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 2011-2014 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
93 #if defined(__sun) || defined(ANDROID)
94 /* Most platforms have posix_memalign, older may only have memalign */
95 #define HAVE_MEMALIGN 1
99 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
100 #include <netinet/in.h>
101 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
104 #if defined(__APPLE__) || defined (BSD)
105 # define MDB_USE_SYSV_SEM 1
106 # define MDB_FDATASYNC fsync
107 #elif defined(ANDROID)
108 # define MDB_FDATASYNC fsync
113 #ifdef MDB_USE_SYSV_SEM
116 #ifdef _SEM_SEMUN_UNDEFINED
119 struct semid_ds *buf;
120 unsigned short *array;
122 #endif /* _SEM_SEMUN_UNDEFINED */
123 #endif /* MDB_USE_SYSV_SEM */
127 #include <valgrind/memcheck.h>
128 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
129 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
130 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
131 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
132 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
134 #define VGMEMP_CREATE(h,r,z)
135 #define VGMEMP_ALLOC(h,a,s)
136 #define VGMEMP_FREE(h,a)
137 #define VGMEMP_DESTROY(h)
138 #define VGMEMP_DEFINED(a,s)
142 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
143 /* Solaris just defines one or the other */
144 # define LITTLE_ENDIAN 1234
145 # define BIG_ENDIAN 4321
146 # ifdef _LITTLE_ENDIAN
147 # define BYTE_ORDER LITTLE_ENDIAN
149 # define BYTE_ORDER BIG_ENDIAN
152 # define BYTE_ORDER __BYTE_ORDER
156 #ifndef LITTLE_ENDIAN
157 #define LITTLE_ENDIAN __LITTLE_ENDIAN
160 #define BIG_ENDIAN __BIG_ENDIAN
163 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
164 #define MISALIGNED_OK 1
170 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
171 # error "Unknown or unsupported endianness (BYTE_ORDER)"
172 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
173 # error "Two's complement, reasonably sized integer types, please"
177 /** Put infrequently used env functions in separate section */
179 # define ESECT __attribute__ ((section("__TEXT,text_env")))
181 # define ESECT __attribute__ ((section("text_env")))
187 /** @defgroup internal LMDB Internals
190 /** @defgroup compat Compatibility Macros
191 * A bunch of macros to minimize the amount of platform-specific ifdefs
192 * needed throughout the rest of the code. When the features this library
193 * needs are similar enough to POSIX to be hidden in a one-or-two line
194 * replacement, this macro approach is used.
198 /** Features under development */
203 #if defined(_WIN32) || (defined(EOWNERDEAD) && !defined(MDB_USE_SYSV_SEM))
204 #define MDB_ROBUST_SUPPORTED 1
207 /** Wrapper around __func__, which is a C99 feature */
208 #if __STDC_VERSION__ >= 199901L
209 # define mdb_func_ __func__
210 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
211 # define mdb_func_ __FUNCTION__
213 /* If a debug message says <mdb_unknown>(), update the #if statements above */
214 # define mdb_func_ "<mdb_unknown>"
218 #define MDB_USE_HASH 1
219 #define MDB_PIDLOCK 0
220 #define THREAD_RET DWORD
221 #define pthread_t HANDLE
222 #define pthread_mutex_t HANDLE
223 #define pthread_cond_t HANDLE
224 typedef HANDLE mdb_mutex_t;
225 #define pthread_key_t DWORD
226 #define pthread_self() GetCurrentThreadId()
227 #define pthread_key_create(x,y) \
228 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
229 #define pthread_key_delete(x) TlsFree(x)
230 #define pthread_getspecific(x) TlsGetValue(x)
231 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
232 #define pthread_mutex_consistent(mutex) 0
233 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
234 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
235 #define pthread_cond_signal(x) SetEvent(*x)
236 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
237 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
238 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
239 #define MDB_MUTEX(env, rw) ((env)->me_##rw##mutex)
240 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
241 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
242 #define getpid() GetCurrentProcessId()
243 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
244 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
245 #define ErrCode() GetLastError()
246 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
247 #define close(fd) (CloseHandle(fd) ? 0 : -1)
248 #define munmap(ptr,len) UnmapViewOfFile(ptr)
249 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
250 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
252 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
256 #define THREAD_RET void *
257 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
258 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
259 #define Z "z" /**< printf format modifier for size_t */
261 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
262 #define MDB_PIDLOCK 1
264 #ifdef MDB_USE_SYSV_SEM
266 typedef struct mdb_mutex {
271 #define MDB_MUTEX(env, rw) (&(env)->me_##rw##mutex)
272 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
273 #define UNLOCK_MUTEX(mutex) do { struct sembuf sb = { mutex->semnum, 1, SEM_UNDO }; semop(mutex->semid, &sb, 1); } while(0)
276 mdb_sem_wait(mdb_mutex_t *sem)
279 struct sembuf sb = { sem->semnum, -1, SEM_UNDO };
280 while ((rc = semop(sem->semid, &sb, 1)) && (rc = errno) == EINTR) ;
285 /** Pointer/HANDLE type of shared mutex/semaphore.
287 typedef pthread_mutex_t mdb_mutex_t;
288 /** Mutex for the reader table (rw = r) or write transaction (rw = w).
290 #define MDB_MUTEX(env, rw) (&(env)->me_txns->mti_##rw##mutex)
291 /** Lock the reader or writer mutex.
292 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
294 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
295 /** Unlock the reader or writer mutex.
297 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
298 #endif /* MDB_USE_SYSV_SEM */
300 /** Get the error code for the last failed system function.
302 #define ErrCode() errno
304 /** An abstraction for a file handle.
305 * On POSIX systems file handles are small integers. On Windows
306 * they're opaque pointers.
310 /** A value for an invalid file handle.
311 * Mainly used to initialize file variables and signify that they are
314 #define INVALID_HANDLE_VALUE (-1)
316 /** Get the size of a memory page for the system.
317 * This is the basic size that the platform's memory manager uses, and is
318 * fundamental to the use of memory-mapped files.
320 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
325 #elif defined(MDB_USE_SYSV_SEM)
328 #define MNAME_LEN (sizeof(pthread_mutex_t))
333 #ifdef MDB_ROBUST_SUPPORTED
334 /** Lock mutex, handle any error, set rc = result.
335 * Return 0 on success, nonzero (not rc) on error.
337 #define LOCK_MUTEX(rc, env, mutex) \
338 (((rc) = LOCK_MUTEX0(mutex)) && \
339 ((rc) = mdb_mutex_failed(env, mutex, rc)))
340 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc);
342 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
343 #define mdb_mutex_failed(env, mutex, rc) (rc)
347 /** A flag for opening a file and requesting synchronous data writes.
348 * This is only used when writing a meta page. It's not strictly needed;
349 * we could just do a normal write and then immediately perform a flush.
350 * But if this flag is available it saves us an extra system call.
352 * @note If O_DSYNC is undefined but exists in /usr/include,
353 * preferably set some compiler flag to get the definition.
354 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
357 # define MDB_DSYNC O_DSYNC
361 /** Function for flushing the data of a file. Define this to fsync
362 * if fdatasync() is not supported.
364 #ifndef MDB_FDATASYNC
365 # define MDB_FDATASYNC fdatasync
369 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
380 /** A page number in the database.
381 * Note that 64 bit page numbers are overkill, since pages themselves
382 * already represent 12-13 bits of addressable memory, and the OS will
383 * always limit applications to a maximum of 63 bits of address space.
385 * @note In the #MDB_node structure, we only store 48 bits of this value,
386 * which thus limits us to only 60 bits of addressable data.
388 typedef MDB_ID pgno_t;
390 /** A transaction ID.
391 * See struct MDB_txn.mt_txnid for details.
393 typedef MDB_ID txnid_t;
395 /** @defgroup debug Debug Macros
399 /** Enable debug output. Needs variable argument macros (a C99 feature).
400 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
401 * read from and written to the database (used for free space management).
407 static int mdb_debug;
408 static txnid_t mdb_debug_start;
410 /** Print a debug message with printf formatting.
411 * Requires double parenthesis around 2 or more args.
413 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
414 # define DPRINTF0(fmt, ...) \
415 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
417 # define DPRINTF(args) ((void) 0)
419 /** Print a debug string.
420 * The string is printed literally, with no format processing.
422 #define DPUTS(arg) DPRINTF(("%s", arg))
423 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
425 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
428 /** @brief The maximum size of a database page.
430 * It is 32k or 64k, since value-PAGEBASE must fit in
431 * #MDB_page.%mp_upper.
433 * LMDB will use database pages < OS pages if needed.
434 * That causes more I/O in write transactions: The OS must
435 * know (read) the whole page before writing a partial page.
437 * Note that we don't currently support Huge pages. On Linux,
438 * regular data files cannot use Huge pages, and in general
439 * Huge pages aren't actually pageable. We rely on the OS
440 * demand-pager to read our data and page it out when memory
441 * pressure from other processes is high. So until OSs have
442 * actual paging support for Huge pages, they're not viable.
444 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
446 /** The minimum number of keys required in a database page.
447 * Setting this to a larger value will place a smaller bound on the
448 * maximum size of a data item. Data items larger than this size will
449 * be pushed into overflow pages instead of being stored directly in
450 * the B-tree node. This value used to default to 4. With a page size
451 * of 4096 bytes that meant that any item larger than 1024 bytes would
452 * go into an overflow page. That also meant that on average 2-3KB of
453 * each overflow page was wasted space. The value cannot be lower than
454 * 2 because then there would no longer be a tree structure. With this
455 * value, items larger than 2KB will go into overflow pages, and on
456 * average only 1KB will be wasted.
458 #define MDB_MINKEYS 2
460 /** A stamp that identifies a file as an LMDB file.
461 * There's nothing special about this value other than that it is easily
462 * recognizable, and it will reflect any byte order mismatches.
464 #define MDB_MAGIC 0xBEEFC0DE
466 /** The version number for a database's datafile format. */
467 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
468 /** The version number for a database's lockfile format. */
469 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
471 /** @brief The max size of a key we can write, or 0 for dynamic max.
473 * Define this as 0 to compute the max from the page size. 511
474 * is default for backwards compat: liblmdb <= 0.9.10 can break
475 * when modifying a DB with keys/dupsort data bigger than its max.
476 * #MDB_DEVEL sets the default to 0.
478 * Data items in an #MDB_DUPSORT database are also limited to
479 * this size, since they're actually keys of a sub-DB. Keys and
480 * #MDB_DUPSORT data items must fit on a node in a regular page.
482 #ifndef MDB_MAXKEYSIZE
483 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
486 /** The maximum size of a key we can write to the environment. */
488 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
490 #define ENV_MAXKEY(env) ((env)->me_maxkey)
493 /** @brief The maximum size of a data item.
495 * We only store a 32 bit value for node sizes.
497 #define MAXDATASIZE 0xffffffffUL
500 /** Key size which fits in a #DKBUF.
503 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
506 * This is used for printing a hex dump of a key's contents.
508 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
509 /** Display a key in hex.
511 * Invoke a function to display a key in hex.
513 #define DKEY(x) mdb_dkey(x, kbuf)
519 /** An invalid page number.
520 * Mainly used to denote an empty tree.
522 #define P_INVALID (~(pgno_t)0)
524 /** Test if the flags \b f are set in a flag word \b w. */
525 #define F_ISSET(w, f) (((w) & (f)) == (f))
527 /** Round \b n up to an even number. */
528 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
530 /** Used for offsets within a single page.
531 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
534 typedef uint16_t indx_t;
536 /** Default size of memory map.
537 * This is certainly too small for any actual applications. Apps should always set
538 * the size explicitly using #mdb_env_set_mapsize().
540 #define DEFAULT_MAPSIZE 1048576
542 /** @defgroup readers Reader Lock Table
543 * Readers don't acquire any locks for their data access. Instead, they
544 * simply record their transaction ID in the reader table. The reader
545 * mutex is needed just to find an empty slot in the reader table. The
546 * slot's address is saved in thread-specific data so that subsequent read
547 * transactions started by the same thread need no further locking to proceed.
549 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
551 * No reader table is used if the database is on a read-only filesystem, or
552 * if #MDB_NOLOCK is set.
554 * Since the database uses multi-version concurrency control, readers don't
555 * actually need any locking. This table is used to keep track of which
556 * readers are using data from which old transactions, so that we'll know
557 * when a particular old transaction is no longer in use. Old transactions
558 * that have discarded any data pages can then have those pages reclaimed
559 * for use by a later write transaction.
561 * The lock table is constructed such that reader slots are aligned with the
562 * processor's cache line size. Any slot is only ever used by one thread.
563 * This alignment guarantees that there will be no contention or cache
564 * thrashing as threads update their own slot info, and also eliminates
565 * any need for locking when accessing a slot.
567 * A writer thread will scan every slot in the table to determine the oldest
568 * outstanding reader transaction. Any freed pages older than this will be
569 * reclaimed by the writer. The writer doesn't use any locks when scanning
570 * this table. This means that there's no guarantee that the writer will
571 * see the most up-to-date reader info, but that's not required for correct
572 * operation - all we need is to know the upper bound on the oldest reader,
573 * we don't care at all about the newest reader. So the only consequence of
574 * reading stale information here is that old pages might hang around a
575 * while longer before being reclaimed. That's actually good anyway, because
576 * the longer we delay reclaiming old pages, the more likely it is that a
577 * string of contiguous pages can be found after coalescing old pages from
578 * many old transactions together.
581 /** Number of slots in the reader table.
582 * This value was chosen somewhat arbitrarily. 126 readers plus a
583 * couple mutexes fit exactly into 8KB on my development machine.
584 * Applications should set the table size using #mdb_env_set_maxreaders().
586 #define DEFAULT_READERS 126
588 /** The size of a CPU cache line in bytes. We want our lock structures
589 * aligned to this size to avoid false cache line sharing in the
591 * This value works for most CPUs. For Itanium this should be 128.
597 /** The information we store in a single slot of the reader table.
598 * In addition to a transaction ID, we also record the process and
599 * thread ID that owns a slot, so that we can detect stale information,
600 * e.g. threads or processes that went away without cleaning up.
601 * @note We currently don't check for stale records. We simply re-init
602 * the table when we know that we're the only process opening the
605 typedef struct MDB_rxbody {
606 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
607 * Multiple readers that start at the same time will probably have the
608 * same ID here. Again, it's not important to exclude them from
609 * anything; all we need to know is which version of the DB they
610 * started from so we can avoid overwriting any data used in that
611 * particular version.
613 volatile txnid_t mrb_txnid;
614 /** The process ID of the process owning this reader txn. */
615 volatile MDB_PID_T mrb_pid;
616 /** The thread ID of the thread owning this txn. */
617 volatile MDB_THR_T mrb_tid;
620 /** The actual reader record, with cacheline padding. */
621 typedef struct MDB_reader {
624 /** shorthand for mrb_txnid */
625 #define mr_txnid mru.mrx.mrb_txnid
626 #define mr_pid mru.mrx.mrb_pid
627 #define mr_tid mru.mrx.mrb_tid
628 /** cache line alignment */
629 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
633 /** The header for the reader table.
634 * The table resides in a memory-mapped file. (This is a different file
635 * than is used for the main database.)
637 * For POSIX the actual mutexes reside in the shared memory of this
638 * mapped file. On Windows, mutexes are named objects allocated by the
639 * kernel; we store the mutex names in this mapped file so that other
640 * processes can grab them. This same approach is also used on
641 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
642 * process-shared POSIX mutexes. For these cases where a named object
643 * is used, the object name is derived from a 64 bit FNV hash of the
644 * environment pathname. As such, naming collisions are extremely
645 * unlikely. If a collision occurs, the results are unpredictable.
647 typedef struct MDB_txbody {
648 /** Stamp identifying this as an LMDB file. It must be set
651 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
654 char mtb_rmname[MNAME_LEN];
655 #elif defined(MDB_USE_SYSV_SEM)
658 /** Mutex protecting access to this table.
659 * This is the #MDB_MUTEX(env,r) reader table lock.
661 pthread_mutex_t mtb_rmutex;
663 /** The ID of the last transaction committed to the database.
664 * This is recorded here only for convenience; the value can always
665 * be determined by reading the main database meta pages.
667 volatile txnid_t mtb_txnid;
668 /** The number of slots that have been used in the reader table.
669 * This always records the maximum count, it is not decremented
670 * when readers release their slots.
672 volatile unsigned mtb_numreaders;
675 /** The actual reader table definition. */
676 typedef struct MDB_txninfo {
679 #define mti_magic mt1.mtb.mtb_magic
680 #define mti_format mt1.mtb.mtb_format
681 #define mti_rmutex mt1.mtb.mtb_rmutex
682 #define mti_rmname mt1.mtb.mtb_rmname
683 #define mti_txnid mt1.mtb.mtb_txnid
684 #define mti_numreaders mt1.mtb.mtb_numreaders
685 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
687 #ifdef MDB_USE_SYSV_SEM
688 #define mti_semid mt1.mtb.mtb_semid
692 char mt2_wmname[MNAME_LEN];
693 #define mti_wmname mt2.mt2_wmname
695 pthread_mutex_t mt2_wmutex;
696 #define mti_wmutex mt2.mt2_wmutex
698 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
701 MDB_reader mti_readers[1];
704 /** Lockfile format signature: version, features and field layout */
705 #define MDB_LOCK_FORMAT \
707 ((MDB_LOCK_VERSION) \
708 /* Flags which describe functionality */ \
709 + (((MNAME_LEN) == 0) << 18) /* MDB_USE_SYSV_SEM */ \
710 + (((MDB_PIDLOCK) != 0) << 16)))
713 /** Common header for all page types.
714 * Overflow records occupy a number of contiguous pages with no
715 * headers on any page after the first.
717 typedef struct MDB_page {
718 #define mp_pgno mp_p.p_pgno
719 #define mp_next mp_p.p_next
721 pgno_t p_pgno; /**< page number */
722 struct MDB_page *p_next; /**< for in-memory list of freed pages */
725 /** @defgroup mdb_page Page Flags
727 * Flags for the page headers.
730 #define P_BRANCH 0x01 /**< branch page */
731 #define P_LEAF 0x02 /**< leaf page */
732 #define P_OVERFLOW 0x04 /**< overflow page */
733 #define P_META 0x08 /**< meta page */
734 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
735 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
736 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
737 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
738 #define P_KEEP 0x8000 /**< leave this page alone during spill */
740 uint16_t mp_flags; /**< @ref mdb_page */
741 #define mp_lower mp_pb.pb.pb_lower
742 #define mp_upper mp_pb.pb.pb_upper
743 #define mp_pages mp_pb.pb_pages
746 indx_t pb_lower; /**< lower bound of free space */
747 indx_t pb_upper; /**< upper bound of free space */
749 uint32_t pb_pages; /**< number of overflow pages */
751 indx_t mp_ptrs[1]; /**< dynamic size */
754 /** Size of the page header, excluding dynamic data at the end */
755 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
757 /** Address of first usable data byte in a page, after the header */
758 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
760 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
761 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
763 /** Number of nodes on a page */
764 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
766 /** The amount of space remaining in the page */
767 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
769 /** The percentage of space used in the page, in tenths of a percent. */
770 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
771 ((env)->me_psize - PAGEHDRSZ))
772 /** The minimum page fill factor, in tenths of a percent.
773 * Pages emptier than this are candidates for merging.
775 #define FILL_THRESHOLD 250
777 /** Test if a page is a leaf page */
778 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
779 /** Test if a page is a LEAF2 page */
780 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
781 /** Test if a page is a branch page */
782 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
783 /** Test if a page is an overflow page */
784 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
785 /** Test if a page is a sub page */
786 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
788 /** The number of overflow pages needed to store the given size. */
789 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
791 /** Link in #MDB_txn.%mt_loose_pgs list */
792 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
794 /** Header for a single key/data pair within a page.
795 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
796 * We guarantee 2-byte alignment for 'MDB_node's.
798 typedef struct MDB_node {
799 /** lo and hi are used for data size on leaf nodes and for
800 * child pgno on branch nodes. On 64 bit platforms, flags
801 * is also used for pgno. (Branch nodes have no flags).
802 * They are in host byte order in case that lets some
803 * accesses be optimized into a 32-bit word access.
805 #if BYTE_ORDER == LITTLE_ENDIAN
806 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
808 unsigned short mn_hi, mn_lo;
810 /** @defgroup mdb_node Node Flags
812 * Flags for node headers.
815 #define F_BIGDATA 0x01 /**< data put on overflow page */
816 #define F_SUBDATA 0x02 /**< data is a sub-database */
817 #define F_DUPDATA 0x04 /**< data has duplicates */
819 /** valid flags for #mdb_node_add() */
820 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
823 unsigned short mn_flags; /**< @ref mdb_node */
824 unsigned short mn_ksize; /**< key size */
825 char mn_data[1]; /**< key and data are appended here */
828 /** Size of the node header, excluding dynamic data at the end */
829 #define NODESIZE offsetof(MDB_node, mn_data)
831 /** Bit position of top word in page number, for shifting mn_flags */
832 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
834 /** Size of a node in a branch page with a given key.
835 * This is just the node header plus the key, there is no data.
837 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
839 /** Size of a node in a leaf page with a given key and data.
840 * This is node header plus key plus data size.
842 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
844 /** Address of node \b i in page \b p */
845 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
847 /** Address of the key for the node */
848 #define NODEKEY(node) (void *)((node)->mn_data)
850 /** Address of the data for a node */
851 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
853 /** Get the page number pointed to by a branch node */
854 #define NODEPGNO(node) \
855 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
856 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
857 /** Set the page number in a branch node */
858 #define SETPGNO(node,pgno) do { \
859 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
860 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
862 /** Get the size of the data in a leaf node */
863 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
864 /** Set the size of the data for a leaf node */
865 #define SETDSZ(node,size) do { \
866 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
867 /** The size of a key in a node */
868 #define NODEKSZ(node) ((node)->mn_ksize)
870 /** Copy a page number from src to dst */
872 #define COPY_PGNO(dst,src) dst = src
874 #if SIZE_MAX > 4294967295UL
875 #define COPY_PGNO(dst,src) do { \
876 unsigned short *s, *d; \
877 s = (unsigned short *)&(src); \
878 d = (unsigned short *)&(dst); \
885 #define COPY_PGNO(dst,src) do { \
886 unsigned short *s, *d; \
887 s = (unsigned short *)&(src); \
888 d = (unsigned short *)&(dst); \
894 /** The address of a key in a LEAF2 page.
895 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
896 * There are no node headers, keys are stored contiguously.
898 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
900 /** Set the \b node's key into \b keyptr, if requested. */
901 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
902 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
904 /** Set the \b node's key into \b key. */
905 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
907 /** Information about a single database in the environment. */
908 typedef struct MDB_db {
909 uint32_t md_pad; /**< also ksize for LEAF2 pages */
910 uint16_t md_flags; /**< @ref mdb_dbi_open */
911 uint16_t md_depth; /**< depth of this tree */
912 pgno_t md_branch_pages; /**< number of internal pages */
913 pgno_t md_leaf_pages; /**< number of leaf pages */
914 pgno_t md_overflow_pages; /**< number of overflow pages */
915 size_t md_entries; /**< number of data items */
916 pgno_t md_root; /**< the root page of this tree */
919 /** mdb_dbi_open flags */
920 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
921 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
922 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
923 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
925 /** Handle for the DB used to track free pages. */
927 /** Handle for the default DB. */
930 /** Meta page content.
931 * A meta page is the start point for accessing a database snapshot.
932 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
934 typedef struct MDB_meta {
935 /** Stamp identifying this as an LMDB file. It must be set
938 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
940 void *mm_address; /**< address for fixed mapping */
941 size_t mm_mapsize; /**< size of mmap region */
942 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
943 /** The size of pages used in this DB */
944 #define mm_psize mm_dbs[0].md_pad
945 /** Any persistent environment flags. @ref mdb_env */
946 #define mm_flags mm_dbs[0].md_flags
947 pgno_t mm_last_pg; /**< last used page in file */
948 volatile txnid_t mm_txnid; /**< txnid that committed this page */
951 /** Buffer for a stack-allocated meta page.
952 * The members define size and alignment, and silence type
953 * aliasing warnings. They are not used directly; that could
954 * mean incorrectly using several union members in parallel.
956 typedef union MDB_metabuf {
959 char mm_pad[PAGEHDRSZ];
964 /** Auxiliary DB info.
965 * The information here is mostly static/read-only. There is
966 * only a single copy of this record in the environment.
968 typedef struct MDB_dbx {
969 MDB_val md_name; /**< name of the database */
970 MDB_cmp_func *md_cmp; /**< function for comparing keys */
971 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
972 MDB_rel_func *md_rel; /**< user relocate function */
973 void *md_relctx; /**< user-provided context for md_rel */
976 /** A database transaction.
977 * Every operation requires a transaction handle.
980 MDB_txn *mt_parent; /**< parent of a nested txn */
981 MDB_txn *mt_child; /**< nested txn under this txn */
982 pgno_t mt_next_pgno; /**< next unallocated page */
983 /** The ID of this transaction. IDs are integers incrementing from 1.
984 * Only committed write transactions increment the ID. If a transaction
985 * aborts, the ID may be re-used by the next writer.
988 MDB_env *mt_env; /**< the DB environment */
989 /** The list of pages that became unused during this transaction.
992 /** The list of loose pages that became unused and may be reused
993 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
995 MDB_page *mt_loose_pgs;
996 /* #Number of loose pages (#mt_loose_pgs) */
998 /** The sorted list of dirty pages we temporarily wrote to disk
999 * because the dirty list was full. page numbers in here are
1000 * shifted left by 1, deleted slots have the LSB set.
