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.
409 # define MDB_DSYNC O_DSYNC
411 # define MDB_DSYNC O_SYNC
416 /** Function for flushing the data of a file. Define this to fsync
417 * if fdatasync() is not supported.
419 #ifndef MDB_FDATASYNC
420 # define MDB_FDATASYNC fdatasync
424 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
435 /** A page number in the database.
436 * Note that 64 bit page numbers are overkill, since pages themselves
437 * already represent 12-13 bits of addressable memory, and the OS will
438 * always limit applications to a maximum of 63 bits of address space.
440 * @note In the #MDB_node structure, we only store 48 bits of this value,
441 * which thus limits us to only 60 bits of addressable data.
443 typedef MDB_ID pgno_t;
445 /** A transaction ID.
446 * See struct MDB_txn.mt_txnid for details.
448 typedef MDB_ID txnid_t;
450 /** @defgroup debug Debug Macros
454 /** Enable debug output. Needs variable argument macros (a C99 feature).
455 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
456 * read from and written to the database (used for free space management).
462 static int mdb_debug;
463 static txnid_t mdb_debug_start;
465 /** Print a debug message with printf formatting.
466 * Requires double parenthesis around 2 or more args.
468 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
469 # define DPRINTF0(fmt, ...) \
470 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
472 # define DPRINTF(args) ((void) 0)
474 /** Print a debug string.
475 * The string is printed literally, with no format processing.
477 #define DPUTS(arg) DPRINTF(("%s", arg))
478 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
480 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
483 /** @brief The maximum size of a database page.
485 * It is 32k or 64k, since value-PAGEBASE must fit in
486 * #MDB_page.%mp_upper.
488 * LMDB will use database pages < OS pages if needed.
489 * That causes more I/O in write transactions: The OS must
490 * know (read) the whole page before writing a partial page.
492 * Note that we don't currently support Huge pages. On Linux,
493 * regular data files cannot use Huge pages, and in general
494 * Huge pages aren't actually pageable. We rely on the OS
495 * demand-pager to read our data and page it out when memory
496 * pressure from other processes is high. So until OSs have
497 * actual paging support for Huge pages, they're not viable.
499 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
501 /** The minimum number of keys required in a database page.
502 * Setting this to a larger value will place a smaller bound on the
503 * maximum size of a data item. Data items larger than this size will
504 * be pushed into overflow pages instead of being stored directly in
505 * the B-tree node. This value used to default to 4. With a page size
506 * of 4096 bytes that meant that any item larger than 1024 bytes would
507 * go into an overflow page. That also meant that on average 2-3KB of
508 * each overflow page was wasted space. The value cannot be lower than
509 * 2 because then there would no longer be a tree structure. With this
510 * value, items larger than 2KB will go into overflow pages, and on
511 * average only 1KB will be wasted.
513 #define MDB_MINKEYS 2
515 /** A stamp that identifies a file as an LMDB file.
516 * There's nothing special about this value other than that it is easily
517 * recognizable, and it will reflect any byte order mismatches.
519 #define MDB_MAGIC 0xBEEFC0DE
521 /** The version number for a database's datafile format. */
522 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
523 /** The version number for a database's lockfile format. */
524 #define MDB_LOCK_VERSION 1
526 /** @brief The max size of a key we can write, or 0 for computed max.
528 * This macro should normally be left alone or set to 0.
529 * Note that a database with big keys or dupsort data cannot be
530 * reliably modified by a liblmdb which uses a smaller max.
531 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
533 * Other values are allowed, for backwards compat. However:
534 * A value bigger than the computed max can break if you do not
535 * know what you are doing, and liblmdb <= 0.9.10 can break when
536 * modifying a DB with keys/dupsort data bigger than its max.
538 * Data items in an #MDB_DUPSORT database are also limited to
539 * this size, since they're actually keys of a sub-DB. Keys and
540 * #MDB_DUPSORT data items must fit on a node in a regular page.
542 #ifndef MDB_MAXKEYSIZE
543 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
546 /** The maximum size of a key we can write to the environment. */
548 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
550 #define ENV_MAXKEY(env) ((env)->me_maxkey)
553 /** @brief The maximum size of a data item.
555 * We only store a 32 bit value for node sizes.
557 #define MAXDATASIZE 0xffffffffUL
560 /** Key size which fits in a #DKBUF.
563 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
566 * This is used for printing a hex dump of a key's contents.
568 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
569 /** Display a key in hex.
571 * Invoke a function to display a key in hex.
573 #define DKEY(x) mdb_dkey(x, kbuf)
579 /** An invalid page number.
580 * Mainly used to denote an empty tree.
582 #define P_INVALID (~(pgno_t)0)
584 /** Test if the flags \b f are set in a flag word \b w. */
585 #define F_ISSET(w, f) (((w) & (f)) == (f))
587 /** Round \b n up to an even number. */
588 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
590 /** Used for offsets within a single page.
591 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
594 typedef uint16_t indx_t;
596 /** Default size of memory map.
597 * This is certainly too small for any actual applications. Apps should always set
598 * the size explicitly using #mdb_env_set_mapsize().
600 #define DEFAULT_MAPSIZE 1048576
602 /** @defgroup readers Reader Lock Table
603 * Readers don't acquire any locks for their data access. Instead, they
604 * simply record their transaction ID in the reader table. The reader
605 * mutex is needed just to find an empty slot in the reader table. The
606 * slot's address is saved in thread-specific data so that subsequent read
607 * transactions started by the same thread need no further locking to proceed.
609 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
611 * No reader table is used if the database is on a read-only filesystem, or
612 * if #MDB_NOLOCK is set.
614 * Since the database uses multi-version concurrency control, readers don't
615 * actually need any locking. This table is used to keep track of which
616 * readers are using data from which old transactions, so that we'll know
617 * when a particular old transaction is no longer in use. Old transactions
618 * that have discarded any data pages can then have those pages reclaimed
619 * for use by a later write transaction.
621 * The lock table is constructed such that reader slots are aligned with the
622 * processor's cache line size. Any slot is only ever used by one thread.
623 * This alignment guarantees that there will be no contention or cache
624 * thrashing as threads update their own slot info, and also eliminates
625 * any need for locking when accessing a slot.
627 * A writer thread will scan every slot in the table to determine the oldest
628 * outstanding reader transaction. Any freed pages older than this will be
629 * reclaimed by the writer. The writer doesn't use any locks when scanning
630 * this table. This means that there's no guarantee that the writer will
631 * see the most up-to-date reader info, but that's not required for correct
632 * operation - all we need is to know the upper bound on the oldest reader,
633 * we don't care at all about the newest reader. So the only consequence of
634 * reading stale information here is that old pages might hang around a
635 * while longer before being reclaimed. That's actually good anyway, because
636 * the longer we delay reclaiming old pages, the more likely it is that a
637 * string of contiguous pages can be found after coalescing old pages from
638 * many old transactions together.
641 /** Number of slots in the reader table.
642 * This value was chosen somewhat arbitrarily. 126 readers plus a
643 * couple mutexes fit exactly into 8KB on my development machine.
644 * Applications should set the table size using #mdb_env_set_maxreaders().
646 #define DEFAULT_READERS 126
648 /** The size of a CPU cache line in bytes. We want our lock structures
649 * aligned to this size to avoid false cache line sharing in the
651 * This value works for most CPUs. For Itanium this should be 128.
657 /** The information we store in a single slot of the reader table.
658 * In addition to a transaction ID, we also record the process and
659 * thread ID that owns a slot, so that we can detect stale information,
660 * e.g. threads or processes that went away without cleaning up.
661 * @note We currently don't check for stale records. We simply re-init
662 * the table when we know that we're the only process opening the
665 typedef struct MDB_rxbody {
666 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
667 * Multiple readers that start at the same time will probably have the
668 * same ID here. Again, it's not important to exclude them from
669 * anything; all we need to know is which version of the DB they
670 * started from so we can avoid overwriting any data used in that
671 * particular version.
673 volatile txnid_t mrb_txnid;
674 /** The process ID of the process owning this reader txn. */
675 volatile MDB_PID_T mrb_pid;
676 /** The thread ID of the thread owning this txn. */
677 volatile MDB_THR_T mrb_tid;
680 /** The actual reader record, with cacheline padding. */
681 typedef struct MDB_reader {
684 /** shorthand for mrb_txnid */
685 #define mr_txnid mru.mrx.mrb_txnid
686 #define mr_pid mru.mrx.mrb_pid
687 #define mr_tid mru.mrx.mrb_tid
688 /** cache line alignment */
689 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
693 /** The header for the reader table.
694 * The table resides in a memory-mapped file. (This is a different file
695 * than is used for the main database.)
697 * For POSIX the actual mutexes reside in the shared memory of this
698 * mapped file. On Windows, mutexes are named objects allocated by the
699 * kernel; we store the mutex names in this mapped file so that other
700 * processes can grab them. This same approach is also used on
701 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
702 * process-shared POSIX mutexes. For these cases where a named object
703 * is used, the object name is derived from a 64 bit FNV hash of the
704 * environment pathname. As such, naming collisions are extremely
705 * unlikely. If a collision occurs, the results are unpredictable.
707 typedef struct MDB_txbody {
708 /** Stamp identifying this as an LMDB file. It must be set
711 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
713 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
714 char mtb_rmname[MNAME_LEN];
716 /** Mutex protecting access to this table.
717 * This is the reader table lock used with LOCK_MUTEX().
719 mdb_mutex_t mtb_rmutex;
721 /** The ID of the last transaction committed to the database.
722 * This is recorded here only for convenience; the value can always
723 * be determined by reading the main database meta pages.
725 volatile txnid_t mtb_txnid;
726 /** The number of slots that have been used in the reader table.
727 * This always records the maximum count, it is not decremented
728 * when readers release their slots.
730 volatile unsigned mtb_numreaders;
733 /** The actual reader table definition. */
734 typedef struct MDB_txninfo {
737 #define mti_magic mt1.mtb.mtb_magic
738 #define mti_format mt1.mtb.mtb_format
739 #define mti_rmutex mt1.mtb.mtb_rmutex
740 #define mti_rmname mt1.mtb.mtb_rmname
741 #define mti_txnid mt1.mtb.mtb_txnid
742 #define mti_numreaders mt1.mtb.mtb_numreaders
743 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
746 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
747 char mt2_wmname[MNAME_LEN];
748 #define mti_wmname mt2.mt2_wmname
750 mdb_mutex_t mt2_wmutex;
751 #define mti_wmutex mt2.mt2_wmutex
753 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
755 MDB_reader mti_readers[1];
758 /** Lockfile format signature: version, features and field layout */
759 #define MDB_LOCK_FORMAT \
761 ((MDB_LOCK_VERSION) \
762 /* Flags which describe functionality */ \
763 + (((MDB_PIDLOCK) != 0) << 16)))
766 /** Common header for all page types.
767 * Overflow records occupy a number of contiguous pages with no
768 * headers on any page after the first.
770 typedef struct MDB_page {
771 #define mp_pgno mp_p.p_pgno
772 #define mp_next mp_p.p_next
774 pgno_t p_pgno; /**< page number */
775 struct MDB_page *p_next; /**< for in-memory list of freed pages */
778 /** @defgroup mdb_page Page Flags
780 * Flags for the page headers.
783 #define P_BRANCH 0x01 /**< branch page */
784 #define P_LEAF 0x02 /**< leaf page */
785 #define P_OVERFLOW 0x04 /**< overflow page */
786 #define P_META 0x08 /**< meta page */
787 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
788 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
789 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
790 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
791 #define P_KEEP 0x8000 /**< leave this page alone during spill */
793 uint16_t mp_flags; /**< @ref mdb_page */
794 #define mp_lower mp_pb.pb.pb_lower
795 #define mp_upper mp_pb.pb.pb_upper
796 #define mp_pages mp_pb.pb_pages
799 indx_t pb_lower; /**< lower bound of free space */
800 indx_t pb_upper; /**< upper bound of free space */
802 uint32_t pb_pages; /**< number of overflow pages */
804 indx_t mp_ptrs[1]; /**< dynamic size */
807 /** Size of the page header, excluding dynamic data at the end */
808 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
810 /** Address of first usable data byte in a page, after the header */
811 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
813 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
814 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
816 /** Number of nodes on a page */
817 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
819 /** The amount of space remaining in the page */
820 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
822 /** The percentage of space used in the page, in tenths of a percent. */
823 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
824 ((env)->me_psize - PAGEHDRSZ))
825 /** The minimum page fill factor, in tenths of a percent.
826 * Pages emptier than this are candidates for merging.
828 #define FILL_THRESHOLD 250
830 /** Test if a page is a leaf page */
831 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
832 /** Test if a page is a LEAF2 page */
833 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
834 /** Test if a page is a branch page */
835 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
836 /** Test if a page is an overflow page */
837 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
838 /** Test if a page is a sub page */
839 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
841 /** The number of overflow pages needed to store the given size. */
842 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
844 /** Link in #MDB_txn.%mt_loose_pgs list */
845 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
847 /** Header for a single key/data pair within a page.
848 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
849 * We guarantee 2-byte alignment for 'MDB_node's.
851 typedef struct MDB_node {
852 /** lo and hi are used for data size on leaf nodes and for
853 * child pgno on branch nodes. On 64 bit platforms, flags
854 * is also used for pgno. (Branch nodes have no flags).
855 * They are in host byte order in case that lets some
856 * accesses be optimized into a 32-bit word access.
858 #if BYTE_ORDER == LITTLE_ENDIAN
859 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
861 unsigned short mn_hi, mn_lo;
863 /** @defgroup mdb_node Node Flags
865 * Flags for node headers.
868 #define F_BIGDATA 0x01 /**< data put on overflow page */
869 #define F_SUBDATA 0x02 /**< data is a sub-database */
870 #define F_DUPDATA 0x04 /**< data has duplicates */
872 /** valid flags for #mdb_node_add() */
873 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
876 unsigned short mn_flags; /**< @ref mdb_node */
877 unsigned short mn_ksize; /**< key size */
878 char mn_data[1]; /**< key and data are appended here */
881 /** Size of the node header, excluding dynamic data at the end */
882 #define NODESIZE offsetof(MDB_node, mn_data)
884 /** Bit position of top word in page number, for shifting mn_flags */
885 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
887 /** Size of a node in a branch page with a given key.
888 * This is just the node header plus the key, there is no data.
890 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
892 /** Size of a node in a leaf page with a given key and data.
893 * This is node header plus key plus data size.
895 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
897 /** Address of node \b i in page \b p */
898 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
900 /** Address of the key for the node */
901 #define NODEKEY(node) (void *)((node)->mn_data)
903 /** Address of the data for a node */
904 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
906 /** Get the page number pointed to by a branch node */
907 #define NODEPGNO(node) \
908 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
909 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
910 /** Set the page number in a branch node */
911 #define SETPGNO(node,pgno) do { \
912 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
913 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
915 /** Get the size of the data in a leaf node */
916 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
917 /** Set the size of the data for a leaf node */
918 #define SETDSZ(node,size) do { \
919 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
920 /** The size of a key in a node */
921 #define NODEKSZ(node) ((node)->mn_ksize)
923 /** Copy a page number from src to dst */
925 #define COPY_PGNO(dst,src) dst = src
927 #if SIZE_MAX > 4294967295UL
928 #define COPY_PGNO(dst,src) do { \
929 unsigned short *s, *d; \
930 s = (unsigned short *)&(src); \
931 d = (unsigned short *)&(dst); \
938 #define COPY_PGNO(dst,src) do { \
939 unsigned short *s, *d; \
940 s = (unsigned short *)&(src); \
941 d = (unsigned short *)&(dst); \
947 /** The address of a key in a LEAF2 page.
948 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
949 * There are no node headers, keys are stored contiguously.
951 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
953 /** Set the \b node's key into \b keyptr, if requested. */
954 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
955 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
957 /** Set the \b node's key into \b key. */
958 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
960 /** Information about a single database in the environment. */
961 typedef struct MDB_db {
962 uint32_t md_pad; /**< also ksize for LEAF2 pages */
963 uint16_t md_flags; /**< @ref mdb_dbi_open */
964 uint16_t md_depth; /**< depth of this tree */
965 pgno_t md_branch_pages; /**< number of internal pages */
966 pgno_t md_leaf_pages; /**< number of leaf pages */
967 pgno_t md_overflow_pages; /**< number of overflow pages */
968 size_t md_entries; /**< number of data items */
969 pgno_t md_root; /**< the root page of this tree */
972 /** mdb_dbi_open flags */
973 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
974 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
975 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
976 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
978 /** Handle for the DB used to track free pages. */
980 /** Handle for the default DB. */
982 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
985 /** Number of meta pages - also hardcoded elsewhere */
988 /** Meta page content.
989 * A meta page is the start point for accessing a database snapshot.
990 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
992 typedef struct MDB_meta {
993 /** Stamp identifying this as an LMDB file. It must be set
996 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
998 void *mm_address; /**< address for fixed mapping */
999 size_t mm_mapsize; /**< size of mmap region */
1000 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1001 /** The size of pages used in this DB */
1002 #define mm_psize mm_dbs[FREE_DBI].md_pad
1003 /** Any persistent environment flags. @ref mdb_env */
1004 #define mm_flags mm_dbs[FREE_DBI].md_flags
1005 pgno_t mm_last_pg; /**< last used page in file */
1006 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1009 /** Buffer for a stack-allocated meta page.
1010 * The members define size and alignment, and silence type
1011 * aliasing warnings. They are not used directly; that could
1012 * mean incorrectly using several union members in parallel.
1014 typedef union MDB_metabuf {
1017 char mm_pad[PAGEHDRSZ];
1022 /** Auxiliary DB info.
1023 * The information here is mostly static/read-only. There is
1024 * only a single copy of this record in the environment.
1026 typedef struct MDB_dbx {
1027 MDB_val md_name; /**< name of the database */
1028 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1029 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1030 MDB_rel_func *md_rel; /**< user relocate function */
1031 void *md_relctx; /**< user-provided context for md_rel */
1034 /** A database transaction.
1035 * Every operation requires a transaction handle.
1038 MDB_txn *mt_parent; /**< parent of a nested txn */
1039 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1041 pgno_t mt_next_pgno; /**< next unallocated page */
1042 /** The ID of this transaction. IDs are integers incrementing from 1.
1043 * Only committed write transactions increment the ID. If a transaction
1044 * aborts, the ID may be re-used by the next writer.
1047 MDB_env *mt_env; /**< the DB environment */
1048 /** The list of pages that became unused during this transaction.
1050 MDB_IDL mt_free_pgs;
1051 /** The list of loose pages that became unused and may be reused
1052 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1054 MDB_page *mt_loose_pgs;
1055 /* #Number of loose pages (#mt_loose_pgs) */
1057 /** The sorted list of dirty pages we temporarily wrote to disk
1058 * because the dirty list was full. page numbers in here are
1059 * shifted left by 1, deleted slots have the LSB set.
1061 MDB_IDL mt_spill_pgs;
1063 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1064 MDB_ID2L dirty_list;
1065 /** For read txns: This thread/txn's reader table slot, or NULL. */
1068 /** Array of records for each DB known in the environment. */
1070 /** Array of MDB_db records for each known DB */
1072 /** Array of sequence numbers for each DB handle */
1073 unsigned int *mt_dbiseqs;
1074 /** @defgroup mt_dbflag Transaction DB Flags
1078 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1079 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1080 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1081 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1082 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1084 /** In write txns, array of cursors for each DB */
1085 MDB_cursor **mt_cursors;
1086 /** Array of flags for each DB */
1087 unsigned char *mt_dbflags;
1088 /** Number of DB records in use, or 0 when the txn is finished.
1089 * This number only ever increments until the txn finishes; we
1090 * don't decrement it when individual DB handles are closed.
1094 /** @defgroup mdb_txn Transaction Flags
1098 /** #mdb_txn_begin() flags */
1099 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1100 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1101 /* internal txn flags */
1102 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1103 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1104 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1105 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1106 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1107 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1108 /** most operations on the txn are currently illegal */
1109 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1111 unsigned int mt_flags; /**< @ref mdb_txn */
1112 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1113 * Includes ancestor txns' dirty pages not hidden by other txns'
1114 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1115 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1117 unsigned int mt_dirty_room;
1120 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1121 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1122 * raise this on a 64 bit machine.
1124 #define CURSOR_STACK 32
1128 /** Cursors are used for all DB operations.
1129 * A cursor holds a path of (page pointer, key index) from the DB
1130 * root to a position in the DB, plus other state. #MDB_DUPSORT
1131 * cursors include an xcursor to the current data item. Write txns
1132 * track their cursors and keep them up to date when data moves.
1133 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1134 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1137 /** Next cursor on this DB in this txn */
1138 MDB_cursor *mc_next;
1139 /** Backup of the original cursor if this cursor is a shadow */
1140 MDB_cursor *mc_backup;
1141 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1142 struct MDB_xcursor *mc_xcursor;
1143 /** The transaction that owns this cursor */
1145 /** The database handle this cursor operates on */
1147 /** The database record for this cursor */
1149 /** The database auxiliary record for this cursor */
1151 /** The @ref mt_dbflag for this database */
1152 unsigned char *mc_dbflag;
1153 unsigned short mc_snum; /**< number of pushed pages */
1154 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1155 /** @defgroup mdb_cursor Cursor Flags
1157 * Cursor state flags.
1160 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1161 #define C_EOF 0x02 /**< No more data */
1162 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1163 #define C_DEL 0x08 /**< last op was a cursor_del */
1164 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1166 unsigned int mc_flags; /**< @ref mdb_cursor */
1167 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1168 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1171 /** Context for sorted-dup records.
1172 * We could have gone to a fully recursive design, with arbitrarily
1173 * deep nesting of sub-databases. But for now we only handle these
1174 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1176 typedef struct MDB_xcursor {
1177 /** A sub-cursor for traversing the Dup DB */
1178 MDB_cursor mx_cursor;
1179 /** The database record for this Dup DB */
1181 /** The auxiliary DB record for this Dup DB */
1183 /** The @ref mt_dbflag for this Dup DB */
1184 unsigned char mx_dbflag;
1187 /** State of FreeDB old pages, stored in the MDB_env */
1188 typedef struct MDB_pgstate {
1189 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1190 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1193 /** The database environment. */
1195 HANDLE me_fd; /**< The main data file */
1196 HANDLE me_lfd; /**< The lock file */
1197 HANDLE me_mfd; /**< just for writing the meta pages */
1198 /** Failed to update the meta page. Probably an I/O error. */
1199 #define MDB_FATAL_ERROR 0x80000000U
1200 /** Some fields are initialized. */
1201 #define MDB_ENV_ACTIVE 0x20000000U
1202 /** me_txkey is set */
1203 #define MDB_ENV_TXKEY 0x10000000U
1204 /** fdatasync is unreliable */
1205 #define MDB_FSYNCONLY 0x08000000U
1206 uint32_t me_flags; /**< @ref mdb_env */
1207 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1208 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1209 unsigned int me_maxreaders; /**< size of the reader table */
1210 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1211 volatile int me_close_readers;
1212 MDB_dbi me_numdbs; /**< number of DBs opened */
1213 MDB_dbi me_maxdbs; /**< size of the DB table */
1214 MDB_PID_T me_pid; /**< process ID of this env */
1215 char *me_path; /**< path to the DB files */
1216 char *me_map; /**< the memory map of the data file */
1217 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1218 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1219 void *me_pbuf; /**< scratch area for DUPSORT put() */
1220 MDB_txn *me_txn; /**< current write transaction */
1221 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1222 size_t me_mapsize; /**< size of the data memory map */
1223 off_t me_size; /**< current file size */
1224 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1225 MDB_dbx *me_dbxs; /**< array of static DB info */
1226 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1227 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1228 pthread_key_t me_txkey; /**< thread-key for readers */
1229 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1230 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1231 # define me_pglast me_pgstate.mf_pglast
1232 # define me_pghead me_pgstate.mf_pghead
1233 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1234 /** IDL of pages that became unused in a write txn */
1235 MDB_IDL me_free_pgs;
1236 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1237 MDB_ID2L me_dirty_list;
1238 /** Max number of freelist items that can fit in a single overflow page */
1240 /** Max size of a node on a page */
1241 unsigned int me_nodemax;
1242 #if !(MDB_MAXKEYSIZE)
1243 unsigned int me_maxkey; /**< max size of a key */
1245 int me_live_reader; /**< have liveness lock in reader table */
1247 int me_pidquery; /**< Used in OpenProcess */
1249 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1250 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1251 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1253 mdb_mutex_t me_rmutex;
1254 mdb_mutex_t me_wmutex;
1256 void *me_userctx; /**< User-settable context */
1257 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1260 /** Nested transaction */
1261 typedef struct MDB_ntxn {
1262 MDB_txn mnt_txn; /**< the transaction */
1263 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1266 /** max number of pages to commit in one writev() call */
1267 #define MDB_COMMIT_PAGES 64
1268 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1269 #undef MDB_COMMIT_PAGES
1270 #define MDB_COMMIT_PAGES IOV_MAX
1273 /** max bytes to write in one call */
1274 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1276 /** Check \b txn and \b dbi arguments to a function */
1277 #define TXN_DBI_EXIST(txn, dbi, validity) \
1278 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1280 /** Check for misused \b dbi handles */
1281 #define TXN_DBI_CHANGED(txn, dbi) \
1282 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1284 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1285 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1286 static int mdb_page_touch(MDB_cursor *mc);
1288 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1289 "reset-tmp", "fail-begin", "fail-beginchild"}
1291 /* mdb_txn_end operation number, for logging */
1292 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1293 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1295 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1296 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1297 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1298 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1299 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1301 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1302 static int mdb_page_search_root(MDB_cursor *mc,
1303 MDB_val *key, int modify);
1304 #define MDB_PS_MODIFY 1
1305 #define MDB_PS_ROOTONLY 2
1306 #define MDB_PS_FIRST 4
1307 #define MDB_PS_LAST 8
1308 static int mdb_page_search(MDB_cursor *mc,
1309 MDB_val *key, int flags);
1310 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1312 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1313 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1314 pgno_t newpgno, unsigned int nflags);
1316 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1317 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1318 static int mdb_env_write_meta(MDB_txn *txn);
1319 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1320 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1322 static void mdb_env_close0(MDB_env *env, int excl);
1324 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1325 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1326 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1327 static void mdb_node_del(MDB_cursor *mc, int ksize);
1328 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1329 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1330 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1331 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1332 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1334 static int mdb_rebalance(MDB_cursor *mc);
1335 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1337 static void mdb_cursor_pop(MDB_cursor *mc);
1338 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1340 static int mdb_cursor_del0(MDB_cursor *mc);
1341 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1342 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1343 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1344 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1345 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1347 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1348 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1350 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1351 static void mdb_xcursor_init0(MDB_cursor *mc);
1352 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1353 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1355 static int mdb_drop0(MDB_cursor *mc, int subs);
1356 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1357 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1360 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1363 /** Compare two items pointing at size_t's of unknown alignment. */
1364 #ifdef MISALIGNED_OK
1365 # define mdb_cmp_clong mdb_cmp_long
1367 # define mdb_cmp_clong mdb_cmp_cint
1371 static SECURITY_DESCRIPTOR mdb_null_sd;
1372 static SECURITY_ATTRIBUTES mdb_all_sa;
1373 static int mdb_sec_inited;
1375 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize);
1378 /** Return the library version info. */
1380 mdb_version(int *major, int *minor, int *patch)
1382 if (major) *major = MDB_VERSION_MAJOR;
1383 if (minor) *minor = MDB_VERSION_MINOR;
1384 if (patch) *patch = MDB_VERSION_PATCH;
1385 return MDB_VERSION_STRING;
1388 /** Table of descriptions for LMDB @ref errors */
1389 static char *const mdb_errstr[] = {
1390 "MDB_KEYEXIST: Key/data pair already exists",
1391 "MDB_NOTFOUND: No matching key/data pair found",
1392 "MDB_PAGE_NOTFOUND: Requested page not found",
1393 "MDB_CORRUPTED: Located page was wrong type",
1394 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1395 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1396 "MDB_INVALID: File is not an LMDB file",
1397 "MDB_MAP_FULL: Environment mapsize limit reached",
1398 "MDB_DBS_FULL: Environment maxdbs limit reached",
1399 "MDB_READERS_FULL: Environment maxreaders limit reached",
1400 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1401 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1402 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1403 "MDB_PAGE_FULL: Internal error - page has no more space",
1404 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1405 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1406 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1407 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1408 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1409 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1413 mdb_strerror(int err)
1416 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1417 * This works as long as no function between the call to mdb_strerror
1418 * and the actual use of the message uses more than 4K of stack.
