2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2016 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #if defined(__WIN64__)
39 #define _FILE_OFFSET_BITS 64
44 #include <wchar.h> /* get wcscpy() */
46 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
47 * as int64 which is wrong. MSVC doesn't define it at all, so just
51 #define MDB_THR_T DWORD
52 #include <sys/types.h>
55 # include <sys/param.h>
57 # define LITTLE_ENDIAN 1234
58 # define BIG_ENDIAN 4321
59 # define BYTE_ORDER LITTLE_ENDIAN
61 # define SSIZE_MAX INT_MAX
65 #include <sys/types.h>
67 #define MDB_PID_T pid_t
68 #define MDB_THR_T pthread_t
69 #include <sys/param.h>
72 #ifdef HAVE_SYS_FILE_H
78 #if defined(__mips) && defined(__linux)
79 /* MIPS has cache coherency issues, requires explicit cache control */
80 #include <asm/cachectl.h>
81 extern int cacheflush(char *addr, int nbytes, int cache);
82 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
84 #define CACHEFLUSH(addr, bytes, cache)
87 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
88 /** fdatasync is broken on ext3/ext4fs on older kernels, see
89 * description in #mdb_env_open2 comments. You can safely
90 * define MDB_FDATASYNC_WORKS if this code will only be run
91 * on kernels 3.6 and newer.
93 #define BROKEN_FDATASYNC
107 typedef SSIZE_T ssize_t;
112 #if defined(__sun) || defined(ANDROID)
113 /* Most platforms have posix_memalign, older may only have memalign */
114 #define HAVE_MEMALIGN 1
118 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
119 #include <netinet/in.h>
120 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
123 #if defined(__APPLE__) || defined (BSD)
124 # define MDB_USE_POSIX_SEM 1
125 # define MDB_FDATASYNC fsync
126 #elif defined(ANDROID)
127 # define MDB_FDATASYNC fsync
132 #ifdef MDB_USE_POSIX_SEM
133 # define MDB_USE_HASH 1
134 #include <semaphore.h>
136 #define MDB_USE_POSIX_MUTEX 1
140 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \
141 + defined(MDB_USE_POSIX_MUTEX) != 1
142 # error "Ambiguous shared-lock implementation"
146 #include <valgrind/memcheck.h>
147 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
148 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
149 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
150 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
151 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
153 #define VGMEMP_CREATE(h,r,z)
154 #define VGMEMP_ALLOC(h,a,s)
155 #define VGMEMP_FREE(h,a)
156 #define VGMEMP_DESTROY(h)
157 #define VGMEMP_DEFINED(a,s)
161 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
162 /* Solaris just defines one or the other */
163 # define LITTLE_ENDIAN 1234
164 # define BIG_ENDIAN 4321
165 # ifdef _LITTLE_ENDIAN
166 # define BYTE_ORDER LITTLE_ENDIAN
168 # define BYTE_ORDER BIG_ENDIAN
171 # define BYTE_ORDER __BYTE_ORDER
175 #ifndef LITTLE_ENDIAN
176 #define LITTLE_ENDIAN __LITTLE_ENDIAN
179 #define BIG_ENDIAN __BIG_ENDIAN
182 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
183 #define MISALIGNED_OK 1
189 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
190 # error "Unknown or unsupported endianness (BYTE_ORDER)"
191 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
192 # error "Two's complement, reasonably sized integer types, please"
196 /** Put infrequently used env functions in separate section */
198 # define ESECT __attribute__ ((section("__TEXT,text_env")))
200 # define ESECT __attribute__ ((section("text_env")))
207 #define CALL_CONV WINAPI
212 /** @defgroup internal LMDB Internals
215 /** @defgroup compat Compatibility Macros
216 * A bunch of macros to minimize the amount of platform-specific ifdefs
217 * needed throughout the rest of the code. When the features this library
218 * needs are similar enough to POSIX to be hidden in a one-or-two line
219 * replacement, this macro approach is used.
223 /** Features under development */
228 /** Wrapper around __func__, which is a C99 feature */
229 #if __STDC_VERSION__ >= 199901L
230 # define mdb_func_ __func__
231 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
232 # define mdb_func_ __FUNCTION__
234 /* If a debug message says <mdb_unknown>(), update the #if statements above */
235 # define mdb_func_ "<mdb_unknown>"
238 /* Internal error codes, not exposed outside liblmdb */
239 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
241 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
242 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
243 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
247 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
249 /** Some platforms define the EOWNERDEAD error code
250 * even though they don't support Robust Mutexes.
251 * Compile with -DMDB_USE_ROBUST=0, or use some other
252 * mechanism like -DMDB_USE_POSIX_SEM instead of
253 * -DMDB_USE_POSIX_MUTEX.
254 * (Posix semaphores are not robust.)
256 #ifndef MDB_USE_ROBUST
257 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
258 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
259 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
260 # define MDB_USE_ROBUST 0
262 # define MDB_USE_ROBUST 1
264 #endif /* !MDB_USE_ROBUST */
266 #if defined(MDB_USE_POSIX_MUTEX) && (MDB_USE_ROBUST)
267 /* glibc < 2.12 only provided _np API */
268 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
269 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
270 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
271 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
272 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
274 #endif /* MDB_USE_POSIX_MUTEX && MDB_USE_ROBUST */
276 #if defined(MDB_OWNERDEAD) && (MDB_USE_ROBUST)
277 #define MDB_ROBUST_SUPPORTED 1
281 #define MDB_USE_HASH 1
282 #define MDB_PIDLOCK 0
283 #define THREAD_RET DWORD
284 #define pthread_t HANDLE
285 #define pthread_mutex_t HANDLE
286 #define pthread_cond_t HANDLE
287 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
288 #define pthread_key_t DWORD
289 #define pthread_self() GetCurrentThreadId()
290 #define pthread_key_create(x,y) \
291 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
292 #define pthread_key_delete(x) TlsFree(x)
293 #define pthread_getspecific(x) TlsGetValue(x)
294 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
295 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
296 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
297 #define pthread_cond_signal(x) SetEvent(*x)
298 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
299 #define THREAD_CREATE(thr,start,arg) \
300 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
301 #define THREAD_FINISH(thr) \
302 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
303 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
304 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
305 #define mdb_mutex_consistent(mutex) 0
306 #define getpid() GetCurrentProcessId()
307 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
308 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
309 #define ErrCode() GetLastError()
310 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
311 #define close(fd) (CloseHandle(fd) ? 0 : -1)
312 #define munmap(ptr,len) UnmapViewOfFile(ptr)
313 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
314 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
316 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
320 #define THREAD_RET void *
321 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
322 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
323 #define Z "z" /**< printf format modifier for size_t */
325 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
326 #define MDB_PIDLOCK 1
328 #ifdef MDB_USE_POSIX_SEM
330 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
331 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
332 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
335 mdb_sem_wait(sem_t *sem)
338 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
342 #else /* MDB_USE_POSIX_MUTEX: */
343 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
344 * local variables keep it (mdb_mutexref_t).
346 * When #mdb_mutexref_t is a pointer declaration and #mdb_mutex_t is
347 * not, then it is array[size 1] so it can be assigned to a pointer.
350 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
352 /** Lock the reader or writer mutex.
353 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
355 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
356 /** Unlock the reader or writer mutex.
358 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
359 /** Mark mutex-protected data as repaired, after death of previous owner.
361 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
362 #endif /* MDB_USE_POSIX_SEM */
364 /** Get the error code for the last failed system function.
366 #define ErrCode() errno
368 /** An abstraction for a file handle.
369 * On POSIX systems file handles are small integers. On Windows
370 * they're opaque pointers.
374 /** A value for an invalid file handle.
375 * Mainly used to initialize file variables and signify that they are
378 #define INVALID_HANDLE_VALUE (-1)
380 /** Get the size of a memory page for the system.
381 * This is the basic size that the platform's memory manager uses, and is
382 * fundamental to the use of memory-mapped files.
384 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
387 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
390 #define MNAME_LEN (sizeof(pthread_mutex_t))
395 #ifdef MDB_ROBUST_SUPPORTED
396 /** Lock mutex, handle any error, set rc = result.
397 * Return 0 on success, nonzero (not rc) on error.
399 #define LOCK_MUTEX(rc, env, mutex) \
400 (((rc) = LOCK_MUTEX0(mutex)) && \
401 ((rc) = mdb_mutex_failed(env, mutex, rc)))
402 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
404 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
405 #define mdb_mutex_failed(env, mutex, rc) (rc)
409 /** A flag for opening a file and requesting synchronous data writes.
410 * This is only used when writing a meta page. It's not strictly needed;
411 * we could just do a normal write and then immediately perform a flush.
412 * But if this flag is available it saves us an extra system call.
414 * @note If O_DSYNC is undefined but exists in /usr/include,
415 * preferably set some compiler flag to get the definition.
419 # define MDB_DSYNC O_DSYNC
421 # define MDB_DSYNC O_SYNC
426 /** Function for flushing the data of a file. Define this to fsync
427 * if fdatasync() is not supported.
429 #ifndef MDB_FDATASYNC
430 # define MDB_FDATASYNC fdatasync
434 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
445 /** A page number in the database.
446 * Note that 64 bit page numbers are overkill, since pages themselves
447 * already represent 12-13 bits of addressable memory, and the OS will
448 * always limit applications to a maximum of 63 bits of address space.
450 * @note In the #MDB_node structure, we only store 48 bits of this value,
451 * which thus limits us to only 60 bits of addressable data.
453 typedef MDB_ID pgno_t;
455 /** A transaction ID.
456 * See struct MDB_txn.mt_txnid for details.
458 typedef MDB_ID txnid_t;
460 /** @defgroup debug Debug Macros
464 /** Enable debug output. Needs variable argument macros (a C99 feature).
465 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
466 * read from and written to the database (used for free space management).
472 static int mdb_debug;
473 static txnid_t mdb_debug_start;
475 /** Print a debug message with printf formatting.
476 * Requires double parenthesis around 2 or more args.
478 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
479 # define DPRINTF0(fmt, ...) \
480 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
482 # define DPRINTF(args) ((void) 0)
484 /** Print a debug string.
485 * The string is printed literally, with no format processing.
487 #define DPUTS(arg) DPRINTF(("%s", arg))
488 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
490 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
493 /** @brief The maximum size of a database page.
495 * It is 32k or 64k, since value-PAGEBASE must fit in
496 * #MDB_page.%mp_upper.
498 * LMDB will use database pages < OS pages if needed.
499 * That causes more I/O in write transactions: The OS must
500 * know (read) the whole page before writing a partial page.
502 * Note that we don't currently support Huge pages. On Linux,
503 * regular data files cannot use Huge pages, and in general
504 * Huge pages aren't actually pageable. We rely on the OS
505 * demand-pager to read our data and page it out when memory
506 * pressure from other processes is high. So until OSs have
507 * actual paging support for Huge pages, they're not viable.
509 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
511 /** The minimum number of keys required in a database page.
512 * Setting this to a larger value will place a smaller bound on the
513 * maximum size of a data item. Data items larger than this size will
514 * be pushed into overflow pages instead of being stored directly in
515 * the B-tree node. This value used to default to 4. With a page size
516 * of 4096 bytes that meant that any item larger than 1024 bytes would
517 * go into an overflow page. That also meant that on average 2-3KB of
518 * each overflow page was wasted space. The value cannot be lower than
519 * 2 because then there would no longer be a tree structure. With this
520 * value, items larger than 2KB will go into overflow pages, and on
521 * average only 1KB will be wasted.
523 #define MDB_MINKEYS 2
525 /** A stamp that identifies a file as an LMDB file.
526 * There's nothing special about this value other than that it is easily
527 * recognizable, and it will reflect any byte order mismatches.
529 #define MDB_MAGIC 0xBEEFC0DE
531 /** The version number for a database's datafile format. */
532 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
533 /** The version number for a database's lockfile format. */
534 #define MDB_LOCK_VERSION 1
536 /** @brief The max size of a key we can write, or 0 for computed max.
538 * This macro should normally be left alone or set to 0.
539 * Note that a database with big keys or dupsort data cannot be
540 * reliably modified by a liblmdb which uses a smaller max.
541 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
543 * Other values are allowed, for backwards compat. However:
544 * A value bigger than the computed max can break if you do not
545 * know what you are doing, and liblmdb <= 0.9.10 can break when
546 * modifying a DB with keys/dupsort data bigger than its max.
548 * Data items in an #MDB_DUPSORT database are also limited to
549 * this size, since they're actually keys of a sub-DB. Keys and
550 * #MDB_DUPSORT data items must fit on a node in a regular page.
552 #ifndef MDB_MAXKEYSIZE
553 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
556 /** The maximum size of a key we can write to the environment. */
558 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
560 #define ENV_MAXKEY(env) ((env)->me_maxkey)
563 /** @brief The maximum size of a data item.
