail */
3511 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3513 parent->mt_flags |= MDB_TXN_ERROR;
3514 mdb_midl_free(txn->mt_spill_pgs);
3515 mdb_midl_sort(parent->mt_spill_pgs);
3517 parent->mt_spill_pgs = txn->mt_spill_pgs;
3521 /* Append our loose page list to parent's */
3522 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3524 *lp = txn->mt_loose_pgs;
3525 parent->mt_loose_count += txn->mt_loose_count;
3527 parent->mt_child = NULL;
3528 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3533 if (txn != env->me_txn) {
3534 DPUTS("attempt to commit unknown transaction");
3539 mdb_cursors_close(txn, 0);
3541 if (!txn->mt_u.dirty_list[0].mid &&
3542 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3545 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3546 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3548 /* Update DB root pointers */
3549 if (txn->mt_numdbs > CORE_DBS) {
3553 data.mv_size = sizeof(MDB_db);
3555 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3556 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3557 if (txn->mt_dbflags[i] & DB_DIRTY) {
3558 if (TXN_DBI_CHANGED(txn, i)) {
3562 data.mv_data = &txn->mt_dbs[i];
3563 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3571 rc = mdb_freelist_save(txn);
3575 mdb_midl_free(env->me_pghead);
3576 env->me_pghead = NULL;
3577 mdb_midl_shrink(&txn->mt_free_pgs);
3583 if ((rc = mdb_page_flush(txn, 0)) ||
3584 (rc = mdb_env_sync(env, 0)) ||
3585 (rc = mdb_env_write_meta(txn)))
3587 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3590 mdb_txn_end(txn, end_mode);
3598 /** Read the environment parameters of a DB environment before
3599 * mapping it into memory.
3600 * @param[in] env the environment handle
3601 * @param[out] meta address of where to store the meta information
3602 * @return 0 on success, non-zero on failure.
3605 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3611 enum { Size = sizeof(pbuf) };
3613 /* We don't know the page size yet, so use a minimum value.
3614 * Read both meta pages so we can use the latest one.
3617 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3621 memset(&ov, 0, sizeof(ov));
3623 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3624 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3627 rc = pread(env->me_fd, &pbuf, Size, off);
3630 if (rc == 0 && off == 0)
3632 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3633 DPRINTF(("read: %s", mdb_strerror(rc)));
3637 p = (MDB_page *)&pbuf;
3639 if (!F_ISSET(p->mp_flags, P_META)) {
3640 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3645 if (m->mm_magic != MDB_MAGIC) {
3646 DPUTS("meta has invalid magic");
3650 if (m->mm_version != MDB_DATA_VERSION) {
3651 DPRINTF(("database is version %u, expected version %u",
3652 m->mm_version, MDB_DATA_VERSION));
3653 return MDB_VERSION_MISMATCH;
3656 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3662 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3664 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3666 meta->mm_magic = MDB_MAGIC;
3667 meta->mm_version = MDB_DATA_VERSION;
3668 meta->mm_mapsize = env->me_mapsize;
3669 meta->mm_psize = env->me_psize;
3670 meta->mm_last_pg = NUM_METAS-1;
3671 meta->mm_flags = env->me_flags & 0xffff;
3672 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3673 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3674 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3677 /** Write the environment parameters of a freshly created DB environment.
3678 * @param[in] env the environment handle
3679 * @param[in] meta the #MDB_meta to write
3680 * @return 0 on success, non-zero on failure.
3683 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3691 memset(&ov, 0, sizeof(ov));
3692 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3694 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3697 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3698 len = pwrite(fd, ptr, size, pos); \
3699 if (len == -1 && ErrCode() == EINTR) continue; \
3700 rc = (len >= 0); break; } while(1)
3703 DPUTS("writing new meta page");
3705 psize = env->me_psize;
3707 p = calloc(NUM_METAS, psize);
3712 p->mp_flags = P_META;
3713 *(MDB_meta *)METADATA(p) = *meta;
3715 q = (MDB_page *)((char *)p + psize);
3717 q->mp_flags = P_META;
3718 *(MDB_meta *)METADATA(q) = *meta;
3720 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3723 else if ((unsigned) len == psize * NUM_METAS)
3731 /** Update the environment info to commit a transaction.
3732 * @param[in] txn the transaction that's being committed
3733 * @return 0 on success, non-zero on failure.
3736 mdb_env_write_meta(MDB_txn *txn)
3739 MDB_meta meta, metab, *mp;
3743 int rc, len, toggle;
3752 toggle = txn->mt_txnid & 1;
3753 DPRINTF(("writing meta page %d for root page %"Z"u",
3754 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3757 flags = env->me_flags;
3758 mp = env->me_metas[toggle];
3759 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3760 /* Persist any increases of mapsize config */
3761 if (mapsize < env->me_mapsize)
3762 mapsize = env->me_mapsize;
3764 if (flags & MDB_WRITEMAP) {
3765 mp->mm_mapsize = mapsize;
3766 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3767 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3768 mp->mm_last_pg = txn->mt_next_pgno - 1;
3769 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3770 !(defined(__i386__) || defined(__x86_64__))
3771 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3772 __sync_synchronize();
3774 mp->mm_txnid = txn->mt_txnid;
3775 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3776 unsigned meta_size = env->me_psize;
3777 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3778 ptr = (char *)mp - PAGEHDRSZ;
3779 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3780 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3784 if (MDB_MSYNC(ptr, meta_size, rc)) {
3791 metab.mm_txnid = mp->mm_txnid;
3792 metab.mm_last_pg = mp->mm_last_pg;
3794 meta.mm_mapsize = mapsize;
3795 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3796 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3797 meta.mm_last_pg = txn->mt_next_pgno - 1;
3798 meta.mm_txnid = txn->mt_txnid;
3800 off = offsetof(MDB_meta, mm_mapsize);
3801 ptr = (char *)&meta + off;
3802 len = sizeof(MDB_meta) - off;
3803 off += (char *)mp - env->me_map;
3805 /* Write to the SYNC fd */
3806 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3809 memset(&ov, 0, sizeof(ov));
3811 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3816 rc = pwrite(mfd, ptr, len, off);
3819 rc = rc < 0 ? ErrCode() : EIO;
3824 DPUTS("write failed, disk error?");
3825 /* On a failure, the pagecache still contains the new data.
3826 * Write some old data back, to prevent it from being used.
3827 * Use the non-SYNC fd; we know it will fail anyway.
3829 meta.mm_last_pg = metab.mm_last_pg;
3830 meta.mm_txnid = metab.mm_txnid;
3832 memset(&ov, 0, sizeof(ov));
3834 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3836 r2 = pwrite(env->me_fd, ptr, len, off);
3837 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3840 env->me_flags |= MDB_FATAL_ERROR;
3843 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3844 CACHEFLUSH(env->me_map + off, len, DCACHE);
3846 /* Memory ordering issues are irrelevant; since the entire writer
3847 * is wrapped by wmutex, all of these changes will become visible
3848 * after the wmutex is unlocked. Since the DB is multi-version,
3849 * readers will get consistent data regardless of how fresh or
3850 * how stale their view of these values is.
3853 env->me_txns->mti_txnid = txn->mt_txnid;
3858 /** Check both meta pages to see which one is newer.
3859 * @param[in] env the environment handle
3860 * @return newest #MDB_meta.
3863 mdb_env_pick_meta(const MDB_env *env)
3865 MDB_meta *const *metas = env->me_metas;
3866 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3870 mdb_env_create(MDB_env **env)
3874 e = calloc(1, sizeof(MDB_env));
3878 e->me_maxreaders = DEFAULT_READERS;
3879 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3880 e->me_fd = INVALID_HANDLE_VALUE;
3881 e->me_lfd = INVALID_HANDLE_VALUE;
3882 e->me_mfd = INVALID_HANDLE_VALUE;
3883 #ifdef MDB_USE_POSIX_SEM
3884 e->me_rmutex = SEM_FAILED;
3885 e->me_wmutex = SEM_FAILED;
3887 e->me_pid = getpid();
3888 GET_PAGESIZE(e->me_os_psize);
3889 VGMEMP_CREATE(e,0,0);
3895 mdb_env_map(MDB_env *env, void *addr)
3898 unsigned int flags = env->me_flags;
3902 LONG sizelo, sizehi;
3905 if (flags & MDB_RDONLY) {
3906 /* Don't set explicit map size, use whatever exists */
3911 msize = env->me_mapsize;
3912 sizelo = msize & 0xffffffff;
3913 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3915 /* Windows won't create mappings for zero length files.
3916 * and won't map more than the file size.
3917 * Just set the maxsize right now.
3919 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3920 || !SetEndOfFile(env->me_fd)
3921 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3925 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3926 PAGE_READWRITE : PAGE_READONLY,
3927 sizehi, sizelo, NULL);
3930 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3931 FILE_MAP_WRITE : FILE_MAP_READ,
3933 rc = env->me_map ? 0 : ErrCode();
3938 int prot = PROT_READ;
3939 if (flags & MDB_WRITEMAP) {
3941 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3944 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3946 if (env->me_map == MAP_FAILED) {
3951 if (flags & MDB_NORDAHEAD) {
3952 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3954 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3956 #ifdef POSIX_MADV_RANDOM
3957 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3958 #endif /* POSIX_MADV_RANDOM */
3959 #endif /* MADV_RANDOM */
3963 /* Can happen because the address argument to mmap() is just a
3964 * hint. mmap() can pick another, e.g. if the range is in use.
3965 * The MAP_FIXED flag would prevent that, but then mmap could
3966 * instead unmap existing pages to make room for the new map.
3968 if (addr && env->me_map != addr)
3969 return EBUSY; /* TODO: Make a new MDB_* error code? */
3971 p = (MDB_page *)env->me_map;
3972 env->me_metas[0] = METADATA(p);
3973 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3979 mdb_env_set_mapsize(MDB_env *env, size_t size)
3981 /* If env is already open, caller is responsible for making
3982 * sure there are no active txns.
3990 meta = mdb_env_pick_meta(env);
3992 size = meta->mm_mapsize;
3994 /* Silently round up to minimum if the size is too small */
3995 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3999 munmap(env->me_map, env->me_mapsize);
4000 env->me_mapsize = size;
4001 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4002 rc = mdb_env_map(env, old);
4006 env->me_mapsize = size;
4008 env->me_maxpg = env->me_mapsize / env->me_psize;
4013 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4017 env->me_maxdbs = dbs + CORE_DBS;
4022 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4024 if (env->me_map || readers < 1)
4026 env->me_maxreaders = readers;
4031 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4033 if (!env || !readers)
4035 *readers = env->me_maxreaders;
4040 mdb_fsize(HANDLE fd, size_t *size)
4043 LARGE_INTEGER fsize;
4045 if (!GetFileSizeEx(fd, &fsize))
4048 *size = fsize.QuadPart;
4060 #ifdef BROKEN_FDATASYNC
4061 #include <sys/utsname.h>
4062 #include <sys/vfs.h>
4065 /** Further setup required for opening an LMDB environment
4068 mdb_env_open2(MDB_env *env)
4070 unsigned int flags = env->me_flags;
4071 int i, newenv = 0, rc;
4075 /* See if we should use QueryLimited */
4077 if ((rc & 0xff) > 5)
4078 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4080 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4083 #ifdef BROKEN_FDATASYNC
4084 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4085 * https://lkml.org/lkml/2012/9/3/83
4086 * Kernels after 3.6-rc6 are known good.
4087 * https://lkml.org/lkml/2012/9/10/556
4088 * See if the DB is on ext3/ext4, then check for new enough kernel
4089 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4094 fstatfs(env->me_fd, &st);
4095 while (st.f_type == 0xEF53) {
4099 if (uts.release[0] < '3') {
4100 if (!strncmp(uts.release, "2.6.32.", 7)) {
4101 i = atoi(uts.release+7);
4103 break; /* 2.6.32.60 and newer is OK */
4104 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4105 i = atoi(uts.release+7);
4107 break; /* 2.6.34.15 and newer is OK */
4109 } else if (uts.release[0] == '3') {
4110 i = atoi(uts.release+2);
4112 break; /* 3.6 and newer is OK */
4114 i = atoi(uts.release+4);
4116 break; /* 3.5.4 and newer is OK */
4117 } else if (i == 2) {
4118 i = atoi(uts.release+4);
4120 break; /* 3.2.30 and newer is OK */
4122 } else { /* 4.x and newer is OK */
4125 env->me_flags |= MDB_FSYNCONLY;
4131 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4134 DPUTS("new mdbenv");
4136 env->me_psize = env->me_os_psize;
4137 if (env->me_psize > MAX_PAGESIZE)
4138 env->me_psize = MAX_PAGESIZE;
4139 memset(&meta, 0, sizeof(meta));
4140 mdb_env_init_meta0(env, &meta);
4141 meta.mm_mapsize = DEFAULT_MAPSIZE;
4143 env->me_psize = meta.mm_psize;
4146 /* Was a mapsize configured? */
4147 if (!env->me_mapsize) {
4148 env->me_mapsize = meta.mm_mapsize;
4151 /* Make sure mapsize >= committed data size. Even when using
4152 * mm_mapsize, which could be broken in old files (ITS#7789).
4154 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4155 if (env->me_mapsize < minsize)
4156 env->me_mapsize = minsize;
4158 meta.mm_mapsize = env->me_mapsize;
4160 if (newenv && !(flags & MDB_FIXEDMAP)) {
4161 /* mdb_env_map() may grow the datafile. Write the metapages
4162 * first, so the file will be valid if initialization fails.
4163 * Except with FIXEDMAP, since we do not yet know mm_address.
4164 * We could fill in mm_address later, but then a different
4165 * program might end up doing that - one with a memory layout
4166 * and map address which does not suit the main program.
4168 rc = mdb_env_init_meta(env, &meta);
4174 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4179 if (flags & MDB_FIXEDMAP)
4180 meta.mm_address = env->me_map;
4181 i = mdb_env_init_meta(env, &meta);
4182 if (i != MDB_SUCCESS) {
4187 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4188 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4190 #if !(MDB_MAXKEYSIZE)
4191 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4193 env->me_maxpg = env->me_mapsize / env->me_psize;
4197 MDB_meta *meta = mdb_env_pick_meta(env);
4198 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4200 DPRINTF(("opened database version %u, pagesize %u",
4201 meta->mm_version, env->me_psize));
4202 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4203 DPRINTF(("depth: %u", db->md_depth));
4204 DPRINTF(("entries: %"Z"u", db->md_entries));
4205 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4206 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4207 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4208 DPRINTF(("root: %"Z"u", db->md_root));
4216 /** Release a reader thread's slot in the reader lock table.
4217 * This function is called automatically when a thread exits.
4218 * @param[in] ptr This points to the slot in the reader lock table.
4221 mdb_env_reader_dest(void *ptr)
4223 MDB_reader *reader = ptr;
4229 /** Junk for arranging thread-specific callbacks on Windows. This is
4230 * necessarily platform and compiler-specific. Windows supports up
4231 * to 1088 keys. Let's assume nobody opens more than 64 environments
4232 * in a single process, for now. They can override this if needed.
4234 #ifndef MAX_TLS_KEYS
4235 #define MAX_TLS_KEYS 64
4237 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4238 static int mdb_tls_nkeys;
4240 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4244 case DLL_PROCESS_ATTACH: break;
4245 case DLL_THREAD_ATTACH: break;
4246 case DLL_THREAD_DETACH:
4247 for (i=0; i<mdb_tls_nkeys; i++) {
4248 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4250 mdb_env_reader_dest(r);
4254 case DLL_PROCESS_DETACH: break;
4259 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4261 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4265 /* Force some symbol references.
4266 * _tls_used forces the linker to create the TLS directory if not already done
4267 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4269 #pragma comment(linker, "/INCLUDE:_tls_used")
4270 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4271 #pragma const_seg(".CRT$XLB")
4272 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4273 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4276 #pragma comment(linker, "/INCLUDE:__tls_used")
4277 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4278 #pragma data_seg(".CRT$XLB")
4279 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4281 #endif /* WIN 32/64 */
4282 #endif /* !__GNUC__ */
4285 /** Downgrade the exclusive lock on the region back to shared */
4287 mdb_env_share_locks(MDB_env *env, int *excl)
4290 MDB_meta *meta = mdb_env_pick_meta(env);
4292 env->me_txns->mti_txnid = meta->mm_txnid;
4297 /* First acquire a shared lock. The Unlock will
4298 * then release the existing exclusive lock.
4300 memset(&ov, 0, sizeof(ov));
4301 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4304 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4310 struct flock lock_info;
4311 /* The shared lock replaces the existing lock */
4312 memset((void *)&lock_info, 0, sizeof(lock_info));
4313 lock_info.l_type = F_RDLCK;
4314 lock_info.l_whence = SEEK_SET;
4315 lock_info.l_start = 0;
4316 lock_info.l_len = 1;
4317 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4318 (rc = ErrCode()) == EINTR) ;
4319 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4326 /** Try to get exclusive lock, otherwise shared.
