2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2015 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
82 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
83 /** fdatasync is broken on ext3/ext4fs on older kernels, see
84 * description in #mdb_env_open2 comments. You can safely
85 * define MDB_FDATASYNC_WORKS if this code will only be run
86 * on kernels 3.6 and newer.
88 #define BROKEN_FDATASYNC
102 typedef SSIZE_T ssize_t;
107 #if defined(__sun) || defined(ANDROID)
108 /* Most platforms have posix_memalign, older may only have memalign */
109 #define HAVE_MEMALIGN 1
113 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
114 #include <netinet/in.h>
115 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
118 #if defined(__APPLE__) || defined (BSD)
119 # define MDB_USE_POSIX_SEM 1
120 # define MDB_FDATASYNC fsync
121 #elif defined(ANDROID)
122 # define MDB_FDATASYNC fsync
127 #ifdef MDB_USE_POSIX_SEM
128 # define MDB_USE_HASH 1
129 #include <semaphore.h>
131 #define MDB_USE_POSIX_MUTEX 1
135 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \
136 + defined(MDB_USE_POSIX_MUTEX) != 1
137 # error "Ambiguous shared-lock implementation"
141 #include <valgrind/memcheck.h>
142 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
143 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
144 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
145 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
146 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
148 #define VGMEMP_CREATE(h,r,z)
149 #define VGMEMP_ALLOC(h,a,s)
150 #define VGMEMP_FREE(h,a)
151 #define VGMEMP_DESTROY(h)
152 #define VGMEMP_DEFINED(a,s)
156 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
157 /* Solaris just defines one or the other */
158 # define LITTLE_ENDIAN 1234
159 # define BIG_ENDIAN 4321
160 # ifdef _LITTLE_ENDIAN
161 # define BYTE_ORDER LITTLE_ENDIAN
163 # define BYTE_ORDER BIG_ENDIAN
166 # define BYTE_ORDER __BYTE_ORDER
170 #ifndef LITTLE_ENDIAN
171 #define LITTLE_ENDIAN __LITTLE_ENDIAN
174 #define BIG_ENDIAN __BIG_ENDIAN
177 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
178 #define MISALIGNED_OK 1
184 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
185 # error "Unknown or unsupported endianness (BYTE_ORDER)"
186 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
187 # error "Two's complement, reasonably sized integer types, please"
191 /** Put infrequently used env functions in separate section */
193 # define ESECT __attribute__ ((section("__TEXT,text_env")))
195 # define ESECT __attribute__ ((section("text_env")))
202 #define CALL_CONV WINAPI
207 /** @defgroup internal LMDB Internals
210 /** @defgroup compat Compatibility Macros
211 * A bunch of macros to minimize the amount of platform-specific ifdefs
212 * needed throughout the rest of the code. When the features this library
213 * needs are similar enough to POSIX to be hidden in a one-or-two line
214 * replacement, this macro approach is used.
218 /** Features under development */
223 /** Wrapper around __func__, which is a C99 feature */
224 #if __STDC_VERSION__ >= 199901L
225 # define mdb_func_ __func__
226 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
227 # define mdb_func_ __FUNCTION__
229 /* If a debug message says <mdb_unknown>(), update the #if statements above */
230 # define mdb_func_ "<mdb_unknown>"
233 /* Internal error codes, not exposed outside liblmdb */
234 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
236 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
237 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
238 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
242 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
244 /** Some platforms define the EOWNERDEAD error code
245 * even though they don't support Robust Mutexes.
246 * Compile with -DMDB_USE_ROBUST=0, or use some other
247 * mechanism like -DMDB_USE_SYSV_SEM instead of
248 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
249 * also Robust, but some systems don't support them
252 #ifndef MDB_USE_ROBUST
253 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
254 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
255 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
256 # define MDB_USE_ROBUST 0
258 # define MDB_USE_ROBUST 1
259 /* glibc < 2.10 only provided _np API */
260 # if defined(__GLIBC__) && GLIBC_VER < 0x02000a
261 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
262 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
263 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
266 #endif /* MDB_USE_ROBUST */
268 #if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
269 #define MDB_ROBUST_SUPPORTED 1
273 #define MDB_USE_HASH 1
274 #define MDB_PIDLOCK 0
275 #define THREAD_RET DWORD
276 #define pthread_t HANDLE
277 #define pthread_mutex_t HANDLE
278 #define pthread_cond_t HANDLE
279 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
280 #define pthread_key_t DWORD
281 #define pthread_self() GetCurrentThreadId()
282 #define pthread_key_create(x,y) \
283 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
284 #define pthread_key_delete(x) TlsFree(x)
285 #define pthread_getspecific(x) TlsGetValue(x)
286 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
287 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
288 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
289 #define pthread_cond_signal(x) SetEvent(*x)
290 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
291 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
292 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
293 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
294 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
295 #define mdb_mutex_consistent(mutex) 0
296 #define getpid() GetCurrentProcessId()
297 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
298 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
299 #define ErrCode() GetLastError()
300 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
301 #define close(fd) (CloseHandle(fd) ? 0 : -1)
302 #define munmap(ptr,len) UnmapViewOfFile(ptr)
303 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
304 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
306 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
310 #define THREAD_RET void *
311 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
312 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
313 #define Z "z" /**< printf format modifier for size_t */
315 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
316 #define MDB_PIDLOCK 1
318 #ifdef MDB_USE_POSIX_SEM
320 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
321 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
322 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
325 mdb_sem_wait(sem_t *sem)
328 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
332 #else /* MDB_USE_POSIX_MUTEX: */
333 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
334 * local variables keep it (mdb_mutexref_t).
336 * When #mdb_mutexref_t is a pointer declaration and #mdb_mutex_t is
337 * not, then it is array[size 1] so it can be assigned to a pointer.
340 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
342 /** Lock the reader or writer mutex.
343 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
345 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
346 /** Unlock the reader or writer mutex.
348 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
349 /** Mark mutex-protected data as repaired, after death of previous owner.
351 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
352 #endif /* MDB_USE_POSIX_SEM */
354 /** Get the error code for the last failed system function.
356 #define ErrCode() errno
358 /** An abstraction for a file handle.
359 * On POSIX systems file handles are small integers. On Windows
360 * they're opaque pointers.
364 /** A value for an invalid file handle.
365 * Mainly used to initialize file variables and signify that they are
368 #define INVALID_HANDLE_VALUE (-1)
370 /** Get the size of a memory page for the system.
371 * This is the basic size that the platform's memory manager uses, and is
372 * fundamental to the use of memory-mapped files.
374 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
377 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
380 #define MNAME_LEN (sizeof(pthread_mutex_t))
385 #ifdef MDB_ROBUST_SUPPORTED
386 /** Lock mutex, handle any error, set rc = result.
387 * Return 0 on success, nonzero (not rc) on error.
389 #define LOCK_MUTEX(rc, env, mutex) \
390 (((rc) = LOCK_MUTEX0(mutex)) && \
391 ((rc) = mdb_mutex_failed(env, mutex, rc)))
392 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
394 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
395 #define mdb_mutex_failed(env, mutex, rc) (rc)
399 /** A flag for opening a file and requesting synchronous data writes.
400 * This is only used when writing a meta page. It's not strictly needed;
401 * we could just do a normal write and then immediately perform a flush.
402 * But if this flag is available it saves us an extra system call.
404 * @note If O_DSYNC is undefined but exists in /usr/include,
405 * preferably set some compiler flag to get the definition.
406 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
409 # define MDB_DSYNC O_DSYNC
413 /** Function for flushing the data of a file. Define this to fsync
414 * if fdatasync() is not supported.
416 #ifndef MDB_FDATASYNC
417 # define MDB_FDATASYNC fdatasync
421 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
432 /** A page number in the database.
433 * Note that 64 bit page numbers are overkill, since pages themselves
434 * already represent 12-13 bits of addressable memory, and the OS will
435 * always limit applications to a maximum of 63 bits of address space.
437 * @note In the #MDB_node structure, we only store 48 bits of this value,
438 * which thus limits us to only 60 bits of addressable data.
440 typedef MDB_ID pgno_t;
442 /** A transaction ID.
443 * See struct MDB_txn.mt_txnid for details.
445 typedef MDB_ID txnid_t;
447 /** @defgroup debug Debug Macros
451 /** Enable debug output. Needs variable argument macros (a C99 feature).
452 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
453 * read from and written to the database (used for free space management).
459 static int mdb_debug;
460 static txnid_t mdb_debug_start;
462 /** Print a debug message with printf formatting.
463 * Requires double parenthesis around 2 or more args.
465 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
466 # define DPRINTF0(fmt, ...) \
467 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
469 # define DPRINTF(args) ((void) 0)
471 /** Print a debug string.
472 * The string is printed literally, with no format processing.
474 #define DPUTS(arg) DPRINTF(("%s", arg))
475 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
477 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
480 /** @brief The maximum size of a database page.
482 * It is 32k or 64k, since value-PAGEBASE must fit in
483 * #MDB_page.%mp_upper.
485 * LMDB will use database pages < OS pages if needed.
486 * That causes more I/O in write transactions: The OS must
487 * know (read) the whole page before writing a partial page.
489 * Note that we don't currently support Huge pages. On Linux,
490 * regular data files cannot use Huge pages, and in general
491 * Huge pages aren't actually pageable. We rely on the OS
492 * demand-pager to read our data and page it out when memory
493 * pressure from other processes is high. So until OSs have
494 * actual paging support for Huge pages, they're not viable.
496 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
498 /** The minimum number of keys required in a database page.
499 * Setting this to a larger value will place a smaller bound on the
500 * maximum size of a data item. Data items larger than this size will
501 * be pushed into overflow pages instead of being stored directly in
502 * the B-tree node. This value used to default to 4. With a page size
503 * of 4096 bytes that meant that any item larger than 1024 bytes would
504 * go into an overflow page. That also meant that on average 2-3KB of
505 * each overflow page was wasted space. The value cannot be lower than
506 * 2 because then there would no longer be a tree structure. With this
507 * value, items larger than 2KB will go into overflow pages, and on
508 * average only 1KB will be wasted.
510 #define MDB_MINKEYS 2
512 /** A stamp that identifies a file as an LMDB file.
513 * There's nothing special about this value other than that it is easily
514 * recognizable, and it will reflect any byte order mismatches.
516 #define MDB_MAGIC 0xBEEFC0DE
518 /** The version number for a database's datafile format. */
519 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
520 /** The version number for a database's lockfile format. */
521 #define MDB_LOCK_VERSION 1
523 /** @brief The max size of a key we can write, or 0 for computed max.
525 * This macro should normally be left alone or set to 0.
526 * Note that a database with big keys or dupsort data cannot be
527 * reliably modified by a liblmdb which uses a smaller max.
528 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
530 * Other values are allowed, for backwards compat. However:
531 * A value bigger than the computed max can break if you do not
532 * know what you are doing, and liblmdb <= 0.9.10 can break when
533 * modifying a DB with keys/dupsort data bigger than its max.
535 * Data items in an #MDB_DUPSORT database are also limited to
536 * this size, since they're actually keys of a sub-DB. Keys and
537 * #MDB_DUPSORT data items must fit on a node in a regular page.
539 #ifndef MDB_MAXKEYSIZE
540 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
543 /** The maximum size of a key we can write to the environment. */
545 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
547 #define ENV_MAXKEY(env) ((env)->me_maxkey)
550 /** @brief The maximum size of a data item.
552 * We only store a 32 bit value for node sizes.
554 #define MAXDATASIZE 0xffffffffUL
557 /** Key size which fits in a #DKBUF.
560 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
563 * This is used for printing a hex dump of a key's contents.
565 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
566 /** Display a key in hex.
568 * Invoke a function to display a key in hex.
570 #define DKEY(x) mdb_dkey(x, kbuf)
576 /** An invalid page number.
577 * Mainly used to denote an empty tree.
579 #define P_INVALID (~(pgno_t)0)
581 /** Test if the flags \b f are set in a flag word \b w. */
582 #define F_ISSET(w, f) (((w) & (f)) == (f))
584 /** Round \b n up to an even number. */
585 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
587 /** Used for offsets within a single page.
588 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
591 typedef uint16_t indx_t;
593 /** Default size of memory map.
594 * This is certainly too small for any actual applications. Apps should always set
595 * the size explicitly using #mdb_env_set_mapsize().
597 #define DEFAULT_MAPSIZE 1048576
599 /** @defgroup readers Reader Lock Table
600 * Readers don't acquire any locks for their data access. Instead, they
601 * simply record their transaction ID in the reader table. The reader
602 * mutex is needed just to find an empty slot in the reader table. The
603 * slot's address is saved in thread-specific data so that subsequent read
604 * transactions started by the same thread need no further locking to proceed.
606 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
608 * No reader table is used if the database is on a read-only filesystem, or
609 * if #MDB_NOLOCK is set.
611 * Since the database uses multi-version concurrency control, readers don't
612 * actually need any locking. This table is used to keep track of which
613 * readers are using data from which old transactions, so that we'll know
614 * when a particular old transaction is no longer in use. Old transactions
615 * that have discarded any data pages can then have those pages reclaimed
616 * for use by a later write transaction.
618 * The lock table is constructed such that reader slots are aligned with the
619 * processor's cache line size. Any slot is only ever used by one thread.
620 * This alignment guarantees that there will be no contention or cache
621 * thrashing as threads update their own slot info, and also eliminates
622 * any need for locking when accessing a slot.
624 * A writer thread will scan every slot in the table to determine the oldest
625 * outstanding reader transaction. Any freed pages older than this will be
626 * reclaimed by the writer. The writer doesn't use any locks when scanning
627 * this table. This means that there's no guarantee that the writer will
628 * see the most up-to-date reader info, but that's not required for correct
629 * operation - all we need is to know the upper bound on the oldest reader,
630 * we don't care at all about the newest reader. So the only consequence of
631 * reading stale information here is that old pages might hang around a
632 * while longer before being reclaimed. That's actually good anyway, because
633 * the longer we delay reclaiming old pages, the more likely it is that a
634 * string of contiguous pages can be found after coalescing old pages from
635 * many old transactions together.
638 /** Number of slots in the reader table.
639 * This value was chosen somewhat arbitrarily. 126 readers plus a
640 * couple mutexes fit exactly into 8KB on my development machine.
641 * Applications should set the table size using #mdb_env_set_maxreaders().
643 #define DEFAULT_READERS 126
645 /** The size of a CPU cache line in bytes. We want our lock structures
646 * aligned to this size to avoid false cache line sharing in the
648 * This value works for most CPUs. For Itanium this should be 128.
654 /** The information we store in a single slot of the reader table.
655 * In addition to a transaction ID, we also record the process and
656 * thread ID that owns a slot, so that we can detect stale information,
657 * e.g. threads or processes that went away without cleaning up.
658 * @note We currently don't check for stale records. We simply re-init
659 * the table when we know that we're the only process opening the
662 typedef struct MDB_rxbody {
663 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
664 * Multiple readers that start at the same time will probably have the
665 * same ID here. Again, it's not important to exclude them from
666 * anything; all we need to know is which version of the DB they
667 * started from so we can avoid overwriting any data used in that
668 * particular version.
670 volatile txnid_t mrb_txnid;
671 /** The process ID of the process owning this reader txn. */
672 volatile MDB_PID_T mrb_pid;
673 /** The thread ID of the thread owning this txn. */
674 volatile MDB_THR_T mrb_tid;
677 /** The actual reader record, with cacheline padding. */
678 typedef struct MDB_reader {
681 /** shorthand for mrb_txnid */
682 #define mr_txnid mru.mrx.mrb_txnid
683 #define mr_pid mru.mrx.mrb_pid
684 #define mr_tid mru.mrx.mrb_tid
685 /** cache line alignment */
686 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
690 /** The header for the reader table.
691 * The table resides in a memory-mapped file. (This is a different file
692 * than is used for the main database.)
694 * For POSIX the actual mutexes reside in the shared memory of this
695 * mapped file. On Windows, mutexes are named objects allocated by the
696 * kernel; we store the mutex names in this mapped file so that other
697 * processes can grab them. This same approach is also used on
698 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
699 * process-shared POSIX mutexes. For these cases where a named object
700 * is used, the object name is derived from a 64 bit FNV hash of the
701 * environment pathname. As such, naming collisions are extremely
702 * unlikely. If a collision occurs, the results are unpredictable.
704 typedef struct MDB_txbody {
705 /** Stamp identifying this as an LMDB file. It must be set
708 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
710 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
711 char mtb_rmname[MNAME_LEN];
713 /** Mutex protecting access to this table.
714 * This is the reader table lock used with LOCK_MUTEX().
716 mdb_mutex_t mtb_rmutex;
718 /** The ID of the last transaction committed to the database.
719 * This is recorded here only for convenience; the value can always
720 * be determined by reading the main database meta pages.
722 volatile txnid_t mtb_txnid;
723 /** The number of slots that have been used in the reader table.
724 * This always records the maximum count, it is not decremented
725 * when readers release their slots.
727 volatile unsigned mtb_numreaders;
730 /** The actual reader table definition. */
731 typedef struct MDB_txninfo {
734 #define mti_magic mt1.mtb.mtb_magic
735 #define mti_format mt1.mtb.mtb_format
736 #define mti_rmutex mt1.mtb.mtb_rmutex
737 #define mti_rmname mt1.mtb.mtb_rmname
738 #define mti_txnid mt1.mtb.mtb_txnid
739 #define mti_numreaders mt1.mtb.mtb_numreaders
740 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
743 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
744 char mt2_wmname[MNAME_LEN];
745 #define mti_wmname mt2.mt2_wmname
747 mdb_mutex_t mt2_wmutex;
748 #define mti_wmutex mt2.mt2_wmutex
750 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
752 MDB_reader mti_readers[1];
755 /** Lockfile format signature: version, features and field layout */
756 #define MDB_LOCK_FORMAT \
758 ((MDB_LOCK_VERSION) \
759 /* Flags which describe functionality */ \
760 + (((MDB_PIDLOCK) != 0) << 16)))
763 /** Common header for all page types.
764 * Overflow records occupy a number of contiguous pages with no
765 * headers on any page after the first.
767 typedef struct MDB_page {
768 #define mp_pgno mp_p.p_pgno
769 #define mp_next mp_p.p_next
771 pgno_t p_pgno; /**< page number */
772 struct MDB_page *p_next; /**< for in-memory list of freed pages */
775 /** @defgroup mdb_page Page Flags
777 * Flags for the page headers.
780 #define P_BRANCH 0x01 /**< branch page */
781 #define P_LEAF 0x02 /**< leaf page */
782 #define P_OVERFLOW 0x04 /**< overflow page */
783 #define P_META 0x08 /**< meta page */
784 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
785 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
786 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
787 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
788 #define P_KEEP 0x8000 /**< leave this page alone during spill */
790 uint16_t mp_flags; /**< @ref mdb_page */
791 #define mp_lower mp_pb.pb.pb_lower
792 #define mp_upper mp_pb.pb.pb_upper
793 #define mp_pages mp_pb.pb_pages
796 indx_t pb_lower; /**< lower bound of free space */
797 indx_t pb_upper; /**< upper bound of free space */
799 uint32_t pb_pages; /**< number of overflow pages */
801 indx_t mp_ptrs[1]; /**< dynamic size */
804 /** Size of the page header, excluding dynamic data at the end */
805 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
807 /** Address of first usable data byte in a page, after the header */
808 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
810 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
811 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
813 /** Number of nodes on a page */
814 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
816 /** The amount of space remaining in the page */
817 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
819 /** The percentage of space used in the page, in tenths of a percent. */
820 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
821 ((env)->me_psize - PAGEHDRSZ))
822 /** The minimum page fill factor, in tenths of a percent.
823 * Pages emptier than this are candidates for merging.
825 #define FILL_THRESHOLD 250
827 /** Test if a page is a leaf page */
828 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
829 /** Test if a page is a LEAF2 page */
830 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
831 /** Test if a page is a branch page */
832 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
833 /** Test if a page is an overflow page */
834 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
835 /** Test if a page is a sub page */
836 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
838 /** The number of overflow pages needed to store the given size. */
839 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
841 /** Link in #MDB_txn.%mt_loose_pgs list */
842 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
844 /** Header for a single key/data pair within a page.
845 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
846 * We guarantee 2-byte alignment for 'MDB_node's.
848 typedef struct MDB_node {
849 /** lo and hi are used for data size on leaf nodes and for
850 * child pgno on branch nodes. On 64 bit platforms, flags
851 * is also used for pgno. (Branch nodes have no flags).
852 * They are in host byte order in case that lets some
853 * accesses be optimized into a 32-bit word access.
855 #if BYTE_ORDER == LITTLE_ENDIAN
856 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
858 unsigned short mn_hi, mn_lo;
860 /** @defgroup mdb_node Node Flags
862 * Flags for node headers.
865 #define F_BIGDATA 0x01 /**< data put on overflow page */
866 #define F_SUBDATA 0x02 /**< data is a sub-database */
867 #define F_DUPDATA 0x04 /**< data has duplicates */
869 /** valid flags for #mdb_node_add() */
870 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
873 unsigned short mn_flags; /**< @ref mdb_node */
874 unsigned short mn_ksize; /**< key size */
875 char mn_data[1]; /**< key and data are appended here */
878 /** Size of the node header, excluding dynamic data at the end */
879 #define NODESIZE offsetof(MDB_node, mn_data)
881 /** Bit position of top word in page number, for shifting mn_flags */
882 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
884 /** Size of a node in a branch page with a given key.
885 * This is just the node header plus the key, there is no data.
887 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
889 /** Size of a node in a leaf page with a given key and data.
890 * This is node header plus key plus data size.
892 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
894 /** Address of node \b i in page \b p */
895 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
897 /** Address of the key for the node */
898 #define NODEKEY(node) (void *)((node)->mn_data)
900 /** Address of the data for a node */
901 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
903 /** Get the page number pointed to by a branch node */
904 #define NODEPGNO(node) \
905 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
906 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
907 /** Set the page number in a branch node */
908 #define SETPGNO(node,pgno) do { \
909 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
910 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
912 /** Get the size of the data in a leaf node */
913 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
914 /** Set the size of the data for a leaf node */
915 #define SETDSZ(node,size) do { \
916 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
917 /** The size of a key in a node */
918 #define NODEKSZ(node) ((node)->mn_ksize)
920 /** Copy a page number from src to dst */
922 #define COPY_PGNO(dst,src) dst = src
924 #if SIZE_MAX > 4294967295UL
925 #define COPY_PGNO(dst,src) do { \
926 unsigned short *s, *d; \
927 s = (unsigned short *)&(src); \
928 d = (unsigned short *)&(dst); \
935 #define COPY_PGNO(dst,src) do { \
936 unsigned short *s, *d; \
937 s = (unsigned short *)&(src); \
938 d = (unsigned short *)&(dst); \
944 /** The address of a key in a LEAF2 page.
945 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
946 * There are no node headers, keys are stored contiguously.
948 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
950 /** Set the \b node's key into \b keyptr, if requested. */
951 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
952 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
954 /** Set the \b node's key into \b key. */
955 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
957 /** Information about a single database in the environment. */
958 typedef struct MDB_db {
959 uint32_t md_pad; /**< also ksize for LEAF2 pages */
960 uint16_t md_flags; /**< @ref mdb_dbi_open */
961 uint16_t md_depth; /**< depth of this tree */
962 pgno_t md_branch_pages; /**< number of internal pages */
963 pgno_t md_leaf_pages; /**< number of leaf pages */
964 pgno_t md_overflow_pages; /**< number of overflow pages */
965 size_t md_entries; /**< number of data items */
966 pgno_t md_root; /**< the root page of this tree */
969 /** mdb_dbi_open flags */
970 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
971 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
972 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
973 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
975 /** Handle for the DB used to track free pages. */
977 /** Handle for the default DB. */
979 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
982 /** Number of meta pages - also hardcoded elsewhere */
985 /** Meta page content.
986 * A meta page is the start point for accessing a database snapshot.
987 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
989 typedef struct MDB_meta {
990 /** Stamp identifying this as an LMDB file. It must be set
993 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
995 void *mm_address; /**< address for fixed mapping */
996 size_t mm_mapsize; /**< size of mmap region */
997 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
998 /** The size of pages used in this DB */
999 #define mm_psize mm_dbs[FREE_DBI].md_pad
1000 /** Any persistent environment flags. @ref mdb_env */
1001 #define mm_flags mm_dbs[FREE_DBI].md_flags
1002 pgno_t mm_last_pg; /**< last used page in file */
1003 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1006 /** Buffer for a stack-allocated meta page.
1007 * The members define size and alignment, and silence type
1008 * aliasing warnings. They are not used directly; that could
1009 * mean incorrectly using several union members in parallel.
1011 typedef union MDB_metabuf {
1014 char mm_pad[PAGEHDRSZ];
1019 /** Auxiliary DB info.
1020 * The information here is mostly static/read-only. There is
1021 * only a single copy of this record in the environment.
1023 typedef struct MDB_dbx {
1024 MDB_val md_name; /**< name of the database */
1025 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1026 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1027 MDB_rel_func *md_rel; /**< user relocate function */
1028 void *md_relctx; /**< user-provided context for md_rel */
1031 /** A database transaction.
1032 * Every operation requires a transaction handle.
1035 MDB_txn *mt_parent; /**< parent of a nested txn */
1036 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1038 pgno_t mt_next_pgno; /**< next unallocated page */
1039 /** The ID of this transaction. IDs are integers incrementing from 1.
1040 * Only committed write transactions increment the ID. If a transaction
1041 * aborts, the ID may be re-used by the next writer.
1044 MDB_env *mt_env; /**< the DB environment */
1045 /** The list of pages that became unused during this transaction.
1047 MDB_IDL mt_free_pgs;
1048 /** The list of loose pages that became unused and may be reused
1049 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1051 MDB_page *mt_loose_pgs;
1052 /* #Number of loose pages (#mt_loose_pgs) */
1054 /** The sorted list of dirty pages we temporarily wrote to disk
1055 * because the dirty list was full. page numbers in here are
1056 * shifted left by 1, deleted slots have the LSB set.
1058 MDB_IDL mt_spill_pgs;
1060 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1061 MDB_ID2L dirty_list;
1062 /** For read txns: This thread/txn's reader table slot, or NULL. */
1065 /** Array of records for each DB known in the environment. */
1067 /** Array of MDB_db records for each known DB */
1069 /** Array of sequence numbers for each DB handle */
1070 unsigned int *mt_dbiseqs;
1071 /** @defgroup mt_dbflag Transaction DB Flags
1075 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1076 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1077 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1078 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1079 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1081 /** In write txns, array of cursors for each DB */
1082 MDB_cursor **mt_cursors;
1083 /** Array of flags for each DB */
1084 unsigned char *mt_dbflags;
1085 /** Number of DB records in use, or 0 when the txn is finished.