565 * We only store a 32 bit value for node sizes.
567 #define MAXDATASIZE 0xffffffffUL
570 /** Key size which fits in a #DKBUF.
573 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
576 * This is used for printing a hex dump of a key's contents.
578 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
579 /** Display a key in hex.
581 * Invoke a function to display a key in hex.
583 #define DKEY(x) mdb_dkey(x, kbuf)
589 /** An invalid page number.
590 * Mainly used to denote an empty tree.
592 #define P_INVALID (~(pgno_t)0)
594 /** Test if the flags \b f are set in a flag word \b w. */
595 #define F_ISSET(w, f) (((w) & (f)) == (f))
597 /** Round \b n up to an even number. */
598 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
600 /** Used for offsets within a single page.
601 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
604 typedef uint16_t indx_t;
606 /** Default size of memory map.
607 * This is certainly too small for any actual applications. Apps should always set
608 * the size explicitly using #mdb_env_set_mapsize().
610 #define DEFAULT_MAPSIZE 1048576
612 /** @defgroup readers Reader Lock Table
613 * Readers don't acquire any locks for their data access. Instead, they
614 * simply record their transaction ID in the reader table. The reader
615 * mutex is needed just to find an empty slot in the reader table. The
616 * slot's address is saved in thread-specific data so that subsequent read
617 * transactions started by the same thread need no further locking to proceed.
619 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
621 * No reader table is used if the database is on a read-only filesystem, or
622 * if #MDB_NOLOCK is set.
624 * Since the database uses multi-version concurrency control, readers don't
625 * actually need any locking. This table is used to keep track of which
626 * readers are using data from which old transactions, so that we'll know
627 * when a particular old transaction is no longer in use. Old transactions
628 * that have discarded any data pages can then have those pages reclaimed
629 * for use by a later write transaction.
631 * The lock table is constructed such that reader slots are aligned with the
632 * processor's cache line size. Any slot is only ever used by one thread.
633 * This alignment guarantees that there will be no contention or cache
634 * thrashing as threads update their own slot info, and also eliminates
635 * any need for locking when accessing a slot.
637 * A writer thread will scan every slot in the table to determine the oldest
638 * outstanding reader transaction. Any freed pages older than this will be
639 * reclaimed by the writer. The writer doesn't use any locks when scanning
640 * this table. This means that there's no guarantee that the writer will
641 * see the most up-to-date reader info, but that's not required for correct
642 * operation - all we need is to know the upper bound on the oldest reader,
643 * we don't care at all about the newest reader. So the only consequence of
644 * reading stale information here is that old pages might hang around a
645 * while longer before being reclaimed. That's actually good anyway, because
646 * the longer we delay reclaiming old pages, the more likely it is that a
647 * string of contiguous pages can be found after coalescing old pages from
648 * many old transactions together.
651 /** Number of slots in the reader table.
652 * This value was chosen somewhat arbitrarily. 126 readers plus a
653 * couple mutexes fit exactly into 8KB on my development machine.
654 * Applications should set the table size using #mdb_env_set_maxreaders().
656 #define DEFAULT_READERS 126
658 /** The size of a CPU cache line in bytes. We want our lock structures
659 * aligned to this size to avoid false cache line sharing in the
661 * This value works for most CPUs. For Itanium this should be 128.
667 /** The information we store in a single slot of the reader table.
668 * In addition to a transaction ID, we also record the process and
669 * thread ID that owns a slot, so that we can detect stale information,
670 * e.g. threads or processes that went away without cleaning up.
671 * @note We currently don't check for stale records. We simply re-init
672 * the table when we know that we're the only process opening the
675 typedef struct MDB_rxbody {
676 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
677 * Multiple readers that start at the same time will probably have the
678 * same ID here. Again, it's not important to exclude them from
679 * anything; all we need to know is which version of the DB they
680 * started from so we can avoid overwriting any data used in that
681 * particular version.
683 volatile txnid_t mrb_txnid;
684 /** The process ID of the process owning this reader txn. */
685 volatile MDB_PID_T mrb_pid;
686 /** The thread ID of the thread owning this txn. */
687 volatile MDB_THR_T mrb_tid;
690 /** The actual reader record, with cacheline padding. */
691 typedef struct MDB_reader {
694 /** shorthand for mrb_txnid */
695 #define mr_txnid mru.mrx.mrb_txnid
696 #define mr_pid mru.mrx.mrb_pid
697 #define mr_tid mru.mrx.mrb_tid
698 /** cache line alignment */
699 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
703 /** The header for the reader table.
704 * The table resides in a memory-mapped file. (This is a different file
705 * than is used for the main database.)
707 * For POSIX the actual mutexes reside in the shared memory of this
708 * mapped file. On Windows, mutexes are named objects allocated by the
709 * kernel; we store the mutex names in this mapped file so that other
710 * processes can grab them. This same approach is also used on
711 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
712 * process-shared POSIX mutexes. For these cases where a named object
713 * is used, the object name is derived from a 64 bit FNV hash of the
714 * environment pathname. As such, naming collisions are extremely
715 * unlikely. If a collision occurs, the results are unpredictable.
717 typedef struct MDB_txbody {
718 /** Stamp identifying this as an LMDB file. It must be set
721 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
723 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
724 char mtb_rmname[MNAME_LEN];
726 /** Mutex protecting access to this table.
727 * This is the reader table lock used with LOCK_MUTEX().
729 mdb_mutex_t mtb_rmutex;
731 /** The ID of the last transaction committed to the database.
732 * This is recorded here only for convenience; the value can always
733 * be determined by reading the main database meta pages.
735 volatile txnid_t mtb_txnid;
736 /** The number of slots that have been used in the reader table.
737 * This always records the maximum count, it is not decremented
738 * when readers release their slots.
740 volatile unsigned mtb_numreaders;
743 /** The actual reader table definition. */
744 typedef struct MDB_txninfo {
747 #define mti_magic mt1.mtb.mtb_magic
748 #define mti_format mt1.mtb.mtb_format
749 #define mti_rmutex mt1.mtb.mtb_rmutex
750 #define mti_rmname mt1.mtb.mtb_rmname
751 #define mti_txnid mt1.mtb.mtb_txnid
752 #define mti_numreaders mt1.mtb.mtb_numreaders
753 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
756 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
757 char mt2_wmname[MNAME_LEN];
758 #define mti_wmname mt2.mt2_wmname
760 mdb_mutex_t mt2_wmutex;
761 #define mti_wmutex mt2.mt2_wmutex
763 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
765 MDB_reader mti_readers[1];
768 /** Lockfile format signature: version, features and field layout */
769 #define MDB_LOCK_FORMAT \
771 ((MDB_LOCK_VERSION) \
772 /* Flags which describe functionality */ \
773 + (((MDB_PIDLOCK) != 0) << 16)))
776 /** Common header for all page types. The page type depends on #mp_flags.
778 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with
779 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages
780 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header.
782 * #P_OVERFLOW records occupy one or more contiguous pages where only the
783 * first has a page header. They hold the real data of #F_BIGDATA nodes.
785 * #P_SUBP sub-pages are small leaf "pages" with duplicate data.
786 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page.
787 * (Duplicate data can also go in sub-databases, which use normal pages.)
789 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot.
791 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once
792 * in the snapshot: Either used by a database or listed in a freeDB record.
794 typedef struct MDB_page {
795 #define mp_pgno mp_p.p_pgno
796 #define mp_next mp_p.p_next
798 pgno_t p_pgno; /**< page number */
799 struct MDB_page *p_next; /**< for in-memory list of freed pages */
801 uint16_t mp_pad; /**< key size if this is a LEAF2 page */
802 /** @defgroup mdb_page Page Flags
804 * Flags for the page headers.
807 #define P_BRANCH 0x01 /**< branch page */
808 #define P_LEAF 0x02 /**< leaf page */
809 #define P_OVERFLOW 0x04 /**< overflow page */
810 #define P_META 0x08 /**< meta page */
811 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
812 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
813 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
814 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
815 #define P_KEEP 0x8000 /**< leave this page alone during spill */
817 uint16_t mp_flags; /**< @ref mdb_page */
818 #define mp_lower mp_pb.pb.pb_lower
819 #define mp_upper mp_pb.pb.pb_upper
820 #define mp_pages mp_pb.pb_pages
823 indx_t pb_lower; /**< lower bound of free space */
824 indx_t pb_upper; /**< upper bound of free space */
826 uint32_t pb_pages; /**< number of overflow pages */
828 indx_t mp_ptrs[1]; /**< dynamic size */
831 /** Size of the page header, excluding dynamic data at the end */
832 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
834 /** Address of first usable data byte in a page, after the header */
835 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
837 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
838 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
840 /** Number of nodes on a page */
841 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
843 /** The amount of space remaining in the page */
844 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
846 /** The percentage of space used in the page, in tenths of a percent. */
847 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
848 ((env)->me_psize - PAGEHDRSZ))
849 /** The minimum page fill factor, in tenths of a percent.
850 * Pages emptier than this are candidates for merging.
852 #define FILL_THRESHOLD 250
854 /** Test if a page is a leaf page */
855 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
856 /** Test if a page is a LEAF2 page */
857 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
858 /** Test if a page is a branch page */
859 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
860 /** Test if a page is an overflow page */
861 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
862 /** Test if a page is a sub page */
863 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
865 /** The number of overflow pages needed to store the given size. */
866 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
868 /** Link in #MDB_txn.%mt_loose_pgs list.
869 * Kept outside the page header, which is needed when reusing the page.
871 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
873 /** Header for a single key/data pair within a page.
874 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
875 * We guarantee 2-byte alignment for 'MDB_node's.
877 typedef struct MDB_node {
878 /** lo and hi are used for data size on leaf nodes and for
879 * child pgno on branch nodes. On 64 bit platforms, flags
880 * is also used for pgno. (Branch nodes have no flags).
881 * They are in host byte order in case that lets some
882 * accesses be optimized into a 32-bit word access.
884 #if BYTE_ORDER == LITTLE_ENDIAN
885 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
887 unsigned short mn_hi, mn_lo;
889 /** @defgroup mdb_node Node Flags
891 * Flags for node headers.
894 #define F_BIGDATA 0x01 /**< data put on overflow page */
895 #define F_SUBDATA 0x02 /**< data is a sub-database */
896 #define F_DUPDATA 0x04 /**< data has duplicates */
898 /** valid flags for #mdb_node_add() */
899 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
902 unsigned short mn_flags; /**< @ref mdb_node */
903 unsigned short mn_ksize; /**< key size */
904 char mn_data[1]; /**< key and data are appended here */
907 /** Size of the node header, excluding dynamic data at the end */
908 #define NODESIZE offsetof(MDB_node, mn_data)
910 /** Bit position of top word in page number, for shifting mn_flags */
911 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
913 /** Size of a node in a branch page with a given key.
914 * This is just the node header plus the key, there is no data.
916 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
918 /** Size of a node in a leaf page with a given key and data.
919 * This is node header plus key plus data size.
921 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
923 /** Address of node \b i in page \b p */
924 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
926 /** Address of the key for the node */
927 #define NODEKEY(node) (void *)((node)->mn_data)
929 /** Address of the data for a node */
930 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
932 /** Get the page number pointed to by a branch node */
933 #define NODEPGNO(node) \
934 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
935 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
936 /** Set the page number in a branch node */
937 #define SETPGNO(node,pgno) do { \
938 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
939 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
941 /** Get the size of the data in a leaf node */
942 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
943 /** Set the size of the data for a leaf node */
944 #define SETDSZ(node,size) do { \
945 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
946 /** The size of a key in a node */
947 #define NODEKSZ(node) ((node)->mn_ksize)
949 /** Copy a page number from src to dst */
951 #define COPY_PGNO(dst,src) dst = src
953 #if SIZE_MAX > 4294967295UL
954 #define COPY_PGNO(dst,src) do { \
955 unsigned short *s, *d; \
956 s = (unsigned short *)&(src); \
957 d = (unsigned short *)&(dst); \
964 #define COPY_PGNO(dst,src) do { \
965 unsigned short *s, *d; \
966 s = (unsigned short *)&(src); \
967 d = (unsigned short *)&(dst); \
973 /** The address of a key in a LEAF2 page.
974 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
975 * There are no node headers, keys are stored contiguously.
977 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
979 /** Set the \b node's key into \b keyptr, if requested. */
980 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
981 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
983 /** Set the \b node's key into \b key. */
984 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
986 /** Information about a single database in the environment. */
987 typedef struct MDB_db {
988 uint32_t md_pad; /**< also ksize for LEAF2 pages */
989 uint16_t md_flags; /**< @ref mdb_dbi_open */
990 uint16_t md_depth; /**< depth of this tree */
991 pgno_t md_branch_pages; /**< number of internal pages */
992 pgno_t md_leaf_pages; /**< number of leaf pages */
993 pgno_t md_overflow_pages; /**< number of overflow pages */
994 size_t md_entries; /**< number of data items */
995 pgno_t md_root; /**< the root page of this tree */
998 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
999 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1000 /** #mdb_dbi_open() flags */
1001 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1002 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1004 /** Handle for the DB used to track free pages. */
1006 /** Handle for the default DB. */
1008 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1011 /** Number of meta pages - also hardcoded elsewhere */
1014 /** Meta page content.