4327 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4330 mdb_env_excl_lock(MDB_env *env, int *excl)
4334 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4338 memset(&ov, 0, sizeof(ov));
4339 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4346 struct flock lock_info;
4347 memset((void *)&lock_info, 0, sizeof(lock_info));
4348 lock_info.l_type = F_WRLCK;
4349 lock_info.l_whence = SEEK_SET;
4350 lock_info.l_start = 0;
4351 lock_info.l_len = 1;
4352 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4353 (rc = ErrCode()) == EINTR) ;
4357 # ifndef MDB_USE_POSIX_MUTEX
4358 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4361 lock_info.l_type = F_RDLCK;
4362 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4363 (rc = ErrCode()) == EINTR) ;
4373 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4375 * @(#) $Revision: 5.1 $
4376 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4377 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4379 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4383 * Please do not copyright this code. This code is in the public domain.
4385 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4386 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4387 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4388 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4389 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4390 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4391 * PERFORMANCE OF THIS SOFTWARE.
4394 * chongo <Landon Curt Noll> /\oo/\
4395 * http://www.isthe.com/chongo/
4397 * Share and Enjoy! :-)
4400 typedef unsigned long long mdb_hash_t;
4401 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4403 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4404 * @param[in] val value to hash
4405 * @param[in] hval initial value for hash
4406 * @return 64 bit hash
4408 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4409 * hval arg on the first call.
4412 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4414 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4415 unsigned char *end = s + val->mv_size;
4417 * FNV-1a hash each octet of the string
4420 /* xor the bottom with the current octet */
4421 hval ^= (mdb_hash_t)*s++;
4423 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4424 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4425 (hval << 7) + (hval << 8) + (hval << 40);
4427 /* return our new hash value */
4431 /** Hash the string and output the encoded hash.
4432 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4433 * very short name limits. We don't care about the encoding being reversible,
4434 * we just want to preserve as many bits of the input as possible in a
4435 * small printable string.
4436 * @param[in] str string to hash
4437 * @param[out] encbuf an array of 11 chars to hold the hash
4439 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4442 mdb_pack85(unsigned long l, char *out)
4446 for (i=0; i<5; i++) {
4447 *out++ = mdb_a85[l % 85];
4453 mdb_hash_enc(MDB_val *val, char *encbuf)
4455 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4457 mdb_pack85(h, encbuf);
4458 mdb_pack85(h>>32, encbuf+5);
4463 /** Open and/or initialize the lock region for the environment.
4464 * @param[in] env The LMDB environment.
4465 * @param[in] lpath The pathname of the file used for the lock region.
4466 * @param[in] mode The Unix permissions for the file, if we create it.
4467 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4468 * @return 0 on success, non-zero on failure.
4471 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4474 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4476 # define MDB_ERRCODE_ROFS EROFS
4477 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4478 # define MDB_CLOEXEC O_CLOEXEC
4481 # define MDB_CLOEXEC 0
4489 rc = utf8_to_utf16(lpath, -1, &wlpath, NULL);
4492 env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
4493 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4494 FILE_ATTRIBUTE_NORMAL, NULL);
4497 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4499 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4501 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4506 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4507 /* Lose record locks when exec*() */
4508 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4509 fcntl(env->me_lfd, F_SETFD, fdflags);
4512 if (!(env->me_flags & MDB_NOTLS)) {
4513 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4516 env->me_flags |= MDB_ENV_TXKEY;
4518 /* Windows TLS callbacks need help finding their TLS info. */
4519 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4523 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4527 /* Try to get exclusive lock. If we succeed, then
4528 * nobody is using the lock region and we should initialize it.
4530 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4533 size = GetFileSize(env->me_lfd, NULL);
4535 size = lseek(env->me_lfd, 0, SEEK_END);
4536 if (size == -1) goto fail_errno;
4538 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4539 if (size < rsize && *excl > 0) {
4541 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4542 || !SetEndOfFile(env->me_lfd))
4545 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4549 size = rsize - sizeof(MDB_txninfo);
4550 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4555 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4557 if (!mh) goto fail_errno;
4558 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4560 if (!env->me_txns) goto fail_errno;
4562 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4564 if (m == MAP_FAILED) goto fail_errno;
4570 BY_HANDLE_FILE_INFORMATION stbuf;
4579 if (!mdb_sec_inited) {
4580 InitializeSecurityDescriptor(&mdb_null_sd,
4581 SECURITY_DESCRIPTOR_REVISION);
4582 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4583 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4584 mdb_all_sa.bInheritHandle = FALSE;
4585 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4588 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4589 idbuf.volume = stbuf.dwVolumeSerialNumber;
4590 idbuf.nhigh = stbuf.nFileIndexHigh;
4591 idbuf.nlow = stbuf.nFileIndexLow;
4592 val.mv_data = &idbuf;
4593 val.mv_size = sizeof(idbuf);
4594 mdb_hash_enc(&val, encbuf);
4595 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4596 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4597 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4598 if (!env->me_rmutex) goto fail_errno;
4599 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4600 if (!env->me_wmutex) goto fail_errno;
4601 #elif defined(MDB_USE_POSIX_SEM)
4610 #if defined(__NetBSD__)
4611 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4613 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4614 idbuf.dev = stbuf.st_dev;
4615 idbuf.ino = stbuf.st_ino;
4616 val.mv_data = &idbuf;
4617 val.mv_size = sizeof(idbuf);
4618 mdb_hash_enc(&val, encbuf);
4619 #ifdef MDB_SHORT_SEMNAMES
4620 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4622 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4623 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4624 /* Clean up after a previous run, if needed: Try to
4625 * remove both semaphores before doing anything else.
4627 sem_unlink(env->me_txns->mti_rmname);
4628 sem_unlink(env->me_txns->mti_wmname);
4629 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4630 O_CREAT|O_EXCL, mode, 1);
4631 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4632 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4633 O_CREAT|O_EXCL, mode, 1);
4634 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4635 #else /* MDB_USE_POSIX_MUTEX: */
4636 pthread_mutexattr_t mattr;
4638 /* Solaris needs this before initing a robust mutex. Otherwise
4639 * it may skip the init and return EBUSY "seems someone already
4640 * inited" or EINVAL "it was inited differently".
4642 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
4643 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
4645 if ((rc = pthread_mutexattr_init(&mattr)))
4648 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
4649 #ifdef MDB_ROBUST_SUPPORTED
4650 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
4652 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
4653 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
4654 pthread_mutexattr_destroy(&mattr);
4657 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4659 env->me_txns->mti_magic = MDB_MAGIC;
4660 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4661 env->me_txns->mti_txnid = 0;
4662 env->me_txns->mti_numreaders = 0;
4665 if (env->me_txns->mti_magic != MDB_MAGIC) {
4666 DPUTS("lock region has invalid magic");
4670 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4671 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4672 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4673 rc = MDB_VERSION_MISMATCH;
4677 if (rc && rc != EACCES && rc != EAGAIN) {
4681 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4682 if (!env->me_rmutex) goto fail_errno;
4683 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4684 if (!env->me_wmutex) goto fail_errno;
4685 #elif defined(MDB_USE_POSIX_SEM)
4686 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4687 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4688 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4689 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4700 /** The name of the lock file in the DB environment */
4701 #define LOCKNAME "/lock.mdb"
4702 /** The name of the data file in the DB environment */
4703 #define DATANAME "/data.mdb"
4704 /** The suffix of the lock file when no subdir is used */
4705 #define LOCKSUFF "-lock"
4706 /** Only a subset of the @ref mdb_env flags can be changed
4707 * at runtime. Changing other flags requires closing the
4708 * environment and re-opening it with the new flags.
4710 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4711 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4712 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4714 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4715 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4719 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4721 int oflags, rc, len, excl = -1;
4722 char *lpath, *dpath;
4727 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4731 if (flags & MDB_NOSUBDIR) {
4732 rc = len + sizeof(LOCKSUFF) + len + 1;
4734 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4739 if (flags & MDB_NOSUBDIR) {
4740 dpath = lpath + len + sizeof(LOCKSUFF);
4741 sprintf(lpath, "%s" LOCKSUFF, path);
4742 strcpy(dpath, path);
4744 dpath = lpath + len + sizeof(LOCKNAME);
4745 sprintf(lpath, "%s" LOCKNAME, path);
4746 sprintf(dpath, "%s" DATANAME, path);
4750 flags |= env->me_flags;
4751 if (flags & MDB_RDONLY) {
4752 /* silently ignore WRITEMAP when we're only getting read access */
4753 flags &= ~MDB_WRITEMAP;
4755 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4756 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4759 env->me_flags = flags |= MDB_ENV_ACTIVE;
4763 env->me_path = strdup(path);
4764 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4765 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4766 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4767 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4771 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4773 /* For RDONLY, get lockfile after we know datafile exists */
4774 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4775 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4781 if (F_ISSET(flags, MDB_RDONLY)) {
4782 oflags = GENERIC_READ;
4783 len = OPEN_EXISTING;
4785 oflags = GENERIC_READ|GENERIC_WRITE;
4788 mode = FILE_ATTRIBUTE_NORMAL;
4789 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
4792 env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4793 NULL, len, mode, NULL);
4796 if (F_ISSET(flags, MDB_RDONLY))
4799 oflags = O_RDWR | O_CREAT;
4801 env->me_fd = open(dpath, oflags, mode);
4803 if (env->me_fd == INVALID_HANDLE_VALUE) {
4808 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4809 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4814 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4815 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4816 env->me_mfd = env->me_fd;
4818 /* Synchronous fd for meta writes. Needed even with
4819 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4822 len = OPEN_EXISTING;
4823 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
4826 env->me_mfd = CreateFileW(wpath, oflags,
4827 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4828 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4832 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4834 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4839 DPRINTF(("opened dbenv %p", (void *) env));
4841 rc = mdb_env_share_locks(env, &excl);
4845 if (!(flags & MDB_RDONLY)) {
4847 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4848 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4849 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4850 (txn = calloc(1, size)))
4852 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4853 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4854 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4855 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4857 txn->mt_dbxs = env->me_dbxs;
4858 txn->mt_flags = MDB_TXN_FINISHED;
4868 mdb_env_close0(env, excl);
4874 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4876 mdb_env_close0(MDB_env *env, int excl)
4880 if (!(env->me_flags & MDB_ENV_ACTIVE))
4883 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4885 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4886 free(env->me_dbxs[i].md_name.mv_data);
4891 free(env->me_dbiseqs);
4892 free(env->me_dbflags);
4894 free(env->me_dirty_list);
4896 mdb_midl_free(env->me_free_pgs);
4898 if (env->me_flags & MDB_ENV_TXKEY) {
4899 pthread_key_delete(env->me_txkey);
4901 /* Delete our key from the global list */
4902 for (i=0; i<mdb_tls_nkeys; i++)
4903 if (mdb_tls_keys[i] == env->me_txkey) {
4904 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4912 munmap(env->me_map, env->me_mapsize);
4914 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4915 (void) close(env->me_mfd);
4916 if (env->me_fd != INVALID_HANDLE_VALUE)
4917 (void) close(env->me_fd);
4919 MDB_PID_T pid = env->me_pid;
4920 /* Clearing readers is done in this function because
4921 * me_txkey with its destructor must be disabled first.
4923 * We skip the the reader mutex, so we touch only
4924 * data owned by this process (me_close_readers and
4925 * our readers), and clear each reader atomically.
4927 for (i = env->me_close_readers; --i >= 0; )
4928 if (env->me_txns->mti_readers[i].mr_pid == pid)
4929 env->me_txns->mti_readers[i].mr_pid = 0;
4931 if (env->me_rmutex) {
4932 CloseHandle(env->me_rmutex);
4933 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4935 /* Windows automatically destroys the mutexes when
4936 * the last handle closes.
4938 #elif defined(MDB_USE_POSIX_SEM)
4939 if (env->me_rmutex != SEM_FAILED) {
4940 sem_close(env->me_rmutex);
4941 if (env->me_wmutex != SEM_FAILED)
4942 sem_close(env->me_wmutex);
4943 /* If we have the filelock: If we are the
4944 * only remaining user, clean up semaphores.
4947 mdb_env_excl_lock(env, &excl);
4949 sem_unlink(env->me_txns->mti_rmname);
4950 sem_unlink(env->me_txns->mti_wmname);
4954 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4956 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4959 /* Unlock the lockfile. Windows would have unlocked it
4960 * after closing anyway, but not necessarily at once.
4962 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4965 (void) close(env->me_lfd);
4968 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4972 mdb_env_close(MDB_env *env)
4979 VGMEMP_DESTROY(env);
4980 while ((dp = env->me_dpages) != NULL) {
4981 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4982 env->me_dpages = dp->mp_next;
4986 mdb_env_close0(env, 0);
4990 /** Compare two items pointing at aligned size_t's */
4992 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4994 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4995 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4998 /** Compare two items pointing at aligned unsigned int's.
5000 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5001 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5004 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5006 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5007 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5010 /** Compare two items pointing at unsigned ints of unknown alignment.
5011 * Nodes and keys are guaranteed to be 2-byte aligned.
5014 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5016 #if BYTE_ORDER == LITTLE_ENDIAN
5017 unsigned short *u, *c;
5020 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5021 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5024 } while(!x && u > (unsigned short *)a->mv_data);
5027 unsigned short *u, *c, *end;
5030 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5031 u = (unsigned short *)a->mv_data;
5032 c = (unsigned short *)b->mv_data;
5035 } while(!x && u < end);
5040 /** Compare two items lexically */
5042 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5049 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5055 diff = memcmp(a->mv_data, b->mv_data, len);
5056 return diff ? diff : len_diff<0 ? -1 : len_diff;
5059 /** Compare two items in reverse byte order */
5061 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5063 const unsigned char *p1, *p2, *p1_lim;
5067 p1_lim = (const unsigned char *)a->mv_data;
5068 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5069 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5071 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5077 while (p1 > p1_lim) {
5078 diff = *--p1 - *--p2;
5082 return len_diff<0 ? -1 : len_diff;
5085 /** Search for key within a page, using binary search.
5086 * Returns the smallest entry larger or equal to the key.
5087 * If exactp is non-null, stores whether the found entry was an exact match
5088 * in *exactp (1 or 0).
5089 * Updates the cursor index with the index of the found entry.
5090 * If no entry larger or equal to the key is found, returns NULL.
5093 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5095 unsigned int i = 0, nkeys;
5098 MDB_page *mp = mc->mc_pg[mc->mc_top];
5099 MDB_node *node = NULL;
5104 nkeys = NUMKEYS(mp);
5106 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5107 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5110 low = IS_LEAF(mp) ? 0 : 1;
5112 cmp = mc->mc_dbx->md_cmp;
5114 /* Branch pages have no data, so if using integer keys,
5115 * alignment is guaranteed. Use faster mdb_cmp_int.
5117 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5118 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5125 nodekey.mv_size = mc->mc_db->md_pad;
5126 node = NODEPTR(mp, 0); /* fake */
5127 while (low <= high) {
5128 i = (low + high) >> 1;
5129 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5130 rc = cmp(key, &nodekey);
5131 DPRINTF(("found leaf index %u [%s], rc = %i",
5132 i, DKEY(&nodekey), rc));
5141 while (low <= high) {
5142 i = (low + high) >> 1;
5144 node = NODEPTR(mp, i);
5145 nodekey.mv_size = NODEKSZ(node);
5146 nodekey.mv_data = NODEKEY(node);
5148 rc = cmp(key, &nodekey);
5151 DPRINTF(("found leaf index %u [%s], rc = %i",
5152 i, DKEY(&nodekey), rc));
5154 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5155 i, DKEY(&nodekey), NODEPGNO(node), rc));
5166 if (rc > 0) { /* Found entry is less than the key. */
5167 i++; /* Skip to get the smallest entry larger than key. */
5169 node = NODEPTR(mp, i);
5172 *exactp = (rc == 0 && nkeys > 0);
5173 /* store the key index */
5174 mc->mc_ki[mc->mc_top] = i;
5176 /* There is no entry larger or equal to the key. */
5179 /* nodeptr is fake for LEAF2 */
5185 mdb_cursor_adjust(MDB_cursor *mc, func)
5189 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5190 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5197 /** Pop a page off the top of the cursor's stack. */
5199 mdb_cursor_pop(MDB_cursor *mc)
5202 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5203 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5209 mc->mc_flags &= ~C_INITIALIZED;
5214 /** Push a page onto the top of the cursor's stack. */
5216 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5218 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5219 DDBI(mc), (void *) mc));
5221 if (mc->mc_snum >= CURSOR_STACK) {
5222 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5223 return MDB_CURSOR_FULL;
5226 mc->mc_top = mc->mc_snum++;
5227 mc->mc_pg[mc->mc_top] = mp;
5228 mc->mc_ki[mc->mc_top] = 0;
5233 /** Find the address of the page corresponding to a given page number.
5234 * @param[in] mc the cursor accessing the page.
5235 * @param[in] pgno the page number for the page to retrieve.
5236 * @param[out] ret address of a pointer where the page's address will be stored.