1421 char buf[1024], *ptr = buf;
1425 return ("Successful return: 0");
1427 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1428 i = err - MDB_KEYEXIST;
1429 return mdb_errstr[i];
1433 /* These are the C-runtime error codes we use. The comment indicates
1434 * their numeric value, and the Win32 error they would correspond to
1435 * if the error actually came from a Win32 API. A major mess, we should
1436 * have used LMDB-specific error codes for everything.
1439 case ENOENT: /* 2, FILE_NOT_FOUND */
1440 case EIO: /* 5, ACCESS_DENIED */
1441 case ENOMEM: /* 12, INVALID_ACCESS */
1442 case EACCES: /* 13, INVALID_DATA */
1443 case EBUSY: /* 16, CURRENT_DIRECTORY */
1444 case EINVAL: /* 22, BAD_COMMAND */
1445 case ENOSPC: /* 28, OUT_OF_PAPER */
1446 return strerror(err);
1451 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1452 FORMAT_MESSAGE_IGNORE_INSERTS,
1453 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1456 return strerror(err);
1460 /** assert(3) variant in cursor context */
1461 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1462 /** assert(3) variant in transaction context */
1463 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1464 /** assert(3) variant in environment context */
1465 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1468 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1469 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1472 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1473 const char *func, const char *file, int line)
1476 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1477 file, line, expr_txt, func);
1478 if (env->me_assert_func)
1479 env->me_assert_func(env, buf);
1480 fprintf(stderr, "%s\n", buf);
1484 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1488 /** Return the page number of \b mp which may be sub-page, for debug output */
1490 mdb_dbg_pgno(MDB_page *mp)
1493 COPY_PGNO(ret, mp->mp_pgno);
1497 /** Display a key in hexadecimal and return the address of the result.
1498 * @param[in] key the key to display
1499 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1500 * @return The key in hexadecimal form.
1503 mdb_dkey(MDB_val *key, char *buf)
1506 unsigned char *c = key->mv_data;
1512 if (key->mv_size > DKBUF_MAXKEYSIZE)
1513 return "MDB_MAXKEYSIZE";
1514 /* may want to make this a dynamic check: if the key is mostly
1515 * printable characters, print it as-is instead of converting to hex.
1519 for (i=0; i<key->mv_size; i++)
1520 ptr += sprintf(ptr, "%02x", *c++);
1522 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1528 mdb_leafnode_type(MDB_node *n)
1530 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1531 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1532 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1535 /** Display all the keys in the page. */
1537 mdb_page_list(MDB_page *mp)
1539 pgno_t pgno = mdb_dbg_pgno(mp);
1540 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1542 unsigned int i, nkeys, nsize, total = 0;
1546 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1547 case P_BRANCH: type = "Branch page"; break;
1548 case P_LEAF: type = "Leaf page"; break;
1549 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1550 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1551 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1553 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1554 pgno, mp->mp_pages, state);
1557 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1558 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1561 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1565 nkeys = NUMKEYS(mp);
1566 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1568 for (i=0; i<nkeys; i++) {
1569 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1570 key.mv_size = nsize = mp->mp_pad;
1571 key.mv_data = LEAF2KEY(mp, i, nsize);
1573 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1576 node = NODEPTR(mp, i);
1577 key.mv_size = node->mn_ksize;
1578 key.mv_data = node->mn_data;
1579 nsize = NODESIZE + key.mv_size;
1580 if (IS_BRANCH(mp)) {
1581 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1585 if (F_ISSET(node->mn_flags, F_BIGDATA))
1586 nsize += sizeof(pgno_t);
1588 nsize += NODEDSZ(node);
1590 nsize += sizeof(indx_t);
1591 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1592 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1594 total = EVEN(total);
1596 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1597 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1601 mdb_cursor_chk(MDB_cursor *mc)
1607 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1608 for (i=0; i<mc->mc_top; i++) {
1610 node = NODEPTR(mp, mc->mc_ki[i]);
1611 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1614 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1620 /** Count all the pages in each DB and in the freelist
1621 * and make sure it matches the actual number of pages
1623 * All named DBs must be open for a correct count.
1625 static void mdb_audit(MDB_txn *txn)
1629 MDB_ID freecount, count;
1634 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1635 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1636 freecount += *(MDB_ID *)data.mv_data;
1637 mdb_tassert(txn, rc == MDB_NOTFOUND);
1640 for (i = 0; i<txn->mt_numdbs; i++) {
1642 if (!(txn->mt_dbflags[i] & DB_VALID))
1644 mdb_cursor_init(&mc, txn, i, &mx);
1645 if (txn->mt_dbs[i].md_root == P_INVALID)
1647 count += txn->mt_dbs[i].md_branch_pages +
1648 txn->mt_dbs[i].md_leaf_pages +
1649 txn->mt_dbs[i].md_overflow_pages;
1650 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1651 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1652 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1655 mp = mc.mc_pg[mc.mc_top];
1656 for (j=0; j<NUMKEYS(mp); j++) {
1657 MDB_node *leaf = NODEPTR(mp, j);
1658 if (leaf->mn_flags & F_SUBDATA) {
1660 memcpy(&db, NODEDATA(leaf), sizeof(db));
1661 count += db.md_branch_pages + db.md_leaf_pages +
1662 db.md_overflow_pages;
1666 mdb_tassert(txn, rc == MDB_NOTFOUND);
1669 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1670 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1671 txn->mt_txnid, freecount, count+NUM_METAS,
1672 freecount+count+NUM_METAS, txn->mt_next_pgno);
1678 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1680 return txn->mt_dbxs[dbi].md_cmp(a, b);
1684 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1686 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1687 #if UINT_MAX < SIZE_MAX
1688 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1689 dcmp = mdb_cmp_clong;
1694 /** Allocate memory for a page.
1695 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1698 mdb_page_malloc(MDB_txn *txn, unsigned num)
1700 MDB_env *env = txn->mt_env;
1701 MDB_page *ret = env->me_dpages;
1702 size_t psize = env->me_psize, sz = psize, off;
1703 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1704 * For a single page alloc, we init everything after the page header.
1705 * For multi-page, we init the final page; if the caller needed that
1706 * many pages they will be filling in at least up to the last page.
1710 VGMEMP_ALLOC(env, ret, sz);
1711 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1712 env->me_dpages = ret->mp_next;
1715 psize -= off = PAGEHDRSZ;
1720 if ((ret = malloc(sz)) != NULL) {
1721 VGMEMP_ALLOC(env, ret, sz);
1722 if (!(env->me_flags & MDB_NOMEMINIT)) {
1723 memset((char *)ret + off, 0, psize);
1727 txn->mt_flags |= MDB_TXN_ERROR;
1731 /** Free a single page.
1732 * Saves single pages to a list, for future reuse.
1733 * (This is not used for multi-page overflow pages.)
1736 mdb_page_free(MDB_env *env, MDB_page *mp)
1738 mp->mp_next = env->me_dpages;
1739 VGMEMP_FREE(env, mp);
1740 env->me_dpages = mp;
1743 /** Free a dirty page */
1745 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1747 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1748 mdb_page_free(env, dp);
1750 /* large pages just get freed directly */
1751 VGMEMP_FREE(env, dp);
1756 /** Return all dirty pages to dpage list */
1758 mdb_dlist_free(MDB_txn *txn)
1760 MDB_env *env = txn->mt_env;
1761 MDB_ID2L dl = txn->mt_u.dirty_list;
1762 unsigned i, n = dl[0].mid;
1764 for (i = 1; i <= n; i++) {
1765 mdb_dpage_free(env, dl[i].mptr);
1770 /** Loosen or free a single page.
1771 * Saves single pages to a list for future reuse
1772 * in this same txn. It has been pulled from the freeDB
1773 * and already resides on the dirty list, but has been
1774 * deleted. Use these pages first before pulling again
1777 * If the page wasn't dirtied in this txn, just add it
1778 * to this txn's free list.
1781 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1784 pgno_t pgno = mp->mp_pgno;
1785 MDB_txn *txn = mc->mc_txn;
1787 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1788 if (txn->mt_parent) {
1789 MDB_ID2 *dl = txn->mt_u.dirty_list;
1790 /* If txn has a parent, make sure the page is in our
1794 unsigned x = mdb_mid2l_search(dl, pgno);
1795 if (x <= dl[0].mid && dl[x].mid == pgno) {
1796 if (mp != dl[x].mptr) { /* bad cursor? */
1797 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1798 txn->mt_flags |= MDB_TXN_ERROR;
1799 return MDB_CORRUPTED;
1806 /* no parent txn, so it's just ours */
1811 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1813 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1814 txn->mt_loose_pgs = mp;
1815 txn->mt_loose_count++;
1816 mp->mp_flags |= P_LOOSE;
1818 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1826 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1827 * @param[in] mc A cursor handle for the current operation.
1828 * @param[in] pflags Flags of the pages to update:
1829 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1830 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1831 * @return 0 on success, non-zero on failure.
1834 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1836 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1837 MDB_txn *txn = mc->mc_txn;
1843 int rc = MDB_SUCCESS, level;
1845 /* Mark pages seen by cursors */
1846 if (mc->mc_flags & C_UNTRACK)
1847 mc = NULL; /* will find mc in mt_cursors */
1848 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1849 for (; mc; mc=mc->mc_next) {
1850 if (!(mc->mc_flags & C_INITIALIZED))
1852 for (m3 = mc;; m3 = &mx->mx_cursor) {
1854 for (j=0; j<m3->mc_snum; j++) {
1856 if ((mp->mp_flags & Mask) == pflags)
1857 mp->mp_flags ^= P_KEEP;
1859 mx = m3->mc_xcursor;
1860 /* Proceed to mx if it is at a sub-database */
1861 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1863 if (! (mp && (mp->mp_flags & P_LEAF)))
1865 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1866 if (!(leaf->mn_flags & F_SUBDATA))
1875 /* Mark dirty root pages */
1876 for (i=0; i<txn->mt_numdbs; i++) {
1877 if (txn->mt_dbflags[i] & DB_DIRTY) {
1878 pgno_t pgno = txn->mt_dbs[i].md_root;
1879 if (pgno == P_INVALID)
1881 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1883 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1884 dp->mp_flags ^= P_KEEP;
1892 static int mdb_page_flush(MDB_txn *txn, int keep);
1894 /** Spill pages from the dirty list back to disk.
1895 * This is intended to prevent running into #MDB_TXN_FULL situations,
1896 * but note that they may still occur in a few cases:
1897 * 1) our estimate of the txn size could be too small. Currently this
1898 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1899 * 2) child txns may run out of space if their parents dirtied a
1900 * lot of pages and never spilled them. TODO: we probably should do
1901 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1902 * the parent's dirty_room is below a given threshold.
1904 * Otherwise, if not using nested txns, it is expected that apps will
1905 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1906 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1907 * If the txn never references them again, they can be left alone.
1908 * If the txn only reads them, they can be used without any fuss.
1909 * If the txn writes them again, they can be dirtied immediately without
1910 * going thru all of the work of #mdb_page_touch(). Such references are
1911 * handled by #mdb_page_unspill().
1913 * Also note, we never spill DB root pages, nor pages of active cursors,
1914 * because we'll need these back again soon anyway. And in nested txns,
1915 * we can't spill a page in a child txn if it was already spilled in a
1916 * parent txn. That would alter the parent txns' data even though
1917 * the child hasn't committed yet, and we'd have no way to undo it if
1918 * the child aborted.
1920 * @param[in] m0 cursor A cursor handle identifying the transaction and
1921 * database for which we are checking space.
1922 * @param[in] key For a put operation, the key being stored.
1923 * @param[in] data For a put operation, the data being stored.
1924 * @return 0 on success, non-zero on failure.
1927 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1929 MDB_txn *txn = m0->mc_txn;
1931 MDB_ID2L dl = txn->mt_u.dirty_list;
1932 unsigned int i, j, need;
1935 if (m0->mc_flags & C_SUB)
1938 /* Estimate how much space this op will take */
1939 i = m0->mc_db->md_depth;
1940 /* Named DBs also dirty the main DB */
1941 if (m0->mc_dbi >= CORE_DBS)
1942 i += txn->mt_dbs[MAIN_DBI].md_depth;
1943 /* For puts, roughly factor in the key+data size */
1945 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1946 i += i; /* double it for good measure */
1949 if (txn->mt_dirty_room > i)
1952 if (!txn->mt_spill_pgs) {
1953 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1954 if (!txn->mt_spill_pgs)
1957 /* purge deleted slots */
1958 MDB_IDL sl = txn->mt_spill_pgs;
1959 unsigned int num = sl[0];
1961 for (i=1; i<=num; i++) {
1968 /* Preserve pages which may soon be dirtied again */
1969 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1972 /* Less aggressive spill - we originally spilled the entire dirty list,
1973 * with a few exceptions for cursor pages and DB root pages. But this
1974 * turns out to be a lot of wasted effort because in a large txn many
1975 * of those pages will need to be used again. So now we spill only 1/8th
1976 * of the dirty pages. Testing revealed this to be a good tradeoff,
1977 * better than 1/2, 1/4, or 1/10.
1979 if (need < MDB_IDL_UM_MAX / 8)
1980 need = MDB_IDL_UM_MAX / 8;
1982 /* Save the page IDs of all the pages we're flushing */
1983 /* flush from the tail forward, this saves a lot of shifting later on. */
1984 for (i=dl[0].mid; i && need; i--) {
1985 MDB_ID pn = dl[i].mid << 1;
1987 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1989 /* Can't spill twice, make sure it's not already in a parent's
1992 if (txn->mt_parent) {
1994 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1995 if (tx2->mt_spill_pgs) {
1996 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1997 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1998 dp->mp_flags |= P_KEEP;
2006 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2010 mdb_midl_sort(txn->mt_spill_pgs);
2012 /* Flush the spilled part of dirty list */
2013 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2016 /* Reset any dirty pages we kept that page_flush didn't see */
2017 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2020 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2024 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2026 mdb_find_oldest(MDB_txn *txn)
2029 txnid_t mr, oldest = txn->mt_txnid - 1;
2030 if (txn->mt_env->me_txns) {
2031 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2032 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2043 /** Add a page to the txn's dirty list */
2045 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2048 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2050 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2051 insert = mdb_mid2l_append;
2053 insert = mdb_mid2l_insert;
2055 mid.mid = mp->mp_pgno;
2057 rc = insert(txn->mt_u.dirty_list, &mid);
2058 mdb_tassert(txn, rc == 0);
2059 txn->mt_dirty_room--;
2062 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2063 * me_pghead and mt_next_pgno.
2065 * If there are free pages available from older transactions, they
2066 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2067 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2068 * and move me_pglast to say which records were consumed. Only this
2069 * function can create me_pghead and move me_pglast/mt_next_pgno.
2070 * @param[in] mc cursor A cursor handle identifying the transaction and
2071 * database for which we are allocating.
2072 * @param[in] num the number of pages to allocate.
2073 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2074 * will always be satisfied by a single contiguous chunk of memory.
2075 * @return 0 on success, non-zero on failure.
2078 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2080 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2081 /* Get at most <Max_retries> more freeDB records once me_pghead
2082 * has enough pages. If not enough, use new pages from the map.
2083 * If <Paranoid> and mc is updating the freeDB, only get new
2084 * records if me_pghead is empty. Then the freelist cannot play
2085 * catch-up with itself by growing while trying to save it.
2087 enum { Paranoid = 1, Max_retries = 500 };
2089 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2091 int rc, retry = num * 60;
2092 MDB_txn *txn = mc->mc_txn;
2093 MDB_env *env = txn->mt_env;
2094 pgno_t pgno, *mop = env->me_pghead;
2095 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2097 txnid_t oldest = 0, last;
2102 /* If there are any loose pages, just use them */
2103 if (num == 1 && txn->mt_loose_pgs) {
2104 np = txn->mt_loose_pgs;
2105 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2106 txn->mt_loose_count--;
2107 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2115 /* If our dirty list is already full, we can't do anything */
2116 if (txn->mt_dirty_room == 0) {
2121 for (op = MDB_FIRST;; op = MDB_NEXT) {
2126 /* Seek a big enough contiguous page range. Prefer
2127 * pages at the tail, just truncating the list.
2133 if (mop[i-n2] == pgno+n2)
2140 if (op == MDB_FIRST) { /* 1st iteration */
2141 /* Prepare to fetch more and coalesce */
2142 last = env->me_pglast;
2143 oldest = env->me_pgoldest;
2144 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2147 key.mv_data = &last; /* will look up last+1 */
2148 key.mv_size = sizeof(last);
2150 if (Paranoid && mc->mc_dbi == FREE_DBI)
2153 if (Paranoid && retry < 0 && mop_len)
2157 /* Do not fetch more if the record will be too recent */
2158 if (oldest <= last) {
2160 oldest = mdb_find_oldest(txn);
2161 env->me_pgoldest = oldest;
2167 rc = mdb_cursor_get(&m2, &key, NULL, op);
2169 if (rc == MDB_NOTFOUND)
2173 last = *(txnid_t*)key.mv_data;
2174 if (oldest <= last) {
2176 oldest = mdb_find_oldest(txn);
2177 env->me_pgoldest = oldest;
2183 np = m2.mc_pg[m2.mc_top];
2184 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2185 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2188 idl = (MDB_ID *) data.mv_data;
2191 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2196 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2198 mop = env->me_pghead;
2200 env->me_pglast = last;
2202 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2203 last, txn->mt_dbs[FREE_DBI].md_root, i));
2205 DPRINTF(("IDL %"Z"u", idl[j]));
2207 /* Merge in descending sorted order */
2208 mdb_midl_xmerge(mop, idl);
2212 /* Use new pages from the map when nothing suitable in the freeDB */
2214 pgno = txn->mt_next_pgno;
2215 if (pgno + num >= env->me_maxpg) {
2216 DPUTS("DB size maxed out");
2222 if (env->me_flags & MDB_WRITEMAP) {
2223 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2225 if (!(np = mdb_page_malloc(txn, num))) {
2231 mop[0] = mop_len -= num;
2232 /* Move any stragglers down */
2233 for (j = i-num; j < mop_len; )
2234 mop[++j] = mop[++i];
2236 txn->mt_next_pgno = pgno + num;
2239 mdb_page_dirty(txn, np);
2245 txn->mt_flags |= MDB_TXN_ERROR;
2249 /** Copy the used portions of a non-overflow page.
2250 * @param[in] dst page to copy into
2251 * @param[in] src page to copy from
2252 * @param[in] psize size of a page
2255 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2257 enum { Align = sizeof(pgno_t) };
2258 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2260 /* If page isn't full, just copy the used portion. Adjust
2261 * alignment so memcpy may copy words instead of bytes.
2263 if ((unused &= -Align) && !IS_LEAF2(src)) {
2264 upper = (upper + PAGEBASE) & -Align;
2265 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2266 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2269 memcpy(dst, src, psize - unused);
2273 /** Pull a page off the txn's spill list, if present.
2274 * If a page being referenced was spilled to disk in this txn, bring
2275 * it back and make it dirty/writable again.
2276 * @param[in] txn the transaction handle.
2277 * @param[in] mp the page being referenced. It must not be dirty.
2278 * @param[out] ret the writable page, if any. ret is unchanged if
2279 * mp wasn't spilled.
2282 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2284 MDB_env *env = txn->mt_env;
2287 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2289 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2290 if (!tx2->mt_spill_pgs)
2292 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2293 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2296 if (txn->mt_dirty_room == 0)
2297 return MDB_TXN_FULL;
2298 if (IS_OVERFLOW(mp))
2302 if (env->me_flags & MDB_WRITEMAP) {
2305 np = mdb_page_malloc(txn, num);
2309 memcpy(np, mp, num * env->me_psize);
2311 mdb_page_copy(np, mp, env->me_psize);
2314 /* If in current txn, this page is no longer spilled.
2315 * If it happens to be the last page, truncate the spill list.
2316 * Otherwise mark it as deleted by setting the LSB.
2318 if (x == txn->mt_spill_pgs[0])
2319 txn->mt_spill_pgs[0]--;
2321 txn->mt_spill_pgs[x] |= 1;
2322 } /* otherwise, if belonging to a parent txn, the
2323 * page remains spilled until child commits
2326 mdb_page_dirty(txn, np);
2327 np->mp_flags |= P_DIRTY;
2335 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2336 * @param[in] mc cursor pointing to the page to be touched
2337 * @return 0 on success, non-zero on failure.
2340 mdb_page_touch(MDB_cursor *mc)
2342 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2343 MDB_txn *txn = mc->mc_txn;
2344 MDB_cursor *m2, *m3;
2348 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2349 if (txn->mt_flags & MDB_TXN_SPILLS) {
2351 rc = mdb_page_unspill(txn, mp, &np);
2357 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2358 (rc = mdb_page_alloc(mc, 1, &np)))
2361 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2362 mp->mp_pgno, pgno));
2363 mdb_cassert(mc, mp->mp_pgno != pgno);
2364 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2365 /* Update the parent page, if any, to point to the new page */
2367 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2368 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2369 SETPGNO(node, pgno);
2371 mc->mc_db->md_root = pgno;
2373 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2374 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2376 /* If txn has a parent, make sure the page is in our
2380 unsigned x = mdb_mid2l_search(dl, pgno);
2381 if (x <= dl[0].mid && dl[x].mid == pgno) {
2382 if (mp != dl[x].mptr) { /* bad cursor? */
2383 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2384 txn->mt_flags |= MDB_TXN_ERROR;
2385 return MDB_CORRUPTED;
2390 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2392 np = mdb_page_malloc(txn, 1);
2397 rc = mdb_mid2l_insert(dl, &mid);
2398 mdb_cassert(mc, rc == 0);
2403 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2405 np->mp_flags |= P_DIRTY;
2408 /* Adjust cursors pointing to mp */
2409 mc->mc_pg[mc->mc_top] = np;
2410 m2 = txn->mt_cursors[mc->mc_dbi];
2411 if (mc->mc_flags & C_SUB) {
2412 for (; m2; m2=m2->mc_next) {
2413 m3 = &m2->mc_xcursor->mx_cursor;
2414 if (m3->mc_snum < mc->mc_snum) continue;
2415 if (m3->mc_pg[mc->mc_top] == mp)
2416 m3->mc_pg[mc->mc_top] = np;
2419 for (; m2; m2=m2->mc_next) {
2420 if (m2->mc_snum < mc->mc_snum) continue;
2421 if (m2->mc_pg[mc->mc_top] == mp) {
2422 m2->mc_pg[mc->mc_top] = np;
2423 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2425 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2427 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2428 if (!(leaf->mn_flags & F_SUBDATA))
2429 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2437 txn->mt_flags |= MDB_TXN_ERROR;
2442 mdb_env_sync(MDB_env *env, int force)
2445 if (env->me_flags & MDB_RDONLY)
2447 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2448 if (env->me_flags & MDB_WRITEMAP) {
2449 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2450 ? MS_ASYNC : MS_SYNC;
2451 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2454 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2458 #ifdef BROKEN_FDATASYNC
2459 if (env->me_flags & MDB_FSYNCONLY) {
2460 if (fsync(env->me_fd))
2464 if (MDB_FDATASYNC(env->me_fd))
2471 /** Back up parent txn's cursors, then grab the originals for tracking */
2473 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2475 MDB_cursor *mc, *bk;
2480 for (i = src->mt_numdbs; --i >= 0; ) {
2481 if ((mc = src->mt_cursors[i]) != NULL) {
2482 size = sizeof(MDB_cursor);
2484 size += sizeof(MDB_xcursor);
2485 for (; mc; mc = bk->mc_next) {
2491 mc->mc_db = &dst->mt_dbs[i];
2492 /* Kill pointers into src to reduce abuse: The
2493 * user may not use mc until dst ends. But we need a valid
2494 * txn pointer here for cursor fixups to keep working.