1015 * A meta page is the start point for accessing a database snapshot.
1016 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1018 typedef struct MDB_meta {
1019 /** Stamp identifying this as an LMDB file. It must be set
1022 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1023 uint32_t mm_version;
1024 void *mm_address; /**< address for fixed mapping */
1025 size_t mm_mapsize; /**< size of mmap region */
1026 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1027 /** The size of pages used in this DB */
1028 #define mm_psize mm_dbs[FREE_DBI].md_pad
1029 /** Any persistent environment flags. @ref mdb_env */
1030 #define mm_flags mm_dbs[FREE_DBI].md_flags
1031 /** Last used page in the datafile.
1032 * Actually the file may be shorter if the freeDB lists the final pages.
1035 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1038 /** Buffer for a stack-allocated meta page.
1039 * The members define size and alignment, and silence type
1040 * aliasing warnings. They are not used directly; that could
1041 * mean incorrectly using several union members in parallel.
1043 typedef union MDB_metabuf {
1046 char mm_pad[PAGEHDRSZ];
1051 /** Auxiliary DB info.
1052 * The information here is mostly static/read-only. There is
1053 * only a single copy of this record in the environment.
1055 typedef struct MDB_dbx {
1056 MDB_val md_name; /**< name of the database */
1057 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1058 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1059 MDB_rel_func *md_rel; /**< user relocate function */
1060 void *md_relctx; /**< user-provided context for md_rel */
1063 /** A database transaction.
1064 * Every operation requires a transaction handle.
1067 MDB_txn *mt_parent; /**< parent of a nested txn */
1068 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1070 pgno_t mt_next_pgno; /**< next unallocated page */
1071 /** The ID of this transaction. IDs are integers incrementing from 1.
1072 * Only committed write transactions increment the ID. If a transaction
1073 * aborts, the ID may be re-used by the next writer.
1076 MDB_env *mt_env; /**< the DB environment */
1077 /** The list of pages that became unused during this transaction.
1079 MDB_IDL mt_free_pgs;
1080 /** The list of loose pages that became unused and may be reused
1081 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1083 MDB_page *mt_loose_pgs;
1084 /** Number of loose pages (#mt_loose_pgs) */
1086 /** The sorted list of dirty pages we temporarily wrote to disk
1087 * because the dirty list was full. page numbers in here are
1088 * shifted left by 1, deleted slots have the LSB set.
1090 MDB_IDL mt_spill_pgs;
1092 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1093 MDB_ID2L dirty_list;
1094 /** For read txns: This thread/txn's reader table slot, or NULL. */
1097 /** Array of records for each DB known in the environment. */
1099 /** Array of MDB_db records for each known DB */
1101 /** Array of sequence numbers for each DB handle */
1102 unsigned int *mt_dbiseqs;
1103 /** @defgroup mt_dbflag Transaction DB Flags
1107 #define DB_DIRTY 0x01 /**< DB was written in this txn */
1108 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1109 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1110 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1111 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1112 #define DB_DUPDATA 0x20 /**< DB is #MDB_DUPSORT data */
1114 /** In write txns, array of cursors for each DB */
1115 MDB_cursor **mt_cursors;
1116 /** Array of flags for each DB */
1117 unsigned char *mt_dbflags;
1118 /** Number of DB records in use, or 0 when the txn is finished.
1119 * This number only ever increments until the txn finishes; we
1120 * don't decrement it when individual DB handles are closed.
1124 /** @defgroup mdb_txn Transaction Flags
1128 /** #mdb_txn_begin() flags */
1129 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1130 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1131 /* internal txn flags */
1132 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1133 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1134 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1135 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1136 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1137 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1138 /** most operations on the txn are currently illegal */
1139 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1141 unsigned int mt_flags; /**< @ref mdb_txn */
1142 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1143 * Includes ancestor txns' dirty pages not hidden by other txns'
1144 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1145 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1147 unsigned int mt_dirty_room;
1150 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1151 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1152 * raise this on a 64 bit machine.
1154 #define CURSOR_STACK 32
1158 /** Cursors are used for all DB operations.
1159 * A cursor holds a path of (page pointer, key index) from the DB
1160 * root to a position in the DB, plus other state. #MDB_DUPSORT
1161 * cursors include an xcursor to the current data item. Write txns
1162 * track their cursors and keep them up to date when data moves.
1163 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1164 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1167 /** Next cursor on this DB in this txn */
1168 MDB_cursor *mc_next;
1169 /** Backup of the original cursor if this cursor is a shadow */
1170 MDB_cursor *mc_backup;
1171 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1172 struct MDB_xcursor *mc_xcursor;
1173 /** The transaction that owns this cursor */
1175 /** The database handle this cursor operates on */
1177 /** The database record for this cursor */
1179 /** The database auxiliary record for this cursor */
1181 /** The @ref mt_dbflag for this database */
1182 unsigned char *mc_dbflag;
1183 unsigned short mc_snum; /**< number of pushed pages */
1184 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1185 /** @defgroup mdb_cursor Cursor Flags
1187 * Cursor state flags.
1190 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1191 #define C_EOF 0x02 /**< No more data */
1192 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1193 #define C_DEL 0x08 /**< last op was a cursor_del */
1194 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1196 unsigned int mc_flags; /**< @ref mdb_cursor */
1197 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1198 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1201 /** Context for sorted-dup records.
1202 * We could have gone to a fully recursive design, with arbitrarily
1203 * deep nesting of sub-databases. But for now we only handle these
1204 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1206 typedef struct MDB_xcursor {
1207 /** A sub-cursor for traversing the Dup DB */
1208 MDB_cursor mx_cursor;
1209 /** The database record for this Dup DB */
1211 /** The auxiliary DB record for this Dup DB */
1213 /** The @ref mt_dbflag for this Dup DB */
1214 unsigned char mx_dbflag;
1217 /** Check if there is an inited xcursor, so #XCURSOR_REFRESH() is proper */
1218 #define XCURSOR_INITED(mc) \
1219 ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
1221 /** Update sub-page pointer, if any, in \b mc->mc_xcursor. Needed
1222 * when the node which contains the sub-page may have moved. Called
1223 * with \b mp = mc->mc_pg[mc->mc_top], \b ki = mc->mc_ki[mc->mc_top].
1225 #define XCURSOR_REFRESH(mc, mp, ki) do { \
1226 MDB_page *xr_pg = (mp); \
1227 MDB_node *xr_node = NODEPTR(xr_pg, ki); \
1228 if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \
1229 (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \
1232 /** State of FreeDB old pages, stored in the MDB_env */
1233 typedef struct MDB_pgstate {
1234 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1235 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1238 /** The database environment. */
1240 HANDLE me_fd; /**< The main data file */
1241 HANDLE me_lfd; /**< The lock file */
1242 HANDLE me_mfd; /**< For writing and syncing the meta pages */
1243 /** Failed to update the meta page. Probably an I/O error. */
1244 #define MDB_FATAL_ERROR 0x80000000U
1245 /** Some fields are initialized. */
1246 #define MDB_ENV_ACTIVE 0x20000000U
1247 /** me_txkey is set */
1248 #define MDB_ENV_TXKEY 0x10000000U
1249 /** fdatasync is unreliable */
1250 #define MDB_FSYNCONLY 0x08000000U
1251 uint32_t me_flags; /**< @ref mdb_env */
1252 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1253 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1254 unsigned int me_maxreaders; /**< size of the reader table */
1255 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1256 volatile int me_close_readers;
1257 MDB_dbi me_numdbs; /**< number of DBs opened */
1258 MDB_dbi me_maxdbs; /**< size of the DB table */
1259 MDB_PID_T me_pid; /**< process ID of this env */
1260 char *me_path; /**< path to the DB files */
1261 char *me_map; /**< the memory map of the data file */
1262 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1263 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1264 void *me_pbuf; /**< scratch area for DUPSORT put() */
1265 MDB_txn *me_txn; /**< current write transaction */
1266 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1267 size_t me_mapsize; /**< size of the data memory map */
1268 off_t me_size; /**< current file size */
1269 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1270 MDB_dbx *me_dbxs; /**< array of static DB info */
1271 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1272 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1273 pthread_key_t me_txkey; /**< thread-key for readers */
1274 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1275 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1276 # define me_pglast me_pgstate.mf_pglast
1277 # define me_pghead me_pgstate.mf_pghead
1278 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1279 /** IDL of pages that became unused in a write txn */
1280 MDB_IDL me_free_pgs;
1281 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1282 MDB_ID2L me_dirty_list;
1283 /** Max number of freelist items that can fit in a single overflow page */
1285 /** Max size of a node on a page */
1286 unsigned int me_nodemax;
1287 #if !(MDB_MAXKEYSIZE)
1288 unsigned int me_maxkey; /**< max size of a key */
1290 int me_live_reader; /**< have liveness lock in reader table */
1292 int me_pidquery; /**< Used in OpenProcess */
1294 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1295 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1296 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1298 mdb_mutex_t me_rmutex;
1299 mdb_mutex_t me_wmutex;
1301 void *me_userctx; /**< User-settable context */
1302 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1305 /** Nested transaction */
1306 typedef struct MDB_ntxn {
1307 MDB_txn mnt_txn; /**< the transaction */
1308 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1311 /** max number of pages to commit in one writev() call */
1312 #define MDB_COMMIT_PAGES 64
1313 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1314 #undef MDB_COMMIT_PAGES
1315 #define MDB_COMMIT_PAGES IOV_MAX
1318 /** max bytes to write in one call */
1319 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1321 /** Check \b txn and \b dbi arguments to a function */
1322 #define TXN_DBI_EXIST(txn, dbi, validity) \
1323 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1325 /** Check for misused \b dbi handles */
1326 #define TXN_DBI_CHANGED(txn, dbi) \
1327 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1329 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1330 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1331 static int mdb_page_touch(MDB_cursor *mc);
1333 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1334 "reset-tmp", "fail-begin", "fail-beginchild"}
1336 /* mdb_txn_end operation number, for logging */
1337 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1338 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1340 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1341 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1342 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1343 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1344 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1346 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1347 static int mdb_page_search_root(MDB_cursor *mc,
1348 MDB_val *key, int modify);
1349 #define MDB_PS_MODIFY 1
1350 #define MDB_PS_ROOTONLY 2
1351 #define MDB_PS_FIRST 4
1352 #define MDB_PS_LAST 8
1353 static int mdb_page_search(MDB_cursor *mc,
1354 MDB_val *key, int flags);
1355 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1357 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1358 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1359 pgno_t newpgno, unsigned int nflags);
1361 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1362 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1363 static int mdb_env_write_meta(MDB_txn *txn);
1364 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1365 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1367 static void mdb_env_close0(MDB_env *env, int excl);
1369 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1370 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1371 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1372 static void mdb_node_del(MDB_cursor *mc, int ksize);
1373 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1374 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1375 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1376 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1377 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1379 static int mdb_rebalance(MDB_cursor *mc);
1380 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1382 static void mdb_cursor_pop(MDB_cursor *mc);
1383 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1385 static int mdb_cursor_del0(MDB_cursor *mc);
1386 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1387 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1388 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1389 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1390 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1392 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1393 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1395 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1396 static void mdb_xcursor_init0(MDB_cursor *mc);
1397 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1398 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1400 static int mdb_drop0(MDB_cursor *mc, int subs);
1401 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1402 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1405 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1408 /** Compare two items pointing at size_t's of unknown alignment. */
1409 #ifdef MISALIGNED_OK
1410 # define mdb_cmp_clong mdb_cmp_long
1412 # define mdb_cmp_clong mdb_cmp_cint
1416 static SECURITY_DESCRIPTOR mdb_null_sd;
1417 static SECURITY_ATTRIBUTES mdb_all_sa;
1418 static int mdb_sec_inited;
1421 static int utf8_to_utf16(const char *src, struct MDB_name *dst, int xtra);
1424 /** Return the library version info. */
1426 mdb_version(int *major, int *minor, int *patch)
1428 if (major) *major = MDB_VERSION_MAJOR;
1429 if (minor) *minor = MDB_VERSION_MINOR;
1430 if (patch) *patch = MDB_VERSION_PATCH;
1431 return MDB_VERSION_STRING;
1434 /** Table of descriptions for LMDB @ref errors */
1435 static char *const mdb_errstr[] = {
1436 "MDB_KEYEXIST: Key/data pair already exists",
1437 "MDB_NOTFOUND: No matching key/data pair found",
1438 "MDB_PAGE_NOTFOUND: Requested page not found",
1439 "MDB_CORRUPTED: Located page was wrong type",
1440 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1441 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1442 "MDB_INVALID: File is not an LMDB file",
1443 "MDB_MAP_FULL: Environment mapsize limit reached",
1444 "MDB_DBS_FULL: Environment maxdbs limit reached",
1445 "MDB_READERS_FULL: Environment maxreaders limit reached",
1446 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1447 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1448 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1449 "MDB_PAGE_FULL: Internal error - page has no more space",
1450 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1451 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1452 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1453 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1454 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1455 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1459 mdb_strerror(int err)
1462 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1463 * This works as long as no function between the call to mdb_strerror
1464 * and the actual use of the message uses more than 4K of stack.