5237 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5238 * @return 0 on success, non-zero on failure.
5241 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
5243 MDB_txn *txn = mc->mc_txn;
5244 MDB_env *env = txn->mt_env;
5248 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5252 MDB_ID2L dl = tx2->mt_u.dirty_list;
5254 /* Spilled pages were dirtied in this txn and flushed
5255 * because the dirty list got full. Bring this page
5256 * back in from the map (but don't unspill it here,
5257 * leave that unless page_touch happens again).
5259 if (tx2->mt_spill_pgs) {
5260 MDB_ID pn = pgno << 1;
5261 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5262 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5263 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5268 unsigned x = mdb_mid2l_search(dl, pgno);
5269 if (x <= dl[0].mid && dl[x].mid == pgno) {
5275 } while ((tx2 = tx2->mt_parent) != NULL);
5278 if (pgno < txn->mt_next_pgno) {
5280 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5282 DPRINTF(("page %"Z"u not found", pgno));
5283 txn->mt_flags |= MDB_TXN_ERROR;
5284 return MDB_PAGE_NOTFOUND;
5294 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5295 * The cursor is at the root page, set up the rest of it.
5298 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5300 MDB_page *mp = mc->mc_pg[mc->mc_top];
5304 while (IS_BRANCH(mp)) {
5308 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5309 /* Don't assert on branch pages in the FreeDB. We can get here
5310 * while in the process of rebalancing a FreeDB branch page; we must
5311 * let that proceed. ITS#8336
5313 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
5314 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5316 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5318 if (flags & MDB_PS_LAST)
5319 i = NUMKEYS(mp) - 1;
5322 node = mdb_node_search(mc, key, &exact);
5324 i = NUMKEYS(mp) - 1;
5326 i = mc->mc_ki[mc->mc_top];
5328 mdb_cassert(mc, i > 0);
5332 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5335 mdb_cassert(mc, i < NUMKEYS(mp));
5336 node = NODEPTR(mp, i);
5338 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5341 mc->mc_ki[mc->mc_top] = i;
5342 if ((rc = mdb_cursor_push(mc, mp)))
5345 if (flags & MDB_PS_MODIFY) {
5346 if ((rc = mdb_page_touch(mc)) != 0)
5348 mp = mc->mc_pg[mc->mc_top];
5353 DPRINTF(("internal error, index points to a %02X page!?",
5355 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5356 return MDB_CORRUPTED;
5359 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5360 key ? DKEY(key) : "null"));
5361 mc->mc_flags |= C_INITIALIZED;
5362 mc->mc_flags &= ~C_EOF;
5367 /** Search for the lowest key under the current branch page.
5368 * This just bypasses a NUMKEYS check in the current page
5369 * before calling mdb_page_search_root(), because the callers
5370 * are all in situations where the current page is known to
5374 mdb_page_search_lowest(MDB_cursor *mc)
5376 MDB_page *mp = mc->mc_pg[mc->mc_top];
5377 MDB_node *node = NODEPTR(mp, 0);
5380 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5383 mc->mc_ki[mc->mc_top] = 0;
5384 if ((rc = mdb_cursor_push(mc, mp)))
5386 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5389 /** Search for the page a given key should be in.
5390 * Push it and its parent pages on the cursor stack.
5391 * @param[in,out] mc the cursor for this operation.
5392 * @param[in] key the key to search for, or NULL for first/last page.
5393 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5394 * are touched (updated with new page numbers).
5395 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5396 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5397 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5398 * @return 0 on success, non-zero on failure.
5401 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5406 /* Make sure the txn is still viable, then find the root from
5407 * the txn's db table and set it as the root of the cursor's stack.
5409 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5410 DPUTS("transaction may not be used now");
5413 /* Make sure we're using an up-to-date root */
5414 if (*mc->mc_dbflag & DB_STALE) {
5416 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5418 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5419 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5426 MDB_node *leaf = mdb_node_search(&mc2,
5427 &mc->mc_dbx->md_name, &exact);
5429 return MDB_NOTFOUND;
5430 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5431 return MDB_INCOMPATIBLE; /* not a named DB */
5432 rc = mdb_node_read(&mc2, leaf, &data);
5435 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5437 /* The txn may not know this DBI, or another process may
5438 * have dropped and recreated the DB with other flags.
5440 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5441 return MDB_INCOMPATIBLE;
5442 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5444 *mc->mc_dbflag &= ~DB_STALE;
5446 root = mc->mc_db->md_root;
5448 if (root == P_INVALID) { /* Tree is empty. */
5449 DPUTS("tree is empty");
5450 return MDB_NOTFOUND;
5454 mdb_cassert(mc, root > 1);
5455 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5456 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
5462 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5463 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5465 if (flags & MDB_PS_MODIFY) {
5466 if ((rc = mdb_page_touch(mc)))
5470 if (flags & MDB_PS_ROOTONLY)
5473 return mdb_page_search_root(mc, key, flags);
5477 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5479 MDB_txn *txn = mc->mc_txn;
5480 pgno_t pg = mp->mp_pgno;
5481 unsigned x = 0, ovpages = mp->mp_pages;
5482 MDB_env *env = txn->mt_env;
5483 MDB_IDL sl = txn->mt_spill_pgs;
5484 MDB_ID pn = pg << 1;
5487 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5488 /* If the page is dirty or on the spill list we just acquired it,
5489 * so we should give it back to our current free list, if any.
5490 * Otherwise put it onto the list of pages we freed in this txn.
5492 * Won't create me_pghead: me_pglast must be inited along with it.
5493 * Unsupported in nested txns: They would need to hide the page
5494 * range in ancestor txns' dirty and spilled lists.
5496 if (env->me_pghead &&
5498 ((mp->mp_flags & P_DIRTY) ||
5499 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5503 MDB_ID2 *dl, ix, iy;
5504 rc = mdb_midl_need(&env->me_pghead, ovpages);
5507 if (!(mp->mp_flags & P_DIRTY)) {
5508 /* This page is no longer spilled */
5515 /* Remove from dirty list */
5516 dl = txn->mt_u.dirty_list;
5518 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5524 mdb_cassert(mc, x > 1);
5526 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5527 txn->mt_flags |= MDB_TXN_ERROR;
5528 return MDB_CORRUPTED;
5531 txn->mt_dirty_room++;
5532 if (!(env->me_flags & MDB_WRITEMAP))
5533 mdb_dpage_free(env, mp);
5535 /* Insert in me_pghead */
5536 mop = env->me_pghead;
5537 j = mop[0] + ovpages;
5538 for (i = mop[0]; i && mop[i] < pg; i--)
5544 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5548 mc->mc_db->md_overflow_pages -= ovpages;
5552 /** Return the data associated with a given node.
5553 * @param[in] mc The cursor for this operation.
5554 * @param[in] leaf The node being read.
5555 * @param[out] data Updated to point to the node's data.
5556 * @return 0 on success, non-zero on failure.
5559 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
5561 MDB_page *omp; /* overflow page */
5565 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5566 data->mv_size = NODEDSZ(leaf);
5567 data->mv_data = NODEDATA(leaf);
5571 /* Read overflow data.
5573 data->mv_size = NODEDSZ(leaf);
5574 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5575 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
5576 DPRINTF(("read overflow page %"Z"u failed", pgno));
5579 data->mv_data = METADATA(omp);
5585 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5586 MDB_val *key, MDB_val *data)
5593 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5595 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5598 if (txn->mt_flags & MDB_TXN_BLOCKED)
5601 mdb_cursor_init(&mc, txn, dbi, &mx);
5602 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5605 /** Find a sibling for a page.
5606 * Replaces the page at the top of the cursor's stack with the
5607 * specified sibling, if one exists.
5608 * @param[in] mc The cursor for this operation.
5609 * @param[in] move_right Non-zero if the right sibling is requested,
5610 * otherwise the left sibling.
5611 * @return 0 on success, non-zero on failure.
5614 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5620 if (mc->mc_snum < 2) {
5621 return MDB_NOTFOUND; /* root has no siblings */
5625 DPRINTF(("parent page is page %"Z"u, index %u",
5626 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5628 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5629 : (mc->mc_ki[mc->mc_top] == 0)) {
5630 DPRINTF(("no more keys left, moving to %s sibling",
5631 move_right ? "right" : "left"));
5632 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5633 /* undo cursor_pop before returning */
5640 mc->mc_ki[mc->mc_top]++;
5642 mc->mc_ki[mc->mc_top]--;
5643 DPRINTF(("just moving to %s index key %u",
5644 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5646 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5648 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5649 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
5650 /* mc will be inconsistent if caller does mc_snum++ as above */
5651 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5655 mdb_cursor_push(mc, mp);
5657 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5662 /** Move the cursor to the next data item. */
5664 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5670 if ((mc->mc_flags & C_EOF) ||
5671 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
5672 return MDB_NOTFOUND;
5674 if (!(mc->mc_flags & C_INITIALIZED))
5675 return mdb_cursor_first(mc, key, data);
5677 mp = mc->mc_pg[mc->mc_top];
5679 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5680 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5681 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5682 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5683 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5684 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5685 if (rc == MDB_SUCCESS)
5686 MDB_GET_KEY(leaf, key);
5691 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5692 if (op == MDB_NEXT_DUP)
5693 return MDB_NOTFOUND;
5697 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5698 mdb_dbg_pgno(mp), (void *) mc));
5699 if (mc->mc_flags & C_DEL) {
5700 mc->mc_flags ^= C_DEL;
5704 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5705 DPUTS("=====> move to next sibling page");
5706 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5707 mc->mc_flags |= C_EOF;
5710 mp = mc->mc_pg[mc->mc_top];
5711 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5713 mc->mc_ki[mc->mc_top]++;
5716 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5717 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5720 key->mv_size = mc->mc_db->md_pad;
5721 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5725 mdb_cassert(mc, IS_LEAF(mp));
5726 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5728 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5729 mdb_xcursor_init1(mc, leaf);
5732 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5735 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5736 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5737 if (rc != MDB_SUCCESS)
5742 MDB_GET_KEY(leaf, key);
5746 /** Move the cursor to the previous data item. */
5748 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5754 if (!(mc->mc_flags & C_INITIALIZED)) {
5755 rc = mdb_cursor_last(mc, key, data);
5758 mc->mc_ki[mc->mc_top]++;
5761 mp = mc->mc_pg[mc->mc_top];
5763 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5764 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5765 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5766 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5767 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5768 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5769 if (rc == MDB_SUCCESS) {
5770 MDB_GET_KEY(leaf, key);
5771 mc->mc_flags &= ~C_EOF;
5777 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5778 if (op == MDB_PREV_DUP)
5779 return MDB_NOTFOUND;
5783 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5784 mdb_dbg_pgno(mp), (void *) mc));
5786 mc->mc_flags &= ~(C_EOF|C_DEL);
5788 if (mc->mc_ki[mc->mc_top] == 0) {
5789 DPUTS("=====> move to prev sibling page");
5790 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5793 mp = mc->mc_pg[mc->mc_top];
5794 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5795 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5797 mc->mc_ki[mc->mc_top]--;
5799 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5800 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5803 key->mv_size = mc->mc_db->md_pad;
5804 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5808 mdb_cassert(mc, IS_LEAF(mp));
5809 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5811 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5812 mdb_xcursor_init1(mc, leaf);
5815 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5818 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5819 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5820 if (rc != MDB_SUCCESS)
5825 MDB_GET_KEY(leaf, key);
5829 /** Set the cursor on a specific data item. */
5831 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5832 MDB_cursor_op op, int *exactp)
5836 MDB_node *leaf = NULL;
5839 if (key->mv_size == 0)
5840 return MDB_BAD_VALSIZE;
5843 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5845 /* See if we're already on the right page */
5846 if (mc->mc_flags & C_INITIALIZED) {
5849 mp = mc->mc_pg[mc->mc_top];
5851 mc->mc_ki[mc->mc_top] = 0;
5852 return MDB_NOTFOUND;
5854 if (mp->mp_flags & P_LEAF2) {
5855 nodekey.mv_size = mc->mc_db->md_pad;
5856 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5858 leaf = NODEPTR(mp, 0);
5859 MDB_GET_KEY2(leaf, nodekey);
5861 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5863 /* Probably happens rarely, but first node on the page
5864 * was the one we wanted.
5866 mc->mc_ki[mc->mc_top] = 0;
5873 unsigned int nkeys = NUMKEYS(mp);
5875 if (mp->mp_flags & P_LEAF2) {
5876 nodekey.mv_data = LEAF2KEY(mp,
5877 nkeys-1, nodekey.mv_size);
5879 leaf = NODEPTR(mp, nkeys-1);
5880 MDB_GET_KEY2(leaf, nodekey);
5882 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5884 /* last node was the one we wanted */
5885 mc->mc_ki[mc->mc_top] = nkeys-1;
5891 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5892 /* This is definitely the right page, skip search_page */
5893 if (mp->mp_flags & P_LEAF2) {
5894 nodekey.mv_data = LEAF2KEY(mp,
5895 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5897 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5898 MDB_GET_KEY2(leaf, nodekey);
5900 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5902 /* current node was the one we wanted */
5912 /* If any parents have right-sibs, search.
5913 * Otherwise, there's nothing further.
5915 for (i=0; i<mc->mc_top; i++)
5917 NUMKEYS(mc->mc_pg[i])-1)
5919 if (i == mc->mc_top) {
5920 /* There are no other pages */
5921 mc->mc_ki[mc->mc_top] = nkeys;
5922 return MDB_NOTFOUND;
5926 /* There are no other pages */
5927 mc->mc_ki[mc->mc_top] = 0;
5928 if (op == MDB_SET_RANGE && !exactp) {
5932 return MDB_NOTFOUND;
5938 rc = mdb_page_search(mc, key, 0);
5939 if (rc != MDB_SUCCESS)
5942 mp = mc->mc_pg[mc->mc_top];
5943 mdb_cassert(mc, IS_LEAF(mp));
5946 leaf = mdb_node_search(mc, key, exactp);
5947 if (exactp != NULL && !*exactp) {
5948 /* MDB_SET specified and not an exact match. */
5949 return MDB_NOTFOUND;
5953 DPUTS("===> inexact leaf not found, goto sibling");
5954 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5955 mc->mc_flags |= C_EOF;
5956 return rc; /* no entries matched */
5958 mp = mc->mc_pg[mc->mc_top];
5959 mdb_cassert(mc, IS_LEAF(mp));
5960 leaf = NODEPTR(mp, 0);
5964 mc->mc_flags |= C_INITIALIZED;
5965 mc->mc_flags &= ~C_EOF;
5968 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5969 key->mv_size = mc->mc_db->md_pad;
5970 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5975 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5976 mdb_xcursor_init1(mc, leaf);
5979 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5980 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5981 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5984 if (op == MDB_GET_BOTH) {
5990 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5991 if (rc != MDB_SUCCESS)
5994 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5997 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
5999 dcmp = mc->mc_dbx->md_dcmp;
6000 #if UINT_MAX < SIZE_MAX
6001 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6002 dcmp = mdb_cmp_clong;
6004 rc = dcmp(data, &olddata);
6006 if (op == MDB_GET_BOTH || rc > 0)
6007 return MDB_NOTFOUND;
6014 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6015 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6020 /* The key already matches in all other cases */
6021 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6022 MDB_GET_KEY(leaf, key);
6023 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6028 /** Move the cursor to the first item in the database. */
6030 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6036 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6038 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6039 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6040 if (rc != MDB_SUCCESS)
6043 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6045 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6046 mc->mc_flags |= C_INITIALIZED;
6047 mc->mc_flags &= ~C_EOF;
6049 mc->mc_ki[mc->mc_top] = 0;
6051 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6052 key->mv_size = mc->mc_db->md_pad;
6053 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6058 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6059 mdb_xcursor_init1(mc, leaf);
6060 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6064 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6068 MDB_GET_KEY(leaf, key);
6072 /** Move the cursor to the last item in the database. */
6074 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6080 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6082 if (!(mc->mc_flags & C_EOF)) {
6084 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6085 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6086 if (rc != MDB_SUCCESS)
6089 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6092 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6093 mc->mc_flags |= C_INITIALIZED|C_EOF;
6094 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6096 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6097 key->mv_size = mc->mc_db->md_pad;
6098 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6103 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6104 mdb_xcursor_init1(mc, leaf);
6105 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6109 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6114 MDB_GET_KEY(leaf, key);
6119 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6124 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6129 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6133 case MDB_GET_CURRENT:
6134 if (!(mc->mc_flags & C_INITIALIZED)) {
6137 MDB_page *mp = mc->mc_pg[mc->mc_top];
6138 int nkeys = NUMKEYS(mp);
6139 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6140 mc->mc_ki[mc->mc_top] = nkeys;
6146 key->mv_size = mc->mc_db->md_pad;
6147 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6149 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6150 MDB_GET_KEY(leaf, key);
6152 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6153 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6155 rc = mdb_node_read(mc, leaf, data);
6162 case MDB_GET_BOTH_RANGE:
6167 if (mc->mc_xcursor == NULL) {
6168 rc = MDB_INCOMPATIBLE;
6178 rc = mdb_cursor_set(mc, key, data, op,
6179 op == MDB_SET_RANGE ? NULL : &exact);
6182 case MDB_GET_MULTIPLE:
6183 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6187 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6188 rc = MDB_INCOMPATIBLE;
6192 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6193 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6196 case MDB_NEXT_MULTIPLE:
6201 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6202 rc = MDB_INCOMPATIBLE;
6205 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6206 if (rc == MDB_SUCCESS) {
6207 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6210 mx = &mc->mc_xcursor->mx_cursor;
6211 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6213 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6214 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6220 case MDB_PREV_MULTIPLE:
6225 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6226 rc = MDB_INCOMPATIBLE;
6229 if (!(mc->mc_flags & C_INITIALIZED))
6230 rc = mdb_cursor_last(mc, key, data);
6233 if (rc == MDB_SUCCESS) {
6234 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
6235 if (mx->mc_flags & C_INITIALIZED) {
6236 rc = mdb_cursor_sibling(mx, 0);
6237 if (rc == MDB_SUCCESS)
6246 case MDB_NEXT_NODUP:
6247 rc = mdb_cursor_next(mc, key, data, op);
6251 case MDB_PREV_NODUP:
6252 rc = mdb_cursor_prev(mc, key, data, op);
6255 rc = mdb_cursor_first(mc, key, data);
6258 mfunc = mdb_cursor_first;
6260 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6264 if (mc->mc_xcursor == NULL) {
6265 rc = MDB_INCOMPATIBLE;
6269 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6270 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6271 MDB_GET_KEY(leaf, key);
6272 rc = mdb_node_read(mc, leaf, data);
6276 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6280 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6283 rc = mdb_cursor_last(mc, key, data);
6286 mfunc = mdb_cursor_last;
6289 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6294 if (mc->mc_flags & C_DEL)
6295 mc->mc_flags ^= C_DEL;
6300 /** Touch all the pages in the cursor stack. Set mc_top.