1086 * This number only ever increments until the txn finishes; we
1087 * don't decrement it when individual DB handles are closed.
1091 /** @defgroup mdb_txn Transaction Flags
1095 /** #mdb_txn_begin() flags */
1096 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1097 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1098 /* internal txn flags */
1099 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1100 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1101 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1102 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1103 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1104 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1105 /** most operations on the txn are currently illegal */
1106 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1108 unsigned int mt_flags; /**< @ref mdb_txn */
1109 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1110 * Includes ancestor txns' dirty pages not hidden by other txns'
1111 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1112 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1114 unsigned int mt_dirty_room;
1117 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1118 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1119 * raise this on a 64 bit machine.
1121 #define CURSOR_STACK 32
1125 /** Cursors are used for all DB operations.
1126 * A cursor holds a path of (page pointer, key index) from the DB
1127 * root to a position in the DB, plus other state. #MDB_DUPSORT
1128 * cursors include an xcursor to the current data item. Write txns
1129 * track their cursors and keep them up to date when data moves.
1130 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1131 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1134 /** Next cursor on this DB in this txn */
1135 MDB_cursor *mc_next;
1136 /** Backup of the original cursor if this cursor is a shadow */
1137 MDB_cursor *mc_backup;
1138 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1139 struct MDB_xcursor *mc_xcursor;
1140 /** The transaction that owns this cursor */
1142 /** The database handle this cursor operates on */
1144 /** The database record for this cursor */
1146 /** The database auxiliary record for this cursor */
1148 /** The @ref mt_dbflag for this database */
1149 unsigned char *mc_dbflag;
1150 unsigned short mc_snum; /**< number of pushed pages */
1151 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1152 /** @defgroup mdb_cursor Cursor Flags
1154 * Cursor state flags.
1157 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1158 #define C_EOF 0x02 /**< No more data */
1159 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1160 #define C_DEL 0x08 /**< last op was a cursor_del */
1161 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1162 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1164 unsigned int mc_flags; /**< @ref mdb_cursor */
1165 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1166 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1169 /** Context for sorted-dup records.
1170 * We could have gone to a fully recursive design, with arbitrarily
1171 * deep nesting of sub-databases. But for now we only handle these
1172 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1174 typedef struct MDB_xcursor {
1175 /** A sub-cursor for traversing the Dup DB */
1176 MDB_cursor mx_cursor;
1177 /** The database record for this Dup DB */
1179 /** The auxiliary DB record for this Dup DB */
1181 /** The @ref mt_dbflag for this Dup DB */
1182 unsigned char mx_dbflag;
1185 /** State of FreeDB old pages, stored in the MDB_env */
1186 typedef struct MDB_pgstate {
1187 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1188 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1191 /** The database environment. */
1193 HANDLE me_fd; /**< The main data file */
1194 HANDLE me_lfd; /**< The lock file */
1195 HANDLE me_mfd; /**< just for writing the meta pages */
1196 /** Failed to update the meta page. Probably an I/O error. */
1197 #define MDB_FATAL_ERROR 0x80000000U
1198 /** Some fields are initialized. */
1199 #define MDB_ENV_ACTIVE 0x20000000U
1200 /** me_txkey is set */
1201 #define MDB_ENV_TXKEY 0x10000000U
1202 /** fdatasync is unreliable */
1203 #define MDB_FSYNCONLY 0x08000000U
1204 uint32_t me_flags; /**< @ref mdb_env */
1205 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1206 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1207 unsigned int me_maxreaders; /**< size of the reader table */
1208 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1209 volatile int me_close_readers;
1210 MDB_dbi me_numdbs; /**< number of DBs opened */
1211 MDB_dbi me_maxdbs; /**< size of the DB table */
1212 MDB_PID_T me_pid; /**< process ID of this env */
1213 char *me_path; /**< path to the DB files */
1214 char *me_map; /**< the memory map of the data file */
1215 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1216 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1217 void *me_pbuf; /**< scratch area for DUPSORT put() */
1218 MDB_txn *me_txn; /**< current write transaction */
1219 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1220 size_t me_mapsize; /**< size of the data memory map */
1221 off_t me_size; /**< current file size */
1222 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1223 MDB_dbx *me_dbxs; /**< array of static DB info */
1224 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1225 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1226 pthread_key_t me_txkey; /**< thread-key for readers */
1227 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1228 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1229 # define me_pglast me_pgstate.mf_pglast
1230 # define me_pghead me_pgstate.mf_pghead
1231 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1232 /** IDL of pages that became unused in a write txn */
1233 MDB_IDL me_free_pgs;
1234 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1235 MDB_ID2L me_dirty_list;
1236 /** Max number of freelist items that can fit in a single overflow page */
1238 /** Max size of a node on a page */
1239 unsigned int me_nodemax;
1240 #if !(MDB_MAXKEYSIZE)
1241 unsigned int me_maxkey; /**< max size of a key */
1243 int me_live_reader; /**< have liveness lock in reader table */
1245 int me_pidquery; /**< Used in OpenProcess */
1247 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1248 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1249 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1251 mdb_mutex_t me_rmutex;
1252 mdb_mutex_t me_wmutex;
1254 void *me_userctx; /**< User-settable context */
1255 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1258 /** Nested transaction */
1259 typedef struct MDB_ntxn {
1260 MDB_txn mnt_txn; /**< the transaction */
1261 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1264 /** max number of pages to commit in one writev() call */
1265 #define MDB_COMMIT_PAGES 64
1266 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1267 #undef MDB_COMMIT_PAGES
1268 #define MDB_COMMIT_PAGES IOV_MAX
1271 /** max bytes to write in one call */
1272 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1274 /** Check \b txn and \b dbi arguments to a function */
1275 #define TXN_DBI_EXIST(txn, dbi, validity) \
1276 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1278 /** Check for misused \b dbi handles */
1279 #define TXN_DBI_CHANGED(txn, dbi) \
1280 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1282 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1283 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1284 static int mdb_page_touch(MDB_cursor *mc);
1286 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1287 "reset-tmp", "fail-begin", "fail-beginchild"}
1289 /* mdb_txn_end operation number, for logging */
1290 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1291 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1293 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1294 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1295 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1296 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1297 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1299 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1300 static int mdb_page_search_root(MDB_cursor *mc,
1301 MDB_val *key, int modify);
1302 #define MDB_PS_MODIFY 1
1303 #define MDB_PS_ROOTONLY 2
1304 #define MDB_PS_FIRST 4
1305 #define MDB_PS_LAST 8
1306 static int mdb_page_search(MDB_cursor *mc,
1307 MDB_val *key, int flags);
1308 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1310 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1311 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1312 pgno_t newpgno, unsigned int nflags);
1314 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1315 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1316 static int mdb_env_write_meta(MDB_txn *txn);
1317 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1318 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1320 static void mdb_env_close0(MDB_env *env, int excl);
1322 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1323 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1324 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1325 static void mdb_node_del(MDB_cursor *mc, int ksize);
1326 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1327 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1328 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1329 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1330 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1332 static int mdb_rebalance(MDB_cursor *mc);
1333 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1335 static void mdb_cursor_pop(MDB_cursor *mc);
1336 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1338 static int mdb_cursor_del0(MDB_cursor *mc);
1339 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1340 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1341 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1342 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1343 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1345 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1346 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1348 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1349 static void mdb_xcursor_init0(MDB_cursor *mc);
1350 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1351 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1353 static int mdb_drop0(MDB_cursor *mc, int subs);
1354 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1355 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1358 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1361 /** Compare two items pointing at size_t's of unknown alignment. */
1362 #ifdef MISALIGNED_OK
1363 # define mdb_cmp_clong mdb_cmp_long
1365 # define mdb_cmp_clong mdb_cmp_cint
1369 static SECURITY_DESCRIPTOR mdb_null_sd;
1370 static SECURITY_ATTRIBUTES mdb_all_sa;
1371 static int mdb_sec_inited;
1374 /** Return the library version info. */
1376 mdb_version(int *major, int *minor, int *patch)
1378 if (major) *major = MDB_VERSION_MAJOR;
1379 if (minor) *minor = MDB_VERSION_MINOR;
1380 if (patch) *patch = MDB_VERSION_PATCH;
1381 return MDB_VERSION_STRING;
1384 /** Table of descriptions for LMDB @ref errors */
1385 static char *const mdb_errstr[] = {
1386 "MDB_KEYEXIST: Key/data pair already exists",
1387 "MDB_NOTFOUND: No matching key/data pair found",
1388 "MDB_PAGE_NOTFOUND: Requested page not found",
1389 "MDB_CORRUPTED: Located page was wrong type",
1390 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1391 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1392 "MDB_INVALID: File is not an LMDB file",
1393 "MDB_MAP_FULL: Environment mapsize limit reached",
1394 "MDB_DBS_FULL: Environment maxdbs limit reached",
1395 "MDB_READERS_FULL: Environment maxreaders limit reached",
1396 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1397 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1398 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1399 "MDB_PAGE_FULL: Internal error - page has no more space",
1400 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1401 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1402 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1403 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1404 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1405 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1409 mdb_strerror(int err)
1412 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1413 * This works as long as no function between the call to mdb_strerror
1414 * and the actual use of the message uses more than 4K of stack.
1417 char buf[1024], *ptr = buf;
1421 return ("Successful return: 0");
1423 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1424 i = err - MDB_KEYEXIST;
1425 return mdb_errstr[i];
1429 /* These are the C-runtime error codes we use. The comment indicates
1430 * their numeric value, and the Win32 error they would correspond to
1431 * if the error actually came from a Win32 API. A major mess, we should
1432 * have used LMDB-specific error codes for everything.
1435 case ENOENT: /* 2, FILE_NOT_FOUND */
1436 case EIO: /* 5, ACCESS_DENIED */
1437 case ENOMEM: /* 12, INVALID_ACCESS */
1438 case EACCES: /* 13, INVALID_DATA */
1439 case EBUSY: /* 16, CURRENT_DIRECTORY */
1440 case EINVAL: /* 22, BAD_COMMAND */
1441 case ENOSPC: /* 28, OUT_OF_PAPER */
1442 return strerror(err);
1447 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1448 FORMAT_MESSAGE_IGNORE_INSERTS,
1449 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1452 return strerror(err);
1456 /** assert(3) variant in cursor context */
1457 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1458 /** assert(3) variant in transaction context */
1459 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1460 /** assert(3) variant in environment context */
1461 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1464 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1465 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1468 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1469 const char *func, const char *file, int line)
1472 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1473 file, line, expr_txt, func);
1474 if (env->me_assert_func)
1475 env->me_assert_func(env, buf);
1476 fprintf(stderr, "%s\n", buf);
1480 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1484 /** Return the page number of \b mp which may be sub-page, for debug output */
1486 mdb_dbg_pgno(MDB_page *mp)
1489 COPY_PGNO(ret, mp->mp_pgno);
1493 /** Display a key in hexadecimal and return the address of the result.
1494 * @param[in] key the key to display
1495 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1496 * @return The key in hexadecimal form.
1499 mdb_dkey(MDB_val *key, char *buf)
1502 unsigned char *c = key->mv_data;
1508 if (key->mv_size > DKBUF_MAXKEYSIZE)
1509 return "MDB_MAXKEYSIZE";
1510 /* may want to make this a dynamic check: if the key is mostly
1511 * printable characters, print it as-is instead of converting to hex.
1515 for (i=0; i<key->mv_size; i++)
1516 ptr += sprintf(ptr, "%02x", *c++);
1518 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1524 mdb_leafnode_type(MDB_node *n)
1526 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1527 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1528 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1531 /** Display all the keys in the page. */
1533 mdb_page_list(MDB_page *mp)
1535 pgno_t pgno = mdb_dbg_pgno(mp);
1536 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1538 unsigned int i, nkeys, nsize, total = 0;
1542 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1543 case P_BRANCH: type = "Branch page"; break;
1544 case P_LEAF: type = "Leaf page"; break;
1545 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1546 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1547 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1549 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1550 pgno, mp->mp_pages, state);
1553 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1554 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1557 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1561 nkeys = NUMKEYS(mp);
1562 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1564 for (i=0; i<nkeys; i++) {
1565 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1566 key.mv_size = nsize = mp->mp_pad;
1567 key.mv_data = LEAF2KEY(mp, i, nsize);
1569 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1572 node = NODEPTR(mp, i);
1573 key.mv_size = node->mn_ksize;
1574 key.mv_data = node->mn_data;
1575 nsize = NODESIZE + key.mv_size;
1576 if (IS_BRANCH(mp)) {
1577 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1581 if (F_ISSET(node->mn_flags, F_BIGDATA))
1582 nsize += sizeof(pgno_t);
1584 nsize += NODEDSZ(node);
1586 nsize += sizeof(indx_t);
1587 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1588 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1590 total = EVEN(total);
1592 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1593 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1597 mdb_cursor_chk(MDB_cursor *mc)
1603 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1604 for (i=0; i<mc->mc_top; i++) {
1606 node = NODEPTR(mp, mc->mc_ki[i]);
1607 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1610 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1616 /** Count all the pages in each DB and in the freelist
1617 * and make sure it matches the actual number of pages
1619 * All named DBs must be open for a correct count.
1621 static void mdb_audit(MDB_txn *txn)
1625 MDB_ID freecount, count;
1630 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1631 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1632 freecount += *(MDB_ID *)data.mv_data;
1633 mdb_tassert(txn, rc == MDB_NOTFOUND);
1636 for (i = 0; i<txn->mt_numdbs; i++) {
1638 if (!(txn->mt_dbflags[i] & DB_VALID))
1640 mdb_cursor_init(&mc, txn, i, &mx);
1641 if (txn->mt_dbs[i].md_root == P_INVALID)
1643 count += txn->mt_dbs[i].md_branch_pages +
1644 txn->mt_dbs[i].md_leaf_pages +
1645 txn->mt_dbs[i].md_overflow_pages;
1646 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1647 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1648 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1651 mp = mc.mc_pg[mc.mc_top];
1652 for (j=0; j<NUMKEYS(mp); j++) {
1653 MDB_node *leaf = NODEPTR(mp, j);
1654 if (leaf->mn_flags & F_SUBDATA) {
1656 memcpy(&db, NODEDATA(leaf), sizeof(db));
1657 count += db.md_branch_pages + db.md_leaf_pages +
1658 db.md_overflow_pages;
1662 mdb_tassert(txn, rc == MDB_NOTFOUND);
1665 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1666 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1667 txn->mt_txnid, freecount, count+NUM_METAS,
1668 freecount+count+NUM_METAS, txn->mt_next_pgno);
1674 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1676 return txn->mt_dbxs[dbi].md_cmp(a, b);
1680 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1682 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1683 #if UINT_MAX < SIZE_MAX
1684 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1685 dcmp = mdb_cmp_clong;
1690 /** Allocate memory for a page.
1691 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1694 mdb_page_malloc(MDB_txn *txn, unsigned num)
1696 MDB_env *env = txn->mt_env;
1697 MDB_page *ret = env->me_dpages;
1698 size_t psize = env->me_psize, sz = psize, off;
1699 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1700 * For a single page alloc, we init everything after the page header.
1701 * For multi-page, we init the final page; if the caller needed that
1702 * many pages they will be filling in at least up to the last page.
1706 VGMEMP_ALLOC(env, ret, sz);
1707 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1708 env->me_dpages = ret->mp_next;
1711 psize -= off = PAGEHDRSZ;
1716 if ((ret = malloc(sz)) != NULL) {
1717 VGMEMP_ALLOC(env, ret, sz);
1718 if (!(env->me_flags & MDB_NOMEMINIT)) {
1719 memset((char *)ret + off, 0, psize);
1723 txn->mt_flags |= MDB_TXN_ERROR;
1727 /** Free a single page.
1728 * Saves single pages to a list, for future reuse.
1729 * (This is not used for multi-page overflow pages.)
1732 mdb_page_free(MDB_env *env, MDB_page *mp)
1734 mp->mp_next = env->me_dpages;
1735 VGMEMP_FREE(env, mp);
1736 env->me_dpages = mp;
1739 /** Free a dirty page */
1741 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1743 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1744 mdb_page_free(env, dp);
1746 /* large pages just get freed directly */
1747 VGMEMP_FREE(env, dp);
1752 /** Return all dirty pages to dpage list */
1754 mdb_dlist_free(MDB_txn *txn)
1756 MDB_env *env = txn->mt_env;
1757 MDB_ID2L dl = txn->mt_u.dirty_list;
1758 unsigned i, n = dl[0].mid;
1760 for (i = 1; i <= n; i++) {
1761 mdb_dpage_free(env, dl[i].mptr);
1766 /** Loosen or free a single page.
1767 * Saves single pages to a list for future reuse
1768 * in this same txn. It has been pulled from the freeDB
1769 * and already resides on the dirty list, but has been
1770 * deleted. Use these pages first before pulling again
1773 * If the page wasn't dirtied in this txn, just add it
1774 * to this txn's free list.
1777 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1780 pgno_t pgno = mp->mp_pgno;
1781 MDB_txn *txn = mc->mc_txn;
1783 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1784 if (txn->mt_parent) {
1785 MDB_ID2 *dl = txn->mt_u.dirty_list;
1786 /* If txn has a parent, make sure the page is in our
1790 unsigned x = mdb_mid2l_search(dl, pgno);
1791 if (x <= dl[0].mid && dl[x].mid == pgno) {
1792 if (mp != dl[x].mptr) { /* bad cursor? */
1793 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1794 txn->mt_flags |= MDB_TXN_ERROR;
1795 return MDB_CORRUPTED;
1802 /* no parent txn, so it's just ours */
1807 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1809 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1810 txn->mt_loose_pgs = mp;
1811 txn->mt_loose_count++;
1812 mp->mp_flags |= P_LOOSE;
1814 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1822 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1823 * @param[in] mc A cursor handle for the current operation.
1824 * @param[in] pflags Flags of the pages to update:
1825 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1826 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1827 * @return 0 on success, non-zero on failure.
1830 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1832 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1833 MDB_txn *txn = mc->mc_txn;
1839 int rc = MDB_SUCCESS, level;
1841 /* Mark pages seen by cursors */
1842 if (mc->mc_flags & C_UNTRACK)
1843 mc = NULL; /* will find mc in mt_cursors */
1844 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1845 for (; mc; mc=mc->mc_next) {
1846 if (!(mc->mc_flags & C_INITIALIZED))
1848 for (m3 = mc;; m3 = &mx->mx_cursor) {
1850 for (j=0; j<m3->mc_snum; j++) {
1852 if ((mp->mp_flags & Mask) == pflags)
1853 mp->mp_flags ^= P_KEEP;
1855 mx = m3->mc_xcursor;
1856 /* Proceed to mx if it is at a sub-database */
1857 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1859 if (! (mp && (mp->mp_flags & P_LEAF)))
1861 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1862 if (!(leaf->mn_flags & F_SUBDATA))
1871 /* Mark dirty root pages */
1872 for (i=0; i<txn->mt_numdbs; i++) {
1873 if (txn->mt_dbflags[i] & DB_DIRTY) {
1874 pgno_t pgno = txn->mt_dbs[i].md_root;
1875 if (pgno == P_INVALID)
1877 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1879 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1880 dp->mp_flags ^= P_KEEP;
1888 static int mdb_page_flush(MDB_txn *txn, int keep);
1890 /** Spill pages from the dirty list back to disk.
1891 * This is intended to prevent running into #MDB_TXN_FULL situations,
1892 * but note that they may still occur in a few cases:
1893 * 1) our estimate of the txn size could be too small. Currently this
1894 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1895 * 2) child txns may run out of space if their parents dirtied a
1896 * lot of pages and never spilled them. TODO: we probably should do
1897 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1898 * the parent's dirty_room is below a given threshold.
1900 * Otherwise, if not using nested txns, it is expected that apps will
1901 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1902 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1903 * If the txn never references them again, they can be left alone.
1904 * If the txn only reads them, they can be used without any fuss.
1905 * If the txn writes them again, they can be dirtied immediately without
1906 * going thru all of the work of #mdb_page_touch(). Such references are
1907 * handled by #mdb_page_unspill().
1909 * Also note, we never spill DB root pages, nor pages of active cursors,
1910 * because we'll need these back again soon anyway. And in nested txns,
1911 * we can't spill a page in a child txn if it was already spilled in a
1912 * parent txn. That would alter the parent txns' data even though
1913 * the child hasn't committed yet, and we'd have no way to undo it if
1914 * the child aborted.
1916 * @param[in] m0 cursor A cursor handle identifying the transaction and
1917 * database for which we are checking space.
1918 * @param[in] key For a put operation, the key being stored.
1919 * @param[in] data For a put operation, the data being stored.
1920 * @return 0 on success, non-zero on failure.
1923 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1925 MDB_txn *txn = m0->mc_txn;
1927 MDB_ID2L dl = txn->mt_u.dirty_list;
1928 unsigned int i, j, need;
1931 if (m0->mc_flags & C_SUB)
1934 /* Estimate how much space this op will take */
1935 i = m0->mc_db->md_depth;
1936 /* Named DBs also dirty the main DB */
1937 if (m0->mc_dbi >= CORE_DBS)
1938 i += txn->mt_dbs[MAIN_DBI].md_depth;
1939 /* For puts, roughly factor in the key+data size */
1941 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1942 i += i; /* double it for good measure */
1945 if (txn->mt_dirty_room > i)
1948 if (!txn->mt_spill_pgs) {
1949 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1950 if (!txn->mt_spill_pgs)
1953 /* purge deleted slots */
1954 MDB_IDL sl = txn->mt_spill_pgs;
1955 unsigned int num = sl[0];
1957 for (i=1; i<=num; i++) {
1964 /* Preserve pages which may soon be dirtied again */
1965 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1968 /* Less aggressive spill - we originally spilled the entire dirty list,
1969 * with a few exceptions for cursor pages and DB root pages. But this
1970 * turns out to be a lot of wasted effort because in a large txn many
1971 * of those pages will need to be used again. So now we spill only 1/8th
1972 * of the dirty pages. Testing revealed this to be a good tradeoff,
1973 * better than 1/2, 1/4, or 1/10.
1975 if (need < MDB_IDL_UM_MAX / 8)
1976 need = MDB_IDL_UM_MAX / 8;
1978 /* Save the page IDs of all the pages we're flushing */
1979 /* flush from the tail forward, this saves a lot of shifting later on. */
1980 for (i=dl[0].mid; i && need; i--) {
1981 MDB_ID pn = dl[i].mid << 1;
1983 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1985 /* Can't spill twice, make sure it's not already in a parent's
1988 if (txn->mt_parent) {
1990 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1991 if (tx2->mt_spill_pgs) {
1992 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1993 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1994 dp->mp_flags |= P_KEEP;
2002 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2006 mdb_midl_sort(txn->mt_spill_pgs);
2008 /* Flush the spilled part of dirty list */
2009 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2012 /* Reset any dirty pages we kept that page_flush didn't see */
2013 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2016 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2020 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2022 mdb_find_oldest(MDB_txn *txn)
2025 txnid_t mr, oldest = txn->mt_txnid - 1;
2026 if (txn->mt_env->me_txns) {
2027 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2028 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2039 /** Add a page to the txn's dirty list */
2041 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2044 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2046 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2047 insert = mdb_mid2l_append;
2049 insert = mdb_mid2l_insert;
2051 mid.mid = mp->mp_pgno;
2053 rc = insert(txn->mt_u.dirty_list, &mid);
2054 mdb_tassert(txn, rc == 0);
2055 txn->mt_dirty_room--;
2058 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2059 * me_pghead and mt_next_pgno.
2061 * If there are free pages available from older transactions, they
2062 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2063 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2064 * and move me_pglast to say which records were consumed. Only this
2065 * function can create me_pghead and move me_pglast/mt_next_pgno.
2066 * @param[in] mc cursor A cursor handle identifying the transaction and
2067 * database for which we are allocating.
2068 * @param[in] num the number of pages to allocate.
2069 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2070 * will always be satisfied by a single contiguous chunk of memory.
2071 * @return 0 on success, non-zero on failure.
2074 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2076 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2077 /* Get at most <Max_retries> more freeDB records once me_pghead
2078 * has enough pages. If not enough, use new pages from the map.
2079 * If <Paranoid> and mc is updating the freeDB, only get new
2080 * records if me_pghead is empty. Then the freelist cannot play
2081 * catch-up with itself by growing while trying to save it.
2083 enum { Paranoid = 1, Max_retries = 500 };
2085 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2087 int rc, retry = num * 60;
2088 MDB_txn *txn = mc->mc_txn;
2089 MDB_env *env = txn->mt_env;
2090 pgno_t pgno, *mop = env->me_pghead;
2091 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2093 txnid_t oldest = 0, last;
2098 /* If there are any loose pages, just use them */
2099 if (num == 1 && txn->mt_loose_pgs) {
2100 np = txn->mt_loose_pgs;
2101 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2102 txn->mt_loose_count--;
2103 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2111 /* If our dirty list is already full, we can't do anything */
2112 if (txn->mt_dirty_room == 0) {
2117 for (op = MDB_FIRST;; op = MDB_NEXT) {
2122 /* Seek a big enough contiguous page range. Prefer
2123 * pages at the tail, just truncating the list.
2129 if (mop[i-n2] == pgno+n2)
2136 if (op == MDB_FIRST) { /* 1st iteration */
2137 /* Prepare to fetch more and coalesce */
2138 last = env->me_pglast;
2139 oldest = env->me_pgoldest;
2140 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2143 key.mv_data = &last; /* will look up last+1 */
2144 key.mv_size = sizeof(last);
2146 if (Paranoid && mc->mc_dbi == FREE_DBI)
2149 if (Paranoid && retry < 0 && mop_len)
2153 /* Do not fetch more if the record will be too recent */
2154 if (oldest <= last) {
2156 oldest = mdb_find_oldest(txn);
2157 env->me_pgoldest = oldest;
2163 rc = mdb_cursor_get(&m2, &key, NULL, op);
2165 if (rc == MDB_NOTFOUND)
2169 last = *(txnid_t*)key.mv_data;
2170 if (oldest <= last) {
2172 oldest = mdb_find_oldest(txn);
2173 env->me_pgoldest = oldest;
2179 np = m2.mc_pg[m2.mc_top];
2180 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2181 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2184 idl = (MDB_ID *) data.mv_data;
2187 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2192 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2194 mop = env->me_pghead;
2196 env->me_pglast = last;
2198 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2199 last, txn->mt_dbs[FREE_DBI].md_root, i));
2201 DPRINTF(("IDL %"Z"u", idl[j]));
2203 /* Merge in descending sorted order */
2204 mdb_midl_xmerge(mop, idl);
2208 /* Use new pages from the map when nothing suitable in the freeDB */
2210 pgno = txn->mt_next_pgno;
2211 if (pgno + num >= env->me_maxpg) {
2212 DPUTS("DB size maxed out");
2218 if (env->me_flags & MDB_WRITEMAP) {
2219 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2221 if (!(np = mdb_page_malloc(txn, num))) {
2227 mop[0] = mop_len -= num;
2228 /* Move any stragglers down */
2229 for (j = i-num; j < mop_len; )
2230 mop[++j] = mop[++i];
2232 txn->mt_next_pgno = pgno + num;
2235 mdb_page_dirty(txn, np);
2241 txn->mt_flags |= MDB_TXN_ERROR;
2245 /** Copy the used portions of a non-overflow page.