2497 mc->mc_dbflag = &dst->mt_dbflags[i];
2498 if ((mx = mc->mc_xcursor) != NULL) {
2499 *(MDB_xcursor *)(bk+1) = *mx;
2500 mx->mx_cursor.mc_txn = dst;
2502 mc->mc_next = dst->mt_cursors[i];
2503 dst->mt_cursors[i] = mc;
2510 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2511 * @param[in] txn the transaction handle.
2512 * @param[in] merge true to keep changes to parent cursors, false to revert.
2513 * @return 0 on success, non-zero on failure.
2516 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2518 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2522 for (i = txn->mt_numdbs; --i >= 0; ) {
2523 for (mc = cursors[i]; mc; mc = next) {
2525 if ((bk = mc->mc_backup) != NULL) {
2527 /* Commit changes to parent txn */
2528 mc->mc_next = bk->mc_next;
2529 mc->mc_backup = bk->mc_backup;
2530 mc->mc_txn = bk->mc_txn;
2531 mc->mc_db = bk->mc_db;
2532 mc->mc_dbflag = bk->mc_dbflag;
2533 if ((mx = mc->mc_xcursor) != NULL)
2534 mx->mx_cursor.mc_txn = bk->mc_txn;
2536 /* Abort nested txn */
2538 if ((mx = mc->mc_xcursor) != NULL)
2539 *mx = *(MDB_xcursor *)(bk+1);
2543 /* Only malloced cursors are permanently tracked. */
2550 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2556 Pidset = F_SETLK, Pidcheck = F_GETLK
2560 /** Set or check a pid lock. Set returns 0 on success.
2561 * Check returns 0 if the process is certainly dead, nonzero if it may
2562 * be alive (the lock exists or an error happened so we do not know).
2564 * On Windows Pidset is a no-op, we merely check for the existence
2565 * of the process with the given pid. On POSIX we use a single byte
2566 * lock on the lockfile, set at an offset equal to the pid.
2569 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2571 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2574 if (op == Pidcheck) {
2575 h = OpenProcess(env->me_pidquery, FALSE, pid);
2576 /* No documented "no such process" code, but other program use this: */
2578 return ErrCode() != ERROR_INVALID_PARAMETER;
2579 /* A process exists until all handles to it close. Has it exited? */
2580 ret = WaitForSingleObject(h, 0) != 0;
2587 struct flock lock_info;
2588 memset(&lock_info, 0, sizeof(lock_info));
2589 lock_info.l_type = F_WRLCK;
2590 lock_info.l_whence = SEEK_SET;
2591 lock_info.l_start = pid;
2592 lock_info.l_len = 1;
2593 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2594 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2596 } else if ((rc = ErrCode()) == EINTR) {
2604 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2605 * @param[in] txn the transaction handle to initialize
2606 * @return 0 on success, non-zero on failure.
2609 mdb_txn_renew0(MDB_txn *txn)
2611 MDB_env *env = txn->mt_env;
2612 MDB_txninfo *ti = env->me_txns;
2614 unsigned int i, nr, flags = txn->mt_flags;
2616 int rc, new_notls = 0;
2618 if ((flags &= MDB_TXN_RDONLY) != 0) {
2620 meta = mdb_env_pick_meta(env);
2621 txn->mt_txnid = meta->mm_txnid;
2622 txn->mt_u.reader = NULL;
2624 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2625 pthread_getspecific(env->me_txkey);
2627 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2628 return MDB_BAD_RSLOT;
2630 MDB_PID_T pid = env->me_pid;
2631 MDB_THR_T tid = pthread_self();
2632 mdb_mutexref_t rmutex = env->me_rmutex;
2634 if (!env->me_live_reader) {
2635 rc = mdb_reader_pid(env, Pidset, pid);
2638 env->me_live_reader = 1;
2641 if (LOCK_MUTEX(rc, env, rmutex))
2643 nr = ti->mti_numreaders;
2644 for (i=0; i<nr; i++)
2645 if (ti->mti_readers[i].mr_pid == 0)
2647 if (i == env->me_maxreaders) {
2648 UNLOCK_MUTEX(rmutex);
2649 return MDB_READERS_FULL;
2651 r = &ti->mti_readers[i];
2652 /* Claim the reader slot, carefully since other code
2653 * uses the reader table un-mutexed: First reset the
2654 * slot, next publish it in mti_numreaders. After
2655 * that, it is safe for mdb_env_close() to touch it.
2656 * When it will be closed, we can finally claim it.
2659 r->mr_txnid = (txnid_t)-1;
2662 ti->mti_numreaders = ++nr;
2663 env->me_close_readers = nr;
2665 UNLOCK_MUTEX(rmutex);
2667 new_notls = (env->me_flags & MDB_NOTLS);
2668 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2673 do /* LY: Retry on a race, ITS#7970. */
2674 r->mr_txnid = ti->mti_txnid;
2675 while(r->mr_txnid != ti->mti_txnid);
2676 txn->mt_txnid = r->mr_txnid;
2677 txn->mt_u.reader = r;
2678 meta = env->me_metas[txn->mt_txnid & 1];
2682 /* Not yet touching txn == env->me_txn0, it may be active */
2684 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2686 txn->mt_txnid = ti->mti_txnid;
2687 meta = env->me_metas[txn->mt_txnid & 1];
2689 meta = mdb_env_pick_meta(env);
2690 txn->mt_txnid = meta->mm_txnid;
2694 if (txn->mt_txnid == mdb_debug_start)
2697 txn->mt_child = NULL;
2698 txn->mt_loose_pgs = NULL;
2699 txn->mt_loose_count = 0;
2700 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2701 txn->mt_u.dirty_list = env->me_dirty_list;
2702 txn->mt_u.dirty_list[0].mid = 0;
2703 txn->mt_free_pgs = env->me_free_pgs;
2704 txn->mt_free_pgs[0] = 0;
2705 txn->mt_spill_pgs = NULL;
2707 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2710 /* Copy the DB info and flags */
2711 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2713 /* Moved to here to avoid a data race in read TXNs */
2714 txn->mt_next_pgno = meta->mm_last_pg+1;
2716 txn->mt_flags = flags;
2719 txn->mt_numdbs = env->me_numdbs;
2720 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2721 x = env->me_dbflags[i];
2722 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2723 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2725 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2726 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2728 if (env->me_flags & MDB_FATAL_ERROR) {
2729 DPUTS("environment had fatal error, must shutdown!");
2731 } else if (env->me_maxpg < txn->mt_next_pgno) {
2732 rc = MDB_MAP_RESIZED;
2736 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2741 mdb_txn_renew(MDB_txn *txn)
2745 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2748 rc = mdb_txn_renew0(txn);
2749 if (rc == MDB_SUCCESS) {
2750 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2751 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2752 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2758 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2762 int rc, size, tsize;
2764 flags &= MDB_TXN_BEGIN_FLAGS;
2765 flags |= env->me_flags & MDB_WRITEMAP;
2767 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2771 /* Nested transactions: Max 1 child, write txns only, no writemap */
2772 flags |= parent->mt_flags;
2773 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2774 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2776 /* Child txns save MDB_pgstate and use own copy of cursors */
2777 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2778 size += tsize = sizeof(MDB_ntxn);
2779 } else if (flags & MDB_RDONLY) {
2780 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2781 size += tsize = sizeof(MDB_txn);
2783 /* Reuse preallocated write txn. However, do not touch it until
2784 * mdb_txn_renew0() succeeds, since it currently may be active.
2789 if ((txn = calloc(1, size)) == NULL) {
2790 DPRINTF(("calloc: %s", strerror(errno)));
2793 txn->mt_dbxs = env->me_dbxs; /* static */
2794 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2795 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2796 txn->mt_flags = flags;
2801 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2802 txn->mt_dbiseqs = parent->mt_dbiseqs;
2803 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2804 if (!txn->mt_u.dirty_list ||
2805 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2807 free(txn->mt_u.dirty_list);
2811 txn->mt_txnid = parent->mt_txnid;
2812 txn->mt_dirty_room = parent->mt_dirty_room;
2813 txn->mt_u.dirty_list[0].mid = 0;
2814 txn->mt_spill_pgs = NULL;
2815 txn->mt_next_pgno = parent->mt_next_pgno;
2816 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2817 parent->mt_child = txn;
2818 txn->mt_parent = parent;
2819 txn->mt_numdbs = parent->mt_numdbs;
2820 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2821 /* Copy parent's mt_dbflags, but clear DB_NEW */
2822 for (i=0; i<txn->mt_numdbs; i++)
2823 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2825 ntxn = (MDB_ntxn *)txn;
2826 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2827 if (env->me_pghead) {
2828 size = MDB_IDL_SIZEOF(env->me_pghead);
2829 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2831 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2836 rc = mdb_cursor_shadow(parent, txn);
2838 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2839 } else { /* MDB_RDONLY */
2840 txn->mt_dbiseqs = env->me_dbiseqs;
2842 rc = mdb_txn_renew0(txn);
2845 if (txn != env->me_txn0)
2848 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2850 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2851 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2852 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2859 mdb_txn_env(MDB_txn *txn)
2861 if(!txn) return NULL;
2866 mdb_txn_id(MDB_txn *txn)
2869 return txn->mt_txnid;
2872 /** Export or close DBI handles opened in this txn. */
2874 mdb_dbis_update(MDB_txn *txn, int keep)
2877 MDB_dbi n = txn->mt_numdbs;
2878 MDB_env *env = txn->mt_env;
2879 unsigned char *tdbflags = txn->mt_dbflags;
2881 for (i = n; --i >= CORE_DBS;) {
2882 if (tdbflags[i] & DB_NEW) {
2884 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2886 char *ptr = env->me_dbxs[i].md_name.mv_data;
2888 env->me_dbxs[i].md_name.mv_data = NULL;
2889 env->me_dbxs[i].md_name.mv_size = 0;
2890 env->me_dbflags[i] = 0;
2891 env->me_dbiseqs[i]++;
2897 if (keep && env->me_numdbs < n)
2901 /** End a transaction, except successful commit of a nested transaction.
2902 * May be called twice for readonly txns: First reset it, then abort.
2903 * @param[in] txn the transaction handle to end
2904 * @param[in] mode why and how to end the transaction
2907 mdb_txn_end(MDB_txn *txn, unsigned mode)
2909 MDB_env *env = txn->mt_env;
2911 static const char *const names[] = MDB_END_NAMES;
2914 /* Export or close DBI handles opened in this txn */
2915 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2917 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2918 names[mode & MDB_END_OPMASK],
2919 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2920 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2922 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2923 if (txn->mt_u.reader) {
2924 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2925 if (!(env->me_flags & MDB_NOTLS)) {
2926 txn->mt_u.reader = NULL; /* txn does not own reader */
2927 } else if (mode & MDB_END_SLOT) {
2928 txn->mt_u.reader->mr_pid = 0;
2929 txn->mt_u.reader = NULL;
2930 } /* else txn owns the slot until it does MDB_END_SLOT */
2932 txn->mt_numdbs = 0; /* prevent further DBI activity */
2933 txn->mt_flags |= MDB_TXN_FINISHED;
2935 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2936 pgno_t *pghead = env->me_pghead;
2938 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2939 mdb_cursors_close(txn, 0);
2940 if (!(env->me_flags & MDB_WRITEMAP)) {
2941 mdb_dlist_free(txn);
2945 txn->mt_flags = MDB_TXN_FINISHED;
2947 if (!txn->mt_parent) {
2948 mdb_midl_shrink(&txn->mt_free_pgs);
2949 env->me_free_pgs = txn->mt_free_pgs;
2951 env->me_pghead = NULL;
2955 mode = 0; /* txn == env->me_txn0, do not free() it */
2957 /* The writer mutex was locked in mdb_txn_begin. */
2959 UNLOCK_MUTEX(env->me_wmutex);
2961 txn->mt_parent->mt_child = NULL;
2962 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2963 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2964 mdb_midl_free(txn->mt_free_pgs);
2965 mdb_midl_free(txn->mt_spill_pgs);
2966 free(txn->mt_u.dirty_list);
2969 mdb_midl_free(pghead);
2972 if (mode & MDB_END_FREE)
2977 mdb_txn_reset(MDB_txn *txn)
2982 /* This call is only valid for read-only txns */
2983 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2986 mdb_txn_end(txn, MDB_END_RESET);
2990 mdb_txn_abort(MDB_txn *txn)
2996 mdb_txn_abort(txn->mt_child);
2998 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3001 /** Save the freelist as of this transaction to the freeDB.
3002 * This changes the freelist. Keep trying until it stabilizes.
3005 mdb_freelist_save(MDB_txn *txn)
3007 /* env->me_pghead[] can grow and shrink during this call.
3008 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3009 * Page numbers cannot disappear from txn->mt_free_pgs[].
3012 MDB_env *env = txn->mt_env;
3013 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3014 txnid_t pglast = 0, head_id = 0;
3015 pgno_t freecnt = 0, *free_pgs, *mop;
3016 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3018 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3020 if (env->me_pghead) {
3021 /* Make sure first page of freeDB is touched and on freelist */
3022 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3023 if (rc && rc != MDB_NOTFOUND)
3027 if (!env->me_pghead && txn->mt_loose_pgs) {
3028 /* Put loose page numbers in mt_free_pgs, since
3029 * we may be unable to return them to me_pghead.
3031 MDB_page *mp = txn->mt_loose_pgs;
3032 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3034 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3035 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3036 txn->mt_loose_pgs = NULL;
3037 txn->mt_loose_count = 0;
3040 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3041 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3042 ? SSIZE_MAX : maxfree_1pg;
3045 /* Come back here after each Put() in case freelist changed */
3050 /* If using records from freeDB which we have not yet
3051 * deleted, delete them and any we reserved for me_pghead.
3053 while (pglast < env->me_pglast) {
3054 rc = mdb_cursor_first(&mc, &key, NULL);
3057 pglast = head_id = *(txnid_t *)key.mv_data;
3058 total_room = head_room = 0;
3059 mdb_tassert(txn, pglast <= env->me_pglast);
3060 rc = mdb_cursor_del(&mc, 0);
3065 /* Save the IDL of pages freed by this txn, to a single record */
3066 if (freecnt < txn->mt_free_pgs[0]) {
3068 /* Make sure last page of freeDB is touched and on freelist */
3069 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3070 if (rc && rc != MDB_NOTFOUND)
3073 free_pgs = txn->mt_free_pgs;
3074 /* Write to last page of freeDB */
3075 key.mv_size = sizeof(txn->mt_txnid);
3076 key.mv_data = &txn->mt_txnid;
3078 freecnt = free_pgs[0];
3079 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3080 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3083 /* Retry if mt_free_pgs[] grew during the Put() */
3084 free_pgs = txn->mt_free_pgs;
3085 } while (freecnt < free_pgs[0]);
3086 mdb_midl_sort(free_pgs);
3087 memcpy(data.mv_data, free_pgs, data.mv_size);
3090 unsigned int i = free_pgs[0];
3091 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3092 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3094 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3100 mop = env->me_pghead;
3101 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3103 /* Reserve records for me_pghead[]. Split it if multi-page,
3104 * to avoid searching freeDB for a page range. Use keys in
3105 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3107 if (total_room >= mop_len) {
3108 if (total_room == mop_len || --more < 0)
3110 } else if (head_room >= maxfree_1pg && head_id > 1) {
3111 /* Keep current record (overflow page), add a new one */
3115 /* (Re)write {key = head_id, IDL length = head_room} */
3116 total_room -= head_room;
3117 head_room = mop_len - total_room;
3118 if (head_room > maxfree_1pg && head_id > 1) {
3119 /* Overflow multi-page for part of me_pghead */
3120 head_room /= head_id; /* amortize page sizes */
3121 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3122 } else if (head_room < 0) {
3123 /* Rare case, not bothering to delete this record */
3126 key.mv_size = sizeof(head_id);
3127 key.mv_data = &head_id;
3128 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3129 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3132 /* IDL is initially empty, zero out at least the length */
3133 pgs = (pgno_t *)data.mv_data;
3134 j = head_room > clean_limit ? head_room : 0;
3138 total_room += head_room;
3141 /* Return loose page numbers to me_pghead, though usually none are
3142 * left at this point. The pages themselves remain in dirty_list.
3144 if (txn->mt_loose_pgs) {
3145 MDB_page *mp = txn->mt_loose_pgs;
3146 unsigned count = txn->mt_loose_count;
3148 /* Room for loose pages + temp IDL with same */
3149 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3151 mop = env->me_pghead;
3152 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3153 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3154 loose[ ++count ] = mp->mp_pgno;
3156 mdb_midl_sort(loose);
3157 mdb_midl_xmerge(mop, loose);
3158 txn->mt_loose_pgs = NULL;
3159 txn->mt_loose_count = 0;
3163 /* Fill in the reserved me_pghead records */
3169 rc = mdb_cursor_first(&mc, &key, &data);
3170 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3171 txnid_t id = *(txnid_t *)key.mv_data;
3172 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3175 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3177 if (len > mop_len) {
3179 data.mv_size = (len + 1) * sizeof(MDB_ID);
3181 data.mv_data = mop -= len;
3184 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3186 if (rc || !(mop_len -= len))
3193 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3194 * @param[in] txn the transaction that's being committed
3195 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3196 * @return 0 on success, non-zero on failure.
3199 mdb_page_flush(MDB_txn *txn, int keep)
3201 MDB_env *env = txn->mt_env;
3202 MDB_ID2L dl = txn->mt_u.dirty_list;
3203 unsigned psize = env->me_psize, j;
3204 int i, pagecount = dl[0].mid, rc;
3205 size_t size = 0, pos = 0;
3207 MDB_page *dp = NULL;
3211 struct iovec iov[MDB_COMMIT_PAGES];
3212 ssize_t wpos = 0, wsize = 0, wres;
3213 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3219 if (env->me_flags & MDB_WRITEMAP) {
3220 /* Clear dirty flags */
3221 while (++i <= pagecount) {
3223 /* Don't flush this page yet */
3224 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3225 dp->mp_flags &= ~P_KEEP;
3229 dp->mp_flags &= ~P_DIRTY;
3234 /* Write the pages */
3236 if (++i <= pagecount) {
3238 /* Don't flush this page yet */
3239 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3240 dp->mp_flags &= ~P_KEEP;
3245 /* clear dirty flag */
3246 dp->mp_flags &= ~P_DIRTY;
3249 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3254 /* Windows actually supports scatter/gather I/O, but only on
3255 * unbuffered file handles. Since we're relying on the OS page
3256 * cache for all our data, that's self-defeating. So we just
3257 * write pages one at a time. We use the ov structure to set
3258 * the write offset, to at least save the overhead of a Seek
3261 DPRINTF(("committing page %"Z"u", pgno));
3262 memset(&ov, 0, sizeof(ov));
3263 ov.Offset = pos & 0xffffffff;
3264 ov.OffsetHigh = pos >> 16 >> 16;
3265 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3267 DPRINTF(("WriteFile: %d", rc));
3271 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3272 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3275 /* Write previous page(s) */
3276 #ifdef MDB_USE_PWRITEV
3277 wres = pwritev(env->me_fd, iov, n, wpos);
3280 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3283 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3287 DPRINTF(("lseek: %s", strerror(rc)));
3290 wres = writev(env->me_fd, iov, n);
3293 if (wres != wsize) {
3298 DPRINTF(("Write error: %s", strerror(rc)));
3300 rc = EIO; /* TODO: Use which error code? */
3301 DPUTS("short write, filesystem full?");
3312 DPRINTF(("committing page %"Z"u", pgno));
3313 next_pos = pos + size;
3314 iov[n].iov_len = size;
3315 iov[n].iov_base = (char *)dp;
3321 /* MIPS has cache coherency issues, this is a no-op everywhere else
3322 * Note: for any size >= on-chip cache size, entire on-chip cache is
3325 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3327 for (i = keep; ++i <= pagecount; ) {
3329 /* This is a page we skipped above */
3332 dl[j].mid = dp->mp_pgno;
3335 mdb_dpage_free(env, dp);
3340 txn->mt_dirty_room += i - j;
3346 mdb_txn_commit(MDB_txn *txn)
3349 unsigned int i, end_mode;
3355 /* mdb_txn_end() mode for a commit which writes nothing */
3356 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3358 if (txn->mt_child) {
3359 rc = mdb_txn_commit(txn->mt_child);
3366 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3370 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3371 DPUTS("txn has failed/finished, can't commit");
3373 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3378 if (txn->mt_parent) {
3379 MDB_txn *parent = txn->mt_parent;
3383 unsigned x, y, len, ps_len;
3385 /* Append our free list to parent's */
3386 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3389 mdb_midl_free(txn->mt_free_pgs);
3390 /* Failures after this must either undo the changes
3391 * to the parent or set MDB_TXN_ERROR in the parent.
3394 parent->mt_next_pgno = txn->mt_next_pgno;
3395 parent->mt_flags = txn->mt_flags;
3397 /* Merge our cursors into parent's and close them */
3398 mdb_cursors_close(txn, 1);
3400 /* Update parent's DB table. */
3401 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3402 parent->mt_numdbs = txn->mt_numdbs;
3403 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3404 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3405 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3406 /* preserve parent's DB_NEW status */
3407 x = parent->mt_dbflags[i] & DB_NEW;
3408 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3411 dst = parent->mt_u.dirty_list;
3412 src = txn->mt_u.dirty_list;
3413 /* Remove anything in our dirty list from parent's spill list */
3414 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3416 pspill[0] = (pgno_t)-1;
3417 /* Mark our dirty pages as deleted in parent spill list */
3418 for (i=0, len=src[0].mid; ++i <= len; ) {
3419 MDB_ID pn = src[i].mid << 1;
3420 while (pn > pspill[x])
3422 if (pn == pspill[x]) {
3427 /* Squash deleted pagenums if we deleted any */
3428 for (x=y; ++x <= ps_len; )
3429 if (!(pspill[x] & 1))
3430 pspill[++y] = pspill[x];
3434 /* Find len = length of merging our dirty list with parent's */
3436 dst[0].mid = 0; /* simplify loops */
3437 if (parent->mt_parent) {
3438 len = x + src[0].mid;
3439 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3440 for (i = x; y && i; y--) {
3441 pgno_t yp = src[y].mid;
3442 while (yp < dst[i].mid)
3444 if (yp == dst[i].mid) {
3449 } else { /* Simplify the above for single-ancestor case */
3450 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3452 /* Merge our dirty list with parent's */
3454 for (i = len; y; dst[i--] = src[y--]) {
3455 pgno_t yp = src[y].mid;
3456 while (yp < dst[x].mid)
3457 dst[i--] = dst[x--];
3458 if (yp == dst[x].mid)
3459 free(dst[x--].mptr);
3461 mdb_tassert(txn, i == x);
3463 free(txn->mt_u.dirty_list);
3464 parent->mt_dirty_room = txn->mt_dirty_room;
3465 if (txn->mt_spill_pgs) {
3466 if (parent->mt_spill_pgs) {
3467 /* TODO: Prevent failure here, so parent does not fail */
3468 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3470 parent->mt_flags |= MDB_TXN_ERROR;
3471 mdb_midl_free(txn->mt_spill_pgs);
3472 mdb_midl_sort(parent->mt_spill_pgs);
3474 parent->mt_spill_pgs = txn->mt_spill_pgs;
3478 /* Append our loose page list to parent's */
3479 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3481 *lp = txn->mt_loose_pgs;
3482 parent->mt_loose_count += txn->mt_loose_count;
3484 parent->mt_child = NULL;
3485 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3490 if (txn != env->me_txn) {
3491 DPUTS("attempt to commit unknown transaction");
3496 mdb_cursors_close(txn, 0);
3498 if (!txn->mt_u.dirty_list[0].mid &&
3499 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3502 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3503 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3505 /* Update DB root pointers */
3506 if (txn->mt_numdbs > CORE_DBS) {
3510 data.mv_size = sizeof(MDB_db);
3512 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3513 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3514 if (txn->mt_dbflags[i] & DB_DIRTY) {
3515 if (TXN_DBI_CHANGED(txn, i)) {
3519 data.mv_data = &txn->mt_dbs[i];
3520 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3528 rc = mdb_freelist_save(txn);
3532 mdb_midl_free(env->me_pghead);
3533 env->me_pghead = NULL;
3534 mdb_midl_shrink(&txn->mt_free_pgs);
3540 if ((rc = mdb_page_flush(txn, 0)) ||
3541 (rc = mdb_env_sync(env, 0)) ||
3542 (rc = mdb_env_write_meta(txn)))
3544 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3547 mdb_txn_end(txn, end_mode);
3555 /** Read the environment parameters of a DB environment before
3556 * mapping it into memory.
3557 * @param[in] env the environment handle
3558 * @param[out] meta address of where to store the meta information
3559 * @return 0 on success, non-zero on failure.
3562 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3568 enum { Size = sizeof(pbuf) };
3570 /* We don't know the page size yet, so use a minimum value.
3571 * Read both meta pages so we can use the latest one.