1466 #define MSGSIZE 1024
1467 #define PADSIZE 4096
1468 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1472 return ("Successful return: 0");
1474 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1475 i = err - MDB_KEYEXIST;
1476 return mdb_errstr[i];
1480 /* These are the C-runtime error codes we use. The comment indicates
1481 * their numeric value, and the Win32 error they would correspond to
1482 * if the error actually came from a Win32 API. A major mess, we should
1483 * have used LMDB-specific error codes for everything.
1486 case ENOENT: /* 2, FILE_NOT_FOUND */
1487 case EIO: /* 5, ACCESS_DENIED */
1488 case ENOMEM: /* 12, INVALID_ACCESS */
1489 case EACCES: /* 13, INVALID_DATA */
1490 case EBUSY: /* 16, CURRENT_DIRECTORY */
1491 case EINVAL: /* 22, BAD_COMMAND */
1492 case ENOSPC: /* 28, OUT_OF_PAPER */
1493 return strerror(err);
1498 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1499 FORMAT_MESSAGE_IGNORE_INSERTS,
1500 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1503 return strerror(err);
1507 /** assert(3) variant in cursor context */
1508 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1509 /** assert(3) variant in transaction context */
1510 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1511 /** assert(3) variant in environment context */
1512 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1515 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1516 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1519 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1520 const char *func, const char *file, int line)
1523 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1524 file, line, expr_txt, func);
1525 if (env->me_assert_func)
1526 env->me_assert_func(env, buf);
1527 fprintf(stderr, "%s\n", buf);
1531 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1535 /** Return the page number of \b mp which may be sub-page, for debug output */
1537 mdb_dbg_pgno(MDB_page *mp)
1540 COPY_PGNO(ret, mp->mp_pgno);
1544 /** Display a key in hexadecimal and return the address of the result.
1545 * @param[in] key the key to display
1546 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1547 * @return The key in hexadecimal form.
1550 mdb_dkey(MDB_val *key, char *buf)
1553 unsigned char *c = key->mv_data;
1559 if (key->mv_size > DKBUF_MAXKEYSIZE)
1560 return "MDB_MAXKEYSIZE";
1561 /* may want to make this a dynamic check: if the key is mostly
1562 * printable characters, print it as-is instead of converting to hex.
1566 for (i=0; i<key->mv_size; i++)
1567 ptr += sprintf(ptr, "%02x", *c++);
1569 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1575 mdb_leafnode_type(MDB_node *n)
1577 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1578 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1579 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1582 /** Display all the keys in the page. */
1584 mdb_page_list(MDB_page *mp)
1586 pgno_t pgno = mdb_dbg_pgno(mp);
1587 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1589 unsigned int i, nkeys, nsize, total = 0;
1593 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1594 case P_BRANCH: type = "Branch page"; break;
1595 case P_LEAF: type = "Leaf page"; break;
1596 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1597 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1598 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1600 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1601 pgno, mp->mp_pages, state);
1604 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1605 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1608 fprintf(stderr, "Bad page %"Z"u flags 0x%X\n", pgno, mp->mp_flags);
1612 nkeys = NUMKEYS(mp);
1613 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1615 for (i=0; i<nkeys; i++) {
1616 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1617 key.mv_size = nsize = mp->mp_pad;
1618 key.mv_data = LEAF2KEY(mp, i, nsize);
1620 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1623 node = NODEPTR(mp, i);
1624 key.mv_size = node->mn_ksize;
1625 key.mv_data = node->mn_data;
1626 nsize = NODESIZE + key.mv_size;
1627 if (IS_BRANCH(mp)) {
1628 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1632 if (F_ISSET(node->mn_flags, F_BIGDATA))
1633 nsize += sizeof(pgno_t);
1635 nsize += NODEDSZ(node);
1637 nsize += sizeof(indx_t);
1638 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1639 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1641 total = EVEN(total);
1643 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1644 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1648 mdb_cursor_chk(MDB_cursor *mc)
1654 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1655 for (i=0; i<mc->mc_top; i++) {
1657 node = NODEPTR(mp, mc->mc_ki[i]);
1658 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1661 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1663 if (XCURSOR_INITED(mc)) {
1664 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1665 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1666 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1674 /** Count all the pages in each DB and in the freelist
1675 * and make sure it matches the actual number of pages
1677 * All named DBs must be open for a correct count.
1679 static void mdb_audit(MDB_txn *txn)
1683 MDB_ID freecount, count;
1688 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1689 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1690 freecount += *(MDB_ID *)data.mv_data;
1691 mdb_tassert(txn, rc == MDB_NOTFOUND);
1694 for (i = 0; i<txn->mt_numdbs; i++) {
1696 if (!(txn->mt_dbflags[i] & DB_VALID))
1698 mdb_cursor_init(&mc, txn, i, &mx);
1699 if (txn->mt_dbs[i].md_root == P_INVALID)
1701 count += txn->mt_dbs[i].md_branch_pages +
1702 txn->mt_dbs[i].md_leaf_pages +
1703 txn->mt_dbs[i].md_overflow_pages;
1704 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1705 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1706 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1709 mp = mc.mc_pg[mc.mc_top];
1710 for (j=0; j<NUMKEYS(mp); j++) {
1711 MDB_node *leaf = NODEPTR(mp, j);
1712 if (leaf->mn_flags & F_SUBDATA) {
1714 memcpy(&db, NODEDATA(leaf), sizeof(db));
1715 count += db.md_branch_pages + db.md_leaf_pages +
1716 db.md_overflow_pages;
1720 mdb_tassert(txn, rc == MDB_NOTFOUND);
1723 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1724 fprintf(stderr, "audit: %"Z"u freecount: %"Z"u count: %"Z"u total: %"Z"u next_pgno: %"Z"u\n",
1725 txn->mt_txnid, freecount, count+NUM_METAS,
1726 freecount+count+NUM_METAS, txn->mt_next_pgno);
1732 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1734 return txn->mt_dbxs[dbi].md_cmp(a, b);
1738 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1740 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1741 #if UINT_MAX < SIZE_MAX
1742 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1743 dcmp = mdb_cmp_clong;
1748 /** Allocate memory for a page.
1749 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1750 * Set #MDB_TXN_ERROR on failure.
1753 mdb_page_malloc(MDB_txn *txn, unsigned num)
1755 MDB_env *env = txn->mt_env;
1756 MDB_page *ret = env->me_dpages;
1757 size_t psize = env->me_psize, sz = psize, off;
1758 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1759 * For a single page alloc, we init everything after the page header.
1760 * For multi-page, we init the final page; if the caller needed that
1761 * many pages they will be filling in at least up to the last page.
1765 VGMEMP_ALLOC(env, ret, sz);
1766 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1767 env->me_dpages = ret->mp_next;
1770 psize -= off = PAGEHDRSZ;
1775 if ((ret = malloc(sz)) != NULL) {
1776 VGMEMP_ALLOC(env, ret, sz);
1777 if (!(env->me_flags & MDB_NOMEMINIT)) {
1778 memset((char *)ret + off, 0, psize);
1782 txn->mt_flags |= MDB_TXN_ERROR;
1786 /** Free a single page.
1787 * Saves single pages to a list, for future reuse.
1788 * (This is not used for multi-page overflow pages.)
1791 mdb_page_free(MDB_env *env, MDB_page *mp)
1793 mp->mp_next = env->me_dpages;
1794 VGMEMP_FREE(env, mp);
1795 env->me_dpages = mp;
1798 /** Free a dirty page */
1800 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1802 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1803 mdb_page_free(env, dp);
1805 /* large pages just get freed directly */
1806 VGMEMP_FREE(env, dp);
1811 /** Return all dirty pages to dpage list */
1813 mdb_dlist_free(MDB_txn *txn)
1815 MDB_env *env = txn->mt_env;
1816 MDB_ID2L dl = txn->mt_u.dirty_list;
1817 unsigned i, n = dl[0].mid;
1819 for (i = 1; i <= n; i++) {
1820 mdb_dpage_free(env, dl[i].mptr);
1825 /** Loosen or free a single page.
1826 * Saves single pages to a list for future reuse
1827 * in this same txn. It has been pulled from the freeDB
1828 * and already resides on the dirty list, but has been
1829 * deleted. Use these pages first before pulling again
1832 * If the page wasn't dirtied in this txn, just add it
1833 * to this txn's free list.
1836 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1839 pgno_t pgno = mp->mp_pgno;
1840 MDB_txn *txn = mc->mc_txn;
1842 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1843 if (txn->mt_parent) {
1844 MDB_ID2 *dl = txn->mt_u.dirty_list;
1845 /* If txn has a parent, make sure the page is in our
1849 unsigned x = mdb_mid2l_search(dl, pgno);
1850 if (x <= dl[0].mid && dl[x].mid == pgno) {
1851 if (mp != dl[x].mptr) { /* bad cursor? */
1852 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1853 txn->mt_flags |= MDB_TXN_ERROR;
1854 return MDB_CORRUPTED;
1861 /* no parent txn, so it's just ours */
1866 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1868 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1869 txn->mt_loose_pgs = mp;
1870 txn->mt_loose_count++;
1871 mp->mp_flags |= P_LOOSE;
1873 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1881 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1882 * @param[in] mc A cursor handle for the current operation.
1883 * @param[in] pflags Flags of the pages to update:
1884 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1885 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1886 * @return 0 on success, non-zero on failure.
1889 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1891 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1892 MDB_txn *txn = mc->mc_txn;
1893 MDB_cursor *m3, *m0 = mc;
1898 int rc = MDB_SUCCESS, level;
1900 /* Mark pages seen by cursors */
1901 if (mc->mc_flags & C_UNTRACK)
1902 mc = NULL; /* will find mc in mt_cursors */
1903 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1904 for (; mc; mc=mc->mc_next) {
1905 if (!(mc->mc_flags & C_INITIALIZED))
1907 for (m3 = mc;; m3 = &mx->mx_cursor) {
1909 for (j=0; j<m3->mc_snum; j++) {
1911 if ((mp->mp_flags & Mask) == pflags)
1912 mp->mp_flags ^= P_KEEP;
1914 mx = m3->mc_xcursor;
1915 /* Proceed to mx if it is at a sub-database */
1916 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1918 if (! (mp && (mp->mp_flags & P_LEAF)))
1920 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1921 if (!(leaf->mn_flags & F_SUBDATA))
1930 /* Mark dirty root pages */
1931 for (i=0; i<txn->mt_numdbs; i++) {
1932 if (txn->mt_dbflags[i] & DB_DIRTY) {
1933 pgno_t pgno = txn->mt_dbs[i].md_root;
1934 if (pgno == P_INVALID)
1936 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
1938 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1939 dp->mp_flags ^= P_KEEP;
1947 static int mdb_page_flush(MDB_txn *txn, int keep);
1949 /** Spill pages from the dirty list back to disk.
1950 * This is intended to prevent running into #MDB_TXN_FULL situations,
1951 * but note that they may still occur in a few cases:
1952 * 1) our estimate of the txn size could be too small. Currently this
1953 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1954 * 2) child txns may run out of space if their parents dirtied a
1955 * lot of pages and never spilled them. TODO: we probably should do
1956 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1957 * the parent's dirty_room is below a given threshold.
1959 * Otherwise, if not using nested txns, it is expected that apps will
1960 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1961 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1962 * If the txn never references them again, they can be left alone.
1963 * If the txn only reads them, they can be used without any fuss.
1964 * If the txn writes them again, they can be dirtied immediately without
1965 * going thru all of the work of #mdb_page_touch(). Such references are
1966 * handled by #mdb_page_unspill().
1968 * Also note, we never spill DB root pages, nor pages of active cursors,
1969 * because we'll need these back again soon anyway. And in nested txns,
1970 * we can't spill a page in a child txn if it was already spilled in a
1971 * parent txn. That would alter the parent txns' data even though
1972 * the child hasn't committed yet, and we'd have no way to undo it if
1973 * the child aborted.
1975 * @param[in] m0 cursor A cursor handle identifying the transaction and
1976 * database for which we are checking space.
1977 * @param[in] key For a put operation, the key being stored.