2246 * @param[in] dst page to copy into
2247 * @param[in] src page to copy from
2248 * @param[in] psize size of a page
2251 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2253 enum { Align = sizeof(pgno_t) };
2254 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2256 /* If page isn't full, just copy the used portion. Adjust
2257 * alignment so memcpy may copy words instead of bytes.
2259 if ((unused &= -Align) && !IS_LEAF2(src)) {
2260 upper = (upper + PAGEBASE) & -Align;
2261 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2262 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2265 memcpy(dst, src, psize - unused);
2269 /** Pull a page off the txn's spill list, if present.
2270 * If a page being referenced was spilled to disk in this txn, bring
2271 * it back and make it dirty/writable again.
2272 * @param[in] txn the transaction handle.
2273 * @param[in] mp the page being referenced. It must not be dirty.
2274 * @param[out] ret the writable page, if any. ret is unchanged if
2275 * mp wasn't spilled.
2278 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2280 MDB_env *env = txn->mt_env;
2283 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2285 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2286 if (!tx2->mt_spill_pgs)
2288 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2289 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2292 if (txn->mt_dirty_room == 0)
2293 return MDB_TXN_FULL;
2294 if (IS_OVERFLOW(mp))
2298 if (env->me_flags & MDB_WRITEMAP) {
2301 np = mdb_page_malloc(txn, num);
2305 memcpy(np, mp, num * env->me_psize);
2307 mdb_page_copy(np, mp, env->me_psize);
2310 /* If in current txn, this page is no longer spilled.
2311 * If it happens to be the last page, truncate the spill list.
2312 * Otherwise mark it as deleted by setting the LSB.
2314 if (x == txn->mt_spill_pgs[0])
2315 txn->mt_spill_pgs[0]--;
2317 txn->mt_spill_pgs[x] |= 1;
2318 } /* otherwise, if belonging to a parent txn, the
2319 * page remains spilled until child commits
2322 mdb_page_dirty(txn, np);
2323 np->mp_flags |= P_DIRTY;
2331 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2332 * @param[in] mc cursor pointing to the page to be touched
2333 * @return 0 on success, non-zero on failure.
2336 mdb_page_touch(MDB_cursor *mc)
2338 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2339 MDB_txn *txn = mc->mc_txn;
2340 MDB_cursor *m2, *m3;
2344 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2345 if (txn->mt_flags & MDB_TXN_SPILLS) {
2347 rc = mdb_page_unspill(txn, mp, &np);
2353 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2354 (rc = mdb_page_alloc(mc, 1, &np)))
2357 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2358 mp->mp_pgno, pgno));
2359 mdb_cassert(mc, mp->mp_pgno != pgno);
2360 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2361 /* Update the parent page, if any, to point to the new page */
2363 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2364 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2365 SETPGNO(node, pgno);
2367 mc->mc_db->md_root = pgno;
2369 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2370 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2372 /* If txn has a parent, make sure the page is in our
2376 unsigned x = mdb_mid2l_search(dl, pgno);
2377 if (x <= dl[0].mid && dl[x].mid == pgno) {
2378 if (mp != dl[x].mptr) { /* bad cursor? */
2379 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2380 txn->mt_flags |= MDB_TXN_ERROR;
2381 return MDB_CORRUPTED;
2386 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2388 np = mdb_page_malloc(txn, 1);
2393 rc = mdb_mid2l_insert(dl, &mid);
2394 mdb_cassert(mc, rc == 0);
2399 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2401 np->mp_flags |= P_DIRTY;
2404 /* Adjust cursors pointing to mp */
2405 mc->mc_pg[mc->mc_top] = np;
2406 m2 = txn->mt_cursors[mc->mc_dbi];
2407 if (mc->mc_flags & C_SUB) {
2408 for (; m2; m2=m2->mc_next) {
2409 m3 = &m2->mc_xcursor->mx_cursor;
2410 if (m3->mc_snum < mc->mc_snum) continue;
2411 if (m3->mc_pg[mc->mc_top] == mp)
2412 m3->mc_pg[mc->mc_top] = np;
2415 for (; m2; m2=m2->mc_next) {
2416 if (m2->mc_snum < mc->mc_snum) continue;
2417 if (m2->mc_pg[mc->mc_top] == mp) {
2418 m2->mc_pg[mc->mc_top] = np;
2419 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2421 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2423 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2424 if (!(leaf->mn_flags & F_SUBDATA))
2425 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2433 txn->mt_flags |= MDB_TXN_ERROR;
2438 mdb_env_sync(MDB_env *env, int force)
2441 if (env->me_flags & MDB_RDONLY)
2443 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2444 if (env->me_flags & MDB_WRITEMAP) {
2445 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2446 ? MS_ASYNC : MS_SYNC;
2447 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2450 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2454 #ifdef BROKEN_FDATASYNC
2455 if (env->me_flags & MDB_FSYNCONLY) {
2456 if (fsync(env->me_fd))
2460 if (MDB_FDATASYNC(env->me_fd))
2467 /** Back up parent txn's cursors, then grab the originals for tracking */
2469 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2471 MDB_cursor *mc, *bk;
2476 for (i = src->mt_numdbs; --i >= 0; ) {
2477 if ((mc = src->mt_cursors[i]) != NULL) {
2478 size = sizeof(MDB_cursor);
2480 size += sizeof(MDB_xcursor);
2481 for (; mc; mc = bk->mc_next) {
2487 mc->mc_db = &dst->mt_dbs[i];
2488 /* Kill pointers into src - and dst to reduce abuse: The
2489 * user may not use mc until dst ends. Otherwise we'd...
2491 mc->mc_txn = NULL; /* ...set this to dst */
2492 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2493 if ((mx = mc->mc_xcursor) != NULL) {
2494 *(MDB_xcursor *)(bk+1) = *mx;
2495 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2497 mc->mc_next = dst->mt_cursors[i];
2498 dst->mt_cursors[i] = mc;
2505 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2506 * @param[in] txn the transaction handle.
2507 * @param[in] merge true to keep changes to parent cursors, false to revert.
2508 * @return 0 on success, non-zero on failure.
2511 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2513 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2517 for (i = txn->mt_numdbs; --i >= 0; ) {
2518 for (mc = cursors[i]; mc; mc = next) {
2520 if ((bk = mc->mc_backup) != NULL) {
2522 /* Commit changes to parent txn */
2523 mc->mc_next = bk->mc_next;
2524 mc->mc_backup = bk->mc_backup;
2525 mc->mc_txn = bk->mc_txn;
2526 mc->mc_db = bk->mc_db;
2527 mc->mc_dbflag = bk->mc_dbflag;
2528 if ((mx = mc->mc_xcursor) != NULL)
2529 mx->mx_cursor.mc_txn = bk->mc_txn;
2531 /* Abort nested txn */
2533 if ((mx = mc->mc_xcursor) != NULL)
2534 *mx = *(MDB_xcursor *)(bk+1);
2538 /* Only malloced cursors are permanently tracked. */
2545 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2551 Pidset = F_SETLK, Pidcheck = F_GETLK
2555 /** Set or check a pid lock. Set returns 0 on success.
2556 * Check returns 0 if the process is certainly dead, nonzero if it may
2557 * be alive (the lock exists or an error happened so we do not know).
2559 * On Windows Pidset is a no-op, we merely check for the existence
2560 * of the process with the given pid. On POSIX we use a single byte
2561 * lock on the lockfile, set at an offset equal to the pid.
2564 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2566 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2569 if (op == Pidcheck) {
2570 h = OpenProcess(env->me_pidquery, FALSE, pid);
2571 /* No documented "no such process" code, but other program use this: */
2573 return ErrCode() != ERROR_INVALID_PARAMETER;
2574 /* A process exists until all handles to it close. Has it exited? */
2575 ret = WaitForSingleObject(h, 0) != 0;
2582 struct flock lock_info;
2583 memset(&lock_info, 0, sizeof(lock_info));
2584 lock_info.l_type = F_WRLCK;
2585 lock_info.l_whence = SEEK_SET;
2586 lock_info.l_start = pid;
2587 lock_info.l_len = 1;
2588 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2589 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2591 } else if ((rc = ErrCode()) == EINTR) {
2599 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2600 * @param[in] txn the transaction handle to initialize
2601 * @return 0 on success, non-zero on failure.
2604 mdb_txn_renew0(MDB_txn *txn)
2606 MDB_env *env = txn->mt_env;
2607 MDB_txninfo *ti = env->me_txns;
2609 unsigned int i, nr, flags = txn->mt_flags;
2611 int rc, new_notls = 0;
2613 if ((flags &= MDB_TXN_RDONLY) != 0) {
2615 meta = mdb_env_pick_meta(env);
2616 txn->mt_txnid = meta->mm_txnid;
2617 txn->mt_u.reader = NULL;
2619 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2620 pthread_getspecific(env->me_txkey);
2622 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2623 return MDB_BAD_RSLOT;
2625 MDB_PID_T pid = env->me_pid;
2626 MDB_THR_T tid = pthread_self();
2627 mdb_mutexref_t rmutex = env->me_rmutex;
2629 if (!env->me_live_reader) {
2630 rc = mdb_reader_pid(env, Pidset, pid);
2633 env->me_live_reader = 1;
2636 if (LOCK_MUTEX(rc, env, rmutex))
2638 nr = ti->mti_numreaders;
2639 for (i=0; i<nr; i++)
2640 if (ti->mti_readers[i].mr_pid == 0)
2642 if (i == env->me_maxreaders) {
2643 UNLOCK_MUTEX(rmutex);
2644 return MDB_READERS_FULL;
2646 r = &ti->mti_readers[i];
2647 /* Claim the reader slot, carefully since other code
2648 * uses the reader table un-mutexed: First reset the
2649 * slot, next publish it in mti_numreaders. After
2650 * that, it is safe for mdb_env_close() to touch it.
2651 * When it will be closed, we can finally claim it.
2654 r->mr_txnid = (txnid_t)-1;
2657 ti->mti_numreaders = ++nr;
2658 env->me_close_readers = nr;
2660 UNLOCK_MUTEX(rmutex);
2662 new_notls = (env->me_flags & MDB_NOTLS);
2663 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2668 do /* LY: Retry on a race, ITS#7970. */
2669 r->mr_txnid = ti->mti_txnid;
2670 while(r->mr_txnid != ti->mti_txnid);
2671 txn->mt_txnid = r->mr_txnid;
2672 txn->mt_u.reader = r;
2673 meta = env->me_metas[txn->mt_txnid & 1];
2677 /* Not yet touching txn == env->me_txn0, it may be active */
2679 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2681 txn->mt_txnid = ti->mti_txnid;
2682 meta = env->me_metas[txn->mt_txnid & 1];
2684 meta = mdb_env_pick_meta(env);
2685 txn->mt_txnid = meta->mm_txnid;
2689 if (txn->mt_txnid == mdb_debug_start)
2692 txn->mt_child = NULL;
2693 txn->mt_loose_pgs = NULL;
2694 txn->mt_loose_count = 0;
2695 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2696 txn->mt_u.dirty_list = env->me_dirty_list;
2697 txn->mt_u.dirty_list[0].mid = 0;
2698 txn->mt_free_pgs = env->me_free_pgs;
2699 txn->mt_free_pgs[0] = 0;
2700 txn->mt_spill_pgs = NULL;
2702 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2705 /* Copy the DB info and flags */
2706 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2708 /* Moved to here to avoid a data race in read TXNs */
2709 txn->mt_next_pgno = meta->mm_last_pg+1;
2711 txn->mt_flags = flags;
2714 txn->mt_numdbs = env->me_numdbs;
2715 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2716 x = env->me_dbflags[i];
2717 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2718 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2720 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2721 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2723 if (env->me_flags & MDB_FATAL_ERROR) {
2724 DPUTS("environment had fatal error, must shutdown!");
2726 } else if (env->me_maxpg < txn->mt_next_pgno) {
2727 rc = MDB_MAP_RESIZED;
2731 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2736 mdb_txn_renew(MDB_txn *txn)
2740 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2743 rc = mdb_txn_renew0(txn);
2744 if (rc == MDB_SUCCESS) {
2745 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2746 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2747 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2753 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2757 int rc, size, tsize;
2759 flags &= MDB_TXN_BEGIN_FLAGS;
2760 flags |= env->me_flags & MDB_WRITEMAP;
2762 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2766 /* Nested transactions: Max 1 child, write txns only, no writemap */
2767 flags |= parent->mt_flags;
2768 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2769 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2771 /* Child txns save MDB_pgstate and use own copy of cursors */
2772 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2773 size += tsize = sizeof(MDB_ntxn);
2774 } else if (flags & MDB_RDONLY) {
2775 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2776 size += tsize = sizeof(MDB_txn);
2778 /* Reuse preallocated write txn. However, do not touch it until
2779 * mdb_txn_renew0() succeeds, since it currently may be active.
2784 if ((txn = calloc(1, size)) == NULL) {
2785 DPRINTF(("calloc: %s", strerror(errno)));
2788 txn->mt_dbxs = env->me_dbxs; /* static */
2789 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2790 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2791 txn->mt_flags = flags;
2796 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2797 txn->mt_dbiseqs = parent->mt_dbiseqs;
2798 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2799 if (!txn->mt_u.dirty_list ||
2800 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2802 free(txn->mt_u.dirty_list);
2806 txn->mt_txnid = parent->mt_txnid;
2807 txn->mt_dirty_room = parent->mt_dirty_room;
2808 txn->mt_u.dirty_list[0].mid = 0;
2809 txn->mt_spill_pgs = NULL;
2810 txn->mt_next_pgno = parent->mt_next_pgno;
2811 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2812 parent->mt_child = txn;
2813 txn->mt_parent = parent;
2814 txn->mt_numdbs = parent->mt_numdbs;
2815 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2816 /* Copy parent's mt_dbflags, but clear DB_NEW */
2817 for (i=0; i<txn->mt_numdbs; i++)
2818 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2820 ntxn = (MDB_ntxn *)txn;
2821 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2822 if (env->me_pghead) {
2823 size = MDB_IDL_SIZEOF(env->me_pghead);
2824 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2826 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2831 rc = mdb_cursor_shadow(parent, txn);
2833 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2834 } else { /* MDB_RDONLY */
2835 txn->mt_dbiseqs = env->me_dbiseqs;
2837 rc = mdb_txn_renew0(txn);
2840 if (txn != env->me_txn0)
2843 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2845 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2846 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2847 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2854 mdb_txn_env(MDB_txn *txn)
2856 if(!txn) return NULL;
2861 mdb_txn_id(MDB_txn *txn)
2864 return txn->mt_txnid;
2867 /** Export or close DBI handles opened in this txn. */
2869 mdb_dbis_update(MDB_txn *txn, int keep)
2872 MDB_dbi n = txn->mt_numdbs;
2873 MDB_env *env = txn->mt_env;
2874 unsigned char *tdbflags = txn->mt_dbflags;
2876 for (i = n; --i >= CORE_DBS;) {
2877 if (tdbflags[i] & DB_NEW) {
2879 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2881 char *ptr = env->me_dbxs[i].md_name.mv_data;
2883 env->me_dbxs[i].md_name.mv_data = NULL;
2884 env->me_dbxs[i].md_name.mv_size = 0;
2885 env->me_dbflags[i] = 0;
2886 env->me_dbiseqs[i]++;
2892 if (keep && env->me_numdbs < n)
2896 /** End a transaction, except successful commit of a nested transaction.
2897 * May be called twice for readonly txns: First reset it, then abort.
2898 * @param[in] txn the transaction handle to end
2899 * @param[in] mode why and how to end the transaction
2902 mdb_txn_end(MDB_txn *txn, unsigned mode)
2904 MDB_env *env = txn->mt_env;
2906 static const char *const names[] = MDB_END_NAMES;
2909 /* Export or close DBI handles opened in this txn */
2910 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2912 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2913 names[mode & MDB_END_OPMASK],
2914 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2915 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2917 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2918 if (txn->mt_u.reader) {
2919 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2920 if (!(env->me_flags & MDB_NOTLS)) {
2921 txn->mt_u.reader = NULL; /* txn does not own reader */
2922 } else if (mode & MDB_END_SLOT) {
2923 txn->mt_u.reader->mr_pid = 0;
2924 txn->mt_u.reader = NULL;
2925 } /* else txn owns the slot until it does MDB_END_SLOT */
2927 txn->mt_numdbs = 0; /* prevent further DBI activity */
2928 txn->mt_flags |= MDB_TXN_FINISHED;
2930 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2931 pgno_t *pghead = env->me_pghead;
2933 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2934 mdb_cursors_close(txn, 0);
2935 if (!(env->me_flags & MDB_WRITEMAP)) {
2936 mdb_dlist_free(txn);
2940 txn->mt_flags = MDB_TXN_FINISHED;
2942 if (!txn->mt_parent) {
2943 mdb_midl_shrink(&txn->mt_free_pgs);
2944 env->me_free_pgs = txn->mt_free_pgs;
2946 env->me_pghead = NULL;
2950 mode = 0; /* txn == env->me_txn0, do not free() it */
2952 /* The writer mutex was locked in mdb_txn_begin. */
2954 UNLOCK_MUTEX(env->me_wmutex);
2956 txn->mt_parent->mt_child = NULL;
2957 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2958 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2959 mdb_midl_free(txn->mt_free_pgs);
2960 mdb_midl_free(txn->mt_spill_pgs);
2961 free(txn->mt_u.dirty_list);
2964 mdb_midl_free(pghead);
2967 if (mode & MDB_END_FREE)
2972 mdb_txn_reset(MDB_txn *txn)
2977 /* This call is only valid for read-only txns */
2978 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2981 mdb_txn_end(txn, MDB_END_RESET);
2985 mdb_txn_abort(MDB_txn *txn)
2991 mdb_txn_abort(txn->mt_child);
2993 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
2996 /** Save the freelist as of this transaction to the freeDB.
2997 * This changes the freelist. Keep trying until it stabilizes.
3000 mdb_freelist_save(MDB_txn *txn)
3002 /* env->me_pghead[] can grow and shrink during this call.
3003 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3004 * Page numbers cannot disappear from txn->mt_free_pgs[].
3007 MDB_env *env = txn->mt_env;
3008 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3009 txnid_t pglast = 0, head_id = 0;
3010 pgno_t freecnt = 0, *free_pgs, *mop;
3011 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3013 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3015 if (env->me_pghead) {
3016 /* Make sure first page of freeDB is touched and on freelist */
3017 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3018 if (rc && rc != MDB_NOTFOUND)
3022 if (!env->me_pghead && txn->mt_loose_pgs) {
3023 /* Put loose page numbers in mt_free_pgs, since
3024 * we may be unable to return them to me_pghead.
3026 MDB_page *mp = txn->mt_loose_pgs;
3027 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3029 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3030 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3031 txn->mt_loose_pgs = NULL;
3032 txn->mt_loose_count = 0;
3035 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3036 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3037 ? SSIZE_MAX : maxfree_1pg;
3040 /* Come back here after each Put() in case freelist changed */
3045 /* If using records from freeDB which we have not yet
3046 * deleted, delete them and any we reserved for me_pghead.
3048 while (pglast < env->me_pglast) {
3049 rc = mdb_cursor_first(&mc, &key, NULL);
3052 pglast = head_id = *(txnid_t *)key.mv_data;
3053 total_room = head_room = 0;
3054 mdb_tassert(txn, pglast <= env->me_pglast);
3055 rc = mdb_cursor_del(&mc, 0);
3060 /* Save the IDL of pages freed by this txn, to a single record */
3061 if (freecnt < txn->mt_free_pgs[0]) {
3063 /* Make sure last page of freeDB is touched and on freelist */
3064 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3065 if (rc && rc != MDB_NOTFOUND)
3068 free_pgs = txn->mt_free_pgs;
3069 /* Write to last page of freeDB */
3070 key.mv_size = sizeof(txn->mt_txnid);
3071 key.mv_data = &txn->mt_txnid;
3073 freecnt = free_pgs[0];
3074 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3075 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3078 /* Retry if mt_free_pgs[] grew during the Put() */
3079 free_pgs = txn->mt_free_pgs;
3080 } while (freecnt < free_pgs[0]);
3081 mdb_midl_sort(free_pgs);
3082 memcpy(data.mv_data, free_pgs, data.mv_size);
3085 unsigned int i = free_pgs[0];
3086 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3087 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3089 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3095 mop = env->me_pghead;
3096 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3098 /* Reserve records for me_pghead[]. Split it if multi-page,
3099 * to avoid searching freeDB for a page range. Use keys in
3100 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3102 if (total_room >= mop_len) {
3103 if (total_room == mop_len || --more < 0)
3105 } else if (head_room >= maxfree_1pg && head_id > 1) {
3106 /* Keep current record (overflow page), add a new one */
3110 /* (Re)write {key = head_id, IDL length = head_room} */
3111 total_room -= head_room;
3112 head_room = mop_len - total_room;
3113 if (head_room > maxfree_1pg && head_id > 1) {
3114 /* Overflow multi-page for part of me_pghead */
3115 head_room /= head_id; /* amortize page sizes */
3116 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3117 } else if (head_room < 0) {
3118 /* Rare case, not bothering to delete this record */
3121 key.mv_size = sizeof(head_id);
3122 key.mv_data = &head_id;
3123 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3124 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3127 /* IDL is initially empty, zero out at least the length */
3128 pgs = (pgno_t *)data.mv_data;
3129 j = head_room > clean_limit ? head_room : 0;
3133 total_room += head_room;
3136 /* Return loose page numbers to me_pghead, though usually none are
3137 * left at this point. The pages themselves remain in dirty_list.
3139 if (txn->mt_loose_pgs) {
3140 MDB_page *mp = txn->mt_loose_pgs;
3141 unsigned count = txn->mt_loose_count;
3143 /* Room for loose pages + temp IDL with same */
3144 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3146 mop = env->me_pghead;
3147 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3148 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3149 loose[ ++count ] = mp->mp_pgno;
3151 mdb_midl_sort(loose);
3152 mdb_midl_xmerge(mop, loose);
3153 txn->mt_loose_pgs = NULL;
3154 txn->mt_loose_count = 0;
3158 /* Fill in the reserved me_pghead records */
3164 rc = mdb_cursor_first(&mc, &key, &data);
3165 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3166 txnid_t id = *(txnid_t *)key.mv_data;
3167 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3170 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3172 if (len > mop_len) {
3174 data.mv_size = (len + 1) * sizeof(MDB_ID);
3176 data.mv_data = mop -= len;
3179 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3181 if (rc || !(mop_len -= len))
3188 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3189 * @param[in] txn the transaction that's being committed
3190 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3191 * @return 0 on success, non-zero on failure.
3194 mdb_page_flush(MDB_txn *txn, int keep)
3196 MDB_env *env = txn->mt_env;
3197 MDB_ID2L dl = txn->mt_u.dirty_list;
3198 unsigned psize = env->me_psize, j;
3199 int i, pagecount = dl[0].mid, rc;
3200 size_t size = 0, pos = 0;
3202 MDB_page *dp = NULL;
3206 struct iovec iov[MDB_COMMIT_PAGES];
3207 ssize_t wpos = 0, wsize = 0, wres;
3208 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3214 if (env->me_flags & MDB_WRITEMAP) {
3215 /* Clear dirty flags */
3216 while (++i <= pagecount) {
3218 /* Don't flush this page yet */
3219 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3220 dp->mp_flags &= ~P_KEEP;
3224 dp->mp_flags &= ~P_DIRTY;
3229 /* Write the pages */
3231 if (++i <= pagecount) {
3233 /* Don't flush this page yet */
3234 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3235 dp->mp_flags &= ~P_KEEP;
3240 /* clear dirty flag */
3241 dp->mp_flags &= ~P_DIRTY;
3244 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3249 /* Windows actually supports scatter/gather I/O, but only on
3250 * unbuffered file handles. Since we're relying on the OS page
3251 * cache for all our data, that's self-defeating. So we just
3252 * write pages one at a time. We use the ov structure to set
3253 * the write offset, to at least save the overhead of a Seek
3256 DPRINTF(("committing page %"Z"u", pgno));
3257 memset(&ov, 0, sizeof(ov));
3258 ov.Offset = pos & 0xffffffff;
3259 ov.OffsetHigh = pos >> 16 >> 16;
3260 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3262 DPRINTF(("WriteFile: %d", rc));
3266 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3267 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3270 /* Write previous page(s) */
3271 #ifdef MDB_USE_PWRITEV
3272 wres = pwritev(env->me_fd, iov, n, wpos);
3275 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3278 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3282 DPRINTF(("lseek: %s", strerror(rc)));
3285 wres = writev(env->me_fd, iov, n);
3288 if (wres != wsize) {
3293 DPRINTF(("Write error: %s", strerror(rc)));
3295 rc = EIO; /* TODO: Use which error code? */
3296 DPUTS("short write, filesystem full?");
3307 DPRINTF(("committing page %"Z"u", pgno));
3308 next_pos = pos + size;
3309 iov[n].iov_len = size;
3310 iov[n].iov_base = (char *)dp;
3316 /* MIPS has cache coherency issues, this is a no-op everywhere else
3317 * Note: for any size >= on-chip cache size, entire on-chip cache is
3320 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3322 for (i = keep; ++i <= pagecount; ) {
3324 /* This is a page we skipped above */
3327 dl[j].mid = dp->mp_pgno;
3330 mdb_dpage_free(env, dp);
3335 txn->mt_dirty_room += i - j;
3341 mdb_txn_commit(MDB_txn *txn)
3344 unsigned int i, end_mode;
3350 /* mdb_txn_end() mode for a commit which writes nothing */
3351 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3353 if (txn->mt_child) {
3354 rc = mdb_txn_commit(txn->mt_child);
3361 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3365 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3366 DPUTS("txn has failed/finished, can't commit");
3368 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3373 if (txn->mt_parent) {
3374 MDB_txn *parent = txn->mt_parent;
3378 unsigned x, y, len, ps_len;
3380 /* Append our free list to parent's */
3381 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3384 mdb_midl_free(txn->mt_free_pgs);
3385 /* Failures after this must either undo the changes
3386 * to the parent or set MDB_TXN_ERROR in the parent.