1002 MDB_IDL mt_spill_pgs;
1004 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1005 MDB_ID2L dirty_list;
1006 /** For read txns: This thread/txn's reader table slot, or NULL. */
1009 /** Array of records for each DB known in the environment. */
1011 /** Array of MDB_db records for each known DB */
1013 /** Array of sequence numbers for each DB handle */
1014 unsigned int *mt_dbiseqs;
1015 /** @defgroup mt_dbflag Transaction DB Flags
1019 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1020 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1021 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1022 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1024 /** In write txns, array of cursors for each DB */
1025 MDB_cursor **mt_cursors;
1026 /** Array of flags for each DB */
1027 unsigned char *mt_dbflags;
1028 /** Number of DB records in use. This number only ever increments;
1029 * we don't decrement it when individual DB handles are closed.
1033 /** @defgroup mdb_txn Transaction Flags
1037 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
1038 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1039 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1040 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1042 unsigned int mt_flags; /**< @ref mdb_txn */
1043 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1044 * Includes ancestor txns' dirty pages not hidden by other txns'
1045 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1046 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1048 unsigned int mt_dirty_room;
1051 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1052 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1053 * raise this on a 64 bit machine.
1055 #define CURSOR_STACK 32
1059 /** Cursors are used for all DB operations.
1060 * A cursor holds a path of (page pointer, key index) from the DB
1061 * root to a position in the DB, plus other state. #MDB_DUPSORT
1062 * cursors include an xcursor to the current data item. Write txns
1063 * track their cursors and keep them up to date when data moves.
1064 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1065 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1068 /** Next cursor on this DB in this txn */
1069 MDB_cursor *mc_next;
1070 /** Backup of the original cursor if this cursor is a shadow */
1071 MDB_cursor *mc_backup;
1072 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1073 struct MDB_xcursor *mc_xcursor;
1074 /** The transaction that owns this cursor */
1076 /** The database handle this cursor operates on */
1078 /** The database record for this cursor */
1080 /** The database auxiliary record for this cursor */
1082 /** The @ref mt_dbflag for this database */
1083 unsigned char *mc_dbflag;
1084 unsigned short mc_snum; /**< number of pushed pages */
1085 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1086 /** @defgroup mdb_cursor Cursor Flags
1088 * Cursor state flags.
1091 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1092 #define C_EOF 0x02 /**< No more data */
1093 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1094 #define C_DEL 0x08 /**< last op was a cursor_del */
1095 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1096 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1098 unsigned int mc_flags; /**< @ref mdb_cursor */
1099 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1100 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1103 /** Context for sorted-dup records.
1104 * We could have gone to a fully recursive design, with arbitrarily
1105 * deep nesting of sub-databases. But for now we only handle these
1106 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1108 typedef struct MDB_xcursor {
1109 /** A sub-cursor for traversing the Dup DB */
1110 MDB_cursor mx_cursor;
1111 /** The database record for this Dup DB */
1113 /** The auxiliary DB record for this Dup DB */
1115 /** The @ref mt_dbflag for this Dup DB */
1116 unsigned char mx_dbflag;
1119 /** State of FreeDB old pages, stored in the MDB_env */
1120 typedef struct MDB_pgstate {
1121 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1122 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1125 /** The database environment. */
1127 HANDLE me_fd; /**< The main data file */
1128 HANDLE me_lfd; /**< The lock file */
1129 HANDLE me_mfd; /**< just for writing the meta pages */
1130 /** Failed to update the meta page. Probably an I/O error. */
1131 #define MDB_FATAL_ERROR 0x80000000U
1132 /** Some fields are initialized. */
1133 #define MDB_ENV_ACTIVE 0x20000000U
1134 /** me_txkey is set */
1135 #define MDB_ENV_TXKEY 0x10000000U
1136 uint32_t me_flags; /**< @ref mdb_env */
1137 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1138 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1139 unsigned int me_maxreaders; /**< size of the reader table */
1140 unsigned int me_numreaders; /**< max numreaders set by this env */
1141 MDB_dbi me_numdbs; /**< number of DBs opened */
1142 MDB_dbi me_maxdbs; /**< size of the DB table */
1143 MDB_PID_T me_pid; /**< process ID of this env */
1144 char *me_path; /**< path to the DB files */
1145 char *me_map; /**< the memory map of the data file */
1146 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1147 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1148 void *me_pbuf; /**< scratch area for DUPSORT put() */
1149 MDB_txn *me_txn; /**< current write transaction */
1150 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1151 size_t me_mapsize; /**< size of the data memory map */
1152 off_t me_size; /**< current file size */
1153 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1154 MDB_dbx *me_dbxs; /**< array of static DB info */
1155 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1156 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1157 pthread_key_t me_txkey; /**< thread-key for readers */
1158 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1159 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1160 # define me_pglast me_pgstate.mf_pglast
1161 # define me_pghead me_pgstate.mf_pghead
1162 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1163 /** IDL of pages that became unused in a write txn */
1164 MDB_IDL me_free_pgs;
1165 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1166 MDB_ID2L me_dirty_list;
1167 /** Max number of freelist items that can fit in a single overflow page */
1169 /** Max size of a node on a page */
1170 unsigned int me_nodemax;
1171 #if !(MDB_MAXKEYSIZE)
1172 unsigned int me_maxkey; /**< max size of a key */
1174 int me_live_reader; /**< have liveness lock in reader table */
1176 int me_pidquery; /**< Used in OpenProcess */
1178 #if defined(_WIN32) || defined(MDB_USE_SYSV_SEM)
1179 /* Windows mutexes/SysV semaphores do not reside in shared mem */
1180 mdb_mutex_t me_rmutex;
1181 mdb_mutex_t me_wmutex;
1183 void *me_userctx; /**< User-settable context */
1184 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1187 /** Nested transaction */
1188 typedef struct MDB_ntxn {
1189 MDB_txn mnt_txn; /**< the transaction */
1190 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1193 /** max number of pages to commit in one writev() call */
1194 #define MDB_COMMIT_PAGES 64
1195 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1196 #undef MDB_COMMIT_PAGES
1197 #define MDB_COMMIT_PAGES IOV_MAX
1200 /** max bytes to write in one call */
1201 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1203 /** Check \b txn and \b dbi arguments to a function */
1204 #define TXN_DBI_EXIST(txn, dbi) \
1205 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1207 /** Check for misused \b dbi handles */
1208 #define TXN_DBI_CHANGED(txn, dbi) \
1209 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1211 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1212 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1213 static int mdb_page_touch(MDB_cursor *mc);
1215 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1216 static int mdb_page_search_root(MDB_cursor *mc,
1217 MDB_val *key, int modify);
1218 #define MDB_PS_MODIFY 1
1219 #define MDB_PS_ROOTONLY 2
1220 #define MDB_PS_FIRST 4
1221 #define MDB_PS_LAST 8
1222 static int mdb_page_search(MDB_cursor *mc,
1223 MDB_val *key, int flags);
1224 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1226 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1227 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1228 pgno_t newpgno, unsigned int nflags);
1230 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1231 static int mdb_env_pick_meta(const MDB_env *env);
1232 static int mdb_env_write_meta(MDB_txn *txn);
1233 #if !(defined(_WIN32) || defined(MDB_USE_SYSV_SEM)) /* Drop unused excl arg */
1234 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1236 static void mdb_env_close0(MDB_env *env, int excl);
1238 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1239 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1240 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1241 static void mdb_node_del(MDB_cursor *mc, int ksize);
1242 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1243 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1244 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1245 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1246 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1248 static int mdb_rebalance(MDB_cursor *mc);
1249 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1251 static void mdb_cursor_pop(MDB_cursor *mc);
1252 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1254 static int mdb_cursor_del0(MDB_cursor *mc);
1255 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1256 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1257 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1258 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1259 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1261 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1262 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1264 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1265 static void mdb_xcursor_init0(MDB_cursor *mc);
1266 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1268 static int mdb_drop0(MDB_cursor *mc, int subs);
1269 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1270 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1273 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1277 static SECURITY_DESCRIPTOR mdb_null_sd;
1278 static SECURITY_ATTRIBUTES mdb_all_sa;
1279 static int mdb_sec_inited;
1282 /** Return the library version info. */
1284 mdb_version(int *major, int *minor, int *patch)
1286 if (major) *major = MDB_VERSION_MAJOR;
1287 if (minor) *minor = MDB_VERSION_MINOR;
1288 if (patch) *patch = MDB_VERSION_PATCH;
1289 return MDB_VERSION_STRING;
1292 /** Table of descriptions for LMDB @ref errors */
1293 static char *const mdb_errstr[] = {
1294 "MDB_KEYEXIST: Key/data pair already exists",
1295 "MDB_NOTFOUND: No matching key/data pair found",
1296 "MDB_PAGE_NOTFOUND: Requested page not found",
1297 "MDB_CORRUPTED: Located page was wrong type",
1298 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1299 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1300 "MDB_INVALID: File is not an LMDB file",
1301 "MDB_MAP_FULL: Environment mapsize limit reached",
1302 "MDB_DBS_FULL: Environment maxdbs limit reached",
1303 "MDB_READERS_FULL: Environment maxreaders limit reached",
1304 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1305 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1306 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1307 "MDB_PAGE_FULL: Internal error - page has no more space",
1308 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1309 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1310 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1311 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1312 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1313 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1317 mdb_strerror(int err)
1320 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1321 * This works as long as no function between the call to mdb_strerror
1322 * and the actual use of the message uses more than 4K of stack.
1325 char buf[1024], *ptr = buf;
1329 return ("Successful return: 0");
1331 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1332 i = err - MDB_KEYEXIST;
1333 return mdb_errstr[i];
1337 /* These are the C-runtime error codes we use. The comment indicates
1338 * their numeric value, and the Win32 error they would correspond to
1339 * if the error actually came from a Win32 API. A major mess, we should
1340 * have used LMDB-specific error codes for everything.
1343 case ENOENT: /* 2, FILE_NOT_FOUND */
1344 case EIO: /* 5, ACCESS_DENIED */
1345 case ENOMEM: /* 12, INVALID_ACCESS */
1346 case EACCES: /* 13, INVALID_DATA */
1347 case EBUSY: /* 16, CURRENT_DIRECTORY */
1348 case EINVAL: /* 22, BAD_COMMAND */
1349 case ENOSPC: /* 28, OUT_OF_PAPER */
1350 return strerror(err);
1355 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1356 FORMAT_MESSAGE_IGNORE_INSERTS,
1357 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1360 return strerror(err);
1364 /** assert(3) variant in cursor context */
1365 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1366 /** assert(3) variant in transaction context */
1367 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1368 /** assert(3) variant in environment context */
1369 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1372 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1373 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1376 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1377 const char *func, const char *file, int line)
1380 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1381 file, line, expr_txt, func);
1382 if (env->me_assert_func)
1383 env->me_assert_func(env, buf);
1384 fprintf(stderr, "%s\n", buf);
1388 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1392 /** Return the page number of \b mp which may be sub-page, for debug output */
1394 mdb_dbg_pgno(MDB_page *mp)
1397 COPY_PGNO(ret, mp->mp_pgno);
1401 /** Display a key in hexadecimal and return the address of the result.
1402 * @param[in] key the key to display
1403 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1404 * @return The key in hexadecimal form.
1407 mdb_dkey(MDB_val *key, char *buf)
1410 unsigned char *c = key->mv_data;
1416 if (key->mv_size > DKBUF_MAXKEYSIZE)
1417 return "MDB_MAXKEYSIZE";
1418 /* may want to make this a dynamic check: if the key is mostly
1419 * printable characters, print it as-is instead of converting to hex.
1423 for (i=0; i<key->mv_size; i++)
1424 ptr += sprintf(ptr, "%02x", *c++);
1426 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1432 mdb_leafnode_type(MDB_node *n)
1434 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1435 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1436 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1439 /** Display all the keys in the page. */
1441 mdb_page_list(MDB_page *mp)
1443 pgno_t pgno = mdb_dbg_pgno(mp);
1444 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1446 unsigned int i, nkeys, nsize, total = 0;
1450 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1451 case P_BRANCH: type = "Branch page"; break;
1452 case P_LEAF: type = "Leaf page"; break;
1453 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1454 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1455 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1457 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1458 pgno, mp->mp_pages, state);
1461 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1462 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1465 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1469 nkeys = NUMKEYS(mp);
1470 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1472 for (i=0; i<nkeys; i++) {
1473 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1474 key.mv_size = nsize = mp->mp_pad;
1475 key.mv_data = LEAF2KEY(mp, i, nsize);
1477 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1480 node = NODEPTR(mp, i);
1481 key.mv_size = node->mn_ksize;
1482 key.mv_data = node->mn_data;
1483 nsize = NODESIZE + key.mv_size;
1484 if (IS_BRANCH(mp)) {
1485 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1489 if (F_ISSET(node->mn_flags, F_BIGDATA))
1490 nsize += sizeof(pgno_t);
1492 nsize += NODEDSZ(node);
1494 nsize += sizeof(indx_t);
1495 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1496 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1498 total = EVEN(total);
1500 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1501 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1505 mdb_cursor_chk(MDB_cursor *mc)
1511 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1512 for (i=0; i<mc->mc_top; i++) {
1514 node = NODEPTR(mp, mc->mc_ki[i]);
1515 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1518 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1524 /** Count all the pages in each DB and in the freelist
1525 * and make sure it matches the actual number of pages
1527 * All named DBs must be open for a correct count.
1529 static void mdb_audit(MDB_txn *txn)
1533 MDB_ID freecount, count;
1538 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1539 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1540 freecount += *(MDB_ID *)data.mv_data;
1541 mdb_tassert(txn, rc == MDB_NOTFOUND);
1544 for (i = 0; i<txn->mt_numdbs; i++) {
1546 if (!(txn->mt_dbflags[i] & DB_VALID))
1548 mdb_cursor_init(&mc, txn, i, &mx);
1549 if (txn->mt_dbs[i].md_root == P_INVALID)
1551 count += txn->mt_dbs[i].md_branch_pages +
1552 txn->mt_dbs[i].md_leaf_pages +
1553 txn->mt_dbs[i].md_overflow_pages;
1554 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1555 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1556 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1559 mp = mc.mc_pg[mc.mc_top];
1560 for (j=0; j<NUMKEYS(mp); j++) {
1561 MDB_node *leaf = NODEPTR(mp, j);
1562 if (leaf->mn_flags & F_SUBDATA) {
1564 memcpy(&db, NODEDATA(leaf), sizeof(db));
1565 count += db.md_branch_pages + db.md_leaf_pages +
1566 db.md_overflow_pages;
1570 mdb_tassert(txn, rc == MDB_NOTFOUND);
1573 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1574 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1575 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1581 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1583 return txn->mt_dbxs[dbi].md_cmp(a, b);
1587 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1589 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1592 /** Allocate memory for a page.
1593 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1596 mdb_page_malloc(MDB_txn *txn, unsigned num)
1598 MDB_env *env = txn->mt_env;
1599 MDB_page *ret = env->me_dpages;
1600 size_t psize = env->me_psize, sz = psize, off;
1601 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1602 * For a single page alloc, we init everything after the page header.
1603 * For multi-page, we init the final page; if the caller needed that
1604 * many pages they will be filling in at least up to the last page.
1608 VGMEMP_ALLOC(env, ret, sz);
1609 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1610 env->me_dpages = ret->mp_next;
1613 psize -= off = PAGEHDRSZ;
1618 if ((ret = malloc(sz)) != NULL) {
1619 VGMEMP_ALLOC(env, ret, sz);
1620 if (!(env->me_flags & MDB_NOMEMINIT)) {
1621 memset((char *)ret + off, 0, psize);
1625 txn->mt_flags |= MDB_TXN_ERROR;
1629 /** Free a single page.
1630 * Saves single pages to a list, for future reuse.
1631 * (This is not used for multi-page overflow pages.)
1634 mdb_page_free(MDB_env *env, MDB_page *mp)
1636 mp->mp_next = env->me_dpages;
1637 VGMEMP_FREE(env, mp);
1638 env->me_dpages = mp;
1641 /** Free a dirty page */
1643 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1645 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1646 mdb_page_free(env, dp);
1648 /* large pages just get freed directly */
1649 VGMEMP_FREE(env, dp);
1654 /** Return all dirty pages to dpage list */
1656 mdb_dlist_free(MDB_txn *txn)
1658 MDB_env *env = txn->mt_env;
1659 MDB_ID2L dl = txn->mt_u.dirty_list;
1660 unsigned i, n = dl[0].mid;
1662 for (i = 1; i <= n; i++) {
1663 mdb_dpage_free(env, dl[i].mptr);
1668 /** Loosen or free a single page.
1669 * Saves single pages to a list for future reuse
1670 * in this same txn. It has been pulled from the freeDB
1671 * and already resides on the dirty list, but has been
1672 * deleted. Use these pages first before pulling again
1675 * If the page wasn't dirtied in this txn, just add it
1676 * to this txn's free list.
1679 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1682 pgno_t pgno = mp->mp_pgno;
1683 MDB_txn *txn = mc->mc_txn;
1685 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1686 if (txn->mt_parent) {
1687 MDB_ID2 *dl = txn->mt_u.dirty_list;
1688 /* If txn has a parent, make sure the page is in our
1692 unsigned x = mdb_mid2l_search(dl, pgno);
1693 if (x <= dl[0].mid && dl[x].mid == pgno) {
1694 if (mp != dl[x].mptr) { /* bad cursor? */
1695 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1696 txn->mt_flags |= MDB_TXN_ERROR;
1697 return MDB_CORRUPTED;
1704 /* no parent txn, so it's just ours */
1709 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1711 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1712 txn->mt_loose_pgs = mp;
1713 txn->mt_loose_count++;
1714 mp->mp_flags |= P_LOOSE;
1716 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1724 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1725 * @param[in] mc A cursor handle for the current operation.
1726 * @param[in] pflags Flags of the pages to update:
1727 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1728 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1729 * @return 0 on success, non-zero on failure.
1732 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1734 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1735 MDB_txn *txn = mc->mc_txn;
1741 int rc = MDB_SUCCESS, level;
1743 /* Mark pages seen by cursors */
1744 if (mc->mc_flags & C_UNTRACK)
1745 mc = NULL; /* will find mc in mt_cursors */
1746 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1747 for (; mc; mc=mc->mc_next) {
1748 if (!(mc->mc_flags & C_INITIALIZED))
1750 for (m3 = mc;; m3 = &mx->mx_cursor) {
1752 for (j=0; j<m3->mc_snum; j++) {
1754 if ((mp->mp_flags & Mask) == pflags)
1755 mp->mp_flags ^= P_KEEP;
1757 mx = m3->mc_xcursor;
1758 /* Proceed to mx if it is at a sub-database */
1759 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1761 if (! (mp && (mp->mp_flags & P_LEAF)))
1763 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1764 if (!(leaf->mn_flags & F_SUBDATA))
1773 /* Mark dirty root pages */
1774 for (i=0; i<txn->mt_numdbs; i++) {
1775 if (txn->mt_dbflags[i] & DB_DIRTY) {
1776 pgno_t pgno = txn->mt_dbs[i].md_root;
1777 if (pgno == P_INVALID)
1779 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1781 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1782 dp->mp_flags ^= P_KEEP;
1790 static int mdb_page_flush(MDB_txn *txn, int keep);
1792 /** Spill pages from the dirty list back to disk.
1793 * This is intended to prevent running into #MDB_TXN_FULL situations,
1794 * but note that they may still occur in a few cases:
1795 * 1) our estimate of the txn size could be too small. Currently this
1796 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1797 * 2) child txns may run out of space if their parents dirtied a
1798 * lot of pages and never spilled them. TODO: we probably should do
1799 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1800 * the parent's dirty_room is below a given threshold.
1802 * Otherwise, if not using nested txns, it is expected that apps will
1803 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1804 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1805 * If the txn never references them again, they can be left alone.
1806 * If the txn only reads them, they can be used without any fuss.
1807 * If the txn writes them again, they can be dirtied immediately without
1808 * going thru all of the work of #mdb_page_touch(). Such references are
1809 * handled by #mdb_page_unspill().
1811 * Also note, we never spill DB root pages, nor pages of active cursors,
1812 * because we'll need these back again soon anyway. And in nested txns,
1813 * we can't spill a page in a child txn if it was already spilled in a
1814 * parent txn. That would alter the parent txns' data even though
1815 * the child hasn't committed yet, and we'd have no way to undo it if
1816 * the child aborted.
1818 * @param[in] m0 cursor A cursor handle identifying the transaction and
1819 * database for which we are checking space.
1820 * @param[in] key For a put operation, the key being stored.
1821 * @param[in] data For a put operation, the data being stored.
1822 * @return 0 on success, non-zero on failure.
1825 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1827 MDB_txn *txn = m0->mc_txn;
1829 MDB_ID2L dl = txn->mt_u.dirty_list;
1830 unsigned int i, j, need;
1833 if (m0->mc_flags & C_SUB)
1836 /* Estimate how much space this op will take */
1837 i = m0->mc_db->md_depth;
1838 /* Named DBs also dirty the main DB */
1839 if (m0->mc_dbi > MAIN_DBI)
1840 i += txn->mt_dbs[MAIN_DBI].md_depth;
1841 /* For puts, roughly factor in the key+data size */
1843 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1844 i += i; /* double it for good measure */
1847 if (txn->mt_dirty_room > i)
1850 if (!txn->mt_spill_pgs) {
1851 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1852 if (!txn->mt_spill_pgs)
1855 /* purge deleted slots */
1856 MDB_IDL sl = txn->mt_spill_pgs;
1857 unsigned int num = sl[0];
1859 for (i=1; i<=num; i++) {
1866 /* Preserve pages which may soon be dirtied again */
1867 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1870 /* Less aggressive spill - we originally spilled the entire dirty list,
1871 * with a few exceptions for cursor pages and DB root pages. But this
1872 * turns out to be a lot of wasted effort because in a large txn many
1873 * of those pages will need to be used again. So now we spill only 1/8th
1874 * of the dirty pages. Testing revealed this to be a good tradeoff,
1875 * better than 1/2, 1/4, or 1/10.
1877 if (need < MDB_IDL_UM_MAX / 8)
1878 need = MDB_IDL_UM_MAX / 8;
1880 /* Save the page IDs of all the pages we're flushing */
1881 /* flush from the tail forward, this saves a lot of shifting later on. */
1882 for (i=dl[0].mid; i && need; i--) {
1883 MDB_ID pn = dl[i].mid << 1;
1885 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1887 /* Can't spill twice, make sure it's not already in a parent's
1890 if (txn->mt_parent) {
1892 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1893 if (tx2->mt_spill_pgs) {
1894 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1895 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1896 dp->mp_flags |= P_KEEP;
1904 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1908 mdb_midl_sort(txn->mt_spill_pgs);
1910 /* Flush the spilled part of dirty list */
1911 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1914 /* Reset any dirty pages we kept that page_flush didn't see */
1915 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1918 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1922 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1924 mdb_find_oldest(MDB_txn *txn)
1927 txnid_t mr, oldest = txn->mt_txnid - 1;
1928 if (txn->mt_env->me_txns) {
1929 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1930 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1941 /** Add a page to the txn's dirty list */
1943 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1946 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1948 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1949 insert = mdb_mid2l_append;
1951 insert = mdb_mid2l_insert;
1953 mid.mid = mp->mp_pgno;
1955 rc = insert(txn->mt_u.dirty_list, &mid);
1956 mdb_tassert(txn, rc == 0);
1957 txn->mt_dirty_room--;
1960 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1961 * me_pghead and mt_next_pgno.
1963 * If there are free pages available from older transactions, they
1964 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1965 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1966 * and move me_pglast to say which records were consumed. Only this
1967 * function can create me_pghead and move me_pglast/mt_next_pgno.
1968 * @param[in] mc cursor A cursor handle identifying the transaction and
1969 * database for which we are allocating.
1970 * @param[in] num the number of pages to allocate.
1971 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1972 * will always be satisfied by a single contiguous chunk of memory.
1973 * @return 0 on success, non-zero on failure.
1976 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1978 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1979 /* Get at most <Max_retries> more freeDB records once me_pghead
1980 * has enough pages. If not enough, use new pages from the map.
1981 * If <Paranoid> and mc is updating the freeDB, only get new
1982 * records if me_pghead is empty. Then the freelist cannot play
1983 * catch-up with itself by growing while trying to save it.
1985 enum { Paranoid = 1, Max_retries = 500 };
1987 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1989 int rc, retry = num * 60;
1990 MDB_txn *txn = mc->mc_txn;
1991 MDB_env *env = txn->mt_env;
1992 pgno_t pgno, *mop = env->me_pghead;
1993 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1995 txnid_t oldest = 0, last;
2000 /* If there are any loose pages, just use them */
2001 if (num == 1 && txn->mt_loose_pgs) {
2002 np = txn->mt_loose_pgs;
2003 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2004 txn->mt_loose_count--;
2005 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2013 /* If our dirty list is already full, we can't do anything */
2014 if (txn->mt_dirty_room == 0) {
2019 for (op = MDB_FIRST;; op = MDB_NEXT) {
2024 /* Seek a big enough contiguous page range. Prefer
2025 * pages at the tail, just truncating the list.