1978 * @param[in] data For a put operation, the data being stored.
1979 * @return 0 on success, non-zero on failure.
1982 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1984 MDB_txn *txn = m0->mc_txn;
1986 MDB_ID2L dl = txn->mt_u.dirty_list;
1987 unsigned int i, j, need;
1990 if (m0->mc_flags & C_SUB)
1993 /* Estimate how much space this op will take */
1994 i = m0->mc_db->md_depth;
1995 /* Named DBs also dirty the main DB */
1996 if (m0->mc_dbi >= CORE_DBS)
1997 i += txn->mt_dbs[MAIN_DBI].md_depth;
1998 /* For puts, roughly factor in the key+data size */
2000 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2001 i += i; /* double it for good measure */
2004 if (txn->mt_dirty_room > i)
2007 if (!txn->mt_spill_pgs) {
2008 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2009 if (!txn->mt_spill_pgs)
2012 /* purge deleted slots */
2013 MDB_IDL sl = txn->mt_spill_pgs;
2014 unsigned int num = sl[0];
2016 for (i=1; i<=num; i++) {
2023 /* Preserve pages which may soon be dirtied again */
2024 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2027 /* Less aggressive spill - we originally spilled the entire dirty list,
2028 * with a few exceptions for cursor pages and DB root pages. But this
2029 * turns out to be a lot of wasted effort because in a large txn many
2030 * of those pages will need to be used again. So now we spill only 1/8th
2031 * of the dirty pages. Testing revealed this to be a good tradeoff,
2032 * better than 1/2, 1/4, or 1/10.
2034 if (need < MDB_IDL_UM_MAX / 8)
2035 need = MDB_IDL_UM_MAX / 8;
2037 /* Save the page IDs of all the pages we're flushing */
2038 /* flush from the tail forward, this saves a lot of shifting later on. */
2039 for (i=dl[0].mid; i && need; i--) {
2040 MDB_ID pn = dl[i].mid << 1;
2042 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2044 /* Can't spill twice, make sure it's not already in a parent's
2047 if (txn->mt_parent) {
2049 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2050 if (tx2->mt_spill_pgs) {
2051 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2052 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2053 dp->mp_flags |= P_KEEP;
2061 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2065 mdb_midl_sort(txn->mt_spill_pgs);
2067 /* Flush the spilled part of dirty list */
2068 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2071 /* Reset any dirty pages we kept that page_flush didn't see */
2072 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2075 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2079 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2081 mdb_find_oldest(MDB_txn *txn)
2084 txnid_t mr, oldest = txn->mt_txnid - 1;
2085 if (txn->mt_env->me_txns) {
2086 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2087 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2098 /** Add a page to the txn's dirty list */
2100 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2103 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2105 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2106 insert = mdb_mid2l_append;
2108 insert = mdb_mid2l_insert;
2110 mid.mid = mp->mp_pgno;
2112 rc = insert(txn->mt_u.dirty_list, &mid);
2113 mdb_tassert(txn, rc == 0);
2114 txn->mt_dirty_room--;
2117 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2118 * me_pghead and mt_next_pgno. Set #MDB_TXN_ERROR on failure.
2120 * If there are free pages available from older transactions, they
2121 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2122 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2123 * and move me_pglast to say which records were consumed. Only this
2124 * function can create me_pghead and move me_pglast/mt_next_pgno.
2125 * @param[in] mc cursor A cursor handle identifying the transaction and
2126 * database for which we are allocating.
2127 * @param[in] num the number of pages to allocate.
2128 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2129 * will always be satisfied by a single contiguous chunk of memory.
2130 * @return 0 on success, non-zero on failure.
2133 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2135 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2136 /* Get at most <Max_retries> more freeDB records once me_pghead
2137 * has enough pages. If not enough, use new pages from the map.
2138 * If <Paranoid> and mc is updating the freeDB, only get new
2139 * records if me_pghead is empty. Then the freelist cannot play
2140 * catch-up with itself by growing while trying to save it.
2142 enum { Paranoid = 1, Max_retries = 500 };
2144 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2146 int rc, retry = num * 60;
2147 MDB_txn *txn = mc->mc_txn;
2148 MDB_env *env = txn->mt_env;
2149 pgno_t pgno, *mop = env->me_pghead;
2150 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2152 txnid_t oldest = 0, last;
2157 /* If there are any loose pages, just use them */
2158 if (num == 1 && txn->mt_loose_pgs) {
2159 np = txn->mt_loose_pgs;
2160 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2161 txn->mt_loose_count--;
2162 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2170 /* If our dirty list is already full, we can't do anything */
2171 if (txn->mt_dirty_room == 0) {
2176 for (op = MDB_FIRST;; op = MDB_NEXT) {
2181 /* Seek a big enough contiguous page range. Prefer
2182 * pages at the tail, just truncating the list.
2188 if (mop[i-n2] == pgno+n2)
2195 if (op == MDB_FIRST) { /* 1st iteration */
2196 /* Prepare to fetch more and coalesce */
2197 last = env->me_pglast;
2198 oldest = env->me_pgoldest;
2199 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2202 key.mv_data = &last; /* will look up last+1 */
2203 key.mv_size = sizeof(last);
2205 if (Paranoid && mc->mc_dbi == FREE_DBI)
2208 if (Paranoid && retry < 0 && mop_len)
2212 /* Do not fetch more if the record will be too recent */
2213 if (oldest <= last) {
2215 oldest = mdb_find_oldest(txn);
2216 env->me_pgoldest = oldest;
2222 rc = mdb_cursor_get(&m2, &key, NULL, op);
2224 if (rc == MDB_NOTFOUND)
2228 last = *(txnid_t*)key.mv_data;
2229 if (oldest <= last) {
2231 oldest = mdb_find_oldest(txn);
2232 env->me_pgoldest = oldest;
2238 np = m2.mc_pg[m2.mc_top];
2239 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2240 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2243 idl = (MDB_ID *) data.mv_data;
2246 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2251 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2253 mop = env->me_pghead;
2255 env->me_pglast = last;
2257 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2258 last, txn->mt_dbs[FREE_DBI].md_root, i));
2260 DPRINTF(("IDL %"Z"u", idl[j]));
2262 /* Merge in descending sorted order */
2263 mdb_midl_xmerge(mop, idl);
2267 /* Use new pages from the map when nothing suitable in the freeDB */
2269 pgno = txn->mt_next_pgno;
2270 if (pgno + num >= env->me_maxpg) {
2271 DPUTS("DB size maxed out");
2277 if (env->me_flags & MDB_WRITEMAP) {
2278 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2280 if (!(np = mdb_page_malloc(txn, num))) {
2286 mop[0] = mop_len -= num;
2287 /* Move any stragglers down */
2288 for (j = i-num; j < mop_len; )
2289 mop[++j] = mop[++i];
2291 txn->mt_next_pgno = pgno + num;
2294 mdb_page_dirty(txn, np);
2300 txn->mt_flags |= MDB_TXN_ERROR;
2304 /** Copy the used portions of a non-overflow page.
2305 * @param[in] dst page to copy into
2306 * @param[in] src page to copy from
2307 * @param[in] psize size of a page
2310 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2312 enum { Align = sizeof(pgno_t) };
2313 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2315 /* If page isn't full, just copy the used portion. Adjust
2316 * alignment so memcpy may copy words instead of bytes.
2318 if ((unused &= -Align) && !IS_LEAF2(src)) {
2319 upper = (upper + PAGEBASE) & -Align;
2320 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2321 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2324 memcpy(dst, src, psize - unused);
2328 /** Pull a page off the txn's spill list, if present.
2329 * If a page being referenced was spilled to disk in this txn, bring
2330 * it back and make it dirty/writable again.
2331 * @param[in] txn the transaction handle.
2332 * @param[in] mp the page being referenced. It must not be dirty.
2333 * @param[out] ret the writable page, if any. ret is unchanged if
2334 * mp wasn't spilled.
2337 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2339 MDB_env *env = txn->mt_env;
2342 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2344 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2345 if (!tx2->mt_spill_pgs)
2347 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2348 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2351 if (txn->mt_dirty_room == 0)
2352 return MDB_TXN_FULL;
2353 if (IS_OVERFLOW(mp))
2357 if (env->me_flags & MDB_WRITEMAP) {
2360 np = mdb_page_malloc(txn, num);
2364 memcpy(np, mp, num * env->me_psize);
2366 mdb_page_copy(np, mp, env->me_psize);
2369 /* If in current txn, this page is no longer spilled.
2370 * If it happens to be the last page, truncate the spill list.
2371 * Otherwise mark it as deleted by setting the LSB.
2373 if (x == txn->mt_spill_pgs[0])
2374 txn->mt_spill_pgs[0]--;
2376 txn->mt_spill_pgs[x] |= 1;
2377 } /* otherwise, if belonging to a parent txn, the
2378 * page remains spilled until child commits
2381 mdb_page_dirty(txn, np);
2382 np->mp_flags |= P_DIRTY;
2390 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2391 * Set #MDB_TXN_ERROR on failure.
2392 * @param[in] mc cursor pointing to the page to be touched
2393 * @return 0 on success, non-zero on failure.
2396 mdb_page_touch(MDB_cursor *mc)
2398 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2399 MDB_txn *txn = mc->mc_txn;
2400 MDB_cursor *m2, *m3;
2404 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2405 if (txn->mt_flags & MDB_TXN_SPILLS) {
2407 rc = mdb_page_unspill(txn, mp, &np);
2413 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2414 (rc = mdb_page_alloc(mc, 1, &np)))
2417 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2418 mp->mp_pgno, pgno));
2419 mdb_cassert(mc, mp->mp_pgno != pgno);
2420 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2421 /* Update the parent page, if any, to point to the new page */
2423 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2424 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2425 SETPGNO(node, pgno);
2427 mc->mc_db->md_root = pgno;
2429 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2430 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2432 /* If txn has a parent, make sure the page is in our
2436 unsigned x = mdb_mid2l_search(dl, pgno);
2437 if (x <= dl[0].mid && dl[x].mid == pgno) {
2438 if (mp != dl[x].mptr) { /* bad cursor? */
2439 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2440 txn->mt_flags |= MDB_TXN_ERROR;
2441 return MDB_CORRUPTED;
2446 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2448 np = mdb_page_malloc(txn, 1);
2453 rc = mdb_mid2l_insert(dl, &mid);
2454 mdb_cassert(mc, rc == 0);
2459 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2461 np->mp_flags |= P_DIRTY;
2464 /* Adjust cursors pointing to mp */
2465 mc->mc_pg[mc->mc_top] = np;
2466 m2 = txn->mt_cursors[mc->mc_dbi];
2467 if (mc->mc_flags & C_SUB) {
2468 for (; m2; m2=m2->mc_next) {
2469 m3 = &m2->mc_xcursor->mx_cursor;
2470 if (m3->mc_snum < mc->mc_snum) continue;
2471 if (m3->mc_pg[mc->mc_top] == mp)
2472 m3->mc_pg[mc->mc_top] = np;
2475 for (; m2; m2=m2->mc_next) {
2476 if (m2->mc_snum < mc->mc_snum) continue;
2477 if (m2 == mc) continue;
2478 if (m2->mc_pg[mc->mc_top] == mp) {
2479 m2->mc_pg[mc->mc_top] = np;
2480 if (XCURSOR_INITED(m2) && IS_LEAF(np))
2481 XCURSOR_REFRESH(m2, np, m2->mc_ki[mc->mc_top]);
2488 txn->mt_flags |= MDB_TXN_ERROR;
2493 mdb_env_sync(MDB_env *env, int force)
2496 if (env->me_flags & MDB_RDONLY)
2498 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2499 if (env->me_flags & MDB_WRITEMAP) {
2500 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2501 ? MS_ASYNC : MS_SYNC;
2502 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2505 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2509 #ifdef BROKEN_FDATASYNC
2510 if (env->me_flags & MDB_FSYNCONLY) {
2511 if (fsync(env->me_fd))
2515 if (MDB_FDATASYNC(env->me_fd))
2522 /** Back up parent txn's cursors, then grab the originals for tracking */
2524 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2526 MDB_cursor *mc, *bk;
2531 for (i = src->mt_numdbs; --i >= 0; ) {
2532 if ((mc = src->mt_cursors[i]) != NULL) {
2533 size = sizeof(MDB_cursor);
2535 size += sizeof(MDB_xcursor);
2536 for (; mc; mc = bk->mc_next) {
2542 mc->mc_db = &dst->mt_dbs[i];
2543 /* Kill pointers into src to reduce abuse: The
2544 * user may not use mc until dst ends. But we need a valid
2545 * txn pointer here for cursor fixups to keep working.