3574 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3578 memset(&ov, 0, sizeof(ov));
3580 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3581 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3584 rc = pread(env->me_fd, &pbuf, Size, off);
3587 if (rc == 0 && off == 0)
3589 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3590 DPRINTF(("read: %s", mdb_strerror(rc)));
3594 p = (MDB_page *)&pbuf;
3596 if (!F_ISSET(p->mp_flags, P_META)) {
3597 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3602 if (m->mm_magic != MDB_MAGIC) {
3603 DPUTS("meta has invalid magic");
3607 if (m->mm_version != MDB_DATA_VERSION) {
3608 DPRINTF(("database is version %u, expected version %u",
3609 m->mm_version, MDB_DATA_VERSION));
3610 return MDB_VERSION_MISMATCH;
3613 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3619 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3621 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3623 meta->mm_magic = MDB_MAGIC;
3624 meta->mm_version = MDB_DATA_VERSION;
3625 meta->mm_mapsize = env->me_mapsize;
3626 meta->mm_psize = env->me_psize;
3627 meta->mm_last_pg = NUM_METAS-1;
3628 meta->mm_flags = env->me_flags & 0xffff;
3629 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3630 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3631 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3634 /** Write the environment parameters of a freshly created DB environment.
3635 * @param[in] env the environment handle
3636 * @param[in] meta the #MDB_meta to write
3637 * @return 0 on success, non-zero on failure.
3640 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3648 memset(&ov, 0, sizeof(ov));
3649 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3651 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3654 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3655 len = pwrite(fd, ptr, size, pos); \
3656 if (len == -1 && ErrCode() == EINTR) continue; \
3657 rc = (len >= 0); break; } while(1)
3660 DPUTS("writing new meta page");
3662 psize = env->me_psize;
3664 p = calloc(NUM_METAS, psize);
3669 p->mp_flags = P_META;
3670 *(MDB_meta *)METADATA(p) = *meta;
3672 q = (MDB_page *)((char *)p + psize);
3674 q->mp_flags = P_META;
3675 *(MDB_meta *)METADATA(q) = *meta;
3677 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3680 else if ((unsigned) len == psize * NUM_METAS)
3688 /** Update the environment info to commit a transaction.
3689 * @param[in] txn the transaction that's being committed
3690 * @return 0 on success, non-zero on failure.
3693 mdb_env_write_meta(MDB_txn *txn)
3696 MDB_meta meta, metab, *mp;
3700 int rc, len, toggle;
3709 toggle = txn->mt_txnid & 1;
3710 DPRINTF(("writing meta page %d for root page %"Z"u",
3711 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3714 flags = env->me_flags;
3715 mp = env->me_metas[toggle];
3716 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3717 /* Persist any increases of mapsize config */
3718 if (mapsize < env->me_mapsize)
3719 mapsize = env->me_mapsize;
3721 if (flags & MDB_WRITEMAP) {
3722 mp->mm_mapsize = mapsize;
3723 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3724 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3725 mp->mm_last_pg = txn->mt_next_pgno - 1;
3726 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3727 !(defined(__i386__) || defined(__x86_64__))
3728 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3729 __sync_synchronize();
3731 mp->mm_txnid = txn->mt_txnid;
3732 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3733 unsigned meta_size = env->me_psize;
3734 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3735 ptr = (char *)mp - PAGEHDRSZ;
3736 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3737 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3741 if (MDB_MSYNC(ptr, meta_size, rc)) {
3748 metab.mm_txnid = mp->mm_txnid;
3749 metab.mm_last_pg = mp->mm_last_pg;
3751 meta.mm_mapsize = mapsize;
3752 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3753 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3754 meta.mm_last_pg = txn->mt_next_pgno - 1;
3755 meta.mm_txnid = txn->mt_txnid;
3757 off = offsetof(MDB_meta, mm_mapsize);
3758 ptr = (char *)&meta + off;
3759 len = sizeof(MDB_meta) - off;
3760 off += (char *)mp - env->me_map;
3762 /* Write to the SYNC fd */
3763 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3766 memset(&ov, 0, sizeof(ov));
3768 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3773 rc = pwrite(mfd, ptr, len, off);
3776 rc = rc < 0 ? ErrCode() : EIO;
3781 DPUTS("write failed, disk error?");
3782 /* On a failure, the pagecache still contains the new data.
3783 * Write some old data back, to prevent it from being used.
3784 * Use the non-SYNC fd; we know it will fail anyway.
3786 meta.mm_last_pg = metab.mm_last_pg;
3787 meta.mm_txnid = metab.mm_txnid;
3789 memset(&ov, 0, sizeof(ov));
3791 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3793 r2 = pwrite(env->me_fd, ptr, len, off);
3794 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3797 env->me_flags |= MDB_FATAL_ERROR;
3800 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3801 CACHEFLUSH(env->me_map + off, len, DCACHE);
3803 /* Memory ordering issues are irrelevant; since the entire writer
3804 * is wrapped by wmutex, all of these changes will become visible
3805 * after the wmutex is unlocked. Since the DB is multi-version,
3806 * readers will get consistent data regardless of how fresh or
3807 * how stale their view of these values is.
3810 env->me_txns->mti_txnid = txn->mt_txnid;
3815 /** Check both meta pages to see which one is newer.
3816 * @param[in] env the environment handle
3817 * @return newest #MDB_meta.
3820 mdb_env_pick_meta(const MDB_env *env)
3822 MDB_meta *const *metas = env->me_metas;
3823 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3827 mdb_env_create(MDB_env **env)
3831 e = calloc(1, sizeof(MDB_env));
3835 e->me_maxreaders = DEFAULT_READERS;
3836 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3837 e->me_fd = INVALID_HANDLE_VALUE;
3838 e->me_lfd = INVALID_HANDLE_VALUE;
3839 e->me_mfd = INVALID_HANDLE_VALUE;
3840 #ifdef MDB_USE_POSIX_SEM
3841 e->me_rmutex = SEM_FAILED;
3842 e->me_wmutex = SEM_FAILED;
3844 e->me_pid = getpid();
3845 GET_PAGESIZE(e->me_os_psize);
3846 VGMEMP_CREATE(e,0,0);
3852 mdb_env_map(MDB_env *env, void *addr)
3855 unsigned int flags = env->me_flags;
3859 LONG sizelo, sizehi;
3862 if (flags & MDB_RDONLY) {
3863 /* Don't set explicit map size, use whatever exists */
3868 msize = env->me_mapsize;
3869 sizelo = msize & 0xffffffff;
3870 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3872 /* Windows won't create mappings for zero length files.
3873 * and won't map more than the file size.
3874 * Just set the maxsize right now.
3876 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3877 || !SetEndOfFile(env->me_fd)
3878 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3882 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3883 PAGE_READWRITE : PAGE_READONLY,
3884 sizehi, sizelo, NULL);
3887 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3888 FILE_MAP_WRITE : FILE_MAP_READ,
3890 rc = env->me_map ? 0 : ErrCode();
3895 int prot = PROT_READ;
3896 if (flags & MDB_WRITEMAP) {
3898 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3901 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3903 if (env->me_map == MAP_FAILED) {
3908 if (flags & MDB_NORDAHEAD) {
3909 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3911 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3913 #ifdef POSIX_MADV_RANDOM
3914 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3915 #endif /* POSIX_MADV_RANDOM */
3916 #endif /* MADV_RANDOM */
3920 /* Can happen because the address argument to mmap() is just a
3921 * hint. mmap() can pick another, e.g. if the range is in use.
3922 * The MAP_FIXED flag would prevent that, but then mmap could
3923 * instead unmap existing pages to make room for the new map.
3925 if (addr && env->me_map != addr)
3926 return EBUSY; /* TODO: Make a new MDB_* error code? */
3928 p = (MDB_page *)env->me_map;
3929 env->me_metas[0] = METADATA(p);
3930 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3936 mdb_env_set_mapsize(MDB_env *env, size_t size)
3938 /* If env is already open, caller is responsible for making
3939 * sure there are no active txns.
3947 meta = mdb_env_pick_meta(env);
3949 size = meta->mm_mapsize;
3951 /* Silently round up to minimum if the size is too small */
3952 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3956 munmap(env->me_map, env->me_mapsize);
3957 env->me_mapsize = size;
3958 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3959 rc = mdb_env_map(env, old);
3963 env->me_mapsize = size;
3965 env->me_maxpg = env->me_mapsize / env->me_psize;
3970 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3974 env->me_maxdbs = dbs + CORE_DBS;
3979 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3981 if (env->me_map || readers < 1)
3983 env->me_maxreaders = readers;
3988 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3990 if (!env || !readers)
3992 *readers = env->me_maxreaders;
3997 mdb_fsize(HANDLE fd, size_t *size)
4000 LARGE_INTEGER fsize;
4002 if (!GetFileSizeEx(fd, &fsize))
4005 *size = fsize.QuadPart;
4017 #ifdef BROKEN_FDATASYNC
4018 #include <sys/utsname.h>
4019 #include <sys/vfs.h>
4022 /** Further setup required for opening an LMDB environment
4025 mdb_env_open2(MDB_env *env)
4027 unsigned int flags = env->me_flags;
4028 int i, newenv = 0, rc;
4032 /* See if we should use QueryLimited */
4034 if ((rc & 0xff) > 5)
4035 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4037 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4040 #ifdef BROKEN_FDATASYNC
4041 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4042 * https://lkml.org/lkml/2012/9/3/83
4043 * Kernels after 3.6-rc6 are known good.
4044 * https://lkml.org/lkml/2012/9/10/556
4045 * See if the DB is on ext3/ext4, then check for new enough kernel
4046 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4051 fstatfs(env->me_fd, &st);
4052 while (st.f_type == 0xEF53) {
4056 if (uts.release[0] < '3') {
4057 if (!strncmp(uts.release, "2.6.32.", 7)) {
4058 i = atoi(uts.release+7);
4060 break; /* 2.6.32.60 and newer is OK */
4061 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4062 i = atoi(uts.release+7);
4064 break; /* 2.6.34.15 and newer is OK */
4066 } else if (uts.release[0] == '3') {
4067 i = atoi(uts.release+2);
4069 break; /* 3.6 and newer is OK */
4071 i = atoi(uts.release+4);
4073 break; /* 3.5.4 and newer is OK */
4074 } else if (i == 2) {
4075 i = atoi(uts.release+4);
4077 break; /* 3.2.30 and newer is OK */
4079 } else { /* 4.x and newer is OK */
4082 env->me_flags |= MDB_FSYNCONLY;
4088 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4091 DPUTS("new mdbenv");
4093 env->me_psize = env->me_os_psize;
4094 if (env->me_psize > MAX_PAGESIZE)
4095 env->me_psize = MAX_PAGESIZE;
4096 memset(&meta, 0, sizeof(meta));
4097 mdb_env_init_meta0(env, &meta);
4098 meta.mm_mapsize = DEFAULT_MAPSIZE;
4100 env->me_psize = meta.mm_psize;
4103 /* Was a mapsize configured? */
4104 if (!env->me_mapsize) {
4105 env->me_mapsize = meta.mm_mapsize;
4108 /* Make sure mapsize >= committed data size. Even when using
4109 * mm_mapsize, which could be broken in old files (ITS#7789).
4111 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4112 if (env->me_mapsize < minsize)
4113 env->me_mapsize = minsize;
4115 meta.mm_mapsize = env->me_mapsize;
4117 if (newenv && !(flags & MDB_FIXEDMAP)) {
4118 /* mdb_env_map() may grow the datafile. Write the metapages
4119 * first, so the file will be valid if initialization fails.
4120 * Except with FIXEDMAP, since we do not yet know mm_address.
4121 * We could fill in mm_address later, but then a different
4122 * program might end up doing that - one with a memory layout
4123 * and map address which does not suit the main program.
4125 rc = mdb_env_init_meta(env, &meta);
4131 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4136 if (flags & MDB_FIXEDMAP)
4137 meta.mm_address = env->me_map;
4138 i = mdb_env_init_meta(env, &meta);
4139 if (i != MDB_SUCCESS) {
4144 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4145 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4147 #if !(MDB_MAXKEYSIZE)
4148 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4150 env->me_maxpg = env->me_mapsize / env->me_psize;
4154 MDB_meta *meta = mdb_env_pick_meta(env);
4155 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4157 DPRINTF(("opened database version %u, pagesize %u",
4158 meta->mm_version, env->me_psize));
4159 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4160 DPRINTF(("depth: %u", db->md_depth));
4161 DPRINTF(("entries: %"Z"u", db->md_entries));
4162 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4163 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4164 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4165 DPRINTF(("root: %"Z"u", db->md_root));
4173 /** Release a reader thread's slot in the reader lock table.
4174 * This function is called automatically when a thread exits.
4175 * @param[in] ptr This points to the slot in the reader lock table.
4178 mdb_env_reader_dest(void *ptr)
4180 MDB_reader *reader = ptr;
4186 /** Junk for arranging thread-specific callbacks on Windows. This is
4187 * necessarily platform and compiler-specific. Windows supports up
4188 * to 1088 keys. Let's assume nobody opens more than 64 environments
4189 * in a single process, for now. They can override this if needed.
4191 #ifndef MAX_TLS_KEYS
4192 #define MAX_TLS_KEYS 64
4194 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4195 static int mdb_tls_nkeys;
4197 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4201 case DLL_PROCESS_ATTACH: break;
4202 case DLL_THREAD_ATTACH: break;
4203 case DLL_THREAD_DETACH:
4204 for (i=0; i<mdb_tls_nkeys; i++) {
4205 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4207 mdb_env_reader_dest(r);
4211 case DLL_PROCESS_DETACH: break;
4216 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4218 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4222 /* Force some symbol references.
4223 * _tls_used forces the linker to create the TLS directory if not already done
4224 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4226 #pragma comment(linker, "/INCLUDE:_tls_used")
4227 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4228 #pragma const_seg(".CRT$XLB")
4229 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4230 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4233 #pragma comment(linker, "/INCLUDE:__tls_used")
4234 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4235 #pragma data_seg(".CRT$XLB")
4236 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4238 #endif /* WIN 32/64 */
4239 #endif /* !__GNUC__ */
4242 /** Downgrade the exclusive lock on the region back to shared */
4244 mdb_env_share_locks(MDB_env *env, int *excl)
4247 MDB_meta *meta = mdb_env_pick_meta(env);
4249 env->me_txns->mti_txnid = meta->mm_txnid;
4254 /* First acquire a shared lock. The Unlock will
4255 * then release the existing exclusive lock.
4257 memset(&ov, 0, sizeof(ov));
4258 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4261 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4267 struct flock lock_info;
4268 /* The shared lock replaces the existing lock */
4269 memset((void *)&lock_info, 0, sizeof(lock_info));
4270 lock_info.l_type = F_RDLCK;
4271 lock_info.l_whence = SEEK_SET;
4272 lock_info.l_start = 0;
4273 lock_info.l_len = 1;
4274 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4275 (rc = ErrCode()) == EINTR) ;
4276 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4283 /** Try to get exclusive lock, otherwise shared.
4284 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4287 mdb_env_excl_lock(MDB_env *env, int *excl)
4291 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4295 memset(&ov, 0, sizeof(ov));
4296 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4303 struct flock lock_info;
4304 memset((void *)&lock_info, 0, sizeof(lock_info));
4305 lock_info.l_type = F_WRLCK;
4306 lock_info.l_whence = SEEK_SET;
4307 lock_info.l_start = 0;
4308 lock_info.l_len = 1;
4309 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4310 (rc = ErrCode()) == EINTR) ;
4314 # ifndef MDB_USE_POSIX_MUTEX
4315 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4318 lock_info.l_type = F_RDLCK;
4319 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4320 (rc = ErrCode()) == EINTR) ;
4330 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4332 * @(#) $Revision: 5.1 $
4333 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4334 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4336 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4340 * Please do not copyright this code. This code is in the public domain.
4342 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4343 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4344 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4345 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4346 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4347 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4348 * PERFORMANCE OF THIS SOFTWARE.
4351 * chongo <Landon Curt Noll> /\oo/\
4352 * http://www.isthe.com/chongo/
4354 * Share and Enjoy! :-)
4357 typedef unsigned long long mdb_hash_t;
4358 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4360 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4361 * @param[in] val value to hash
4362 * @param[in] hval initial value for hash
4363 * @return 64 bit hash
4365 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4366 * hval arg on the first call.
4369 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4371 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4372 unsigned char *end = s + val->mv_size;
4374 * FNV-1a hash each octet of the string
4377 /* xor the bottom with the current octet */
4378 hval ^= (mdb_hash_t)*s++;
4380 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4381 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4382 (hval << 7) + (hval << 8) + (hval << 40);
4384 /* return our new hash value */
4388 /** Hash the string and output the encoded hash.
4389 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4390 * very short name limits. We don't care about the encoding being reversible,
4391 * we just want to preserve as many bits of the input as possible in a
4392 * small printable string.
4393 * @param[in] str string to hash
4394 * @param[out] encbuf an array of 11 chars to hold the hash
4396 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4399 mdb_pack85(unsigned long l, char *out)
4403 for (i=0; i<5; i++) {
4404 *out++ = mdb_a85[l % 85];
4410 mdb_hash_enc(MDB_val *val, char *encbuf)
4412 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4414 mdb_pack85(h, encbuf);
4415 mdb_pack85(h>>32, encbuf+5);
4420 /** Open and/or initialize the lock region for the environment.
4421 * @param[in] env The LMDB environment.
4422 * @param[in] lpath The pathname of the file used for the lock region.
4423 * @param[in] mode The Unix permissions for the file, if we create it.
4424 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4425 * @return 0 on success, non-zero on failure.
4428 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4431 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4433 # define MDB_ERRCODE_ROFS EROFS
4434 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4435 # define MDB_CLOEXEC O_CLOEXEC
4438 # define MDB_CLOEXEC 0
4446 utf8_to_utf16(lpath, -1, &wlpath, NULL);
4447 env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
4448 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4449 FILE_ATTRIBUTE_NORMAL, NULL);
4452 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4454 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4456 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4461 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4462 /* Lose record locks when exec*() */
4463 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4464 fcntl(env->me_lfd, F_SETFD, fdflags);
4467 if (!(env->me_flags & MDB_NOTLS)) {
4468 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4471 env->me_flags |= MDB_ENV_TXKEY;
4473 /* Windows TLS callbacks need help finding their TLS info. */
4474 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4478 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4482 /* Try to get exclusive lock. If we succeed, then
4483 * nobody is using the lock region and we should initialize it.
4485 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4488 size = GetFileSize(env->me_lfd, NULL);
4490 size = lseek(env->me_lfd, 0, SEEK_END);
4491 if (size == -1) goto fail_errno;
4493 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4494 if (size < rsize && *excl > 0) {
4496 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4497 || !SetEndOfFile(env->me_lfd))
4500 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4504 size = rsize - sizeof(MDB_txninfo);
4505 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4510 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4512 if (!mh) goto fail_errno;
4513 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4515 if (!env->me_txns) goto fail_errno;
4517 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4519 if (m == MAP_FAILED) goto fail_errno;
4525 BY_HANDLE_FILE_INFORMATION stbuf;
4534 if (!mdb_sec_inited) {
4535 InitializeSecurityDescriptor(&mdb_null_sd,
4536 SECURITY_DESCRIPTOR_REVISION);
4537 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4538 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4539 mdb_all_sa.bInheritHandle = FALSE;
4540 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4543 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4544 idbuf.volume = stbuf.dwVolumeSerialNumber;
4545 idbuf.nhigh = stbuf.nFileIndexHigh;
4546 idbuf.nlow = stbuf.nFileIndexLow;
4547 val.mv_data = &idbuf;
4548 val.mv_size = sizeof(idbuf);
4549 mdb_hash_enc(&val, encbuf);
4550 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4551 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4552 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4553 if (!env->me_rmutex) goto fail_errno;
4554 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4555 if (!env->me_wmutex) goto fail_errno;
4556 #elif defined(MDB_USE_POSIX_SEM)
4565 #if defined(__NetBSD__)
4566 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4568 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4569 idbuf.dev = stbuf.st_dev;
4570 idbuf.ino = stbuf.st_ino;
4571 val.mv_data = &idbuf;
4572 val.mv_size = sizeof(idbuf);
4573 mdb_hash_enc(&val, encbuf);
4574 #ifdef MDB_SHORT_SEMNAMES
4575 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4577 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4578 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4579 /* Clean up after a previous run, if needed: Try to
4580 * remove both semaphores before doing anything else.
4582 sem_unlink(env->me_txns->mti_rmname);
4583 sem_unlink(env->me_txns->mti_wmname);
4584 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4585 O_CREAT|O_EXCL, mode, 1);
4586 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4587 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4588 O_CREAT|O_EXCL, mode, 1);
4589 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4590 #else /* MDB_USE_POSIX_MUTEX: */
4591 pthread_mutexattr_t mattr;
4593 if ((rc = pthread_mutexattr_init(&mattr))
4594 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4595 #ifdef MDB_ROBUST_SUPPORTED
4596 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4598 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4599 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4601 pthread_mutexattr_destroy(&mattr);
4602 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4604 env->me_txns->mti_magic = MDB_MAGIC;
4605 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4606 env->me_txns->mti_txnid = 0;
4607 env->me_txns->mti_numreaders = 0;
4610 if (env->me_txns->mti_magic != MDB_MAGIC) {
4611 DPUTS("lock region has invalid magic");
4615 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4616 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4617 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4618 rc = MDB_VERSION_MISMATCH;
4622 if (rc && rc != EACCES && rc != EAGAIN) {
4626 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4627 if (!env->me_rmutex) goto fail_errno;
4628 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4629 if (!env->me_wmutex) goto fail_errno;
4630 #elif defined(MDB_USE_POSIX_SEM)
4631 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4632 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4633 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4634 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4645 /** The name of the lock file in the DB environment */
4646 #define LOCKNAME "/lock.mdb"
4647 /** The name of the data file in the DB environment */
4648 #define DATANAME "/data.mdb"
4649 /** The suffix of the lock file when no subdir is used */
4650 #define LOCKSUFF "-lock"
4651 /** Only a subset of the @ref mdb_env flags can be changed
4652 * at runtime. Changing other flags requires closing the
4653 * environment and re-opening it with the new flags.
4655 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4656 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4657 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4659 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4660 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4664 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4666 int oflags, rc, len, excl = -1;
4667 char *lpath, *dpath;
4672 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4676 if (flags & MDB_NOSUBDIR) {
4677 rc = len + sizeof(LOCKSUFF) + len + 1;
4679 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4684 if (flags & MDB_NOSUBDIR) {
4685 dpath = lpath + len + sizeof(LOCKSUFF);
4686 sprintf(lpath, "%s" LOCKSUFF, path);
4687 strcpy(dpath, path);
4689 dpath = lpath + len + sizeof(LOCKNAME);
4690 sprintf(lpath, "%s" LOCKNAME, path);
4691 sprintf(dpath, "%s" DATANAME, path);
4695 flags |= env->me_flags;
4696 if (flags & MDB_RDONLY) {
4697 /* silently ignore WRITEMAP when we're only getting read access */
4698 flags &= ~MDB_WRITEMAP;
4700 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4701 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4704 env->me_flags = flags |= MDB_ENV_ACTIVE;
4708 env->me_path = strdup(path);
4709 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4710 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4711 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4712 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4716 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4718 /* For RDONLY, get lockfile after we know datafile exists */
4719 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4720 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4726 if (F_ISSET(flags, MDB_RDONLY)) {
4727 oflags = GENERIC_READ;
4728 len = OPEN_EXISTING;
4730 oflags = GENERIC_READ|GENERIC_WRITE;
4733 mode = FILE_ATTRIBUTE_NORMAL;
4734 utf8_to_utf16(dpath, -1, &wpath, NULL);
4735 env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4736 NULL, len, mode, NULL);
4739 if (F_ISSET(flags, MDB_RDONLY))
4742 oflags = O_RDWR | O_CREAT;
4744 env->me_fd = open(dpath, oflags, mode);
4746 if (env->me_fd == INVALID_HANDLE_VALUE) {
4751 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4752 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4757 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4758 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4759 env->me_mfd = env->me_fd;
4761 /* Synchronous fd for meta writes. Needed even with
4762 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4765 len = OPEN_EXISTING;
4766 utf8_to_utf16(dpath, -1, &wpath, NULL);
4767 env->me_mfd = CreateFileW(wpath, oflags,
4768 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4769 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4773 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4775 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4780 DPRINTF(("opened dbenv %p", (void *) env));
4782 rc = mdb_env_share_locks(env, &excl);
4786 if (!(flags & MDB_RDONLY)) {
4788 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4789 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4790 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4791 (txn = calloc(1, size)))
4793 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4794 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4795 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4796 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4798 txn->mt_dbxs = env->me_dbxs;
4799 txn->mt_flags = MDB_TXN_FINISHED;
4809 mdb_env_close0(env, excl);
4815 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4817 mdb_env_close0(MDB_env *env, int excl)
4821 if (!(env->me_flags & MDB_ENV_ACTIVE))
4824 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4826 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4827 free(env->me_dbxs[i].md_name.mv_data);
4832 free(env->me_dbiseqs);
4833 free(env->me_dbflags);
4835 free(env->me_dirty_list);
4837 mdb_midl_free(env->me_free_pgs);
4839 if (env->me_flags & MDB_ENV_TXKEY) {
4840 pthread_key_delete(env->me_txkey);
4842 /* Delete our key from the global list */
4843 for (i=0; i<mdb_tls_nkeys; i++)
4844 if (mdb_tls_keys[i] == env->me_txkey) {
4845 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4853 munmap(env->me_map, env->me_mapsize);
4855 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4856 (void) close(env->me_mfd);
4857 if (env->me_fd != INVALID_HANDLE_VALUE)
4858 (void) close(env->me_fd);
4860 MDB_PID_T pid = env->me_pid;
4861 /* Clearing readers is done in this function because
4862 * me_txkey with its destructor must be disabled first.
4864 * We skip the the reader mutex, so we touch only
4865 * data owned by this process (me_close_readers and
4866 * our readers), and clear each reader atomically.
4868 for (i = env->me_close_readers; --i >= 0; )
4869 if (env->me_txns->mti_readers[i].mr_pid == pid)
4870 env->me_txns->mti_readers[i].mr_pid = 0;
4872 if (env->me_rmutex) {
4873 CloseHandle(env->me_rmutex);
4874 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4876 /* Windows automatically destroys the mutexes when
4877 * the last handle closes.