6301 * Makes sure all the pages are writable, before attempting a write operation.
6302 * @param[in] mc The cursor to operate on.
6305 mdb_cursor_touch(MDB_cursor *mc)
6307 int rc = MDB_SUCCESS;
6309 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & (DB_DIRTY|DB_DUPDATA))) {
6310 /* Touch DB record of named DB */
6313 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6315 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6316 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6319 *mc->mc_dbflag |= DB_DIRTY;
6324 rc = mdb_page_touch(mc);
6325 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6326 mc->mc_top = mc->mc_snum-1;
6331 /** Do not spill pages to disk if txn is getting full, may fail instead */
6332 #define MDB_NOSPILL 0x8000
6335 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6339 MDB_node *leaf = NULL;
6340 MDB_page *fp, *mp, *sub_root = NULL;
6342 MDB_val xdata, *rdata, dkey, olddata;
6344 int do_sub = 0, insert_key, insert_data;
6345 unsigned int mcount = 0, dcount = 0, nospill;
6348 unsigned int nflags;
6351 if (mc == NULL || key == NULL)
6354 env = mc->mc_txn->mt_env;
6356 /* Check this first so counter will always be zero on any
6359 if (flags & MDB_MULTIPLE) {
6360 dcount = data[1].mv_size;
6361 data[1].mv_size = 0;
6362 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6363 return MDB_INCOMPATIBLE;
6366 nospill = flags & MDB_NOSPILL;
6367 flags &= ~MDB_NOSPILL;
6369 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6370 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6372 if (key->mv_size-1 >= ENV_MAXKEY(env))
6373 return MDB_BAD_VALSIZE;
6375 #if SIZE_MAX > MAXDATASIZE
6376 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6377 return MDB_BAD_VALSIZE;
6379 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6380 return MDB_BAD_VALSIZE;
6383 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6384 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6388 if (flags == MDB_CURRENT) {
6389 if (!(mc->mc_flags & C_INITIALIZED))
6392 } else if (mc->mc_db->md_root == P_INVALID) {
6393 /* new database, cursor has nothing to point to */
6396 mc->mc_flags &= ~C_INITIALIZED;
6401 if (flags & MDB_APPEND) {
6403 rc = mdb_cursor_last(mc, &k2, &d2);
6405 rc = mc->mc_dbx->md_cmp(key, &k2);
6408 mc->mc_ki[mc->mc_top]++;
6410 /* new key is <= last key */
6415 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6417 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6418 DPRINTF(("duplicate key [%s]", DKEY(key)));
6420 return MDB_KEYEXIST;
6422 if (rc && rc != MDB_NOTFOUND)
6426 if (mc->mc_flags & C_DEL)
6427 mc->mc_flags ^= C_DEL;
6429 /* Cursor is positioned, check for room in the dirty list */
6431 if (flags & MDB_MULTIPLE) {
6433 xdata.mv_size = data->mv_size * dcount;
6437 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6441 if (rc == MDB_NO_ROOT) {
6443 /* new database, write a root leaf page */
6444 DPUTS("allocating new root leaf page");
6445 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6448 mdb_cursor_push(mc, np);
6449 mc->mc_db->md_root = np->mp_pgno;
6450 mc->mc_db->md_depth++;
6451 *mc->mc_dbflag |= DB_DIRTY;
6452 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6454 np->mp_flags |= P_LEAF2;
6455 mc->mc_flags |= C_INITIALIZED;
6457 /* make sure all cursor pages are writable */
6458 rc2 = mdb_cursor_touch(mc);
6463 insert_key = insert_data = rc;
6465 /* The key does not exist */
6466 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6467 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6468 LEAFSIZE(key, data) > env->me_nodemax)
6470 /* Too big for a node, insert in sub-DB. Set up an empty
6471 * "old sub-page" for prep_subDB to expand to a full page.
6473 fp_flags = P_LEAF|P_DIRTY;
6475 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6476 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6477 olddata.mv_size = PAGEHDRSZ;
6481 /* there's only a key anyway, so this is a no-op */
6482 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6484 unsigned int ksize = mc->mc_db->md_pad;
6485 if (key->mv_size != ksize)
6486 return MDB_BAD_VALSIZE;
6487 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6488 memcpy(ptr, key->mv_data, ksize);
6490 /* if overwriting slot 0 of leaf, need to
6491 * update branch key if there is a parent page
6493 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6494 unsigned short dtop = 1;
6496 /* slot 0 is always an empty key, find real slot */
6497 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6501 if (mc->mc_ki[mc->mc_top])
6502 rc2 = mdb_update_key(mc, key);
6513 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6514 olddata.mv_size = NODEDSZ(leaf);
6515 olddata.mv_data = NODEDATA(leaf);
6518 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6519 /* Prepare (sub-)page/sub-DB to accept the new item,
6520 * if needed. fp: old sub-page or a header faking
6521 * it. mp: new (sub-)page. offset: growth in page
6522 * size. xdata: node data with new page or DB.
6524 unsigned i, offset = 0;
6525 mp = fp = xdata.mv_data = env->me_pbuf;
6526 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6528 /* Was a single item before, must convert now */
6529 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6531 /* Just overwrite the current item */
6532 if (flags == MDB_CURRENT)
6534 dcmp = mc->mc_dbx->md_dcmp;
6535 #if UINT_MAX < SIZE_MAX
6536 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6537 dcmp = mdb_cmp_clong;
6539 /* does data match? */
6540 if (!dcmp(data, &olddata)) {
6541 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
6542 return MDB_KEYEXIST;
6547 /* Back up original data item */
6548 dkey.mv_size = olddata.mv_size;
6549 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6551 /* Make sub-page header for the dup items, with dummy body */
6552 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6553 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6554 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6555 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6556 fp->mp_flags |= P_LEAF2;
6557 fp->mp_pad = data->mv_size;
6558 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6560 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6561 (dkey.mv_size & 1) + (data->mv_size & 1);
6563 fp->mp_upper = xdata.mv_size - PAGEBASE;
6564 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6565 } else if (leaf->mn_flags & F_SUBDATA) {
6566 /* Data is on sub-DB, just store it */
6567 flags |= F_DUPDATA|F_SUBDATA;
6570 /* Data is on sub-page */
6571 fp = olddata.mv_data;
6574 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6575 offset = EVEN(NODESIZE + sizeof(indx_t) +
6579 offset = fp->mp_pad;
6580 if (SIZELEFT(fp) < offset) {
6581 offset *= 4; /* space for 4 more */
6584 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6586 fp->mp_flags |= P_DIRTY;
6587 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6588 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6592 xdata.mv_size = olddata.mv_size + offset;
6595 fp_flags = fp->mp_flags;
6596 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6597 /* Too big for a sub-page, convert to sub-DB */
6598 fp_flags &= ~P_SUBP;
6600 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6601 fp_flags |= P_LEAF2;
6602 dummy.md_pad = fp->mp_pad;
6603 dummy.md_flags = MDB_DUPFIXED;
6604 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6605 dummy.md_flags |= MDB_INTEGERKEY;
6611 dummy.md_branch_pages = 0;
6612 dummy.md_leaf_pages = 1;
6613 dummy.md_overflow_pages = 0;
6614 dummy.md_entries = NUMKEYS(fp);
6615 xdata.mv_size = sizeof(MDB_db);
6616 xdata.mv_data = &dummy;
6617 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6619 offset = env->me_psize - olddata.mv_size;
6620 flags |= F_DUPDATA|F_SUBDATA;
6621 dummy.md_root = mp->mp_pgno;
6625 mp->mp_flags = fp_flags | P_DIRTY;
6626 mp->mp_pad = fp->mp_pad;
6627 mp->mp_lower = fp->mp_lower;
6628 mp->mp_upper = fp->mp_upper + offset;
6629 if (fp_flags & P_LEAF2) {
6630 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6632 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6633 olddata.mv_size - fp->mp_upper - PAGEBASE);
6634 for (i=0; i<NUMKEYS(fp); i++)
6635 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6643 mdb_node_del(mc, 0);
6647 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6648 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6649 return MDB_INCOMPATIBLE;
6650 /* overflow page overwrites need special handling */
6651 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6654 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6656 memcpy(&pg, olddata.mv_data, sizeof(pg));
6657 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
6659 ovpages = omp->mp_pages;
6661 /* Is the ov page large enough? */
6662 if (ovpages >= dpages) {
6663 if (!(omp->mp_flags & P_DIRTY) &&
6664 (level || (env->me_flags & MDB_WRITEMAP)))
6666 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6669 level = 0; /* dirty in this txn or clean */
6672 if (omp->mp_flags & P_DIRTY) {
6673 /* yes, overwrite it. Note in this case we don't
6674 * bother to try shrinking the page if the new data
6675 * is smaller than the overflow threshold.
6678 /* It is writable only in a parent txn */
6679 size_t sz = (size_t) env->me_psize * ovpages, off;
6680 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6686 /* Note - this page is already counted in parent's dirty_room */
6687 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6688 mdb_cassert(mc, rc2 == 0);
6689 /* Currently we make the page look as with put() in the
6690 * parent txn, in case the user peeks at MDB_RESERVEd
6691 * or unused parts. Some users treat ovpages specially.
6693 if (!(flags & MDB_RESERVE)) {
6694 /* Skip the part where LMDB will put *data.
6695 * Copy end of page, adjusting alignment so
6696 * compiler may copy words instead of bytes.
6698 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6699 memcpy((size_t *)((char *)np + off),
6700 (size_t *)((char *)omp + off), sz - off);
6703 memcpy(np, omp, sz); /* Copy beginning of page */
6706 SETDSZ(leaf, data->mv_size);
6707 if (F_ISSET(flags, MDB_RESERVE))
6708 data->mv_data = METADATA(omp);
6710 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6714 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6716 } else if (data->mv_size == olddata.mv_size) {
6717 /* same size, just replace it. Note that we could
6718 * also reuse this node if the new data is smaller,
6719 * but instead we opt to shrink the node in that case.
6721 if (F_ISSET(flags, MDB_RESERVE))
6722 data->mv_data = olddata.mv_data;
6723 else if (!(mc->mc_flags & C_SUB))
6724 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6726 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6731 mdb_node_del(mc, 0);
6737 nflags = flags & NODE_ADD_FLAGS;
6738 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6739 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6740 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6741 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6743 nflags |= MDB_SPLIT_REPLACE;
6744 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6746 /* There is room already in this leaf page. */
6747 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6749 /* Adjust other cursors pointing to mp */
6750 MDB_cursor *m2, *m3;
6751 MDB_dbi dbi = mc->mc_dbi;
6752 unsigned i = mc->mc_top;
6753 MDB_page *mp = mc->mc_pg[i];
6755 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6756 if (mc->mc_flags & C_SUB)
6757 m3 = &m2->mc_xcursor->mx_cursor;
6760 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
6761 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
6764 if (XCURSOR_INITED(m3))
6765 XCURSOR_REFRESH(m3, mp, m3->mc_ki[i]);
6770 if (rc == MDB_SUCCESS) {
6771 /* Now store the actual data in the child DB. Note that we're
6772 * storing the user data in the keys field, so there are strict
6773 * size limits on dupdata. The actual data fields of the child
6774 * DB are all zero size.
6777 int xflags, new_dupdata;
6782 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6783 if (flags & MDB_CURRENT) {
6784 xflags = MDB_CURRENT|MDB_NOSPILL;
6786 mdb_xcursor_init1(mc, leaf);
6787 xflags = (flags & MDB_NODUPDATA) ?
6788 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6791 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6792 new_dupdata = (int)dkey.mv_size;
6793 /* converted, write the original data first */
6795 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6798 /* we've done our job */
6801 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6802 /* Adjust other cursors pointing to mp */
6804 MDB_xcursor *mx = mc->mc_xcursor;
6805 unsigned i = mc->mc_top;
6806 MDB_page *mp = mc->mc_pg[i];
6807 int nkeys = NUMKEYS(mp);
6809 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6810 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6811 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6812 if (m2->mc_pg[i] == mp) {
6813 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6814 mdb_xcursor_init2(m2, mx, new_dupdata);
6815 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
6816 XCURSOR_REFRESH(m2, mp, m2->mc_ki[i]);
6821 ecount = mc->mc_xcursor->mx_db.md_entries;
6822 if (flags & MDB_APPENDDUP)
6823 xflags |= MDB_APPEND;
6824 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6825 if (flags & F_SUBDATA) {
6826 void *db = NODEDATA(leaf);
6827 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6829 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6831 /* Increment count unless we just replaced an existing item. */
6833 mc->mc_db->md_entries++;
6835 /* Invalidate txn if we created an empty sub-DB */
6838 /* If we succeeded and the key didn't exist before,
6839 * make sure the cursor is marked valid.
6841 mc->mc_flags |= C_INITIALIZED;
6843 if (flags & MDB_MULTIPLE) {
6846 /* let caller know how many succeeded, if any */
6847 data[1].mv_size = mcount;
6848 if (mcount < dcount) {
6849 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6850 insert_key = insert_data = 0;
6857 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6860 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6865 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6871 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6872 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6874 if (!(mc->mc_flags & C_INITIALIZED))
6877 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6878 return MDB_NOTFOUND;
6880 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6883 rc = mdb_cursor_touch(mc);
6887 mp = mc->mc_pg[mc->mc_top];
6890 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6892 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6893 if (flags & MDB_NODUPDATA) {
6894 /* mdb_cursor_del0() will subtract the final entry */
6895 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6896 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
6898 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6899 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6901 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6904 /* If sub-DB still has entries, we're done */
6905 if (mc->mc_xcursor->mx_db.md_entries) {
6906 if (leaf->mn_flags & F_SUBDATA) {
6907 /* update subDB info */
6908 void *db = NODEDATA(leaf);
6909 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6912 /* shrink fake page */
6913 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6914 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6915 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6916 /* fix other sub-DB cursors pointed at fake pages on this page */
6917 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6918 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6919 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6920 if (m2->mc_pg[mc->mc_top] == mp) {
6921 MDB_node *n2 = leaf;
6922 if (m2->mc_ki[mc->mc_top] != mc->mc_ki[mc->mc_top]) {
6923 n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6924 if (n2->mn_flags & F_SUBDATA) continue;
6926 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6930 mc->mc_db->md_entries--;
6933 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
6935 /* otherwise fall thru and delete the sub-DB */
6938 if (leaf->mn_flags & F_SUBDATA) {
6939 /* add all the child DB's pages to the free list */
6940 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6945 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6946 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6947 rc = MDB_INCOMPATIBLE;
6951 /* add overflow pages to free list */
6952 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6956 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6957 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
6958 (rc = mdb_ovpage_free(mc, omp)))
6963 return mdb_cursor_del0(mc);
6966 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6970 /** Allocate and initialize new pages for a database.
6971 * @param[in] mc a cursor on the database being added to.
6972 * @param[in] flags flags defining what type of page is being allocated.
6973 * @param[in] num the number of pages to allocate. This is usually 1,
6974 * unless allocating overflow pages for a large record.
6975 * @param[out] mp Address of a page, or NULL on failure.
6976 * @return 0 on success, non-zero on failure.