2031 if (mop[i-n2] == pgno+n2)
2038 if (op == MDB_FIRST) { /* 1st iteration */
2039 /* Prepare to fetch more and coalesce */
2040 last = env->me_pglast;
2041 oldest = env->me_pgoldest;
2042 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2045 key.mv_data = &last; /* will look up last+1 */
2046 key.mv_size = sizeof(last);
2048 if (Paranoid && mc->mc_dbi == FREE_DBI)
2051 if (Paranoid && retry < 0 && mop_len)
2055 /* Do not fetch more if the record will be too recent */
2056 if (oldest <= last) {
2058 oldest = mdb_find_oldest(txn);
2059 env->me_pgoldest = oldest;
2065 rc = mdb_cursor_get(&m2, &key, NULL, op);
2067 if (rc == MDB_NOTFOUND)
2071 last = *(txnid_t*)key.mv_data;
2072 if (oldest <= last) {
2074 oldest = mdb_find_oldest(txn);
2075 env->me_pgoldest = oldest;
2081 np = m2.mc_pg[m2.mc_top];
2082 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2083 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2086 idl = (MDB_ID *) data.mv_data;
2089 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2094 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2096 mop = env->me_pghead;
2098 env->me_pglast = last;
2100 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2101 last, txn->mt_dbs[FREE_DBI].md_root, i));
2103 DPRINTF(("IDL %"Z"u", idl[j]));
2105 /* Merge in descending sorted order */
2106 mdb_midl_xmerge(mop, idl);
2110 /* Use new pages from the map when nothing suitable in the freeDB */
2112 pgno = txn->mt_next_pgno;
2113 if (pgno + num >= env->me_maxpg) {
2114 DPUTS("DB size maxed out");
2120 if (env->me_flags & MDB_WRITEMAP) {
2121 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2123 if (!(np = mdb_page_malloc(txn, num))) {
2129 mop[0] = mop_len -= num;
2130 /* Move any stragglers down */
2131 for (j = i-num; j < mop_len; )
2132 mop[++j] = mop[++i];
2134 txn->mt_next_pgno = pgno + num;
2137 mdb_page_dirty(txn, np);
2143 txn->mt_flags |= MDB_TXN_ERROR;
2147 /** Copy the used portions of a non-overflow page.
2148 * @param[in] dst page to copy into
2149 * @param[in] src page to copy from
2150 * @param[in] psize size of a page
2153 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2155 enum { Align = sizeof(pgno_t) };
2156 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2158 /* If page isn't full, just copy the used portion. Adjust
2159 * alignment so memcpy may copy words instead of bytes.
2161 if ((unused &= -Align) && !IS_LEAF2(src)) {
2162 upper = (upper + PAGEBASE) & -Align;
2163 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2164 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2167 memcpy(dst, src, psize - unused);
2171 /** Pull a page off the txn's spill list, if present.
2172 * If a page being referenced was spilled to disk in this txn, bring
2173 * it back and make it dirty/writable again.
2174 * @param[in] txn the transaction handle.
2175 * @param[in] mp the page being referenced. It must not be dirty.
2176 * @param[out] ret the writable page, if any. ret is unchanged if
2177 * mp wasn't spilled.
2180 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2182 MDB_env *env = txn->mt_env;
2185 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2187 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2188 if (!tx2->mt_spill_pgs)
2190 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2191 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2194 if (txn->mt_dirty_room == 0)
2195 return MDB_TXN_FULL;
2196 if (IS_OVERFLOW(mp))
2200 if (env->me_flags & MDB_WRITEMAP) {
2203 np = mdb_page_malloc(txn, num);
2207 memcpy(np, mp, num * env->me_psize);
2209 mdb_page_copy(np, mp, env->me_psize);
2212 /* If in current txn, this page is no longer spilled.
2213 * If it happens to be the last page, truncate the spill list.
2214 * Otherwise mark it as deleted by setting the LSB.
2216 if (x == txn->mt_spill_pgs[0])
2217 txn->mt_spill_pgs[0]--;
2219 txn->mt_spill_pgs[x] |= 1;
2220 } /* otherwise, if belonging to a parent txn, the
2221 * page remains spilled until child commits
2224 mdb_page_dirty(txn, np);
2225 np->mp_flags |= P_DIRTY;
2233 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2234 * @param[in] mc cursor pointing to the page to be touched
2235 * @return 0 on success, non-zero on failure.
2238 mdb_page_touch(MDB_cursor *mc)
2240 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2241 MDB_txn *txn = mc->mc_txn;
2242 MDB_cursor *m2, *m3;
2246 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2247 if (txn->mt_flags & MDB_TXN_SPILLS) {
2249 rc = mdb_page_unspill(txn, mp, &np);
2255 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2256 (rc = mdb_page_alloc(mc, 1, &np)))
2259 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2260 mp->mp_pgno, pgno));
2261 mdb_cassert(mc, mp->mp_pgno != pgno);
2262 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2263 /* Update the parent page, if any, to point to the new page */
2265 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2266 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2267 SETPGNO(node, pgno);
2269 mc->mc_db->md_root = pgno;
2271 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2272 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2274 /* If txn has a parent, make sure the page is in our
2278 unsigned x = mdb_mid2l_search(dl, pgno);
2279 if (x <= dl[0].mid && dl[x].mid == pgno) {
2280 if (mp != dl[x].mptr) { /* bad cursor? */
2281 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2282 txn->mt_flags |= MDB_TXN_ERROR;
2283 return MDB_CORRUPTED;
2288 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2290 np = mdb_page_malloc(txn, 1);
2295 rc = mdb_mid2l_insert(dl, &mid);
2296 mdb_cassert(mc, rc == 0);
2301 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2303 np->mp_flags |= P_DIRTY;
2306 /* Adjust cursors pointing to mp */
2307 mc->mc_pg[mc->mc_top] = np;
2308 m2 = txn->mt_cursors[mc->mc_dbi];
2309 if (mc->mc_flags & C_SUB) {
2310 for (; m2; m2=m2->mc_next) {
2311 m3 = &m2->mc_xcursor->mx_cursor;
2312 if (m3->mc_snum < mc->mc_snum) continue;
2313 if (m3->mc_pg[mc->mc_top] == mp)
2314 m3->mc_pg[mc->mc_top] = np;
2317 for (; m2; m2=m2->mc_next) {
2318 if (m2->mc_snum < mc->mc_snum) continue;
2319 if (m2->mc_pg[mc->mc_top] == mp) {
2320 m2->mc_pg[mc->mc_top] = np;
2321 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2323 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2325 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2326 if (!(leaf->mn_flags & F_SUBDATA))
2327 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2335 txn->mt_flags |= MDB_TXN_ERROR;
2340 mdb_env_sync(MDB_env *env, int force)
2343 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2344 if (env->me_flags & MDB_WRITEMAP) {
2345 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2346 ? MS_ASYNC : MS_SYNC;
2347 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2350 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2354 if (MDB_FDATASYNC(env->me_fd))
2361 /** Back up parent txn's cursors, then grab the originals for tracking */
2363 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2365 MDB_cursor *mc, *bk;
2370 for (i = src->mt_numdbs; --i >= 0; ) {
2371 if ((mc = src->mt_cursors[i]) != NULL) {
2372 size = sizeof(MDB_cursor);
2374 size += sizeof(MDB_xcursor);
2375 for (; mc; mc = bk->mc_next) {
2381 mc->mc_db = &dst->mt_dbs[i];
2382 /* Kill pointers into src - and dst to reduce abuse: The
2383 * user may not use mc until dst ends. Otherwise we'd...
2385 mc->mc_txn = NULL; /* ...set this to dst */
2386 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2387 if ((mx = mc->mc_xcursor) != NULL) {
2388 *(MDB_xcursor *)(bk+1) = *mx;
2389 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2391 mc->mc_next = dst->mt_cursors[i];
2392 dst->mt_cursors[i] = mc;
2399 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2400 * @param[in] txn the transaction handle.
2401 * @param[in] merge true to keep changes to parent cursors, false to revert.
2402 * @return 0 on success, non-zero on failure.
2405 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2407 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2411 for (i = txn->mt_numdbs; --i >= 0; ) {
2412 for (mc = cursors[i]; mc; mc = next) {
2414 if ((bk = mc->mc_backup) != NULL) {
2416 /* Commit changes to parent txn */
2417 mc->mc_next = bk->mc_next;
2418 mc->mc_backup = bk->mc_backup;
2419 mc->mc_txn = bk->mc_txn;
2420 mc->mc_db = bk->mc_db;
2421 mc->mc_dbflag = bk->mc_dbflag;
2422 if ((mx = mc->mc_xcursor) != NULL)
2423 mx->mx_cursor.mc_txn = bk->mc_txn;
2425 /* Abort nested txn */
2427 if ((mx = mc->mc_xcursor) != NULL)
2428 *mx = *(MDB_xcursor *)(bk+1);
2432 /* Only malloced cursors are permanently tracked. */
2440 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2443 mdb_txn_reset0(MDB_txn *txn, const char *act);
2445 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2451 Pidset = F_SETLK, Pidcheck = F_GETLK
2455 /** Set or check a pid lock. Set returns 0 on success.
2456 * Check returns 0 if the process is certainly dead, nonzero if it may
2457 * be alive (the lock exists or an error happened so we do not know).
2459 * On Windows Pidset is a no-op, we merely check for the existence
2460 * of the process with the given pid. On POSIX we use a single byte
2461 * lock on the lockfile, set at an offset equal to the pid.
2464 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2466 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2469 if (op == Pidcheck) {
2470 h = OpenProcess(env->me_pidquery, FALSE, pid);
2471 /* No documented "no such process" code, but other program use this: */
2473 return ErrCode() != ERROR_INVALID_PARAMETER;
2474 /* A process exists until all handles to it close. Has it exited? */
2475 ret = WaitForSingleObject(h, 0) != 0;
2482 struct flock lock_info;
2483 memset(&lock_info, 0, sizeof(lock_info));
2484 lock_info.l_type = F_WRLCK;
2485 lock_info.l_whence = SEEK_SET;
2486 lock_info.l_start = pid;
2487 lock_info.l_len = 1;
2488 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2489 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2491 } else if ((rc = ErrCode()) == EINTR) {
2499 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2500 * @param[in] txn the transaction handle to initialize
2501 * @return 0 on success, non-zero on failure.
2504 mdb_txn_renew0(MDB_txn *txn)
2506 MDB_env *env = txn->mt_env;
2507 MDB_txninfo *ti = env->me_txns;
2511 int rc, new_notls = 0;
2513 if (txn->mt_flags & MDB_TXN_RDONLY) {
2515 txn->mt_numdbs = env->me_numdbs;
2516 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2518 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2519 txn->mt_txnid = meta->mm_txnid;
2520 txn->mt_u.reader = NULL;
2522 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2523 pthread_getspecific(env->me_txkey);
2525 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2526 return MDB_BAD_RSLOT;
2528 MDB_PID_T pid = env->me_pid;
2529 MDB_THR_T tid = pthread_self();
2530 mdb_mutex_t *rmutex = MDB_MUTEX(env, r);
2532 if (!env->me_live_reader) {
2533 rc = mdb_reader_pid(env, Pidset, pid);
2536 env->me_live_reader = 1;
2539 if (LOCK_MUTEX(rc, env, rmutex))
2541 nr = ti->mti_numreaders;
2542 for (i=0; i<nr; i++)
2543 if (ti->mti_readers[i].mr_pid == 0)
2545 if (i == env->me_maxreaders) {
2546 UNLOCK_MUTEX(rmutex);
2547 return MDB_READERS_FULL;
2549 r = &ti->mti_readers[i];
2550 r->mr_txnid = (txnid_t)-1;
2552 r->mr_pid = pid; /* should be written last, see ITS#7971. */
2554 ti->mti_numreaders = ++nr;
2555 /* Save numreaders for un-mutexed mdb_env_close() */
2556 env->me_numreaders = nr;
2557 UNLOCK_MUTEX(rmutex);
2559 new_notls = (env->me_flags & MDB_NOTLS);
2560 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2565 do /* LY: Retry on a race, ITS#7970. */
2566 r->mr_txnid = ti->mti_txnid;
2567 while(r->mr_txnid != ti->mti_txnid);
2568 txn->mt_txnid = r->mr_txnid;
2569 txn->mt_u.reader = r;
2570 meta = env->me_metas[txn->mt_txnid & 1];
2574 if (LOCK_MUTEX(rc, env, MDB_MUTEX(env, w)))
2577 txn->mt_txnid = ti->mti_txnid;
2578 meta = env->me_metas[txn->mt_txnid & 1];
2580 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2581 txn->mt_txnid = meta->mm_txnid;
2584 txn->mt_numdbs = env->me_numdbs;
2587 if (txn->mt_txnid == mdb_debug_start)
2591 txn->mt_child = NULL;
2592 txn->mt_loose_pgs = NULL;
2593 txn->mt_loose_count = 0;
2594 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2595 txn->mt_u.dirty_list = env->me_dirty_list;
2596 txn->mt_u.dirty_list[0].mid = 0;
2597 txn->mt_free_pgs = env->me_free_pgs;
2598 txn->mt_free_pgs[0] = 0;
2599 txn->mt_spill_pgs = NULL;
2601 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2604 /* Copy the DB info and flags */
2605 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2607 /* Moved to here to avoid a data race in read TXNs */
2608 txn->mt_next_pgno = meta->mm_last_pg+1;
2610 for (i=2; i<txn->mt_numdbs; i++) {
2611 x = env->me_dbflags[i];
2612 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2613 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2615 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2617 if (env->me_maxpg < txn->mt_next_pgno) {
2618 mdb_txn_reset0(txn, "renew0-mapfail");
2620 txn->mt_u.reader->mr_pid = 0;
2621 txn->mt_u.reader = NULL;
2623 return MDB_MAP_RESIZED;
2630 mdb_txn_renew(MDB_txn *txn)
2634 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2637 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2638 DPUTS("environment had fatal error, must shutdown!");
2642 rc = mdb_txn_renew0(txn);
2643 if (rc == MDB_SUCCESS) {
2644 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2645 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2646 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2652 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2656 int rc, size, tsize = sizeof(MDB_txn);
2658 if (env->me_flags & MDB_FATAL_ERROR) {
2659 DPUTS("environment had fatal error, must shutdown!");
2662 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2665 /* Nested transactions: Max 1 child, write txns only, no writemap */
2666 if (parent->mt_child ||
2667 (flags & MDB_RDONLY) ||
2668 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2669 (env->me_flags & MDB_WRITEMAP))
2671 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2673 tsize = sizeof(MDB_ntxn);
2676 if (!(flags & MDB_RDONLY)) {
2678 txn = env->me_txn0; /* just reuse preallocated write txn */
2681 /* child txns use own copy of cursors */
2682 size += env->me_maxdbs * sizeof(MDB_cursor *);
2684 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2686 if ((txn = calloc(1, size)) == NULL) {
2687 DPRINTF(("calloc: %s", strerror(errno)));
2690 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2691 if (flags & MDB_RDONLY) {
2692 txn->mt_flags |= MDB_TXN_RDONLY;
2693 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2694 txn->mt_dbiseqs = env->me_dbiseqs;
2696 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2698 txn->mt_dbiseqs = parent->mt_dbiseqs;
2699 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2701 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2702 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2710 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2711 if (!txn->mt_u.dirty_list ||
2712 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2714 free(txn->mt_u.dirty_list);
2718 txn->mt_txnid = parent->mt_txnid;
2719 txn->mt_dirty_room = parent->mt_dirty_room;
2720 txn->mt_u.dirty_list[0].mid = 0;
2721 txn->mt_spill_pgs = NULL;
2722 txn->mt_next_pgno = parent->mt_next_pgno;
2723 parent->mt_child = txn;
2724 txn->mt_parent = parent;
2725 txn->mt_numdbs = parent->mt_numdbs;
2726 txn->mt_flags = parent->mt_flags;
2727 txn->mt_dbxs = parent->mt_dbxs;
2728 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2729 /* Copy parent's mt_dbflags, but clear DB_NEW */
2730 for (i=0; i<txn->mt_numdbs; i++)
2731 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2733 ntxn = (MDB_ntxn *)txn;
2734 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2735 if (env->me_pghead) {
2736 size = MDB_IDL_SIZEOF(env->me_pghead);
2737 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2739 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2744 rc = mdb_cursor_shadow(parent, txn);
2746 mdb_txn_reset0(txn, "beginchild-fail");
2748 rc = mdb_txn_renew0(txn);
2751 if (txn != env->me_txn0)
2755 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2756 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2757 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2764 mdb_txn_env(MDB_txn *txn)
2766 if(!txn) return NULL;
2770 /** Export or close DBI handles opened in this txn. */
2772 mdb_dbis_update(MDB_txn *txn, int keep)
2775 MDB_dbi n = txn->mt_numdbs;
2776 MDB_env *env = txn->mt_env;
2777 unsigned char *tdbflags = txn->mt_dbflags;
2779 for (i = n; --i >= 2;) {
2780 if (tdbflags[i] & DB_NEW) {
2782 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2784 char *ptr = env->me_dbxs[i].md_name.mv_data;
2786 env->me_dbxs[i].md_name.mv_data = NULL;
2787 env->me_dbxs[i].md_name.mv_size = 0;
2788 env->me_dbflags[i] = 0;
2789 env->me_dbiseqs[i]++;
2795 if (keep && env->me_numdbs < n)
2799 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2800 * May be called twice for readonly txns: First reset it, then abort.
2801 * @param[in] txn the transaction handle to reset
2802 * @param[in] act why the transaction is being reset
2805 mdb_txn_reset0(MDB_txn *txn, const char *act)
2807 MDB_env *env = txn->mt_env;
2809 /* Close any DBI handles opened in this txn */
2810 mdb_dbis_update(txn, 0);
2812 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2813 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2814 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2816 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2817 if (txn->mt_u.reader) {
2818 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2819 if (!(env->me_flags & MDB_NOTLS))
2820 txn->mt_u.reader = NULL; /* txn does not own reader */
2822 txn->mt_numdbs = 0; /* close nothing if called again */
2823 txn->mt_dbxs = NULL; /* mark txn as reset */
2825 pgno_t *pghead = env->me_pghead;
2827 mdb_cursors_close(txn, 0);
2828 if (!(env->me_flags & MDB_WRITEMAP)) {
2829 mdb_dlist_free(txn);
2832 if (!txn->mt_parent) {
2833 if (mdb_midl_shrink(&txn->mt_free_pgs))
2834 env->me_free_pgs = txn->mt_free_pgs;
2836 env->me_pghead = NULL;
2840 /* The writer mutex was locked in mdb_txn_begin. */
2842 UNLOCK_MUTEX(MDB_MUTEX(env, w));
2844 txn->mt_parent->mt_child = NULL;
2845 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2846 mdb_midl_free(txn->mt_free_pgs);
2847 mdb_midl_free(txn->mt_spill_pgs);
2848 free(txn->mt_u.dirty_list);
2851 mdb_midl_free(pghead);
2856 mdb_txn_reset(MDB_txn *txn)
2861 /* This call is only valid for read-only txns */
2862 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2865 mdb_txn_reset0(txn, "reset");
2869 mdb_txn_abort(MDB_txn *txn)
2875 mdb_txn_abort(txn->mt_child);
2877 mdb_txn_reset0(txn, "abort");
2878 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2879 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2880 txn->mt_u.reader->mr_pid = 0;
2882 if (txn != txn->mt_env->me_txn0)
2886 /** Save the freelist as of this transaction to the freeDB.
2887 * This changes the freelist. Keep trying until it stabilizes.
2890 mdb_freelist_save(MDB_txn *txn)
2892 /* env->me_pghead[] can grow and shrink during this call.
2893 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2894 * Page numbers cannot disappear from txn->mt_free_pgs[].
2897 MDB_env *env = txn->mt_env;
2898 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2899 txnid_t pglast = 0, head_id = 0;
2900 pgno_t freecnt = 0, *free_pgs, *mop;
2901 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2903 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2905 if (env->me_pghead) {
2906 /* Make sure first page of freeDB is touched and on freelist */
2907 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2908 if (rc && rc != MDB_NOTFOUND)
2912 if (!env->me_pghead && txn->mt_loose_pgs) {
2913 /* Put loose page numbers in mt_free_pgs, since
2914 * we may be unable to return them to me_pghead.
2916 MDB_page *mp = txn->mt_loose_pgs;
2917 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2919 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2920 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2921 txn->mt_loose_pgs = NULL;
2922 txn->mt_loose_count = 0;
2925 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2926 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2927 ? SSIZE_MAX : maxfree_1pg;
2930 /* Come back here after each Put() in case freelist changed */
2935 /* If using records from freeDB which we have not yet
2936 * deleted, delete them and any we reserved for me_pghead.
2938 while (pglast < env->me_pglast) {
2939 rc = mdb_cursor_first(&mc, &key, NULL);
2942 pglast = head_id = *(txnid_t *)key.mv_data;
2943 total_room = head_room = 0;
2944 mdb_tassert(txn, pglast <= env->me_pglast);
2945 rc = mdb_cursor_del(&mc, 0);
2950 /* Save the IDL of pages freed by this txn, to a single record */
2951 if (freecnt < txn->mt_free_pgs[0]) {
2953 /* Make sure last page of freeDB is touched and on freelist */
2954 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2955 if (rc && rc != MDB_NOTFOUND)
2958 free_pgs = txn->mt_free_pgs;
2959 /* Write to last page of freeDB */
2960 key.mv_size = sizeof(txn->mt_txnid);
2961 key.mv_data = &txn->mt_txnid;
2963 freecnt = free_pgs[0];
2964 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2965 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2968 /* Retry if mt_free_pgs[] grew during the Put() */
2969 free_pgs = txn->mt_free_pgs;
2970 } while (freecnt < free_pgs[0]);
2971 mdb_midl_sort(free_pgs);
2972 memcpy(data.mv_data, free_pgs, data.mv_size);
2975 unsigned int i = free_pgs[0];
2976 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2977 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2979 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2985 mop = env->me_pghead;
2986 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2988 /* Reserve records for me_pghead[]. Split it if multi-page,
2989 * to avoid searching freeDB for a page range. Use keys in
2990 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2992 if (total_room >= mop_len) {
2993 if (total_room == mop_len || --more < 0)
2995 } else if (head_room >= maxfree_1pg && head_id > 1) {
2996 /* Keep current record (overflow page), add a new one */
3000 /* (Re)write {key = head_id, IDL length = head_room} */
3001 total_room -= head_room;
3002 head_room = mop_len - total_room;
3003 if (head_room > maxfree_1pg && head_id > 1) {
3004 /* Overflow multi-page for part of me_pghead */
3005 head_room /= head_id; /* amortize page sizes */
3006 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3007 } else if (head_room < 0) {
3008 /* Rare case, not bothering to delete this record */
3011 key.mv_size = sizeof(head_id);
3012 key.mv_data = &head_id;
3013 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3014 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3017 /* IDL is initially empty, zero out at least the length */
3018 pgs = (pgno_t *)data.mv_data;
3019 j = head_room > clean_limit ? head_room : 0;
3023 total_room += head_room;
3026 /* Return loose page numbers to me_pghead, though usually none are
3027 * left at this point. The pages themselves remain in dirty_list.
3029 if (txn->mt_loose_pgs) {
3030 MDB_page *mp = txn->mt_loose_pgs;
3031 unsigned count = txn->mt_loose_count;
3033 /* Room for loose pages + temp IDL with same */
3034 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3036 mop = env->me_pghead;
3037 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3038 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3039 loose[ ++count ] = mp->mp_pgno;
3041 mdb_midl_sort(loose);
3042 mdb_midl_xmerge(mop, loose);
3043 txn->mt_loose_pgs = NULL;
3044 txn->mt_loose_count = 0;
3048 /* Fill in the reserved me_pghead records */
3054 rc = mdb_cursor_first(&mc, &key, &data);
3055 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3056 txnid_t id = *(txnid_t *)key.mv_data;
3057 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3060 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3062 if (len > mop_len) {
3064 data.mv_size = (len + 1) * sizeof(MDB_ID);
3066 data.mv_data = mop -= len;
3069 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3071 if (rc || !(mop_len -= len))
3078 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3079 * @param[in] txn the transaction that's being committed
3080 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3081 * @return 0 on success, non-zero on failure.
3084 mdb_page_flush(MDB_txn *txn, int keep)
3086 MDB_env *env = txn->mt_env;
3087 MDB_ID2L dl = txn->mt_u.dirty_list;
3088 unsigned psize = env->me_psize, j;
3089 int i, pagecount = dl[0].mid, rc;
3090 size_t size = 0, pos = 0;
3092 MDB_page *dp = NULL;
3096 struct iovec iov[MDB_COMMIT_PAGES];
3097 ssize_t wpos = 0, wsize = 0, wres;
3098 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3104 if (env->me_flags & MDB_WRITEMAP) {
3105 /* Clear dirty flags */
3106 while (++i <= pagecount) {
3108 /* Don't flush this page yet */
3109 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3110 dp->mp_flags &= ~P_KEEP;
3114 dp->mp_flags &= ~P_DIRTY;
3119 /* Write the pages */
3121 if (++i <= pagecount) {
3123 /* Don't flush this page yet */
3124 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3125 dp->mp_flags &= ~P_KEEP;
3130 /* clear dirty flag */
3131 dp->mp_flags &= ~P_DIRTY;
3134 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3139 /* Windows actually supports scatter/gather I/O, but only on
3140 * unbuffered file handles. Since we're relying on the OS page
3141 * cache for all our data, that's self-defeating. So we just
3142 * write pages one at a time. We use the ov structure to set
3143 * the write offset, to at least save the overhead of a Seek
3146 DPRINTF(("committing page %"Z"u", pgno));
3147 memset(&ov, 0, sizeof(ov));
3148 ov.Offset = pos & 0xffffffff;
3149 ov.OffsetHigh = pos >> 16 >> 16;
3150 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3152 DPRINTF(("WriteFile: %d", rc));
3156 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3157 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3159 /* Write previous page(s) */
3160 #ifdef MDB_USE_PWRITEV
3161 wres = pwritev(env->me_fd, iov, n, wpos);
3164 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3166 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3168 DPRINTF(("lseek: %s", strerror(rc)));
3171 wres = writev(env->me_fd, iov, n);
3174 if (wres != wsize) {
3177 DPRINTF(("Write error: %s", strerror(rc)));
3179 rc = EIO; /* TODO: Use which error code? */
3180 DPUTS("short write, filesystem full?");
3191 DPRINTF(("committing page %"Z"u", pgno));
3192 next_pos = pos + size;
3193 iov[n].iov_len = size;
3194 iov[n].iov_base = (char *)dp;
3200 /* MIPS has cache coherency issues, this is a no-op everywhere else
3201 * Note: for any size >= on-chip cache size, entire on-chip cache is
3204 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3206 for (i = keep; ++i <= pagecount; ) {
3208 /* This is a page we skipped above */
3211 dl[j].mid = dp->mp_pgno;
3214 mdb_dpage_free(env, dp);
3219 txn->mt_dirty_room += i - j;
3225 mdb_txn_commit(MDB_txn *txn)
3231 if (txn == NULL || txn->mt_env == NULL)
3234 if (txn->mt_child) {
3235 rc = mdb_txn_commit(txn->mt_child);
3236 txn->mt_child = NULL;
3243 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3244 mdb_dbis_update(txn, 1);
3245 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3250 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3251 DPUTS("error flag is set, can't commit");
3253 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3258 if (txn->mt_parent) {
3259 MDB_txn *parent = txn->mt_parent;
3263 unsigned x, y, len, ps_len;
3265 /* Append our free list to parent's */
3266 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3269 mdb_midl_free(txn->mt_free_pgs);
3270 /* Failures after this must either undo the changes
3271 * to the parent or set MDB_TXN_ERROR in the parent.