4879 #elif defined(MDB_USE_POSIX_SEM)
4880 if (env->me_rmutex != SEM_FAILED) {
4881 sem_close(env->me_rmutex);
4882 if (env->me_wmutex != SEM_FAILED)
4883 sem_close(env->me_wmutex);
4884 /* If we have the filelock: If we are the
4885 * only remaining user, clean up semaphores.
4888 mdb_env_excl_lock(env, &excl);
4890 sem_unlink(env->me_txns->mti_rmname);
4891 sem_unlink(env->me_txns->mti_wmname);
4895 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4897 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4900 /* Unlock the lockfile. Windows would have unlocked it
4901 * after closing anyway, but not necessarily at once.
4903 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4906 (void) close(env->me_lfd);
4909 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4913 mdb_env_close(MDB_env *env)
4920 VGMEMP_DESTROY(env);
4921 while ((dp = env->me_dpages) != NULL) {
4922 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4923 env->me_dpages = dp->mp_next;
4927 mdb_env_close0(env, 0);
4931 /** Compare two items pointing at aligned size_t's */
4933 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4935 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4936 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4939 /** Compare two items pointing at aligned unsigned int's.
4941 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4942 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4945 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4947 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4948 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4951 /** Compare two items pointing at unsigned ints of unknown alignment.
4952 * Nodes and keys are guaranteed to be 2-byte aligned.
4955 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4957 #if BYTE_ORDER == LITTLE_ENDIAN
4958 unsigned short *u, *c;
4961 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4962 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4965 } while(!x && u > (unsigned short *)a->mv_data);
4968 unsigned short *u, *c, *end;
4971 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4972 u = (unsigned short *)a->mv_data;
4973 c = (unsigned short *)b->mv_data;
4976 } while(!x && u < end);
4981 /** Compare two items lexically */
4983 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4990 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4996 diff = memcmp(a->mv_data, b->mv_data, len);
4997 return diff ? diff : len_diff<0 ? -1 : len_diff;
5000 /** Compare two items in reverse byte order */
5002 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5004 const unsigned char *p1, *p2, *p1_lim;
5008 p1_lim = (const unsigned char *)a->mv_data;
5009 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5010 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5012 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5018 while (p1 > p1_lim) {
5019 diff = *--p1 - *--p2;
5023 return len_diff<0 ? -1 : len_diff;
5026 /** Search for key within a page, using binary search.
5027 * Returns the smallest entry larger or equal to the key.
5028 * If exactp is non-null, stores whether the found entry was an exact match
5029 * in *exactp (1 or 0).
5030 * Updates the cursor index with the index of the found entry.
5031 * If no entry larger or equal to the key is found, returns NULL.
5034 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5036 unsigned int i = 0, nkeys;
5039 MDB_page *mp = mc->mc_pg[mc->mc_top];
5040 MDB_node *node = NULL;
5045 nkeys = NUMKEYS(mp);
5047 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5048 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5051 low = IS_LEAF(mp) ? 0 : 1;
5053 cmp = mc->mc_dbx->md_cmp;
5055 /* Branch pages have no data, so if using integer keys,
5056 * alignment is guaranteed. Use faster mdb_cmp_int.
5058 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5059 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5066 nodekey.mv_size = mc->mc_db->md_pad;
5067 node = NODEPTR(mp, 0); /* fake */
5068 while (low <= high) {
5069 i = (low + high) >> 1;
5070 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5071 rc = cmp(key, &nodekey);
5072 DPRINTF(("found leaf index %u [%s], rc = %i",
5073 i, DKEY(&nodekey), rc));
5082 while (low <= high) {
5083 i = (low + high) >> 1;
5085 node = NODEPTR(mp, i);
5086 nodekey.mv_size = NODEKSZ(node);
5087 nodekey.mv_data = NODEKEY(node);
5089 rc = cmp(key, &nodekey);
5092 DPRINTF(("found leaf index %u [%s], rc = %i",
5093 i, DKEY(&nodekey), rc));
5095 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5096 i, DKEY(&nodekey), NODEPGNO(node), rc));
5107 if (rc > 0) { /* Found entry is less than the key. */
5108 i++; /* Skip to get the smallest entry larger than key. */
5110 node = NODEPTR(mp, i);
5113 *exactp = (rc == 0 && nkeys > 0);
5114 /* store the key index */
5115 mc->mc_ki[mc->mc_top] = i;
5117 /* There is no entry larger or equal to the key. */
5120 /* nodeptr is fake for LEAF2 */
5126 mdb_cursor_adjust(MDB_cursor *mc, func)
5130 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5131 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5138 /** Pop a page off the top of the cursor's stack. */
5140 mdb_cursor_pop(MDB_cursor *mc)
5143 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5144 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5152 /** Push a page onto the top of the cursor's stack. */
5154 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5156 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5157 DDBI(mc), (void *) mc));
5159 if (mc->mc_snum >= CURSOR_STACK) {
5160 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5161 return MDB_CURSOR_FULL;
5164 mc->mc_top = mc->mc_snum++;
5165 mc->mc_pg[mc->mc_top] = mp;
5166 mc->mc_ki[mc->mc_top] = 0;
5171 /** Find the address of the page corresponding to a given page number.
5172 * @param[in] txn the transaction for this access.
5173 * @param[in] pgno the page number for the page to retrieve.
5174 * @param[out] ret address of a pointer where the page's address will be stored.
5175 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5176 * @return 0 on success, non-zero on failure.
5179 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5181 MDB_env *env = txn->mt_env;
5185 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5189 MDB_ID2L dl = tx2->mt_u.dirty_list;
5191 /* Spilled pages were dirtied in this txn and flushed
5192 * because the dirty list got full. Bring this page
5193 * back in from the map (but don't unspill it here,
5194 * leave that unless page_touch happens again).
5196 if (tx2->mt_spill_pgs) {
5197 MDB_ID pn = pgno << 1;
5198 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5199 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5200 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5205 unsigned x = mdb_mid2l_search(dl, pgno);
5206 if (x <= dl[0].mid && dl[x].mid == pgno) {
5212 } while ((tx2 = tx2->mt_parent) != NULL);
5215 if (pgno < txn->mt_next_pgno) {
5217 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5219 DPRINTF(("page %"Z"u not found", pgno));
5220 txn->mt_flags |= MDB_TXN_ERROR;
5221 return MDB_PAGE_NOTFOUND;
5231 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5232 * The cursor is at the root page, set up the rest of it.
5235 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5237 MDB_page *mp = mc->mc_pg[mc->mc_top];
5241 while (IS_BRANCH(mp)) {
5245 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5246 mdb_cassert(mc, NUMKEYS(mp) > 1);
5247 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5249 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5251 if (flags & MDB_PS_LAST)
5252 i = NUMKEYS(mp) - 1;
5255 node = mdb_node_search(mc, key, &exact);
5257 i = NUMKEYS(mp) - 1;
5259 i = mc->mc_ki[mc->mc_top];
5261 mdb_cassert(mc, i > 0);
5265 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5268 mdb_cassert(mc, i < NUMKEYS(mp));
5269 node = NODEPTR(mp, i);
5271 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5274 mc->mc_ki[mc->mc_top] = i;
5275 if ((rc = mdb_cursor_push(mc, mp)))
5278 if (flags & MDB_PS_MODIFY) {
5279 if ((rc = mdb_page_touch(mc)) != 0)
5281 mp = mc->mc_pg[mc->mc_top];
5286 DPRINTF(("internal error, index points to a %02X page!?",
5288 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5289 return MDB_CORRUPTED;
5292 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5293 key ? DKEY(key) : "null"));
5294 mc->mc_flags |= C_INITIALIZED;
5295 mc->mc_flags &= ~C_EOF;
5300 /** Search for the lowest key under the current branch page.
5301 * This just bypasses a NUMKEYS check in the current page
5302 * before calling mdb_page_search_root(), because the callers
5303 * are all in situations where the current page is known to
5307 mdb_page_search_lowest(MDB_cursor *mc)
5309 MDB_page *mp = mc->mc_pg[mc->mc_top];
5310 MDB_node *node = NODEPTR(mp, 0);
5313 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5316 mc->mc_ki[mc->mc_top] = 0;
5317 if ((rc = mdb_cursor_push(mc, mp)))
5319 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5322 /** Search for the page a given key should be in.
5323 * Push it and its parent pages on the cursor stack.
5324 * @param[in,out] mc the cursor for this operation.
5325 * @param[in] key the key to search for, or NULL for first/last page.
5326 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5327 * are touched (updated with new page numbers).
5328 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5329 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5330 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5331 * @return 0 on success, non-zero on failure.
5334 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5339 /* Make sure the txn is still viable, then find the root from
5340 * the txn's db table and set it as the root of the cursor's stack.
5342 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5343 DPUTS("transaction may not be used now");
5346 /* Make sure we're using an up-to-date root */
5347 if (*mc->mc_dbflag & DB_STALE) {
5349 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5351 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5352 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5359 MDB_node *leaf = mdb_node_search(&mc2,
5360 &mc->mc_dbx->md_name, &exact);
5362 return MDB_NOTFOUND;
5363 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5364 return MDB_INCOMPATIBLE; /* not a named DB */
5365 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5368 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5370 /* The txn may not know this DBI, or another process may
5371 * have dropped and recreated the DB with other flags.
5373 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5374 return MDB_INCOMPATIBLE;
5375 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5377 *mc->mc_dbflag &= ~DB_STALE;
5379 root = mc->mc_db->md_root;
5381 if (root == P_INVALID) { /* Tree is empty. */
5382 DPUTS("tree is empty");
5383 return MDB_NOTFOUND;
5387 mdb_cassert(mc, root > 1);
5388 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5389 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5395 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5396 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5398 if (flags & MDB_PS_MODIFY) {
5399 if ((rc = mdb_page_touch(mc)))
5403 if (flags & MDB_PS_ROOTONLY)
5406 return mdb_page_search_root(mc, key, flags);
5410 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5412 MDB_txn *txn = mc->mc_txn;
5413 pgno_t pg = mp->mp_pgno;
5414 unsigned x = 0, ovpages = mp->mp_pages;
5415 MDB_env *env = txn->mt_env;
5416 MDB_IDL sl = txn->mt_spill_pgs;
5417 MDB_ID pn = pg << 1;
5420 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5421 /* If the page is dirty or on the spill list we just acquired it,
5422 * so we should give it back to our current free list, if any.
5423 * Otherwise put it onto the list of pages we freed in this txn.
5425 * Won't create me_pghead: me_pglast must be inited along with it.
5426 * Unsupported in nested txns: They would need to hide the page
5427 * range in ancestor txns' dirty and spilled lists.
5429 if (env->me_pghead &&
5431 ((mp->mp_flags & P_DIRTY) ||
5432 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5436 MDB_ID2 *dl, ix, iy;
5437 rc = mdb_midl_need(&env->me_pghead, ovpages);
5440 if (!(mp->mp_flags & P_DIRTY)) {
5441 /* This page is no longer spilled */
5448 /* Remove from dirty list */
5449 dl = txn->mt_u.dirty_list;
5451 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5457 mdb_cassert(mc, x > 1);
5459 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5460 txn->mt_flags |= MDB_TXN_ERROR;
5461 return MDB_CORRUPTED;
5464 txn->mt_dirty_room++;
5465 if (!(env->me_flags & MDB_WRITEMAP))
5466 mdb_dpage_free(env, mp);
5468 /* Insert in me_pghead */
5469 mop = env->me_pghead;
5470 j = mop[0] + ovpages;
5471 for (i = mop[0]; i && mop[i] < pg; i--)
5477 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5481 mc->mc_db->md_overflow_pages -= ovpages;
5485 /** Return the data associated with a given node.
5486 * @param[in] txn The transaction for this operation.
5487 * @param[in] leaf The node being read.
5488 * @param[out] data Updated to point to the node's data.
5489 * @return 0 on success, non-zero on failure.
5492 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5494 MDB_page *omp; /* overflow page */
5498 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5499 data->mv_size = NODEDSZ(leaf);
5500 data->mv_data = NODEDATA(leaf);
5504 /* Read overflow data.
5506 data->mv_size = NODEDSZ(leaf);
5507 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5508 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5509 DPRINTF(("read overflow page %"Z"u failed", pgno));
5512 data->mv_data = METADATA(omp);
5518 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5519 MDB_val *key, MDB_val *data)
5526 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5528 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5531 if (txn->mt_flags & MDB_TXN_BLOCKED)
5534 mdb_cursor_init(&mc, txn, dbi, &mx);
5535 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5538 /** Find a sibling for a page.
5539 * Replaces the page at the top of the cursor's stack with the
5540 * specified sibling, if one exists.
5541 * @param[in] mc The cursor for this operation.
5542 * @param[in] move_right Non-zero if the right sibling is requested,
5543 * otherwise the left sibling.
5544 * @return 0 on success, non-zero on failure.
5547 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5553 if (mc->mc_snum < 2) {
5554 return MDB_NOTFOUND; /* root has no siblings */
5558 DPRINTF(("parent page is page %"Z"u, index %u",
5559 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5561 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5562 : (mc->mc_ki[mc->mc_top] == 0)) {
5563 DPRINTF(("no more keys left, moving to %s sibling",
5564 move_right ? "right" : "left"));
5565 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5566 /* undo cursor_pop before returning */
5573 mc->mc_ki[mc->mc_top]++;
5575 mc->mc_ki[mc->mc_top]--;
5576 DPRINTF(("just moving to %s index key %u",
5577 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5579 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5581 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5582 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5583 /* mc will be inconsistent if caller does mc_snum++ as above */
5584 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5588 mdb_cursor_push(mc, mp);
5590 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5595 /** Move the cursor to the next data item. */
5597 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5603 if (mc->mc_flags & C_EOF) {
5604 return MDB_NOTFOUND;
5607 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5609 mp = mc->mc_pg[mc->mc_top];
5611 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5612 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5613 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5614 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5615 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5616 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5617 if (rc == MDB_SUCCESS)
5618 MDB_GET_KEY(leaf, key);
5623 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5624 if (op == MDB_NEXT_DUP)
5625 return MDB_NOTFOUND;
5629 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5630 mdb_dbg_pgno(mp), (void *) mc));
5631 if (mc->mc_flags & C_DEL)
5634 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5635 DPUTS("=====> move to next sibling page");
5636 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5637 mc->mc_flags |= C_EOF;
5640 mp = mc->mc_pg[mc->mc_top];
5641 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5643 mc->mc_ki[mc->mc_top]++;
5646 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5647 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5650 key->mv_size = mc->mc_db->md_pad;
5651 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5655 mdb_cassert(mc, IS_LEAF(mp));
5656 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5658 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5659 mdb_xcursor_init1(mc, leaf);
5662 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5665 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5666 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5667 if (rc != MDB_SUCCESS)
5672 MDB_GET_KEY(leaf, key);
5676 /** Move the cursor to the previous data item. */
5678 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5684 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5686 mp = mc->mc_pg[mc->mc_top];
5688 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5689 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5690 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5691 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5692 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5693 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5694 if (rc == MDB_SUCCESS) {
5695 MDB_GET_KEY(leaf, key);
5696 mc->mc_flags &= ~C_EOF;
5702 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5703 if (op == MDB_PREV_DUP)
5704 return MDB_NOTFOUND;
5708 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5709 mdb_dbg_pgno(mp), (void *) mc));
5711 if (mc->mc_ki[mc->mc_top] == 0) {
5712 DPUTS("=====> move to prev sibling page");
5713 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5716 mp = mc->mc_pg[mc->mc_top];
5717 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5718 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5720 mc->mc_ki[mc->mc_top]--;
5722 mc->mc_flags &= ~C_EOF;
5724 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5725 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5728 key->mv_size = mc->mc_db->md_pad;
5729 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5733 mdb_cassert(mc, IS_LEAF(mp));
5734 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5736 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5737 mdb_xcursor_init1(mc, leaf);
5740 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5743 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5744 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5745 if (rc != MDB_SUCCESS)
5750 MDB_GET_KEY(leaf, key);
5754 /** Set the cursor on a specific data item. */
5756 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5757 MDB_cursor_op op, int *exactp)
5761 MDB_node *leaf = NULL;
5764 if (key->mv_size == 0)
5765 return MDB_BAD_VALSIZE;
5768 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5770 /* See if we're already on the right page */
5771 if (mc->mc_flags & C_INITIALIZED) {
5774 mp = mc->mc_pg[mc->mc_top];
5776 mc->mc_ki[mc->mc_top] = 0;
5777 return MDB_NOTFOUND;
5779 if (mp->mp_flags & P_LEAF2) {
5780 nodekey.mv_size = mc->mc_db->md_pad;
5781 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5783 leaf = NODEPTR(mp, 0);
5784 MDB_GET_KEY2(leaf, nodekey);
5786 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5788 /* Probably happens rarely, but first node on the page
5789 * was the one we wanted.
5791 mc->mc_ki[mc->mc_top] = 0;
5798 unsigned int nkeys = NUMKEYS(mp);
5800 if (mp->mp_flags & P_LEAF2) {
5801 nodekey.mv_data = LEAF2KEY(mp,
5802 nkeys-1, nodekey.mv_size);
5804 leaf = NODEPTR(mp, nkeys-1);
5805 MDB_GET_KEY2(leaf, nodekey);
5807 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5809 /* last node was the one we wanted */
5810 mc->mc_ki[mc->mc_top] = nkeys-1;
5816 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5817 /* This is definitely the right page, skip search_page */
5818 if (mp->mp_flags & P_LEAF2) {
5819 nodekey.mv_data = LEAF2KEY(mp,
5820 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5822 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5823 MDB_GET_KEY2(leaf, nodekey);
5825 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5827 /* current node was the one we wanted */
5837 /* If any parents have right-sibs, search.
5838 * Otherwise, there's nothing further.
5840 for (i=0; i<mc->mc_top; i++)
5842 NUMKEYS(mc->mc_pg[i])-1)
5844 if (i == mc->mc_top) {
5845 /* There are no other pages */
5846 mc->mc_ki[mc->mc_top] = nkeys;
5847 return MDB_NOTFOUND;
5851 /* There are no other pages */
5852 mc->mc_ki[mc->mc_top] = 0;
5853 if (op == MDB_SET_RANGE && !exactp) {
5857 return MDB_NOTFOUND;
5861 rc = mdb_page_search(mc, key, 0);
5862 if (rc != MDB_SUCCESS)
5865 mp = mc->mc_pg[mc->mc_top];
5866 mdb_cassert(mc, IS_LEAF(mp));
5869 leaf = mdb_node_search(mc, key, exactp);
5870 if (exactp != NULL && !*exactp) {
5871 /* MDB_SET specified and not an exact match. */
5872 return MDB_NOTFOUND;
5876 DPUTS("===> inexact leaf not found, goto sibling");
5877 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5878 mc->mc_flags |= C_EOF;
5879 return rc; /* no entries matched */
5881 mp = mc->mc_pg[mc->mc_top];
5882 mdb_cassert(mc, IS_LEAF(mp));
5883 leaf = NODEPTR(mp, 0);
5887 mc->mc_flags |= C_INITIALIZED;
5888 mc->mc_flags &= ~C_EOF;
5891 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5892 key->mv_size = mc->mc_db->md_pad;
5893 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5898 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5899 mdb_xcursor_init1(mc, leaf);
5902 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5903 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5904 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5907 if (op == MDB_GET_BOTH) {
5913 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5914 if (rc != MDB_SUCCESS)
5917 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5920 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5922 dcmp = mc->mc_dbx->md_dcmp;
5923 #if UINT_MAX < SIZE_MAX
5924 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5925 dcmp = mdb_cmp_clong;
5927 rc = dcmp(data, &olddata);
5929 if (op == MDB_GET_BOTH || rc > 0)
5930 return MDB_NOTFOUND;
5937 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5938 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5943 /* The key already matches in all other cases */
5944 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5945 MDB_GET_KEY(leaf, key);
5946 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5951 /** Move the cursor to the first item in the database. */
5953 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5959 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5961 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5962 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5963 if (rc != MDB_SUCCESS)
5966 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5968 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5969 mc->mc_flags |= C_INITIALIZED;
5970 mc->mc_flags &= ~C_EOF;
5972 mc->mc_ki[mc->mc_top] = 0;
5974 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5975 key->mv_size = mc->mc_db->md_pad;
5976 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5981 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5982 mdb_xcursor_init1(mc, leaf);
5983 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5987 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5991 MDB_GET_KEY(leaf, key);
5995 /** Move the cursor to the last item in the database. */
5997 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6003 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6005 if (!(mc->mc_flags & C_EOF)) {
6007 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6008 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6009 if (rc != MDB_SUCCESS)
6012 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6015 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6016 mc->mc_flags |= C_INITIALIZED|C_EOF;
6017 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6019 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6020 key->mv_size = mc->mc_db->md_pad;
6021 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6026 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6027 mdb_xcursor_init1(mc, leaf);
6028 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6032 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6037 MDB_GET_KEY(leaf, key);
6042 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6047 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6052 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6056 case MDB_GET_CURRENT:
6057 if (!(mc->mc_flags & C_INITIALIZED)) {
6060 MDB_page *mp = mc->mc_pg[mc->mc_top];
6061 int nkeys = NUMKEYS(mp);
6062 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6063 mc->mc_ki[mc->mc_top] = nkeys;
6069 key->mv_size = mc->mc_db->md_pad;
6070 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6072 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6073 MDB_GET_KEY(leaf, key);
6075 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6076 if (mc->mc_flags & C_DEL)
6077 mdb_xcursor_init1(mc, leaf);
6078 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6080 rc = mdb_node_read(mc->mc_txn, leaf, data);
6087 case MDB_GET_BOTH_RANGE:
6092 if (mc->mc_xcursor == NULL) {
6093 rc = MDB_INCOMPATIBLE;
6103 rc = mdb_cursor_set(mc, key, data, op,
6104 op == MDB_SET_RANGE ? NULL : &exact);
6107 case MDB_GET_MULTIPLE:
6108 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6112 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6113 rc = MDB_INCOMPATIBLE;
6117 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6118 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6121 case MDB_NEXT_MULTIPLE:
6126 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6127 rc = MDB_INCOMPATIBLE;
6130 if (!(mc->mc_flags & C_INITIALIZED))
6131 rc = mdb_cursor_first(mc, key, data);
6133 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6134 if (rc == MDB_SUCCESS) {
6135 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6138 mx = &mc->mc_xcursor->mx_cursor;
6139 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6141 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6142 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6150 case MDB_NEXT_NODUP:
6151 if (!(mc->mc_flags & C_INITIALIZED))
6152 rc = mdb_cursor_first(mc, key, data);
6154 rc = mdb_cursor_next(mc, key, data, op);
6158 case MDB_PREV_NODUP:
6159 if (!(mc->mc_flags & C_INITIALIZED)) {
6160 rc = mdb_cursor_last(mc, key, data);
6163 mc->mc_flags |= C_INITIALIZED;
6164 mc->mc_ki[mc->mc_top]++;
6166 rc = mdb_cursor_prev(mc, key, data, op);
6169 rc = mdb_cursor_first(mc, key, data);
6172 mfunc = mdb_cursor_first;
6174 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6178 if (mc->mc_xcursor == NULL) {
6179 rc = MDB_INCOMPATIBLE;
6183 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6184 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6185 MDB_GET_KEY(leaf, key);
6186 rc = mdb_node_read(mc->mc_txn, leaf, data);
6190 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6194 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6197 rc = mdb_cursor_last(mc, key, data);
6200 mfunc = mdb_cursor_last;
6203 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6208 if (mc->mc_flags & C_DEL)
6209 mc->mc_flags ^= C_DEL;
6214 /** Touch all the pages in the cursor stack. Set mc_top.
6215 * Makes sure all the pages are writable, before attempting a write operation.
6216 * @param[in] mc The cursor to operate on.
6219 mdb_cursor_touch(MDB_cursor *mc)
6221 int rc = MDB_SUCCESS;
6223 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6226 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6228 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6229 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6232 *mc->mc_dbflag |= DB_DIRTY;
6237 rc = mdb_page_touch(mc);
6238 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6239 mc->mc_top = mc->mc_snum-1;
6244 /** Do not spill pages to disk if txn is getting full, may fail instead */
6245 #define MDB_NOSPILL 0x8000
6248 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6252 MDB_node *leaf = NULL;
6253 MDB_page *fp, *mp, *sub_root = NULL;
6255 MDB_val xdata, *rdata, dkey, olddata;
6257 int do_sub = 0, insert_key, insert_data;
6258 unsigned int mcount = 0, dcount = 0, nospill;
6261 unsigned int nflags;
6264 if (mc == NULL || key == NULL)
6267 env = mc->mc_txn->mt_env;
6269 /* Check this first so counter will always be zero on any
6272 if (flags & MDB_MULTIPLE) {
6273 dcount = data[1].mv_size;
6274 data[1].mv_size = 0;
6275 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6276 return MDB_INCOMPATIBLE;
6279 nospill = flags & MDB_NOSPILL;
6280 flags &= ~MDB_NOSPILL;
6282 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6283 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6285 if (key->mv_size-1 >= ENV_MAXKEY(env))
6286 return MDB_BAD_VALSIZE;
6288 #if SIZE_MAX > MAXDATASIZE
6289 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6290 return MDB_BAD_VALSIZE;
6292 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6293 return MDB_BAD_VALSIZE;
6296 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6297 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6301 if (flags == MDB_CURRENT) {
6302 if (!(mc->mc_flags & C_INITIALIZED))
6305 } else if (mc->mc_db->md_root == P_INVALID) {
6306 /* new database, cursor has nothing to point to */
6309 mc->mc_flags &= ~C_INITIALIZED;
6314 if (flags & MDB_APPEND) {
6316 rc = mdb_cursor_last(mc, &k2, &d2);
6318 rc = mc->mc_dbx->md_cmp(key, &k2);
6321 mc->mc_ki[mc->mc_top]++;
6323 /* new key is <= last key */
6328 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6330 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6331 DPRINTF(("duplicate key [%s]", DKEY(key)));
6333 return MDB_KEYEXIST;
6335 if (rc && rc != MDB_NOTFOUND)
6339 if (mc->mc_flags & C_DEL)
6340 mc->mc_flags ^= C_DEL;
6342 /* Cursor is positioned, check for room in the dirty list */
6344 if (flags & MDB_MULTIPLE) {
6346 xdata.mv_size = data->mv_size * dcount;
6350 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6354 if (rc == MDB_NO_ROOT) {
6356 /* new database, write a root leaf page */
6357 DPUTS("allocating new root leaf page");
6358 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6361 mdb_cursor_push(mc, np);
6362 mc->mc_db->md_root = np->mp_pgno;
6363 mc->mc_db->md_depth++;
6364 *mc->mc_dbflag |= DB_DIRTY;
6365 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6367 np->mp_flags |= P_LEAF2;
6368 mc->mc_flags |= C_INITIALIZED;
6370 /* make sure all cursor pages are writable */
6371 rc2 = mdb_cursor_touch(mc);
6376 insert_key = insert_data = rc;
6378 /* The key does not exist */
6379 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6380 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6381 LEAFSIZE(key, data) > env->me_nodemax)
6383 /* Too big for a node, insert in sub-DB. Set up an empty
6384 * "old sub-page" for prep_subDB to expand to a full page.