6979 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6984 if ((rc = mdb_page_alloc(mc, num, &np)))
6986 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6987 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6988 np->mp_flags = flags | P_DIRTY;
6989 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6990 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6993 mc->mc_db->md_branch_pages++;
6994 else if (IS_LEAF(np))
6995 mc->mc_db->md_leaf_pages++;
6996 else if (IS_OVERFLOW(np)) {
6997 mc->mc_db->md_overflow_pages += num;
7005 /** Calculate the size of a leaf node.
7006 * The size depends on the environment's page size; if a data item
7007 * is too large it will be put onto an overflow page and the node
7008 * size will only include the key and not the data. Sizes are always
7009 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7010 * of the #MDB_node headers.
7011 * @param[in] env The environment handle.
7012 * @param[in] key The key for the node.
7013 * @param[in] data The data for the node.
7014 * @return The number of bytes needed to store the node.
7017 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7021 sz = LEAFSIZE(key, data);
7022 if (sz > env->me_nodemax) {
7023 /* put on overflow page */
7024 sz -= data->mv_size - sizeof(pgno_t);
7027 return EVEN(sz + sizeof(indx_t));
7030 /** Calculate the size of a branch node.
7031 * The size should depend on the environment's page size but since
7032 * we currently don't support spilling large keys onto overflow
7033 * pages, it's simply the size of the #MDB_node header plus the
7034 * size of the key. Sizes are always rounded up to an even number
7035 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7036 * @param[in] env The environment handle.
7037 * @param[in] key The key for the node.
7038 * @return The number of bytes needed to store the node.
7041 mdb_branch_size(MDB_env *env, MDB_val *key)
7046 if (sz > env->me_nodemax) {
7047 /* put on overflow page */
7048 /* not implemented */
7049 /* sz -= key->size - sizeof(pgno_t); */
7052 return sz + sizeof(indx_t);
7055 /** Add a node to the page pointed to by the cursor.
7056 * @param[in] mc The cursor for this operation.
7057 * @param[in] indx The index on the page where the new node should be added.
7058 * @param[in] key The key for the new node.
7059 * @param[in] data The data for the new node, if any.
7060 * @param[in] pgno The page number, if adding a branch node.
7061 * @param[in] flags Flags for the node.
7062 * @return 0 on success, non-zero on failure. Possible errors are:
7064 * <li>ENOMEM - failed to allocate overflow pages for the node.
7065 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7066 * should never happen since all callers already calculate the
7067 * page's free space before calling this function.
7071 mdb_node_add(MDB_cursor *mc, indx_t indx,
7072 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7075 size_t node_size = NODESIZE;
7079 MDB_page *mp = mc->mc_pg[mc->mc_top];
7080 MDB_page *ofp = NULL; /* overflow page */
7084 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7086 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7087 IS_LEAF(mp) ? "leaf" : "branch",
7088 IS_SUBP(mp) ? "sub-" : "",
7089 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7090 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7093 /* Move higher keys up one slot. */
7094 int ksize = mc->mc_db->md_pad, dif;
7095 char *ptr = LEAF2KEY(mp, indx, ksize);
7096 dif = NUMKEYS(mp) - indx;
7098 memmove(ptr+ksize, ptr, dif*ksize);
7099 /* insert new key */
7100 memcpy(ptr, key->mv_data, ksize);
7102 /* Just using these for counting */
7103 mp->mp_lower += sizeof(indx_t);
7104 mp->mp_upper -= ksize - sizeof(indx_t);
7108 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7110 node_size += key->mv_size;
7112 mdb_cassert(mc, key && data);
7113 if (F_ISSET(flags, F_BIGDATA)) {
7114 /* Data already on overflow page. */
7115 node_size += sizeof(pgno_t);
7116 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7117 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7119 /* Put data on overflow page. */
7120 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7121 data->mv_size, node_size+data->mv_size));
7122 node_size = EVEN(node_size + sizeof(pgno_t));
7123 if ((ssize_t)node_size > room)
7125 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7127 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7131 node_size += data->mv_size;
7134 node_size = EVEN(node_size);
7135 if ((ssize_t)node_size > room)
7139 /* Move higher pointers up one slot. */
7140 for (i = NUMKEYS(mp); i > indx; i--)
7141 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7143 /* Adjust free space offsets. */
7144 ofs = mp->mp_upper - node_size;
7145 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7146 mp->mp_ptrs[indx] = ofs;
7148 mp->mp_lower += sizeof(indx_t);
7150 /* Write the node data. */
7151 node = NODEPTR(mp, indx);
7152 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7153 node->mn_flags = flags;
7155 SETDSZ(node,data->mv_size);
7160 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7163 ndata = NODEDATA(node);
7165 if (F_ISSET(flags, F_BIGDATA))
7166 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7167 else if (F_ISSET(flags, MDB_RESERVE))
7168 data->mv_data = ndata;
7170 memcpy(ndata, data->mv_data, data->mv_size);
7172 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7173 ndata = METADATA(ofp);
7174 if (F_ISSET(flags, MDB_RESERVE))
7175 data->mv_data = ndata;
7177 memcpy(ndata, data->mv_data, data->mv_size);
7184 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7185 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7186 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7187 DPRINTF(("node size = %"Z"u", node_size));
7188 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7189 return MDB_PAGE_FULL;
7192 /** Delete the specified node from a page.
7193 * @param[in] mc Cursor pointing to the node to delete.
7194 * @param[in] ksize The size of a node. Only used if the page is
7195 * part of a #MDB_DUPFIXED database.
7198 mdb_node_del(MDB_cursor *mc, int ksize)
7200 MDB_page *mp = mc->mc_pg[mc->mc_top];
7201 indx_t indx = mc->mc_ki[mc->mc_top];
7203 indx_t i, j, numkeys, ptr;
7207 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7208 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7209 numkeys = NUMKEYS(mp);
7210 mdb_cassert(mc, indx < numkeys);
7213 int x = numkeys - 1 - indx;
7214 base = LEAF2KEY(mp, indx, ksize);
7216 memmove(base, base + ksize, x * ksize);
7217 mp->mp_lower -= sizeof(indx_t);
7218 mp->mp_upper += ksize - sizeof(indx_t);
7222 node = NODEPTR(mp, indx);
7223 sz = NODESIZE + node->mn_ksize;
7225 if (F_ISSET(node->mn_flags, F_BIGDATA))
7226 sz += sizeof(pgno_t);
7228 sz += NODEDSZ(node);
7232 ptr = mp->mp_ptrs[indx];
7233 for (i = j = 0; i < numkeys; i++) {
7235 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7236 if (mp->mp_ptrs[i] < ptr)
7237 mp->mp_ptrs[j] += sz;
7242 base = (char *)mp + mp->mp_upper + PAGEBASE;
7243 memmove(base + sz, base, ptr - mp->mp_upper);
7245 mp->mp_lower -= sizeof(indx_t);
7249 /** Compact the main page after deleting a node on a subpage.
7250 * @param[in] mp The main page to operate on.
7251 * @param[in] indx The index of the subpage on the main page.
7254 mdb_node_shrink(MDB_page *mp, indx_t indx)
7259 indx_t delta, nsize, len, ptr;
7262 node = NODEPTR(mp, indx);
7263 sp = (MDB_page *)NODEDATA(node);
7264 delta = SIZELEFT(sp);
7265 nsize = NODEDSZ(node) - delta;
7267 /* Prepare to shift upward, set len = length(subpage part to shift) */
7271 return; /* do not make the node uneven-sized */
7273 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7274 for (i = NUMKEYS(sp); --i >= 0; )
7275 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7278 sp->mp_upper = sp->mp_lower;
7279 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7280 SETDSZ(node, nsize);
7282 /* Shift <lower nodes...initial part of subpage> upward */
7283 base = (char *)mp + mp->mp_upper + PAGEBASE;
7284 memmove(base + delta, base, (char *)sp + len - base);
7286 ptr = mp->mp_ptrs[indx];
7287 for (i = NUMKEYS(mp); --i >= 0; ) {
7288 if (mp->mp_ptrs[i] <= ptr)
7289 mp->mp_ptrs[i] += delta;
7291 mp->mp_upper += delta;
7294 /** Initial setup of a sorted-dups cursor.
7295 * Sorted duplicates are implemented as a sub-database for the given key.
7296 * The duplicate data items are actually keys of the sub-database.
7297 * Operations on the duplicate data items are performed using a sub-cursor
7298 * initialized when the sub-database is first accessed. This function does
7299 * the preliminary setup of the sub-cursor, filling in the fields that
7300 * depend only on the parent DB.
7301 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7304 mdb_xcursor_init0(MDB_cursor *mc)
7306 MDB_xcursor *mx = mc->mc_xcursor;
7308 mx->mx_cursor.mc_xcursor = NULL;
7309 mx->mx_cursor.mc_txn = mc->mc_txn;
7310 mx->mx_cursor.mc_db = &mx->mx_db;
7311 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7312 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7313 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7314 mx->mx_cursor.mc_snum = 0;
7315 mx->mx_cursor.mc_top = 0;
7316 mx->mx_cursor.mc_flags = C_SUB;
7317 mx->mx_dbx.md_name.mv_size = 0;
7318 mx->mx_dbx.md_name.mv_data = NULL;
7319 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7320 mx->mx_dbx.md_dcmp = NULL;
7321 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7324 /** Final setup of a sorted-dups cursor.
7325 * Sets up the fields that depend on the data from the main cursor.
7326 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7327 * @param[in] node The data containing the #MDB_db record for the
7328 * sorted-dup database.
7331 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7333 MDB_xcursor *mx = mc->mc_xcursor;
7335 if (node->mn_flags & F_SUBDATA) {
7336 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7337 mx->mx_cursor.mc_pg[0] = 0;
7338 mx->mx_cursor.mc_snum = 0;
7339 mx->mx_cursor.mc_top = 0;
7340 mx->mx_cursor.mc_flags = C_SUB;
7342 MDB_page *fp = NODEDATA(node);
7343 mx->mx_db.md_pad = 0;
7344 mx->mx_db.md_flags = 0;
7345 mx->mx_db.md_depth = 1;
7346 mx->mx_db.md_branch_pages = 0;
7347 mx->mx_db.md_leaf_pages = 1;
7348 mx->mx_db.md_overflow_pages = 0;
7349 mx->mx_db.md_entries = NUMKEYS(fp);
7350 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7351 mx->mx_cursor.mc_snum = 1;
7352 mx->mx_cursor.mc_top = 0;
7353 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7354 mx->mx_cursor.mc_pg[0] = fp;
7355 mx->mx_cursor.mc_ki[0] = 0;
7356 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7357 mx->mx_db.md_flags = MDB_DUPFIXED;
7358 mx->mx_db.md_pad = fp->mp_pad;
7359 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7360 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7363 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7364 mx->mx_db.md_root));
7365 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7366 #if UINT_MAX < SIZE_MAX
7367 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7368 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7373 /** Fixup a sorted-dups cursor due to underlying update.
7374 * Sets up some fields that depend on the data from the main cursor.
7375 * Almost the same as init1, but skips initialization steps if the
7376 * xcursor had already been used.
7377 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7378 * @param[in] src_mx The xcursor of an up-to-date cursor.
7379 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7382 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7384 MDB_xcursor *mx = mc->mc_xcursor;
7387 mx->mx_cursor.mc_snum = 1;
7388 mx->mx_cursor.mc_top = 0;
7389 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7390 mx->mx_cursor.mc_ki[0] = 0;
7391 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7392 #if UINT_MAX < SIZE_MAX
7393 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7395 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7398 mx->mx_db = src_mx->mx_db;
7399 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7400 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7401 mx->mx_db.md_root));
7404 /** Initialize a cursor for a given transaction and database. */
7406 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7409 mc->mc_backup = NULL;
7412 mc->mc_db = &txn->mt_dbs[dbi];
7413 mc->mc_dbx = &txn->mt_dbxs[dbi];
7414 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7420 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7421 mdb_tassert(txn, mx != NULL);
7422 mc->mc_xcursor = mx;
7423 mdb_xcursor_init0(mc);
7425 mc->mc_xcursor = NULL;
7427 if (*mc->mc_dbflag & DB_STALE) {
7428 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7433 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7436 size_t size = sizeof(MDB_cursor);
7438 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7441 if (txn->mt_flags & MDB_TXN_BLOCKED)
7444 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7447 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7448 size += sizeof(MDB_xcursor);
7450 if ((mc = malloc(size)) != NULL) {
7451 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7452 if (txn->mt_cursors) {
7453 mc->mc_next = txn->mt_cursors[dbi];
7454 txn->mt_cursors[dbi] = mc;
7455 mc->mc_flags |= C_UNTRACK;
7467 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7469 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7472 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7475 if (txn->mt_flags & MDB_TXN_BLOCKED)
7478 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7482 /* Return the count of duplicate data items for the current key */
7484 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7488 if (mc == NULL || countp == NULL)
7491 if (mc->mc_xcursor == NULL)
7492 return MDB_INCOMPATIBLE;
7494 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7497 if (!(mc->mc_flags & C_INITIALIZED))
7500 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7501 return MDB_NOTFOUND;
7503 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7504 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7507 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7510 *countp = mc->mc_xcursor->mx_db.md_entries;
7516 mdb_cursor_close(MDB_cursor *mc)
7518 if (mc && !mc->mc_backup) {
7519 /* remove from txn, if tracked */
7520 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7521 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7522 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7524 *prev = mc->mc_next;
7531 mdb_cursor_txn(MDB_cursor *mc)
7533 if (!mc) return NULL;
7538 mdb_cursor_dbi(MDB_cursor *mc)
7543 /** Replace the key for a branch node with a new key.
7544 * @param[in] mc Cursor pointing to the node to operate on.
7545 * @param[in] key The new key to use.
7546 * @return 0 on success, non-zero on failure.
7549 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7555 int delta, ksize, oksize;
7556 indx_t ptr, i, numkeys, indx;
7559 indx = mc->mc_ki[mc->mc_top];
7560 mp = mc->mc_pg[mc->mc_top];
7561 node = NODEPTR(mp, indx);
7562 ptr = mp->mp_ptrs[indx];
7566 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7567 k2.mv_data = NODEKEY(node);
7568 k2.mv_size = node->mn_ksize;
7569 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7571 mdb_dkey(&k2, kbuf2),
7577 /* Sizes must be 2-byte aligned. */
7578 ksize = EVEN(key->mv_size);
7579 oksize = EVEN(node->mn_ksize);
7580 delta = ksize - oksize;
7582 /* Shift node contents if EVEN(key length) changed. */
7584 if (delta > 0 && SIZELEFT(mp) < delta) {
7586 /* not enough space left, do a delete and split */
7587 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7588 pgno = NODEPGNO(node);
7589 mdb_node_del(mc, 0);
7590 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7593 numkeys = NUMKEYS(mp);
7594 for (i = 0; i < numkeys; i++) {
7595 if (mp->mp_ptrs[i] <= ptr)
7596 mp->mp_ptrs[i] -= delta;
7599 base = (char *)mp + mp->mp_upper + PAGEBASE;
7600 len = ptr - mp->mp_upper + NODESIZE;
7601 memmove(base - delta, base, len);
7602 mp->mp_upper -= delta;
7604 node = NODEPTR(mp, indx);
7607 /* But even if no shift was needed, update ksize */
7608 if (node->mn_ksize != key->mv_size)
7609 node->mn_ksize = key->mv_size;
7612 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7618 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7620 /** Perform \b act while tracking temporary cursor \b mn */
7621 #define WITH_CURSOR_TRACKING(mn, act) do { \
7622 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
7623 if ((mn).mc_flags & C_SUB) { \
7624 dummy.mc_flags = C_INITIALIZED; \
7625 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
7630 tracked->mc_next = *tp; \
7633 *tp = tracked->mc_next; \
7636 /** Move a node from csrc to cdst.
7639 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
7646 unsigned short flags;
7650 /* Mark src and dst as dirty. */
7651 if ((rc = mdb_page_touch(csrc)) ||
7652 (rc = mdb_page_touch(cdst)))
7655 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7656 key.mv_size = csrc->mc_db->md_pad;
7657 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7659 data.mv_data = NULL;
7663 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7664 mdb_cassert(csrc, !((size_t)srcnode & 1));
7665 srcpg = NODEPGNO(srcnode);
7666 flags = srcnode->mn_flags;
7667 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7668 unsigned int snum = csrc->mc_snum;
7670 /* must find the lowest key below src */
7671 rc = mdb_page_search_lowest(csrc);
7674 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7675 key.mv_size = csrc->mc_db->md_pad;
7676 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7678 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7679 key.mv_size = NODEKSZ(s2);
7680 key.mv_data = NODEKEY(s2);
7682 csrc->mc_snum = snum--;
7683 csrc->mc_top = snum;
7685 key.mv_size = NODEKSZ(srcnode);
7686 key.mv_data = NODEKEY(srcnode);
7688 data.mv_size = NODEDSZ(srcnode);
7689 data.mv_data = NODEDATA(srcnode);
7691 mn.mc_xcursor = NULL;
7692 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7693 unsigned int snum = cdst->mc_snum;
7696 /* must find the lowest key below dst */
7697 mdb_cursor_copy(cdst, &mn);
7698 rc = mdb_page_search_lowest(&mn);
7701 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7702 bkey.mv_size = mn.mc_db->md_pad;
7703 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7705 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7706 bkey.mv_size = NODEKSZ(s2);
7707 bkey.mv_data = NODEKEY(s2);
7709 mn.mc_snum = snum--;
7712 rc = mdb_update_key(&mn, &bkey);
7717 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7718 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7719 csrc->mc_ki[csrc->mc_top],
7721 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7722 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7724 /* Add the node to the destination page.