3274 parent->mt_next_pgno = txn->mt_next_pgno;
3275 parent->mt_flags = txn->mt_flags;
3277 /* Merge our cursors into parent's and close them */
3278 mdb_cursors_close(txn, 1);
3280 /* Update parent's DB table. */
3281 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3282 parent->mt_numdbs = txn->mt_numdbs;
3283 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3284 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3285 for (i=2; i<txn->mt_numdbs; i++) {
3286 /* preserve parent's DB_NEW status */
3287 x = parent->mt_dbflags[i] & DB_NEW;
3288 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3291 dst = parent->mt_u.dirty_list;
3292 src = txn->mt_u.dirty_list;
3293 /* Remove anything in our dirty list from parent's spill list */
3294 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3296 pspill[0] = (pgno_t)-1;
3297 /* Mark our dirty pages as deleted in parent spill list */
3298 for (i=0, len=src[0].mid; ++i <= len; ) {
3299 MDB_ID pn = src[i].mid << 1;
3300 while (pn > pspill[x])
3302 if (pn == pspill[x]) {
3307 /* Squash deleted pagenums if we deleted any */
3308 for (x=y; ++x <= ps_len; )
3309 if (!(pspill[x] & 1))
3310 pspill[++y] = pspill[x];
3314 /* Find len = length of merging our dirty list with parent's */
3316 dst[0].mid = 0; /* simplify loops */
3317 if (parent->mt_parent) {
3318 len = x + src[0].mid;
3319 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3320 for (i = x; y && i; y--) {
3321 pgno_t yp = src[y].mid;
3322 while (yp < dst[i].mid)
3324 if (yp == dst[i].mid) {
3329 } else { /* Simplify the above for single-ancestor case */
3330 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3332 /* Merge our dirty list with parent's */
3334 for (i = len; y; dst[i--] = src[y--]) {
3335 pgno_t yp = src[y].mid;
3336 while (yp < dst[x].mid)
3337 dst[i--] = dst[x--];
3338 if (yp == dst[x].mid)
3339 free(dst[x--].mptr);
3341 mdb_tassert(txn, i == x);
3343 free(txn->mt_u.dirty_list);
3344 parent->mt_dirty_room = txn->mt_dirty_room;
3345 if (txn->mt_spill_pgs) {
3346 if (parent->mt_spill_pgs) {
3347 /* TODO: Prevent failure here, so parent does not fail */
3348 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3350 parent->mt_flags |= MDB_TXN_ERROR;
3351 mdb_midl_free(txn->mt_spill_pgs);
3352 mdb_midl_sort(parent->mt_spill_pgs);
3354 parent->mt_spill_pgs = txn->mt_spill_pgs;
3358 /* Append our loose page list to parent's */
3359 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3361 *lp = txn->mt_loose_pgs;
3362 parent->mt_loose_count += txn->mt_loose_count;
3364 parent->mt_child = NULL;
3365 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3370 if (txn != env->me_txn) {
3371 DPUTS("attempt to commit unknown transaction");
3376 mdb_cursors_close(txn, 0);
3378 if (!txn->mt_u.dirty_list[0].mid &&
3379 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3382 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3383 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3385 /* Update DB root pointers */
3386 if (txn->mt_numdbs > 2) {
3390 data.mv_size = sizeof(MDB_db);
3392 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3393 for (i = 2; i < txn->mt_numdbs; i++) {
3394 if (txn->mt_dbflags[i] & DB_DIRTY) {
3395 if (TXN_DBI_CHANGED(txn, i)) {
3399 data.mv_data = &txn->mt_dbs[i];
3400 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3407 rc = mdb_freelist_save(txn);
3411 mdb_midl_free(env->me_pghead);
3412 env->me_pghead = NULL;
3413 if (mdb_midl_shrink(&txn->mt_free_pgs))
3414 env->me_free_pgs = txn->mt_free_pgs;
3420 if ((rc = mdb_page_flush(txn, 0)) ||
3421 (rc = mdb_env_sync(env, 0)) ||
3422 (rc = mdb_env_write_meta(txn)))
3425 /* Free P_LOOSE pages left behind in dirty_list */
3426 if (!(env->me_flags & MDB_WRITEMAP))
3427 mdb_dlist_free(txn);
3432 mdb_dbis_update(txn, 1);
3435 UNLOCK_MUTEX(MDB_MUTEX(env, w));
3436 if (txn != env->me_txn0)
3446 /** Read the environment parameters of a DB environment before
3447 * mapping it into memory.
3448 * @param[in] env the environment handle
3449 * @param[out] meta address of where to store the meta information
3450 * @return 0 on success, non-zero on failure.
3453 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3459 enum { Size = sizeof(pbuf) };
3461 /* We don't know the page size yet, so use a minimum value.
3462 * Read both meta pages so we can use the latest one.
3465 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3469 memset(&ov, 0, sizeof(ov));
3471 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3472 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3475 rc = pread(env->me_fd, &pbuf, Size, off);
3478 if (rc == 0 && off == 0)
3480 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3481 DPRINTF(("read: %s", mdb_strerror(rc)));
3485 p = (MDB_page *)&pbuf;
3487 if (!F_ISSET(p->mp_flags, P_META)) {
3488 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3493 if (m->mm_magic != MDB_MAGIC) {
3494 DPUTS("meta has invalid magic");
3498 if (m->mm_version != MDB_DATA_VERSION) {
3499 DPRINTF(("database is version %u, expected version %u",
3500 m->mm_version, MDB_DATA_VERSION));
3501 return MDB_VERSION_MISMATCH;
3504 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3510 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3512 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3514 meta->mm_magic = MDB_MAGIC;
3515 meta->mm_version = MDB_DATA_VERSION;
3516 meta->mm_mapsize = env->me_mapsize;
3517 meta->mm_psize = env->me_psize;
3518 meta->mm_last_pg = 1;
3519 meta->mm_flags = env->me_flags & 0xffff;
3520 meta->mm_flags |= MDB_INTEGERKEY;
3521 meta->mm_dbs[0].md_root = P_INVALID;
3522 meta->mm_dbs[1].md_root = P_INVALID;
3525 /** Write the environment parameters of a freshly created DB environment.
3526 * @param[in] env the environment handle
3527 * @param[in] meta the #MDB_meta to write
3528 * @return 0 on success, non-zero on failure.
3531 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3539 memset(&ov, 0, sizeof(ov));
3540 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3542 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3545 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3546 len = pwrite(fd, ptr, size, pos); \
3547 rc = (len >= 0); } while(0)
3550 DPUTS("writing new meta page");
3552 psize = env->me_psize;
3554 p = calloc(2, psize);
3556 p->mp_flags = P_META;
3557 *(MDB_meta *)METADATA(p) = *meta;
3559 q = (MDB_page *)((char *)p + psize);
3561 q->mp_flags = P_META;
3562 *(MDB_meta *)METADATA(q) = *meta;
3564 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3567 else if ((unsigned) len == psize * 2)
3575 /** Update the environment info to commit a transaction.
3576 * @param[in] txn the transaction that's being committed
3577 * @return 0 on success, non-zero on failure.
3580 mdb_env_write_meta(MDB_txn *txn)
3583 MDB_meta meta, metab, *mp;
3586 int rc, len, toggle;
3595 toggle = txn->mt_txnid & 1;
3596 DPRINTF(("writing meta page %d for root page %"Z"u",
3597 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3600 mp = env->me_metas[toggle];
3601 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3602 /* Persist any increases of mapsize config */
3603 if (mapsize < env->me_mapsize)
3604 mapsize = env->me_mapsize;
3606 if (env->me_flags & MDB_WRITEMAP) {
3607 mp->mm_mapsize = mapsize;
3608 mp->mm_dbs[0] = txn->mt_dbs[0];
3609 mp->mm_dbs[1] = txn->mt_dbs[1];
3610 mp->mm_last_pg = txn->mt_next_pgno - 1;
3611 #if !(defined(_MSC_VER) || defined(__i386__) || defined(__x86_64__))
3612 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3613 __sync_synchronize();
3615 mp->mm_txnid = txn->mt_txnid;
3616 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3617 unsigned meta_size = env->me_psize;
3618 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3621 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3622 if (meta_size < env->me_os_psize)
3623 meta_size += meta_size;
3628 if (MDB_MSYNC(ptr, meta_size, rc)) {
3635 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3636 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3638 meta.mm_mapsize = mapsize;
3639 meta.mm_dbs[0] = txn->mt_dbs[0];
3640 meta.mm_dbs[1] = txn->mt_dbs[1];
3641 meta.mm_last_pg = txn->mt_next_pgno - 1;
3642 meta.mm_txnid = txn->mt_txnid;
3644 off = offsetof(MDB_meta, mm_mapsize);
3645 ptr = (char *)&meta + off;
3646 len = sizeof(MDB_meta) - off;
3648 off += env->me_psize;
3651 /* Write to the SYNC fd */
3652 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3653 env->me_fd : env->me_mfd;
3656 memset(&ov, 0, sizeof(ov));
3658 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3662 rc = pwrite(mfd, ptr, len, off);
3665 rc = rc < 0 ? ErrCode() : EIO;
3666 DPUTS("write failed, disk error?");
3667 /* On a failure, the pagecache still contains the new data.
3668 * Write some old data back, to prevent it from being used.
3669 * Use the non-SYNC fd; we know it will fail anyway.
3671 meta.mm_last_pg = metab.mm_last_pg;
3672 meta.mm_txnid = metab.mm_txnid;
3674 memset(&ov, 0, sizeof(ov));
3676 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3678 r2 = pwrite(env->me_fd, ptr, len, off);
3679 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3682 env->me_flags |= MDB_FATAL_ERROR;
3685 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3686 CACHEFLUSH(env->me_map + off, len, DCACHE);
3688 /* Memory ordering issues are irrelevant; since the entire writer
3689 * is wrapped by wmutex, all of these changes will become visible
3690 * after the wmutex is unlocked. Since the DB is multi-version,
3691 * readers will get consistent data regardless of how fresh or
3692 * how stale their view of these values is.
3695 env->me_txns->mti_txnid = txn->mt_txnid;
3700 /** Check both meta pages to see which one is newer.
3701 * @param[in] env the environment handle
3702 * @return meta toggle (0 or 1).
3705 mdb_env_pick_meta(const MDB_env *env)
3707 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3711 mdb_env_create(MDB_env **env)
3715 e = calloc(1, sizeof(MDB_env));
3719 e->me_maxreaders = DEFAULT_READERS;
3720 e->me_maxdbs = e->me_numdbs = 2;
3721 e->me_fd = INVALID_HANDLE_VALUE;
3722 e->me_lfd = INVALID_HANDLE_VALUE;
3723 e->me_mfd = INVALID_HANDLE_VALUE;
3724 #ifdef MDB_USE_SYSV_SEM
3725 e->me_rmutex.semid = -1;
3726 e->me_wmutex.semid = -1;
3728 e->me_pid = getpid();
3729 GET_PAGESIZE(e->me_os_psize);
3730 VGMEMP_CREATE(e,0,0);
3736 mdb_env_map(MDB_env *env, void *addr)
3739 unsigned int flags = env->me_flags;
3743 LONG sizelo, sizehi;
3746 if (flags & MDB_RDONLY) {
3747 /* Don't set explicit map size, use whatever exists */
3752 msize = env->me_mapsize;
3753 sizelo = msize & 0xffffffff;
3754 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3756 /* Windows won't create mappings for zero length files.
3757 * and won't map more than the file size.
3758 * Just set the maxsize right now.
3760 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3761 || !SetEndOfFile(env->me_fd)
3762 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3766 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3767 PAGE_READWRITE : PAGE_READONLY,
3768 sizehi, sizelo, NULL);
3771 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3772 FILE_MAP_WRITE : FILE_MAP_READ,
3774 rc = env->me_map ? 0 : ErrCode();
3779 int prot = PROT_READ;
3780 if (flags & MDB_WRITEMAP) {
3782 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3785 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3787 if (env->me_map == MAP_FAILED) {
3792 if (flags & MDB_NORDAHEAD) {
3793 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3795 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3797 #ifdef POSIX_MADV_RANDOM
3798 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3799 #endif /* POSIX_MADV_RANDOM */
3800 #endif /* MADV_RANDOM */
3804 /* Can happen because the address argument to mmap() is just a
3805 * hint. mmap() can pick another, e.g. if the range is in use.
3806 * The MAP_FIXED flag would prevent that, but then mmap could
3807 * instead unmap existing pages to make room for the new map.
3809 if (addr && env->me_map != addr)
3810 return EBUSY; /* TODO: Make a new MDB_* error code? */
3812 p = (MDB_page *)env->me_map;
3813 env->me_metas[0] = METADATA(p);
3814 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3820 mdb_env_set_mapsize(MDB_env *env, size_t size)
3822 /* If env is already open, caller is responsible for making
3823 * sure there are no active txns.
3831 meta = env->me_metas[mdb_env_pick_meta(env)];
3833 size = meta->mm_mapsize;
3835 /* Silently round up to minimum if the size is too small */
3836 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3840 munmap(env->me_map, env->me_mapsize);
3841 env->me_mapsize = size;
3842 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3843 rc = mdb_env_map(env, old);
3847 env->me_mapsize = size;
3849 env->me_maxpg = env->me_mapsize / env->me_psize;
3854 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3858 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3863 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3865 if (env->me_map || readers < 1)
3867 env->me_maxreaders = readers;
3872 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3874 if (!env || !readers)
3876 *readers = env->me_maxreaders;
3881 mdb_fsize(HANDLE fd, size_t *size)
3884 LARGE_INTEGER fsize;
3886 if (!GetFileSizeEx(fd, &fsize))
3889 *size = fsize.QuadPart;
3901 /** Further setup required for opening an LMDB environment
3904 mdb_env_open2(MDB_env *env)
3906 unsigned int flags = env->me_flags;
3907 int i, newenv = 0, rc;
3911 /* See if we should use QueryLimited */
3913 if ((rc & 0xff) > 5)
3914 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3916 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3919 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3922 DPUTS("new mdbenv");
3924 env->me_psize = env->me_os_psize;
3925 if (env->me_psize > MAX_PAGESIZE)
3926 env->me_psize = MAX_PAGESIZE;
3927 memset(&meta, 0, sizeof(meta));
3928 mdb_env_init_meta0(env, &meta);
3929 meta.mm_mapsize = DEFAULT_MAPSIZE;
3931 env->me_psize = meta.mm_psize;
3934 /* Was a mapsize configured? */
3935 if (!env->me_mapsize) {
3936 env->me_mapsize = meta.mm_mapsize;
3939 /* Make sure mapsize >= committed data size. Even when using
3940 * mm_mapsize, which could be broken in old files (ITS#7789).
3942 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3943 if (env->me_mapsize < minsize)
3944 env->me_mapsize = minsize;
3946 meta.mm_mapsize = env->me_mapsize;
3948 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3953 if (flags & MDB_FIXEDMAP)
3954 meta.mm_address = env->me_map;
3955 i = mdb_env_init_meta(env, &meta);
3956 if (i != MDB_SUCCESS) {
3961 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3962 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3964 #if !(MDB_MAXKEYSIZE)
3965 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3967 env->me_maxpg = env->me_mapsize / env->me_psize;
3971 int toggle = mdb_env_pick_meta(env);
3972 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3974 DPRINTF(("opened database version %u, pagesize %u",
3975 env->me_metas[0]->mm_version, env->me_psize));
3976 DPRINTF(("using meta page %d", toggle));
3977 DPRINTF(("depth: %u", db->md_depth));
3978 DPRINTF(("entries: %"Z"u", db->md_entries));
3979 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3980 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3981 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3982 DPRINTF(("root: %"Z"u", db->md_root));
3990 /** Release a reader thread's slot in the reader lock table.
3991 * This function is called automatically when a thread exits.
3992 * @param[in] ptr This points to the slot in the reader lock table.
3995 mdb_env_reader_dest(void *ptr)
3997 MDB_reader *reader = ptr;
4003 /** Junk for arranging thread-specific callbacks on Windows. This is
4004 * necessarily platform and compiler-specific. Windows supports up
4005 * to 1088 keys. Let's assume nobody opens more than 64 environments
4006 * in a single process, for now. They can override this if needed.
4008 #ifndef MAX_TLS_KEYS
4009 #define MAX_TLS_KEYS 64
4011 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4012 static int mdb_tls_nkeys;
4014 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4018 case DLL_PROCESS_ATTACH: break;
4019 case DLL_THREAD_ATTACH: break;
4020 case DLL_THREAD_DETACH:
4021 for (i=0; i<mdb_tls_nkeys; i++) {
4022 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4024 mdb_env_reader_dest(r);
4028 case DLL_PROCESS_DETACH: break;
4033 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4035 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4039 /* Force some symbol references.
4040 * _tls_used forces the linker to create the TLS directory if not already done
4041 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4043 #pragma comment(linker, "/INCLUDE:_tls_used")
4044 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4045 #pragma const_seg(".CRT$XLB")
4046 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4047 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4050 #pragma comment(linker, "/INCLUDE:__tls_used")
4051 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4052 #pragma data_seg(".CRT$XLB")
4053 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4055 #endif /* WIN 32/64 */
4056 #endif /* !__GNUC__ */
4059 /** Downgrade the exclusive lock on the region back to shared */
4061 mdb_env_share_locks(MDB_env *env, int *excl)
4063 int rc = 0, toggle = mdb_env_pick_meta(env);
4065 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4070 /* First acquire a shared lock. The Unlock will
4071 * then release the existing exclusive lock.
4073 memset(&ov, 0, sizeof(ov));
4074 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4077 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4083 struct flock lock_info;
4084 /* The shared lock replaces the existing lock */
4085 memset((void *)&lock_info, 0, sizeof(lock_info));
4086 lock_info.l_type = F_RDLCK;
4087 lock_info.l_whence = SEEK_SET;
4088 lock_info.l_start = 0;
4089 lock_info.l_len = 1;
4090 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4091 (rc = ErrCode()) == EINTR) ;
4092 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4099 /** Try to get exclusive lock, otherwise shared.
4100 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4103 mdb_env_excl_lock(MDB_env *env, int *excl)
4107 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4111 memset(&ov, 0, sizeof(ov));
4112 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4119 struct flock lock_info;
4120 memset((void *)&lock_info, 0, sizeof(lock_info));
4121 lock_info.l_type = F_WRLCK;
4122 lock_info.l_whence = SEEK_SET;
4123 lock_info.l_start = 0;
4124 lock_info.l_len = 1;
4125 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4126 (rc = ErrCode()) == EINTR) ;
4130 # ifdef MDB_USE_SYSV_SEM
4131 if (*excl < 0) /* always true when !MDB_USE_SYSV_SEM */
4134 lock_info.l_type = F_RDLCK;
4135 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4136 (rc = ErrCode()) == EINTR) ;
4146 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4148 * @(#) $Revision: 5.1 $
4149 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4150 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4152 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4156 * Please do not copyright this code. This code is in the public domain.
4158 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4159 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4160 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4161 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4162 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4163 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4164 * PERFORMANCE OF THIS SOFTWARE.
4167 * chongo <Landon Curt Noll> /\oo/\
4168 * http://www.isthe.com/chongo/
4170 * Share and Enjoy! :-)
4173 typedef unsigned long long mdb_hash_t;
4174 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4176 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4177 * @param[in] val value to hash
4178 * @param[in] hval initial value for hash
4179 * @return 64 bit hash
4181 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4182 * hval arg on the first call.
4185 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4187 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4188 unsigned char *end = s + val->mv_size;
4190 * FNV-1a hash each octet of the string
4193 /* xor the bottom with the current octet */
4194 hval ^= (mdb_hash_t)*s++;
4196 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4197 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4198 (hval << 7) + (hval << 8) + (hval << 40);
4200 /* return our new hash value */
4204 /** Hash the string and output the encoded hash.
4205 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4206 * very short name limits. We don't care about the encoding being reversible,
4207 * we just want to preserve as many bits of the input as possible in a
4208 * small printable string.
4209 * @param[in] str string to hash
4210 * @param[out] encbuf an array of 11 chars to hold the hash
4212 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4215 mdb_pack85(unsigned long l, char *out)
4219 for (i=0; i<5; i++) {
4220 *out++ = mdb_a85[l % 85];
4226 mdb_hash_enc(MDB_val *val, char *encbuf)
4228 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4230 mdb_pack85(h, encbuf);
4231 mdb_pack85(h>>32, encbuf+5);
4236 /** Open and/or initialize the lock region for the environment.
4237 * @param[in] env The LMDB environment.
4238 * @param[in] lpath The pathname of the file used for the lock region.
4239 * @param[in] mode The Unix permissions for the file, if we create it.
4240 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4241 * @return 0 on success, non-zero on failure.
4244 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4247 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4249 # define MDB_ERRCODE_ROFS EROFS
4250 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4251 # define MDB_CLOEXEC O_CLOEXEC
4254 # define MDB_CLOEXEC 0
4261 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4262 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4263 FILE_ATTRIBUTE_NORMAL, NULL);
4265 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4267 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4269 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4274 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4275 /* Lose record locks when exec*() */
4276 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4277 fcntl(env->me_lfd, F_SETFD, fdflags);
4280 if (!(env->me_flags & MDB_NOTLS)) {
4281 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4284 env->me_flags |= MDB_ENV_TXKEY;
4286 /* Windows TLS callbacks need help finding their TLS info. */
4287 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4291 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4295 /* Try to get exclusive lock. If we succeed, then
4296 * nobody is using the lock region and we should initialize it.
4298 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4301 size = GetFileSize(env->me_lfd, NULL);
4303 size = lseek(env->me_lfd, 0, SEEK_END);
4304 if (size == -1) goto fail_errno;
4306 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4307 if (size < rsize && *excl > 0) {
4309 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4310 || !SetEndOfFile(env->me_lfd))
4313 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4317 size = rsize - sizeof(MDB_txninfo);
4318 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4323 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4325 if (!mh) goto fail_errno;
4326 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4328 if (!env->me_txns) goto fail_errno;
4330 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4332 if (m == MAP_FAILED) goto fail_errno;
4338 BY_HANDLE_FILE_INFORMATION stbuf;
4347 if (!mdb_sec_inited) {
4348 InitializeSecurityDescriptor(&mdb_null_sd,
4349 SECURITY_DESCRIPTOR_REVISION);
4350 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4351 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4352 mdb_all_sa.bInheritHandle = FALSE;
4353 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4356 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4357 idbuf.volume = stbuf.dwVolumeSerialNumber;
4358 idbuf.nhigh = stbuf.nFileIndexHigh;
4359 idbuf.nlow = stbuf.nFileIndexLow;
4360 val.mv_data = &idbuf;
4361 val.mv_size = sizeof(idbuf);
4362 mdb_hash_enc(&val, encbuf);
4363 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4364 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4365 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4366 if (!env->me_rmutex) goto fail_errno;
4367 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4368 if (!env->me_wmutex) goto fail_errno;
4369 #elif defined(MDB_USE_SYSV_SEM)
4371 unsigned short vals[2] = {1, 1};
4372 int semid = semget(IPC_PRIVATE, 2, mode);
4376 env->me_rmutex.semid = semid;
4377 env->me_wmutex.semid = semid;
4378 env->me_rmutex.semnum = 0;
4379 env->me_wmutex.semnum = 1;
4382 if (semctl(semid, 0, SETALL, semu) < 0)
4384 env->me_txns->mti_semid = semid;
4385 #else /* MDB_USE_SYSV_SEM */
4386 pthread_mutexattr_t mattr;
4388 if ((rc = pthread_mutexattr_init(&mattr))
4389 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4390 #ifdef MDB_ROBUST_SUPPORTED
4391 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4393 || (rc = pthread_mutex_init(&env->me_txns->mti_rmutex, &mattr))
4394 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4396 pthread_mutexattr_destroy(&mattr);
4397 #endif /* _WIN32 || MDB_USE_SYSV_SEM */
4399 env->me_txns->mti_magic = MDB_MAGIC;
4400 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4401 env->me_txns->mti_txnid = 0;
4402 env->me_txns->mti_numreaders = 0;
4405 if (env->me_txns->mti_magic != MDB_MAGIC) {
4406 DPUTS("lock region has invalid magic");
4410 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4411 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4412 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4413 rc = MDB_VERSION_MISMATCH;
4417 if (rc && rc != EACCES && rc != EAGAIN) {
4421 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4422 if (!env->me_rmutex) goto fail_errno;
4423 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4424 if (!env->me_wmutex) goto fail_errno;
4425 #elif defined(MDB_USE_SYSV_SEM)
4426 struct semid_ds buf;
4428 int semid = env->me_txns->mti_semid;
4431 /* check for read access */
4432 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4434 /* check for write access */
4435 if (semctl(semid, 0, IPC_SET, semu) < 0)
4438 env->me_rmutex.semid = semid;
4439 env->me_wmutex.semid = semid;
4440 env->me_rmutex.semnum = 0;
4441 env->me_wmutex.semnum = 1;
4452 /** The name of the lock file in the DB environment */
4453 #define LOCKNAME "/lock.mdb"
4454 /** The name of the data file in the DB environment */
4455 #define DATANAME "/data.mdb"
4456 /** The suffix of the lock file when no subdir is used */
4457 #define LOCKSUFF "-lock"
4458 /** Only a subset of the @ref mdb_env flags can be changed
4459 * at runtime. Changing other flags requires closing the
4460 * environment and re-opening it with the new flags.