3389 parent->mt_next_pgno = txn->mt_next_pgno;
3390 parent->mt_flags = txn->mt_flags;
3392 /* Merge our cursors into parent's and close them */
3393 mdb_cursors_close(txn, 1);
3395 /* Update parent's DB table. */
3396 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3397 parent->mt_numdbs = txn->mt_numdbs;
3398 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3399 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3400 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3401 /* preserve parent's DB_NEW status */
3402 x = parent->mt_dbflags[i] & DB_NEW;
3403 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3406 dst = parent->mt_u.dirty_list;
3407 src = txn->mt_u.dirty_list;
3408 /* Remove anything in our dirty list from parent's spill list */
3409 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3411 pspill[0] = (pgno_t)-1;
3412 /* Mark our dirty pages as deleted in parent spill list */
3413 for (i=0, len=src[0].mid; ++i <= len; ) {
3414 MDB_ID pn = src[i].mid << 1;
3415 while (pn > pspill[x])
3417 if (pn == pspill[x]) {
3422 /* Squash deleted pagenums if we deleted any */
3423 for (x=y; ++x <= ps_len; )
3424 if (!(pspill[x] & 1))
3425 pspill[++y] = pspill[x];
3429 /* Find len = length of merging our dirty list with parent's */
3431 dst[0].mid = 0; /* simplify loops */
3432 if (parent->mt_parent) {
3433 len = x + src[0].mid;
3434 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3435 for (i = x; y && i; y--) {
3436 pgno_t yp = src[y].mid;
3437 while (yp < dst[i].mid)
3439 if (yp == dst[i].mid) {
3444 } else { /* Simplify the above for single-ancestor case */
3445 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3447 /* Merge our dirty list with parent's */
3449 for (i = len; y; dst[i--] = src[y--]) {
3450 pgno_t yp = src[y].mid;
3451 while (yp < dst[x].mid)
3452 dst[i--] = dst[x--];
3453 if (yp == dst[x].mid)
3454 free(dst[x--].mptr);
3456 mdb_tassert(txn, i == x);
3458 free(txn->mt_u.dirty_list);
3459 parent->mt_dirty_room = txn->mt_dirty_room;
3460 if (txn->mt_spill_pgs) {
3461 if (parent->mt_spill_pgs) {
3462 /* TODO: Prevent failure here, so parent does not fail */
3463 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3465 parent->mt_flags |= MDB_TXN_ERROR;
3466 mdb_midl_free(txn->mt_spill_pgs);
3467 mdb_midl_sort(parent->mt_spill_pgs);
3469 parent->mt_spill_pgs = txn->mt_spill_pgs;
3473 /* Append our loose page list to parent's */
3474 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3476 *lp = txn->mt_loose_pgs;
3477 parent->mt_loose_count += txn->mt_loose_count;
3479 parent->mt_child = NULL;
3480 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3485 if (txn != env->me_txn) {
3486 DPUTS("attempt to commit unknown transaction");
3491 mdb_cursors_close(txn, 0);
3493 if (!txn->mt_u.dirty_list[0].mid &&
3494 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3497 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3498 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3500 /* Update DB root pointers */
3501 if (txn->mt_numdbs > CORE_DBS) {
3505 data.mv_size = sizeof(MDB_db);
3507 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3508 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3509 if (txn->mt_dbflags[i] & DB_DIRTY) {
3510 if (TXN_DBI_CHANGED(txn, i)) {
3514 data.mv_data = &txn->mt_dbs[i];
3515 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3523 rc = mdb_freelist_save(txn);
3527 mdb_midl_free(env->me_pghead);
3528 env->me_pghead = NULL;
3529 mdb_midl_shrink(&txn->mt_free_pgs);
3535 if ((rc = mdb_page_flush(txn, 0)) ||
3536 (rc = mdb_env_sync(env, 0)) ||
3537 (rc = mdb_env_write_meta(txn)))
3539 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3542 mdb_txn_end(txn, end_mode);
3550 /** Read the environment parameters of a DB environment before
3551 * mapping it into memory.
3552 * @param[in] env the environment handle
3553 * @param[out] meta address of where to store the meta information
3554 * @return 0 on success, non-zero on failure.
3557 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3563 enum { Size = sizeof(pbuf) };
3565 /* We don't know the page size yet, so use a minimum value.
3566 * Read both meta pages so we can use the latest one.
3569 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3573 memset(&ov, 0, sizeof(ov));
3575 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3576 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3579 rc = pread(env->me_fd, &pbuf, Size, off);
3582 if (rc == 0 && off == 0)
3584 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3585 DPRINTF(("read: %s", mdb_strerror(rc)));
3589 p = (MDB_page *)&pbuf;
3591 if (!F_ISSET(p->mp_flags, P_META)) {
3592 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3597 if (m->mm_magic != MDB_MAGIC) {
3598 DPUTS("meta has invalid magic");
3602 if (m->mm_version != MDB_DATA_VERSION) {
3603 DPRINTF(("database is version %u, expected version %u",
3604 m->mm_version, MDB_DATA_VERSION));
3605 return MDB_VERSION_MISMATCH;
3608 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3614 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3616 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3618 meta->mm_magic = MDB_MAGIC;
3619 meta->mm_version = MDB_DATA_VERSION;
3620 meta->mm_mapsize = env->me_mapsize;
3621 meta->mm_psize = env->me_psize;
3622 meta->mm_last_pg = NUM_METAS-1;
3623 meta->mm_flags = env->me_flags & 0xffff;
3624 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3625 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3626 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3629 /** Write the environment parameters of a freshly created DB environment.
3630 * @param[in] env the environment handle
3631 * @param[in] meta the #MDB_meta to write
3632 * @return 0 on success, non-zero on failure.
3635 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3643 memset(&ov, 0, sizeof(ov));
3644 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3646 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3649 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3650 len = pwrite(fd, ptr, size, pos); \
3651 if (len == -1 && ErrCode() == EINTR) continue; \
3652 rc = (len >= 0); break; } while(1)
3655 DPUTS("writing new meta page");
3657 psize = env->me_psize;
3659 p = calloc(NUM_METAS, psize);
3664 p->mp_flags = P_META;
3665 *(MDB_meta *)METADATA(p) = *meta;
3667 q = (MDB_page *)((char *)p + psize);
3669 q->mp_flags = P_META;
3670 *(MDB_meta *)METADATA(q) = *meta;
3672 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3675 else if ((unsigned) len == psize * NUM_METAS)
3683 /** Update the environment info to commit a transaction.
3684 * @param[in] txn the transaction that's being committed
3685 * @return 0 on success, non-zero on failure.
3688 mdb_env_write_meta(MDB_txn *txn)
3691 MDB_meta meta, metab, *mp;
3695 int rc, len, toggle;
3704 toggle = txn->mt_txnid & 1;
3705 DPRINTF(("writing meta page %d for root page %"Z"u",
3706 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3709 flags = env->me_flags;
3710 mp = env->me_metas[toggle];
3711 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3712 /* Persist any increases of mapsize config */
3713 if (mapsize < env->me_mapsize)
3714 mapsize = env->me_mapsize;
3716 if (flags & MDB_WRITEMAP) {
3717 mp->mm_mapsize = mapsize;
3718 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3719 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3720 mp->mm_last_pg = txn->mt_next_pgno - 1;
3721 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3722 !(defined(__i386__) || defined(__x86_64__))
3723 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3724 __sync_synchronize();
3726 mp->mm_txnid = txn->mt_txnid;
3727 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3728 unsigned meta_size = env->me_psize;
3729 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3730 ptr = (char *)mp - PAGEHDRSZ;
3731 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3732 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3736 if (MDB_MSYNC(ptr, meta_size, rc)) {
3743 metab.mm_txnid = mp->mm_txnid;
3744 metab.mm_last_pg = mp->mm_last_pg;
3746 meta.mm_mapsize = mapsize;
3747 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3748 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3749 meta.mm_last_pg = txn->mt_next_pgno - 1;
3750 meta.mm_txnid = txn->mt_txnid;
3752 off = offsetof(MDB_meta, mm_mapsize);
3753 ptr = (char *)&meta + off;
3754 len = sizeof(MDB_meta) - off;
3755 off += (char *)mp - env->me_map;
3757 /* Write to the SYNC fd */
3758 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3761 memset(&ov, 0, sizeof(ov));
3763 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3768 rc = pwrite(mfd, ptr, len, off);
3771 rc = rc < 0 ? ErrCode() : EIO;
3776 DPUTS("write failed, disk error?");
3777 /* On a failure, the pagecache still contains the new data.
3778 * Write some old data back, to prevent it from being used.
3779 * Use the non-SYNC fd; we know it will fail anyway.
3781 meta.mm_last_pg = metab.mm_last_pg;
3782 meta.mm_txnid = metab.mm_txnid;
3784 memset(&ov, 0, sizeof(ov));
3786 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3788 r2 = pwrite(env->me_fd, ptr, len, off);
3789 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3792 env->me_flags |= MDB_FATAL_ERROR;
3795 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3796 CACHEFLUSH(env->me_map + off, len, DCACHE);
3798 /* Memory ordering issues are irrelevant; since the entire writer
3799 * is wrapped by wmutex, all of these changes will become visible
3800 * after the wmutex is unlocked. Since the DB is multi-version,
3801 * readers will get consistent data regardless of how fresh or
3802 * how stale their view of these values is.
3805 env->me_txns->mti_txnid = txn->mt_txnid;
3810 /** Check both meta pages to see which one is newer.
3811 * @param[in] env the environment handle
3812 * @return newest #MDB_meta.
3815 mdb_env_pick_meta(const MDB_env *env)
3817 MDB_meta *const *metas = env->me_metas;
3818 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3822 mdb_env_create(MDB_env **env)
3826 e = calloc(1, sizeof(MDB_env));
3830 e->me_maxreaders = DEFAULT_READERS;
3831 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3832 e->me_fd = INVALID_HANDLE_VALUE;
3833 e->me_lfd = INVALID_HANDLE_VALUE;
3834 e->me_mfd = INVALID_HANDLE_VALUE;
3835 #ifdef MDB_USE_POSIX_SEM
3836 e->me_rmutex = SEM_FAILED;
3837 e->me_wmutex = SEM_FAILED;
3839 e->me_pid = getpid();
3840 GET_PAGESIZE(e->me_os_psize);
3841 VGMEMP_CREATE(e,0,0);
3847 mdb_env_map(MDB_env *env, void *addr)
3850 unsigned int flags = env->me_flags;
3854 LONG sizelo, sizehi;
3857 if (flags & MDB_RDONLY) {
3858 /* Don't set explicit map size, use whatever exists */
3863 msize = env->me_mapsize;
3864 sizelo = msize & 0xffffffff;
3865 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3867 /* Windows won't create mappings for zero length files.
3868 * and won't map more than the file size.
3869 * Just set the maxsize right now.
3871 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3872 || !SetEndOfFile(env->me_fd)
3873 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3877 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3878 PAGE_READWRITE : PAGE_READONLY,
3879 sizehi, sizelo, NULL);
3882 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3883 FILE_MAP_WRITE : FILE_MAP_READ,
3885 rc = env->me_map ? 0 : ErrCode();
3890 int prot = PROT_READ;
3891 if (flags & MDB_WRITEMAP) {
3893 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3896 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3898 if (env->me_map == MAP_FAILED) {
3903 if (flags & MDB_NORDAHEAD) {
3904 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3906 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3908 #ifdef POSIX_MADV_RANDOM
3909 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3910 #endif /* POSIX_MADV_RANDOM */
3911 #endif /* MADV_RANDOM */
3915 /* Can happen because the address argument to mmap() is just a
3916 * hint. mmap() can pick another, e.g. if the range is in use.
3917 * The MAP_FIXED flag would prevent that, but then mmap could
3918 * instead unmap existing pages to make room for the new map.
3920 if (addr && env->me_map != addr)
3921 return EBUSY; /* TODO: Make a new MDB_* error code? */
3923 p = (MDB_page *)env->me_map;
3924 env->me_metas[0] = METADATA(p);
3925 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3931 mdb_env_set_mapsize(MDB_env *env, size_t size)
3933 /* If env is already open, caller is responsible for making
3934 * sure there are no active txns.
3942 meta = mdb_env_pick_meta(env);
3944 size = meta->mm_mapsize;
3946 /* Silently round up to minimum if the size is too small */
3947 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3951 munmap(env->me_map, env->me_mapsize);
3952 env->me_mapsize = size;
3953 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3954 rc = mdb_env_map(env, old);
3958 env->me_mapsize = size;
3960 env->me_maxpg = env->me_mapsize / env->me_psize;
3965 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3969 env->me_maxdbs = dbs + CORE_DBS;
3974 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3976 if (env->me_map || readers < 1)
3978 env->me_maxreaders = readers;
3983 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3985 if (!env || !readers)
3987 *readers = env->me_maxreaders;
3992 mdb_fsize(HANDLE fd, size_t *size)
3995 LARGE_INTEGER fsize;
3997 if (!GetFileSizeEx(fd, &fsize))
4000 *size = fsize.QuadPart;
4012 #ifdef BROKEN_FDATASYNC
4013 #include <sys/utsname.h>
4014 #include <sys/vfs.h>
4017 /** Further setup required for opening an LMDB environment
4020 mdb_env_open2(MDB_env *env)
4022 unsigned int flags = env->me_flags;
4023 int i, newenv = 0, rc;
4027 /* See if we should use QueryLimited */
4029 if ((rc & 0xff) > 5)
4030 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4032 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4035 #ifdef BROKEN_FDATASYNC
4036 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4037 * https://lkml.org/lkml/2012/9/3/83
4038 * Kernels after 3.6-rc6 are known good.
4039 * https://lkml.org/lkml/2012/9/10/556
4040 * See if the DB is on ext3/ext4, then check for new enough kernel
4041 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4046 fstatfs(env->me_fd, &st);
4047 while (st.f_type == 0xEF53) {
4051 if (uts.release[0] < '3') {
4052 if (!strncmp(uts.release, "2.6.32.", 7)) {
4053 i = atoi(uts.release+7);
4055 break; /* 2.6.32.60 and newer is OK */
4056 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4057 i = atoi(uts.release+7);
4059 break; /* 2.6.34.15 and newer is OK */
4061 } else if (uts.release[0] == '3') {
4062 i = atoi(uts.release+2);
4064 break; /* 3.6 and newer is OK */
4066 i = atoi(uts.release+4);
4068 break; /* 3.5.4 and newer is OK */
4069 } else if (i == 2) {
4070 i = atoi(uts.release+4);
4072 break; /* 3.2.30 and newer is OK */
4074 } else { /* 4.x and newer is OK */
4077 env->me_flags |= MDB_FSYNCONLY;
4083 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4086 DPUTS("new mdbenv");
4088 env->me_psize = env->me_os_psize;
4089 if (env->me_psize > MAX_PAGESIZE)
4090 env->me_psize = MAX_PAGESIZE;
4091 memset(&meta, 0, sizeof(meta));
4092 mdb_env_init_meta0(env, &meta);
4093 meta.mm_mapsize = DEFAULT_MAPSIZE;
4095 env->me_psize = meta.mm_psize;
4098 /* Was a mapsize configured? */
4099 if (!env->me_mapsize) {
4100 env->me_mapsize = meta.mm_mapsize;
4103 /* Make sure mapsize >= committed data size. Even when using
4104 * mm_mapsize, which could be broken in old files (ITS#7789).
4106 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4107 if (env->me_mapsize < minsize)
4108 env->me_mapsize = minsize;
4110 meta.mm_mapsize = env->me_mapsize;
4112 if (newenv && !(flags & MDB_FIXEDMAP)) {
4113 /* mdb_env_map() may grow the datafile. Write the metapages
4114 * first, so the file will be valid if initialization fails.
4115 * Except with FIXEDMAP, since we do not yet know mm_address.
4116 * We could fill in mm_address later, but then a different
4117 * program might end up doing that - one with a memory layout
4118 * and map address which does not suit the main program.
4120 rc = mdb_env_init_meta(env, &meta);
4126 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4131 if (flags & MDB_FIXEDMAP)
4132 meta.mm_address = env->me_map;
4133 i = mdb_env_init_meta(env, &meta);
4134 if (i != MDB_SUCCESS) {
4139 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4140 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4142 #if !(MDB_MAXKEYSIZE)
4143 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4145 env->me_maxpg = env->me_mapsize / env->me_psize;
4149 MDB_meta *meta = mdb_env_pick_meta(env);
4150 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4152 DPRINTF(("opened database version %u, pagesize %u",
4153 meta->mm_version, env->me_psize));
4154 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4155 DPRINTF(("depth: %u", db->md_depth));
4156 DPRINTF(("entries: %"Z"u", db->md_entries));
4157 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4158 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4159 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4160 DPRINTF(("root: %"Z"u", db->md_root));
4168 /** Release a reader thread's slot in the reader lock table.
4169 * This function is called automatically when a thread exits.
4170 * @param[in] ptr This points to the slot in the reader lock table.
4173 mdb_env_reader_dest(void *ptr)
4175 MDB_reader *reader = ptr;
4181 /** Junk for arranging thread-specific callbacks on Windows. This is
4182 * necessarily platform and compiler-specific. Windows supports up
4183 * to 1088 keys. Let's assume nobody opens more than 64 environments
4184 * in a single process, for now. They can override this if needed.
4186 #ifndef MAX_TLS_KEYS
4187 #define MAX_TLS_KEYS 64
4189 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4190 static int mdb_tls_nkeys;
4192 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4196 case DLL_PROCESS_ATTACH: break;
4197 case DLL_THREAD_ATTACH: break;
4198 case DLL_THREAD_DETACH:
4199 for (i=0; i<mdb_tls_nkeys; i++) {
4200 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4202 mdb_env_reader_dest(r);
4206 case DLL_PROCESS_DETACH: break;
4211 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4213 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4217 /* Force some symbol references.
4218 * _tls_used forces the linker to create the TLS directory if not already done
4219 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4221 #pragma comment(linker, "/INCLUDE:_tls_used")
4222 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4223 #pragma const_seg(".CRT$XLB")
4224 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4225 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4228 #pragma comment(linker, "/INCLUDE:__tls_used")
4229 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4230 #pragma data_seg(".CRT$XLB")
4231 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4233 #endif /* WIN 32/64 */
4234 #endif /* !__GNUC__ */
4237 /** Downgrade the exclusive lock on the region back to shared */
4239 mdb_env_share_locks(MDB_env *env, int *excl)
4242 MDB_meta *meta = mdb_env_pick_meta(env);
4244 env->me_txns->mti_txnid = meta->mm_txnid;
4249 /* First acquire a shared lock. The Unlock will
4250 * then release the existing exclusive lock.
4252 memset(&ov, 0, sizeof(ov));
4253 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4256 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4262 struct flock lock_info;
4263 /* The shared lock replaces the existing lock */
4264 memset((void *)&lock_info, 0, sizeof(lock_info));
4265 lock_info.l_type = F_RDLCK;
4266 lock_info.l_whence = SEEK_SET;
4267 lock_info.l_start = 0;
4268 lock_info.l_len = 1;
4269 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4270 (rc = ErrCode()) == EINTR) ;
4271 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4278 /** Try to get exclusive lock, otherwise shared.
4279 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4282 mdb_env_excl_lock(MDB_env *env, int *excl)
4286 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4290 memset(&ov, 0, sizeof(ov));
4291 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4298 struct flock lock_info;
4299 memset((void *)&lock_info, 0, sizeof(lock_info));
4300 lock_info.l_type = F_WRLCK;
4301 lock_info.l_whence = SEEK_SET;
4302 lock_info.l_start = 0;
4303 lock_info.l_len = 1;
4304 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4305 (rc = ErrCode()) == EINTR) ;
4309 # ifndef MDB_USE_POSIX_MUTEX
4310 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4313 lock_info.l_type = F_RDLCK;
4314 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4315 (rc = ErrCode()) == EINTR) ;
4325 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4327 * @(#) $Revision: 5.1 $
4328 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4329 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4331 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4335 * Please do not copyright this code. This code is in the public domain.
4337 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4338 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4339 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4340 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4341 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4342 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4343 * PERFORMANCE OF THIS SOFTWARE.
4346 * chongo <Landon Curt Noll> /\oo/\
4347 * http://www.isthe.com/chongo/
4349 * Share and Enjoy! :-)
4352 typedef unsigned long long mdb_hash_t;
4353 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4355 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4356 * @param[in] val value to hash
4357 * @param[in] hval initial value for hash
4358 * @return 64 bit hash
4360 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4361 * hval arg on the first call.
4364 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4366 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4367 unsigned char *end = s + val->mv_size;
4369 * FNV-1a hash each octet of the string
4372 /* xor the bottom with the current octet */
4373 hval ^= (mdb_hash_t)*s++;
4375 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4376 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4377 (hval << 7) + (hval << 8) + (hval << 40);
4379 /* return our new hash value */
4383 /** Hash the string and output the encoded hash.
4384 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4385 * very short name limits. We don't care about the encoding being reversible,
4386 * we just want to preserve as many bits of the input as possible in a
4387 * small printable string.
4388 * @param[in] str string to hash
4389 * @param[out] encbuf an array of 11 chars to hold the hash
4391 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4394 mdb_pack85(unsigned long l, char *out)
4398 for (i=0; i<5; i++) {
4399 *out++ = mdb_a85[l % 85];
4405 mdb_hash_enc(MDB_val *val, char *encbuf)
4407 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4409 mdb_pack85(h, encbuf);
4410 mdb_pack85(h>>32, encbuf+5);
4415 /** Open and/or initialize the lock region for the environment.
4416 * @param[in] env The LMDB environment.
4417 * @param[in] lpath The pathname of the file used for the lock region.
4418 * @param[in] mode The Unix permissions for the file, if we create it.
4419 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4420 * @return 0 on success, non-zero on failure.
4423 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4426 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4428 # define MDB_ERRCODE_ROFS EROFS
4429 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4430 # define MDB_CLOEXEC O_CLOEXEC
4433 # define MDB_CLOEXEC 0
4440 env->me_lfd = CreateFileA(lpath, GENERIC_READ|GENERIC_WRITE,
4441 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4442 FILE_ATTRIBUTE_NORMAL, NULL);
4444 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4446 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4448 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4453 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4454 /* Lose record locks when exec*() */
4455 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4456 fcntl(env->me_lfd, F_SETFD, fdflags);
4459 if (!(env->me_flags & MDB_NOTLS)) {
4460 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4463 env->me_flags |= MDB_ENV_TXKEY;
4465 /* Windows TLS callbacks need help finding their TLS info. */
4466 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4470 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4474 /* Try to get exclusive lock. If we succeed, then
4475 * nobody is using the lock region and we should initialize it.
4477 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4480 size = GetFileSize(env->me_lfd, NULL);
4482 size = lseek(env->me_lfd, 0, SEEK_END);
4483 if (size == -1) goto fail_errno;
4485 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4486 if (size < rsize && *excl > 0) {
4488 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4489 || !SetEndOfFile(env->me_lfd))
4492 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4496 size = rsize - sizeof(MDB_txninfo);
4497 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4502 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4504 if (!mh) goto fail_errno;
4505 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4507 if (!env->me_txns) goto fail_errno;
4509 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4511 if (m == MAP_FAILED) goto fail_errno;
4517 BY_HANDLE_FILE_INFORMATION stbuf;
4526 if (!mdb_sec_inited) {
4527 InitializeSecurityDescriptor(&mdb_null_sd,
4528 SECURITY_DESCRIPTOR_REVISION);
4529 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4530 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4531 mdb_all_sa.bInheritHandle = FALSE;
4532 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4535 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4536 idbuf.volume = stbuf.dwVolumeSerialNumber;
4537 idbuf.nhigh = stbuf.nFileIndexHigh;
4538 idbuf.nlow = stbuf.nFileIndexLow;
4539 val.mv_data = &idbuf;
4540 val.mv_size = sizeof(idbuf);
4541 mdb_hash_enc(&val, encbuf);
4542 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4543 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4544 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4545 if (!env->me_rmutex) goto fail_errno;
4546 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4547 if (!env->me_wmutex) goto fail_errno;
4548 #elif defined(MDB_USE_POSIX_SEM)
4557 #if defined(__NetBSD__)
4558 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4560 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4561 idbuf.dev = stbuf.st_dev;
4562 idbuf.ino = stbuf.st_ino;
4563 val.mv_data = &idbuf;
4564 val.mv_size = sizeof(idbuf);
4565 mdb_hash_enc(&val, encbuf);
4566 #ifdef MDB_SHORT_SEMNAMES
4567 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4569 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4570 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4571 /* Clean up after a previous run, if needed: Try to
4572 * remove both semaphores before doing anything else.
4574 sem_unlink(env->me_txns->mti_rmname);
4575 sem_unlink(env->me_txns->mti_wmname);
4576 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4577 O_CREAT|O_EXCL, mode, 1);
4578 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4579 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4580 O_CREAT|O_EXCL, mode, 1);
4581 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4582 #else /* MDB_USE_POSIX_MUTEX: */
4583 pthread_mutexattr_t mattr;
4585 if ((rc = pthread_mutexattr_init(&mattr))
4586 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4587 #ifdef MDB_ROBUST_SUPPORTED
4588 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4590 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4591 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4593 pthread_mutexattr_destroy(&mattr);
4594 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4596 env->me_txns->mti_magic = MDB_MAGIC;
4597 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4598 env->me_txns->mti_txnid = 0;
4599 env->me_txns->mti_numreaders = 0;
4602 if (env->me_txns->mti_magic != MDB_MAGIC) {
4603 DPUTS("lock region has invalid magic");
4607 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4608 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4609 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4610 rc = MDB_VERSION_MISMATCH;
4614 if (rc && rc != EACCES && rc != EAGAIN) {
4618 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4619 if (!env->me_rmutex) goto fail_errno;
4620 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4621 if (!env->me_wmutex) goto fail_errno;
4622 #elif defined(MDB_USE_POSIX_SEM)
4623 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4624 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4625 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4626 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4637 /** The name of the lock file in the DB environment */
4638 #define LOCKNAME "/lock.mdb"
4639 /** The name of the data file in the DB environment */
4640 #define DATANAME "/data.mdb"
4641 /** The suffix of the lock file when no subdir is used */
4642 #define LOCKSUFF "-lock"
4643 /** Only a subset of the @ref mdb_env flags can be changed
4644 * at runtime. Changing other flags requires closing the
4645 * environment and re-opening it with the new flags.