7726 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7727 if (rc != MDB_SUCCESS)
7730 /* Delete the node from the source page.
7732 mdb_node_del(csrc, key.mv_size);
7735 /* Adjust other cursors pointing to mp */
7736 MDB_cursor *m2, *m3;
7737 MDB_dbi dbi = csrc->mc_dbi;
7738 MDB_page *mpd, *mps;
7740 mps = csrc->mc_pg[csrc->mc_top];
7741 /* If we're adding on the left, bump others up */
7743 mpd = cdst->mc_pg[csrc->mc_top];
7744 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7745 if (csrc->mc_flags & C_SUB)
7746 m3 = &m2->mc_xcursor->mx_cursor;
7749 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7752 m3->mc_pg[csrc->mc_top] == mpd &&
7753 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7754 m3->mc_ki[csrc->mc_top]++;
7757 m3->mc_pg[csrc->mc_top] == mps &&
7758 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7759 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7760 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7761 m3->mc_ki[csrc->mc_top-1]++;
7763 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7764 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7767 /* Adding on the right, bump others down */
7769 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7770 if (csrc->mc_flags & C_SUB)
7771 m3 = &m2->mc_xcursor->mx_cursor;
7774 if (m3 == csrc) continue;
7775 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7777 if (m3->mc_pg[csrc->mc_top] == mps) {
7778 if (!m3->mc_ki[csrc->mc_top]) {
7779 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7780 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7781 m3->mc_ki[csrc->mc_top-1]--;
7783 m3->mc_ki[csrc->mc_top]--;
7785 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7786 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7792 /* Update the parent separators.
7794 if (csrc->mc_ki[csrc->mc_top] == 0) {
7795 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7796 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7797 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7799 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7800 key.mv_size = NODEKSZ(srcnode);
7801 key.mv_data = NODEKEY(srcnode);
7803 DPRINTF(("update separator for source page %"Z"u to [%s]",
7804 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7805 mdb_cursor_copy(csrc, &mn);
7808 /* We want mdb_rebalance to find mn when doing fixups */
7809 WITH_CURSOR_TRACKING(mn,
7810 rc = mdb_update_key(&mn, &key));
7814 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7816 indx_t ix = csrc->mc_ki[csrc->mc_top];
7817 nullkey.mv_size = 0;
7818 csrc->mc_ki[csrc->mc_top] = 0;
7819 rc = mdb_update_key(csrc, &nullkey);
7820 csrc->mc_ki[csrc->mc_top] = ix;
7821 mdb_cassert(csrc, rc == MDB_SUCCESS);
7825 if (cdst->mc_ki[cdst->mc_top] == 0) {
7826 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7827 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7828 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7830 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7831 key.mv_size = NODEKSZ(srcnode);
7832 key.mv_data = NODEKEY(srcnode);
7834 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7835 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7836 mdb_cursor_copy(cdst, &mn);
7839 /* We want mdb_rebalance to find mn when doing fixups */
7840 WITH_CURSOR_TRACKING(mn,
7841 rc = mdb_update_key(&mn, &key));
7845 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7847 indx_t ix = cdst->mc_ki[cdst->mc_top];
7848 nullkey.mv_size = 0;
7849 cdst->mc_ki[cdst->mc_top] = 0;
7850 rc = mdb_update_key(cdst, &nullkey);
7851 cdst->mc_ki[cdst->mc_top] = ix;
7852 mdb_cassert(cdst, rc == MDB_SUCCESS);
7859 /** Merge one page into another.
7860 * The nodes from the page pointed to by \b csrc will
7861 * be copied to the page pointed to by \b cdst and then
7862 * the \b csrc page will be freed.
7863 * @param[in] csrc Cursor pointing to the source page.
7864 * @param[in] cdst Cursor pointing to the destination page.
7865 * @return 0 on success, non-zero on failure.
7868 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7870 MDB_page *psrc, *pdst;
7877 psrc = csrc->mc_pg[csrc->mc_top];
7878 pdst = cdst->mc_pg[cdst->mc_top];
7880 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7882 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7883 mdb_cassert(csrc, cdst->mc_snum > 1);
7885 /* Mark dst as dirty. */
7886 if ((rc = mdb_page_touch(cdst)))
7889 /* get dst page again now that we've touched it. */
7890 pdst = cdst->mc_pg[cdst->mc_top];
7892 /* Move all nodes from src to dst.
7894 j = nkeys = NUMKEYS(pdst);
7895 if (IS_LEAF2(psrc)) {
7896 key.mv_size = csrc->mc_db->md_pad;
7897 key.mv_data = METADATA(psrc);
7898 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7899 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7900 if (rc != MDB_SUCCESS)
7902 key.mv_data = (char *)key.mv_data + key.mv_size;
7905 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7906 srcnode = NODEPTR(psrc, i);
7907 if (i == 0 && IS_BRANCH(psrc)) {
7910 mdb_cursor_copy(csrc, &mn);
7911 mn.mc_xcursor = NULL;
7912 /* must find the lowest key below src */
7913 rc = mdb_page_search_lowest(&mn);
7916 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7917 key.mv_size = mn.mc_db->md_pad;
7918 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7920 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7921 key.mv_size = NODEKSZ(s2);
7922 key.mv_data = NODEKEY(s2);
7925 key.mv_size = srcnode->mn_ksize;
7926 key.mv_data = NODEKEY(srcnode);
7929 data.mv_size = NODEDSZ(srcnode);
7930 data.mv_data = NODEDATA(srcnode);
7931 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7932 if (rc != MDB_SUCCESS)
7937 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7938 pdst->mp_pgno, NUMKEYS(pdst),
7939 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7941 /* Unlink the src page from parent and add to free list.
7944 mdb_node_del(csrc, 0);
7945 if (csrc->mc_ki[csrc->mc_top] == 0) {
7947 rc = mdb_update_key(csrc, &key);
7955 psrc = csrc->mc_pg[csrc->mc_top];
7956 /* If not operating on FreeDB, allow this page to be reused
7957 * in this txn. Otherwise just add to free list.
7959 rc = mdb_page_loose(csrc, psrc);
7963 csrc->mc_db->md_leaf_pages--;
7965 csrc->mc_db->md_branch_pages--;
7967 /* Adjust other cursors pointing to mp */
7968 MDB_cursor *m2, *m3;
7969 MDB_dbi dbi = csrc->mc_dbi;
7970 unsigned int top = csrc->mc_top;
7972 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7973 if (csrc->mc_flags & C_SUB)
7974 m3 = &m2->mc_xcursor->mx_cursor;
7977 if (m3 == csrc) continue;
7978 if (m3->mc_snum < csrc->mc_snum) continue;
7979 if (m3->mc_pg[top] == psrc) {
7980 m3->mc_pg[top] = pdst;
7981 m3->mc_ki[top] += nkeys;
7982 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
7983 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
7984 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
7987 if (XCURSOR_INITED(m3) && IS_LEAF(psrc))
7988 XCURSOR_REFRESH(m3, m3->mc_pg[top], m3->mc_ki[top]);
7992 unsigned int snum = cdst->mc_snum;
7993 uint16_t depth = cdst->mc_db->md_depth;
7994 mdb_cursor_pop(cdst);
7995 rc = mdb_rebalance(cdst);
7996 /* Did the tree height change? */
7997 if (depth != cdst->mc_db->md_depth)
7998 snum += cdst->mc_db->md_depth - depth;
7999 cdst->mc_snum = snum;
8000 cdst->mc_top = snum-1;
8005 /** Copy the contents of a cursor.
8006 * @param[in] csrc The cursor to copy from.
8007 * @param[out] cdst The cursor to copy to.
8010 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8014 cdst->mc_txn = csrc->mc_txn;
8015 cdst->mc_dbi = csrc->mc_dbi;
8016 cdst->mc_db = csrc->mc_db;
8017 cdst->mc_dbx = csrc->mc_dbx;
8018 cdst->mc_snum = csrc->mc_snum;
8019 cdst->mc_top = csrc->mc_top;
8020 cdst->mc_flags = csrc->mc_flags;
8022 for (i=0; i<csrc->mc_snum; i++) {
8023 cdst->mc_pg[i] = csrc->mc_pg[i];
8024 cdst->mc_ki[i] = csrc->mc_ki[i];
8028 /** Rebalance the tree after a delete operation.
8029 * @param[in] mc Cursor pointing to the page where rebalancing
8031 * @return 0 on success, non-zero on failure.
8034 mdb_rebalance(MDB_cursor *mc)
8038 unsigned int ptop, minkeys, thresh;
8042 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8047 thresh = FILL_THRESHOLD;
8049 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
8050 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8051 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8052 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8054 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8055 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8056 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
8057 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8061 if (mc->mc_snum < 2) {
8062 MDB_page *mp = mc->mc_pg[0];
8064 DPUTS("Can't rebalance a subpage, ignoring");
8067 if (NUMKEYS(mp) == 0) {
8068 DPUTS("tree is completely empty");
8069 mc->mc_db->md_root = P_INVALID;
8070 mc->mc_db->md_depth = 0;
8071 mc->mc_db->md_leaf_pages = 0;
8072 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8075 /* Adjust cursors pointing to mp */
8078 mc->mc_flags &= ~C_INITIALIZED;
8080 MDB_cursor *m2, *m3;
8081 MDB_dbi dbi = mc->mc_dbi;
8083 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8084 if (mc->mc_flags & C_SUB)
8085 m3 = &m2->mc_xcursor->mx_cursor;
8088 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8090 if (m3->mc_pg[0] == mp) {
8093 m3->mc_flags &= ~C_INITIALIZED;
8097 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8099 DPUTS("collapsing root page!");
8100 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8103 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8104 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
8107 mc->mc_db->md_depth--;
8108 mc->mc_db->md_branch_pages--;
8109 mc->mc_ki[0] = mc->mc_ki[1];
8110 for (i = 1; i<mc->mc_db->md_depth; i++) {
8111 mc->mc_pg[i] = mc->mc_pg[i+1];
8112 mc->mc_ki[i] = mc->mc_ki[i+1];
8115 /* Adjust other cursors pointing to mp */
8116 MDB_cursor *m2, *m3;
8117 MDB_dbi dbi = mc->mc_dbi;
8119 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8120 if (mc->mc_flags & C_SUB)
8121 m3 = &m2->mc_xcursor->mx_cursor;
8124 if (m3 == mc) continue;
8125 if (!(m3->mc_flags & C_INITIALIZED))
8127 if (m3->mc_pg[0] == mp) {
8128 for (i=0; i<mc->mc_db->md_depth; i++) {
8129 m3->mc_pg[i] = m3->mc_pg[i+1];
8130 m3->mc_ki[i] = m3->mc_ki[i+1];
8138 DPUTS("root page doesn't need rebalancing");
8142 /* The parent (branch page) must have at least 2 pointers,
8143 * otherwise the tree is invalid.
8145 ptop = mc->mc_top-1;
8146 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8148 /* Leaf page fill factor is below the threshold.
8149 * Try to move keys from left or right neighbor, or
8150 * merge with a neighbor page.
8155 mdb_cursor_copy(mc, &mn);
8156 mn.mc_xcursor = NULL;
8158 oldki = mc->mc_ki[mc->mc_top];
8159 if (mc->mc_ki[ptop] == 0) {
8160 /* We're the leftmost leaf in our parent.
8162 DPUTS("reading right neighbor");
8164 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8165 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8168 mn.mc_ki[mn.mc_top] = 0;
8169 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8172 /* There is at least one neighbor to the left.
8174 DPUTS("reading left neighbor");
8176 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8177 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8180 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8181 mc->mc_ki[mc->mc_top] = 0;
8185 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8186 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8187 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8189 /* If the neighbor page is above threshold and has enough keys,
8190 * move one key from it. Otherwise we should try to merge them.
8191 * (A branch page must never have less than 2 keys.)
8193 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8194 rc = mdb_node_move(&mn, mc, fromleft);
8196 /* if we inserted on left, bump position up */
8201 rc = mdb_page_merge(&mn, mc);
8203 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8204 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8205 /* We want mdb_rebalance to find mn when doing fixups */
8206 WITH_CURSOR_TRACKING(mn,
8207 rc = mdb_page_merge(mc, &mn));
8208 mdb_cursor_copy(&mn, mc);
8210 mc->mc_flags &= ~C_EOF;
8212 mc->mc_ki[mc->mc_top] = oldki;
8216 /** Complete a delete operation started by #mdb_cursor_del(). */
8218 mdb_cursor_del0(MDB_cursor *mc)
8224 MDB_cursor *m2, *m3;
8225 MDB_dbi dbi = mc->mc_dbi;
8227 ki = mc->mc_ki[mc->mc_top];
8228 mp = mc->mc_pg[mc->mc_top];
8229 mdb_node_del(mc, mc->mc_db->md_pad);
8230 mc->mc_db->md_entries--;
8232 /* Adjust other cursors pointing to mp */
8233 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8234 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8235 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8237 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8239 if (m3->mc_pg[mc->mc_top] == mp) {
8240 if (m3->mc_ki[mc->mc_top] == ki) {
8241 m3->mc_flags |= C_DEL;
8242 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8243 /* Sub-cursor referred into dataset which is gone */
8244 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
8247 } else if (m3->mc_ki[mc->mc_top] > ki) {
8248 m3->mc_ki[mc->mc_top]--;
8250 if (XCURSOR_INITED(m3))
8251 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8255 rc = mdb_rebalance(mc);
8257 if (rc == MDB_SUCCESS) {
8258 /* DB is totally empty now, just bail out.
8259 * Other cursors adjustments were already done
8260 * by mdb_rebalance and aren't needed here.
8265 mp = mc->mc_pg[mc->mc_top];
8266 nkeys = NUMKEYS(mp);
8268 /* Adjust other cursors pointing to mp */
8269 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8270 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8271 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8273 if (m3->mc_snum < mc->mc_snum)
8275 if (m3->mc_pg[mc->mc_top] == mp) {
8276 /* if m3 points past last node in page, find next sibling */
8277 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
8278 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8279 rc = mdb_cursor_sibling(m3, 1);
8280 if (rc == MDB_NOTFOUND) {
8281 m3->mc_flags |= C_EOF;
8286 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8287 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
8288 /* If this node is a fake page, it needs to be reinited
8289 * because its data has moved. But just reset mc_pg[0]
8290 * if the xcursor is already live.
8292 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) {
8293 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)
8294 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8296 mdb_xcursor_init1(m3, node);
8302 mc->mc_flags |= C_DEL;
8306 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8311 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8312 MDB_val *key, MDB_val *data)
8314 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8317 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8318 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8320 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8321 /* must ignore any data */
8325 return mdb_del0(txn, dbi, key, data, 0);
8329 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8330 MDB_val *key, MDB_val *data, unsigned flags)
8335 MDB_val rdata, *xdata;
8339 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8341 mdb_cursor_init(&mc, txn, dbi, &mx);
8350 flags |= MDB_NODUPDATA;
8352 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8354 /* let mdb_page_split know about this cursor if needed:
8355 * delete will trigger a rebalance; if it needs to move
8356 * a node from one page to another, it will have to
8357 * update the parent's separator key(s). If the new sepkey
8358 * is larger than the current one, the parent page may
8359 * run out of space, triggering a split. We need this
8360 * cursor to be consistent until the end of the rebalance.
8362 mc.mc_flags |= C_UNTRACK;
8363 mc.mc_next = txn->mt_cursors[dbi];
8364 txn->mt_cursors[dbi] = &mc;
8365 rc = mdb_cursor_del(&mc, flags);
8366 txn->mt_cursors[dbi] = mc.mc_next;
8371 /** Split a page and insert a new node.
8372 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8373 * The cursor will be updated to point to the actual page and index where
8374 * the node got inserted after the split.
8375 * @param[in] newkey The key for the newly inserted node.
8376 * @param[in] newdata The data for the newly inserted node.
8377 * @param[in] newpgno The page number, if the new node is a branch node.
8378 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8379 * @return 0 on success, non-zero on failure.
8382 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8383 unsigned int nflags)
8386 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8389 int i, j, split_indx, nkeys, pmax;
8390 MDB_env *env = mc->mc_txn->mt_env;
8392 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8393 MDB_page *copy = NULL;
8394 MDB_page *mp, *rp, *pp;
8399 mp = mc->mc_pg[mc->mc_top];
8400 newindx = mc->mc_ki[mc->mc_top];
8401 nkeys = NUMKEYS(mp);
8403 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8404 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8405 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8407 /* Create a right sibling. */
8408 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8410 rp->mp_pad = mp->mp_pad;
8411 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8413 /* Usually when splitting the root page, the cursor
8414 * height is 1. But when called from mdb_update_key,
8415 * the cursor height may be greater because it walks
8416 * up the stack while finding the branch slot to update.