4462 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4463 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4464 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4466 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4467 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4471 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4473 int oflags, rc, len, excl = -1;
4474 char *lpath, *dpath;
4476 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4480 if (flags & MDB_NOSUBDIR) {
4481 rc = len + sizeof(LOCKSUFF) + len + 1;
4483 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4488 if (flags & MDB_NOSUBDIR) {
4489 dpath = lpath + len + sizeof(LOCKSUFF);
4490 sprintf(lpath, "%s" LOCKSUFF, path);
4491 strcpy(dpath, path);
4493 dpath = lpath + len + sizeof(LOCKNAME);
4494 sprintf(lpath, "%s" LOCKNAME, path);
4495 sprintf(dpath, "%s" DATANAME, path);
4499 flags |= env->me_flags;
4500 if (flags & MDB_RDONLY) {
4501 /* silently ignore WRITEMAP when we're only getting read access */
4502 flags &= ~MDB_WRITEMAP;
4504 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4505 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4508 env->me_flags = flags |= MDB_ENV_ACTIVE;
4512 env->me_path = strdup(path);
4513 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4514 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4515 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4516 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4521 /* For RDONLY, get lockfile after we know datafile exists */
4522 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4523 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4529 if (F_ISSET(flags, MDB_RDONLY)) {
4530 oflags = GENERIC_READ;
4531 len = OPEN_EXISTING;
4533 oflags = GENERIC_READ|GENERIC_WRITE;
4536 mode = FILE_ATTRIBUTE_NORMAL;
4537 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4538 NULL, len, mode, NULL);
4540 if (F_ISSET(flags, MDB_RDONLY))
4543 oflags = O_RDWR | O_CREAT;
4545 env->me_fd = open(dpath, oflags, mode);
4547 if (env->me_fd == INVALID_HANDLE_VALUE) {
4552 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4553 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4558 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4559 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4560 env->me_mfd = env->me_fd;
4562 /* Synchronous fd for meta writes. Needed even with
4563 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4566 len = OPEN_EXISTING;
4567 env->me_mfd = CreateFile(dpath, oflags,
4568 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4569 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4572 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4574 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4579 DPRINTF(("opened dbenv %p", (void *) env));
4581 rc = mdb_env_share_locks(env, &excl);
4585 if (!((flags & MDB_RDONLY) ||
4586 (env->me_pbuf = calloc(1, env->me_psize))))
4588 if (!(flags & MDB_RDONLY)) {
4590 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4591 (sizeof(MDB_db)+sizeof(MDB_cursor)+sizeof(unsigned int)+1);
4592 txn = calloc(1, size);
4594 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4595 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4596 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4597 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4599 txn->mt_dbxs = env->me_dbxs;
4609 mdb_env_close0(env, excl);
4615 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4617 mdb_env_close0(MDB_env *env, int excl)
4621 if (!(env->me_flags & MDB_ENV_ACTIVE))
4624 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4625 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4626 free(env->me_dbxs[i].md_name.mv_data);
4629 free(env->me_dbiseqs);
4630 free(env->me_dbflags);
4633 free(env->me_dirty_list);
4635 mdb_midl_free(env->me_free_pgs);
4637 if (env->me_flags & MDB_ENV_TXKEY) {
4638 pthread_key_delete(env->me_txkey);
4640 /* Delete our key from the global list */
4641 for (i=0; i<mdb_tls_nkeys; i++)
4642 if (mdb_tls_keys[i] == env->me_txkey) {
4643 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4651 munmap(env->me_map, env->me_mapsize);
4653 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4654 (void) close(env->me_mfd);
4655 if (env->me_fd != INVALID_HANDLE_VALUE)
4656 (void) close(env->me_fd);
4658 MDB_PID_T pid = env->me_pid;
4659 /* Clearing readers is done in this function because
4660 * me_txkey with its destructor must be disabled first.
4662 for (i = env->me_numreaders; --i >= 0; )
4663 if (env->me_txns->mti_readers[i].mr_pid == pid)
4664 env->me_txns->mti_readers[i].mr_pid = 0;
4666 if (env->me_rmutex) {
4667 CloseHandle(env->me_rmutex);
4668 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4670 /* Windows automatically destroys the mutexes when
4671 * the last handle closes.
4673 #elif defined(MDB_USE_SYSV_SEM)
4674 if (env->me_rmutex.semid != -1) {
4675 /* If we have the filelock: If we are the
4676 * only remaining user, clean up semaphores.
4679 mdb_env_excl_lock(env, &excl);
4681 semctl(env->me_rmutex.semid, 0, IPC_RMID);
4684 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4686 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4689 /* Unlock the lockfile. Windows would have unlocked it
4690 * after closing anyway, but not necessarily at once.
4692 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4695 (void) close(env->me_lfd);
4698 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4702 mdb_env_close(MDB_env *env)
4709 VGMEMP_DESTROY(env);
4710 while ((dp = env->me_dpages) != NULL) {
4711 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4712 env->me_dpages = dp->mp_next;
4716 mdb_env_close0(env, 0);
4720 /** Compare two items pointing at aligned size_t's */
4722 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4724 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4725 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4728 /** Compare two items pointing at aligned unsigned int's */
4730 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4732 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4733 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4736 /** Compare two items pointing at unsigned ints of unknown alignment.
4737 * Nodes and keys are guaranteed to be 2-byte aligned.
4740 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4742 #if BYTE_ORDER == LITTLE_ENDIAN
4743 unsigned short *u, *c;
4746 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4747 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4750 } while(!x && u > (unsigned short *)a->mv_data);
4753 unsigned short *u, *c, *end;
4756 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4757 u = (unsigned short *)a->mv_data;
4758 c = (unsigned short *)b->mv_data;
4761 } while(!x && u < end);
4766 /** Compare two items pointing at size_t's of unknown alignment. */
4767 #ifdef MISALIGNED_OK
4768 # define mdb_cmp_clong mdb_cmp_long
4770 # define mdb_cmp_clong mdb_cmp_cint
4773 /** Compare two items lexically */
4775 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4782 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4788 diff = memcmp(a->mv_data, b->mv_data, len);
4789 return diff ? diff : len_diff<0 ? -1 : len_diff;
4792 /** Compare two items in reverse byte order */
4794 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4796 const unsigned char *p1, *p2, *p1_lim;
4800 p1_lim = (const unsigned char *)a->mv_data;
4801 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4802 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4804 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4810 while (p1 > p1_lim) {
4811 diff = *--p1 - *--p2;
4815 return len_diff<0 ? -1 : len_diff;
4818 /** Search for key within a page, using binary search.
4819 * Returns the smallest entry larger or equal to the key.
4820 * If exactp is non-null, stores whether the found entry was an exact match
4821 * in *exactp (1 or 0).
4822 * Updates the cursor index with the index of the found entry.
4823 * If no entry larger or equal to the key is found, returns NULL.
4826 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4828 unsigned int i = 0, nkeys;
4831 MDB_page *mp = mc->mc_pg[mc->mc_top];
4832 MDB_node *node = NULL;
4837 nkeys = NUMKEYS(mp);
4839 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4840 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4843 low = IS_LEAF(mp) ? 0 : 1;
4845 cmp = mc->mc_dbx->md_cmp;
4847 /* Branch pages have no data, so if using integer keys,
4848 * alignment is guaranteed. Use faster mdb_cmp_int.
4850 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4851 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4858 nodekey.mv_size = mc->mc_db->md_pad;
4859 node = NODEPTR(mp, 0); /* fake */
4860 while (low <= high) {
4861 i = (low + high) >> 1;
4862 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4863 rc = cmp(key, &nodekey);
4864 DPRINTF(("found leaf index %u [%s], rc = %i",
4865 i, DKEY(&nodekey), rc));
4874 while (low <= high) {
4875 i = (low + high) >> 1;
4877 node = NODEPTR(mp, i);
4878 nodekey.mv_size = NODEKSZ(node);
4879 nodekey.mv_data = NODEKEY(node);
4881 rc = cmp(key, &nodekey);
4884 DPRINTF(("found leaf index %u [%s], rc = %i",
4885 i, DKEY(&nodekey), rc));
4887 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4888 i, DKEY(&nodekey), NODEPGNO(node), rc));
4899 if (rc > 0) { /* Found entry is less than the key. */
4900 i++; /* Skip to get the smallest entry larger than key. */
4902 node = NODEPTR(mp, i);
4905 *exactp = (rc == 0 && nkeys > 0);
4906 /* store the key index */
4907 mc->mc_ki[mc->mc_top] = i;
4909 /* There is no entry larger or equal to the key. */
4912 /* nodeptr is fake for LEAF2 */
4918 mdb_cursor_adjust(MDB_cursor *mc, func)
4922 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4923 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4930 /** Pop a page off the top of the cursor's stack. */
4932 mdb_cursor_pop(MDB_cursor *mc)
4936 MDB_page *top = mc->mc_pg[mc->mc_top];
4942 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4943 DDBI(mc), (void *) mc));
4947 /** Push a page onto the top of the cursor's stack. */
4949 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4951 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4952 DDBI(mc), (void *) mc));
4954 if (mc->mc_snum >= CURSOR_STACK) {
4955 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4956 return MDB_CURSOR_FULL;
4959 mc->mc_top = mc->mc_snum++;
4960 mc->mc_pg[mc->mc_top] = mp;
4961 mc->mc_ki[mc->mc_top] = 0;
4966 /** Find the address of the page corresponding to a given page number.
4967 * @param[in] txn the transaction for this access.
4968 * @param[in] pgno the page number for the page to retrieve.
4969 * @param[out] ret address of a pointer where the page's address will be stored.
4970 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4971 * @return 0 on success, non-zero on failure.
4974 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4976 MDB_env *env = txn->mt_env;
4980 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4984 MDB_ID2L dl = tx2->mt_u.dirty_list;
4986 /* Spilled pages were dirtied in this txn and flushed
4987 * because the dirty list got full. Bring this page
4988 * back in from the map (but don't unspill it here,
4989 * leave that unless page_touch happens again).
4991 if (tx2->mt_spill_pgs) {
4992 MDB_ID pn = pgno << 1;
4993 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4994 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4995 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5000 unsigned x = mdb_mid2l_search(dl, pgno);
5001 if (x <= dl[0].mid && dl[x].mid == pgno) {
5007 } while ((tx2 = tx2->mt_parent) != NULL);
5010 if (pgno < txn->mt_next_pgno) {
5012 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5014 DPRINTF(("page %"Z"u not found", pgno));
5015 txn->mt_flags |= MDB_TXN_ERROR;
5016 return MDB_PAGE_NOTFOUND;
5026 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5027 * The cursor is at the root page, set up the rest of it.
5030 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5032 MDB_page *mp = mc->mc_pg[mc->mc_top];
5036 while (IS_BRANCH(mp)) {
5040 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5041 mdb_cassert(mc, NUMKEYS(mp) > 1);
5042 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5044 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5046 if (flags & MDB_PS_LAST)
5047 i = NUMKEYS(mp) - 1;
5050 node = mdb_node_search(mc, key, &exact);
5052 i = NUMKEYS(mp) - 1;
5054 i = mc->mc_ki[mc->mc_top];
5056 mdb_cassert(mc, i > 0);
5060 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5063 mdb_cassert(mc, i < NUMKEYS(mp));
5064 node = NODEPTR(mp, i);
5066 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5069 mc->mc_ki[mc->mc_top] = i;
5070 if ((rc = mdb_cursor_push(mc, mp)))
5073 if (flags & MDB_PS_MODIFY) {
5074 if ((rc = mdb_page_touch(mc)) != 0)
5076 mp = mc->mc_pg[mc->mc_top];
5081 DPRINTF(("internal error, index points to a %02X page!?",
5083 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5084 return MDB_CORRUPTED;
5087 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5088 key ? DKEY(key) : "null"));
5089 mc->mc_flags |= C_INITIALIZED;
5090 mc->mc_flags &= ~C_EOF;
5095 /** Search for the lowest key under the current branch page.
5096 * This just bypasses a NUMKEYS check in the current page
5097 * before calling mdb_page_search_root(), because the callers
5098 * are all in situations where the current page is known to
5102 mdb_page_search_lowest(MDB_cursor *mc)
5104 MDB_page *mp = mc->mc_pg[mc->mc_top];
5105 MDB_node *node = NODEPTR(mp, 0);
5108 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5111 mc->mc_ki[mc->mc_top] = 0;
5112 if ((rc = mdb_cursor_push(mc, mp)))
5114 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5117 /** Search for the page a given key should be in.
5118 * Push it and its parent pages on the cursor stack.
5119 * @param[in,out] mc the cursor for this operation.
5120 * @param[in] key the key to search for, or NULL for first/last page.
5121 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5122 * are touched (updated with new page numbers).
5123 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5124 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5125 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5126 * @return 0 on success, non-zero on failure.
5129 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5134 /* Make sure the txn is still viable, then find the root from
5135 * the txn's db table and set it as the root of the cursor's stack.
5137 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5138 DPUTS("transaction has failed, must abort");
5141 /* Make sure we're using an up-to-date root */
5142 if (*mc->mc_dbflag & DB_STALE) {
5144 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5146 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5147 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5154 MDB_node *leaf = mdb_node_search(&mc2,
5155 &mc->mc_dbx->md_name, &exact);
5157 return MDB_NOTFOUND;
5158 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5161 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5163 /* The txn may not know this DBI, or another process may
5164 * have dropped and recreated the DB with other flags.
5166 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5167 return MDB_INCOMPATIBLE;
5168 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5170 *mc->mc_dbflag &= ~DB_STALE;
5172 root = mc->mc_db->md_root;
5174 if (root == P_INVALID) { /* Tree is empty. */
5175 DPUTS("tree is empty");
5176 return MDB_NOTFOUND;
5180 mdb_cassert(mc, root > 1);
5181 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5182 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5188 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5189 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5191 if (flags & MDB_PS_MODIFY) {
5192 if ((rc = mdb_page_touch(mc)))
5196 if (flags & MDB_PS_ROOTONLY)
5199 return mdb_page_search_root(mc, key, flags);
5203 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5205 MDB_txn *txn = mc->mc_txn;
5206 pgno_t pg = mp->mp_pgno;
5207 unsigned x = 0, ovpages = mp->mp_pages;
5208 MDB_env *env = txn->mt_env;
5209 MDB_IDL sl = txn->mt_spill_pgs;
5210 MDB_ID pn = pg << 1;
5213 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5214 /* If the page is dirty or on the spill list we just acquired it,
5215 * so we should give it back to our current free list, if any.
5216 * Otherwise put it onto the list of pages we freed in this txn.
5218 * Won't create me_pghead: me_pglast must be inited along with it.
5219 * Unsupported in nested txns: They would need to hide the page
5220 * range in ancestor txns' dirty and spilled lists.
5222 if (env->me_pghead &&
5224 ((mp->mp_flags & P_DIRTY) ||
5225 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5229 MDB_ID2 *dl, ix, iy;
5230 rc = mdb_midl_need(&env->me_pghead, ovpages);
5233 if (!(mp->mp_flags & P_DIRTY)) {
5234 /* This page is no longer spilled */
5241 /* Remove from dirty list */
5242 dl = txn->mt_u.dirty_list;
5244 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5250 mdb_cassert(mc, x > 1);
5252 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5253 txn->mt_flags |= MDB_TXN_ERROR;
5254 return MDB_CORRUPTED;
5257 if (!(env->me_flags & MDB_WRITEMAP))
5258 mdb_dpage_free(env, mp);
5260 /* Insert in me_pghead */
5261 mop = env->me_pghead;
5262 j = mop[0] + ovpages;
5263 for (i = mop[0]; i && mop[i] < pg; i--)
5269 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5273 mc->mc_db->md_overflow_pages -= ovpages;
5277 /** Return the data associated with a given node.
5278 * @param[in] txn The transaction for this operation.
5279 * @param[in] leaf The node being read.
5280 * @param[out] data Updated to point to the node's data.
5281 * @return 0 on success, non-zero on failure.
5284 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5286 MDB_page *omp; /* overflow page */
5290 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5291 data->mv_size = NODEDSZ(leaf);
5292 data->mv_data = NODEDATA(leaf);
5296 /* Read overflow data.
5298 data->mv_size = NODEDSZ(leaf);
5299 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5300 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5301 DPRINTF(("read overflow page %"Z"u failed", pgno));
5304 data->mv_data = METADATA(omp);
5310 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5311 MDB_val *key, MDB_val *data)
5318 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5320 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5323 if (txn->mt_flags & MDB_TXN_ERROR)
5326 mdb_cursor_init(&mc, txn, dbi, &mx);
5327 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5330 /** Find a sibling for a page.
5331 * Replaces the page at the top of the cursor's stack with the
5332 * specified sibling, if one exists.
5333 * @param[in] mc The cursor for this operation.
5334 * @param[in] move_right Non-zero if the right sibling is requested,
5335 * otherwise the left sibling.
5336 * @return 0 on success, non-zero on failure.
5339 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5345 if (mc->mc_snum < 2) {
5346 return MDB_NOTFOUND; /* root has no siblings */
5350 DPRINTF(("parent page is page %"Z"u, index %u",
5351 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5353 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5354 : (mc->mc_ki[mc->mc_top] == 0)) {
5355 DPRINTF(("no more keys left, moving to %s sibling",
5356 move_right ? "right" : "left"));
5357 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5358 /* undo cursor_pop before returning */
5365 mc->mc_ki[mc->mc_top]++;
5367 mc->mc_ki[mc->mc_top]--;
5368 DPRINTF(("just moving to %s index key %u",
5369 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5371 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5373 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5374 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5375 /* mc will be inconsistent if caller does mc_snum++ as above */
5376 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5380 mdb_cursor_push(mc, mp);
5382 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5387 /** Move the cursor to the next data item. */
5389 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5395 if (mc->mc_flags & C_EOF) {
5396 return MDB_NOTFOUND;
5399 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5401 mp = mc->mc_pg[mc->mc_top];
5403 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5404 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5405 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5406 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5407 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5408 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5409 if (rc == MDB_SUCCESS)
5410 MDB_GET_KEY(leaf, key);
5415 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5416 if (op == MDB_NEXT_DUP)
5417 return MDB_NOTFOUND;
5421 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5422 mdb_dbg_pgno(mp), (void *) mc));
5423 if (mc->mc_flags & C_DEL)
5426 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5427 DPUTS("=====> move to next sibling page");
5428 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5429 mc->mc_flags |= C_EOF;
5432 mp = mc->mc_pg[mc->mc_top];
5433 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5435 mc->mc_ki[mc->mc_top]++;
5438 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5439 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5442 key->mv_size = mc->mc_db->md_pad;
5443 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5447 mdb_cassert(mc, IS_LEAF(mp));
5448 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5450 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5451 mdb_xcursor_init1(mc, leaf);
5454 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5457 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5458 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5459 if (rc != MDB_SUCCESS)
5464 MDB_GET_KEY(leaf, key);
5468 /** Move the cursor to the previous data item. */
5470 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5476 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5478 mp = mc->mc_pg[mc->mc_top];
5480 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5481 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5482 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5483 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5484 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5485 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5486 if (rc == MDB_SUCCESS) {
5487 MDB_GET_KEY(leaf, key);
5488 mc->mc_flags &= ~C_EOF;
5494 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5495 if (op == MDB_PREV_DUP)
5496 return MDB_NOTFOUND;
5500 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5501 mdb_dbg_pgno(mp), (void *) mc));
5503 if (mc->mc_ki[mc->mc_top] == 0) {
5504 DPUTS("=====> move to prev sibling page");
5505 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5508 mp = mc->mc_pg[mc->mc_top];
5509 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5510 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5512 mc->mc_ki[mc->mc_top]--;
5514 mc->mc_flags &= ~C_EOF;
5516 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5517 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5520 key->mv_size = mc->mc_db->md_pad;
5521 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5525 mdb_cassert(mc, IS_LEAF(mp));
5526 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5528 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5529 mdb_xcursor_init1(mc, leaf);
5532 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5535 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5536 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5537 if (rc != MDB_SUCCESS)
5542 MDB_GET_KEY(leaf, key);
5546 /** Set the cursor on a specific data item. */
5548 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5549 MDB_cursor_op op, int *exactp)
5553 MDB_node *leaf = NULL;
5556 if (key->mv_size == 0)
5557 return MDB_BAD_VALSIZE;
5560 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5562 /* See if we're already on the right page */
5563 if (mc->mc_flags & C_INITIALIZED) {
5566 mp = mc->mc_pg[mc->mc_top];
5568 mc->mc_ki[mc->mc_top] = 0;
5569 return MDB_NOTFOUND;
5571 if (mp->mp_flags & P_LEAF2) {
5572 nodekey.mv_size = mc->mc_db->md_pad;
5573 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5575 leaf = NODEPTR(mp, 0);
5576 MDB_GET_KEY2(leaf, nodekey);
5578 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5580 /* Probably happens rarely, but first node on the page
5581 * was the one we wanted.
5583 mc->mc_ki[mc->mc_top] = 0;
5590 unsigned int nkeys = NUMKEYS(mp);
5592 if (mp->mp_flags & P_LEAF2) {
5593 nodekey.mv_data = LEAF2KEY(mp,
5594 nkeys-1, nodekey.mv_size);
5596 leaf = NODEPTR(mp, nkeys-1);
5597 MDB_GET_KEY2(leaf, nodekey);
5599 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5601 /* last node was the one we wanted */
5602 mc->mc_ki[mc->mc_top] = nkeys-1;
5608 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5609 /* This is definitely the right page, skip search_page */
5610 if (mp->mp_flags & P_LEAF2) {
5611 nodekey.mv_data = LEAF2KEY(mp,
5612 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5614 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5615 MDB_GET_KEY2(leaf, nodekey);
5617 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5619 /* current node was the one we wanted */
5629 /* If any parents have right-sibs, search.
5630 * Otherwise, there's nothing further.
5632 for (i=0; i<mc->mc_top; i++)
5634 NUMKEYS(mc->mc_pg[i])-1)
5636 if (i == mc->mc_top) {
5637 /* There are no other pages */
5638 mc->mc_ki[mc->mc_top] = nkeys;
5639 return MDB_NOTFOUND;
5643 /* There are no other pages */
5644 mc->mc_ki[mc->mc_top] = 0;
5645 if (op == MDB_SET_RANGE && !exactp) {
5649 return MDB_NOTFOUND;
5653 rc = mdb_page_search(mc, key, 0);
5654 if (rc != MDB_SUCCESS)
5657 mp = mc->mc_pg[mc->mc_top];
5658 mdb_cassert(mc, IS_LEAF(mp));
5661 leaf = mdb_node_search(mc, key, exactp);
5662 if (exactp != NULL && !*exactp) {
5663 /* MDB_SET specified and not an exact match. */
5664 return MDB_NOTFOUND;
5668 DPUTS("===> inexact leaf not found, goto sibling");
5669 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5670 return rc; /* no entries matched */
5671 mp = mc->mc_pg[mc->mc_top];
5672 mdb_cassert(mc, IS_LEAF(mp));
5673 leaf = NODEPTR(mp, 0);
5677 mc->mc_flags |= C_INITIALIZED;
5678 mc->mc_flags &= ~C_EOF;
5681 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5682 key->mv_size = mc->mc_db->md_pad;
5683 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5688 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5689 mdb_xcursor_init1(mc, leaf);
5692 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5693 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5694 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5697 if (op == MDB_GET_BOTH) {
5703 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5704 if (rc != MDB_SUCCESS)
5707 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5709 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5711 rc = mc->mc_dbx->md_dcmp(data, &d2);
5713 if (op == MDB_GET_BOTH || rc > 0)
5714 return MDB_NOTFOUND;
5721 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5722 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5727 /* The key already matches in all other cases */
5728 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5729 MDB_GET_KEY(leaf, key);
5730 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5735 /** Move the cursor to the first item in the database. */
5737 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5743 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5745 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5746 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5747 if (rc != MDB_SUCCESS)
5750 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5752 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5753 mc->mc_flags |= C_INITIALIZED;
5754 mc->mc_flags &= ~C_EOF;
5756 mc->mc_ki[mc->mc_top] = 0;
5758 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5759 key->mv_size = mc->mc_db->md_pad;
5760 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5765 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5766 mdb_xcursor_init1(mc, leaf);
5767 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5771 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5775 MDB_GET_KEY(leaf, key);
5779 /** Move the cursor to the last item in the database. */
5781 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5787 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5789 if (!(mc->mc_flags & C_EOF)) {
5791 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5792 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5793 if (rc != MDB_SUCCESS)
5796 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5799 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5800 mc->mc_flags |= C_INITIALIZED|C_EOF;
5801 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5803 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5804 key->mv_size = mc->mc_db->md_pad;
5805 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5810 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5811 mdb_xcursor_init1(mc, leaf);
5812 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5816 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5821 MDB_GET_KEY(leaf, key);
5826 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5831 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5836 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5840 case MDB_GET_CURRENT:
5841 if (!(mc->mc_flags & C_INITIALIZED)) {
5844 MDB_page *mp = mc->mc_pg[mc->mc_top];
5845 int nkeys = NUMKEYS(mp);
5846 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5847 mc->mc_ki[mc->mc_top] = nkeys;
5853 key->mv_size = mc->mc_db->md_pad;
5854 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5856 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5857 MDB_GET_KEY(leaf, key);
5859 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5860 if (mc->mc_flags & C_DEL)
5861 mdb_xcursor_init1(mc, leaf);
5862 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5864 rc = mdb_node_read(mc->mc_txn, leaf, data);
5871 case MDB_GET_BOTH_RANGE:
5876 if (mc->mc_xcursor == NULL) {
5877 rc = MDB_INCOMPATIBLE;
5887 rc = mdb_cursor_set(mc, key, data, op,
5888 op == MDB_SET_RANGE ? NULL : &exact);
5891 case MDB_GET_MULTIPLE:
5892 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5896 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5897 rc = MDB_INCOMPATIBLE;
5901 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5902 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5905 case MDB_NEXT_MULTIPLE:
5910 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5911 rc = MDB_INCOMPATIBLE;
5914 if (!(mc->mc_flags & C_INITIALIZED))
5915 rc = mdb_cursor_first(mc, key, data);
5917 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5918 if (rc == MDB_SUCCESS) {
5919 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5922 mx = &mc->mc_xcursor->mx_cursor;
5923 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5925 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5926 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5934 case MDB_NEXT_NODUP:
5935 if (!(mc->mc_flags & C_INITIALIZED))
5936 rc = mdb_cursor_first(mc, key, data);
5938 rc = mdb_cursor_next(mc, key, data, op);
5942 case MDB_PREV_NODUP:
5943 if (!(mc->mc_flags & C_INITIALIZED)) {
5944 rc = mdb_cursor_last(mc, key, data);
5947 mc->mc_flags |= C_INITIALIZED;
5948 mc->mc_ki[mc->mc_top]++;
5950 rc = mdb_cursor_prev(mc, key, data, op);
5953 rc = mdb_cursor_first(mc, key, data);
5956 mfunc = mdb_cursor_first;
5958 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5962 if (mc->mc_xcursor == NULL) {
5963 rc = MDB_INCOMPATIBLE;
5967 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5968 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5969 MDB_GET_KEY(leaf, key);
5970 rc = mdb_node_read(mc->mc_txn, leaf, data);
5974 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5978 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5981 rc = mdb_cursor_last(mc, key, data);
5984 mfunc = mdb_cursor_last;
5987 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5992 if (mc->mc_flags & C_DEL)
5993 mc->mc_flags ^= C_DEL;
5998 /** Touch all the pages in the cursor stack. Set mc_top.