4647 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4648 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4649 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4651 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4652 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4656 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4658 int oflags, rc, len, excl = -1;
4659 char *lpath, *dpath;
4661 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4665 if (flags & MDB_NOSUBDIR) {
4666 rc = len + sizeof(LOCKSUFF) + len + 1;
4668 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4673 if (flags & MDB_NOSUBDIR) {
4674 dpath = lpath + len + sizeof(LOCKSUFF);
4675 sprintf(lpath, "%s" LOCKSUFF, path);
4676 strcpy(dpath, path);
4678 dpath = lpath + len + sizeof(LOCKNAME);
4679 sprintf(lpath, "%s" LOCKNAME, path);
4680 sprintf(dpath, "%s" DATANAME, path);
4684 flags |= env->me_flags;
4685 if (flags & MDB_RDONLY) {
4686 /* silently ignore WRITEMAP when we're only getting read access */
4687 flags &= ~MDB_WRITEMAP;
4689 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4690 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4693 env->me_flags = flags |= MDB_ENV_ACTIVE;
4697 env->me_path = strdup(path);
4698 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4699 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4700 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4701 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4705 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4707 /* For RDONLY, get lockfile after we know datafile exists */
4708 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4709 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4715 if (F_ISSET(flags, MDB_RDONLY)) {
4716 oflags = GENERIC_READ;
4717 len = OPEN_EXISTING;
4719 oflags = GENERIC_READ|GENERIC_WRITE;
4722 mode = FILE_ATTRIBUTE_NORMAL;
4723 env->me_fd = CreateFileA(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4724 NULL, len, mode, NULL);
4726 if (F_ISSET(flags, MDB_RDONLY))
4729 oflags = O_RDWR | O_CREAT;
4731 env->me_fd = open(dpath, oflags, mode);
4733 if (env->me_fd == INVALID_HANDLE_VALUE) {
4738 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4739 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4744 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4745 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4746 env->me_mfd = env->me_fd;
4748 /* Synchronous fd for meta writes. Needed even with
4749 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4752 len = OPEN_EXISTING;
4753 env->me_mfd = CreateFileA(dpath, oflags,
4754 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4755 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4758 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4760 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4765 DPRINTF(("opened dbenv %p", (void *) env));
4767 rc = mdb_env_share_locks(env, &excl);
4771 if (!(flags & MDB_RDONLY)) {
4773 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4774 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4775 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4776 (txn = calloc(1, size)))
4778 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4779 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4780 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4781 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4783 txn->mt_dbxs = env->me_dbxs;
4784 txn->mt_flags = MDB_TXN_FINISHED;
4794 mdb_env_close0(env, excl);
4800 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4802 mdb_env_close0(MDB_env *env, int excl)
4806 if (!(env->me_flags & MDB_ENV_ACTIVE))
4809 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4811 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4812 free(env->me_dbxs[i].md_name.mv_data);
4817 free(env->me_dbiseqs);
4818 free(env->me_dbflags);
4820 free(env->me_dirty_list);
4822 mdb_midl_free(env->me_free_pgs);
4824 if (env->me_flags & MDB_ENV_TXKEY) {
4825 pthread_key_delete(env->me_txkey);
4827 /* Delete our key from the global list */
4828 for (i=0; i<mdb_tls_nkeys; i++)
4829 if (mdb_tls_keys[i] == env->me_txkey) {
4830 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4838 munmap(env->me_map, env->me_mapsize);
4840 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4841 (void) close(env->me_mfd);
4842 if (env->me_fd != INVALID_HANDLE_VALUE)
4843 (void) close(env->me_fd);
4845 MDB_PID_T pid = env->me_pid;
4846 /* Clearing readers is done in this function because
4847 * me_txkey with its destructor must be disabled first.
4849 * We skip the the reader mutex, so we touch only
4850 * data owned by this process (me_close_readers and
4851 * our readers), and clear each reader atomically.
4853 for (i = env->me_close_readers; --i >= 0; )
4854 if (env->me_txns->mti_readers[i].mr_pid == pid)
4855 env->me_txns->mti_readers[i].mr_pid = 0;
4857 if (env->me_rmutex) {
4858 CloseHandle(env->me_rmutex);
4859 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4861 /* Windows automatically destroys the mutexes when
4862 * the last handle closes.
4864 #elif defined(MDB_USE_POSIX_SEM)
4865 if (env->me_rmutex != SEM_FAILED) {
4866 sem_close(env->me_rmutex);
4867 if (env->me_wmutex != SEM_FAILED)
4868 sem_close(env->me_wmutex);
4869 /* If we have the filelock: If we are the
4870 * only remaining user, clean up semaphores.
4873 mdb_env_excl_lock(env, &excl);
4875 sem_unlink(env->me_txns->mti_rmname);
4876 sem_unlink(env->me_txns->mti_wmname);
4880 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4882 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4885 /* Unlock the lockfile. Windows would have unlocked it
4886 * after closing anyway, but not necessarily at once.
4888 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4891 (void) close(env->me_lfd);
4894 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4898 mdb_env_close(MDB_env *env)
4905 VGMEMP_DESTROY(env);
4906 while ((dp = env->me_dpages) != NULL) {
4907 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4908 env->me_dpages = dp->mp_next;
4912 mdb_env_close0(env, 0);
4916 /** Compare two items pointing at aligned size_t's */
4918 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4920 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4921 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4924 /** Compare two items pointing at aligned unsigned int's.
4926 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4927 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4930 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4932 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4933 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4936 /** Compare two items pointing at unsigned ints of unknown alignment.
4937 * Nodes and keys are guaranteed to be 2-byte aligned.
4940 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4942 #if BYTE_ORDER == LITTLE_ENDIAN
4943 unsigned short *u, *c;
4946 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4947 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4950 } while(!x && u > (unsigned short *)a->mv_data);
4953 unsigned short *u, *c, *end;
4956 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4957 u = (unsigned short *)a->mv_data;
4958 c = (unsigned short *)b->mv_data;
4961 } while(!x && u < end);
4966 /** Compare two items lexically */
4968 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4975 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4981 diff = memcmp(a->mv_data, b->mv_data, len);
4982 return diff ? diff : len_diff<0 ? -1 : len_diff;
4985 /** Compare two items in reverse byte order */
4987 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4989 const unsigned char *p1, *p2, *p1_lim;
4993 p1_lim = (const unsigned char *)a->mv_data;
4994 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4995 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4997 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5003 while (p1 > p1_lim) {
5004 diff = *--p1 - *--p2;
5008 return len_diff<0 ? -1 : len_diff;
5011 /** Search for key within a page, using binary search.
5012 * Returns the smallest entry larger or equal to the key.
5013 * If exactp is non-null, stores whether the found entry was an exact match
5014 * in *exactp (1 or 0).
5015 * Updates the cursor index with the index of the found entry.
5016 * If no entry larger or equal to the key is found, returns NULL.
5019 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5021 unsigned int i = 0, nkeys;
5024 MDB_page *mp = mc->mc_pg[mc->mc_top];
5025 MDB_node *node = NULL;
5030 nkeys = NUMKEYS(mp);
5032 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5033 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5036 low = IS_LEAF(mp) ? 0 : 1;
5038 cmp = mc->mc_dbx->md_cmp;
5040 /* Branch pages have no data, so if using integer keys,
5041 * alignment is guaranteed. Use faster mdb_cmp_int.
5043 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5044 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5051 nodekey.mv_size = mc->mc_db->md_pad;
5052 node = NODEPTR(mp, 0); /* fake */
5053 while (low <= high) {
5054 i = (low + high) >> 1;
5055 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5056 rc = cmp(key, &nodekey);
5057 DPRINTF(("found leaf index %u [%s], rc = %i",
5058 i, DKEY(&nodekey), rc));
5067 while (low <= high) {
5068 i = (low + high) >> 1;
5070 node = NODEPTR(mp, i);
5071 nodekey.mv_size = NODEKSZ(node);
5072 nodekey.mv_data = NODEKEY(node);
5074 rc = cmp(key, &nodekey);
5077 DPRINTF(("found leaf index %u [%s], rc = %i",
5078 i, DKEY(&nodekey), rc));
5080 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5081 i, DKEY(&nodekey), NODEPGNO(node), rc));
5092 if (rc > 0) { /* Found entry is less than the key. */
5093 i++; /* Skip to get the smallest entry larger than key. */
5095 node = NODEPTR(mp, i);
5098 *exactp = (rc == 0 && nkeys > 0);
5099 /* store the key index */
5100 mc->mc_ki[mc->mc_top] = i;
5102 /* There is no entry larger or equal to the key. */
5105 /* nodeptr is fake for LEAF2 */
5111 mdb_cursor_adjust(MDB_cursor *mc, func)
5115 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5116 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5123 /** Pop a page off the top of the cursor's stack. */
5125 mdb_cursor_pop(MDB_cursor *mc)
5128 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5129 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5137 /** Push a page onto the top of the cursor's stack. */
5139 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5141 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5142 DDBI(mc), (void *) mc));
5144 if (mc->mc_snum >= CURSOR_STACK) {
5145 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5146 return MDB_CURSOR_FULL;
5149 mc->mc_top = mc->mc_snum++;
5150 mc->mc_pg[mc->mc_top] = mp;
5151 mc->mc_ki[mc->mc_top] = 0;
5156 /** Find the address of the page corresponding to a given page number.
5157 * @param[in] txn the transaction for this access.
5158 * @param[in] pgno the page number for the page to retrieve.
5159 * @param[out] ret address of a pointer where the page's address will be stored.
5160 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5161 * @return 0 on success, non-zero on failure.
5164 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5166 MDB_env *env = txn->mt_env;
5170 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5174 MDB_ID2L dl = tx2->mt_u.dirty_list;
5176 /* Spilled pages were dirtied in this txn and flushed
5177 * because the dirty list got full. Bring this page
5178 * back in from the map (but don't unspill it here,
5179 * leave that unless page_touch happens again).
5181 if (tx2->mt_spill_pgs) {
5182 MDB_ID pn = pgno << 1;
5183 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5184 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5185 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5190 unsigned x = mdb_mid2l_search(dl, pgno);
5191 if (x <= dl[0].mid && dl[x].mid == pgno) {
5197 } while ((tx2 = tx2->mt_parent) != NULL);
5200 if (pgno < txn->mt_next_pgno) {
5202 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5204 DPRINTF(("page %"Z"u not found", pgno));
5205 txn->mt_flags |= MDB_TXN_ERROR;
5206 return MDB_PAGE_NOTFOUND;
5216 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5217 * The cursor is at the root page, set up the rest of it.
5220 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5222 MDB_page *mp = mc->mc_pg[mc->mc_top];
5226 while (IS_BRANCH(mp)) {
5230 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5231 mdb_cassert(mc, NUMKEYS(mp) > 1);
5232 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5234 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5236 if (flags & MDB_PS_LAST)
5237 i = NUMKEYS(mp) - 1;
5240 node = mdb_node_search(mc, key, &exact);
5242 i = NUMKEYS(mp) - 1;
5244 i = mc->mc_ki[mc->mc_top];
5246 mdb_cassert(mc, i > 0);
5250 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5253 mdb_cassert(mc, i < NUMKEYS(mp));
5254 node = NODEPTR(mp, i);
5256 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5259 mc->mc_ki[mc->mc_top] = i;
5260 if ((rc = mdb_cursor_push(mc, mp)))
5263 if (flags & MDB_PS_MODIFY) {
5264 if ((rc = mdb_page_touch(mc)) != 0)
5266 mp = mc->mc_pg[mc->mc_top];
5271 DPRINTF(("internal error, index points to a %02X page!?",
5273 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5274 return MDB_CORRUPTED;
5277 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5278 key ? DKEY(key) : "null"));
5279 mc->mc_flags |= C_INITIALIZED;
5280 mc->mc_flags &= ~C_EOF;
5285 /** Search for the lowest key under the current branch page.
5286 * This just bypasses a NUMKEYS check in the current page
5287 * before calling mdb_page_search_root(), because the callers
5288 * are all in situations where the current page is known to
5292 mdb_page_search_lowest(MDB_cursor *mc)
5294 MDB_page *mp = mc->mc_pg[mc->mc_top];
5295 MDB_node *node = NODEPTR(mp, 0);
5298 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5301 mc->mc_ki[mc->mc_top] = 0;
5302 if ((rc = mdb_cursor_push(mc, mp)))
5304 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5307 /** Search for the page a given key should be in.
5308 * Push it and its parent pages on the cursor stack.
5309 * @param[in,out] mc the cursor for this operation.
5310 * @param[in] key the key to search for, or NULL for first/last page.
5311 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5312 * are touched (updated with new page numbers).
5313 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5314 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5315 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5316 * @return 0 on success, non-zero on failure.
5319 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5324 /* Make sure the txn is still viable, then find the root from
5325 * the txn's db table and set it as the root of the cursor's stack.
5327 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5328 DPUTS("transaction may not be used now");
5331 /* Make sure we're using an up-to-date root */
5332 if (*mc->mc_dbflag & DB_STALE) {
5334 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5336 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5337 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5344 MDB_node *leaf = mdb_node_search(&mc2,
5345 &mc->mc_dbx->md_name, &exact);
5347 return MDB_NOTFOUND;
5348 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5349 return MDB_INCOMPATIBLE; /* not a named DB */
5350 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5353 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5355 /* The txn may not know this DBI, or another process may
5356 * have dropped and recreated the DB with other flags.
5358 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5359 return MDB_INCOMPATIBLE;
5360 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5362 *mc->mc_dbflag &= ~DB_STALE;
5364 root = mc->mc_db->md_root;
5366 if (root == P_INVALID) { /* Tree is empty. */
5367 DPUTS("tree is empty");
5368 return MDB_NOTFOUND;
5372 mdb_cassert(mc, root > 1);
5373 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5374 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5380 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5381 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5383 if (flags & MDB_PS_MODIFY) {
5384 if ((rc = mdb_page_touch(mc)))
5388 if (flags & MDB_PS_ROOTONLY)
5391 return mdb_page_search_root(mc, key, flags);
5395 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5397 MDB_txn *txn = mc->mc_txn;
5398 pgno_t pg = mp->mp_pgno;
5399 unsigned x = 0, ovpages = mp->mp_pages;
5400 MDB_env *env = txn->mt_env;
5401 MDB_IDL sl = txn->mt_spill_pgs;
5402 MDB_ID pn = pg << 1;
5405 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5406 /* If the page is dirty or on the spill list we just acquired it,
5407 * so we should give it back to our current free list, if any.
5408 * Otherwise put it onto the list of pages we freed in this txn.
5410 * Won't create me_pghead: me_pglast must be inited along with it.
5411 * Unsupported in nested txns: They would need to hide the page
5412 * range in ancestor txns' dirty and spilled lists.
5414 if (env->me_pghead &&
5416 ((mp->mp_flags & P_DIRTY) ||
5417 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5421 MDB_ID2 *dl, ix, iy;
5422 rc = mdb_midl_need(&env->me_pghead, ovpages);
5425 if (!(mp->mp_flags & P_DIRTY)) {
5426 /* This page is no longer spilled */
5433 /* Remove from dirty list */
5434 dl = txn->mt_u.dirty_list;
5436 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5442 mdb_cassert(mc, x > 1);
5444 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5445 txn->mt_flags |= MDB_TXN_ERROR;
5446 return MDB_CORRUPTED;
5449 if (!(env->me_flags & MDB_WRITEMAP))
5450 mdb_dpage_free(env, mp);
5452 /* Insert in me_pghead */
5453 mop = env->me_pghead;
5454 j = mop[0] + ovpages;
5455 for (i = mop[0]; i && mop[i] < pg; i--)
5461 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5465 mc->mc_db->md_overflow_pages -= ovpages;
5469 /** Return the data associated with a given node.
5470 * @param[in] txn The transaction for this operation.
5471 * @param[in] leaf The node being read.
5472 * @param[out] data Updated to point to the node's data.
5473 * @return 0 on success, non-zero on failure.
5476 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5478 MDB_page *omp; /* overflow page */
5482 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5483 data->mv_size = NODEDSZ(leaf);
5484 data->mv_data = NODEDATA(leaf);
5488 /* Read overflow data.
5490 data->mv_size = NODEDSZ(leaf);
5491 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5492 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5493 DPRINTF(("read overflow page %"Z"u failed", pgno));
5496 data->mv_data = METADATA(omp);
5502 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5503 MDB_val *key, MDB_val *data)
5510 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5512 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5515 if (txn->mt_flags & MDB_TXN_BLOCKED)
5518 mdb_cursor_init(&mc, txn, dbi, &mx);
5519 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5522 /** Find a sibling for a page.
5523 * Replaces the page at the top of the cursor's stack with the
5524 * specified sibling, if one exists.
5525 * @param[in] mc The cursor for this operation.
5526 * @param[in] move_right Non-zero if the right sibling is requested,
5527 * otherwise the left sibling.
5528 * @return 0 on success, non-zero on failure.
5531 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5537 if (mc->mc_snum < 2) {
5538 return MDB_NOTFOUND; /* root has no siblings */
5542 DPRINTF(("parent page is page %"Z"u, index %u",
5543 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5545 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5546 : (mc->mc_ki[mc->mc_top] == 0)) {
5547 DPRINTF(("no more keys left, moving to %s sibling",
5548 move_right ? "right" : "left"));
5549 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5550 /* undo cursor_pop before returning */
5557 mc->mc_ki[mc->mc_top]++;
5559 mc->mc_ki[mc->mc_top]--;
5560 DPRINTF(("just moving to %s index key %u",
5561 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5563 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5565 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5566 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5567 /* mc will be inconsistent if caller does mc_snum++ as above */
5568 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5572 mdb_cursor_push(mc, mp);
5574 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5579 /** Move the cursor to the next data item. */
5581 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5587 if (mc->mc_flags & C_EOF) {
5588 return MDB_NOTFOUND;
5591 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5593 mp = mc->mc_pg[mc->mc_top];
5595 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5596 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5597 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5598 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5599 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5600 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5601 if (rc == MDB_SUCCESS)
5602 MDB_GET_KEY(leaf, key);
5607 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5608 if (op == MDB_NEXT_DUP)
5609 return MDB_NOTFOUND;
5613 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5614 mdb_dbg_pgno(mp), (void *) mc));
5615 if (mc->mc_flags & C_DEL)
5618 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5619 DPUTS("=====> move to next sibling page");
5620 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5621 mc->mc_flags |= C_EOF;
5624 mp = mc->mc_pg[mc->mc_top];
5625 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5627 mc->mc_ki[mc->mc_top]++;
5630 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5631 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5634 key->mv_size = mc->mc_db->md_pad;
5635 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5639 mdb_cassert(mc, IS_LEAF(mp));
5640 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5642 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5643 mdb_xcursor_init1(mc, leaf);
5646 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5649 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5650 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5651 if (rc != MDB_SUCCESS)
5656 MDB_GET_KEY(leaf, key);
5660 /** Move the cursor to the previous data item. */
5662 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5668 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5670 mp = mc->mc_pg[mc->mc_top];
5672 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5673 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5674 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5675 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5676 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5677 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5678 if (rc == MDB_SUCCESS) {
5679 MDB_GET_KEY(leaf, key);
5680 mc->mc_flags &= ~C_EOF;
5686 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5687 if (op == MDB_PREV_DUP)
5688 return MDB_NOTFOUND;
5692 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5693 mdb_dbg_pgno(mp), (void *) mc));
5695 if (mc->mc_ki[mc->mc_top] == 0) {
5696 DPUTS("=====> move to prev sibling page");
5697 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5700 mp = mc->mc_pg[mc->mc_top];
5701 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5702 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5704 mc->mc_ki[mc->mc_top]--;
5706 mc->mc_flags &= ~C_EOF;
5708 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5709 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5712 key->mv_size = mc->mc_db->md_pad;
5713 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5717 mdb_cassert(mc, IS_LEAF(mp));
5718 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5720 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5721 mdb_xcursor_init1(mc, leaf);
5724 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5727 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5728 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5729 if (rc != MDB_SUCCESS)
5734 MDB_GET_KEY(leaf, key);
5738 /** Set the cursor on a specific data item. */
5740 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5741 MDB_cursor_op op, int *exactp)
5745 MDB_node *leaf = NULL;
5748 if (key->mv_size == 0)
5749 return MDB_BAD_VALSIZE;
5752 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5754 /* See if we're already on the right page */
5755 if (mc->mc_flags & C_INITIALIZED) {
5758 mp = mc->mc_pg[mc->mc_top];
5760 mc->mc_ki[mc->mc_top] = 0;
5761 return MDB_NOTFOUND;
5763 if (mp->mp_flags & P_LEAF2) {
5764 nodekey.mv_size = mc->mc_db->md_pad;
5765 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5767 leaf = NODEPTR(mp, 0);
5768 MDB_GET_KEY2(leaf, nodekey);
5770 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5772 /* Probably happens rarely, but first node on the page
5773 * was the one we wanted.
5775 mc->mc_ki[mc->mc_top] = 0;
5782 unsigned int nkeys = NUMKEYS(mp);
5784 if (mp->mp_flags & P_LEAF2) {
5785 nodekey.mv_data = LEAF2KEY(mp,
5786 nkeys-1, nodekey.mv_size);
5788 leaf = NODEPTR(mp, nkeys-1);
5789 MDB_GET_KEY2(leaf, nodekey);
5791 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5793 /* last node was the one we wanted */
5794 mc->mc_ki[mc->mc_top] = nkeys-1;
5800 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5801 /* This is definitely the right page, skip search_page */
5802 if (mp->mp_flags & P_LEAF2) {
5803 nodekey.mv_data = LEAF2KEY(mp,
5804 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5806 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5807 MDB_GET_KEY2(leaf, nodekey);
5809 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5811 /* current node was the one we wanted */
5821 /* If any parents have right-sibs, search.
5822 * Otherwise, there's nothing further.
5824 for (i=0; i<mc->mc_top; i++)
5826 NUMKEYS(mc->mc_pg[i])-1)
5828 if (i == mc->mc_top) {
5829 /* There are no other pages */
5830 mc->mc_ki[mc->mc_top] = nkeys;
5831 return MDB_NOTFOUND;
5835 /* There are no other pages */
5836 mc->mc_ki[mc->mc_top] = 0;
5837 if (op == MDB_SET_RANGE && !exactp) {
5841 return MDB_NOTFOUND;
5845 rc = mdb_page_search(mc, key, 0);
5846 if (rc != MDB_SUCCESS)
5849 mp = mc->mc_pg[mc->mc_top];
5850 mdb_cassert(mc, IS_LEAF(mp));
5853 leaf = mdb_node_search(mc, key, exactp);
5854 if (exactp != NULL && !*exactp) {
5855 /* MDB_SET specified and not an exact match. */
5856 return MDB_NOTFOUND;
5860 DPUTS("===> inexact leaf not found, goto sibling");
5861 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5862 mc->mc_flags |= C_EOF;
5863 return rc; /* no entries matched */
5865 mp = mc->mc_pg[mc->mc_top];
5866 mdb_cassert(mc, IS_LEAF(mp));
5867 leaf = NODEPTR(mp, 0);
5871 mc->mc_flags |= C_INITIALIZED;
5872 mc->mc_flags &= ~C_EOF;
5875 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5876 key->mv_size = mc->mc_db->md_pad;
5877 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5882 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5883 mdb_xcursor_init1(mc, leaf);
5886 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5887 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5888 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5891 if (op == MDB_GET_BOTH) {
5897 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5898 if (rc != MDB_SUCCESS)
5901 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5904 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5906 dcmp = mc->mc_dbx->md_dcmp;
5907 #if UINT_MAX < SIZE_MAX
5908 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5909 dcmp = mdb_cmp_clong;
5911 rc = dcmp(data, &olddata);
5913 if (op == MDB_GET_BOTH || rc > 0)
5914 return MDB_NOTFOUND;
5921 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5922 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5927 /* The key already matches in all other cases */
5928 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5929 MDB_GET_KEY(leaf, key);
5930 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5935 /** Move the cursor to the first item in the database. */
5937 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5943 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5945 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5946 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5947 if (rc != MDB_SUCCESS)
5950 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5952 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5953 mc->mc_flags |= C_INITIALIZED;
5954 mc->mc_flags &= ~C_EOF;
5956 mc->mc_ki[mc->mc_top] = 0;
5958 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5959 key->mv_size = mc->mc_db->md_pad;
5960 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5965 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5966 mdb_xcursor_init1(mc, leaf);
5967 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5971 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5975 MDB_GET_KEY(leaf, key);
5979 /** Move the cursor to the last item in the database. */
5981 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5987 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5989 if (!(mc->mc_flags & C_EOF)) {
5991 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5992 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5993 if (rc != MDB_SUCCESS)
5996 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5999 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6000 mc->mc_flags |= C_INITIALIZED|C_EOF;
6001 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6003 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6004 key->mv_size = mc->mc_db->md_pad;
6005 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6010 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6011 mdb_xcursor_init1(mc, leaf);
6012 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6016 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6021 MDB_GET_KEY(leaf, key);
6026 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6031 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6036 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6040 case MDB_GET_CURRENT:
6041 if (!(mc->mc_flags & C_INITIALIZED)) {
6044 MDB_page *mp = mc->mc_pg[mc->mc_top];
6045 int nkeys = NUMKEYS(mp);
6046 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6047 mc->mc_ki[mc->mc_top] = nkeys;
6053 key->mv_size = mc->mc_db->md_pad;
6054 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6056 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6057 MDB_GET_KEY(leaf, key);
6059 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6060 if (mc->mc_flags & C_DEL)
6061 mdb_xcursor_init1(mc, leaf);
6062 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6064 rc = mdb_node_read(mc->mc_txn, leaf, data);
6071 case MDB_GET_BOTH_RANGE:
6076 if (mc->mc_xcursor == NULL) {
6077 rc = MDB_INCOMPATIBLE;
6087 rc = mdb_cursor_set(mc, key, data, op,
6088 op == MDB_SET_RANGE ? NULL : &exact);
6091 case MDB_GET_MULTIPLE:
6092 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6096 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6097 rc = MDB_INCOMPATIBLE;
6101 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6102 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6105 case MDB_NEXT_MULTIPLE:
6110 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6111 rc = MDB_INCOMPATIBLE;
6114 if (!(mc->mc_flags & C_INITIALIZED))
6115 rc = mdb_cursor_first(mc, key, data);
6117 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6118 if (rc == MDB_SUCCESS) {
6119 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6122 mx = &mc->mc_xcursor->mx_cursor;
6123 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6125 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6126 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6134 case MDB_NEXT_NODUP:
6135 if (!(mc->mc_flags & C_INITIALIZED))
6136 rc = mdb_cursor_first(mc, key, data);
6138 rc = mdb_cursor_next(mc, key, data, op);
6142 case MDB_PREV_NODUP:
6143 if (!(mc->mc_flags & C_INITIALIZED)) {
6144 rc = mdb_cursor_last(mc, key, data);
6147 mc->mc_flags |= C_INITIALIZED;
6148 mc->mc_ki[mc->mc_top]++;
6150 rc = mdb_cursor_prev(mc, key, data, op);
6153 rc = mdb_cursor_first(mc, key, data);
6156 mfunc = mdb_cursor_first;
6158 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6162 if (mc->mc_xcursor == NULL) {
6163 rc = MDB_INCOMPATIBLE;
6167 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6168 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6169 MDB_GET_KEY(leaf, key);
6170 rc = mdb_node_read(mc->mc_txn, leaf, data);
6174 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6178 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6181 rc = mdb_cursor_last(mc, key, data);
6184 mfunc = mdb_cursor_last;
6187 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6192 if (mc->mc_flags & C_DEL)
6193 mc->mc_flags ^= C_DEL;
6198 /** Touch all the pages in the cursor stack. Set mc_top.