8418 if (mc->mc_top < 1) {
8419 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8421 /* shift current top to make room for new parent */
8422 for (i=mc->mc_snum; i>0; i--) {
8423 mc->mc_pg[i] = mc->mc_pg[i-1];
8424 mc->mc_ki[i] = mc->mc_ki[i-1];
8428 mc->mc_db->md_root = pp->mp_pgno;
8429 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8430 new_root = mc->mc_db->md_depth++;
8432 /* Add left (implicit) pointer. */
8433 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8434 /* undo the pre-push */
8435 mc->mc_pg[0] = mc->mc_pg[1];
8436 mc->mc_ki[0] = mc->mc_ki[1];
8437 mc->mc_db->md_root = mp->mp_pgno;
8438 mc->mc_db->md_depth--;
8445 ptop = mc->mc_top-1;
8446 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8449 mdb_cursor_copy(mc, &mn);
8450 mn.mc_xcursor = NULL;
8451 mn.mc_pg[mn.mc_top] = rp;
8452 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8454 if (nflags & MDB_APPEND) {
8455 mn.mc_ki[mn.mc_top] = 0;
8457 split_indx = newindx;
8461 split_indx = (nkeys+1) / 2;
8466 unsigned int lsize, rsize, ksize;
8467 /* Move half of the keys to the right sibling */
8468 x = mc->mc_ki[mc->mc_top] - split_indx;
8469 ksize = mc->mc_db->md_pad;
8470 split = LEAF2KEY(mp, split_indx, ksize);
8471 rsize = (nkeys - split_indx) * ksize;
8472 lsize = (nkeys - split_indx) * sizeof(indx_t);
8473 mp->mp_lower -= lsize;
8474 rp->mp_lower += lsize;
8475 mp->mp_upper += rsize - lsize;
8476 rp->mp_upper -= rsize - lsize;
8477 sepkey.mv_size = ksize;
8478 if (newindx == split_indx) {
8479 sepkey.mv_data = newkey->mv_data;
8481 sepkey.mv_data = split;
8484 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8485 memcpy(rp->mp_ptrs, split, rsize);
8486 sepkey.mv_data = rp->mp_ptrs;
8487 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8488 memcpy(ins, newkey->mv_data, ksize);
8489 mp->mp_lower += sizeof(indx_t);
8490 mp->mp_upper -= ksize - sizeof(indx_t);
8493 memcpy(rp->mp_ptrs, split, x * ksize);
8494 ins = LEAF2KEY(rp, x, ksize);
8495 memcpy(ins, newkey->mv_data, ksize);
8496 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8497 rp->mp_lower += sizeof(indx_t);
8498 rp->mp_upper -= ksize - sizeof(indx_t);
8499 mc->mc_ki[mc->mc_top] = x;
8502 int psize, nsize, k;
8503 /* Maximum free space in an empty page */
8504 pmax = env->me_psize - PAGEHDRSZ;
8506 nsize = mdb_leaf_size(env, newkey, newdata);
8508 nsize = mdb_branch_size(env, newkey);
8509 nsize = EVEN(nsize);
8511 /* grab a page to hold a temporary copy */
8512 copy = mdb_page_malloc(mc->mc_txn, 1);
8517 copy->mp_pgno = mp->mp_pgno;
8518 copy->mp_flags = mp->mp_flags;
8519 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8520 copy->mp_upper = env->me_psize - PAGEBASE;
8522 /* prepare to insert */
8523 for (i=0, j=0; i<nkeys; i++) {
8525 copy->mp_ptrs[j++] = 0;
8527 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8530 /* When items are relatively large the split point needs
8531 * to be checked, because being off-by-one will make the
8532 * difference between success or failure in mdb_node_add.
8534 * It's also relevant if a page happens to be laid out
8535 * such that one half of its nodes are all "small" and
8536 * the other half of its nodes are "large." If the new
8537 * item is also "large" and falls on the half with
8538 * "large" nodes, it also may not fit.
8540 * As a final tweak, if the new item goes on the last
8541 * spot on the page (and thus, onto the new page), bias
8542 * the split so the new page is emptier than the old page.
8543 * This yields better packing during sequential inserts.
8545 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8546 /* Find split point */
8548 if (newindx <= split_indx || newindx >= nkeys) {
8550 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8555 for (; i!=k; i+=j) {
8560 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8561 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8563 if (F_ISSET(node->mn_flags, F_BIGDATA))
8564 psize += sizeof(pgno_t);
8566 psize += NODEDSZ(node);
8568 psize = EVEN(psize);
8570 if (psize > pmax || i == k-j) {
8571 split_indx = i + (j<0);
8576 if (split_indx == newindx) {
8577 sepkey.mv_size = newkey->mv_size;
8578 sepkey.mv_data = newkey->mv_data;
8580 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8581 sepkey.mv_size = node->mn_ksize;
8582 sepkey.mv_data = NODEKEY(node);
8587 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8589 /* Copy separator key to the parent.
8591 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8592 int snum = mc->mc_snum;
8596 /* We want other splits to find mn when doing fixups */
8597 WITH_CURSOR_TRACKING(mn,
8598 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
8603 if (mc->mc_snum > snum) {
8606 /* Right page might now have changed parent.
8607 * Check if left page also changed parent.
8609 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8610 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8611 for (i=0; i<ptop; i++) {
8612 mc->mc_pg[i] = mn.mc_pg[i];
8613 mc->mc_ki[i] = mn.mc_ki[i];
8615 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8616 if (mn.mc_ki[ptop]) {
8617 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8619 /* find right page's left sibling */
8620 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8621 mdb_cursor_sibling(mc, 0);
8626 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8629 if (rc != MDB_SUCCESS) {
8632 if (nflags & MDB_APPEND) {
8633 mc->mc_pg[mc->mc_top] = rp;
8634 mc->mc_ki[mc->mc_top] = 0;
8635 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8638 for (i=0; i<mc->mc_top; i++)
8639 mc->mc_ki[i] = mn.mc_ki[i];
8640 } else if (!IS_LEAF2(mp)) {
8642 mc->mc_pg[mc->mc_top] = rp;
8647 rkey.mv_data = newkey->mv_data;
8648 rkey.mv_size = newkey->mv_size;
8654 /* Update index for the new key. */
8655 mc->mc_ki[mc->mc_top] = j;
8657 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8658 rkey.mv_data = NODEKEY(node);
8659 rkey.mv_size = node->mn_ksize;
8661 xdata.mv_data = NODEDATA(node);
8662 xdata.mv_size = NODEDSZ(node);
8665 pgno = NODEPGNO(node);
8666 flags = node->mn_flags;
8669 if (!IS_LEAF(mp) && j == 0) {
8670 /* First branch index doesn't need key data. */
8674 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8680 mc->mc_pg[mc->mc_top] = copy;
8685 } while (i != split_indx);
8687 nkeys = NUMKEYS(copy);
8688 for (i=0; i<nkeys; i++)
8689 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8690 mp->mp_lower = copy->mp_lower;
8691 mp->mp_upper = copy->mp_upper;
8692 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8693 env->me_psize - copy->mp_upper - PAGEBASE);
8695 /* reset back to original page */
8696 if (newindx < split_indx) {
8697 mc->mc_pg[mc->mc_top] = mp;
8699 mc->mc_pg[mc->mc_top] = rp;
8701 /* Make sure mc_ki is still valid.
8703 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8704 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8705 for (i=0; i<=ptop; i++) {
8706 mc->mc_pg[i] = mn.mc_pg[i];
8707 mc->mc_ki[i] = mn.mc_ki[i];
8711 if (nflags & MDB_RESERVE) {
8712 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8713 if (!(node->mn_flags & F_BIGDATA))
8714 newdata->mv_data = NODEDATA(node);
8717 if (newindx >= split_indx) {
8718 mc->mc_pg[mc->mc_top] = rp;
8720 /* Make sure mc_ki is still valid.
8722 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8723 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8724 for (i=0; i<=ptop; i++) {
8725 mc->mc_pg[i] = mn.mc_pg[i];
8726 mc->mc_ki[i] = mn.mc_ki[i];
8733 /* Adjust other cursors pointing to mp */
8734 MDB_cursor *m2, *m3;
8735 MDB_dbi dbi = mc->mc_dbi;
8736 nkeys = NUMKEYS(mp);
8738 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8739 if (mc->mc_flags & C_SUB)
8740 m3 = &m2->mc_xcursor->mx_cursor;
8745 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8749 /* sub cursors may be on different DB */
8750 if (m3->mc_pg[0] != mp)
8753 for (k=new_root; k>=0; k--) {
8754 m3->mc_ki[k+1] = m3->mc_ki[k];
8755 m3->mc_pg[k+1] = m3->mc_pg[k];
8757 if (m3->mc_ki[0] >= nkeys) {
8762 m3->mc_pg[0] = mc->mc_pg[0];
8766 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8767 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8768 m3->mc_ki[mc->mc_top]++;
8769 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8770 m3->mc_pg[mc->mc_top] = rp;
8771 m3->mc_ki[mc->mc_top] -= nkeys;
8772 for (i=0; i<mc->mc_top; i++) {
8773 m3->mc_ki[i] = mn.mc_ki[i];
8774 m3->mc_pg[i] = mn.mc_pg[i];
8777 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8778 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8781 if (XCURSOR_INITED(m3) && IS_LEAF(mp))
8782 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8785 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8788 if (copy) /* tmp page */
8789 mdb_page_free(env, copy);
8791 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8796 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8797 MDB_val *key, MDB_val *data, unsigned int flags)
8803 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8806 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8809 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8810 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8812 mdb_cursor_init(&mc, txn, dbi, &mx);
8813 mc.mc_next = txn->mt_cursors[dbi];
8814 txn->mt_cursors[dbi] = &mc;
8815 rc = mdb_cursor_put(&mc, key, data, flags);
8816 txn->mt_cursors[dbi] = mc.mc_next;
8821 #define MDB_WBUF (1024*1024)
8823 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
8825 /** State needed for a double-buffering compacting copy. */
8826 typedef struct mdb_copy {
8829 pthread_mutex_t mc_mutex;
8830 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
8835 pgno_t mc_next_pgno;
8837 int mc_toggle; /**< Buffer number in provider */
8838 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
8839 /** Error code. Never cleared if set. Both threads can set nonzero
8840 * to fail the copy. Not mutex-protected, LMDB expects atomic int.
8842 volatile int mc_error;
8845 /** Dedicated writer thread for compacting copy. */
8846 static THREAD_RET ESECT CALL_CONV
8847 mdb_env_copythr(void *arg)
8851 int toggle = 0, wsize, rc;
8854 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8857 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8860 pthread_mutex_lock(&my->mc_mutex);
8863 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8864 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
8866 wsize = my->mc_wlen[toggle];
8867 ptr = my->mc_wbuf[toggle];
8870 while (wsize > 0 && !my->mc_error) {
8871 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8875 } else if (len > 0) {
8888 /* If there's an overflow page tail, write it too */
8889 if (my->mc_olen[toggle]) {
8890 wsize = my->mc_olen[toggle];
8891 ptr = my->mc_over[toggle];
8892 my->mc_olen[toggle] = 0;
8895 my->mc_wlen[toggle] = 0;
8897 /* Return the empty buffer to provider */
8899 pthread_cond_signal(&my->mc_cond);
8901 pthread_mutex_unlock(&my->mc_mutex);
8902 return (THREAD_RET)0;
8906 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
8908 * @param[in] my control structure.
8909 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
8912 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
8914 pthread_mutex_lock(&my->mc_mutex);
8915 my->mc_new += adjust;
8916 pthread_cond_signal(&my->mc_cond);
8917 while (my->mc_new & 2) /* both buffers in use */
8918 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8919 pthread_mutex_unlock(&my->mc_mutex);
8921 my->mc_toggle ^= (adjust & 1);
8922 /* Both threads reset mc_wlen, to be safe from threading errors */
8923 my->mc_wlen[my->mc_toggle] = 0;
8924 return my->mc_error;
8927 /** Depth-first tree traversal for compacting copy.
8928 * @param[in] my control structure.
8929 * @param[in,out] pg database root.
8930 * @param[in] flags includes #F_DUPDATA if it is a sorted-duplicate sub-DB.
8933 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8935 MDB_cursor mc = {0};
8937 MDB_page *mo, *mp, *leaf;
8942 /* Empty DB, nothing to do */
8943 if (*pg == P_INVALID)
8947 mc.mc_txn = my->mc_txn;
8949 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
8952 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8956 /* Make cursor pages writable */
8957 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8961 for (i=0; i<mc.mc_top; i++) {
8962 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8963 mc.mc_pg[i] = (MDB_page *)ptr;
8964 ptr += my->mc_env->me_psize;
8967 /* This is writable space for a leaf page. Usually not needed. */
8968 leaf = (MDB_page *)ptr;
8970 toggle = my->mc_toggle;
8971 while (mc.mc_snum > 0) {
8973 mp = mc.mc_pg[mc.mc_top];
8977 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8978 for (i=0; i<n; i++) {
8979 ni = NODEPTR(mp, i);
8980 if (ni->mn_flags & F_BIGDATA) {
8984 /* Need writable leaf */
8986 mc.mc_pg[mc.mc_top] = leaf;
8987 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8989 ni = NODEPTR(mp, i);
8992 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8993 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
8994 rc = mdb_page_get(&mc, pg, &omp, NULL);
8997 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8998 rc = mdb_env_cthr_toggle(my, 1);
9001 toggle = my->mc_toggle;
9003 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9004 memcpy(mo, omp, my->mc_env->me_psize);
9005 mo->mp_pgno = my->mc_next_pgno;
9006 my->mc_next_pgno += omp->mp_pages;
9007 my->mc_wlen[toggle] += my->mc_env->me_psize;
9008 if (omp->mp_pages > 1) {
9009 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9010 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9011 rc = mdb_env_cthr_toggle(my, 1);
9014 toggle = my->mc_toggle;
9016 } else if (ni->mn_flags & F_SUBDATA) {
9019 /* Need writable leaf */
9021 mc.mc_pg[mc.mc_top] = leaf;
9022 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9024 ni = NODEPTR(mp, i);
9027 memcpy(&db, NODEDATA(ni), sizeof(db));
9028 my->mc_toggle = toggle;
9029 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9032 toggle = my->mc_toggle;
9033 memcpy(NODEDATA(ni), &db, sizeof(db));
9038 mc.mc_ki[mc.mc_top]++;
9039 if (mc.mc_ki[mc.mc_top] < n) {
9042 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9044 rc = mdb_page_get(&mc, pg, &mp, NULL);
9049 mc.mc_ki[mc.mc_top] = 0;
9050 if (IS_BRANCH(mp)) {
9051 /* Whenever we advance to a sibling branch page,
9052 * we must proceed all the way down to its first leaf.
9054 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9057 mc.mc_pg[mc.mc_top] = mp;
9061 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9062 rc = mdb_env_cthr_toggle(my, 1);
9065 toggle = my->mc_toggle;
9067 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9068 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9069 mo->mp_pgno = my->mc_next_pgno++;
9070 my->mc_wlen[toggle] += my->mc_env->me_psize;
9072 /* Update parent if there is one */
9073 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9074 SETPGNO(ni, mo->mp_pgno);
9075 mdb_cursor_pop(&mc);
9077 /* Otherwise we're done */
9087 /** Copy environment with compaction. */
9089 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9094 MDB_txn *txn = NULL;
9096 pgno_t root, new_root;
9097 int rc = MDB_SUCCESS;
9100 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
9101 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
9105 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9106 if (my.mc_wbuf[0] == NULL) {
9107 /* _aligned_malloc() sets errno, but we use Windows error codes */
9108 rc = ERROR_NOT_ENOUGH_MEMORY;
9112 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
9114 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
9116 #ifdef HAVE_MEMALIGN
9117 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9118 if (my.mc_wbuf[0] == NULL) {
9125 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
9131 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9132 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9133 my.mc_next_pgno = NUM_METAS;
9136 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
9140 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9144 mp = (MDB_page *)my.mc_wbuf[0];
9145 memset(mp, 0, NUM_METAS * env->me_psize);
9147 mp->mp_flags = P_META;
9148 mm = (MDB_meta *)METADATA(mp);
9149 mdb_env_init_meta0(env, mm);
9150 mm->mm_address = env->me_metas[0]->mm_address;
9152 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9154 mp->mp_flags = P_META;
9155 *(MDB_meta *)METADATA(mp) = *mm;
9156 mm = (MDB_meta *)METADATA(mp);
9158 /* Set metapage 1 with current main DB */
9159 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
9160 if (root != P_INVALID) {
9161 /* Count free pages + freeDB pages. Subtract from last_pg
9162 * to find the new last_pg, which also becomes the new root.