5999 * Makes sure all the pages are writable, before attempting a write operation.
6000 * @param[in] mc The cursor to operate on.
6003 mdb_cursor_touch(MDB_cursor *mc)
6005 int rc = MDB_SUCCESS;
6007 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6010 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6012 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6013 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6016 *mc->mc_dbflag |= DB_DIRTY;
6021 rc = mdb_page_touch(mc);
6022 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6023 mc->mc_top = mc->mc_snum-1;
6028 /** Do not spill pages to disk if txn is getting full, may fail instead */
6029 #define MDB_NOSPILL 0x8000
6032 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6035 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
6037 MDB_node *leaf = NULL;
6040 MDB_val xdata, *rdata, dkey, olddata;
6042 int do_sub = 0, insert_key, insert_data;
6043 unsigned int mcount = 0, dcount = 0, nospill;
6046 unsigned int nflags;
6049 if (mc == NULL || key == NULL)
6052 env = mc->mc_txn->mt_env;
6054 /* Check this first so counter will always be zero on any
6057 if (flags & MDB_MULTIPLE) {
6058 dcount = data[1].mv_size;
6059 data[1].mv_size = 0;
6060 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6061 return MDB_INCOMPATIBLE;
6064 nospill = flags & MDB_NOSPILL;
6065 flags &= ~MDB_NOSPILL;
6067 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6068 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6070 if (key->mv_size-1 >= ENV_MAXKEY(env))
6071 return MDB_BAD_VALSIZE;
6073 #if SIZE_MAX > MAXDATASIZE
6074 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6075 return MDB_BAD_VALSIZE;
6077 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6078 return MDB_BAD_VALSIZE;
6081 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6082 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6086 if (flags == MDB_CURRENT) {
6087 if (!(mc->mc_flags & C_INITIALIZED))
6090 } else if (mc->mc_db->md_root == P_INVALID) {
6091 /* new database, cursor has nothing to point to */
6094 mc->mc_flags &= ~C_INITIALIZED;
6099 if (flags & MDB_APPEND) {
6101 rc = mdb_cursor_last(mc, &k2, &d2);
6103 rc = mc->mc_dbx->md_cmp(key, &k2);
6106 mc->mc_ki[mc->mc_top]++;
6108 /* new key is <= last key */
6113 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6115 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6116 DPRINTF(("duplicate key [%s]", DKEY(key)));
6118 return MDB_KEYEXIST;
6120 if (rc && rc != MDB_NOTFOUND)
6124 if (mc->mc_flags & C_DEL)
6125 mc->mc_flags ^= C_DEL;
6127 /* Cursor is positioned, check for room in the dirty list */
6129 if (flags & MDB_MULTIPLE) {
6131 xdata.mv_size = data->mv_size * dcount;
6135 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6139 if (rc == MDB_NO_ROOT) {
6141 /* new database, write a root leaf page */
6142 DPUTS("allocating new root leaf page");
6143 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6146 mdb_cursor_push(mc, np);
6147 mc->mc_db->md_root = np->mp_pgno;
6148 mc->mc_db->md_depth++;
6149 *mc->mc_dbflag |= DB_DIRTY;
6150 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6152 np->mp_flags |= P_LEAF2;
6153 mc->mc_flags |= C_INITIALIZED;
6155 /* make sure all cursor pages are writable */
6156 rc2 = mdb_cursor_touch(mc);
6161 insert_key = insert_data = rc;
6163 /* The key does not exist */
6164 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6165 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6166 LEAFSIZE(key, data) > env->me_nodemax)
6168 /* Too big for a node, insert in sub-DB. Set up an empty
6169 * "old sub-page" for prep_subDB to expand to a full page.
6171 fp_flags = P_LEAF|P_DIRTY;
6173 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6174 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6175 olddata.mv_size = PAGEHDRSZ;
6179 /* there's only a key anyway, so this is a no-op */
6180 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6182 unsigned int ksize = mc->mc_db->md_pad;
6183 if (key->mv_size != ksize)
6184 return MDB_BAD_VALSIZE;
6185 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6186 memcpy(ptr, key->mv_data, ksize);
6188 /* if overwriting slot 0 of leaf, need to
6189 * update branch key if there is a parent page
6191 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6192 unsigned short top = mc->mc_top;
6194 /* slot 0 is always an empty key, find real slot */
6195 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6197 if (mc->mc_ki[mc->mc_top])
6198 rc2 = mdb_update_key(mc, key);
6209 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6210 olddata.mv_size = NODEDSZ(leaf);
6211 olddata.mv_data = NODEDATA(leaf);
6214 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6215 /* Prepare (sub-)page/sub-DB to accept the new item,
6216 * if needed. fp: old sub-page or a header faking
6217 * it. mp: new (sub-)page. offset: growth in page
6218 * size. xdata: node data with new page or DB.
6220 unsigned i, offset = 0;
6221 mp = fp = xdata.mv_data = env->me_pbuf;
6222 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6224 /* Was a single item before, must convert now */
6225 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6226 /* Just overwrite the current item */
6227 if (flags == MDB_CURRENT)
6230 #if UINT_MAX < SIZE_MAX
6231 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6232 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6234 /* does data match? */
6235 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6236 if (flags & MDB_NODUPDATA)
6237 return MDB_KEYEXIST;
6242 /* Back up original data item */
6243 dkey.mv_size = olddata.mv_size;
6244 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6246 /* Make sub-page header for the dup items, with dummy body */
6247 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6248 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6249 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6250 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6251 fp->mp_flags |= P_LEAF2;
6252 fp->mp_pad = data->mv_size;
6253 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6255 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6256 (dkey.mv_size & 1) + (data->mv_size & 1);
6258 fp->mp_upper = xdata.mv_size - PAGEBASE;
6259 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6260 } else if (leaf->mn_flags & F_SUBDATA) {
6261 /* Data is on sub-DB, just store it */
6262 flags |= F_DUPDATA|F_SUBDATA;
6265 /* Data is on sub-page */
6266 fp = olddata.mv_data;
6269 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6270 offset = EVEN(NODESIZE + sizeof(indx_t) +
6274 offset = fp->mp_pad;
6275 if (SIZELEFT(fp) < offset) {
6276 offset *= 4; /* space for 4 more */
6279 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6281 fp->mp_flags |= P_DIRTY;
6282 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6283 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6287 xdata.mv_size = olddata.mv_size + offset;
6290 fp_flags = fp->mp_flags;
6291 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6292 /* Too big for a sub-page, convert to sub-DB */
6293 fp_flags &= ~P_SUBP;
6295 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6296 fp_flags |= P_LEAF2;
6297 dummy.md_pad = fp->mp_pad;
6298 dummy.md_flags = MDB_DUPFIXED;
6299 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6300 dummy.md_flags |= MDB_INTEGERKEY;
6306 dummy.md_branch_pages = 0;
6307 dummy.md_leaf_pages = 1;
6308 dummy.md_overflow_pages = 0;
6309 dummy.md_entries = NUMKEYS(fp);
6310 xdata.mv_size = sizeof(MDB_db);
6311 xdata.mv_data = &dummy;
6312 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6314 offset = env->me_psize - olddata.mv_size;
6315 flags |= F_DUPDATA|F_SUBDATA;
6316 dummy.md_root = mp->mp_pgno;
6319 mp->mp_flags = fp_flags | P_DIRTY;
6320 mp->mp_pad = fp->mp_pad;
6321 mp->mp_lower = fp->mp_lower;
6322 mp->mp_upper = fp->mp_upper + offset;
6323 if (fp_flags & P_LEAF2) {
6324 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6326 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6327 olddata.mv_size - fp->mp_upper - PAGEBASE);
6328 for (i=0; i<NUMKEYS(fp); i++)
6329 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6337 mdb_node_del(mc, 0);
6341 /* overflow page overwrites need special handling */
6342 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6345 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6347 memcpy(&pg, olddata.mv_data, sizeof(pg));
6348 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6350 ovpages = omp->mp_pages;
6352 /* Is the ov page large enough? */
6353 if (ovpages >= dpages) {
6354 if (!(omp->mp_flags & P_DIRTY) &&
6355 (level || (env->me_flags & MDB_WRITEMAP)))
6357 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6360 level = 0; /* dirty in this txn or clean */
6363 if (omp->mp_flags & P_DIRTY) {
6364 /* yes, overwrite it. Note in this case we don't
6365 * bother to try shrinking the page if the new data
6366 * is smaller than the overflow threshold.
6369 /* It is writable only in a parent txn */
6370 size_t sz = (size_t) env->me_psize * ovpages, off;
6371 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6377 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6378 mdb_cassert(mc, rc2 == 0);
6379 if (!(flags & MDB_RESERVE)) {
6380 /* Copy end of page, adjusting alignment so
6381 * compiler may copy words instead of bytes.
6383 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6384 memcpy((size_t *)((char *)np + off),
6385 (size_t *)((char *)omp + off), sz - off);
6388 memcpy(np, omp, sz); /* Copy beginning of page */
6391 SETDSZ(leaf, data->mv_size);
6392 if (F_ISSET(flags, MDB_RESERVE))
6393 data->mv_data = METADATA(omp);
6395 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6399 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6401 } else if (data->mv_size == olddata.mv_size) {
6402 /* same size, just replace it. Note that we could
6403 * also reuse this node if the new data is smaller,
6404 * but instead we opt to shrink the node in that case.
6406 if (F_ISSET(flags, MDB_RESERVE))
6407 data->mv_data = olddata.mv_data;
6408 else if (!(mc->mc_flags & C_SUB))
6409 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6411 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6416 mdb_node_del(mc, 0);
6422 nflags = flags & NODE_ADD_FLAGS;
6423 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6424 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6425 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6426 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6428 nflags |= MDB_SPLIT_REPLACE;
6429 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6431 /* There is room already in this leaf page. */
6432 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6433 if (rc == 0 && insert_key) {
6434 /* Adjust other cursors pointing to mp */
6435 MDB_cursor *m2, *m3;
6436 MDB_dbi dbi = mc->mc_dbi;
6437 unsigned i = mc->mc_top;
6438 MDB_page *mp = mc->mc_pg[i];
6440 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6441 if (mc->mc_flags & C_SUB)
6442 m3 = &m2->mc_xcursor->mx_cursor;
6445 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6446 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6453 if (rc == MDB_SUCCESS) {
6454 /* Now store the actual data in the child DB. Note that we're
6455 * storing the user data in the keys field, so there are strict
6456 * size limits on dupdata. The actual data fields of the child
6457 * DB are all zero size.
6465 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6466 if (flags & MDB_CURRENT) {
6467 xflags = MDB_CURRENT|MDB_NOSPILL;
6469 mdb_xcursor_init1(mc, leaf);
6470 xflags = (flags & MDB_NODUPDATA) ?
6471 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6473 /* converted, write the original data first */
6475 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6479 /* Adjust other cursors pointing to mp */
6481 unsigned i = mc->mc_top;
6482 MDB_page *mp = mc->mc_pg[i];
6484 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6485 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6486 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6487 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6488 mdb_xcursor_init1(m2, leaf);
6492 /* we've done our job */
6495 ecount = mc->mc_xcursor->mx_db.md_entries;
6496 if (flags & MDB_APPENDDUP)
6497 xflags |= MDB_APPEND;
6498 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6499 if (flags & F_SUBDATA) {
6500 void *db = NODEDATA(leaf);
6501 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6503 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6505 /* Increment count unless we just replaced an existing item. */
6507 mc->mc_db->md_entries++;
6509 /* Invalidate txn if we created an empty sub-DB */
6512 /* If we succeeded and the key didn't exist before,
6513 * make sure the cursor is marked valid.
6515 mc->mc_flags |= C_INITIALIZED;
6517 if (flags & MDB_MULTIPLE) {
6520 /* let caller know how many succeeded, if any */
6521 data[1].mv_size = mcount;
6522 if (mcount < dcount) {
6523 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6524 insert_key = insert_data = 0;
6531 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6534 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6539 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6545 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6546 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6548 if (!(mc->mc_flags & C_INITIALIZED))
6551 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6552 return MDB_NOTFOUND;
6554 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6557 rc = mdb_cursor_touch(mc);
6561 mp = mc->mc_pg[mc->mc_top];
6564 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6566 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6567 if (flags & MDB_NODUPDATA) {
6568 /* mdb_cursor_del0() will subtract the final entry */
6569 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6571 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6572 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6574 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6577 /* If sub-DB still has entries, we're done */
6578 if (mc->mc_xcursor->mx_db.md_entries) {
6579 if (leaf->mn_flags & F_SUBDATA) {
6580 /* update subDB info */
6581 void *db = NODEDATA(leaf);
6582 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6585 /* shrink fake page */
6586 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6587 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6588 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6589 /* fix other sub-DB cursors pointed at this fake page */
6590 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6591 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6592 if (m2->mc_pg[mc->mc_top] == mp &&
6593 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6594 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6597 mc->mc_db->md_entries--;
6598 mc->mc_flags |= C_DEL;
6601 /* otherwise fall thru and delete the sub-DB */
6604 if (leaf->mn_flags & F_SUBDATA) {
6605 /* add all the child DB's pages to the free list */
6606 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6612 /* add overflow pages to free list */
6613 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6617 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6618 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6619 (rc = mdb_ovpage_free(mc, omp)))
6624 return mdb_cursor_del0(mc);
6627 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6631 /** Allocate and initialize new pages for a database.
6632 * @param[in] mc a cursor on the database being added to.
6633 * @param[in] flags flags defining what type of page is being allocated.
6634 * @param[in] num the number of pages to allocate. This is usually 1,
6635 * unless allocating overflow pages for a large record.
6636 * @param[out] mp Address of a page, or NULL on failure.
6637 * @return 0 on success, non-zero on failure.
6640 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6645 if ((rc = mdb_page_alloc(mc, num, &np)))
6647 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6648 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6649 np->mp_flags = flags | P_DIRTY;
6650 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6651 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6654 mc->mc_db->md_branch_pages++;
6655 else if (IS_LEAF(np))
6656 mc->mc_db->md_leaf_pages++;
6657 else if (IS_OVERFLOW(np)) {
6658 mc->mc_db->md_overflow_pages += num;
6666 /** Calculate the size of a leaf node.
6667 * The size depends on the environment's page size; if a data item
6668 * is too large it will be put onto an overflow page and the node
6669 * size will only include the key and not the data. Sizes are always
6670 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6671 * of the #MDB_node headers.
6672 * @param[in] env The environment handle.
6673 * @param[in] key The key for the node.
6674 * @param[in] data The data for the node.
6675 * @return The number of bytes needed to store the node.
6678 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6682 sz = LEAFSIZE(key, data);
6683 if (sz > env->me_nodemax) {
6684 /* put on overflow page */
6685 sz -= data->mv_size - sizeof(pgno_t);
6688 return EVEN(sz + sizeof(indx_t));
6691 /** Calculate the size of a branch node.
6692 * The size should depend on the environment's page size but since
6693 * we currently don't support spilling large keys onto overflow
6694 * pages, it's simply the size of the #MDB_node header plus the
6695 * size of the key. Sizes are always rounded up to an even number
6696 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6697 * @param[in] env The environment handle.
6698 * @param[in] key The key for the node.
6699 * @return The number of bytes needed to store the node.
6702 mdb_branch_size(MDB_env *env, MDB_val *key)
6707 if (sz > env->me_nodemax) {
6708 /* put on overflow page */
6709 /* not implemented */
6710 /* sz -= key->size - sizeof(pgno_t); */
6713 return sz + sizeof(indx_t);
6716 /** Add a node to the page pointed to by the cursor.
6717 * @param[in] mc The cursor for this operation.
6718 * @param[in] indx The index on the page where the new node should be added.
6719 * @param[in] key The key for the new node.
6720 * @param[in] data The data for the new node, if any.
6721 * @param[in] pgno The page number, if adding a branch node.
6722 * @param[in] flags Flags for the node.
6723 * @return 0 on success, non-zero on failure. Possible errors are:
6725 * <li>ENOMEM - failed to allocate overflow pages for the node.
6726 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6727 * should never happen since all callers already calculate the
6728 * page's free space before calling this function.
6732 mdb_node_add(MDB_cursor *mc, indx_t indx,
6733 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6736 size_t node_size = NODESIZE;
6740 MDB_page *mp = mc->mc_pg[mc->mc_top];
6741 MDB_page *ofp = NULL; /* overflow page */
6744 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6746 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6747 IS_LEAF(mp) ? "leaf" : "branch",
6748 IS_SUBP(mp) ? "sub-" : "",
6749 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6750 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6753 /* Move higher keys up one slot. */
6754 int ksize = mc->mc_db->md_pad, dif;
6755 char *ptr = LEAF2KEY(mp, indx, ksize);
6756 dif = NUMKEYS(mp) - indx;
6758 memmove(ptr+ksize, ptr, dif*ksize);
6759 /* insert new key */
6760 memcpy(ptr, key->mv_data, ksize);
6762 /* Just using these for counting */
6763 mp->mp_lower += sizeof(indx_t);
6764 mp->mp_upper -= ksize - sizeof(indx_t);
6768 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6770 node_size += key->mv_size;
6772 mdb_cassert(mc, data);
6773 if (F_ISSET(flags, F_BIGDATA)) {
6774 /* Data already on overflow page. */
6775 node_size += sizeof(pgno_t);
6776 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6777 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6779 /* Put data on overflow page. */
6780 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6781 data->mv_size, node_size+data->mv_size));
6782 node_size = EVEN(node_size + sizeof(pgno_t));
6783 if ((ssize_t)node_size > room)
6785 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6787 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6791 node_size += data->mv_size;
6794 node_size = EVEN(node_size);
6795 if ((ssize_t)node_size > room)
6799 /* Move higher pointers up one slot. */
6800 for (i = NUMKEYS(mp); i > indx; i--)
6801 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6803 /* Adjust free space offsets. */
6804 ofs = mp->mp_upper - node_size;
6805 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6806 mp->mp_ptrs[indx] = ofs;
6808 mp->mp_lower += sizeof(indx_t);
6810 /* Write the node data. */
6811 node = NODEPTR(mp, indx);
6812 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6813 node->mn_flags = flags;
6815 SETDSZ(node,data->mv_size);
6820 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6823 mdb_cassert(mc, key);
6825 if (F_ISSET(flags, F_BIGDATA))
6826 memcpy(node->mn_data + key->mv_size, data->mv_data,
6828 else if (F_ISSET(flags, MDB_RESERVE))
6829 data->mv_data = node->mn_data + key->mv_size;
6831 memcpy(node->mn_data + key->mv_size, data->mv_data,
6834 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6836 if (F_ISSET(flags, MDB_RESERVE))
6837 data->mv_data = METADATA(ofp);
6839 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6846 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6847 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6848 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6849 DPRINTF(("node size = %"Z"u", node_size));
6850 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6851 return MDB_PAGE_FULL;
6854 /** Delete the specified node from a page.
6855 * @param[in] mc Cursor pointing to the node to delete.
6856 * @param[in] ksize The size of a node. Only used if the page is
6857 * part of a #MDB_DUPFIXED database.
6860 mdb_node_del(MDB_cursor *mc, int ksize)
6862 MDB_page *mp = mc->mc_pg[mc->mc_top];
6863 indx_t indx = mc->mc_ki[mc->mc_top];
6865 indx_t i, j, numkeys, ptr;
6869 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6870 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6871 numkeys = NUMKEYS(mp);
6872 mdb_cassert(mc, indx < numkeys);
6875 int x = numkeys - 1 - indx;
6876 base = LEAF2KEY(mp, indx, ksize);
6878 memmove(base, base + ksize, x * ksize);
6879 mp->mp_lower -= sizeof(indx_t);
6880 mp->mp_upper += ksize - sizeof(indx_t);
6884 node = NODEPTR(mp, indx);
6885 sz = NODESIZE + node->mn_ksize;
6887 if (F_ISSET(node->mn_flags, F_BIGDATA))
6888 sz += sizeof(pgno_t);
6890 sz += NODEDSZ(node);
6894 ptr = mp->mp_ptrs[indx];
6895 for (i = j = 0; i < numkeys; i++) {
6897 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6898 if (mp->mp_ptrs[i] < ptr)
6899 mp->mp_ptrs[j] += sz;
6904 base = (char *)mp + mp->mp_upper + PAGEBASE;
6905 memmove(base + sz, base, ptr - mp->mp_upper);
6907 mp->mp_lower -= sizeof(indx_t);
6911 /** Compact the main page after deleting a node on a subpage.
6912 * @param[in] mp The main page to operate on.
6913 * @param[in] indx The index of the subpage on the main page.
6916 mdb_node_shrink(MDB_page *mp, indx_t indx)
6922 indx_t i, numkeys, ptr;
6924 node = NODEPTR(mp, indx);
6925 sp = (MDB_page *)NODEDATA(node);
6926 delta = SIZELEFT(sp);
6927 xp = (MDB_page *)((char *)sp + delta);
6929 /* shift subpage upward */
6931 nsize = NUMKEYS(sp) * sp->mp_pad;
6933 return; /* do not make the node uneven-sized */
6934 memmove(METADATA(xp), METADATA(sp), nsize);
6937 numkeys = NUMKEYS(sp);
6938 for (i=numkeys-1; i>=0; i--)
6939 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6941 xp->mp_upper = sp->mp_lower;
6942 xp->mp_lower = sp->mp_lower;
6943 xp->mp_flags = sp->mp_flags;
6944 xp->mp_pad = sp->mp_pad;
6945 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6947 nsize = NODEDSZ(node) - delta;
6948 SETDSZ(node, nsize);
6950 /* shift lower nodes upward */
6951 ptr = mp->mp_ptrs[indx];
6952 numkeys = NUMKEYS(mp);
6953 for (i = 0; i < numkeys; i++) {
6954 if (mp->mp_ptrs[i] <= ptr)
6955 mp->mp_ptrs[i] += delta;
6958 base = (char *)mp + mp->mp_upper + PAGEBASE;
6959 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6960 mp->mp_upper += delta;
6963 /** Initial setup of a sorted-dups cursor.
6964 * Sorted duplicates are implemented as a sub-database for the given key.
6965 * The duplicate data items are actually keys of the sub-database.
6966 * Operations on the duplicate data items are performed using a sub-cursor
6967 * initialized when the sub-database is first accessed. This function does
6968 * the preliminary setup of the sub-cursor, filling in the fields that
6969 * depend only on the parent DB.
6970 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6973 mdb_xcursor_init0(MDB_cursor *mc)
6975 MDB_xcursor *mx = mc->mc_xcursor;
6977 mx->mx_cursor.mc_xcursor = NULL;
6978 mx->mx_cursor.mc_txn = mc->mc_txn;
6979 mx->mx_cursor.mc_db = &mx->mx_db;
6980 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6981 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6982 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6983 mx->mx_cursor.mc_snum = 0;
6984 mx->mx_cursor.mc_top = 0;
6985 mx->mx_cursor.mc_flags = C_SUB;
6986 mx->mx_dbx.md_name.mv_size = 0;
6987 mx->mx_dbx.md_name.mv_data = NULL;
6988 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6989 mx->mx_dbx.md_dcmp = NULL;
6990 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6993 /** Final setup of a sorted-dups cursor.
6994 * Sets up the fields that depend on the data from the main cursor.
6995 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6996 * @param[in] node The data containing the #MDB_db record for the
6997 * sorted-dup database.