2548 mc->mc_dbflag = &dst->mt_dbflags[i];
2549 if ((mx = mc->mc_xcursor) != NULL) {
2550 *(MDB_xcursor *)(bk+1) = *mx;
2551 mx->mx_cursor.mc_txn = dst;
2553 mc->mc_next = dst->mt_cursors[i];
2554 dst->mt_cursors[i] = mc;
2561 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2562 * @param[in] txn the transaction handle.
2563 * @param[in] merge true to keep changes to parent cursors, false to revert.
2564 * @return 0 on success, non-zero on failure.
2567 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2569 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2573 for (i = txn->mt_numdbs; --i >= 0; ) {
2574 for (mc = cursors[i]; mc; mc = next) {
2576 if ((bk = mc->mc_backup) != NULL) {
2578 /* Commit changes to parent txn */
2579 mc->mc_next = bk->mc_next;
2580 mc->mc_backup = bk->mc_backup;
2581 mc->mc_txn = bk->mc_txn;
2582 mc->mc_db = bk->mc_db;
2583 mc->mc_dbflag = bk->mc_dbflag;
2584 if ((mx = mc->mc_xcursor) != NULL)
2585 mx->mx_cursor.mc_txn = bk->mc_txn;
2587 /* Abort nested txn */
2589 if ((mx = mc->mc_xcursor) != NULL)
2590 *mx = *(MDB_xcursor *)(bk+1);
2594 /* Only malloced cursors are permanently tracked. */
2601 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2607 Pidset = F_SETLK, Pidcheck = F_GETLK
2611 /** Set or check a pid lock. Set returns 0 on success.
2612 * Check returns 0 if the process is certainly dead, nonzero if it may
2613 * be alive (the lock exists or an error happened so we do not know).
2615 * On Windows Pidset is a no-op, we merely check for the existence
2616 * of the process with the given pid. On POSIX we use a single byte
2617 * lock on the lockfile, set at an offset equal to the pid.
2620 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2622 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2625 if (op == Pidcheck) {
2626 h = OpenProcess(env->me_pidquery, FALSE, pid);
2627 /* No documented "no such process" code, but other program use this: */
2629 return ErrCode() != ERROR_INVALID_PARAMETER;
2630 /* A process exists until all handles to it close. Has it exited? */
2631 ret = WaitForSingleObject(h, 0) != 0;
2638 struct flock lock_info;
2639 memset(&lock_info, 0, sizeof(lock_info));
2640 lock_info.l_type = F_WRLCK;
2641 lock_info.l_whence = SEEK_SET;
2642 lock_info.l_start = pid;
2643 lock_info.l_len = 1;
2644 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2645 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2647 } else if ((rc = ErrCode()) == EINTR) {
2655 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2656 * @param[in] txn the transaction handle to initialize
2657 * @return 0 on success, non-zero on failure.
2660 mdb_txn_renew0(MDB_txn *txn)
2662 MDB_env *env = txn->mt_env;
2663 MDB_txninfo *ti = env->me_txns;
2665 unsigned int i, nr, flags = txn->mt_flags;
2667 int rc, new_notls = 0;
2669 if ((flags &= MDB_TXN_RDONLY) != 0) {
2671 meta = mdb_env_pick_meta(env);
2672 txn->mt_txnid = meta->mm_txnid;
2673 txn->mt_u.reader = NULL;
2675 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2676 pthread_getspecific(env->me_txkey);
2678 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2679 return MDB_BAD_RSLOT;
2681 MDB_PID_T pid = env->me_pid;
2682 MDB_THR_T tid = pthread_self();
2683 mdb_mutexref_t rmutex = env->me_rmutex;
2685 if (!env->me_live_reader) {
2686 rc = mdb_reader_pid(env, Pidset, pid);
2689 env->me_live_reader = 1;
2692 if (LOCK_MUTEX(rc, env, rmutex))
2694 nr = ti->mti_numreaders;
2695 for (i=0; i<nr; i++)
2696 if (ti->mti_readers[i].mr_pid == 0)
2698 if (i == env->me_maxreaders) {
2699 UNLOCK_MUTEX(rmutex);
2700 return MDB_READERS_FULL;
2702 r = &ti->mti_readers[i];
2703 /* Claim the reader slot, carefully since other code
2704 * uses the reader table un-mutexed: First reset the
2705 * slot, next publish it in mti_numreaders. After
2706 * that, it is safe for mdb_env_close() to touch it.
2707 * When it will be closed, we can finally claim it.
2710 r->mr_txnid = (txnid_t)-1;
2713 ti->mti_numreaders = ++nr;
2714 env->me_close_readers = nr;
2716 UNLOCK_MUTEX(rmutex);
2718 new_notls = (env->me_flags & MDB_NOTLS);
2719 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2724 do /* LY: Retry on a race, ITS#7970. */
2725 r->mr_txnid = ti->mti_txnid;
2726 while(r->mr_txnid != ti->mti_txnid);
2727 txn->mt_txnid = r->mr_txnid;
2728 txn->mt_u.reader = r;
2729 meta = env->me_metas[txn->mt_txnid & 1];
2733 /* Not yet touching txn == env->me_txn0, it may be active */
2735 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2737 txn->mt_txnid = ti->mti_txnid;
2738 meta = env->me_metas[txn->mt_txnid & 1];
2740 meta = mdb_env_pick_meta(env);
2741 txn->mt_txnid = meta->mm_txnid;
2745 if (txn->mt_txnid == mdb_debug_start)
2748 txn->mt_child = NULL;
2749 txn->mt_loose_pgs = NULL;
2750 txn->mt_loose_count = 0;
2751 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2752 txn->mt_u.dirty_list = env->me_dirty_list;
2753 txn->mt_u.dirty_list[0].mid = 0;
2754 txn->mt_free_pgs = env->me_free_pgs;
2755 txn->mt_free_pgs[0] = 0;
2756 txn->mt_spill_pgs = NULL;
2758 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2761 /* Copy the DB info and flags */
2762 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2764 /* Moved to here to avoid a data race in read TXNs */
2765 txn->mt_next_pgno = meta->mm_last_pg+1;
2767 txn->mt_flags = flags;
2770 txn->mt_numdbs = env->me_numdbs;
2771 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2772 x = env->me_dbflags[i];
2773 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2774 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2776 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2777 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2779 if (env->me_flags & MDB_FATAL_ERROR) {
2780 DPUTS("environment had fatal error, must shutdown!");
2782 } else if (env->me_maxpg < txn->mt_next_pgno) {
2783 rc = MDB_MAP_RESIZED;
2787 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2792 mdb_txn_renew(MDB_txn *txn)
2796 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2799 rc = mdb_txn_renew0(txn);
2800 if (rc == MDB_SUCCESS) {
2801 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2802 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2803 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2809 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2813 int rc, size, tsize;
2815 flags &= MDB_TXN_BEGIN_FLAGS;
2816 flags |= env->me_flags & MDB_WRITEMAP;
2818 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2822 /* Nested transactions: Max 1 child, write txns only, no writemap */
2823 flags |= parent->mt_flags;
2824 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2825 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2827 /* Child txns save MDB_pgstate and use own copy of cursors */
2828 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2829 size += tsize = sizeof(MDB_ntxn);
2830 } else if (flags & MDB_RDONLY) {
2831 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2832 size += tsize = sizeof(MDB_txn);
2834 /* Reuse preallocated write txn. However, do not touch it until
2835 * mdb_txn_renew0() succeeds, since it currently may be active.
2840 if ((txn = calloc(1, size)) == NULL) {
2841 DPRINTF(("calloc: %s", strerror(errno)));
2844 txn->mt_dbxs = env->me_dbxs; /* static */
2845 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2846 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2847 txn->mt_flags = flags;
2852 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2853 txn->mt_dbiseqs = parent->mt_dbiseqs;
2854 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2855 if (!txn->mt_u.dirty_list ||
2856 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2858 free(txn->mt_u.dirty_list);
2862 txn->mt_txnid = parent->mt_txnid;
2863 txn->mt_dirty_room = parent->mt_dirty_room;
2864 txn->mt_u.dirty_list[0].mid = 0;
2865 txn->mt_spill_pgs = NULL;
2866 txn->mt_next_pgno = parent->mt_next_pgno;
2867 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2868 parent->mt_child = txn;
2869 txn->mt_parent = parent;
2870 txn->mt_numdbs = parent->mt_numdbs;
2871 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2872 /* Copy parent's mt_dbflags, but clear DB_NEW */
2873 for (i=0; i<txn->mt_numdbs; i++)
2874 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2876 ntxn = (MDB_ntxn *)txn;
2877 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2878 if (env->me_pghead) {
2879 size = MDB_IDL_SIZEOF(env->me_pghead);
2880 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2882 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2887 rc = mdb_cursor_shadow(parent, txn);
2889 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2890 } else { /* MDB_RDONLY */
2891 txn->mt_dbiseqs = env->me_dbiseqs;
2893 rc = mdb_txn_renew0(txn);
2896 if (txn != env->me_txn0)
2899 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2901 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2902 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2903 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2910 mdb_txn_env(MDB_txn *txn)
2912 if(!txn) return NULL;
2917 mdb_txn_id(MDB_txn *txn)
2920 return txn->mt_txnid;
2923 /** Export or close DBI handles opened in this txn. */
2925 mdb_dbis_update(MDB_txn *txn, int keep)
2928 MDB_dbi n = txn->mt_numdbs;
2929 MDB_env *env = txn->mt_env;
2930 unsigned char *tdbflags = txn->mt_dbflags;
2932 for (i = n; --i >= CORE_DBS;) {
2933 if (tdbflags[i] & DB_NEW) {
2935 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2937 char *ptr = env->me_dbxs[i].md_name.mv_data;
2939 env->me_dbxs[i].md_name.mv_data = NULL;
2940 env->me_dbxs[i].md_name.mv_size = 0;
2941 env->me_dbflags[i] = 0;
2942 env->me_dbiseqs[i]++;
2948 if (keep && env->me_numdbs < n)
2952 /** End a transaction, except successful commit of a nested transaction.
2953 * May be called twice for readonly txns: First reset it, then abort.
2954 * @param[in] txn the transaction handle to end
2955 * @param[in] mode why and how to end the transaction
2958 mdb_txn_end(MDB_txn *txn, unsigned mode)
2960 MDB_env *env = txn->mt_env;
2962 static const char *const names[] = MDB_END_NAMES;
2965 /* Export or close DBI handles opened in this txn */
2966 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2968 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2969 names[mode & MDB_END_OPMASK],
2970 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2971 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2973 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2974 if (txn->mt_u.reader) {
2975 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2976 if (!(env->me_flags & MDB_NOTLS)) {
2977 txn->mt_u.reader = NULL; /* txn does not own reader */
2978 } else if (mode & MDB_END_SLOT) {
2979 txn->mt_u.reader->mr_pid = 0;
2980 txn->mt_u.reader = NULL;
2981 } /* else txn owns the slot until it does MDB_END_SLOT */
2983 txn->mt_numdbs = 0; /* prevent further DBI activity */
2984 txn->mt_flags |= MDB_TXN_FINISHED;
2986 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2987 pgno_t *pghead = env->me_pghead;
2989 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2990 mdb_cursors_close(txn, 0);
2991 if (!(env->me_flags & MDB_WRITEMAP)) {
2992 mdb_dlist_free(txn);
2996 txn->mt_flags = MDB_TXN_FINISHED;
2998 if (!txn->mt_parent) {
2999 mdb_midl_shrink(&txn->mt_free_pgs);
3000 env->me_free_pgs = txn->mt_free_pgs;
3002 env->me_pghead = NULL;
3006 mode = 0; /* txn == env->me_txn0, do not free() it */
3008 /* The writer mutex was locked in mdb_txn_begin. */
3010 UNLOCK_MUTEX(env->me_wmutex);
3012 txn->mt_parent->mt_child = NULL;
3013 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3014 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3015 mdb_midl_free(txn->mt_free_pgs);
3016 mdb_midl_free(txn->mt_spill_pgs);
3017 free(txn->mt_u.dirty_list);
3020 mdb_midl_free(pghead);
3023 if (mode & MDB_END_FREE)
3028 mdb_txn_reset(MDB_txn *txn)
3033 /* This call is only valid for read-only txns */
3034 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3037 mdb_txn_end(txn, MDB_END_RESET);
3041 mdb_txn_abort(MDB_txn *txn)
3047 mdb_txn_abort(txn->mt_child);
3049 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3052 /** Save the freelist as of this transaction to the freeDB.
3053 * This changes the freelist. Keep trying until it stabilizes.
3056 mdb_freelist_save(MDB_txn *txn)
3058 /* env->me_pghead[] can grow and shrink during this call.
3059 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3060 * Page numbers cannot disappear from txn->mt_free_pgs[].