6199 * Makes sure all the pages are writable, before attempting a write operation.
6200 * @param[in] mc The cursor to operate on.
6203 mdb_cursor_touch(MDB_cursor *mc)
6205 int rc = MDB_SUCCESS;
6207 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6210 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6212 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6213 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6216 *mc->mc_dbflag |= DB_DIRTY;
6221 rc = mdb_page_touch(mc);
6222 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6223 mc->mc_top = mc->mc_snum-1;
6228 /** Do not spill pages to disk if txn is getting full, may fail instead */
6229 #define MDB_NOSPILL 0x8000
6232 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6236 MDB_node *leaf = NULL;
6237 MDB_page *fp, *mp, *sub_root = NULL;
6239 MDB_val xdata, *rdata, dkey, olddata;
6241 int do_sub = 0, insert_key, insert_data;
6242 unsigned int mcount = 0, dcount = 0, nospill;
6245 unsigned int nflags;
6248 if (mc == NULL || key == NULL)
6251 env = mc->mc_txn->mt_env;
6253 /* Check this first so counter will always be zero on any
6256 if (flags & MDB_MULTIPLE) {
6257 dcount = data[1].mv_size;
6258 data[1].mv_size = 0;
6259 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6260 return MDB_INCOMPATIBLE;
6263 nospill = flags & MDB_NOSPILL;
6264 flags &= ~MDB_NOSPILL;
6266 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6267 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6269 if (key->mv_size-1 >= ENV_MAXKEY(env))
6270 return MDB_BAD_VALSIZE;
6272 #if SIZE_MAX > MAXDATASIZE
6273 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6274 return MDB_BAD_VALSIZE;
6276 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6277 return MDB_BAD_VALSIZE;
6280 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6281 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6285 if (flags == MDB_CURRENT) {
6286 if (!(mc->mc_flags & C_INITIALIZED))
6289 } else if (mc->mc_db->md_root == P_INVALID) {
6290 /* new database, cursor has nothing to point to */
6293 mc->mc_flags &= ~C_INITIALIZED;
6298 if (flags & MDB_APPEND) {
6300 rc = mdb_cursor_last(mc, &k2, &d2);
6302 rc = mc->mc_dbx->md_cmp(key, &k2);
6305 mc->mc_ki[mc->mc_top]++;
6307 /* new key is <= last key */
6312 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6314 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6315 DPRINTF(("duplicate key [%s]", DKEY(key)));
6317 return MDB_KEYEXIST;
6319 if (rc && rc != MDB_NOTFOUND)
6323 if (mc->mc_flags & C_DEL)
6324 mc->mc_flags ^= C_DEL;
6326 /* Cursor is positioned, check for room in the dirty list */
6328 if (flags & MDB_MULTIPLE) {
6330 xdata.mv_size = data->mv_size * dcount;
6334 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6338 if (rc == MDB_NO_ROOT) {
6340 /* new database, write a root leaf page */
6341 DPUTS("allocating new root leaf page");
6342 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6345 mdb_cursor_push(mc, np);
6346 mc->mc_db->md_root = np->mp_pgno;
6347 mc->mc_db->md_depth++;
6348 *mc->mc_dbflag |= DB_DIRTY;
6349 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6351 np->mp_flags |= P_LEAF2;
6352 mc->mc_flags |= C_INITIALIZED;
6354 /* make sure all cursor pages are writable */
6355 rc2 = mdb_cursor_touch(mc);
6360 insert_key = insert_data = rc;
6362 /* The key does not exist */
6363 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6364 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6365 LEAFSIZE(key, data) > env->me_nodemax)
6367 /* Too big for a node, insert in sub-DB. Set up an empty
6368 * "old sub-page" for prep_subDB to expand to a full page.
6370 fp_flags = P_LEAF|P_DIRTY;
6372 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6373 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6374 olddata.mv_size = PAGEHDRSZ;
6378 /* there's only a key anyway, so this is a no-op */
6379 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6381 unsigned int ksize = mc->mc_db->md_pad;
6382 if (key->mv_size != ksize)
6383 return MDB_BAD_VALSIZE;
6384 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6385 memcpy(ptr, key->mv_data, ksize);
6387 /* if overwriting slot 0 of leaf, need to
6388 * update branch key if there is a parent page
6390 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6391 unsigned short top = mc->mc_top;
6393 /* slot 0 is always an empty key, find real slot */
6394 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6396 if (mc->mc_ki[mc->mc_top])
6397 rc2 = mdb_update_key(mc, key);
6408 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6409 olddata.mv_size = NODEDSZ(leaf);
6410 olddata.mv_data = NODEDATA(leaf);
6413 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6414 /* Prepare (sub-)page/sub-DB to accept the new item,
6415 * if needed. fp: old sub-page or a header faking
6416 * it. mp: new (sub-)page. offset: growth in page
6417 * size. xdata: node data with new page or DB.
6419 unsigned i, offset = 0;
6420 mp = fp = xdata.mv_data = env->me_pbuf;
6421 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6423 /* Was a single item before, must convert now */
6424 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6426 /* Just overwrite the current item */
6427 if (flags == MDB_CURRENT)
6429 dcmp = mc->mc_dbx->md_dcmp;
6430 #if UINT_MAX < SIZE_MAX
6431 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6432 dcmp = mdb_cmp_clong;
6434 /* does data match? */
6435 if (!dcmp(data, &olddata)) {
6436 if (flags & MDB_NODUPDATA)
6437 return MDB_KEYEXIST;
6442 /* Back up original data item */
6443 dkey.mv_size = olddata.mv_size;
6444 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6446 /* Make sub-page header for the dup items, with dummy body */
6447 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6448 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6449 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6450 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6451 fp->mp_flags |= P_LEAF2;
6452 fp->mp_pad = data->mv_size;
6453 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6455 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6456 (dkey.mv_size & 1) + (data->mv_size & 1);
6458 fp->mp_upper = xdata.mv_size - PAGEBASE;
6459 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6460 } else if (leaf->mn_flags & F_SUBDATA) {
6461 /* Data is on sub-DB, just store it */
6462 flags |= F_DUPDATA|F_SUBDATA;
6465 /* Data is on sub-page */
6466 fp = olddata.mv_data;
6469 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6470 offset = EVEN(NODESIZE + sizeof(indx_t) +
6474 offset = fp->mp_pad;
6475 if (SIZELEFT(fp) < offset) {
6476 offset *= 4; /* space for 4 more */
6479 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6481 fp->mp_flags |= P_DIRTY;
6482 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6483 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6487 xdata.mv_size = olddata.mv_size + offset;
6490 fp_flags = fp->mp_flags;
6491 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6492 /* Too big for a sub-page, convert to sub-DB */
6493 fp_flags &= ~P_SUBP;
6495 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6496 fp_flags |= P_LEAF2;
6497 dummy.md_pad = fp->mp_pad;
6498 dummy.md_flags = MDB_DUPFIXED;
6499 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6500 dummy.md_flags |= MDB_INTEGERKEY;
6506 dummy.md_branch_pages = 0;
6507 dummy.md_leaf_pages = 1;
6508 dummy.md_overflow_pages = 0;
6509 dummy.md_entries = NUMKEYS(fp);
6510 xdata.mv_size = sizeof(MDB_db);
6511 xdata.mv_data = &dummy;
6512 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6514 offset = env->me_psize - olddata.mv_size;
6515 flags |= F_DUPDATA|F_SUBDATA;
6516 dummy.md_root = mp->mp_pgno;
6520 mp->mp_flags = fp_flags | P_DIRTY;
6521 mp->mp_pad = fp->mp_pad;
6522 mp->mp_lower = fp->mp_lower;
6523 mp->mp_upper = fp->mp_upper + offset;
6524 if (fp_flags & P_LEAF2) {
6525 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6527 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6528 olddata.mv_size - fp->mp_upper - PAGEBASE);
6529 for (i=0; i<NUMKEYS(fp); i++)
6530 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6538 mdb_node_del(mc, 0);
6542 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6543 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6544 return MDB_INCOMPATIBLE;
6545 /* overflow page overwrites need special handling */
6546 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6549 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6551 memcpy(&pg, olddata.mv_data, sizeof(pg));
6552 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6554 ovpages = omp->mp_pages;
6556 /* Is the ov page large enough? */
6557 if (ovpages >= dpages) {
6558 if (!(omp->mp_flags & P_DIRTY) &&
6559 (level || (env->me_flags & MDB_WRITEMAP)))
6561 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6564 level = 0; /* dirty in this txn or clean */
6567 if (omp->mp_flags & P_DIRTY) {
6568 /* yes, overwrite it. Note in this case we don't
6569 * bother to try shrinking the page if the new data
6570 * is smaller than the overflow threshold.
6573 /* It is writable only in a parent txn */
6574 size_t sz = (size_t) env->me_psize * ovpages, off;
6575 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6581 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6582 mdb_cassert(mc, rc2 == 0);
6583 if (!(flags & MDB_RESERVE)) {
6584 /* Copy end of page, adjusting alignment so
6585 * compiler may copy words instead of bytes.
6587 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6588 memcpy((size_t *)((char *)np + off),
6589 (size_t *)((char *)omp + off), sz - off);
6592 memcpy(np, omp, sz); /* Copy beginning of page */
6595 SETDSZ(leaf, data->mv_size);
6596 if (F_ISSET(flags, MDB_RESERVE))
6597 data->mv_data = METADATA(omp);
6599 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6603 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6605 } else if (data->mv_size == olddata.mv_size) {
6606 /* same size, just replace it. Note that we could
6607 * also reuse this node if the new data is smaller,
6608 * but instead we opt to shrink the node in that case.
6610 if (F_ISSET(flags, MDB_RESERVE))
6611 data->mv_data = olddata.mv_data;
6612 else if (!(mc->mc_flags & C_SUB))
6613 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6615 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6620 mdb_node_del(mc, 0);
6626 nflags = flags & NODE_ADD_FLAGS;
6627 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6628 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6629 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6630 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6632 nflags |= MDB_SPLIT_REPLACE;
6633 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6635 /* There is room already in this leaf page. */
6636 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6637 if (rc == 0 && insert_key) {
6638 /* Adjust other cursors pointing to mp */
6639 MDB_cursor *m2, *m3;
6640 MDB_dbi dbi = mc->mc_dbi;
6641 unsigned i = mc->mc_top;
6642 MDB_page *mp = mc->mc_pg[i];
6644 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6645 if (mc->mc_flags & C_SUB)
6646 m3 = &m2->mc_xcursor->mx_cursor;
6649 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6650 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6657 if (rc == MDB_SUCCESS) {
6658 /* Now store the actual data in the child DB. Note that we're
6659 * storing the user data in the keys field, so there are strict
6660 * size limits on dupdata. The actual data fields of the child
6661 * DB are all zero size.
6664 int xflags, new_dupdata;
6669 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6670 if (flags & MDB_CURRENT) {
6671 xflags = MDB_CURRENT|MDB_NOSPILL;
6673 mdb_xcursor_init1(mc, leaf);
6674 xflags = (flags & MDB_NODUPDATA) ?
6675 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6678 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6679 new_dupdata = (int)dkey.mv_size;
6680 /* converted, write the original data first */
6682 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6685 /* we've done our job */
6688 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6689 /* Adjust other cursors pointing to mp */
6691 MDB_xcursor *mx = mc->mc_xcursor;
6692 unsigned i = mc->mc_top;
6693 MDB_page *mp = mc->mc_pg[i];
6695 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6696 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6697 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6698 if (m2->mc_pg[i] == mp) {
6699 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6700 mdb_xcursor_init2(m2, mx, new_dupdata);
6701 } else if (!insert_key) {
6702 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6703 if (!(n2->mn_flags & F_SUBDATA))
6704 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6709 ecount = mc->mc_xcursor->mx_db.md_entries;
6710 if (flags & MDB_APPENDDUP)
6711 xflags |= MDB_APPEND;
6712 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6713 if (flags & F_SUBDATA) {
6714 void *db = NODEDATA(leaf);
6715 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6717 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6719 /* Increment count unless we just replaced an existing item. */
6721 mc->mc_db->md_entries++;
6723 /* Invalidate txn if we created an empty sub-DB */
6726 /* If we succeeded and the key didn't exist before,
6727 * make sure the cursor is marked valid.
6729 mc->mc_flags |= C_INITIALIZED;
6731 if (flags & MDB_MULTIPLE) {
6734 /* let caller know how many succeeded, if any */
6735 data[1].mv_size = mcount;
6736 if (mcount < dcount) {
6737 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6738 insert_key = insert_data = 0;
6745 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6748 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6753 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6759 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6760 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6762 if (!(mc->mc_flags & C_INITIALIZED))
6765 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6766 return MDB_NOTFOUND;
6768 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6771 rc = mdb_cursor_touch(mc);
6775 mp = mc->mc_pg[mc->mc_top];
6778 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6780 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6781 if (flags & MDB_NODUPDATA) {
6782 /* mdb_cursor_del0() will subtract the final entry */
6783 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6785 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6786 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6788 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6791 /* If sub-DB still has entries, we're done */
6792 if (mc->mc_xcursor->mx_db.md_entries) {
6793 if (leaf->mn_flags & F_SUBDATA) {
6794 /* update subDB info */
6795 void *db = NODEDATA(leaf);
6796 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6799 /* shrink fake page */
6800 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6801 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6802 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6803 /* fix other sub-DB cursors pointed at fake pages on this page */
6804 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6805 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6806 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6807 if (m2->mc_pg[mc->mc_top] == mp) {
6808 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6809 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6811 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6812 if (!(n2->mn_flags & F_SUBDATA))
6813 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6818 mc->mc_db->md_entries--;
6819 mc->mc_flags |= C_DEL;
6822 /* otherwise fall thru and delete the sub-DB */
6825 if (leaf->mn_flags & F_SUBDATA) {
6826 /* add all the child DB's pages to the free list */
6827 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6832 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6833 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6834 rc = MDB_INCOMPATIBLE;
6838 /* add overflow pages to free list */
6839 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6843 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6844 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6845 (rc = mdb_ovpage_free(mc, omp)))
6850 return mdb_cursor_del0(mc);
6853 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6857 /** Allocate and initialize new pages for a database.
6858 * @param[in] mc a cursor on the database being added to.
6859 * @param[in] flags flags defining what type of page is being allocated.
6860 * @param[in] num the number of pages to allocate. This is usually 1,
6861 * unless allocating overflow pages for a large record.
6862 * @param[out] mp Address of a page, or NULL on failure.
6863 * @return 0 on success, non-zero on failure.
6866 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6871 if ((rc = mdb_page_alloc(mc, num, &np)))
6873 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6874 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6875 np->mp_flags = flags | P_DIRTY;
6876 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6877 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6880 mc->mc_db->md_branch_pages++;
6881 else if (IS_LEAF(np))
6882 mc->mc_db->md_leaf_pages++;
6883 else if (IS_OVERFLOW(np)) {
6884 mc->mc_db->md_overflow_pages += num;
6892 /** Calculate the size of a leaf node.
6893 * The size depends on the environment's page size; if a data item
6894 * is too large it will be put onto an overflow page and the node
6895 * size will only include the key and not the data. Sizes are always
6896 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6897 * of the #MDB_node headers.
6898 * @param[in] env The environment handle.
6899 * @param[in] key The key for the node.
6900 * @param[in] data The data for the node.
6901 * @return The number of bytes needed to store the node.
6904 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6908 sz = LEAFSIZE(key, data);
6909 if (sz > env->me_nodemax) {
6910 /* put on overflow page */
6911 sz -= data->mv_size - sizeof(pgno_t);
6914 return EVEN(sz + sizeof(indx_t));
6917 /** Calculate the size of a branch node.
6918 * The size should depend on the environment's page size but since
6919 * we currently don't support spilling large keys onto overflow
6920 * pages, it's simply the size of the #MDB_node header plus the
6921 * size of the key. Sizes are always rounded up to an even number
6922 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6923 * @param[in] env The environment handle.
6924 * @param[in] key The key for the node.
6925 * @return The number of bytes needed to store the node.
6928 mdb_branch_size(MDB_env *env, MDB_val *key)
6933 if (sz > env->me_nodemax) {
6934 /* put on overflow page */
6935 /* not implemented */
6936 /* sz -= key->size - sizeof(pgno_t); */
6939 return sz + sizeof(indx_t);
6942 /** Add a node to the page pointed to by the cursor.
6943 * @param[in] mc The cursor for this operation.
6944 * @param[in] indx The index on the page where the new node should be added.
6945 * @param[in] key The key for the new node.
6946 * @param[in] data The data for the new node, if any.
6947 * @param[in] pgno The page number, if adding a branch node.
6948 * @param[in] flags Flags for the node.
6949 * @return 0 on success, non-zero on failure. Possible errors are:
6951 * <li>ENOMEM - failed to allocate overflow pages for the node.
6952 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6953 * should never happen since all callers already calculate the
6954 * page's free space before calling this function.
6958 mdb_node_add(MDB_cursor *mc, indx_t indx,
6959 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6962 size_t node_size = NODESIZE;
6966 MDB_page *mp = mc->mc_pg[mc->mc_top];
6967 MDB_page *ofp = NULL; /* overflow page */
6971 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6973 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6974 IS_LEAF(mp) ? "leaf" : "branch",
6975 IS_SUBP(mp) ? "sub-" : "",
6976 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6977 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6980 /* Move higher keys up one slot. */
6981 int ksize = mc->mc_db->md_pad, dif;
6982 char *ptr = LEAF2KEY(mp, indx, ksize);
6983 dif = NUMKEYS(mp) - indx;
6985 memmove(ptr+ksize, ptr, dif*ksize);
6986 /* insert new key */
6987 memcpy(ptr, key->mv_data, ksize);
6989 /* Just using these for counting */
6990 mp->mp_lower += sizeof(indx_t);
6991 mp->mp_upper -= ksize - sizeof(indx_t);
6995 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6997 node_size += key->mv_size;
6999 mdb_cassert(mc, key && data);
7000 if (F_ISSET(flags, F_BIGDATA)) {
7001 /* Data already on overflow page. */
7002 node_size += sizeof(pgno_t);
7003 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7004 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7006 /* Put data on overflow page. */
7007 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7008 data->mv_size, node_size+data->mv_size));
7009 node_size = EVEN(node_size + sizeof(pgno_t));
7010 if ((ssize_t)node_size > room)
7012 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7014 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7018 node_size += data->mv_size;
7021 node_size = EVEN(node_size);
7022 if ((ssize_t)node_size > room)
7026 /* Move higher pointers up one slot. */
7027 for (i = NUMKEYS(mp); i > indx; i--)
7028 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7030 /* Adjust free space offsets. */
7031 ofs = mp->mp_upper - node_size;
7032 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7033 mp->mp_ptrs[indx] = ofs;
7035 mp->mp_lower += sizeof(indx_t);
7037 /* Write the node data. */
7038 node = NODEPTR(mp, indx);
7039 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7040 node->mn_flags = flags;
7042 SETDSZ(node,data->mv_size);
7047 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7050 ndata = NODEDATA(node);
7052 if (F_ISSET(flags, F_BIGDATA))
7053 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7054 else if (F_ISSET(flags, MDB_RESERVE))
7055 data->mv_data = ndata;
7057 memcpy(ndata, data->mv_data, data->mv_size);
7059 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7060 ndata = METADATA(ofp);
7061 if (F_ISSET(flags, MDB_RESERVE))
7062 data->mv_data = ndata;
7064 memcpy(ndata, data->mv_data, data->mv_size);
7071 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7072 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7073 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7074 DPRINTF(("node size = %"Z"u", node_size));
7075 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7076 return MDB_PAGE_FULL;
7079 /** Delete the specified node from a page.
7080 * @param[in] mc Cursor pointing to the node to delete.
7081 * @param[in] ksize The size of a node. Only used if the page is
7082 * part of a #MDB_DUPFIXED database.
7085 mdb_node_del(MDB_cursor *mc, int ksize)
7087 MDB_page *mp = mc->mc_pg[mc->mc_top];
7088 indx_t indx = mc->mc_ki[mc->mc_top];
7090 indx_t i, j, numkeys, ptr;
7094 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7095 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7096 numkeys = NUMKEYS(mp);
7097 mdb_cassert(mc, indx < numkeys);
7100 int x = numkeys - 1 - indx;
7101 base = LEAF2KEY(mp, indx, ksize);
7103 memmove(base, base + ksize, x * ksize);
7104 mp->mp_lower -= sizeof(indx_t);
7105 mp->mp_upper += ksize - sizeof(indx_t);
7109 node = NODEPTR(mp, indx);
7110 sz = NODESIZE + node->mn_ksize;
7112 if (F_ISSET(node->mn_flags, F_BIGDATA))
7113 sz += sizeof(pgno_t);
7115 sz += NODEDSZ(node);
7119 ptr = mp->mp_ptrs[indx];
7120 for (i = j = 0; i < numkeys; i++) {
7122 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7123 if (mp->mp_ptrs[i] < ptr)
7124 mp->mp_ptrs[j] += sz;
7129 base = (char *)mp + mp->mp_upper + PAGEBASE;
7130 memmove(base + sz, base, ptr - mp->mp_upper);
7132 mp->mp_lower -= sizeof(indx_t);
7136 /** Compact the main page after deleting a node on a subpage.
7137 * @param[in] mp The main page to operate on.
7138 * @param[in] indx The index of the subpage on the main page.
7141 mdb_node_shrink(MDB_page *mp, indx_t indx)
7146 indx_t delta, nsize, len, ptr;
7149 node = NODEPTR(mp, indx);
7150 sp = (MDB_page *)NODEDATA(node);
7151 delta = SIZELEFT(sp);
7152 nsize = NODEDSZ(node) - delta;
7154 /* Prepare to shift upward, set len = length(subpage part to shift) */
7158 return; /* do not make the node uneven-sized */
7160 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7161 for (i = NUMKEYS(sp); --i >= 0; )
7162 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7165 sp->mp_upper = sp->mp_lower;
7166 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7167 SETDSZ(node, nsize);
7169 /* Shift <lower nodes...initial part of subpage> upward */
7170 base = (char *)mp + mp->mp_upper + PAGEBASE;
7171 memmove(base + delta, base, (char *)sp + len - base);
7173 ptr = mp->mp_ptrs[indx];
7174 for (i = NUMKEYS(mp); --i >= 0; ) {
7175 if (mp->mp_ptrs[i] <= ptr)
7176 mp->mp_ptrs[i] += delta;
7178 mp->mp_upper += delta;
7181 /** Initial setup of a sorted-dups cursor.
7182 * Sorted duplicates are implemented as a sub-database for the given key.
7183 * The duplicate data items are actually keys of the sub-database.
7184 * Operations on the duplicate data items are performed using a sub-cursor
7185 * initialized when the sub-database is first accessed. This function does
7186 * the preliminary setup of the sub-cursor, filling in the fields that
7187 * depend only on the parent DB.
7188 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7191 mdb_xcursor_init0(MDB_cursor *mc)
7193 MDB_xcursor *mx = mc->mc_xcursor;
7195 mx->mx_cursor.mc_xcursor = NULL;
7196 mx->mx_cursor.mc_txn = mc->mc_txn;
7197 mx->mx_cursor.mc_db = &mx->mx_db;
7198 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7199 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7200 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7201 mx->mx_cursor.mc_snum = 0;
7202 mx->mx_cursor.mc_top = 0;
7203 mx->mx_cursor.mc_flags = C_SUB;
7204 mx->mx_dbx.md_name.mv_size = 0;
7205 mx->mx_dbx.md_name.mv_data = NULL;
7206 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7207 mx->mx_dbx.md_dcmp = NULL;
7208 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7211 /** Final setup of a sorted-dups cursor.
7212 * Sets up the fields that depend on the data from the main cursor.
7213 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7214 * @param[in] node The data containing the #MDB_db record for the
7215 * sorted-dup database.
7218 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7220 MDB_xcursor *mx = mc->mc_xcursor;
7222 if (node->mn_flags & F_SUBDATA) {
7223 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7224 mx->mx_cursor.mc_pg[0] = 0;
7225 mx->mx_cursor.mc_snum = 0;
7226 mx->mx_cursor.mc_top = 0;
7227 mx->mx_cursor.mc_flags = C_SUB;
7229 MDB_page *fp = NODEDATA(node);
7230 mx->mx_db.md_pad = 0;
7231 mx->mx_db.md_flags = 0;
7232 mx->mx_db.md_depth = 1;
7233 mx->mx_db.md_branch_pages = 0;
7234 mx->mx_db.md_leaf_pages = 1;
7235 mx->mx_db.md_overflow_pages = 0;
7236 mx->mx_db.md_entries = NUMKEYS(fp);
7237 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7238 mx->mx_cursor.mc_snum = 1;
7239 mx->mx_cursor.mc_top = 0;
7240 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7241 mx->mx_cursor.mc_pg[0] = fp;
7242 mx->mx_cursor.mc_ki[0] = 0;
7243 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7244 mx->mx_db.md_flags = MDB_DUPFIXED;
7245 mx->mx_db.md_pad = fp->mp_pad;
7246 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7247 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7250 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7251 mx->mx_db.md_root));
7252 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7253 #if UINT_MAX < SIZE_MAX
7254 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7255 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7260 /** Fixup a sorted-dups cursor due to underlying update.