6386 fp_flags = P_LEAF|P_DIRTY;
6388 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6389 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6390 olddata.mv_size = PAGEHDRSZ;
6394 /* there's only a key anyway, so this is a no-op */
6395 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6397 unsigned int ksize = mc->mc_db->md_pad;
6398 if (key->mv_size != ksize)
6399 return MDB_BAD_VALSIZE;
6400 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6401 memcpy(ptr, key->mv_data, ksize);
6403 /* if overwriting slot 0 of leaf, need to
6404 * update branch key if there is a parent page
6406 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6407 unsigned short dtop = 1;
6409 /* slot 0 is always an empty key, find real slot */
6410 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6414 if (mc->mc_ki[mc->mc_top])
6415 rc2 = mdb_update_key(mc, key);
6426 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6427 olddata.mv_size = NODEDSZ(leaf);
6428 olddata.mv_data = NODEDATA(leaf);
6431 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6432 /* Prepare (sub-)page/sub-DB to accept the new item,
6433 * if needed. fp: old sub-page or a header faking
6434 * it. mp: new (sub-)page. offset: growth in page
6435 * size. xdata: node data with new page or DB.
6437 unsigned i, offset = 0;
6438 mp = fp = xdata.mv_data = env->me_pbuf;
6439 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6441 /* Was a single item before, must convert now */
6442 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6444 /* Just overwrite the current item */
6445 if (flags == MDB_CURRENT)
6447 dcmp = mc->mc_dbx->md_dcmp;
6448 #if UINT_MAX < SIZE_MAX
6449 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6450 dcmp = mdb_cmp_clong;
6452 /* does data match? */
6453 if (!dcmp(data, &olddata)) {
6454 if (flags & MDB_NODUPDATA)
6455 return MDB_KEYEXIST;
6460 /* Back up original data item */
6461 dkey.mv_size = olddata.mv_size;
6462 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6464 /* Make sub-page header for the dup items, with dummy body */
6465 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6466 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6467 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6468 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6469 fp->mp_flags |= P_LEAF2;
6470 fp->mp_pad = data->mv_size;
6471 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6473 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6474 (dkey.mv_size & 1) + (data->mv_size & 1);
6476 fp->mp_upper = xdata.mv_size - PAGEBASE;
6477 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6478 } else if (leaf->mn_flags & F_SUBDATA) {
6479 /* Data is on sub-DB, just store it */
6480 flags |= F_DUPDATA|F_SUBDATA;
6483 /* Data is on sub-page */
6484 fp = olddata.mv_data;
6487 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6488 offset = EVEN(NODESIZE + sizeof(indx_t) +
6492 offset = fp->mp_pad;
6493 if (SIZELEFT(fp) < offset) {
6494 offset *= 4; /* space for 4 more */
6497 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6499 fp->mp_flags |= P_DIRTY;
6500 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6501 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6505 xdata.mv_size = olddata.mv_size + offset;
6508 fp_flags = fp->mp_flags;
6509 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6510 /* Too big for a sub-page, convert to sub-DB */
6511 fp_flags &= ~P_SUBP;
6513 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6514 fp_flags |= P_LEAF2;
6515 dummy.md_pad = fp->mp_pad;
6516 dummy.md_flags = MDB_DUPFIXED;
6517 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6518 dummy.md_flags |= MDB_INTEGERKEY;
6524 dummy.md_branch_pages = 0;
6525 dummy.md_leaf_pages = 1;
6526 dummy.md_overflow_pages = 0;
6527 dummy.md_entries = NUMKEYS(fp);
6528 xdata.mv_size = sizeof(MDB_db);
6529 xdata.mv_data = &dummy;
6530 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6532 offset = env->me_psize - olddata.mv_size;
6533 flags |= F_DUPDATA|F_SUBDATA;
6534 dummy.md_root = mp->mp_pgno;
6538 mp->mp_flags = fp_flags | P_DIRTY;
6539 mp->mp_pad = fp->mp_pad;
6540 mp->mp_lower = fp->mp_lower;
6541 mp->mp_upper = fp->mp_upper + offset;
6542 if (fp_flags & P_LEAF2) {
6543 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6545 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6546 olddata.mv_size - fp->mp_upper - PAGEBASE);
6547 for (i=0; i<NUMKEYS(fp); i++)
6548 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6556 mdb_node_del(mc, 0);
6560 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6561 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6562 return MDB_INCOMPATIBLE;
6563 /* overflow page overwrites need special handling */
6564 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6567 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6569 memcpy(&pg, olddata.mv_data, sizeof(pg));
6570 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6572 ovpages = omp->mp_pages;
6574 /* Is the ov page large enough? */
6575 if (ovpages >= dpages) {
6576 if (!(omp->mp_flags & P_DIRTY) &&
6577 (level || (env->me_flags & MDB_WRITEMAP)))
6579 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6582 level = 0; /* dirty in this txn or clean */
6585 if (omp->mp_flags & P_DIRTY) {
6586 /* yes, overwrite it. Note in this case we don't
6587 * bother to try shrinking the page if the new data
6588 * is smaller than the overflow threshold.
6591 /* It is writable only in a parent txn */
6592 size_t sz = (size_t) env->me_psize * ovpages, off;
6593 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6599 /* Note - this page is already counted in parent's dirty_room */
6600 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6601 mdb_cassert(mc, rc2 == 0);
6602 if (!(flags & MDB_RESERVE)) {
6603 /* Copy end of page, adjusting alignment so
6604 * compiler may copy words instead of bytes.
6606 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6607 memcpy((size_t *)((char *)np + off),
6608 (size_t *)((char *)omp + off), sz - off);
6611 memcpy(np, omp, sz); /* Copy beginning of page */
6614 SETDSZ(leaf, data->mv_size);
6615 if (F_ISSET(flags, MDB_RESERVE))
6616 data->mv_data = METADATA(omp);
6618 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6622 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6624 } else if (data->mv_size == olddata.mv_size) {
6625 /* same size, just replace it. Note that we could
6626 * also reuse this node if the new data is smaller,
6627 * but instead we opt to shrink the node in that case.
6629 if (F_ISSET(flags, MDB_RESERVE))
6630 data->mv_data = olddata.mv_data;
6631 else if (!(mc->mc_flags & C_SUB))
6632 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6634 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6639 mdb_node_del(mc, 0);
6645 nflags = flags & NODE_ADD_FLAGS;
6646 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6647 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6648 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6649 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6651 nflags |= MDB_SPLIT_REPLACE;
6652 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6654 /* There is room already in this leaf page. */
6655 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6656 if (rc == 0 && insert_key) {
6657 /* Adjust other cursors pointing to mp */
6658 MDB_cursor *m2, *m3;
6659 MDB_dbi dbi = mc->mc_dbi;
6660 unsigned i = mc->mc_top;
6661 MDB_page *mp = mc->mc_pg[i];
6663 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6664 if (mc->mc_flags & C_SUB)
6665 m3 = &m2->mc_xcursor->mx_cursor;
6668 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6669 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6676 if (rc == MDB_SUCCESS) {
6677 /* Now store the actual data in the child DB. Note that we're
6678 * storing the user data in the keys field, so there are strict
6679 * size limits on dupdata. The actual data fields of the child
6680 * DB are all zero size.
6683 int xflags, new_dupdata;
6688 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6689 if (flags & MDB_CURRENT) {
6690 xflags = MDB_CURRENT|MDB_NOSPILL;
6692 mdb_xcursor_init1(mc, leaf);
6693 xflags = (flags & MDB_NODUPDATA) ?
6694 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6697 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6698 new_dupdata = (int)dkey.mv_size;
6699 /* converted, write the original data first */
6701 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6704 /* we've done our job */
6707 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6708 /* Adjust other cursors pointing to mp */
6710 MDB_xcursor *mx = mc->mc_xcursor;
6711 unsigned i = mc->mc_top;
6712 MDB_page *mp = mc->mc_pg[i];
6713 int nkeys = NUMKEYS(mp);
6715 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6716 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6717 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6718 if (m2->mc_pg[i] == mp) {
6719 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6720 mdb_xcursor_init2(m2, mx, new_dupdata);
6721 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
6722 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6723 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
6724 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6729 ecount = mc->mc_xcursor->mx_db.md_entries;
6730 if (flags & MDB_APPENDDUP)
6731 xflags |= MDB_APPEND;
6732 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6733 if (flags & F_SUBDATA) {
6734 void *db = NODEDATA(leaf);
6735 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6737 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6739 /* Increment count unless we just replaced an existing item. */
6741 mc->mc_db->md_entries++;
6743 /* Invalidate txn if we created an empty sub-DB */
6746 /* If we succeeded and the key didn't exist before,
6747 * make sure the cursor is marked valid.
6749 mc->mc_flags |= C_INITIALIZED;
6751 if (flags & MDB_MULTIPLE) {
6754 /* let caller know how many succeeded, if any */
6755 data[1].mv_size = mcount;
6756 if (mcount < dcount) {
6757 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6758 insert_key = insert_data = 0;
6765 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6768 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6773 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6779 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6780 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6782 if (!(mc->mc_flags & C_INITIALIZED))
6785 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6786 return MDB_NOTFOUND;
6788 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6791 rc = mdb_cursor_touch(mc);
6795 mp = mc->mc_pg[mc->mc_top];
6798 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6800 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6801 if (flags & MDB_NODUPDATA) {
6802 /* mdb_cursor_del0() will subtract the final entry */
6803 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6805 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6806 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6808 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6811 /* If sub-DB still has entries, we're done */
6812 if (mc->mc_xcursor->mx_db.md_entries) {
6813 if (leaf->mn_flags & F_SUBDATA) {
6814 /* update subDB info */
6815 void *db = NODEDATA(leaf);
6816 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6819 /* shrink fake page */
6820 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6821 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6822 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6823 /* fix other sub-DB cursors pointed at fake pages on this page */
6824 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6825 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6826 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6827 if (m2->mc_pg[mc->mc_top] == mp) {
6828 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6829 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6831 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6832 if (!(n2->mn_flags & F_SUBDATA))
6833 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6838 mc->mc_db->md_entries--;
6839 mc->mc_flags |= C_DEL;
6842 /* otherwise fall thru and delete the sub-DB */
6845 if (leaf->mn_flags & F_SUBDATA) {
6846 /* add all the child DB's pages to the free list */
6847 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6852 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6853 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6854 rc = MDB_INCOMPATIBLE;
6858 /* add overflow pages to free list */
6859 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6863 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6864 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6865 (rc = mdb_ovpage_free(mc, omp)))
6870 return mdb_cursor_del0(mc);
6873 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6877 /** Allocate and initialize new pages for a database.
6878 * @param[in] mc a cursor on the database being added to.
6879 * @param[in] flags flags defining what type of page is being allocated.
6880 * @param[in] num the number of pages to allocate. This is usually 1,
6881 * unless allocating overflow pages for a large record.
6882 * @param[out] mp Address of a page, or NULL on failure.
6883 * @return 0 on success, non-zero on failure.
6886 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6891 if ((rc = mdb_page_alloc(mc, num, &np)))
6893 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6894 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6895 np->mp_flags = flags | P_DIRTY;
6896 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6897 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6900 mc->mc_db->md_branch_pages++;
6901 else if (IS_LEAF(np))
6902 mc->mc_db->md_leaf_pages++;
6903 else if (IS_OVERFLOW(np)) {
6904 mc->mc_db->md_overflow_pages += num;
6912 /** Calculate the size of a leaf node.
6913 * The size depends on the environment's page size; if a data item
6914 * is too large it will be put onto an overflow page and the node
6915 * size will only include the key and not the data. Sizes are always
6916 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6917 * of the #MDB_node headers.
6918 * @param[in] env The environment handle.
6919 * @param[in] key The key for the node.
6920 * @param[in] data The data for the node.
6921 * @return The number of bytes needed to store the node.
6924 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6928 sz = LEAFSIZE(key, data);
6929 if (sz > env->me_nodemax) {
6930 /* put on overflow page */
6931 sz -= data->mv_size - sizeof(pgno_t);
6934 return EVEN(sz + sizeof(indx_t));
6937 /** Calculate the size of a branch node.
6938 * The size should depend on the environment's page size but since
6939 * we currently don't support spilling large keys onto overflow
6940 * pages, it's simply the size of the #MDB_node header plus the
6941 * size of the key. Sizes are always rounded up to an even number
6942 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6943 * @param[in] env The environment handle.
6944 * @param[in] key The key for the node.
6945 * @return The number of bytes needed to store the node.
6948 mdb_branch_size(MDB_env *env, MDB_val *key)
6953 if (sz > env->me_nodemax) {
6954 /* put on overflow page */
6955 /* not implemented */
6956 /* sz -= key->size - sizeof(pgno_t); */
6959 return sz + sizeof(indx_t);
6962 /** Add a node to the page pointed to by the cursor.
6963 * @param[in] mc The cursor for this operation.
6964 * @param[in] indx The index on the page where the new node should be added.
6965 * @param[in] key The key for the new node.
6966 * @param[in] data The data for the new node, if any.
6967 * @param[in] pgno The page number, if adding a branch node.
6968 * @param[in] flags Flags for the node.
6969 * @return 0 on success, non-zero on failure. Possible errors are:
6971 * <li>ENOMEM - failed to allocate overflow pages for the node.
6972 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6973 * should never happen since all callers already calculate the
6974 * page's free space before calling this function.
6978 mdb_node_add(MDB_cursor *mc, indx_t indx,
6979 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6982 size_t node_size = NODESIZE;
6986 MDB_page *mp = mc->mc_pg[mc->mc_top];
6987 MDB_page *ofp = NULL; /* overflow page */
6991 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6993 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6994 IS_LEAF(mp) ? "leaf" : "branch",
6995 IS_SUBP(mp) ? "sub-" : "",
6996 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6997 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7000 /* Move higher keys up one slot. */
7001 int ksize = mc->mc_db->md_pad, dif;
7002 char *ptr = LEAF2KEY(mp, indx, ksize);
7003 dif = NUMKEYS(mp) - indx;
7005 memmove(ptr+ksize, ptr, dif*ksize);
7006 /* insert new key */
7007 memcpy(ptr, key->mv_data, ksize);
7009 /* Just using these for counting */
7010 mp->mp_lower += sizeof(indx_t);
7011 mp->mp_upper -= ksize - sizeof(indx_t);
7015 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7017 node_size += key->mv_size;
7019 mdb_cassert(mc, key && data);
7020 if (F_ISSET(flags, F_BIGDATA)) {
7021 /* Data already on overflow page. */
7022 node_size += sizeof(pgno_t);
7023 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7024 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7026 /* Put data on overflow page. */
7027 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7028 data->mv_size, node_size+data->mv_size));
7029 node_size = EVEN(node_size + sizeof(pgno_t));
7030 if ((ssize_t)node_size > room)
7032 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7034 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7038 node_size += data->mv_size;
7041 node_size = EVEN(node_size);
7042 if ((ssize_t)node_size > room)
7046 /* Move higher pointers up one slot. */
7047 for (i = NUMKEYS(mp); i > indx; i--)
7048 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7050 /* Adjust free space offsets. */
7051 ofs = mp->mp_upper - node_size;
7052 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7053 mp->mp_ptrs[indx] = ofs;
7055 mp->mp_lower += sizeof(indx_t);
7057 /* Write the node data. */
7058 node = NODEPTR(mp, indx);
7059 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7060 node->mn_flags = flags;
7062 SETDSZ(node,data->mv_size);
7067 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7070 ndata = NODEDATA(node);
7072 if (F_ISSET(flags, F_BIGDATA))
7073 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7074 else if (F_ISSET(flags, MDB_RESERVE))
7075 data->mv_data = ndata;
7077 memcpy(ndata, data->mv_data, data->mv_size);
7079 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7080 ndata = METADATA(ofp);
7081 if (F_ISSET(flags, MDB_RESERVE))
7082 data->mv_data = ndata;
7084 memcpy(ndata, data->mv_data, data->mv_size);
7091 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7092 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7093 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7094 DPRINTF(("node size = %"Z"u", node_size));
7095 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7096 return MDB_PAGE_FULL;
7099 /** Delete the specified node from a page.
7100 * @param[in] mc Cursor pointing to the node to delete.
7101 * @param[in] ksize The size of a node. Only used if the page is
7102 * part of a #MDB_DUPFIXED database.
7105 mdb_node_del(MDB_cursor *mc, int ksize)
7107 MDB_page *mp = mc->mc_pg[mc->mc_top];
7108 indx_t indx = mc->mc_ki[mc->mc_top];
7110 indx_t i, j, numkeys, ptr;
7114 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7115 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7116 numkeys = NUMKEYS(mp);
7117 mdb_cassert(mc, indx < numkeys);
7120 int x = numkeys - 1 - indx;
7121 base = LEAF2KEY(mp, indx, ksize);
7123 memmove(base, base + ksize, x * ksize);
7124 mp->mp_lower -= sizeof(indx_t);
7125 mp->mp_upper += ksize - sizeof(indx_t);
7129 node = NODEPTR(mp, indx);
7130 sz = NODESIZE + node->mn_ksize;
7132 if (F_ISSET(node->mn_flags, F_BIGDATA))
7133 sz += sizeof(pgno_t);
7135 sz += NODEDSZ(node);
7139 ptr = mp->mp_ptrs[indx];
7140 for (i = j = 0; i < numkeys; i++) {
7142 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7143 if (mp->mp_ptrs[i] < ptr)
7144 mp->mp_ptrs[j] += sz;
7149 base = (char *)mp + mp->mp_upper + PAGEBASE;
7150 memmove(base + sz, base, ptr - mp->mp_upper);
7152 mp->mp_lower -= sizeof(indx_t);
7156 /** Compact the main page after deleting a node on a subpage.
7157 * @param[in] mp The main page to operate on.
7158 * @param[in] indx The index of the subpage on the main page.
7161 mdb_node_shrink(MDB_page *mp, indx_t indx)
7166 indx_t delta, nsize, len, ptr;
7169 node = NODEPTR(mp, indx);
7170 sp = (MDB_page *)NODEDATA(node);
7171 delta = SIZELEFT(sp);
7172 nsize = NODEDSZ(node) - delta;
7174 /* Prepare to shift upward, set len = length(subpage part to shift) */
7178 return; /* do not make the node uneven-sized */
7180 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7181 for (i = NUMKEYS(sp); --i >= 0; )
7182 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7185 sp->mp_upper = sp->mp_lower;
7186 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7187 SETDSZ(node, nsize);
7189 /* Shift <lower nodes...initial part of subpage> upward */
7190 base = (char *)mp + mp->mp_upper + PAGEBASE;
7191 memmove(base + delta, base, (char *)sp + len - base);
7193 ptr = mp->mp_ptrs[indx];
7194 for (i = NUMKEYS(mp); --i >= 0; ) {
7195 if (mp->mp_ptrs[i] <= ptr)
7196 mp->mp_ptrs[i] += delta;
7198 mp->mp_upper += delta;
7201 /** Initial setup of a sorted-dups cursor.
7202 * Sorted duplicates are implemented as a sub-database for the given key.
7203 * The duplicate data items are actually keys of the sub-database.
7204 * Operations on the duplicate data items are performed using a sub-cursor
7205 * initialized when the sub-database is first accessed. This function does
7206 * the preliminary setup of the sub-cursor, filling in the fields that
7207 * depend only on the parent DB.
7208 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7211 mdb_xcursor_init0(MDB_cursor *mc)
7213 MDB_xcursor *mx = mc->mc_xcursor;
7215 mx->mx_cursor.mc_xcursor = NULL;
7216 mx->mx_cursor.mc_txn = mc->mc_txn;
7217 mx->mx_cursor.mc_db = &mx->mx_db;
7218 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7219 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7220 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7221 mx->mx_cursor.mc_snum = 0;
7222 mx->mx_cursor.mc_top = 0;
7223 mx->mx_cursor.mc_flags = C_SUB;
7224 mx->mx_dbx.md_name.mv_size = 0;
7225 mx->mx_dbx.md_name.mv_data = NULL;
7226 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7227 mx->mx_dbx.md_dcmp = NULL;
7228 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7231 /** Final setup of a sorted-dups cursor.
7232 * Sets up the fields that depend on the data from the main cursor.
7233 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7234 * @param[in] node The data containing the #MDB_db record for the
7235 * sorted-dup database.
7238 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7240 MDB_xcursor *mx = mc->mc_xcursor;
7242 if (node->mn_flags & F_SUBDATA) {
7243 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7244 mx->mx_cursor.mc_pg[0] = 0;
7245 mx->mx_cursor.mc_snum = 0;
7246 mx->mx_cursor.mc_top = 0;
7247 mx->mx_cursor.mc_flags = C_SUB;
7249 MDB_page *fp = NODEDATA(node);
7250 mx->mx_db.md_pad = 0;
7251 mx->mx_db.md_flags = 0;
7252 mx->mx_db.md_depth = 1;
7253 mx->mx_db.md_branch_pages = 0;
7254 mx->mx_db.md_leaf_pages = 1;
7255 mx->mx_db.md_overflow_pages = 0;
7256 mx->mx_db.md_entries = NUMKEYS(fp);
7257 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7258 mx->mx_cursor.mc_snum = 1;
7259 mx->mx_cursor.mc_top = 0;
7260 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7261 mx->mx_cursor.mc_pg[0] = fp;
7262 mx->mx_cursor.mc_ki[0] = 0;
7263 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7264 mx->mx_db.md_flags = MDB_DUPFIXED;
7265 mx->mx_db.md_pad = fp->mp_pad;
7266 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7267 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7270 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7271 mx->mx_db.md_root));
7272 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7273 #if UINT_MAX < SIZE_MAX
7274 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7275 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7280 /** Fixup a sorted-dups cursor due to underlying update.
7281 * Sets up some fields that depend on the data from the main cursor.
7282 * Almost the same as init1, but skips initialization steps if the
7283 * xcursor had already been used.
7284 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7285 * @param[in] src_mx The xcursor of an up-to-date cursor.
7286 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7289 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7291 MDB_xcursor *mx = mc->mc_xcursor;
7294 mx->mx_cursor.mc_snum = 1;
7295 mx->mx_cursor.mc_top = 0;
7296 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7297 mx->mx_cursor.mc_ki[0] = 0;
7298 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7299 #if UINT_MAX < SIZE_MAX
7300 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7302 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7305 mx->mx_db = src_mx->mx_db;
7306 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7307 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7308 mx->mx_db.md_root));
7311 /** Initialize a cursor for a given transaction and database. */
7313 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7316 mc->mc_backup = NULL;
7319 mc->mc_db = &txn->mt_dbs[dbi];
7320 mc->mc_dbx = &txn->mt_dbxs[dbi];
7321 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7327 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7328 mdb_tassert(txn, mx != NULL);
7329 mc->mc_xcursor = mx;
7330 mdb_xcursor_init0(mc);
7332 mc->mc_xcursor = NULL;
7334 if (*mc->mc_dbflag & DB_STALE) {
7335 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7340 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7343 size_t size = sizeof(MDB_cursor);
7345 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7348 if (txn->mt_flags & MDB_TXN_BLOCKED)
7351 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7354 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7355 size += sizeof(MDB_xcursor);
7357 if ((mc = malloc(size)) != NULL) {
7358 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7359 if (txn->mt_cursors) {
7360 mc->mc_next = txn->mt_cursors[dbi];
7361 txn->mt_cursors[dbi] = mc;
7362 mc->mc_flags |= C_UNTRACK;
7374 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7376 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7379 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7382 if (txn->mt_flags & MDB_TXN_BLOCKED)
7385 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7389 /* Return the count of duplicate data items for the current key */
7391 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7395 if (mc == NULL || countp == NULL)
7398 if (mc->mc_xcursor == NULL)
7399 return MDB_INCOMPATIBLE;
7401 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7404 if (!(mc->mc_flags & C_INITIALIZED))
7407 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7408 return MDB_NOTFOUND;
7410 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7411 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7414 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7417 *countp = mc->mc_xcursor->mx_db.md_entries;
7423 mdb_cursor_close(MDB_cursor *mc)
7425 if (mc && !mc->mc_backup) {
7426 /* remove from txn, if tracked */
7427 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7428 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7429 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7431 *prev = mc->mc_next;
7438 mdb_cursor_txn(MDB_cursor *mc)
7440 if (!mc) return NULL;
7445 mdb_cursor_dbi(MDB_cursor *mc)
7450 /** Replace the key for a branch node with a new key.
7451 * @param[in] mc Cursor pointing to the node to operate on.
7452 * @param[in] key The new key to use.
7453 * @return 0 on success, non-zero on failure.