3063 MDB_env *env = txn->mt_env;
3064 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3065 txnid_t pglast = 0, head_id = 0;
3066 pgno_t freecnt = 0, *free_pgs, *mop;
3067 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3069 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3071 if (env->me_pghead) {
3072 /* Make sure first page of freeDB is touched and on freelist */
3073 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3074 if (rc && rc != MDB_NOTFOUND)
3078 if (!env->me_pghead && txn->mt_loose_pgs) {
3079 /* Put loose page numbers in mt_free_pgs, since
3080 * we may be unable to return them to me_pghead.
3082 MDB_page *mp = txn->mt_loose_pgs;
3083 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3085 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3086 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3087 txn->mt_loose_pgs = NULL;
3088 txn->mt_loose_count = 0;
3091 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3092 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3093 ? SSIZE_MAX : maxfree_1pg;
3096 /* Come back here after each Put() in case freelist changed */
3101 /* If using records from freeDB which we have not yet
3102 * deleted, delete them and any we reserved for me_pghead.
3104 while (pglast < env->me_pglast) {
3105 rc = mdb_cursor_first(&mc, &key, NULL);
3108 pglast = head_id = *(txnid_t *)key.mv_data;
3109 total_room = head_room = 0;
3110 mdb_tassert(txn, pglast <= env->me_pglast);
3111 rc = mdb_cursor_del(&mc, 0);
3116 /* Save the IDL of pages freed by this txn, to a single record */
3117 if (freecnt < txn->mt_free_pgs[0]) {
3119 /* Make sure last page of freeDB is touched and on freelist */
3120 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3121 if (rc && rc != MDB_NOTFOUND)
3124 free_pgs = txn->mt_free_pgs;
3125 /* Write to last page of freeDB */
3126 key.mv_size = sizeof(txn->mt_txnid);
3127 key.mv_data = &txn->mt_txnid;
3129 freecnt = free_pgs[0];
3130 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3131 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3134 /* Retry if mt_free_pgs[] grew during the Put() */
3135 free_pgs = txn->mt_free_pgs;
3136 } while (freecnt < free_pgs[0]);
3137 mdb_midl_sort(free_pgs);
3138 memcpy(data.mv_data, free_pgs, data.mv_size);
3141 unsigned int i = free_pgs[0];
3142 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3143 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3145 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3151 mop = env->me_pghead;
3152 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3154 /* Reserve records for me_pghead[]. Split it if multi-page,
3155 * to avoid searching freeDB for a page range. Use keys in
3156 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3158 if (total_room >= mop_len) {
3159 if (total_room == mop_len || --more < 0)
3161 } else if (head_room >= maxfree_1pg && head_id > 1) {
3162 /* Keep current record (overflow page), add a new one */
3166 /* (Re)write {key = head_id, IDL length = head_room} */
3167 total_room -= head_room;
3168 head_room = mop_len - total_room;
3169 if (head_room > maxfree_1pg && head_id > 1) {
3170 /* Overflow multi-page for part of me_pghead */
3171 head_room /= head_id; /* amortize page sizes */
3172 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3173 } else if (head_room < 0) {
3174 /* Rare case, not bothering to delete this record */
3177 key.mv_size = sizeof(head_id);
3178 key.mv_data = &head_id;
3179 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3180 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3183 /* IDL is initially empty, zero out at least the length */
3184 pgs = (pgno_t *)data.mv_data;
3185 j = head_room > clean_limit ? head_room : 0;
3189 total_room += head_room;
3192 /* Return loose page numbers to me_pghead, though usually none are
3193 * left at this point. The pages themselves remain in dirty_list.
3195 if (txn->mt_loose_pgs) {
3196 MDB_page *mp = txn->mt_loose_pgs;
3197 unsigned count = txn->mt_loose_count;
3199 /* Room for loose pages + temp IDL with same */
3200 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3202 mop = env->me_pghead;
3203 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3204 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3205 loose[ ++count ] = mp->mp_pgno;
3207 mdb_midl_sort(loose);
3208 mdb_midl_xmerge(mop, loose);
3209 txn->mt_loose_pgs = NULL;
3210 txn->mt_loose_count = 0;
3214 /* Fill in the reserved me_pghead records */
3220 rc = mdb_cursor_first(&mc, &key, &data);
3221 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3222 txnid_t id = *(txnid_t *)key.mv_data;
3223 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3226 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3228 if (len > mop_len) {
3230 data.mv_size = (len + 1) * sizeof(MDB_ID);
3232 data.mv_data = mop -= len;
3235 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3237 if (rc || !(mop_len -= len))
3244 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3245 * @param[in] txn the transaction that's being committed
3246 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3247 * @return 0 on success, non-zero on failure.
3250 mdb_page_flush(MDB_txn *txn, int keep)
3252 MDB_env *env = txn->mt_env;
3253 MDB_ID2L dl = txn->mt_u.dirty_list;
3254 unsigned psize = env->me_psize, j;
3255 int i, pagecount = dl[0].mid, rc;
3256 size_t size = 0, pos = 0;
3258 MDB_page *dp = NULL;
3262 struct iovec iov[MDB_COMMIT_PAGES];
3263 ssize_t wpos = 0, wsize = 0, wres;
3264 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3270 if (env->me_flags & MDB_WRITEMAP) {
3271 /* Clear dirty flags */
3272 while (++i <= pagecount) {
3274 /* Don't flush this page yet */
3275 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3276 dp->mp_flags &= ~P_KEEP;
3280 dp->mp_flags &= ~P_DIRTY;
3285 /* Write the pages */
3287 if (++i <= pagecount) {
3289 /* Don't flush this page yet */
3290 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3291 dp->mp_flags &= ~P_KEEP;
3296 /* clear dirty flag */
3297 dp->mp_flags &= ~P_DIRTY;
3300 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3305 /* Windows actually supports scatter/gather I/O, but only on
3306 * unbuffered file handles. Since we're relying on the OS page
3307 * cache for all our data, that's self-defeating. So we just
3308 * write pages one at a time. We use the ov structure to set
3309 * the write offset, to at least save the overhead of a Seek
3312 DPRINTF(("committing page %"Z"u", pgno));
3313 memset(&ov, 0, sizeof(ov));
3314 ov.Offset = pos & 0xffffffff;
3315 ov.OffsetHigh = pos >> 16 >> 16;
3316 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3318 DPRINTF(("WriteFile: %d", rc));
3322 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3323 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3326 /* Write previous page(s) */
3327 #ifdef MDB_USE_PWRITEV
3328 wres = pwritev(env->me_fd, iov, n, wpos);
3331 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3334 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3338 DPRINTF(("lseek: %s", strerror(rc)));
3341 wres = writev(env->me_fd, iov, n);
3344 if (wres != wsize) {
3349 DPRINTF(("Write error: %s", strerror(rc)));
3351 rc = EIO; /* TODO: Use which error code? */
3352 DPUTS("short write, filesystem full?");
3363 DPRINTF(("committing page %"Z"u", pgno));
3364 next_pos = pos + size;
3365 iov[n].iov_len = size;
3366 iov[n].iov_base = (char *)dp;
3372 /* MIPS has cache coherency issues, this is a no-op everywhere else
3373 * Note: for any size >= on-chip cache size, entire on-chip cache is
3376 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3378 for (i = keep; ++i <= pagecount; ) {
3380 /* This is a page we skipped above */
3383 dl[j].mid = dp->mp_pgno;
3386 mdb_dpage_free(env, dp);
3391 txn->mt_dirty_room += i - j;
3397 mdb_txn_commit(MDB_txn *txn)
3400 unsigned int i, end_mode;
3406 /* mdb_txn_end() mode for a commit which writes nothing */
3407 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3409 if (txn->mt_child) {
3410 rc = mdb_txn_commit(txn->mt_child);
3417 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3421 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3422 DPUTS("txn has failed/finished, can't commit");
3424 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3429 if (txn->mt_parent) {
3430 MDB_txn *parent = txn->mt_parent;
3434 unsigned x, y, len, ps_len;
3436 /* Append our free list to parent's */
3437 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3440 mdb_midl_free(txn->mt_free_pgs);
3441 /* Failures after this must either undo the changes
3442 * to the parent or set MDB_TXN_ERROR in the parent.
3445 parent->mt_next_pgno = txn->mt_next_pgno;
3446 parent->mt_flags = txn->mt_flags;
3448 /* Merge our cursors into parent's and close them */
3449 mdb_cursors_close(txn, 1);
3451 /* Update parent's DB table. */
3452 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3453 parent->mt_numdbs = txn->mt_numdbs;
3454 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3455 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3456 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3457 /* preserve parent's DB_NEW status */
3458 x = parent->mt_dbflags[i] & DB_NEW;
3459 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3462 dst = parent->mt_u.dirty_list;
3463 src = txn->mt_u.dirty_list;
3464 /* Remove anything in our dirty list from parent's spill list */
3465 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3467 pspill[0] = (pgno_t)-1;
3468 /* Mark our dirty pages as deleted in parent spill list */
3469 for (i=0, len=src[0].mid; ++i <= len; ) {
3470 MDB_ID pn = src[i].mid << 1;
3471 while (pn > pspill[x])
3473 if (pn == pspill[x]) {
3478 /* Squash deleted pagenums if we deleted any */
3479 for (x=y; ++x <= ps_len; )
3480 if (!(pspill[x] & 1))
3481 pspill[++y] = pspill[x];
3485 /* Remove anything in our spill list from parent's dirty list */
3486 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3487 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3488 MDB_ID pn = txn->mt_spill_pgs[i];
3490 continue; /* deleted spillpg */
3492 y = mdb_mid2l_search(dst, pn);
3493 if (y <= dst[0].mid && dst[y].mid == pn) {
3495 while (y < dst[0].mid) {
3504 /* Find len = length of merging our dirty list with parent's */
3506 dst[0].mid = 0; /* simplify loops */
3507 if (parent->mt_parent) {
3508 len = x + src[0].mid;
3509 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3510 for (i = x; y && i; y--) {
3511 pgno_t yp = src[y].mid;
3512 while (yp < dst[i].mid)
3514 if (yp == dst[i].mid) {
3519 } else { /* Simplify the above for single-ancestor case */
3520 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3522 /* Merge our dirty list with parent's */
3524 for (i = len; y; dst[i--] = src[y--]) {
3525 pgno_t yp = src[y].mid;
3526 while (yp < dst[x].mid)
3527 dst[i--] = dst[x--];
3528 if (yp == dst[x].mid)
3529 free(dst[x--].mptr);
3531 mdb_tassert(txn, i == x);
3533 free(txn->mt_u.dirty_list);
3534 parent->mt_dirty_room = txn->mt_dirty_room;
3535 if (txn->mt_spill_pgs) {
3536 if (parent->mt_spill_pgs) {
3537 /* TODO: Prevent failure here, so parent does not fail */
3538 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3540 parent->mt_flags |= MDB_TXN_ERROR;
3541 mdb_midl_free(txn->mt_spill_pgs);
3542 mdb_midl_sort(parent->mt_spill_pgs);
3544 parent->mt_spill_pgs = txn->mt_spill_pgs;
3548 /* Append our loose page list to parent's */
3549 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3551 *lp = txn->mt_loose_pgs;
3552 parent->mt_loose_count += txn->mt_loose_count;
3554 parent->mt_child = NULL;
3555 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3560 if (txn != env->me_txn) {
3561 DPUTS("attempt to commit unknown transaction");
3566 mdb_cursors_close(txn, 0);
3568 if (!txn->mt_u.dirty_list[0].mid &&
3569 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3572 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3573 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3575 /* Update DB root pointers */
3576 if (txn->mt_numdbs > CORE_DBS) {
3580 data.mv_size = sizeof(MDB_db);
3582 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3583 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3584 if (txn->mt_dbflags[i] & DB_DIRTY) {
3585 if (TXN_DBI_CHANGED(txn, i)) {
3589 data.mv_data = &txn->mt_dbs[i];
3590 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3598 rc = mdb_freelist_save(txn);
3602 mdb_midl_free(env->me_pghead);
3603 env->me_pghead = NULL;
3604 mdb_midl_shrink(&txn->mt_free_pgs);
3610 if ((rc = mdb_page_flush(txn, 0)) ||
3611 (rc = mdb_env_sync(env, 0)) ||
3612 (rc = mdb_env_write_meta(txn)))
3614 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3617 mdb_txn_end(txn, end_mode);
3625 /** Read the environment parameters of a DB environment before
3626 * mapping it into memory.
3627 * @param[in] env the environment handle
3628 * @param[out] meta address of where to store the meta information
3629 * @return 0 on success, non-zero on failure.
3632 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3638 enum { Size = sizeof(pbuf) };
3640 /* We don't know the page size yet, so use a minimum value.
3641 * Read both meta pages so we can use the latest one.