9164 MDB_ID freecount = 0;
9167 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9168 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9169 freecount += *(MDB_ID *)data.mv_data;
9170 if (rc != MDB_NOTFOUND)
9172 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9173 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9174 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9176 new_root = txn->mt_next_pgno - 1 - freecount;
9177 mm->mm_last_pg = new_root;
9178 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9179 mm->mm_dbs[MAIN_DBI].md_root = new_root;
9181 /* When the DB is empty, handle it specially to
9182 * fix any breakage like page leaks from ITS#8174.
9184 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
9186 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
9187 mm->mm_txnid = 1; /* use metapage 1 */
9190 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9192 rc = mdb_env_cwalk(&my, &root, 0);
9193 if (rc == MDB_SUCCESS && root != new_root) {
9194 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
9200 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
9201 rc = THREAD_FINISH(thr);
9206 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
9207 if (my.mc_cond) CloseHandle(my.mc_cond);
9208 if (my.mc_mutex) CloseHandle(my.mc_mutex);
9210 free(my.mc_wbuf[0]);
9211 pthread_cond_destroy(&my.mc_cond);
9213 pthread_mutex_destroy(&my.mc_mutex);
9215 return rc ? rc : my.mc_error;
9218 /** Copy environment as-is. */
9220 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9222 MDB_txn *txn = NULL;
9223 mdb_mutexref_t wmutex = NULL;
9229 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9233 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9236 /* Do the lock/unlock of the reader mutex before starting the
9237 * write txn. Otherwise other read txns could block writers.
9239 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9244 /* We must start the actual read txn after blocking writers */
9245 mdb_txn_end(txn, MDB_END_RESET_TMP);
9247 /* Temporarily block writers until we snapshot the meta pages */
9248 wmutex = env->me_wmutex;
9249 if (LOCK_MUTEX(rc, env, wmutex))
9252 rc = mdb_txn_renew0(txn);
9254 UNLOCK_MUTEX(wmutex);
9259 wsize = env->me_psize * NUM_METAS;
9263 DO_WRITE(rc, fd, ptr, w2, len);
9267 } else if (len > 0) {
9273 /* Non-blocking or async handles are not supported */
9279 UNLOCK_MUTEX(wmutex);
9284 w3 = txn->mt_next_pgno * env->me_psize;
9287 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9294 if (wsize > MAX_WRITE)
9298 DO_WRITE(rc, fd, ptr, w2, len);
9302 } else if (len > 0) {
9319 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9321 if (flags & MDB_CP_COMPACT)
9322 return mdb_env_copyfd1(env, fd);
9324 return mdb_env_copyfd0(env, fd);
9328 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9330 return mdb_env_copyfd2(env, fd, 0);
9334 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9338 HANDLE newfd = INVALID_HANDLE_VALUE;
9343 if (env->me_flags & MDB_NOSUBDIR) {
9344 lpath = (char *)path;
9347 len += sizeof(DATANAME);
9348 lpath = malloc(len);
9351 sprintf(lpath, "%s" DATANAME, path);
9354 /* The destination path must exist, but the destination file must not.
9355 * We don't want the OS to cache the writes, since the source data is
9356 * already in the OS cache.
9359 rc = utf8_to_utf16(lpath, -1, &wpath, NULL);
9362 newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9363 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9366 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9368 if (newfd == INVALID_HANDLE_VALUE) {
9373 if (env->me_psize >= env->me_os_psize) {
9375 /* Set O_DIRECT if the file system supports it */
9376 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9377 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9379 #ifdef F_NOCACHE /* __APPLE__ */
9380 rc = fcntl(newfd, F_NOCACHE, 1);
9388 rc = mdb_env_copyfd2(env, newfd, flags);
9391 if (!(env->me_flags & MDB_NOSUBDIR))
9393 if (newfd != INVALID_HANDLE_VALUE)
9394 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9401 mdb_env_copy(MDB_env *env, const char *path)
9403 return mdb_env_copy2(env, path, 0);
9407 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9409 if (flag & ~CHANGEABLE)
9412 env->me_flags |= flag;
9414 env->me_flags &= ~flag;
9419 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9424 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9429 mdb_env_set_userctx(MDB_env *env, void *ctx)
9433 env->me_userctx = ctx;
9438 mdb_env_get_userctx(MDB_env *env)
9440 return env ? env->me_userctx : NULL;
9444 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9449 env->me_assert_func = func;
9455 mdb_env_get_path(MDB_env *env, const char **arg)
9460 *arg = env->me_path;
9465 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9474 /** Common code for #mdb_stat() and #mdb_env_stat().
9475 * @param[in] env the environment to operate in.
9476 * @param[in] db the #MDB_db record containing the stats to return.
9477 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9478 * @return 0, this function always succeeds.
9481 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9483 arg->ms_psize = env->me_psize;
9484 arg->ms_depth = db->md_depth;
9485 arg->ms_branch_pages = db->md_branch_pages;
9486 arg->ms_leaf_pages = db->md_leaf_pages;
9487 arg->ms_overflow_pages = db->md_overflow_pages;
9488 arg->ms_entries = db->md_entries;
9494 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9498 if (env == NULL || arg == NULL)
9501 meta = mdb_env_pick_meta(env);
9503 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9507 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9511 if (env == NULL || arg == NULL)
9514 meta = mdb_env_pick_meta(env);
9515 arg->me_mapaddr = meta->mm_address;
9516 arg->me_last_pgno = meta->mm_last_pg;
9517 arg->me_last_txnid = meta->mm_txnid;
9519 arg->me_mapsize = env->me_mapsize;
9520 arg->me_maxreaders = env->me_maxreaders;
9521 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9525 /** Set the default comparison functions for a database.
9526 * Called immediately after a database is opened to set the defaults.
9527 * The user can then override them with #mdb_set_compare() or
9528 * #mdb_set_dupsort().
9529 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9530 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9533 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9535 uint16_t f = txn->mt_dbs[dbi].md_flags;
9537 txn->mt_dbxs[dbi].md_cmp =
9538 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9539 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9541 txn->mt_dbxs[dbi].md_dcmp =
9542 !(f & MDB_DUPSORT) ? 0 :
9543 ((f & MDB_INTEGERDUP)
9544 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9545 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9548 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9554 int rc, dbflag, exact;
9555 unsigned int unused = 0, seq;
9559 if (flags & ~VALID_FLAGS)
9561 if (txn->mt_flags & MDB_TXN_BLOCKED)
9567 if (flags & PERSISTENT_FLAGS) {
9568 uint16_t f2 = flags & PERSISTENT_FLAGS;
9569 /* make sure flag changes get committed */
9570 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9571 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9572 txn->mt_flags |= MDB_TXN_DIRTY;
9575 mdb_default_cmp(txn, MAIN_DBI);
9579 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9580 mdb_default_cmp(txn, MAIN_DBI);
9583 /* Is the DB already open? */
9585 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9586 if (!txn->mt_dbxs[i].md_name.mv_size) {
9587 /* Remember this free slot */
9588 if (!unused) unused = i;
9591 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9592 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9598 /* If no free slot and max hit, fail */
9599 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9600 return MDB_DBS_FULL;
9602 /* Cannot mix named databases with some mainDB flags */
9603 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9604 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9606 /* Find the DB info */
9607 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9610 key.mv_data = (void *)name;
9611 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9612 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9613 if (rc == MDB_SUCCESS) {
9614 /* make sure this is actually a DB */
9615 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9616 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9617 return MDB_INCOMPATIBLE;
9618 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
9622 /* Done here so we cannot fail after creating a new DB */
9623 if ((namedup = strdup(name)) == NULL)
9627 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
9628 data.mv_size = sizeof(MDB_db);
9629 data.mv_data = &dummy;
9630 memset(&dummy, 0, sizeof(dummy));
9631 dummy.md_root = P_INVALID;
9632 dummy.md_flags = flags & PERSISTENT_FLAGS;
9633 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9640 /* Got info, register DBI in this txn */
9641 unsigned int slot = unused ? unused : txn->mt_numdbs;
9642 txn->mt_dbxs[slot].md_name.mv_data = namedup;
9643 txn->mt_dbxs[slot].md_name.mv_size = len;
9644 txn->mt_dbxs[slot].md_rel = NULL;
9645 txn->mt_dbflags[slot] = dbflag;
9646 /* txn-> and env-> are the same in read txns, use
9647 * tmp variable to avoid undefined assignment
9649 seq = ++txn->mt_env->me_dbiseqs[slot];
9650 txn->mt_dbiseqs[slot] = seq;
9652 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9654 mdb_default_cmp(txn, slot);
9664 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9666 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9669 if (txn->mt_flags & MDB_TXN_BLOCKED)
9672 if (txn->mt_dbflags[dbi] & DB_STALE) {
9675 /* Stale, must read the DB's root. cursor_init does it for us. */
9676 mdb_cursor_init(&mc, txn, dbi, &mx);
9678 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9681 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9684 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9686 ptr = env->me_dbxs[dbi].md_name.mv_data;
9687 /* If there was no name, this was already closed */
9689 env->me_dbxs[dbi].md_name.mv_data = NULL;
9690 env->me_dbxs[dbi].md_name.mv_size = 0;
9691 env->me_dbflags[dbi] = 0;
9692 env->me_dbiseqs[dbi]++;
9697 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9699 /* We could return the flags for the FREE_DBI too but what's the point? */
9700 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9702 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9706 /** Add all the DB's pages to the free list.
9707 * @param[in] mc Cursor on the DB to free.
9708 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9709 * @return 0 on success, non-zero on failure.
9712 mdb_drop0(MDB_cursor *mc, int subs)
9716 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9717 if (rc == MDB_SUCCESS) {
9718 MDB_txn *txn = mc->mc_txn;
9723 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9724 * This also avoids any P_LEAF2 pages, which have no nodes.
9725 * Also if the DB doesn't have sub-DBs and has no overflow
9726 * pages, omit scanning leaves.
9728 if ((mc->mc_flags & C_SUB) ||
9729 (!subs && !mc->mc_db->md_overflow_pages))
9732 mdb_cursor_copy(mc, &mx);
9733 while (mc->mc_snum > 0) {
9734 MDB_page *mp = mc->mc_pg[mc->mc_top];
9735 unsigned n = NUMKEYS(mp);
9737 for (i=0; i<n; i++) {
9738 ni = NODEPTR(mp, i);
9739 if (ni->mn_flags & F_BIGDATA) {
9742 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9743 rc = mdb_page_get(mc, pg, &omp, NULL);
9746 mdb_cassert(mc, IS_OVERFLOW(omp));
9747 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9751 mc->mc_db->md_overflow_pages -= omp->mp_pages;
9752 if (!mc->mc_db->md_overflow_pages && !subs)
9754 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9755 mdb_xcursor_init1(mc, ni);
9756 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9761 if (!subs && !mc->mc_db->md_overflow_pages)
9764 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9766 for (i=0; i<n; i++) {
9768 ni = NODEPTR(mp, i);
9771 mdb_midl_xappend(txn->mt_free_pgs, pg);
9776 mc->mc_ki[mc->mc_top] = i;
9777 rc = mdb_cursor_sibling(mc, 1);
9779 if (rc != MDB_NOTFOUND)
9781 /* no more siblings, go back to beginning
9782 * of previous level.
9787 for (i=1; i<mc->mc_snum; i++) {
9789 mc->mc_pg[i] = mx.mc_pg[i];
9794 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9797 txn->mt_flags |= MDB_TXN_ERROR;
9798 } else if (rc == MDB_NOTFOUND) {
9801 mc->mc_flags &= ~C_INITIALIZED;
9805 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9807 MDB_cursor *mc, *m2;
9810 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9813 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9816 if (TXN_DBI_CHANGED(txn, dbi))
9819 rc = mdb_cursor_open(txn, dbi, &mc);
9823 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9824 /* Invalidate the dropped DB's cursors */
9825 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9826 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9830 /* Can't delete the main DB */
9831 if (del && dbi >= CORE_DBS) {
9832 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9834 txn->mt_dbflags[dbi] = DB_STALE;
9835 mdb_dbi_close(txn->mt_env, dbi);
9837 txn->mt_flags |= MDB_TXN_ERROR;
9840 /* reset the DB record, mark it dirty */
9841 txn->mt_dbflags[dbi] |= DB_DIRTY;
9842 txn->mt_dbs[dbi].md_depth = 0;
9843 txn->mt_dbs[dbi].md_branch_pages = 0;
9844 txn->mt_dbs[dbi].md_leaf_pages = 0;
9845 txn->mt_dbs[dbi].md_overflow_pages = 0;
9846 txn->mt_dbs[dbi].md_entries = 0;
9847 txn->mt_dbs[dbi].md_root = P_INVALID;
9849 txn->mt_flags |= MDB_TXN_DIRTY;
9852 mdb_cursor_close(mc);
9856 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9858 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9861 txn->mt_dbxs[dbi].md_cmp = cmp;
9865 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9867 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9870 txn->mt_dbxs[dbi].md_dcmp = cmp;
9874 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9876 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9879 txn->mt_dbxs[dbi].md_rel = rel;
9883 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9885 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9888 txn->mt_dbxs[dbi].md_relctx = ctx;
9893 mdb_env_get_maxkeysize(MDB_env *env)
9895 return ENV_MAXKEY(env);
9899 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9901 unsigned int i, rdrs;
9904 int rc = 0, first = 1;
9908 if (!env->me_txns) {
9909 return func("(no reader locks)\n", ctx);
9911 rdrs = env->me_txns->mti_numreaders;
9912 mr = env->me_txns->mti_readers;
9913 for (i=0; i<rdrs; i++) {
9915 txnid_t txnid = mr[i].mr_txnid;
9916 sprintf(buf, txnid == (txnid_t)-1 ?
9917 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9918 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9921 rc = func(" pid thread txnid\n", ctx);
9925 rc = func(buf, ctx);
9931 rc = func("(no active readers)\n", ctx);
9936 /** Insert pid into list if not already present.
9937 * return -1 if already present.
9940 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9942 /* binary search of pid in list */
9944 unsigned cursor = 1;
9946 unsigned n = ids[0];
9949 unsigned pivot = n >> 1;
9950 cursor = base + pivot + 1;
9951 val = pid - ids[cursor];
9956 } else if ( val > 0 ) {
9961 /* found, so it's a duplicate */
9970 for (n = ids[0]; n > cursor; n--)
9977 mdb_reader_check(MDB_env *env, int *dead)
9983 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9986 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9988 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9990 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9991 unsigned int i, j, rdrs;
9993 MDB_PID_T *pids, pid;
9994 int rc = MDB_SUCCESS, count = 0;
9996 rdrs = env->me_txns->mti_numreaders;
9997 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10001 mr = env->me_txns->mti_readers;
10002 for (i=0; i<rdrs; i++) {
10003 pid = mr[i].mr_pid;
10004 if (pid && pid != env->me_pid) {
10005 if (mdb_pid_insert(pids, pid) == 0) {
10006 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10007 /* Stale reader found */
10010 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10011 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10013 rdrs = 0; /* the above checked all readers */
10015 /* Recheck, a new process may have reused pid */
10016 if (mdb_reader_pid(env, Pidcheck, pid))
10020 for (; j<rdrs; j++)
10021 if (mr[j].mr_pid == pid) {
10022 DPRINTF(("clear stale reader pid %u txn %"Z"d",
10023 (unsigned) pid, mr[j].mr_txnid));
10028 UNLOCK_MUTEX(rmutex);
10039 #ifdef MDB_ROBUST_SUPPORTED
10040 /** Handle #LOCK_MUTEX0() failure.
10041 * Try to repair the lock file if the mutex owner died.
10042 * @param[in] env the environment handle
10043 * @param[in] mutex LOCK_MUTEX0() mutex
10044 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10045 * @return 0 on success with the mutex locked, or an error code on failure.
10048 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10053 if (rc == MDB_OWNERDEAD) {
10054 /* We own the mutex. Clean up after dead previous owner. */
10056 rlocked = (mutex == env->me_rmutex);
10058 /* Keep mti_txnid updated, otherwise next writer can
10059 * overwrite data which latest meta page refers to.
10061 meta = mdb_env_pick_meta(env);
10062 env->me_txns->mti_txnid = meta->mm_txnid;
10063 /* env is hosed if the dead thread was ours */
10065 env->me_flags |= MDB_FATAL_ERROR;
10066 env->me_txn = NULL;
10070 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10071 (rc ? "this process' env is hosed" : "recovering")));
10072 rc2 = mdb_reader_check0(env, rlocked, NULL);
10074 rc2 = mdb_mutex_consistent(mutex);
10075 if (rc || (rc = rc2)) {
10076 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10077 UNLOCK_MUTEX(mutex);
10083 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10088 #endif /* MDB_ROBUST_SUPPORTED */
10091 #if defined(_WIN32)
10092 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
10096 need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
10097 if (need == 0xFFFD)
10101 result = malloc(sizeof(wchar_t) * need);
10104 MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
10110 #endif /* defined(_WIN32) */