7456 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7462 int delta, ksize, oksize;
7463 indx_t ptr, i, numkeys, indx;
7466 indx = mc->mc_ki[mc->mc_top];
7467 mp = mc->mc_pg[mc->mc_top];
7468 node = NODEPTR(mp, indx);
7469 ptr = mp->mp_ptrs[indx];
7473 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7474 k2.mv_data = NODEKEY(node);
7475 k2.mv_size = node->mn_ksize;
7476 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7478 mdb_dkey(&k2, kbuf2),
7484 /* Sizes must be 2-byte aligned. */
7485 ksize = EVEN(key->mv_size);
7486 oksize = EVEN(node->mn_ksize);
7487 delta = ksize - oksize;
7489 /* Shift node contents if EVEN(key length) changed. */
7491 if (delta > 0 && SIZELEFT(mp) < delta) {
7493 /* not enough space left, do a delete and split */
7494 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7495 pgno = NODEPGNO(node);
7496 mdb_node_del(mc, 0);
7497 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7500 numkeys = NUMKEYS(mp);
7501 for (i = 0; i < numkeys; i++) {
7502 if (mp->mp_ptrs[i] <= ptr)
7503 mp->mp_ptrs[i] -= delta;
7506 base = (char *)mp + mp->mp_upper + PAGEBASE;
7507 len = ptr - mp->mp_upper + NODESIZE;
7508 memmove(base - delta, base, len);
7509 mp->mp_upper -= delta;
7511 node = NODEPTR(mp, indx);
7514 /* But even if no shift was needed, update ksize */
7515 if (node->mn_ksize != key->mv_size)
7516 node->mn_ksize = key->mv_size;
7519 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7525 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7527 /** Track a temporary cursor */
7528 #define CURSOR_TMP_TRACK(mc, mn, dummy, tracked) \
7529 if (mc->mc_flags & C_SUB) { \
7530 dummy.mc_flags = C_INITIALIZED; \
7531 dummy.mc_xcursor = (MDB_xcursor *)&mn; \
7536 tracked->mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi]; \
7537 mc->mc_txn->mt_cursors[mc->mc_dbi] = tracked
7539 /** Stop tracking a temporary cursor */
7540 #define CURSOR_TMP_UNTRACK(mc, tracked) \
7541 mc->mc_txn->mt_cursors[mc->mc_dbi] = tracked->mc_next
7543 /** Move a node from csrc to cdst.
7546 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
7553 unsigned short flags;
7557 /* Mark src and dst as dirty. */
7558 if ((rc = mdb_page_touch(csrc)) ||
7559 (rc = mdb_page_touch(cdst)))
7562 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7563 key.mv_size = csrc->mc_db->md_pad;
7564 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7566 data.mv_data = NULL;
7570 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7571 mdb_cassert(csrc, !((size_t)srcnode & 1));
7572 srcpg = NODEPGNO(srcnode);
7573 flags = srcnode->mn_flags;
7574 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7575 unsigned int snum = csrc->mc_snum;
7577 /* must find the lowest key below src */
7578 rc = mdb_page_search_lowest(csrc);
7581 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7582 key.mv_size = csrc->mc_db->md_pad;
7583 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7585 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7586 key.mv_size = NODEKSZ(s2);
7587 key.mv_data = NODEKEY(s2);
7589 csrc->mc_snum = snum--;
7590 csrc->mc_top = snum;
7592 key.mv_size = NODEKSZ(srcnode);
7593 key.mv_data = NODEKEY(srcnode);
7595 data.mv_size = NODEDSZ(srcnode);
7596 data.mv_data = NODEDATA(srcnode);
7598 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7599 unsigned int snum = cdst->mc_snum;
7602 /* must find the lowest key below dst */
7603 mdb_cursor_copy(cdst, &mn);
7604 rc = mdb_page_search_lowest(&mn);
7607 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7608 bkey.mv_size = mn.mc_db->md_pad;
7609 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7611 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7612 bkey.mv_size = NODEKSZ(s2);
7613 bkey.mv_data = NODEKEY(s2);
7615 mn.mc_snum = snum--;
7618 rc = mdb_update_key(&mn, &bkey);
7623 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7624 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7625 csrc->mc_ki[csrc->mc_top],
7627 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7628 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7630 /* Add the node to the destination page.
7632 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7633 if (rc != MDB_SUCCESS)
7636 /* Delete the node from the source page.
7638 mdb_node_del(csrc, key.mv_size);
7641 /* Adjust other cursors pointing to mp */
7642 MDB_cursor *m2, *m3;
7643 MDB_dbi dbi = csrc->mc_dbi;
7644 MDB_page *mpd, *mps;
7646 mps = csrc->mc_pg[csrc->mc_top];
7647 /* If we're adding on the left, bump others up */
7649 mpd = cdst->mc_pg[csrc->mc_top];
7650 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7651 if (csrc->mc_flags & C_SUB)
7652 m3 = &m2->mc_xcursor->mx_cursor;
7655 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7658 m3->mc_pg[csrc->mc_top] == mpd &&
7659 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7660 m3->mc_ki[csrc->mc_top]++;
7663 m3->mc_pg[csrc->mc_top] == mps &&
7664 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7665 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7666 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7667 m3->mc_ki[csrc->mc_top-1]++;
7671 /* Adding on the right, bump others down */
7673 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7674 if (csrc->mc_flags & C_SUB)
7675 m3 = &m2->mc_xcursor->mx_cursor;
7678 if (m3 == csrc) continue;
7679 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7681 if (m3->mc_pg[csrc->mc_top] == mps) {
7682 if (!m3->mc_ki[csrc->mc_top]) {
7683 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7684 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7685 m3->mc_ki[csrc->mc_top-1]--;
7687 m3->mc_ki[csrc->mc_top]--;
7694 /* Update the parent separators.
7696 if (csrc->mc_ki[csrc->mc_top] == 0) {
7697 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7698 MDB_cursor dummy, *tracked;
7699 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7700 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7702 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7703 key.mv_size = NODEKSZ(srcnode);
7704 key.mv_data = NODEKEY(srcnode);
7706 DPRINTF(("update separator for source page %"Z"u to [%s]",
7707 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7708 mdb_cursor_copy(csrc, &mn);
7711 /* We want mdb_rebalance to find mn when doing fixups */
7712 CURSOR_TMP_TRACK(csrc, mn, dummy, tracked);
7713 rc = mdb_update_key(&mn, &key);
7714 CURSOR_TMP_UNTRACK(csrc, tracked);
7718 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7720 indx_t ix = csrc->mc_ki[csrc->mc_top];
7721 nullkey.mv_size = 0;
7722 csrc->mc_ki[csrc->mc_top] = 0;
7723 rc = mdb_update_key(csrc, &nullkey);
7724 csrc->mc_ki[csrc->mc_top] = ix;
7725 mdb_cassert(csrc, rc == MDB_SUCCESS);
7729 if (cdst->mc_ki[cdst->mc_top] == 0) {
7730 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7731 MDB_cursor dummy, *tracked;
7732 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7733 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7735 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7736 key.mv_size = NODEKSZ(srcnode);
7737 key.mv_data = NODEKEY(srcnode);
7739 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7740 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7741 mdb_cursor_copy(cdst, &mn);
7744 /* We want mdb_rebalance to find mn when doing fixups */
7745 CURSOR_TMP_TRACK(cdst, mn, dummy, tracked);
7746 rc = mdb_update_key(&mn, &key);
7747 CURSOR_TMP_UNTRACK(cdst, tracked);
7751 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7753 indx_t ix = cdst->mc_ki[cdst->mc_top];
7754 nullkey.mv_size = 0;
7755 cdst->mc_ki[cdst->mc_top] = 0;
7756 rc = mdb_update_key(cdst, &nullkey);
7757 cdst->mc_ki[cdst->mc_top] = ix;
7758 mdb_cassert(cdst, rc == MDB_SUCCESS);
7765 /** Merge one page into another.
7766 * The nodes from the page pointed to by \b csrc will
7767 * be copied to the page pointed to by \b cdst and then
7768 * the \b csrc page will be freed.
7769 * @param[in] csrc Cursor pointing to the source page.
7770 * @param[in] cdst Cursor pointing to the destination page.
7771 * @return 0 on success, non-zero on failure.
7774 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7776 MDB_page *psrc, *pdst;
7783 psrc = csrc->mc_pg[csrc->mc_top];
7784 pdst = cdst->mc_pg[cdst->mc_top];
7786 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7788 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7789 mdb_cassert(csrc, cdst->mc_snum > 1);
7791 /* Mark dst as dirty. */
7792 if ((rc = mdb_page_touch(cdst)))
7795 /* get dst page again now that we've touched it. */
7796 pdst = cdst->mc_pg[cdst->mc_top];
7798 /* Move all nodes from src to dst.
7800 j = nkeys = NUMKEYS(pdst);
7801 if (IS_LEAF2(psrc)) {
7802 key.mv_size = csrc->mc_db->md_pad;
7803 key.mv_data = METADATA(psrc);
7804 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7805 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7806 if (rc != MDB_SUCCESS)
7808 key.mv_data = (char *)key.mv_data + key.mv_size;
7811 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7812 srcnode = NODEPTR(psrc, i);
7813 if (i == 0 && IS_BRANCH(psrc)) {
7816 mdb_cursor_copy(csrc, &mn);
7817 /* must find the lowest key below src */
7818 rc = mdb_page_search_lowest(&mn);
7821 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7822 key.mv_size = mn.mc_db->md_pad;
7823 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7825 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7826 key.mv_size = NODEKSZ(s2);
7827 key.mv_data = NODEKEY(s2);
7830 key.mv_size = srcnode->mn_ksize;
7831 key.mv_data = NODEKEY(srcnode);
7834 data.mv_size = NODEDSZ(srcnode);
7835 data.mv_data = NODEDATA(srcnode);
7836 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7837 if (rc != MDB_SUCCESS)
7842 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7843 pdst->mp_pgno, NUMKEYS(pdst),
7844 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7846 /* Unlink the src page from parent and add to free list.
7849 mdb_node_del(csrc, 0);
7850 if (csrc->mc_ki[csrc->mc_top] == 0) {
7852 rc = mdb_update_key(csrc, &key);
7860 psrc = csrc->mc_pg[csrc->mc_top];
7861 /* If not operating on FreeDB, allow this page to be reused
7862 * in this txn. Otherwise just add to free list.
7864 rc = mdb_page_loose(csrc, psrc);
7868 csrc->mc_db->md_leaf_pages--;
7870 csrc->mc_db->md_branch_pages--;
7872 /* Adjust other cursors pointing to mp */
7873 MDB_cursor *m2, *m3;
7874 MDB_dbi dbi = csrc->mc_dbi;
7875 unsigned int top = csrc->mc_top;
7877 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7878 if (csrc->mc_flags & C_SUB)
7879 m3 = &m2->mc_xcursor->mx_cursor;
7882 if (m3 == csrc) continue;
7883 if (m3->mc_snum < csrc->mc_snum) continue;
7884 if (m3->mc_pg[top] == psrc) {
7885 m3->mc_pg[top] = pdst;
7886 m3->mc_ki[top] += nkeys;
7887 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
7888 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
7889 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
7895 unsigned int snum = cdst->mc_snum;
7896 uint16_t depth = cdst->mc_db->md_depth;
7897 mdb_cursor_pop(cdst);
7898 rc = mdb_rebalance(cdst);
7899 /* Did the tree height change? */
7900 if (depth != cdst->mc_db->md_depth)
7901 snum += cdst->mc_db->md_depth - depth;
7902 cdst->mc_snum = snum;
7903 cdst->mc_top = snum-1;
7908 /** Copy the contents of a cursor.
7909 * @param[in] csrc The cursor to copy from.
7910 * @param[out] cdst The cursor to copy to.
7913 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7917 cdst->mc_txn = csrc->mc_txn;
7918 cdst->mc_dbi = csrc->mc_dbi;
7919 cdst->mc_db = csrc->mc_db;
7920 cdst->mc_dbx = csrc->mc_dbx;
7921 cdst->mc_snum = csrc->mc_snum;
7922 cdst->mc_top = csrc->mc_top;
7923 cdst->mc_flags = csrc->mc_flags;
7925 for (i=0; i<csrc->mc_snum; i++) {
7926 cdst->mc_pg[i] = csrc->mc_pg[i];
7927 cdst->mc_ki[i] = csrc->mc_ki[i];
7931 /** Rebalance the tree after a delete operation.
7932 * @param[in] mc Cursor pointing to the page where rebalancing
7934 * @return 0 on success, non-zero on failure.
7937 mdb_rebalance(MDB_cursor *mc)
7941 unsigned int ptop, minkeys, thresh;
7945 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7950 thresh = FILL_THRESHOLD;
7952 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7953 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7954 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7955 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7957 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7958 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7959 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7960 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7964 if (mc->mc_snum < 2) {
7965 MDB_page *mp = mc->mc_pg[0];
7967 DPUTS("Can't rebalance a subpage, ignoring");
7970 if (NUMKEYS(mp) == 0) {
7971 DPUTS("tree is completely empty");
7972 mc->mc_db->md_root = P_INVALID;
7973 mc->mc_db->md_depth = 0;
7974 mc->mc_db->md_leaf_pages = 0;
7975 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7978 /* Adjust cursors pointing to mp */
7981 mc->mc_flags &= ~C_INITIALIZED;
7983 MDB_cursor *m2, *m3;
7984 MDB_dbi dbi = mc->mc_dbi;
7986 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7987 if (mc->mc_flags & C_SUB)
7988 m3 = &m2->mc_xcursor->mx_cursor;
7991 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
7993 if (m3->mc_pg[0] == mp) {
7996 m3->mc_flags &= ~C_INITIALIZED;
8000 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8002 DPUTS("collapsing root page!");
8003 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8006 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8007 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
8010 mc->mc_db->md_depth--;
8011 mc->mc_db->md_branch_pages--;
8012 mc->mc_ki[0] = mc->mc_ki[1];
8013 for (i = 1; i<mc->mc_db->md_depth; i++) {
8014 mc->mc_pg[i] = mc->mc_pg[i+1];
8015 mc->mc_ki[i] = mc->mc_ki[i+1];
8018 /* Adjust other cursors pointing to mp */
8019 MDB_cursor *m2, *m3;
8020 MDB_dbi dbi = mc->mc_dbi;
8022 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8023 if (mc->mc_flags & C_SUB)
8024 m3 = &m2->mc_xcursor->mx_cursor;
8027 if (m3 == mc) continue;
8028 if (!(m3->mc_flags & C_INITIALIZED))
8030 if (m3->mc_pg[0] == mp) {
8031 for (i=0; i<mc->mc_db->md_depth; i++) {
8032 m3->mc_pg[i] = m3->mc_pg[i+1];
8033 m3->mc_ki[i] = m3->mc_ki[i+1];
8041 DPUTS("root page doesn't need rebalancing");
8045 /* The parent (branch page) must have at least 2 pointers,
8046 * otherwise the tree is invalid.
8048 ptop = mc->mc_top-1;
8049 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8051 /* Leaf page fill factor is below the threshold.
8052 * Try to move keys from left or right neighbor, or
8053 * merge with a neighbor page.
8058 mdb_cursor_copy(mc, &mn);
8059 mn.mc_xcursor = NULL;
8061 oldki = mc->mc_ki[mc->mc_top];
8062 if (mc->mc_ki[ptop] == 0) {
8063 /* We're the leftmost leaf in our parent.
8065 DPUTS("reading right neighbor");
8067 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8068 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8071 mn.mc_ki[mn.mc_top] = 0;
8072 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8075 /* There is at least one neighbor to the left.
8077 DPUTS("reading left neighbor");
8079 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8080 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8083 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8084 mc->mc_ki[mc->mc_top] = 0;
8088 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8089 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8090 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8092 /* If the neighbor page is above threshold and has enough keys,
8093 * move one key from it. Otherwise we should try to merge them.
8094 * (A branch page must never have less than 2 keys.)
8096 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8097 rc = mdb_node_move(&mn, mc, fromleft);
8099 /* if we inserted on left, bump position up */
8104 rc = mdb_page_merge(&mn, mc);
8106 MDB_cursor dummy, *tracked;
8107 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8108 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8109 /* We want mdb_rebalance to find mn when doing fixups */
8110 CURSOR_TMP_TRACK(mc, mn, dummy, tracked);
8111 rc = mdb_page_merge(mc, &mn);
8112 CURSOR_TMP_UNTRACK(mc, tracked);
8113 mdb_cursor_copy(&mn, mc);
8115 mc->mc_flags &= ~C_EOF;
8117 mc->mc_ki[mc->mc_top] = oldki;
8121 /** Complete a delete operation started by #mdb_cursor_del(). */
8123 mdb_cursor_del0(MDB_cursor *mc)
8129 MDB_cursor *m2, *m3;
8130 MDB_dbi dbi = mc->mc_dbi;
8132 ki = mc->mc_ki[mc->mc_top];
8133 mp = mc->mc_pg[mc->mc_top];
8134 mdb_node_del(mc, mc->mc_db->md_pad);
8135 mc->mc_db->md_entries--;
8137 /* Adjust other cursors pointing to mp */
8138 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8139 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8140 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8142 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8144 if (m3->mc_pg[mc->mc_top] == mp) {
8145 if (m3->mc_ki[mc->mc_top] >= ki) {
8146 m3->mc_flags |= C_DEL;
8147 if (m3->mc_ki[mc->mc_top] > ki)
8148 m3->mc_ki[mc->mc_top]--;
8149 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8150 m3->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
8155 rc = mdb_rebalance(mc);
8157 if (rc == MDB_SUCCESS) {
8158 /* DB is totally empty now, just bail out.
8159 * Other cursors adjustments were already done
8160 * by mdb_rebalance and aren't needed here.
8165 mp = mc->mc_pg[mc->mc_top];
8166 nkeys = NUMKEYS(mp);
8168 /* Adjust other cursors pointing to mp */
8169 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8170 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8171 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8173 if (m3->mc_snum < mc->mc_snum)
8175 if (m3->mc_pg[mc->mc_top] == mp) {
8176 /* if m3 points past last node in page, find next sibling */
8177 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8178 rc = mdb_cursor_sibling(m3, 1);
8179 if (rc == MDB_NOTFOUND) {
8180 m3->mc_flags |= C_EOF;
8186 mc->mc_flags |= C_DEL;
8190 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8195 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8196 MDB_val *key, MDB_val *data)
8198 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8201 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8202 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8204 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8205 /* must ignore any data */
8209 return mdb_del0(txn, dbi, key, data, 0);
8213 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8214 MDB_val *key, MDB_val *data, unsigned flags)
8219 MDB_val rdata, *xdata;
8223 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8225 mdb_cursor_init(&mc, txn, dbi, &mx);
8234 flags |= MDB_NODUPDATA;
8236 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8238 /* let mdb_page_split know about this cursor if needed:
8239 * delete will trigger a rebalance; if it needs to move
8240 * a node from one page to another, it will have to
8241 * update the parent's separator key(s). If the new sepkey
8242 * is larger than the current one, the parent page may
8243 * run out of space, triggering a split. We need this
8244 * cursor to be consistent until the end of the rebalance.
8246 mc.mc_flags |= C_UNTRACK;
8247 mc.mc_next = txn->mt_cursors[dbi];
8248 txn->mt_cursors[dbi] = &mc;
8249 rc = mdb_cursor_del(&mc, flags);
8250 txn->mt_cursors[dbi] = mc.mc_next;
8255 /** Split a page and insert a new node.
8256 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8257 * The cursor will be updated to point to the actual page and index where
8258 * the node got inserted after the split.
8259 * @param[in] newkey The key for the newly inserted node.
8260 * @param[in] newdata The data for the newly inserted node.
8261 * @param[in] newpgno The page number, if the new node is a branch node.
8262 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8263 * @return 0 on success, non-zero on failure.
8266 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8267 unsigned int nflags)
8270 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8273 int i, j, split_indx, nkeys, pmax;
8274 MDB_env *env = mc->mc_txn->mt_env;
8276 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8277 MDB_page *copy = NULL;
8278 MDB_page *mp, *rp, *pp;
8283 mp = mc->mc_pg[mc->mc_top];
8284 newindx = mc->mc_ki[mc->mc_top];
8285 nkeys = NUMKEYS(mp);
8287 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8288 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8289 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8291 /* Create a right sibling. */
8292 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8294 rp->mp_pad = mp->mp_pad;
8295 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8297 /* Usually when splitting the root page, the cursor
8298 * height is 1. But when called from mdb_update_key,
8299 * the cursor height may be greater because it walks
8300 * up the stack while finding the branch slot to update.
8302 if (mc->mc_top < 1) {
8303 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8305 /* shift current top to make room for new parent */
8306 for (i=mc->mc_snum; i>0; i--) {
8307 mc->mc_pg[i] = mc->mc_pg[i-1];
8308 mc->mc_ki[i] = mc->mc_ki[i-1];
8312 mc->mc_db->md_root = pp->mp_pgno;
8313 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8314 new_root = mc->mc_db->md_depth++;
8316 /* Add left (implicit) pointer. */
8317 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8318 /* undo the pre-push */
8319 mc->mc_pg[0] = mc->mc_pg[1];
8320 mc->mc_ki[0] = mc->mc_ki[1];
8321 mc->mc_db->md_root = mp->mp_pgno;
8322 mc->mc_db->md_depth--;
8329 ptop = mc->mc_top-1;
8330 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8333 mdb_cursor_copy(mc, &mn);
8334 mn.mc_pg[mn.mc_top] = rp;
8335 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8337 if (nflags & MDB_APPEND) {
8338 mn.mc_ki[mn.mc_top] = 0;
8340 split_indx = newindx;
8344 split_indx = (nkeys+1) / 2;
8349 unsigned int lsize, rsize, ksize;
8350 /* Move half of the keys to the right sibling */
8351 x = mc->mc_ki[mc->mc_top] - split_indx;
8352 ksize = mc->mc_db->md_pad;
8353 split = LEAF2KEY(mp, split_indx, ksize);
8354 rsize = (nkeys - split_indx) * ksize;
8355 lsize = (nkeys - split_indx) * sizeof(indx_t);
8356 mp->mp_lower -= lsize;
8357 rp->mp_lower += lsize;
8358 mp->mp_upper += rsize - lsize;
8359 rp->mp_upper -= rsize - lsize;
8360 sepkey.mv_size = ksize;
8361 if (newindx == split_indx) {
8362 sepkey.mv_data = newkey->mv_data;
8364 sepkey.mv_data = split;
8367 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8368 memcpy(rp->mp_ptrs, split, rsize);
8369 sepkey.mv_data = rp->mp_ptrs;
8370 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8371 memcpy(ins, newkey->mv_data, ksize);
8372 mp->mp_lower += sizeof(indx_t);
8373 mp->mp_upper -= ksize - sizeof(indx_t);
8376 memcpy(rp->mp_ptrs, split, x * ksize);
8377 ins = LEAF2KEY(rp, x, ksize);
8378 memcpy(ins, newkey->mv_data, ksize);
8379 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8380 rp->mp_lower += sizeof(indx_t);
8381 rp->mp_upper -= ksize - sizeof(indx_t);
8382 mc->mc_ki[mc->mc_top] = x;
8385 int psize, nsize, k;
8386 /* Maximum free space in an empty page */
8387 pmax = env->me_psize - PAGEHDRSZ;
8389 nsize = mdb_leaf_size(env, newkey, newdata);
8391 nsize = mdb_branch_size(env, newkey);
8392 nsize = EVEN(nsize);
8394 /* grab a page to hold a temporary copy */
8395 copy = mdb_page_malloc(mc->mc_txn, 1);
8400 copy->mp_pgno = mp->mp_pgno;
8401 copy->mp_flags = mp->mp_flags;
8402 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8403 copy->mp_upper = env->me_psize - PAGEBASE;
8405 /* prepare to insert */
8406 for (i=0, j=0; i<nkeys; i++) {
8408 copy->mp_ptrs[j++] = 0;
8410 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8413 /* When items are relatively large the split point needs
8414 * to be checked, because being off-by-one will make the
8415 * difference between success or failure in mdb_node_add.
8417 * It's also relevant if a page happens to be laid out
8418 * such that one half of its nodes are all "small" and
8419 * the other half of its nodes are "large." If the new
8420 * item is also "large" and falls on the half with
8421 * "large" nodes, it also may not fit.
8423 * As a final tweak, if the new item goes on the last
8424 * spot on the page (and thus, onto the new page), bias
8425 * the split so the new page is emptier than the old page.
8426 * This yields better packing during sequential inserts.
8428 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8429 /* Find split point */
8431 if (newindx <= split_indx || newindx >= nkeys) {
8433 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8438 for (; i!=k; i+=j) {
8443 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8444 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8446 if (F_ISSET(node->mn_flags, F_BIGDATA))
8447 psize += sizeof(pgno_t);
8449 psize += NODEDSZ(node);
8451 psize = EVEN(psize);
8453 if (psize > pmax || i == k-j) {
8454 split_indx = i + (j<0);
8459 if (split_indx == newindx) {
8460 sepkey.mv_size = newkey->mv_size;
8461 sepkey.mv_data = newkey->mv_data;
8463 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8464 sepkey.mv_size = node->mn_ksize;
8465 sepkey.mv_data = NODEKEY(node);
8470 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8472 /* Copy separator key to the parent.
8474 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8475 int snum = mc->mc_snum;
8476 MDB_cursor dummy, *tracked;
8480 /* We want other splits to find mn when doing fixups */
8481 CURSOR_TMP_TRACK(mc, mn, dummy, tracked);
8482 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8483 CURSOR_TMP_UNTRACK(mc, tracked);
8488 if (mc->mc_snum > snum) {
8491 /* Right page might now have changed parent.
8492 * Check if left page also changed parent.