7261 * Sets up some fields that depend on the data from the main cursor.
7262 * Almost the same as init1, but skips initialization steps if the
7263 * xcursor had already been used.
7264 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7265 * @param[in] src_mx The xcursor of an up-to-date cursor.
7266 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7269 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7271 MDB_xcursor *mx = mc->mc_xcursor;
7274 mx->mx_cursor.mc_snum = 1;
7275 mx->mx_cursor.mc_top = 0;
7276 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7277 mx->mx_cursor.mc_ki[0] = 0;
7278 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7279 #if UINT_MAX < SIZE_MAX
7280 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7282 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7285 mx->mx_db = src_mx->mx_db;
7286 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7287 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7288 mx->mx_db.md_root));
7291 /** Initialize a cursor for a given transaction and database. */
7293 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7296 mc->mc_backup = NULL;
7299 mc->mc_db = &txn->mt_dbs[dbi];
7300 mc->mc_dbx = &txn->mt_dbxs[dbi];
7301 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7307 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7308 mdb_tassert(txn, mx != NULL);
7309 mc->mc_xcursor = mx;
7310 mdb_xcursor_init0(mc);
7312 mc->mc_xcursor = NULL;
7314 if (*mc->mc_dbflag & DB_STALE) {
7315 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7320 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7323 size_t size = sizeof(MDB_cursor);
7325 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7328 if (txn->mt_flags & MDB_TXN_BLOCKED)
7331 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7334 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7335 size += sizeof(MDB_xcursor);
7337 if ((mc = malloc(size)) != NULL) {
7338 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7339 if (txn->mt_cursors) {
7340 mc->mc_next = txn->mt_cursors[dbi];
7341 txn->mt_cursors[dbi] = mc;
7342 mc->mc_flags |= C_UNTRACK;
7354 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7356 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7359 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7362 if (txn->mt_flags & MDB_TXN_BLOCKED)
7365 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7369 /* Return the count of duplicate data items for the current key */
7371 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7375 if (mc == NULL || countp == NULL)
7378 if (mc->mc_xcursor == NULL)
7379 return MDB_INCOMPATIBLE;
7381 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7384 if (!(mc->mc_flags & C_INITIALIZED))
7387 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7388 return MDB_NOTFOUND;
7390 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7391 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7394 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7397 *countp = mc->mc_xcursor->mx_db.md_entries;
7403 mdb_cursor_close(MDB_cursor *mc)
7405 if (mc && !mc->mc_backup) {
7406 /* remove from txn, if tracked */
7407 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7408 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7409 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7411 *prev = mc->mc_next;
7418 mdb_cursor_txn(MDB_cursor *mc)
7420 if (!mc) return NULL;
7425 mdb_cursor_dbi(MDB_cursor *mc)
7430 /** Replace the key for a branch node with a new key.
7431 * @param[in] mc Cursor pointing to the node to operate on.
7432 * @param[in] key The new key to use.
7433 * @return 0 on success, non-zero on failure.
7436 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7442 int delta, ksize, oksize;
7443 indx_t ptr, i, numkeys, indx;
7446 indx = mc->mc_ki[mc->mc_top];
7447 mp = mc->mc_pg[mc->mc_top];
7448 node = NODEPTR(mp, indx);
7449 ptr = mp->mp_ptrs[indx];
7453 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7454 k2.mv_data = NODEKEY(node);
7455 k2.mv_size = node->mn_ksize;
7456 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7458 mdb_dkey(&k2, kbuf2),
7464 /* Sizes must be 2-byte aligned. */
7465 ksize = EVEN(key->mv_size);
7466 oksize = EVEN(node->mn_ksize);
7467 delta = ksize - oksize;
7469 /* Shift node contents if EVEN(key length) changed. */
7471 if (delta > 0 && SIZELEFT(mp) < delta) {
7473 /* not enough space left, do a delete and split */
7474 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7475 pgno = NODEPGNO(node);
7476 mdb_node_del(mc, 0);
7477 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7480 numkeys = NUMKEYS(mp);
7481 for (i = 0; i < numkeys; i++) {
7482 if (mp->mp_ptrs[i] <= ptr)
7483 mp->mp_ptrs[i] -= delta;
7486 base = (char *)mp + mp->mp_upper + PAGEBASE;
7487 len = ptr - mp->mp_upper + NODESIZE;
7488 memmove(base - delta, base, len);
7489 mp->mp_upper -= delta;
7491 node = NODEPTR(mp, indx);
7494 /* But even if no shift was needed, update ksize */
7495 if (node->mn_ksize != key->mv_size)
7496 node->mn_ksize = key->mv_size;
7499 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7505 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7507 /** Move a node from csrc to cdst.
7510 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7517 unsigned short flags;
7521 /* Mark src and dst as dirty. */
7522 if ((rc = mdb_page_touch(csrc)) ||
7523 (rc = mdb_page_touch(cdst)))
7526 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7527 key.mv_size = csrc->mc_db->md_pad;
7528 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7530 data.mv_data = NULL;
7534 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7535 mdb_cassert(csrc, !((size_t)srcnode & 1));
7536 srcpg = NODEPGNO(srcnode);
7537 flags = srcnode->mn_flags;
7538 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7539 unsigned int snum = csrc->mc_snum;
7541 /* must find the lowest key below src */
7542 rc = mdb_page_search_lowest(csrc);
7545 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7546 key.mv_size = csrc->mc_db->md_pad;
7547 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7549 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7550 key.mv_size = NODEKSZ(s2);
7551 key.mv_data = NODEKEY(s2);
7553 csrc->mc_snum = snum--;
7554 csrc->mc_top = snum;
7556 key.mv_size = NODEKSZ(srcnode);
7557 key.mv_data = NODEKEY(srcnode);
7559 data.mv_size = NODEDSZ(srcnode);
7560 data.mv_data = NODEDATA(srcnode);
7562 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7563 unsigned int snum = cdst->mc_snum;
7566 /* must find the lowest key below dst */
7567 mdb_cursor_copy(cdst, &mn);
7568 rc = mdb_page_search_lowest(&mn);
7571 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7572 bkey.mv_size = mn.mc_db->md_pad;
7573 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7575 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7576 bkey.mv_size = NODEKSZ(s2);
7577 bkey.mv_data = NODEKEY(s2);
7579 mn.mc_snum = snum--;
7582 rc = mdb_update_key(&mn, &bkey);
7587 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7588 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7589 csrc->mc_ki[csrc->mc_top],
7591 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7592 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7594 /* Add the node to the destination page.
7596 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7597 if (rc != MDB_SUCCESS)
7600 /* Delete the node from the source page.
7602 mdb_node_del(csrc, key.mv_size);
7605 /* Adjust other cursors pointing to mp */
7606 MDB_cursor *m2, *m3;
7607 MDB_dbi dbi = csrc->mc_dbi;
7608 MDB_page *mpd, *mps;
7610 mps = csrc->mc_pg[csrc->mc_top];
7611 /* If we're adding on the left, bump others up */
7612 if (!cdst->mc_ki[csrc->mc_top]) {
7613 mpd = cdst->mc_pg[csrc->mc_top];
7614 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7615 if (csrc->mc_flags & C_SUB)
7616 m3 = &m2->mc_xcursor->mx_cursor;
7620 m3->mc_pg[csrc->mc_top] == mpd &&
7621 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7622 m3->mc_ki[csrc->mc_top]++;
7625 m3->mc_pg[csrc->mc_top] == mps &&
7626 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7627 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7628 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7629 m3->mc_ki[csrc->mc_top-1]++;
7633 /* Adding on the right, bump others down */
7635 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7636 if (csrc->mc_flags & C_SUB)
7637 m3 = &m2->mc_xcursor->mx_cursor;
7640 if (m3 == csrc) continue;
7641 if (m3->mc_pg[csrc->mc_top] == mps) {
7642 if (!m3->mc_ki[csrc->mc_top]) {
7643 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7644 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7645 m3->mc_ki[csrc->mc_top-1]--;
7647 m3->mc_ki[csrc->mc_top]--;
7654 /* Update the parent separators.
7656 if (csrc->mc_ki[csrc->mc_top] == 0) {
7657 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7658 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7659 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7661 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7662 key.mv_size = NODEKSZ(srcnode);
7663 key.mv_data = NODEKEY(srcnode);
7665 DPRINTF(("update separator for source page %"Z"u to [%s]",
7666 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7667 mdb_cursor_copy(csrc, &mn);
7670 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7673 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7675 indx_t ix = csrc->mc_ki[csrc->mc_top];
7676 nullkey.mv_size = 0;
7677 csrc->mc_ki[csrc->mc_top] = 0;
7678 rc = mdb_update_key(csrc, &nullkey);
7679 csrc->mc_ki[csrc->mc_top] = ix;
7680 mdb_cassert(csrc, rc == MDB_SUCCESS);
7684 if (cdst->mc_ki[cdst->mc_top] == 0) {
7685 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7686 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7687 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7689 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7690 key.mv_size = NODEKSZ(srcnode);
7691 key.mv_data = NODEKEY(srcnode);
7693 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7694 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7695 mdb_cursor_copy(cdst, &mn);
7698 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7701 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7703 indx_t ix = cdst->mc_ki[cdst->mc_top];
7704 nullkey.mv_size = 0;
7705 cdst->mc_ki[cdst->mc_top] = 0;
7706 rc = mdb_update_key(cdst, &nullkey);
7707 cdst->mc_ki[cdst->mc_top] = ix;
7708 mdb_cassert(cdst, rc == MDB_SUCCESS);
7715 /** Merge one page into another.
7716 * The nodes from the page pointed to by \b csrc will
7717 * be copied to the page pointed to by \b cdst and then
7718 * the \b csrc page will be freed.
7719 * @param[in] csrc Cursor pointing to the source page.
7720 * @param[in] cdst Cursor pointing to the destination page.
7721 * @return 0 on success, non-zero on failure.
7724 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7726 MDB_page *psrc, *pdst;
7733 psrc = csrc->mc_pg[csrc->mc_top];
7734 pdst = cdst->mc_pg[cdst->mc_top];
7736 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7738 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7739 mdb_cassert(csrc, cdst->mc_snum > 1);
7741 /* Mark dst as dirty. */
7742 if ((rc = mdb_page_touch(cdst)))
7745 /* get dst page again now that we've touched it. */
7746 pdst = cdst->mc_pg[cdst->mc_top];
7748 /* Move all nodes from src to dst.
7750 j = nkeys = NUMKEYS(pdst);
7751 if (IS_LEAF2(psrc)) {
7752 key.mv_size = csrc->mc_db->md_pad;
7753 key.mv_data = METADATA(psrc);
7754 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7755 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7756 if (rc != MDB_SUCCESS)
7758 key.mv_data = (char *)key.mv_data + key.mv_size;
7761 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7762 srcnode = NODEPTR(psrc, i);
7763 if (i == 0 && IS_BRANCH(psrc)) {
7766 mdb_cursor_copy(csrc, &mn);
7767 /* must find the lowest key below src */
7768 rc = mdb_page_search_lowest(&mn);
7771 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7772 key.mv_size = mn.mc_db->md_pad;
7773 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7775 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7776 key.mv_size = NODEKSZ(s2);
7777 key.mv_data = NODEKEY(s2);
7780 key.mv_size = srcnode->mn_ksize;
7781 key.mv_data = NODEKEY(srcnode);
7784 data.mv_size = NODEDSZ(srcnode);
7785 data.mv_data = NODEDATA(srcnode);
7786 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7787 if (rc != MDB_SUCCESS)
7792 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7793 pdst->mp_pgno, NUMKEYS(pdst),
7794 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7796 /* Unlink the src page from parent and add to free list.
7799 mdb_node_del(csrc, 0);
7800 if (csrc->mc_ki[csrc->mc_top] == 0) {
7802 rc = mdb_update_key(csrc, &key);
7810 psrc = csrc->mc_pg[csrc->mc_top];
7811 /* If not operating on FreeDB, allow this page to be reused
7812 * in this txn. Otherwise just add to free list.
7814 rc = mdb_page_loose(csrc, psrc);
7818 csrc->mc_db->md_leaf_pages--;
7820 csrc->mc_db->md_branch_pages--;
7822 /* Adjust other cursors pointing to mp */
7823 MDB_cursor *m2, *m3;
7824 MDB_dbi dbi = csrc->mc_dbi;
7825 unsigned int top = csrc->mc_top;
7827 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7828 if (csrc->mc_flags & C_SUB)
7829 m3 = &m2->mc_xcursor->mx_cursor;
7832 if (m3 == csrc) continue;
7833 if (m3->mc_snum < csrc->mc_snum) continue;
7834 if (m3->mc_pg[top] == psrc) {
7835 m3->mc_pg[top] = pdst;
7836 m3->mc_ki[top] += nkeys;
7837 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
7842 unsigned int snum = cdst->mc_snum;
7843 uint16_t depth = cdst->mc_db->md_depth;
7844 mdb_cursor_pop(cdst);
7845 rc = mdb_rebalance(cdst);
7846 /* Did the tree height change? */
7847 if (depth != cdst->mc_db->md_depth)
7848 snum += cdst->mc_db->md_depth - depth;
7849 cdst->mc_snum = snum;
7850 cdst->mc_top = snum-1;
7855 /** Copy the contents of a cursor.
7856 * @param[in] csrc The cursor to copy from.
7857 * @param[out] cdst The cursor to copy to.
7860 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7864 cdst->mc_txn = csrc->mc_txn;
7865 cdst->mc_dbi = csrc->mc_dbi;
7866 cdst->mc_db = csrc->mc_db;
7867 cdst->mc_dbx = csrc->mc_dbx;
7868 cdst->mc_snum = csrc->mc_snum;
7869 cdst->mc_top = csrc->mc_top;
7870 cdst->mc_flags = csrc->mc_flags;
7872 for (i=0; i<csrc->mc_snum; i++) {
7873 cdst->mc_pg[i] = csrc->mc_pg[i];
7874 cdst->mc_ki[i] = csrc->mc_ki[i];
7878 /** Rebalance the tree after a delete operation.
7879 * @param[in] mc Cursor pointing to the page where rebalancing
7881 * @return 0 on success, non-zero on failure.
7884 mdb_rebalance(MDB_cursor *mc)
7888 unsigned int ptop, minkeys, thresh;
7892 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7897 thresh = FILL_THRESHOLD;
7899 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7900 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7901 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7902 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7904 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7905 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7906 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7907 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7911 if (mc->mc_snum < 2) {
7912 MDB_page *mp = mc->mc_pg[0];
7914 DPUTS("Can't rebalance a subpage, ignoring");
7917 if (NUMKEYS(mp) == 0) {
7918 DPUTS("tree is completely empty");
7919 mc->mc_db->md_root = P_INVALID;
7920 mc->mc_db->md_depth = 0;
7921 mc->mc_db->md_leaf_pages = 0;
7922 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7925 /* Adjust cursors pointing to mp */
7928 mc->mc_flags &= ~C_INITIALIZED;
7930 MDB_cursor *m2, *m3;
7931 MDB_dbi dbi = mc->mc_dbi;
7933 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7934 if (mc->mc_flags & C_SUB)
7935 m3 = &m2->mc_xcursor->mx_cursor;
7938 if (m3->mc_snum < mc->mc_snum) continue;
7939 if (m3->mc_pg[0] == mp) {
7942 m3->mc_flags &= ~C_INITIALIZED;
7946 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7948 DPUTS("collapsing root page!");
7949 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7952 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7953 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7956 mc->mc_db->md_depth--;
7957 mc->mc_db->md_branch_pages--;
7958 mc->mc_ki[0] = mc->mc_ki[1];
7959 for (i = 1; i<mc->mc_db->md_depth; i++) {
7960 mc->mc_pg[i] = mc->mc_pg[i+1];
7961 mc->mc_ki[i] = mc->mc_ki[i+1];
7964 /* Adjust other cursors pointing to mp */
7965 MDB_cursor *m2, *m3;
7966 MDB_dbi dbi = mc->mc_dbi;
7968 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7969 if (mc->mc_flags & C_SUB)
7970 m3 = &m2->mc_xcursor->mx_cursor;
7973 if (m3 == mc) continue;
7974 if (m3->mc_pg[0] == mp) {
7975 for (i=0; i<mc->mc_db->md_depth; i++) {
7976 m3->mc_pg[i] = m3->mc_pg[i+1];
7977 m3->mc_ki[i] = m3->mc_ki[i+1];
7985 DPUTS("root page doesn't need rebalancing");
7989 /* The parent (branch page) must have at least 2 pointers,
7990 * otherwise the tree is invalid.
7992 ptop = mc->mc_top-1;
7993 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7995 /* Leaf page fill factor is below the threshold.
7996 * Try to move keys from left or right neighbor, or
7997 * merge with a neighbor page.
8002 mdb_cursor_copy(mc, &mn);
8003 mn.mc_xcursor = NULL;
8005 oldki = mc->mc_ki[mc->mc_top];
8006 if (mc->mc_ki[ptop] == 0) {
8007 /* We're the leftmost leaf in our parent.
8009 DPUTS("reading right neighbor");
8011 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8012 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8015 mn.mc_ki[mn.mc_top] = 0;
8016 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8018 /* There is at least one neighbor to the left.
8020 DPUTS("reading left neighbor");
8022 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8023 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8026 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8027 mc->mc_ki[mc->mc_top] = 0;
8030 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8031 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8032 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8034 /* If the neighbor page is above threshold and has enough keys,
8035 * move one key from it. Otherwise we should try to merge them.
8036 * (A branch page must never have less than 2 keys.)
8038 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8039 rc = mdb_node_move(&mn, mc);
8040 if (!mc->mc_ki[mc->mc_top]) {
8041 /* if we inserted on left, bump position up */
8045 if (mc->mc_ki[ptop] == 0) {
8046 rc = mdb_page_merge(&mn, mc);
8049 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8050 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8051 /* We want mdb_rebalance to find mn when doing fixups */
8052 if (mc->mc_flags & C_SUB) {
8053 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8054 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8055 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8057 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8058 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8060 rc = mdb_page_merge(mc, &mn);
8061 if (mc->mc_flags & C_SUB)
8062 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8064 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8065 mdb_cursor_copy(&mn, mc);
8067 mc->mc_flags &= ~C_EOF;
8069 mc->mc_ki[mc->mc_top] = oldki;
8073 /** Complete a delete operation started by #mdb_cursor_del(). */
8075 mdb_cursor_del0(MDB_cursor *mc)
8081 MDB_cursor *m2, *m3;
8082 MDB_dbi dbi = mc->mc_dbi;
8084 ki = mc->mc_ki[mc->mc_top];
8085 mp = mc->mc_pg[mc->mc_top];
8086 mdb_node_del(mc, mc->mc_db->md_pad);
8087 mc->mc_db->md_entries--;
8089 /* Adjust other cursors pointing to mp */
8090 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8091 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8092 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8094 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8096 if (m3->mc_pg[mc->mc_top] == mp) {
8097 if (m3->mc_ki[mc->mc_top] >= ki) {
8098 m3->mc_flags |= C_DEL;
8099 if (m3->mc_ki[mc->mc_top] > ki)
8100 m3->mc_ki[mc->mc_top]--;
8101 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8102 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8107 rc = mdb_rebalance(mc);
8109 if (rc == MDB_SUCCESS) {
8110 /* DB is totally empty now, just bail out.
8111 * Other cursors adjustments were already done
8112 * by mdb_rebalance and aren't needed here.
8117 mp = mc->mc_pg[mc->mc_top];
8118 nkeys = NUMKEYS(mp);
8120 /* Adjust other cursors pointing to mp */
8121 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8122 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8123 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8125 if (m3->mc_snum < mc->mc_snum)
8127 if (m3->mc_pg[mc->mc_top] == mp) {
8128 /* if m3 points past last node in page, find next sibling */
8129 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8130 rc = mdb_cursor_sibling(m3, 1);
8131 if (rc == MDB_NOTFOUND) {
8132 m3->mc_flags |= C_EOF;
8138 mc->mc_flags |= C_DEL;
8142 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8147 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8148 MDB_val *key, MDB_val *data)
8150 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8153 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8154 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8156 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8157 /* must ignore any data */
8161 return mdb_del0(txn, dbi, key, data, 0);
8165 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8166 MDB_val *key, MDB_val *data, unsigned flags)
8171 MDB_val rdata, *xdata;
8175 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8177 mdb_cursor_init(&mc, txn, dbi, &mx);
8186 flags |= MDB_NODUPDATA;
8188 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8190 /* let mdb_page_split know about this cursor if needed:
8191 * delete will trigger a rebalance; if it needs to move
8192 * a node from one page to another, it will have to
8193 * update the parent's separator key(s). If the new sepkey
8194 * is larger than the current one, the parent page may
8195 * run out of space, triggering a split. We need this
8196 * cursor to be consistent until the end of the rebalance.
8198 mc.mc_flags |= C_UNTRACK;
8199 mc.mc_next = txn->mt_cursors[dbi];
8200 txn->mt_cursors[dbi] = &mc;
8201 rc = mdb_cursor_del(&mc, flags);
8202 txn->mt_cursors[dbi] = mc.mc_next;
8207 /** Split a page and insert a new node.
8208 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8209 * The cursor will be updated to point to the actual page and index where
8210 * the node got inserted after the split.
8211 * @param[in] newkey The key for the newly inserted node.
8212 * @param[in] newdata The data for the newly inserted node.
8213 * @param[in] newpgno The page number, if the new node is a branch node.
8214 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8215 * @return 0 on success, non-zero on failure.
8218 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8219 unsigned int nflags)
8222 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8225 int i, j, split_indx, nkeys, pmax;
8226 MDB_env *env = mc->mc_txn->mt_env;
8228 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8229 MDB_page *copy = NULL;
8230 MDB_page *mp, *rp, *pp;
8235 mp = mc->mc_pg[mc->mc_top];
8236 newindx = mc->mc_ki[mc->mc_top];
8237 nkeys = NUMKEYS(mp);
8239 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8240 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8241 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8243 /* Create a right sibling. */
8244 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8246 rp->mp_pad = mp->mp_pad;
8247 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8249 if (mc->mc_snum < 2) {
8250 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8252 /* shift current top to make room for new parent */
8253 mc->mc_pg[1] = mc->mc_pg[0];
8254 mc->mc_ki[1] = mc->mc_ki[0];
8257 mc->mc_db->md_root = pp->mp_pgno;
8258 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8259 new_root = mc->mc_db->md_depth++;
8261 /* Add left (implicit) pointer. */
8262 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8263 /* undo the pre-push */
8264 mc->mc_pg[0] = mc->mc_pg[1];
8265 mc->mc_ki[0] = mc->mc_ki[1];
8266 mc->mc_db->md_root = mp->mp_pgno;
8267 mc->mc_db->md_depth--;
8274 ptop = mc->mc_top-1;
8275 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8278 mc->mc_flags |= C_SPLITTING;
8279 mdb_cursor_copy(mc, &mn);
8280 mn.mc_pg[mn.mc_top] = rp;
8281 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8283 if (nflags & MDB_APPEND) {
8284 mn.mc_ki[mn.mc_top] = 0;
8286 split_indx = newindx;
8290 split_indx = (nkeys+1) / 2;
8295 unsigned int lsize, rsize, ksize;
8296 /* Move half of the keys to the right sibling */
8297 x = mc->mc_ki[mc->mc_top] - split_indx;
8298 ksize = mc->mc_db->md_pad;
8299 split = LEAF2KEY(mp, split_indx, ksize);
8300 rsize = (nkeys - split_indx) * ksize;
8301 lsize = (nkeys - split_indx) * sizeof(indx_t);
8302 mp->mp_lower -= lsize;
8303 rp->mp_lower += lsize;
8304 mp->mp_upper += rsize - lsize;
8305 rp->mp_upper -= rsize - lsize;
8306 sepkey.mv_size = ksize;
8307 if (newindx == split_indx) {
8308 sepkey.mv_data = newkey->mv_data;
8310 sepkey.mv_data = split;
8313 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8314 memcpy(rp->mp_ptrs, split, rsize);
8315 sepkey.mv_data = rp->mp_ptrs;
8316 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8317 memcpy(ins, newkey->mv_data, ksize);
8318 mp->mp_lower += sizeof(indx_t);
8319 mp->mp_upper -= ksize - sizeof(indx_t);
8322 memcpy(rp->mp_ptrs, split, x * ksize);
8323 ins = LEAF2KEY(rp, x, ksize);
8324 memcpy(ins, newkey->mv_data, ksize);
8325 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8326 rp->mp_lower += sizeof(indx_t);
8327 rp->mp_upper -= ksize - sizeof(indx_t);
8328 mc->mc_ki[mc->mc_top] = x;
8329 mc->mc_pg[mc->mc_top] = rp;
8333 int psize, nsize, k;
8334 /* Maximum free space in an empty page */
8335 pmax = env->me_psize - PAGEHDRSZ;
8337 nsize = mdb_leaf_size(env, newkey, newdata);
8339 nsize = mdb_branch_size(env, newkey);
8340 nsize = EVEN(nsize);
8342 /* grab a page to hold a temporary copy */
8343 copy = mdb_page_malloc(mc->mc_txn, 1);
8348 copy->mp_pgno = mp->mp_pgno;
8349 copy->mp_flags = mp->mp_flags;
8350 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8351 copy->mp_upper = env->me_psize - PAGEBASE;
8353 /* prepare to insert */
8354 for (i=0, j=0; i<nkeys; i++) {
8356 copy->mp_ptrs[j++] = 0;
8358 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8361 /* When items are relatively large the split point needs
8362 * to be checked, because being off-by-one will make the
8363 * difference between success or failure in mdb_node_add.
8365 * It's also relevant if a page happens to be laid out
8366 * such that one half of its nodes are all "small" and
8367 * the other half of its nodes are "large." If the new
8368 * item is also "large" and falls on the half with
8369 * "large" nodes, it also may not fit.
8371 * As a final tweak, if the new item goes on the last
8372 * spot on the page (and thus, onto the new page), bias
8373 * the split so the new page is emptier than the old page.
8374 * This yields better packing during sequential inserts.
8376 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8377 /* Find split point */
8379 if (newindx <= split_indx || newindx >= nkeys) {
8381 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8386 for (; i!=k; i+=j) {
8391 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8392 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8394 if (F_ISSET(node->mn_flags, F_BIGDATA))
8395 psize += sizeof(pgno_t);
8397 psize += NODEDSZ(node);
8399 psize = EVEN(psize);
8401 if (psize > pmax || i == k-j) {
8402 split_indx = i + (j<0);
8407 if (split_indx == newindx) {
8408 sepkey.mv_size = newkey->mv_size;
8409 sepkey.mv_data = newkey->mv_data;
8411 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8412 sepkey.mv_size = node->mn_ksize;
8413 sepkey.mv_data = NODEKEY(node);
8418 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8420 /* Copy separator key to the parent.