7000 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7002 MDB_xcursor *mx = mc->mc_xcursor;
7004 if (node->mn_flags & F_SUBDATA) {
7005 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7006 mx->mx_cursor.mc_pg[0] = 0;
7007 mx->mx_cursor.mc_snum = 0;
7008 mx->mx_cursor.mc_top = 0;
7009 mx->mx_cursor.mc_flags = C_SUB;
7011 MDB_page *fp = NODEDATA(node);
7012 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7013 mx->mx_db.md_flags = 0;
7014 mx->mx_db.md_depth = 1;
7015 mx->mx_db.md_branch_pages = 0;
7016 mx->mx_db.md_leaf_pages = 1;
7017 mx->mx_db.md_overflow_pages = 0;
7018 mx->mx_db.md_entries = NUMKEYS(fp);
7019 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7020 mx->mx_cursor.mc_snum = 1;
7021 mx->mx_cursor.mc_top = 0;
7022 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7023 mx->mx_cursor.mc_pg[0] = fp;
7024 mx->mx_cursor.mc_ki[0] = 0;
7025 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7026 mx->mx_db.md_flags = MDB_DUPFIXED;
7027 mx->mx_db.md_pad = fp->mp_pad;
7028 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7029 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7032 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7033 mx->mx_db.md_root));
7034 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7035 #if UINT_MAX < SIZE_MAX
7036 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7037 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7041 /** Initialize a cursor for a given transaction and database. */
7043 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7046 mc->mc_backup = NULL;
7049 mc->mc_db = &txn->mt_dbs[dbi];
7050 mc->mc_dbx = &txn->mt_dbxs[dbi];
7051 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7056 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7057 mdb_tassert(txn, mx != NULL);
7058 mc->mc_xcursor = mx;
7059 mdb_xcursor_init0(mc);
7061 mc->mc_xcursor = NULL;
7063 if (*mc->mc_dbflag & DB_STALE) {
7064 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7069 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7072 size_t size = sizeof(MDB_cursor);
7074 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7077 if (txn->mt_flags & MDB_TXN_ERROR)
7080 /* Allow read access to the freelist */
7081 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7084 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7085 size += sizeof(MDB_xcursor);
7087 if ((mc = malloc(size)) != NULL) {
7088 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7089 if (txn->mt_cursors) {
7090 mc->mc_next = txn->mt_cursors[dbi];
7091 txn->mt_cursors[dbi] = mc;
7092 mc->mc_flags |= C_UNTRACK;
7104 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7106 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7109 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7112 if (txn->mt_flags & MDB_TXN_ERROR)
7115 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7119 /* Return the count of duplicate data items for the current key */
7121 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7125 if (mc == NULL || countp == NULL)
7128 if (mc->mc_xcursor == NULL)
7129 return MDB_INCOMPATIBLE;
7131 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7134 if (!(mc->mc_flags & C_INITIALIZED))
7137 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7138 return MDB_NOTFOUND;
7140 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7141 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7144 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7147 *countp = mc->mc_xcursor->mx_db.md_entries;
7153 mdb_cursor_close(MDB_cursor *mc)
7155 if (mc && !mc->mc_backup) {
7156 /* remove from txn, if tracked */
7157 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7158 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7159 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7161 *prev = mc->mc_next;
7168 mdb_cursor_txn(MDB_cursor *mc)
7170 if (!mc) return NULL;
7175 mdb_cursor_dbi(MDB_cursor *mc)
7180 /** Replace the key for a branch node with a new key.
7181 * @param[in] mc Cursor pointing to the node to operate on.
7182 * @param[in] key The new key to use.
7183 * @return 0 on success, non-zero on failure.
7186 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7192 int delta, ksize, oksize;
7193 indx_t ptr, i, numkeys, indx;
7196 indx = mc->mc_ki[mc->mc_top];
7197 mp = mc->mc_pg[mc->mc_top];
7198 node = NODEPTR(mp, indx);
7199 ptr = mp->mp_ptrs[indx];
7203 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7204 k2.mv_data = NODEKEY(node);
7205 k2.mv_size = node->mn_ksize;
7206 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7208 mdb_dkey(&k2, kbuf2),
7214 /* Sizes must be 2-byte aligned. */
7215 ksize = EVEN(key->mv_size);
7216 oksize = EVEN(node->mn_ksize);
7217 delta = ksize - oksize;
7219 /* Shift node contents if EVEN(key length) changed. */
7221 if (delta > 0 && SIZELEFT(mp) < delta) {
7223 /* not enough space left, do a delete and split */
7224 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7225 pgno = NODEPGNO(node);
7226 mdb_node_del(mc, 0);
7227 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7230 numkeys = NUMKEYS(mp);
7231 for (i = 0; i < numkeys; i++) {
7232 if (mp->mp_ptrs[i] <= ptr)
7233 mp->mp_ptrs[i] -= delta;
7236 base = (char *)mp + mp->mp_upper + PAGEBASE;
7237 len = ptr - mp->mp_upper + NODESIZE;
7238 memmove(base - delta, base, len);
7239 mp->mp_upper -= delta;
7241 node = NODEPTR(mp, indx);
7244 /* But even if no shift was needed, update ksize */
7245 if (node->mn_ksize != key->mv_size)
7246 node->mn_ksize = key->mv_size;
7249 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7255 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7257 /** Move a node from csrc to cdst.
7260 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7267 unsigned short flags;
7271 /* Mark src and dst as dirty. */
7272 if ((rc = mdb_page_touch(csrc)) ||
7273 (rc = mdb_page_touch(cdst)))
7276 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7277 key.mv_size = csrc->mc_db->md_pad;
7278 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7280 data.mv_data = NULL;
7284 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7285 mdb_cassert(csrc, !((size_t)srcnode & 1));
7286 srcpg = NODEPGNO(srcnode);
7287 flags = srcnode->mn_flags;
7288 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7289 unsigned int snum = csrc->mc_snum;
7291 /* must find the lowest key below src */
7292 rc = mdb_page_search_lowest(csrc);
7295 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7296 key.mv_size = csrc->mc_db->md_pad;
7297 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7299 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7300 key.mv_size = NODEKSZ(s2);
7301 key.mv_data = NODEKEY(s2);
7303 csrc->mc_snum = snum--;
7304 csrc->mc_top = snum;
7306 key.mv_size = NODEKSZ(srcnode);
7307 key.mv_data = NODEKEY(srcnode);
7309 data.mv_size = NODEDSZ(srcnode);
7310 data.mv_data = NODEDATA(srcnode);
7312 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7313 unsigned int snum = cdst->mc_snum;
7316 /* must find the lowest key below dst */
7317 mdb_cursor_copy(cdst, &mn);
7318 rc = mdb_page_search_lowest(&mn);
7321 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7322 bkey.mv_size = mn.mc_db->md_pad;
7323 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7325 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7326 bkey.mv_size = NODEKSZ(s2);
7327 bkey.mv_data = NODEKEY(s2);
7329 mn.mc_snum = snum--;
7332 rc = mdb_update_key(&mn, &bkey);
7337 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7338 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7339 csrc->mc_ki[csrc->mc_top],
7341 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7342 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7344 /* Add the node to the destination page.
7346 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7347 if (rc != MDB_SUCCESS)
7350 /* Delete the node from the source page.
7352 mdb_node_del(csrc, key.mv_size);
7355 /* Adjust other cursors pointing to mp */
7356 MDB_cursor *m2, *m3;
7357 MDB_dbi dbi = csrc->mc_dbi;
7358 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7360 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7361 if (csrc->mc_flags & C_SUB)
7362 m3 = &m2->mc_xcursor->mx_cursor;
7365 if (m3 == csrc) continue;
7366 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7367 csrc->mc_ki[csrc->mc_top]) {
7368 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7369 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7374 /* Update the parent separators.
7376 if (csrc->mc_ki[csrc->mc_top] == 0) {
7377 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7378 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7379 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7381 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7382 key.mv_size = NODEKSZ(srcnode);
7383 key.mv_data = NODEKEY(srcnode);
7385 DPRINTF(("update separator for source page %"Z"u to [%s]",
7386 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7387 mdb_cursor_copy(csrc, &mn);
7390 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7393 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7395 indx_t ix = csrc->mc_ki[csrc->mc_top];
7396 nullkey.mv_size = 0;
7397 csrc->mc_ki[csrc->mc_top] = 0;
7398 rc = mdb_update_key(csrc, &nullkey);
7399 csrc->mc_ki[csrc->mc_top] = ix;
7400 mdb_cassert(csrc, rc == MDB_SUCCESS);
7404 if (cdst->mc_ki[cdst->mc_top] == 0) {
7405 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7406 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7407 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7409 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7410 key.mv_size = NODEKSZ(srcnode);
7411 key.mv_data = NODEKEY(srcnode);
7413 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7414 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7415 mdb_cursor_copy(cdst, &mn);
7418 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7421 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7423 indx_t ix = cdst->mc_ki[cdst->mc_top];
7424 nullkey.mv_size = 0;
7425 cdst->mc_ki[cdst->mc_top] = 0;
7426 rc = mdb_update_key(cdst, &nullkey);
7427 cdst->mc_ki[cdst->mc_top] = ix;
7428 mdb_cassert(csrc, rc == MDB_SUCCESS);
7435 /** Merge one page into another.
7436 * The nodes from the page pointed to by \b csrc will
7437 * be copied to the page pointed to by \b cdst and then
7438 * the \b csrc page will be freed.
7439 * @param[in] csrc Cursor pointing to the source page.
7440 * @param[in] cdst Cursor pointing to the destination page.
7441 * @return 0 on success, non-zero on failure.
7444 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7446 MDB_page *psrc, *pdst;
7453 psrc = csrc->mc_pg[csrc->mc_top];
7454 pdst = cdst->mc_pg[cdst->mc_top];
7456 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7458 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7459 mdb_cassert(csrc, cdst->mc_snum > 1);
7461 /* Mark dst as dirty. */
7462 if ((rc = mdb_page_touch(cdst)))
7465 /* Move all nodes from src to dst.
7467 j = nkeys = NUMKEYS(pdst);
7468 if (IS_LEAF2(psrc)) {
7469 key.mv_size = csrc->mc_db->md_pad;
7470 key.mv_data = METADATA(psrc);
7471 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7472 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7473 if (rc != MDB_SUCCESS)
7475 key.mv_data = (char *)key.mv_data + key.mv_size;
7478 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7479 srcnode = NODEPTR(psrc, i);
7480 if (i == 0 && IS_BRANCH(psrc)) {
7483 mdb_cursor_copy(csrc, &mn);
7484 /* must find the lowest key below src */
7485 rc = mdb_page_search_lowest(&mn);
7488 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7489 key.mv_size = mn.mc_db->md_pad;
7490 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7492 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7493 key.mv_size = NODEKSZ(s2);
7494 key.mv_data = NODEKEY(s2);
7497 key.mv_size = srcnode->mn_ksize;
7498 key.mv_data = NODEKEY(srcnode);
7501 data.mv_size = NODEDSZ(srcnode);
7502 data.mv_data = NODEDATA(srcnode);
7503 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7504 if (rc != MDB_SUCCESS)
7509 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7510 pdst->mp_pgno, NUMKEYS(pdst),
7511 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7513 /* Unlink the src page from parent and add to free list.
7516 mdb_node_del(csrc, 0);
7517 if (csrc->mc_ki[csrc->mc_top] == 0) {
7519 rc = mdb_update_key(csrc, &key);
7527 psrc = csrc->mc_pg[csrc->mc_top];
7528 /* If not operating on FreeDB, allow this page to be reused
7529 * in this txn. Otherwise just add to free list.
7531 rc = mdb_page_loose(csrc, psrc);
7535 csrc->mc_db->md_leaf_pages--;
7537 csrc->mc_db->md_branch_pages--;
7539 /* Adjust other cursors pointing to mp */
7540 MDB_cursor *m2, *m3;
7541 MDB_dbi dbi = csrc->mc_dbi;
7543 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7544 if (csrc->mc_flags & C_SUB)
7545 m3 = &m2->mc_xcursor->mx_cursor;
7548 if (m3 == csrc) continue;
7549 if (m3->mc_snum < csrc->mc_snum) continue;
7550 if (m3->mc_pg[csrc->mc_top] == psrc) {
7551 m3->mc_pg[csrc->mc_top] = pdst;
7552 m3->mc_ki[csrc->mc_top] += nkeys;
7557 unsigned int snum = cdst->mc_snum;
7558 uint16_t depth = cdst->mc_db->md_depth;
7559 mdb_cursor_pop(cdst);
7560 rc = mdb_rebalance(cdst);
7561 /* Did the tree shrink? */
7562 if (depth > cdst->mc_db->md_depth)
7564 cdst->mc_snum = snum;
7565 cdst->mc_top = snum-1;
7570 /** Copy the contents of a cursor.
7571 * @param[in] csrc The cursor to copy from.
7572 * @param[out] cdst The cursor to copy to.
7575 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7579 cdst->mc_txn = csrc->mc_txn;
7580 cdst->mc_dbi = csrc->mc_dbi;
7581 cdst->mc_db = csrc->mc_db;
7582 cdst->mc_dbx = csrc->mc_dbx;
7583 cdst->mc_snum = csrc->mc_snum;
7584 cdst->mc_top = csrc->mc_top;
7585 cdst->mc_flags = csrc->mc_flags;
7587 for (i=0; i<csrc->mc_snum; i++) {
7588 cdst->mc_pg[i] = csrc->mc_pg[i];
7589 cdst->mc_ki[i] = csrc->mc_ki[i];
7593 /** Rebalance the tree after a delete operation.
7594 * @param[in] mc Cursor pointing to the page where rebalancing
7596 * @return 0 on success, non-zero on failure.
7599 mdb_rebalance(MDB_cursor *mc)
7603 unsigned int ptop, minkeys;
7607 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7608 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7609 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7610 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7611 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7613 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7614 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7615 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7616 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7620 if (mc->mc_snum < 2) {
7621 MDB_page *mp = mc->mc_pg[0];
7623 DPUTS("Can't rebalance a subpage, ignoring");
7626 if (NUMKEYS(mp) == 0) {
7627 DPUTS("tree is completely empty");
7628 mc->mc_db->md_root = P_INVALID;
7629 mc->mc_db->md_depth = 0;
7630 mc->mc_db->md_leaf_pages = 0;
7631 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7634 /* Adjust cursors pointing to mp */
7637 mc->mc_flags &= ~C_INITIALIZED;
7639 MDB_cursor *m2, *m3;
7640 MDB_dbi dbi = mc->mc_dbi;
7642 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7643 if (mc->mc_flags & C_SUB)
7644 m3 = &m2->mc_xcursor->mx_cursor;
7647 if (m3->mc_snum < mc->mc_snum) continue;
7648 if (m3->mc_pg[0] == mp) {
7651 m3->mc_flags &= ~C_INITIALIZED;
7655 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7657 DPUTS("collapsing root page!");
7658 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7661 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7662 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7665 mc->mc_db->md_depth--;
7666 mc->mc_db->md_branch_pages--;
7667 mc->mc_ki[0] = mc->mc_ki[1];
7668 for (i = 1; i<mc->mc_db->md_depth; i++) {
7669 mc->mc_pg[i] = mc->mc_pg[i+1];
7670 mc->mc_ki[i] = mc->mc_ki[i+1];
7673 /* Adjust other cursors pointing to mp */
7674 MDB_cursor *m2, *m3;
7675 MDB_dbi dbi = mc->mc_dbi;
7677 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7678 if (mc->mc_flags & C_SUB)
7679 m3 = &m2->mc_xcursor->mx_cursor;
7682 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7683 if (m3->mc_pg[0] == mp) {
7686 for (i=0; i<m3->mc_snum; i++) {
7687 m3->mc_pg[i] = m3->mc_pg[i+1];
7688 m3->mc_ki[i] = m3->mc_ki[i+1];
7694 DPUTS("root page doesn't need rebalancing");
7698 /* The parent (branch page) must have at least 2 pointers,
7699 * otherwise the tree is invalid.
7701 ptop = mc->mc_top-1;
7702 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7704 /* Leaf page fill factor is below the threshold.
7705 * Try to move keys from left or right neighbor, or
7706 * merge with a neighbor page.
7711 mdb_cursor_copy(mc, &mn);
7712 mn.mc_xcursor = NULL;
7714 oldki = mc->mc_ki[mc->mc_top];
7715 if (mc->mc_ki[ptop] == 0) {
7716 /* We're the leftmost leaf in our parent.
7718 DPUTS("reading right neighbor");
7720 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7721 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7724 mn.mc_ki[mn.mc_top] = 0;
7725 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7727 /* There is at least one neighbor to the left.
7729 DPUTS("reading left neighbor");
7731 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7732 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7735 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7736 mc->mc_ki[mc->mc_top] = 0;
7739 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7740 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7741 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7743 /* If the neighbor page is above threshold and has enough keys,
7744 * move one key from it. Otherwise we should try to merge them.
7745 * (A branch page must never have less than 2 keys.)
7747 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7748 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7749 rc = mdb_node_move(&mn, mc);
7750 if (mc->mc_ki[ptop]) {
7754 if (mc->mc_ki[ptop] == 0) {
7755 rc = mdb_page_merge(&mn, mc);
7757 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7758 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7759 rc = mdb_page_merge(mc, &mn);
7760 mdb_cursor_copy(&mn, mc);
7762 mc->mc_flags &= ~C_EOF;
7764 mc->mc_ki[mc->mc_top] = oldki;
7768 /** Complete a delete operation started by #mdb_cursor_del(). */
7770 mdb_cursor_del0(MDB_cursor *mc)
7777 ki = mc->mc_ki[mc->mc_top];
7778 mdb_node_del(mc, mc->mc_db->md_pad);
7779 mc->mc_db->md_entries--;
7780 rc = mdb_rebalance(mc);
7782 if (rc == MDB_SUCCESS) {
7783 MDB_cursor *m2, *m3;
7784 MDB_dbi dbi = mc->mc_dbi;
7786 mp = mc->mc_pg[mc->mc_top];
7787 nkeys = NUMKEYS(mp);
7789 /* if mc points past last node in page, find next sibling */
7790 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7791 rc = mdb_cursor_sibling(mc, 1);
7792 if (rc == MDB_NOTFOUND) {
7793 mc->mc_flags |= C_EOF;
7798 /* Adjust other cursors pointing to mp */
7799 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7800 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7801 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7803 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7805 if (m3->mc_pg[mc->mc_top] == mp) {
7806 if (m3->mc_ki[mc->mc_top] >= ki) {
7807 m3->mc_flags |= C_DEL;
7808 if (m3->mc_ki[mc->mc_top] > ki)
7809 m3->mc_ki[mc->mc_top]--;
7810 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7811 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7813 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7814 rc = mdb_cursor_sibling(m3, 1);
7815 if (rc == MDB_NOTFOUND) {
7816 m3->mc_flags |= C_EOF;
7822 mc->mc_flags |= C_DEL;
7826 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7831 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7832 MDB_val *key, MDB_val *data)
7834 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7837 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7838 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7840 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7841 /* must ignore any data */
7845 return mdb_del0(txn, dbi, key, data, 0);
7849 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7850 MDB_val *key, MDB_val *data, unsigned flags)
7855 MDB_val rdata, *xdata;
7859 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7861 mdb_cursor_init(&mc, txn, dbi, &mx);
7870 flags |= MDB_NODUPDATA;
7872 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7874 /* let mdb_page_split know about this cursor if needed:
7875 * delete will trigger a rebalance; if it needs to move
7876 * a node from one page to another, it will have to
7877 * update the parent's separator key(s). If the new sepkey
7878 * is larger than the current one, the parent page may
7879 * run out of space, triggering a split. We need this
7880 * cursor to be consistent until the end of the rebalance.
7882 mc.mc_flags |= C_UNTRACK;
7883 mc.mc_next = txn->mt_cursors[dbi];
7884 txn->mt_cursors[dbi] = &mc;
7885 rc = mdb_cursor_del(&mc, flags);
7886 txn->mt_cursors[dbi] = mc.mc_next;
7891 /** Split a page and insert a new node.
7892 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7893 * The cursor will be updated to point to the actual page and index where
7894 * the node got inserted after the split.
7895 * @param[in] newkey The key for the newly inserted node.
7896 * @param[in] newdata The data for the newly inserted node.
7897 * @param[in] newpgno The page number, if the new node is a branch node.
7898 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7899 * @return 0 on success, non-zero on failure.
7902 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7903 unsigned int nflags)
7906 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7909 int i, j, split_indx, nkeys, pmax;
7910 MDB_env *env = mc->mc_txn->mt_env;
7912 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7913 MDB_page *copy = NULL;
7914 MDB_page *mp, *rp, *pp;
7919 mp = mc->mc_pg[mc->mc_top];
7920 newindx = mc->mc_ki[mc->mc_top];
7921 nkeys = NUMKEYS(mp);
7923 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7924 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7925 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7927 /* Create a right sibling. */
7928 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7930 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7932 if (mc->mc_snum < 2) {
7933 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7935 /* shift current top to make room for new parent */
7936 mc->mc_pg[1] = mc->mc_pg[0];
7937 mc->mc_ki[1] = mc->mc_ki[0];
7940 mc->mc_db->md_root = pp->mp_pgno;
7941 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7942 mc->mc_db->md_depth++;
7945 /* Add left (implicit) pointer. */
7946 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7947 /* undo the pre-push */
7948 mc->mc_pg[0] = mc->mc_pg[1];
7949 mc->mc_ki[0] = mc->mc_ki[1];
7950 mc->mc_db->md_root = mp->mp_pgno;
7951 mc->mc_db->md_depth--;
7958 ptop = mc->mc_top-1;
7959 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7962 mc->mc_flags |= C_SPLITTING;
7963 mdb_cursor_copy(mc, &mn);
7964 mn.mc_pg[mn.mc_top] = rp;
7965 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7967 if (nflags & MDB_APPEND) {
7968 mn.mc_ki[mn.mc_top] = 0;
7970 split_indx = newindx;
7974 split_indx = (nkeys+1) / 2;
7979 unsigned int lsize, rsize, ksize;
7980 /* Move half of the keys to the right sibling */
7981 x = mc->mc_ki[mc->mc_top] - split_indx;
7982 ksize = mc->mc_db->md_pad;
7983 split = LEAF2KEY(mp, split_indx, ksize);
7984 rsize = (nkeys - split_indx) * ksize;
7985 lsize = (nkeys - split_indx) * sizeof(indx_t);
7986 mp->mp_lower -= lsize;
7987 rp->mp_lower += lsize;
7988 mp->mp_upper += rsize - lsize;
7989 rp->mp_upper -= rsize - lsize;
7990 sepkey.mv_size = ksize;
7991 if (newindx == split_indx) {
7992 sepkey.mv_data = newkey->mv_data;
7994 sepkey.mv_data = split;
7997 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7998 memcpy(rp->mp_ptrs, split, rsize);
7999 sepkey.mv_data = rp->mp_ptrs;
8000 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8001 memcpy(ins, newkey->mv_data, ksize);
8002 mp->mp_lower += sizeof(indx_t);
8003 mp->mp_upper -= ksize - sizeof(indx_t);
8006 memcpy(rp->mp_ptrs, split, x * ksize);
8007 ins = LEAF2KEY(rp, x, ksize);
8008 memcpy(ins, newkey->mv_data, ksize);
8009 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8010 rp->mp_lower += sizeof(indx_t);
8011 rp->mp_upper -= ksize - sizeof(indx_t);
8012 mc->mc_ki[mc->mc_top] = x;
8013 mc->mc_pg[mc->mc_top] = rp;
8016 int psize, nsize, k;
8017 /* Maximum free space in an empty page */
8018 pmax = env->me_psize - PAGEHDRSZ;
8020 nsize = mdb_leaf_size(env, newkey, newdata);
8022 nsize = mdb_branch_size(env, newkey);
8023 nsize = EVEN(nsize);
8025 /* grab a page to hold a temporary copy */
8026 copy = mdb_page_malloc(mc->mc_txn, 1);
8031 copy->mp_pgno = mp->mp_pgno;
8032 copy->mp_flags = mp->mp_flags;
8033 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8034 copy->mp_upper = env->me_psize - PAGEBASE;
8036 /* prepare to insert */
8037 for (i=0, j=0; i<nkeys; i++) {
8039 copy->mp_ptrs[j++] = 0;
8041 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8044 /* When items are relatively large the split point needs
8045 * to be checked, because being off-by-one will make the
8046 * difference between success or failure in mdb_node_add.
8048 * It's also relevant if a page happens to be laid out
8049 * such that one half of its nodes are all "small" and
8050 * the other half of its nodes are "large." If the new
8051 * item is also "large" and falls on the half with
8052 * "large" nodes, it also may not fit.
8054 * As a final tweak, if the new item goes on the last
8055 * spot on the page (and thus, onto the new page), bias
8056 * the split so the new page is emptier than the old page.
8057 * This yields better packing during sequential inserts.
8059 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8060 /* Find split point */
8062 if (newindx <= split_indx || newindx >= nkeys) {
8064 k = newindx >= nkeys ? nkeys : split_indx+2;
8069 for (; i!=k; i+=j) {
8074 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8075 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8077 if (F_ISSET(node->mn_flags, F_BIGDATA))
8078 psize += sizeof(pgno_t);
8080 psize += NODEDSZ(node);
8082 psize = EVEN(psize);
8084 if (psize > pmax || i == k-j) {
8085 split_indx = i + (j<0);
8090 if (split_indx == newindx) {
8091 sepkey.mv_size = newkey->mv_size;
8092 sepkey.mv_data = newkey->mv_data;
8094 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8095 sepkey.mv_size = node->mn_ksize;
8096 sepkey.mv_data = NODEKEY(node);
8101 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8103 /* Copy separator key to the parent.
8105 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8109 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8114 if (mn.mc_snum == mc->mc_snum) {
8115 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8116 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8117 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8118 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8123 /* Right page might now have changed parent.
8124 * Check if left page also changed parent.