8494 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8495 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8496 for (i=0; i<ptop; i++) {
8497 mc->mc_pg[i] = mn.mc_pg[i];
8498 mc->mc_ki[i] = mn.mc_ki[i];
8500 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8501 if (mn.mc_ki[ptop]) {
8502 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8504 /* find right page's left sibling */
8505 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8506 mdb_cursor_sibling(mc, 0);
8511 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8514 if (rc != MDB_SUCCESS) {
8517 if (nflags & MDB_APPEND) {
8518 mc->mc_pg[mc->mc_top] = rp;
8519 mc->mc_ki[mc->mc_top] = 0;
8520 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8523 for (i=0; i<mc->mc_top; i++)
8524 mc->mc_ki[i] = mn.mc_ki[i];
8525 } else if (!IS_LEAF2(mp)) {
8527 mc->mc_pg[mc->mc_top] = rp;
8532 rkey.mv_data = newkey->mv_data;
8533 rkey.mv_size = newkey->mv_size;
8539 /* Update index for the new key. */
8540 mc->mc_ki[mc->mc_top] = j;
8542 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8543 rkey.mv_data = NODEKEY(node);
8544 rkey.mv_size = node->mn_ksize;
8546 xdata.mv_data = NODEDATA(node);
8547 xdata.mv_size = NODEDSZ(node);
8550 pgno = NODEPGNO(node);
8551 flags = node->mn_flags;
8554 if (!IS_LEAF(mp) && j == 0) {
8555 /* First branch index doesn't need key data. */
8559 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8565 mc->mc_pg[mc->mc_top] = copy;
8570 } while (i != split_indx);
8572 nkeys = NUMKEYS(copy);
8573 for (i=0; i<nkeys; i++)
8574 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8575 mp->mp_lower = copy->mp_lower;
8576 mp->mp_upper = copy->mp_upper;
8577 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8578 env->me_psize - copy->mp_upper - PAGEBASE);
8580 /* reset back to original page */
8581 if (newindx < split_indx) {
8582 mc->mc_pg[mc->mc_top] = mp;
8584 mc->mc_pg[mc->mc_top] = rp;
8586 /* Make sure mc_ki is still valid.
8588 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8589 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8590 for (i=0; i<=ptop; i++) {
8591 mc->mc_pg[i] = mn.mc_pg[i];
8592 mc->mc_ki[i] = mn.mc_ki[i];
8596 if (nflags & MDB_RESERVE) {
8597 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8598 if (!(node->mn_flags & F_BIGDATA))
8599 newdata->mv_data = NODEDATA(node);
8602 if (newindx >= split_indx) {
8603 mc->mc_pg[mc->mc_top] = rp;
8605 /* Make sure mc_ki is still valid.
8607 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8608 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8609 for (i=0; i<=ptop; i++) {
8610 mc->mc_pg[i] = mn.mc_pg[i];
8611 mc->mc_ki[i] = mn.mc_ki[i];
8618 /* Adjust other cursors pointing to mp */
8619 MDB_cursor *m2, *m3;
8620 MDB_dbi dbi = mc->mc_dbi;
8621 nkeys = NUMKEYS(mp);
8623 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8624 if (mc->mc_flags & C_SUB)
8625 m3 = &m2->mc_xcursor->mx_cursor;
8630 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8634 /* sub cursors may be on different DB */
8635 if (m3->mc_pg[0] != mp)
8638 for (k=new_root; k>=0; k--) {
8639 m3->mc_ki[k+1] = m3->mc_ki[k];
8640 m3->mc_pg[k+1] = m3->mc_pg[k];
8642 if (m3->mc_ki[0] > nkeys) {
8647 m3->mc_pg[0] = mc->mc_pg[0];
8651 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8652 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8653 m3->mc_ki[mc->mc_top]++;
8654 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8655 m3->mc_pg[mc->mc_top] = rp;
8656 m3->mc_ki[mc->mc_top] -= nkeys;
8657 for (i=0; i<mc->mc_top; i++) {
8658 m3->mc_ki[i] = mn.mc_ki[i];
8659 m3->mc_pg[i] = mn.mc_pg[i];
8662 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8663 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8668 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8671 if (copy) /* tmp page */
8672 mdb_page_free(env, copy);
8674 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8679 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8680 MDB_val *key, MDB_val *data, unsigned int flags)
8685 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8688 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8691 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8692 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8694 mdb_cursor_init(&mc, txn, dbi, &mx);
8695 return mdb_cursor_put(&mc, key, data, flags);
8699 #define MDB_WBUF (1024*1024)
8702 /** State needed for a compacting copy. */
8703 typedef struct mdb_copy {
8704 pthread_mutex_t mc_mutex;
8705 pthread_cond_t mc_cond;
8712 pgno_t mc_next_pgno;
8715 volatile int mc_new;
8720 /** Dedicated writer thread for compacting copy. */
8721 static THREAD_RET ESECT CALL_CONV
8722 mdb_env_copythr(void *arg)
8726 int toggle = 0, wsize, rc;
8729 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8732 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8735 pthread_mutex_lock(&my->mc_mutex);
8737 pthread_cond_signal(&my->mc_cond);
8740 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8741 if (my->mc_new < 0) {
8746 wsize = my->mc_wlen[toggle];
8747 ptr = my->mc_wbuf[toggle];
8750 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8754 } else if (len > 0) {
8768 /* If there's an overflow page tail, write it too */
8769 if (my->mc_olen[toggle]) {
8770 wsize = my->mc_olen[toggle];
8771 ptr = my->mc_over[toggle];
8772 my->mc_olen[toggle] = 0;
8775 my->mc_wlen[toggle] = 0;
8777 pthread_cond_signal(&my->mc_cond);
8779 pthread_cond_signal(&my->mc_cond);
8780 pthread_mutex_unlock(&my->mc_mutex);
8781 return (THREAD_RET)0;
8785 /** Tell the writer thread there's a buffer ready to write */
8787 mdb_env_cthr_toggle(mdb_copy *my, int st)
8789 int toggle = my->mc_toggle ^ 1;
8790 pthread_mutex_lock(&my->mc_mutex);
8791 if (my->mc_status) {
8792 pthread_mutex_unlock(&my->mc_mutex);
8793 return my->mc_status;
8795 while (my->mc_new == 1)
8796 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8798 my->mc_toggle = toggle;
8799 pthread_cond_signal(&my->mc_cond);
8800 pthread_mutex_unlock(&my->mc_mutex);
8804 /** Depth-first tree traversal for compacting copy. */
8806 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8809 MDB_txn *txn = my->mc_txn;
8811 MDB_page *mo, *mp, *leaf;
8816 /* Empty DB, nothing to do */
8817 if (*pg == P_INVALID)
8824 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8827 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8831 /* Make cursor pages writable */
8832 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8836 for (i=0; i<mc.mc_top; i++) {
8837 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8838 mc.mc_pg[i] = (MDB_page *)ptr;
8839 ptr += my->mc_env->me_psize;
8842 /* This is writable space for a leaf page. Usually not needed. */
8843 leaf = (MDB_page *)ptr;
8845 toggle = my->mc_toggle;
8846 while (mc.mc_snum > 0) {
8848 mp = mc.mc_pg[mc.mc_top];
8852 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8853 for (i=0; i<n; i++) {
8854 ni = NODEPTR(mp, i);
8855 if (ni->mn_flags & F_BIGDATA) {
8859 /* Need writable leaf */
8861 mc.mc_pg[mc.mc_top] = leaf;
8862 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8864 ni = NODEPTR(mp, i);
8867 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8868 rc = mdb_page_get(txn, pg, &omp, NULL);
8871 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8872 rc = mdb_env_cthr_toggle(my, 1);
8875 toggle = my->mc_toggle;
8877 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8878 memcpy(mo, omp, my->mc_env->me_psize);
8879 mo->mp_pgno = my->mc_next_pgno;
8880 my->mc_next_pgno += omp->mp_pages;
8881 my->mc_wlen[toggle] += my->mc_env->me_psize;
8882 if (omp->mp_pages > 1) {
8883 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8884 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8885 rc = mdb_env_cthr_toggle(my, 1);
8888 toggle = my->mc_toggle;
8890 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8891 } else if (ni->mn_flags & F_SUBDATA) {
8894 /* Need writable leaf */
8896 mc.mc_pg[mc.mc_top] = leaf;
8897 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8899 ni = NODEPTR(mp, i);
8902 memcpy(&db, NODEDATA(ni), sizeof(db));
8903 my->mc_toggle = toggle;
8904 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8907 toggle = my->mc_toggle;
8908 memcpy(NODEDATA(ni), &db, sizeof(db));
8913 mc.mc_ki[mc.mc_top]++;
8914 if (mc.mc_ki[mc.mc_top] < n) {
8917 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8919 rc = mdb_page_get(txn, pg, &mp, NULL);
8924 mc.mc_ki[mc.mc_top] = 0;
8925 if (IS_BRANCH(mp)) {
8926 /* Whenever we advance to a sibling branch page,
8927 * we must proceed all the way down to its first leaf.
8929 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8932 mc.mc_pg[mc.mc_top] = mp;
8936 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8937 rc = mdb_env_cthr_toggle(my, 1);
8940 toggle = my->mc_toggle;
8942 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8943 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8944 mo->mp_pgno = my->mc_next_pgno++;
8945 my->mc_wlen[toggle] += my->mc_env->me_psize;
8947 /* Update parent if there is one */
8948 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8949 SETPGNO(ni, mo->mp_pgno);
8950 mdb_cursor_pop(&mc);
8952 /* Otherwise we're done */
8962 /** Copy environment with compaction. */
8964 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8969 MDB_txn *txn = NULL;
8974 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8975 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8976 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8977 if (my.mc_wbuf[0] == NULL)
8980 pthread_mutex_init(&my.mc_mutex, NULL);
8981 pthread_cond_init(&my.mc_cond, NULL);
8982 #ifdef HAVE_MEMALIGN
8983 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8984 if (my.mc_wbuf[0] == NULL)
8987 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8992 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8993 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8998 my.mc_next_pgno = NUM_METAS;
9004 THREAD_CREATE(thr, mdb_env_copythr, &my);
9006 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9010 mp = (MDB_page *)my.mc_wbuf[0];
9011 memset(mp, 0, NUM_METAS * env->me_psize);
9013 mp->mp_flags = P_META;
9014 mm = (MDB_meta *)METADATA(mp);
9015 mdb_env_init_meta0(env, mm);
9016 mm->mm_address = env->me_metas[0]->mm_address;
9018 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9020 mp->mp_flags = P_META;
9021 *(MDB_meta *)METADATA(mp) = *mm;
9022 mm = (MDB_meta *)METADATA(mp);
9024 /* Count the number of free pages, subtract from lastpg to find
9025 * number of active pages
9028 MDB_ID freecount = 0;
9031 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9032 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9033 freecount += *(MDB_ID *)data.mv_data;
9034 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9035 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9036 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9038 /* Set metapage 1 */
9039 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
9040 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9041 if (mm->mm_last_pg > NUM_METAS-1) {
9042 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
9045 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
9048 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9050 pthread_mutex_lock(&my.mc_mutex);
9052 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9053 pthread_mutex_unlock(&my.mc_mutex);
9054 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
9055 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
9056 rc = mdb_env_cthr_toggle(&my, 1);
9057 mdb_env_cthr_toggle(&my, -1);
9058 pthread_mutex_lock(&my.mc_mutex);
9060 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9061 pthread_mutex_unlock(&my.mc_mutex);
9066 CloseHandle(my.mc_cond);
9067 CloseHandle(my.mc_mutex);
9068 _aligned_free(my.mc_wbuf[0]);
9070 pthread_cond_destroy(&my.mc_cond);
9071 pthread_mutex_destroy(&my.mc_mutex);
9072 free(my.mc_wbuf[0]);
9077 /** Copy environment as-is. */
9079 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9081 MDB_txn *txn = NULL;
9082 mdb_mutexref_t wmutex = NULL;
9088 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9092 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9095 /* Do the lock/unlock of the reader mutex before starting the
9096 * write txn. Otherwise other read txns could block writers.
9098 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9103 /* We must start the actual read txn after blocking writers */
9104 mdb_txn_end(txn, MDB_END_RESET_TMP);
9106 /* Temporarily block writers until we snapshot the meta pages */
9107 wmutex = env->me_wmutex;
9108 if (LOCK_MUTEX(rc, env, wmutex))
9111 rc = mdb_txn_renew0(txn);
9113 UNLOCK_MUTEX(wmutex);
9118 wsize = env->me_psize * NUM_METAS;
9122 DO_WRITE(rc, fd, ptr, w2, len);
9126 } else if (len > 0) {
9132 /* Non-blocking or async handles are not supported */
9138 UNLOCK_MUTEX(wmutex);
9143 w2 = txn->mt_next_pgno * env->me_psize;
9146 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9153 if (wsize > MAX_WRITE)
9157 DO_WRITE(rc, fd, ptr, w2, len);
9161 } else if (len > 0) {
9178 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9180 if (flags & MDB_CP_COMPACT)
9181 return mdb_env_copyfd1(env, fd);
9183 return mdb_env_copyfd0(env, fd);
9187 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9189 return mdb_env_copyfd2(env, fd, 0);
9193 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9197 HANDLE newfd = INVALID_HANDLE_VALUE;
9202 if (env->me_flags & MDB_NOSUBDIR) {
9203 lpath = (char *)path;
9206 len += sizeof(DATANAME);
9207 lpath = malloc(len);
9210 sprintf(lpath, "%s" DATANAME, path);
9213 /* The destination path must exist, but the destination file must not.
9214 * We don't want the OS to cache the writes, since the source data is
9215 * already in the OS cache.
9218 utf8_to_utf16(lpath, -1, &wpath, NULL);
9219 newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9220 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9223 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9225 if (newfd == INVALID_HANDLE_VALUE) {
9230 if (env->me_psize >= env->me_os_psize) {
9232 /* Set O_DIRECT if the file system supports it */
9233 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9234 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9236 #ifdef F_NOCACHE /* __APPLE__ */
9237 rc = fcntl(newfd, F_NOCACHE, 1);
9245 rc = mdb_env_copyfd2(env, newfd, flags);
9248 if (!(env->me_flags & MDB_NOSUBDIR))
9250 if (newfd != INVALID_HANDLE_VALUE)
9251 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9258 mdb_env_copy(MDB_env *env, const char *path)
9260 return mdb_env_copy2(env, path, 0);
9264 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9266 if (flag & ~CHANGEABLE)
9269 env->me_flags |= flag;
9271 env->me_flags &= ~flag;
9276 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9281 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9286 mdb_env_set_userctx(MDB_env *env, void *ctx)
9290 env->me_userctx = ctx;
9295 mdb_env_get_userctx(MDB_env *env)
9297 return env ? env->me_userctx : NULL;
9301 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9306 env->me_assert_func = func;
9312 mdb_env_get_path(MDB_env *env, const char **arg)
9317 *arg = env->me_path;
9322 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9331 /** Common code for #mdb_stat() and #mdb_env_stat().
9332 * @param[in] env the environment to operate in.
9333 * @param[in] db the #MDB_db record containing the stats to return.
9334 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9335 * @return 0, this function always succeeds.
9338 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9340 arg->ms_psize = env->me_psize;
9341 arg->ms_depth = db->md_depth;
9342 arg->ms_branch_pages = db->md_branch_pages;
9343 arg->ms_leaf_pages = db->md_leaf_pages;
9344 arg->ms_overflow_pages = db->md_overflow_pages;
9345 arg->ms_entries = db->md_entries;
9351 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9355 if (env == NULL || arg == NULL)
9358 meta = mdb_env_pick_meta(env);
9360 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9364 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9368 if (env == NULL || arg == NULL)
9371 meta = mdb_env_pick_meta(env);
9372 arg->me_mapaddr = meta->mm_address;
9373 arg->me_last_pgno = meta->mm_last_pg;
9374 arg->me_last_txnid = meta->mm_txnid;
9376 arg->me_mapsize = env->me_mapsize;
9377 arg->me_maxreaders = env->me_maxreaders;
9378 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9382 /** Set the default comparison functions for a database.
9383 * Called immediately after a database is opened to set the defaults.
9384 * The user can then override them with #mdb_set_compare() or
9385 * #mdb_set_dupsort().
9386 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9387 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9390 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9392 uint16_t f = txn->mt_dbs[dbi].md_flags;
9394 txn->mt_dbxs[dbi].md_cmp =
9395 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9396 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9398 txn->mt_dbxs[dbi].md_dcmp =
9399 !(f & MDB_DUPSORT) ? 0 :
9400 ((f & MDB_INTEGERDUP)
9401 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9402 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9405 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9411 int rc, dbflag, exact;
9412 unsigned int unused = 0, seq;
9415 if (flags & ~VALID_FLAGS)
9417 if (txn->mt_flags & MDB_TXN_BLOCKED)
9423 if (flags & PERSISTENT_FLAGS) {
9424 uint16_t f2 = flags & PERSISTENT_FLAGS;
9425 /* make sure flag changes get committed */
9426 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9427 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9428 txn->mt_flags |= MDB_TXN_DIRTY;
9431 mdb_default_cmp(txn, MAIN_DBI);
9435 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9436 mdb_default_cmp(txn, MAIN_DBI);
9439 /* Is the DB already open? */
9441 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9442 if (!txn->mt_dbxs[i].md_name.mv_size) {
9443 /* Remember this free slot */
9444 if (!unused) unused = i;
9447 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9448 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9454 /* If no free slot and max hit, fail */
9455 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9456 return MDB_DBS_FULL;
9458 /* Cannot mix named databases with some mainDB flags */
9459 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9460 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9462 /* Find the DB info */
9463 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9466 key.mv_data = (void *)name;
9467 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9468 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9469 if (rc == MDB_SUCCESS) {
9470 /* make sure this is actually a DB */
9471 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9472 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9473 return MDB_INCOMPATIBLE;
9474 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9475 /* Create if requested */
9476 data.mv_size = sizeof(MDB_db);
9477 data.mv_data = &dummy;
9478 memset(&dummy, 0, sizeof(dummy));
9479 dummy.md_root = P_INVALID;
9480 dummy.md_flags = flags & PERSISTENT_FLAGS;
9481 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9485 /* OK, got info, add to table */
9486 if (rc == MDB_SUCCESS) {
9487 unsigned int slot = unused ? unused : txn->mt_numdbs;
9488 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9489 txn->mt_dbxs[slot].md_name.mv_size = len;
9490 txn->mt_dbxs[slot].md_rel = NULL;
9491 txn->mt_dbflags[slot] = dbflag;
9492 /* txn-> and env-> are the same in read txns, use
9493 * tmp variable to avoid undefined assignment
9495 seq = ++txn->mt_env->me_dbiseqs[slot];
9496 txn->mt_dbiseqs[slot] = seq;
9498 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9500 mdb_default_cmp(txn, slot);
9510 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9512 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9515 if (txn->mt_flags & MDB_TXN_BLOCKED)
9518 if (txn->mt_dbflags[dbi] & DB_STALE) {
9521 /* Stale, must read the DB's root. cursor_init does it for us. */
9522 mdb_cursor_init(&mc, txn, dbi, &mx);
9524 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9527 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9530 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9532 ptr = env->me_dbxs[dbi].md_name.mv_data;
9533 /* If there was no name, this was already closed */
9535 env->me_dbxs[dbi].md_name.mv_data = NULL;
9536 env->me_dbxs[dbi].md_name.mv_size = 0;
9537 env->me_dbflags[dbi] = 0;
9538 env->me_dbiseqs[dbi]++;
9543 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9545 /* We could return the flags for the FREE_DBI too but what's the point? */
9546 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9548 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9552 /** Add all the DB's pages to the free list.
9553 * @param[in] mc Cursor on the DB to free.
9554 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9555 * @return 0 on success, non-zero on failure.
9558 mdb_drop0(MDB_cursor *mc, int subs)
9562 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9563 if (rc == MDB_SUCCESS) {
9564 MDB_txn *txn = mc->mc_txn;
9569 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9570 * This also avoids any P_LEAF2 pages, which have no nodes.
9572 if (mc->mc_flags & C_SUB)
9575 mdb_cursor_copy(mc, &mx);
9576 while (mc->mc_snum > 0) {
9577 MDB_page *mp = mc->mc_pg[mc->mc_top];
9578 unsigned n = NUMKEYS(mp);
9580 for (i=0; i<n; i++) {
9581 ni = NODEPTR(mp, i);
9582 if (ni->mn_flags & F_BIGDATA) {
9585 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9586 rc = mdb_page_get(txn, pg, &omp, NULL);
9589 mdb_cassert(mc, IS_OVERFLOW(omp));
9590 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9594 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9595 mdb_xcursor_init1(mc, ni);
9596 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9602 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9604 for (i=0; i<n; i++) {
9606 ni = NODEPTR(mp, i);
9609 mdb_midl_xappend(txn->mt_free_pgs, pg);
9614 mc->mc_ki[mc->mc_top] = i;
9615 rc = mdb_cursor_sibling(mc, 1);
9617 if (rc != MDB_NOTFOUND)
9619 /* no more siblings, go back to beginning
9620 * of previous level.
9624 for (i=1; i<mc->mc_snum; i++) {
9626 mc->mc_pg[i] = mx.mc_pg[i];
9631 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9634 txn->mt_flags |= MDB_TXN_ERROR;
9635 } else if (rc == MDB_NOTFOUND) {
9641 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9643 MDB_cursor *mc, *m2;
9646 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9649 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9652 if (TXN_DBI_CHANGED(txn, dbi))
9655 rc = mdb_cursor_open(txn, dbi, &mc);
9659 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9660 /* Invalidate the dropped DB's cursors */
9661 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9662 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9666 /* Can't delete the main DB */
9667 if (del && dbi >= CORE_DBS) {
9668 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9670 txn->mt_dbflags[dbi] = DB_STALE;
9671 mdb_dbi_close(txn->mt_env, dbi);
9673 txn->mt_flags |= MDB_TXN_ERROR;
9676 /* reset the DB record, mark it dirty */
9677 txn->mt_dbflags[dbi] |= DB_DIRTY;
9678 txn->mt_dbs[dbi].md_depth = 0;
9679 txn->mt_dbs[dbi].md_branch_pages = 0;
9680 txn->mt_dbs[dbi].md_leaf_pages = 0;
9681 txn->mt_dbs[dbi].md_overflow_pages = 0;
9682 txn->mt_dbs[dbi].md_entries = 0;
9683 txn->mt_dbs[dbi].md_root = P_INVALID;
9685 txn->mt_flags |= MDB_TXN_DIRTY;
9688 mdb_cursor_close(mc);
9692 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9694 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9697 txn->mt_dbxs[dbi].md_cmp = cmp;
9701 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9703 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9706 txn->mt_dbxs[dbi].md_dcmp = cmp;
9710 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9712 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9715 txn->mt_dbxs[dbi].md_rel = rel;
9719 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9721 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9724 txn->mt_dbxs[dbi].md_relctx = ctx;
9729 mdb_env_get_maxkeysize(MDB_env *env)
9731 return ENV_MAXKEY(env);
9735 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9737 unsigned int i, rdrs;
9740 int rc = 0, first = 1;
9744 if (!env->me_txns) {
9745 return func("(no reader locks)\n", ctx);
9747 rdrs = env->me_txns->mti_numreaders;
9748 mr = env->me_txns->mti_readers;
9749 for (i=0; i<rdrs; i++) {
9751 txnid_t txnid = mr[i].mr_txnid;
9752 sprintf(buf, txnid == (txnid_t)-1 ?
9753 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9754 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9757 rc = func(" pid thread txnid\n", ctx);
9761 rc = func(buf, ctx);
9767 rc = func("(no active readers)\n", ctx);
9772 /** Insert pid into list if not already present.
9773 * return -1 if already present.
9776 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9778 /* binary search of pid in list */
9780 unsigned cursor = 1;
9782 unsigned n = ids[0];
9785 unsigned pivot = n >> 1;
9786 cursor = base + pivot + 1;
9787 val = pid - ids[cursor];
9792 } else if ( val > 0 ) {
9797 /* found, so it's a duplicate */
9806 for (n = ids[0]; n > cursor; n--)
9813 mdb_reader_check(MDB_env *env, int *dead)
9819 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9822 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9824 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9826 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9827 unsigned int i, j, rdrs;
9829 MDB_PID_T *pids, pid;
9830 int rc = MDB_SUCCESS, count = 0;
9832 rdrs = env->me_txns->mti_numreaders;
9833 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9837 mr = env->me_txns->mti_readers;
9838 for (i=0; i<rdrs; i++) {
9840 if (pid && pid != env->me_pid) {
9841 if (mdb_pid_insert(pids, pid) == 0) {
9842 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9843 /* Stale reader found */
9846 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9847 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9849 rdrs = 0; /* the above checked all readers */
9851 /* Recheck, a new process may have reused pid */
9852 if (mdb_reader_pid(env, Pidcheck, pid))
9857 if (mr[j].mr_pid == pid) {
9858 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9859 (unsigned) pid, mr[j].mr_txnid));
9864 UNLOCK_MUTEX(rmutex);
9875 #ifdef MDB_ROBUST_SUPPORTED
9876 /** Handle #LOCK_MUTEX0() failure.
9877 * Try to repair the lock file if the mutex owner died.
9878 * @param[in] env the environment handle
9879 * @param[in] mutex LOCK_MUTEX0() mutex
9880 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9881 * @return 0 on success with the mutex locked, or an error code on failure.
9884 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9889 if (rc == MDB_OWNERDEAD) {
9890 /* We own the mutex. Clean up after dead previous owner. */
9892 rlocked = (mutex == env->me_rmutex);
9894 /* Keep mti_txnid updated, otherwise next writer can
9895 * overwrite data which latest meta page refers to.
9897 meta = mdb_env_pick_meta(env);
9898 env->me_txns->mti_txnid = meta->mm_txnid;
9899 /* env is hosed if the dead thread was ours */
9901 env->me_flags |= MDB_FATAL_ERROR;
9906 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9907 (rc ? "this process' env is hosed" : "recovering")));
9908 rc2 = mdb_reader_check0(env, rlocked, NULL);
9910 rc2 = mdb_mutex_consistent(mutex);
9911 if (rc || (rc = rc2)) {
9912 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9913 UNLOCK_MUTEX(mutex);
9919 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9924 #endif /* MDB_ROBUST_SUPPORTED */
9928 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
9932 need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
9937 result = malloc(sizeof(wchar_t) * need);
9938 MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
9944 #endif /* defined(_WIN32) */