8422 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8426 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8431 if (mn.mc_snum == mc->mc_snum) {
8432 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8433 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8434 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8435 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8440 /* Right page might now have changed parent.
8441 * Check if left page also changed parent.
8443 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8444 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8445 for (i=0; i<ptop; i++) {
8446 mc->mc_pg[i] = mn.mc_pg[i];
8447 mc->mc_ki[i] = mn.mc_ki[i];
8449 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8450 if (mn.mc_ki[ptop]) {
8451 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8453 /* find right page's left sibling */
8454 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8455 mdb_cursor_sibling(mc, 0);
8460 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8463 mc->mc_flags ^= C_SPLITTING;
8464 if (rc != MDB_SUCCESS) {
8467 if (nflags & MDB_APPEND) {
8468 mc->mc_pg[mc->mc_top] = rp;
8469 mc->mc_ki[mc->mc_top] = 0;
8470 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8473 for (i=0; i<mc->mc_top; i++)
8474 mc->mc_ki[i] = mn.mc_ki[i];
8475 } else if (!IS_LEAF2(mp)) {
8477 mc->mc_pg[mc->mc_top] = rp;
8482 rkey.mv_data = newkey->mv_data;
8483 rkey.mv_size = newkey->mv_size;
8489 /* Update index for the new key. */
8490 mc->mc_ki[mc->mc_top] = j;
8492 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8493 rkey.mv_data = NODEKEY(node);
8494 rkey.mv_size = node->mn_ksize;
8496 xdata.mv_data = NODEDATA(node);
8497 xdata.mv_size = NODEDSZ(node);
8500 pgno = NODEPGNO(node);
8501 flags = node->mn_flags;
8504 if (!IS_LEAF(mp) && j == 0) {
8505 /* First branch index doesn't need key data. */
8509 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8515 mc->mc_pg[mc->mc_top] = copy;
8520 } while (i != split_indx);
8522 nkeys = NUMKEYS(copy);
8523 for (i=0; i<nkeys; i++)
8524 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8525 mp->mp_lower = copy->mp_lower;
8526 mp->mp_upper = copy->mp_upper;
8527 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8528 env->me_psize - copy->mp_upper - PAGEBASE);
8530 /* reset back to original page */
8531 if (newindx < split_indx) {
8532 mc->mc_pg[mc->mc_top] = mp;
8533 if (nflags & MDB_RESERVE) {
8534 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8535 if (!(node->mn_flags & F_BIGDATA))
8536 newdata->mv_data = NODEDATA(node);
8539 mc->mc_pg[mc->mc_top] = rp;
8541 /* Make sure mc_ki is still valid.
8543 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8544 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8545 for (i=0; i<=ptop; i++) {
8546 mc->mc_pg[i] = mn.mc_pg[i];
8547 mc->mc_ki[i] = mn.mc_ki[i];
8554 /* Adjust other cursors pointing to mp */
8555 MDB_cursor *m2, *m3;
8556 MDB_dbi dbi = mc->mc_dbi;
8557 int fixup = NUMKEYS(mp);
8559 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8560 if (mc->mc_flags & C_SUB)
8561 m3 = &m2->mc_xcursor->mx_cursor;
8566 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8568 if (m3->mc_flags & C_SPLITTING)
8573 for (k=new_root; k>=0; k--) {
8574 m3->mc_ki[k+1] = m3->mc_ki[k];
8575 m3->mc_pg[k+1] = m3->mc_pg[k];
8577 if (m3->mc_ki[0] >= split_indx) {
8582 m3->mc_pg[0] = mc->mc_pg[0];
8586 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8587 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8588 m3->mc_ki[mc->mc_top]++;
8589 if (m3->mc_ki[mc->mc_top] >= fixup) {
8590 m3->mc_pg[mc->mc_top] = rp;
8591 m3->mc_ki[mc->mc_top] -= fixup;
8592 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8594 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8595 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8600 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8603 if (copy) /* tmp page */
8604 mdb_page_free(env, copy);
8606 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8611 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8612 MDB_val *key, MDB_val *data, unsigned int flags)
8617 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8620 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8623 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8624 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8626 mdb_cursor_init(&mc, txn, dbi, &mx);
8627 return mdb_cursor_put(&mc, key, data, flags);
8631 #define MDB_WBUF (1024*1024)
8634 /** State needed for a compacting copy. */
8635 typedef struct mdb_copy {
8636 pthread_mutex_t mc_mutex;
8637 pthread_cond_t mc_cond;
8644 pgno_t mc_next_pgno;
8647 volatile int mc_new;
8652 /** Dedicated writer thread for compacting copy. */
8653 static THREAD_RET ESECT CALL_CONV
8654 mdb_env_copythr(void *arg)
8658 int toggle = 0, wsize, rc;
8661 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8664 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8667 pthread_mutex_lock(&my->mc_mutex);
8669 pthread_cond_signal(&my->mc_cond);
8672 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8673 if (my->mc_new < 0) {
8678 wsize = my->mc_wlen[toggle];
8679 ptr = my->mc_wbuf[toggle];
8682 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8686 } else if (len > 0) {
8700 /* If there's an overflow page tail, write it too */
8701 if (my->mc_olen[toggle]) {
8702 wsize = my->mc_olen[toggle];
8703 ptr = my->mc_over[toggle];
8704 my->mc_olen[toggle] = 0;
8707 my->mc_wlen[toggle] = 0;
8709 pthread_cond_signal(&my->mc_cond);
8711 pthread_cond_signal(&my->mc_cond);
8712 pthread_mutex_unlock(&my->mc_mutex);
8713 return (THREAD_RET)0;
8717 /** Tell the writer thread there's a buffer ready to write */
8719 mdb_env_cthr_toggle(mdb_copy *my, int st)
8721 int toggle = my->mc_toggle ^ 1;
8722 pthread_mutex_lock(&my->mc_mutex);
8723 if (my->mc_status) {
8724 pthread_mutex_unlock(&my->mc_mutex);
8725 return my->mc_status;
8727 while (my->mc_new == 1)
8728 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8730 my->mc_toggle = toggle;
8731 pthread_cond_signal(&my->mc_cond);
8732 pthread_mutex_unlock(&my->mc_mutex);
8736 /** Depth-first tree traversal for compacting copy. */
8738 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8741 MDB_txn *txn = my->mc_txn;
8743 MDB_page *mo, *mp, *leaf;
8748 /* Empty DB, nothing to do */
8749 if (*pg == P_INVALID)
8756 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8759 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8763 /* Make cursor pages writable */
8764 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8768 for (i=0; i<mc.mc_top; i++) {
8769 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8770 mc.mc_pg[i] = (MDB_page *)ptr;
8771 ptr += my->mc_env->me_psize;
8774 /* This is writable space for a leaf page. Usually not needed. */
8775 leaf = (MDB_page *)ptr;
8777 toggle = my->mc_toggle;
8778 while (mc.mc_snum > 0) {
8780 mp = mc.mc_pg[mc.mc_top];
8784 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8785 for (i=0; i<n; i++) {
8786 ni = NODEPTR(mp, i);
8787 if (ni->mn_flags & F_BIGDATA) {
8791 /* Need writable leaf */
8793 mc.mc_pg[mc.mc_top] = leaf;
8794 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8796 ni = NODEPTR(mp, i);
8799 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8800 rc = mdb_page_get(txn, pg, &omp, NULL);
8803 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8804 rc = mdb_env_cthr_toggle(my, 1);
8807 toggle = my->mc_toggle;
8809 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8810 memcpy(mo, omp, my->mc_env->me_psize);
8811 mo->mp_pgno = my->mc_next_pgno;
8812 my->mc_next_pgno += omp->mp_pages;
8813 my->mc_wlen[toggle] += my->mc_env->me_psize;
8814 if (omp->mp_pages > 1) {
8815 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8816 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8817 rc = mdb_env_cthr_toggle(my, 1);
8820 toggle = my->mc_toggle;
8822 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8823 } else if (ni->mn_flags & F_SUBDATA) {
8826 /* Need writable leaf */
8828 mc.mc_pg[mc.mc_top] = leaf;
8829 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8831 ni = NODEPTR(mp, i);
8834 memcpy(&db, NODEDATA(ni), sizeof(db));
8835 my->mc_toggle = toggle;
8836 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8839 toggle = my->mc_toggle;
8840 memcpy(NODEDATA(ni), &db, sizeof(db));
8845 mc.mc_ki[mc.mc_top]++;
8846 if (mc.mc_ki[mc.mc_top] < n) {
8849 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8851 rc = mdb_page_get(txn, pg, &mp, NULL);
8856 mc.mc_ki[mc.mc_top] = 0;
8857 if (IS_BRANCH(mp)) {
8858 /* Whenever we advance to a sibling branch page,
8859 * we must proceed all the way down to its first leaf.
8861 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8864 mc.mc_pg[mc.mc_top] = mp;
8868 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8869 rc = mdb_env_cthr_toggle(my, 1);
8872 toggle = my->mc_toggle;
8874 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8875 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8876 mo->mp_pgno = my->mc_next_pgno++;
8877 my->mc_wlen[toggle] += my->mc_env->me_psize;
8879 /* Update parent if there is one */
8880 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8881 SETPGNO(ni, mo->mp_pgno);
8882 mdb_cursor_pop(&mc);
8884 /* Otherwise we're done */
8894 /** Copy environment with compaction. */
8896 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8901 MDB_txn *txn = NULL;
8906 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8907 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8908 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8909 if (my.mc_wbuf[0] == NULL)
8912 pthread_mutex_init(&my.mc_mutex, NULL);
8913 pthread_cond_init(&my.mc_cond, NULL);
8914 #ifdef HAVE_MEMALIGN
8915 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8916 if (my.mc_wbuf[0] == NULL)
8919 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8924 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8925 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8930 my.mc_next_pgno = NUM_METAS;
8936 THREAD_CREATE(thr, mdb_env_copythr, &my);
8938 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8942 mp = (MDB_page *)my.mc_wbuf[0];
8943 memset(mp, 0, NUM_METAS * env->me_psize);
8945 mp->mp_flags = P_META;
8946 mm = (MDB_meta *)METADATA(mp);
8947 mdb_env_init_meta0(env, mm);
8948 mm->mm_address = env->me_metas[0]->mm_address;
8950 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8952 mp->mp_flags = P_META;
8953 *(MDB_meta *)METADATA(mp) = *mm;
8954 mm = (MDB_meta *)METADATA(mp);
8956 /* Count the number of free pages, subtract from lastpg to find
8957 * number of active pages
8960 MDB_ID freecount = 0;
8963 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8964 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8965 freecount += *(MDB_ID *)data.mv_data;
8966 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
8967 txn->mt_dbs[FREE_DBI].md_leaf_pages +
8968 txn->mt_dbs[FREE_DBI].md_overflow_pages;
8970 /* Set metapage 1 */
8971 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8972 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
8973 if (mm->mm_last_pg > NUM_METAS-1) {
8974 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
8977 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
8980 my.mc_wlen[0] = env->me_psize * NUM_METAS;
8982 pthread_mutex_lock(&my.mc_mutex);
8984 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8985 pthread_mutex_unlock(&my.mc_mutex);
8986 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
8987 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8988 rc = mdb_env_cthr_toggle(&my, 1);
8989 mdb_env_cthr_toggle(&my, -1);
8990 pthread_mutex_lock(&my.mc_mutex);
8992 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8993 pthread_mutex_unlock(&my.mc_mutex);
8998 CloseHandle(my.mc_cond);
8999 CloseHandle(my.mc_mutex);
9000 _aligned_free(my.mc_wbuf[0]);
9002 pthread_cond_destroy(&my.mc_cond);
9003 pthread_mutex_destroy(&my.mc_mutex);
9004 free(my.mc_wbuf[0]);
9009 /** Copy environment as-is. */
9011 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9013 MDB_txn *txn = NULL;
9014 mdb_mutexref_t wmutex = NULL;
9020 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9024 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9027 /* Do the lock/unlock of the reader mutex before starting the
9028 * write txn. Otherwise other read txns could block writers.
9030 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9035 /* We must start the actual read txn after blocking writers */
9036 mdb_txn_end(txn, MDB_END_RESET_TMP);
9038 /* Temporarily block writers until we snapshot the meta pages */
9039 wmutex = env->me_wmutex;
9040 if (LOCK_MUTEX(rc, env, wmutex))
9043 rc = mdb_txn_renew0(txn);
9045 UNLOCK_MUTEX(wmutex);
9050 wsize = env->me_psize * NUM_METAS;
9054 DO_WRITE(rc, fd, ptr, w2, len);
9058 } else if (len > 0) {
9064 /* Non-blocking or async handles are not supported */
9070 UNLOCK_MUTEX(wmutex);
9075 w2 = txn->mt_next_pgno * env->me_psize;
9078 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9085 if (wsize > MAX_WRITE)
9089 DO_WRITE(rc, fd, ptr, w2, len);
9093 } else if (len > 0) {
9110 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9112 if (flags & MDB_CP_COMPACT)
9113 return mdb_env_copyfd1(env, fd);
9115 return mdb_env_copyfd0(env, fd);
9119 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9121 return mdb_env_copyfd2(env, fd, 0);
9125 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9129 HANDLE newfd = INVALID_HANDLE_VALUE;
9131 if (env->me_flags & MDB_NOSUBDIR) {
9132 lpath = (char *)path;
9135 len += sizeof(DATANAME);
9136 lpath = malloc(len);
9139 sprintf(lpath, "%s" DATANAME, path);
9142 /* The destination path must exist, but the destination file must not.
9143 * We don't want the OS to cache the writes, since the source data is
9144 * already in the OS cache.
9147 newfd = CreateFileA(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9148 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9150 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9152 if (newfd == INVALID_HANDLE_VALUE) {
9157 if (env->me_psize >= env->me_os_psize) {
9159 /* Set O_DIRECT if the file system supports it */
9160 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9161 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9163 #ifdef F_NOCACHE /* __APPLE__ */
9164 rc = fcntl(newfd, F_NOCACHE, 1);
9172 rc = mdb_env_copyfd2(env, newfd, flags);
9175 if (!(env->me_flags & MDB_NOSUBDIR))
9177 if (newfd != INVALID_HANDLE_VALUE)
9178 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9185 mdb_env_copy(MDB_env *env, const char *path)
9187 return mdb_env_copy2(env, path, 0);
9191 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9193 if (flag & ~CHANGEABLE)
9196 env->me_flags |= flag;
9198 env->me_flags &= ~flag;
9203 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9208 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9213 mdb_env_set_userctx(MDB_env *env, void *ctx)
9217 env->me_userctx = ctx;
9222 mdb_env_get_userctx(MDB_env *env)
9224 return env ? env->me_userctx : NULL;
9228 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9233 env->me_assert_func = func;
9239 mdb_env_get_path(MDB_env *env, const char **arg)
9244 *arg = env->me_path;
9249 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9258 /** Common code for #mdb_stat() and #mdb_env_stat().
9259 * @param[in] env the environment to operate in.
9260 * @param[in] db the #MDB_db record containing the stats to return.
9261 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9262 * @return 0, this function always succeeds.
9265 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9267 arg->ms_psize = env->me_psize;
9268 arg->ms_depth = db->md_depth;
9269 arg->ms_branch_pages = db->md_branch_pages;
9270 arg->ms_leaf_pages = db->md_leaf_pages;
9271 arg->ms_overflow_pages = db->md_overflow_pages;
9272 arg->ms_entries = db->md_entries;
9278 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9282 if (env == NULL || arg == NULL)
9285 meta = mdb_env_pick_meta(env);
9287 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9291 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9295 if (env == NULL || arg == NULL)
9298 meta = mdb_env_pick_meta(env);
9299 arg->me_mapaddr = meta->mm_address;
9300 arg->me_last_pgno = meta->mm_last_pg;
9301 arg->me_last_txnid = meta->mm_txnid;
9303 arg->me_mapsize = env->me_mapsize;
9304 arg->me_maxreaders = env->me_maxreaders;
9305 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9309 /** Set the default comparison functions for a database.
9310 * Called immediately after a database is opened to set the defaults.
9311 * The user can then override them with #mdb_set_compare() or
9312 * #mdb_set_dupsort().
9313 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9314 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9317 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9319 uint16_t f = txn->mt_dbs[dbi].md_flags;
9321 txn->mt_dbxs[dbi].md_cmp =
9322 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9323 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9325 txn->mt_dbxs[dbi].md_dcmp =
9326 !(f & MDB_DUPSORT) ? 0 :
9327 ((f & MDB_INTEGERDUP)
9328 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9329 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9332 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9338 int rc, dbflag, exact;
9339 unsigned int unused = 0, seq;
9342 if (flags & ~VALID_FLAGS)
9344 if (txn->mt_flags & MDB_TXN_BLOCKED)
9350 if (flags & PERSISTENT_FLAGS) {
9351 uint16_t f2 = flags & PERSISTENT_FLAGS;
9352 /* make sure flag changes get committed */
9353 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9354 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9355 txn->mt_flags |= MDB_TXN_DIRTY;
9358 mdb_default_cmp(txn, MAIN_DBI);
9362 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9363 mdb_default_cmp(txn, MAIN_DBI);
9366 /* Is the DB already open? */
9368 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9369 if (!txn->mt_dbxs[i].md_name.mv_size) {
9370 /* Remember this free slot */
9371 if (!unused) unused = i;
9374 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9375 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9381 /* If no free slot and max hit, fail */
9382 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9383 return MDB_DBS_FULL;
9385 /* Cannot mix named databases with some mainDB flags */
9386 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9387 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9389 /* Find the DB info */
9390 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9393 key.mv_data = (void *)name;
9394 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9395 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9396 if (rc == MDB_SUCCESS) {
9397 /* make sure this is actually a DB */
9398 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9399 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9400 return MDB_INCOMPATIBLE;
9401 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9402 /* Create if requested */
9403 data.mv_size = sizeof(MDB_db);
9404 data.mv_data = &dummy;
9405 memset(&dummy, 0, sizeof(dummy));
9406 dummy.md_root = P_INVALID;
9407 dummy.md_flags = flags & PERSISTENT_FLAGS;
9408 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9412 /* OK, got info, add to table */
9413 if (rc == MDB_SUCCESS) {
9414 unsigned int slot = unused ? unused : txn->mt_numdbs;
9415 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9416 txn->mt_dbxs[slot].md_name.mv_size = len;
9417 txn->mt_dbxs[slot].md_rel = NULL;
9418 txn->mt_dbflags[slot] = dbflag;
9419 /* txn-> and env-> are the same in read txns, use
9420 * tmp variable to avoid undefined assignment
9422 seq = ++txn->mt_env->me_dbiseqs[slot];
9423 txn->mt_dbiseqs[slot] = seq;
9425 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9427 mdb_default_cmp(txn, slot);
9437 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9439 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9442 if (txn->mt_flags & MDB_TXN_BLOCKED)
9445 if (txn->mt_dbflags[dbi] & DB_STALE) {
9448 /* Stale, must read the DB's root. cursor_init does it for us. */
9449 mdb_cursor_init(&mc, txn, dbi, &mx);
9451 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9454 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9457 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9459 ptr = env->me_dbxs[dbi].md_name.mv_data;
9460 /* If there was no name, this was already closed */
9462 env->me_dbxs[dbi].md_name.mv_data = NULL;
9463 env->me_dbxs[dbi].md_name.mv_size = 0;
9464 env->me_dbflags[dbi] = 0;
9465 env->me_dbiseqs[dbi]++;
9470 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9472 /* We could return the flags for the FREE_DBI too but what's the point? */
9473 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9475 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9479 /** Add all the DB's pages to the free list.
9480 * @param[in] mc Cursor on the DB to free.
9481 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9482 * @return 0 on success, non-zero on failure.
9485 mdb_drop0(MDB_cursor *mc, int subs)
9489 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9490 if (rc == MDB_SUCCESS) {
9491 MDB_txn *txn = mc->mc_txn;
9496 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9497 * This also avoids any P_LEAF2 pages, which have no nodes.
9499 if (mc->mc_flags & C_SUB)
9502 mdb_cursor_copy(mc, &mx);
9503 while (mc->mc_snum > 0) {
9504 MDB_page *mp = mc->mc_pg[mc->mc_top];
9505 unsigned n = NUMKEYS(mp);
9507 for (i=0; i<n; i++) {
9508 ni = NODEPTR(mp, i);
9509 if (ni->mn_flags & F_BIGDATA) {
9512 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9513 rc = mdb_page_get(txn, pg, &omp, NULL);
9516 mdb_cassert(mc, IS_OVERFLOW(omp));
9517 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9521 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9522 mdb_xcursor_init1(mc, ni);
9523 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9529 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9531 for (i=0; i<n; i++) {
9533 ni = NODEPTR(mp, i);
9536 mdb_midl_xappend(txn->mt_free_pgs, pg);
9541 mc->mc_ki[mc->mc_top] = i;
9542 rc = mdb_cursor_sibling(mc, 1);
9544 if (rc != MDB_NOTFOUND)
9546 /* no more siblings, go back to beginning
9547 * of previous level.
9551 for (i=1; i<mc->mc_snum; i++) {
9553 mc->mc_pg[i] = mx.mc_pg[i];
9558 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9561 txn->mt_flags |= MDB_TXN_ERROR;
9562 } else if (rc == MDB_NOTFOUND) {
9568 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9570 MDB_cursor *mc, *m2;
9573 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9576 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9579 if (TXN_DBI_CHANGED(txn, dbi))
9582 rc = mdb_cursor_open(txn, dbi, &mc);
9586 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9587 /* Invalidate the dropped DB's cursors */
9588 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9589 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9593 /* Can't delete the main DB */
9594 if (del && dbi >= CORE_DBS) {
9595 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9597 txn->mt_dbflags[dbi] = DB_STALE;
9598 mdb_dbi_close(txn->mt_env, dbi);
9600 txn->mt_flags |= MDB_TXN_ERROR;
9603 /* reset the DB record, mark it dirty */
9604 txn->mt_dbflags[dbi] |= DB_DIRTY;
9605 txn->mt_dbs[dbi].md_depth = 0;
9606 txn->mt_dbs[dbi].md_branch_pages = 0;
9607 txn->mt_dbs[dbi].md_leaf_pages = 0;
9608 txn->mt_dbs[dbi].md_overflow_pages = 0;
9609 txn->mt_dbs[dbi].md_entries = 0;
9610 txn->mt_dbs[dbi].md_root = P_INVALID;
9612 txn->mt_flags |= MDB_TXN_DIRTY;
9615 mdb_cursor_close(mc);
9619 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9621 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9624 txn->mt_dbxs[dbi].md_cmp = cmp;
9628 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9630 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9633 txn->mt_dbxs[dbi].md_dcmp = cmp;
9637 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9639 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9642 txn->mt_dbxs[dbi].md_rel = rel;
9646 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9648 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9651 txn->mt_dbxs[dbi].md_relctx = ctx;
9656 mdb_env_get_maxkeysize(MDB_env *env)
9658 return ENV_MAXKEY(env);
9662 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9664 unsigned int i, rdrs;
9667 int rc = 0, first = 1;
9671 if (!env->me_txns) {
9672 return func("(no reader locks)\n", ctx);
9674 rdrs = env->me_txns->mti_numreaders;
9675 mr = env->me_txns->mti_readers;
9676 for (i=0; i<rdrs; i++) {
9678 txnid_t txnid = mr[i].mr_txnid;
9679 sprintf(buf, txnid == (txnid_t)-1 ?
9680 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9681 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9684 rc = func(" pid thread txnid\n", ctx);
9688 rc = func(buf, ctx);
9694 rc = func("(no active readers)\n", ctx);
9699 /** Insert pid into list if not already present.
9700 * return -1 if already present.
9703 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9705 /* binary search of pid in list */
9707 unsigned cursor = 1;
9709 unsigned n = ids[0];
9712 unsigned pivot = n >> 1;
9713 cursor = base + pivot + 1;
9714 val = pid - ids[cursor];
9719 } else if ( val > 0 ) {
9724 /* found, so it's a duplicate */
9733 for (n = ids[0]; n > cursor; n--)
9740 mdb_reader_check(MDB_env *env, int *dead)
9746 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9749 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9751 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9753 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9754 unsigned int i, j, rdrs;
9756 MDB_PID_T *pids, pid;
9757 int rc = MDB_SUCCESS, count = 0;
9759 rdrs = env->me_txns->mti_numreaders;
9760 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9764 mr = env->me_txns->mti_readers;
9765 for (i=0; i<rdrs; i++) {
9767 if (pid && pid != env->me_pid) {
9768 if (mdb_pid_insert(pids, pid) == 0) {
9769 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9770 /* Stale reader found */
9773 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9774 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9776 rdrs = 0; /* the above checked all readers */
9778 /* Recheck, a new process may have reused pid */
9779 if (mdb_reader_pid(env, Pidcheck, pid))
9784 if (mr[j].mr_pid == pid) {
9785 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9786 (unsigned) pid, mr[j].mr_txnid));
9791 UNLOCK_MUTEX(rmutex);
9802 #ifdef MDB_ROBUST_SUPPORTED
9803 /** Handle #LOCK_MUTEX0() failure.
9804 * Try to repair the lock file if the mutex owner died.
9805 * @param[in] env the environment handle
9806 * @param[in] mutex LOCK_MUTEX0() mutex
9807 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9808 * @return 0 on success with the mutex locked, or an error code on failure.
9811 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9816 if (rc == MDB_OWNERDEAD) {
9817 /* We own the mutex. Clean up after dead previous owner. */
9819 rlocked = (mutex == env->me_rmutex);
9821 /* Keep mti_txnid updated, otherwise next writer can
9822 * overwrite data which latest meta page refers to.
9824 meta = mdb_env_pick_meta(env);
9825 env->me_txns->mti_txnid = meta->mm_txnid;
9826 /* env is hosed if the dead thread was ours */
9828 env->me_flags |= MDB_FATAL_ERROR;
9833 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9834 (rc ? "this process' env is hosed" : "recovering")));
9835 rc2 = mdb_reader_check0(env, rlocked, NULL);
9837 rc2 = mdb_mutex_consistent(mutex);
9838 if (rc || (rc = rc2)) {
9839 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9840 UNLOCK_MUTEX(mutex);
9846 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9851 #endif /* MDB_ROBUST_SUPPORTED */