8126 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8127 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8128 for (i=0; i<ptop; i++) {
8129 mc->mc_pg[i] = mn.mc_pg[i];
8130 mc->mc_ki[i] = mn.mc_ki[i];
8132 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8133 if (mn.mc_ki[ptop]) {
8134 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8136 /* find right page's left sibling */
8137 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8138 mdb_cursor_sibling(mc, 0);
8143 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8146 mc->mc_flags ^= C_SPLITTING;
8147 if (rc != MDB_SUCCESS) {
8150 if (nflags & MDB_APPEND) {
8151 mc->mc_pg[mc->mc_top] = rp;
8152 mc->mc_ki[mc->mc_top] = 0;
8153 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8156 for (i=0; i<mc->mc_top; i++)
8157 mc->mc_ki[i] = mn.mc_ki[i];
8158 } else if (!IS_LEAF2(mp)) {
8160 mc->mc_pg[mc->mc_top] = rp;
8165 rkey.mv_data = newkey->mv_data;
8166 rkey.mv_size = newkey->mv_size;
8172 /* Update index for the new key. */
8173 mc->mc_ki[mc->mc_top] = j;
8175 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8176 rkey.mv_data = NODEKEY(node);
8177 rkey.mv_size = node->mn_ksize;
8179 xdata.mv_data = NODEDATA(node);
8180 xdata.mv_size = NODEDSZ(node);
8183 pgno = NODEPGNO(node);
8184 flags = node->mn_flags;
8187 if (!IS_LEAF(mp) && j == 0) {
8188 /* First branch index doesn't need key data. */
8192 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8198 mc->mc_pg[mc->mc_top] = copy;
8203 } while (i != split_indx);
8205 nkeys = NUMKEYS(copy);
8206 for (i=0; i<nkeys; i++)
8207 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8208 mp->mp_lower = copy->mp_lower;
8209 mp->mp_upper = copy->mp_upper;
8210 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8211 env->me_psize - copy->mp_upper - PAGEBASE);
8213 /* reset back to original page */
8214 if (newindx < split_indx) {
8215 mc->mc_pg[mc->mc_top] = mp;
8216 if (nflags & MDB_RESERVE) {
8217 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8218 if (!(node->mn_flags & F_BIGDATA))
8219 newdata->mv_data = NODEDATA(node);
8222 mc->mc_pg[mc->mc_top] = rp;
8224 /* Make sure mc_ki is still valid.
8226 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8227 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8228 for (i=0; i<=ptop; i++) {
8229 mc->mc_pg[i] = mn.mc_pg[i];
8230 mc->mc_ki[i] = mn.mc_ki[i];
8237 /* Adjust other cursors pointing to mp */
8238 MDB_cursor *m2, *m3;
8239 MDB_dbi dbi = mc->mc_dbi;
8240 int fixup = NUMKEYS(mp);
8242 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8243 if (mc->mc_flags & C_SUB)
8244 m3 = &m2->mc_xcursor->mx_cursor;
8249 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8251 if (m3->mc_flags & C_SPLITTING)
8256 for (k=m3->mc_top; k>=0; k--) {
8257 m3->mc_ki[k+1] = m3->mc_ki[k];
8258 m3->mc_pg[k+1] = m3->mc_pg[k];
8260 if (m3->mc_ki[0] >= split_indx) {
8265 m3->mc_pg[0] = mc->mc_pg[0];
8269 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8270 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8271 m3->mc_ki[mc->mc_top]++;
8272 if (m3->mc_ki[mc->mc_top] >= fixup) {
8273 m3->mc_pg[mc->mc_top] = rp;
8274 m3->mc_ki[mc->mc_top] -= fixup;
8275 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8277 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8278 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8283 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8286 if (copy) /* tmp page */
8287 mdb_page_free(env, copy);
8289 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8294 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8295 MDB_val *key, MDB_val *data, unsigned int flags)
8300 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8303 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8306 mdb_cursor_init(&mc, txn, dbi, &mx);
8307 return mdb_cursor_put(&mc, key, data, flags);
8311 #define MDB_WBUF (1024*1024)
8314 /** State needed for a compacting copy. */
8315 typedef struct mdb_copy {
8316 pthread_mutex_t mc_mutex;
8317 pthread_cond_t mc_cond;
8324 pgno_t mc_next_pgno;
8327 volatile int mc_new;
8332 /** Dedicated writer thread for compacting copy. */
8333 static THREAD_RET ESECT
8334 mdb_env_copythr(void *arg)
8338 int toggle = 0, wsize, rc;
8341 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8344 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8347 pthread_mutex_lock(&my->mc_mutex);
8349 pthread_cond_signal(&my->mc_cond);
8352 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8353 if (my->mc_new < 0) {
8358 wsize = my->mc_wlen[toggle];
8359 ptr = my->mc_wbuf[toggle];
8362 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8366 } else if (len > 0) {
8380 /* If there's an overflow page tail, write it too */
8381 if (my->mc_olen[toggle]) {
8382 wsize = my->mc_olen[toggle];
8383 ptr = my->mc_over[toggle];
8384 my->mc_olen[toggle] = 0;
8387 my->mc_wlen[toggle] = 0;
8389 pthread_cond_signal(&my->mc_cond);
8391 pthread_cond_signal(&my->mc_cond);
8392 pthread_mutex_unlock(&my->mc_mutex);
8393 return (THREAD_RET)0;
8397 /** Tell the writer thread there's a buffer ready to write */
8399 mdb_env_cthr_toggle(mdb_copy *my, int st)
8401 int toggle = my->mc_toggle ^ 1;
8402 pthread_mutex_lock(&my->mc_mutex);
8403 if (my->mc_status) {
8404 pthread_mutex_unlock(&my->mc_mutex);
8405 return my->mc_status;
8407 while (my->mc_new == 1)
8408 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8410 my->mc_toggle = toggle;
8411 pthread_cond_signal(&my->mc_cond);
8412 pthread_mutex_unlock(&my->mc_mutex);
8416 /** Depth-first tree traversal for compacting copy. */
8418 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8421 MDB_txn *txn = my->mc_txn;
8423 MDB_page *mo, *mp, *leaf;
8428 /* Empty DB, nothing to do */
8429 if (*pg == P_INVALID)
8436 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8439 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8443 /* Make cursor pages writable */
8444 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8448 for (i=0; i<mc.mc_top; i++) {
8449 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8450 mc.mc_pg[i] = (MDB_page *)ptr;
8451 ptr += my->mc_env->me_psize;
8454 /* This is writable space for a leaf page. Usually not needed. */
8455 leaf = (MDB_page *)ptr;
8457 toggle = my->mc_toggle;
8458 while (mc.mc_snum > 0) {
8460 mp = mc.mc_pg[mc.mc_top];
8464 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8465 for (i=0; i<n; i++) {
8466 ni = NODEPTR(mp, i);
8467 if (ni->mn_flags & F_BIGDATA) {
8471 /* Need writable leaf */
8473 mc.mc_pg[mc.mc_top] = leaf;
8474 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8476 ni = NODEPTR(mp, i);
8479 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8480 rc = mdb_page_get(txn, pg, &omp, NULL);
8483 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8484 rc = mdb_env_cthr_toggle(my, 1);
8487 toggle = my->mc_toggle;
8489 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8490 memcpy(mo, omp, my->mc_env->me_psize);
8491 mo->mp_pgno = my->mc_next_pgno;
8492 my->mc_next_pgno += omp->mp_pages;
8493 my->mc_wlen[toggle] += my->mc_env->me_psize;
8494 if (omp->mp_pages > 1) {
8495 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8496 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8497 rc = mdb_env_cthr_toggle(my, 1);
8500 toggle = my->mc_toggle;
8502 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8503 } else if (ni->mn_flags & F_SUBDATA) {
8506 /* Need writable leaf */
8508 mc.mc_pg[mc.mc_top] = leaf;
8509 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8511 ni = NODEPTR(mp, i);
8514 memcpy(&db, NODEDATA(ni), sizeof(db));
8515 my->mc_toggle = toggle;
8516 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8519 toggle = my->mc_toggle;
8520 memcpy(NODEDATA(ni), &db, sizeof(db));
8525 mc.mc_ki[mc.mc_top]++;
8526 if (mc.mc_ki[mc.mc_top] < n) {
8529 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8531 rc = mdb_page_get(txn, pg, &mp, NULL);
8536 mc.mc_ki[mc.mc_top] = 0;
8537 if (IS_BRANCH(mp)) {
8538 /* Whenever we advance to a sibling branch page,
8539 * we must proceed all the way down to its first leaf.
8541 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8544 mc.mc_pg[mc.mc_top] = mp;
8548 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8549 rc = mdb_env_cthr_toggle(my, 1);
8552 toggle = my->mc_toggle;
8554 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8555 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8556 mo->mp_pgno = my->mc_next_pgno++;
8557 my->mc_wlen[toggle] += my->mc_env->me_psize;
8559 /* Update parent if there is one */
8560 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8561 SETPGNO(ni, mo->mp_pgno);
8562 mdb_cursor_pop(&mc);
8564 /* Otherwise we're done */
8574 /** Copy environment with compaction. */
8576 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8581 MDB_txn *txn = NULL;
8586 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8587 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8588 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8589 if (my.mc_wbuf[0] == NULL)
8592 pthread_mutex_init(&my.mc_mutex, NULL);
8593 pthread_cond_init(&my.mc_cond, NULL);
8594 #ifdef HAVE_MEMALIGN
8595 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8596 if (my.mc_wbuf[0] == NULL)
8599 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8604 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8605 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8610 my.mc_next_pgno = 2;
8616 THREAD_CREATE(thr, mdb_env_copythr, &my);
8618 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8622 mp = (MDB_page *)my.mc_wbuf[0];
8623 memset(mp, 0, 2*env->me_psize);
8625 mp->mp_flags = P_META;
8626 mm = (MDB_meta *)METADATA(mp);
8627 mdb_env_init_meta0(env, mm);
8628 mm->mm_address = env->me_metas[0]->mm_address;
8630 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8632 mp->mp_flags = P_META;
8633 *(MDB_meta *)METADATA(mp) = *mm;
8634 mm = (MDB_meta *)METADATA(mp);
8636 /* Count the number of free pages, subtract from lastpg to find
8637 * number of active pages
8640 MDB_ID freecount = 0;
8643 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8644 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8645 freecount += *(MDB_ID *)data.mv_data;
8646 freecount += txn->mt_dbs[0].md_branch_pages +
8647 txn->mt_dbs[0].md_leaf_pages +
8648 txn->mt_dbs[0].md_overflow_pages;
8650 /* Set metapage 1 */
8651 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8652 mm->mm_dbs[1] = txn->mt_dbs[1];
8653 if (mm->mm_last_pg > 1) {
8654 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8657 mm->mm_dbs[1].md_root = P_INVALID;
8660 my.mc_wlen[0] = env->me_psize * 2;
8662 pthread_mutex_lock(&my.mc_mutex);
8664 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8665 pthread_mutex_unlock(&my.mc_mutex);
8666 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8667 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8668 rc = mdb_env_cthr_toggle(&my, 1);
8669 mdb_env_cthr_toggle(&my, -1);
8670 pthread_mutex_lock(&my.mc_mutex);
8672 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8673 pthread_mutex_unlock(&my.mc_mutex);
8678 CloseHandle(my.mc_cond);
8679 CloseHandle(my.mc_mutex);
8680 _aligned_free(my.mc_wbuf[0]);
8682 pthread_cond_destroy(&my.mc_cond);
8683 pthread_mutex_destroy(&my.mc_mutex);
8684 free(my.mc_wbuf[0]);
8689 /** Copy environment as-is. */
8691 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8693 MDB_txn *txn = NULL;
8694 mdb_mutex_t *wmutex = NULL;
8700 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8704 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8707 /* Do the lock/unlock of the reader mutex before starting the
8708 * write txn. Otherwise other read txns could block writers.
8710 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8715 /* We must start the actual read txn after blocking writers */
8716 mdb_txn_reset0(txn, "reset-stage1");
8718 /* Temporarily block writers until we snapshot the meta pages */
8719 wmutex = MDB_MUTEX(env, w);
8720 if (LOCK_MUTEX(rc, env, wmutex))
8723 rc = mdb_txn_renew0(txn);
8725 UNLOCK_MUTEX(wmutex);
8730 wsize = env->me_psize * 2;
8734 DO_WRITE(rc, fd, ptr, w2, len);
8738 } else if (len > 0) {
8744 /* Non-blocking or async handles are not supported */
8750 UNLOCK_MUTEX(wmutex);
8755 w2 = txn->mt_next_pgno * env->me_psize;
8758 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8765 if (wsize > MAX_WRITE)
8769 DO_WRITE(rc, fd, ptr, w2, len);
8773 } else if (len > 0) {
8790 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8792 if (flags & MDB_CP_COMPACT)
8793 return mdb_env_copyfd1(env, fd);
8795 return mdb_env_copyfd0(env, fd);
8799 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8801 return mdb_env_copyfd2(env, fd, 0);
8805 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8809 HANDLE newfd = INVALID_HANDLE_VALUE;
8811 if (env->me_flags & MDB_NOSUBDIR) {
8812 lpath = (char *)path;
8815 len += sizeof(DATANAME);
8816 lpath = malloc(len);
8819 sprintf(lpath, "%s" DATANAME, path);
8822 /* The destination path must exist, but the destination file must not.
8823 * We don't want the OS to cache the writes, since the source data is
8824 * already in the OS cache.
8827 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8828 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8830 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8832 if (newfd == INVALID_HANDLE_VALUE) {
8837 if (env->me_psize >= env->me_os_psize) {
8839 /* Set O_DIRECT if the file system supports it */
8840 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8841 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8843 #ifdef F_NOCACHE /* __APPLE__ */
8844 rc = fcntl(newfd, F_NOCACHE, 1);
8852 rc = mdb_env_copyfd2(env, newfd, flags);
8855 if (!(env->me_flags & MDB_NOSUBDIR))
8857 if (newfd != INVALID_HANDLE_VALUE)
8858 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8865 mdb_env_copy(MDB_env *env, const char *path)
8867 return mdb_env_copy2(env, path, 0);
8871 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8873 if (flag & (env->me_map ? ~CHANGEABLE : ~(CHANGEABLE|CHANGELESS)))
8876 env->me_flags |= flag;
8878 env->me_flags &= ~flag;
8883 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8888 *arg = env->me_flags;
8893 mdb_env_set_userctx(MDB_env *env, void *ctx)
8897 env->me_userctx = ctx;
8902 mdb_env_get_userctx(MDB_env *env)
8904 return env ? env->me_userctx : NULL;
8908 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8913 env->me_assert_func = func;
8919 mdb_env_get_path(MDB_env *env, const char **arg)
8924 *arg = env->me_path;
8929 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8938 /** Common code for #mdb_stat() and #mdb_env_stat().
8939 * @param[in] env the environment to operate in.
8940 * @param[in] db the #MDB_db record containing the stats to return.
8941 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8942 * @return 0, this function always succeeds.
8945 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8947 arg->ms_psize = env->me_psize;
8948 arg->ms_depth = db->md_depth;
8949 arg->ms_branch_pages = db->md_branch_pages;
8950 arg->ms_leaf_pages = db->md_leaf_pages;
8951 arg->ms_overflow_pages = db->md_overflow_pages;
8952 arg->ms_entries = db->md_entries;
8958 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8962 if (env == NULL || arg == NULL)
8965 toggle = mdb_env_pick_meta(env);
8967 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8971 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8975 if (env == NULL || arg == NULL)
8978 toggle = mdb_env_pick_meta(env);
8979 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8980 arg->me_mapsize = env->me_mapsize;
8981 arg->me_maxreaders = env->me_maxreaders;
8983 /* me_numreaders may be zero if this process never used any readers. Use
8984 * the shared numreader count if it exists.
8986 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8988 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8989 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8993 /** Set the default comparison functions for a database.
8994 * Called immediately after a database is opened to set the defaults.
8995 * The user can then override them with #mdb_set_compare() or
8996 * #mdb_set_dupsort().
8997 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8998 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9001 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9003 uint16_t f = txn->mt_dbs[dbi].md_flags;
9005 txn->mt_dbxs[dbi].md_cmp =
9006 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9007 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9009 txn->mt_dbxs[dbi].md_dcmp =
9010 !(f & MDB_DUPSORT) ? 0 :
9011 ((f & MDB_INTEGERDUP)
9012 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9013 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9016 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9022 int rc, dbflag, exact;
9023 unsigned int unused = 0, seq;
9026 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9027 mdb_default_cmp(txn, FREE_DBI);
9030 if ((flags & VALID_FLAGS) != flags)
9032 if (txn->mt_flags & MDB_TXN_ERROR)
9038 if (flags & PERSISTENT_FLAGS) {
9039 uint16_t f2 = flags & PERSISTENT_FLAGS;
9040 /* make sure flag changes get committed */
9041 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9042 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9043 txn->mt_flags |= MDB_TXN_DIRTY;
9046 mdb_default_cmp(txn, MAIN_DBI);
9050 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9051 mdb_default_cmp(txn, MAIN_DBI);
9054 /* Is the DB already open? */
9056 for (i=2; i<txn->mt_numdbs; i++) {
9057 if (!txn->mt_dbxs[i].md_name.mv_size) {
9058 /* Remember this free slot */
9059 if (!unused) unused = i;
9062 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9063 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9069 /* If no free slot and max hit, fail */
9070 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9071 return MDB_DBS_FULL;
9073 /* Cannot mix named databases with some mainDB flags */
9074 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9075 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9077 /* Find the DB info */
9078 dbflag = DB_NEW|DB_VALID;
9081 key.mv_data = (void *)name;
9082 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9083 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9084 if (rc == MDB_SUCCESS) {
9085 /* make sure this is actually a DB */
9086 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9087 if (!(node->mn_flags & F_SUBDATA))
9088 return MDB_INCOMPATIBLE;
9089 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9090 /* Create if requested */
9091 data.mv_size = sizeof(MDB_db);
9092 data.mv_data = &dummy;
9093 memset(&dummy, 0, sizeof(dummy));
9094 dummy.md_root = P_INVALID;
9095 dummy.md_flags = flags & PERSISTENT_FLAGS;
9096 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9100 /* OK, got info, add to table */
9101 if (rc == MDB_SUCCESS) {
9102 unsigned int slot = unused ? unused : txn->mt_numdbs;
9103 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9104 txn->mt_dbxs[slot].md_name.mv_size = len;
9105 txn->mt_dbxs[slot].md_rel = NULL;
9106 txn->mt_dbflags[slot] = dbflag;
9107 /* txn-> and env-> are the same in read txns, use
9108 * tmp variable to avoid undefined assignment
9110 seq = ++txn->mt_env->me_dbiseqs[slot];
9111 txn->mt_dbiseqs[slot] = seq;
9113 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9115 mdb_default_cmp(txn, slot);
9124 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9126 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9129 if (txn->mt_flags & MDB_TXN_ERROR)
9132 if (txn->mt_dbflags[dbi] & DB_STALE) {
9135 /* Stale, must read the DB's root. cursor_init does it for us. */
9136 mdb_cursor_init(&mc, txn, dbi, &mx);
9138 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9141 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9144 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9146 ptr = env->me_dbxs[dbi].md_name.mv_data;
9147 /* If there was no name, this was already closed */
9149 env->me_dbxs[dbi].md_name.mv_data = NULL;
9150 env->me_dbxs[dbi].md_name.mv_size = 0;
9151 env->me_dbflags[dbi] = 0;
9152 env->me_dbiseqs[dbi]++;
9157 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9159 /* We could return the flags for the FREE_DBI too but what's the point? */
9160 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9162 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9166 /** Add all the DB's pages to the free list.
9167 * @param[in] mc Cursor on the DB to free.
9168 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9169 * @return 0 on success, non-zero on failure.
9172 mdb_drop0(MDB_cursor *mc, int subs)
9176 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9177 if (rc == MDB_SUCCESS) {
9178 MDB_txn *txn = mc->mc_txn;
9183 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9184 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9187 mdb_cursor_copy(mc, &mx);
9188 while (mc->mc_snum > 0) {
9189 MDB_page *mp = mc->mc_pg[mc->mc_top];
9190 unsigned n = NUMKEYS(mp);
9192 for (i=0; i<n; i++) {
9193 ni = NODEPTR(mp, i);
9194 if (ni->mn_flags & F_BIGDATA) {
9197 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9198 rc = mdb_page_get(txn, pg, &omp, NULL);
9201 mdb_cassert(mc, IS_OVERFLOW(omp));
9202 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9206 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9207 mdb_xcursor_init1(mc, ni);
9208 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9214 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9216 for (i=0; i<n; i++) {
9218 ni = NODEPTR(mp, i);
9221 mdb_midl_xappend(txn->mt_free_pgs, pg);
9226 mc->mc_ki[mc->mc_top] = i;
9227 rc = mdb_cursor_sibling(mc, 1);
9229 if (rc != MDB_NOTFOUND)
9231 /* no more siblings, go back to beginning
9232 * of previous level.
9236 for (i=1; i<mc->mc_snum; i++) {
9238 mc->mc_pg[i] = mx.mc_pg[i];
9243 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9246 txn->mt_flags |= MDB_TXN_ERROR;
9247 } else if (rc == MDB_NOTFOUND) {
9253 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9255 MDB_cursor *mc, *m2;
9258 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9261 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9264 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9267 rc = mdb_cursor_open(txn, dbi, &mc);
9271 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9272 /* Invalidate the dropped DB's cursors */
9273 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9274 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9278 /* Can't delete the main DB */
9279 if (del && dbi > MAIN_DBI) {
9280 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9282 txn->mt_dbflags[dbi] = DB_STALE;
9283 mdb_dbi_close(txn->mt_env, dbi);
9285 txn->mt_flags |= MDB_TXN_ERROR;
9288 /* reset the DB record, mark it dirty */
9289 txn->mt_dbflags[dbi] |= DB_DIRTY;
9290 txn->mt_dbs[dbi].md_depth = 0;
9291 txn->mt_dbs[dbi].md_branch_pages = 0;
9292 txn->mt_dbs[dbi].md_leaf_pages = 0;
9293 txn->mt_dbs[dbi].md_overflow_pages = 0;
9294 txn->mt_dbs[dbi].md_entries = 0;
9295 txn->mt_dbs[dbi].md_root = P_INVALID;
9297 txn->mt_flags |= MDB_TXN_DIRTY;
9300 mdb_cursor_close(mc);
9304 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9306 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9309 txn->mt_dbxs[dbi].md_cmp = cmp;
9313 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9315 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9318 txn->mt_dbxs[dbi].md_dcmp = cmp;
9322 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9324 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9327 txn->mt_dbxs[dbi].md_rel = rel;
9331 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9333 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9336 txn->mt_dbxs[dbi].md_relctx = ctx;
9341 mdb_env_get_maxkeysize(MDB_env *env)
9343 return ENV_MAXKEY(env);
9347 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9349 unsigned int i, rdrs;
9352 int rc = 0, first = 1;
9356 if (!env->me_txns) {
9357 return func("(no reader locks)\n", ctx);
9359 rdrs = env->me_txns->mti_numreaders;
9360 mr = env->me_txns->mti_readers;
9361 for (i=0; i<rdrs; i++) {
9363 txnid_t txnid = mr[i].mr_txnid;
9364 sprintf(buf, txnid == (txnid_t)-1 ?
9365 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9366 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9369 rc = func(" pid thread txnid\n", ctx);
9373 rc = func(buf, ctx);
9379 rc = func("(no active readers)\n", ctx);
9384 /** Insert pid into list if not already present.
9385 * return -1 if already present.
9388 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9390 /* binary search of pid in list */
9392 unsigned cursor = 1;
9394 unsigned n = ids[0];
9397 unsigned pivot = n >> 1;
9398 cursor = base + pivot + 1;
9399 val = pid - ids[cursor];
9404 } else if ( val > 0 ) {
9409 /* found, so it's a duplicate */
9418 for (n = ids[0]; n > cursor; n--)
9425 mdb_reader_check(MDB_env *env, int *dead)
9431 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9434 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9435 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9437 mdb_mutex_t *rmutex = rlocked ? NULL : MDB_MUTEX(env, r);
9438 unsigned int i, j, rdrs;
9440 MDB_PID_T *pids, pid;
9441 int rc = MDB_SUCCESS, count = 0;
9443 rdrs = env->me_txns->mti_numreaders;
9444 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9448 mr = env->me_txns->mti_readers;
9449 for (i=0; i<rdrs; i++) {
9451 if (pid && pid != env->me_pid) {
9452 if (mdb_pid_insert(pids, pid) == 0) {
9453 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9454 /* Stale reader found */
9457 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9458 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9460 rdrs = 0; /* the above checked all readers */
9462 /* Recheck, a new process may have reused pid */
9463 if (mdb_reader_pid(env, Pidcheck, pid))
9468 if (mr[j].mr_pid == pid) {
9469 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9470 (unsigned) pid, mr[j].mr_txnid));
9475 UNLOCK_MUTEX(rmutex);
9486 #ifdef MDB_ROBUST_SUPPORTED
9487 /** Handle #LOCK_MUTEX0() failure.
9488 * With #MDB_ROBUST, try to repair the lock file if the mutex owner died.
9489 * @param[in] env the environment handle
9490 * @param[in] mutex LOCK_MUTEX0() mutex
9491 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9492 * @return 0 on success with the mutex locked, or an error code on failure.
9494 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc)
9496 int toggle, rlocked, rc2;
9498 enum { WAIT_ABANDONED = EOWNERDEAD };
9501 if (rc == (int) WAIT_ABANDONED) {
9502 /* We own the mutex. Clean up after dead previous owner. */
9504 rlocked = (mutex == MDB_MUTEX(env, r));
9506 /* Keep mti_txnid updated, otherwise next writer can
9507 * overwrite data which latest meta page refers to.
9509 toggle = mdb_env_pick_meta(env);
9510 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9511 /* env is hosed if the dead thread was ours */
9513 env->me_flags |= MDB_FATAL_ERROR;
9518 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9519 (rc ? "this process' env is hosed" : "recovering")));
9520 rc2 = mdb_reader_check0(env, rlocked, NULL);
9522 rc2 = pthread_mutex_consistent(mutex);
9523 if (rc || (rc = rc2)) {
9524 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9525 UNLOCK_MUTEX(mutex);
9531 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9536 #endif /* MDB_ROBUST_SUPPORTED */