3644 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3648 memset(&ov, 0, sizeof(ov));
3650 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3651 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3654 rc = pread(env->me_fd, &pbuf, Size, off);
3657 if (rc == 0 && off == 0)
3659 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3660 DPRINTF(("read: %s", mdb_strerror(rc)));
3664 p = (MDB_page *)&pbuf;
3666 if (!F_ISSET(p->mp_flags, P_META)) {
3667 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3672 if (m->mm_magic != MDB_MAGIC) {
3673 DPUTS("meta has invalid magic");
3677 if (m->mm_version != MDB_DATA_VERSION) {
3678 DPRINTF(("database is version %u, expected version %u",
3679 m->mm_version, MDB_DATA_VERSION));
3680 return MDB_VERSION_MISMATCH;
3683 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3689 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3691 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3693 meta->mm_magic = MDB_MAGIC;
3694 meta->mm_version = MDB_DATA_VERSION;
3695 meta->mm_mapsize = env->me_mapsize;
3696 meta->mm_psize = env->me_psize;
3697 meta->mm_last_pg = NUM_METAS-1;
3698 meta->mm_flags = env->me_flags & 0xffff;
3699 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3700 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3701 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3704 /** Write the environment parameters of a freshly created DB environment.
3705 * @param[in] env the environment handle
3706 * @param[in] meta the #MDB_meta to write
3707 * @return 0 on success, non-zero on failure.
3710 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3718 memset(&ov, 0, sizeof(ov));
3719 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3721 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3724 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3725 len = pwrite(fd, ptr, size, pos); \
3726 if (len == -1 && ErrCode() == EINTR) continue; \
3727 rc = (len >= 0); break; } while(1)
3730 DPUTS("writing new meta page");
3732 psize = env->me_psize;
3734 p = calloc(NUM_METAS, psize);
3739 p->mp_flags = P_META;
3740 *(MDB_meta *)METADATA(p) = *meta;
3742 q = (MDB_page *)((char *)p + psize);
3744 q->mp_flags = P_META;
3745 *(MDB_meta *)METADATA(q) = *meta;
3747 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3750 else if ((unsigned) len == psize * NUM_METAS)
3758 /** Update the environment info to commit a transaction.
3759 * @param[in] txn the transaction that's being committed
3760 * @return 0 on success, non-zero on failure.
3763 mdb_env_write_meta(MDB_txn *txn)
3766 MDB_meta meta, metab, *mp;
3770 int rc, len, toggle;
3779 toggle = txn->mt_txnid & 1;
3780 DPRINTF(("writing meta page %d for root page %"Z"u",
3781 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3784 flags = env->me_flags;
3785 mp = env->me_metas[toggle];
3786 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3787 /* Persist any increases of mapsize config */
3788 if (mapsize < env->me_mapsize)
3789 mapsize = env->me_mapsize;
3791 if (flags & MDB_WRITEMAP) {
3792 mp->mm_mapsize = mapsize;
3793 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3794 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3795 mp->mm_last_pg = txn->mt_next_pgno - 1;
3796 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3797 !(defined(__i386__) || defined(__x86_64__))
3798 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3799 __sync_synchronize();
3801 mp->mm_txnid = txn->mt_txnid;
3802 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3803 unsigned meta_size = env->me_psize;
3804 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3805 ptr = (char *)mp - PAGEHDRSZ;
3806 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3807 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3811 if (MDB_MSYNC(ptr, meta_size, rc)) {
3818 metab.mm_txnid = mp->mm_txnid;
3819 metab.mm_last_pg = mp->mm_last_pg;
3821 meta.mm_mapsize = mapsize;
3822 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3823 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3824 meta.mm_last_pg = txn->mt_next_pgno - 1;
3825 meta.mm_txnid = txn->mt_txnid;
3827 off = offsetof(MDB_meta, mm_mapsize);
3828 ptr = (char *)&meta + off;
3829 len = sizeof(MDB_meta) - off;
3830 off += (char *)mp - env->me_map;
3832 /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC.
3833 * (me_mfd goes to the same file as me_fd, but writing to it
3834 * also syncs to disk. Avoids a separate fdatasync() call.)
3836 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3839 memset(&ov, 0, sizeof(ov));
3841 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3846 rc = pwrite(mfd, ptr, len, off);
3849 rc = rc < 0 ? ErrCode() : EIO;
3854 DPUTS("write failed, disk error?");
3855 /* On a failure, the pagecache still contains the new data.
3856 * Write some old data back, to prevent it from being used.
3857 * Use the non-SYNC fd; we know it will fail anyway.
3859 meta.mm_last_pg = metab.mm_last_pg;
3860 meta.mm_txnid = metab.mm_txnid;
3862 memset(&ov, 0, sizeof(ov));
3864 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3866 r2 = pwrite(env->me_fd, ptr, len, off);
3867 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3870 env->me_flags |= MDB_FATAL_ERROR;
3873 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3874 CACHEFLUSH(env->me_map + off, len, DCACHE);
3876 /* Memory ordering issues are irrelevant; since the entire writer
3877 * is wrapped by wmutex, all of these changes will become visible
3878 * after the wmutex is unlocked. Since the DB is multi-version,
3879 * readers will get consistent data regardless of how fresh or
3880 * how stale their view of these values is.
3883 env->me_txns->mti_txnid = txn->mt_txnid;
3888 /** Check both meta pages to see which one is newer.
3889 * @param[in] env the environment handle
3890 * @return newest #MDB_meta.
3893 mdb_env_pick_meta(const MDB_env *env)
3895 MDB_meta *const *metas = env->me_metas;
3896 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3900 mdb_env_create(MDB_env **env)
3904 e = calloc(1, sizeof(MDB_env));
3908 e->me_maxreaders = DEFAULT_READERS;
3909 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3910 e->me_fd = INVALID_HANDLE_VALUE;
3911 e->me_lfd = INVALID_HANDLE_VALUE;
3912 e->me_mfd = INVALID_HANDLE_VALUE;
3913 #ifdef MDB_USE_POSIX_SEM
3914 e->me_rmutex = SEM_FAILED;
3915 e->me_wmutex = SEM_FAILED;
3917 e->me_pid = getpid();
3918 GET_PAGESIZE(e->me_os_psize);
3919 VGMEMP_CREATE(e,0,0);
3925 mdb_env_map(MDB_env *env, void *addr)
3928 unsigned int flags = env->me_flags;
3932 LONG sizelo, sizehi;
3935 if (flags & MDB_RDONLY) {
3936 /* Don't set explicit map size, use whatever exists */
3941 msize = env->me_mapsize;
3942 sizelo = msize & 0xffffffff;
3943 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3945 /* Windows won't create mappings for zero length files.
3946 * and won't map more than the file size.
3947 * Just set the maxsize right now.
3949 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3950 || !SetEndOfFile(env->me_fd)
3951 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3955 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3956 PAGE_READWRITE : PAGE_READONLY,
3957 sizehi, sizelo, NULL);
3960 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3961 FILE_MAP_WRITE : FILE_MAP_READ,
3963 rc = env->me_map ? 0 : ErrCode();
3968 int prot = PROT_READ;
3969 if (flags & MDB_WRITEMAP) {
3971 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3974 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3976 if (env->me_map == MAP_FAILED) {
3981 if (flags & MDB_NORDAHEAD) {
3982 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3984 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3986 #ifdef POSIX_MADV_RANDOM
3987 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3988 #endif /* POSIX_MADV_RANDOM */
3989 #endif /* MADV_RANDOM */
3993 /* Can happen because the address argument to mmap() is just a
3994 * hint. mmap() can pick another, e.g. if the range is in use.
3995 * The MAP_FIXED flag would prevent that, but then mmap could
3996 * instead unmap existing pages to make room for the new map.
3998 if (addr && env->me_map != addr)
3999 return EBUSY; /* TODO: Make a new MDB_* error code? */
4001 p = (MDB_page *)env->me_map;
4002 env->me_metas[0] = METADATA(p);
4003 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4009 mdb_env_set_mapsize(MDB_env *env, size_t size)
4011 /* If env is already open, caller is responsible for making
4012 * sure there are no active txns.
4020 meta = mdb_env_pick_meta(env);
4022 size = meta->mm_mapsize;
4024 /* Silently round up to minimum if the size is too small */
4025 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4029 munmap(env->me_map, env->me_mapsize);
4030 env->me_mapsize = size;
4031 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4032 rc = mdb_env_map(env, old);
4036 env->me_mapsize = size;
4038 env->me_maxpg = env->me_mapsize / env->me_psize;
4043 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4047 env->me_maxdbs = dbs + CORE_DBS;
4052 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4054 if (env->me_map || readers < 1)
4056 env->me_maxreaders = readers;
4061 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4063 if (!env || !readers)
4065 *readers = env->me_maxreaders;
4070 mdb_fsize(HANDLE fd, size_t *size)
4073 LARGE_INTEGER fsize;
4075 if (!GetFileSizeEx(fd, &fsize))
4078 *size = fsize.QuadPart;
4092 typedef wchar_t mdb_nchar_t;
4093 # define MDB_NAME(str) L##str
4094 # define mdb_name_cpy wcscpy
4096 /** Character type for file names: char on Unix, wchar_t on Windows */
4097 typedef char mdb_nchar_t;
4098 # define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */
4099 # define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */
4102 /** Filename - string of #mdb_nchar_t[] */
4103 typedef struct MDB_name {
4104 int mn_len; /**< Length */
4105 int mn_alloced; /**< True if #mn_val was malloced */
4106 mdb_nchar_t *mn_val; /**< Contents */
4109 /** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */
4110 static const mdb_nchar_t *const mdb_suffixes[2][2] = {
4111 { MDB_NAME("/data.mdb"), MDB_NAME("") },
4112 { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") }
4115 #define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */
4117 /** Set up filename + scratch area for filename suffix, for opening files.
4118 * It should be freed with #mdb_fname_destroy().
4119 * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16.
4121 * @param[in] path Pathname for #mdb_env_open().
4122 * @param[in] envflags Whether a subdir and/or lockfile will be used.
4123 * @param[out] fname Resulting filename, with room for a suffix if necessary.
4126 mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname)
4128 int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK);
4129 fname->mn_alloced = 0;
4131 return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN);
4133 fname->mn_len = strlen(path);
4135 fname->mn_val = (char *) path;
4136 else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) {
4137 fname->mn_alloced = 1;
4138 strcpy(fname->mn_val, path);
4146 /** Destroy \b fname from #mdb_fname_init() */
4147 #define mdb_fname_destroy(fname) \
4148 do { if ((fname).mn_alloced) free((fname).mn_val); } while (0)
4150 #ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */
4151 # define MDB_CLOEXEC O_CLOEXEC
4153 # define MDB_CLOEXEC 0
4156 /** File type, access mode etc. for #mdb_fopen() */
4157 enum mdb_fopen_type {
4159 MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS
4161 /* A comment in mdb_fopen() explains some O_* flag choices. */
4162 MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */
4163 MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */
4164 MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */
4165 MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */
4166 /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits
4167 * distinguish otherwise-equal MDB_O_* constants from each other.
4169 MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY,
4170 MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */
4174 /** Open an LMDB file.
4175 * @param[in] env The LMDB environment.
4176 * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is
4177 * appended if necessary to create the filename, without changing mn_len.
4178 * @param[in] which Determines file type, access mode, etc.
4179 * @param[in] mode The Unix permissions for the file, if we create it.
4180 * @param[out] res Resulting file handle.
4181 * @return 0 on success, non-zero on failure.
4184 mdb_fopen(const MDB_env *env, MDB_name *fname,
4185 enum mdb_fopen_type which, mdb_mode_t mode,
4188 int rc = MDB_SUCCESS;
4191 DWORD acc, share, disp, attrs;
4196 if (fname->mn_alloced) /* modifiable copy */
4197 mdb_name_cpy(fname->mn_val + fname->mn_len,
4198 mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]);
4200 /* The directory must already exist. Usually the file need not.
4201 * MDB_O_META requires the file because we already created it using
4202 * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file.
4204 * With MDB_O_COPY we do not want the OS to cache the writes, since
4205 * the source data is already in the OS cache.
4207 * The lockfile needs FD_CLOEXEC (close file descriptor on exec*())
4208 * to avoid the flock() issues noted under Caveats in lmdb.h.
4209 * Also set it for other filehandles which the user cannot get at
4210 * and close himself, which he may need after fork(). I.e. all but
4211 * me_fd, which programs do use via mdb_env_get_fd().
4215 acc = GENERIC_READ|GENERIC_WRITE;
4216 share = FILE_SHARE_READ|FILE_SHARE_WRITE;
4218 attrs = FILE_ATTRIBUTE_NORMAL;
4220 case MDB_O_RDONLY: /* read-only datafile */
4222 disp = OPEN_EXISTING;
4224 case MDB_O_META: /* for writing metapages */
4225 acc = GENERIC_WRITE;
4226 disp = OPEN_EXISTING;
4227 attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH;
4229 case MDB_O_COPY: /* mdb_env_copy() & co */
4230 acc = GENERIC_WRITE;
4233 attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH;
4235 default: break; /* silence gcc -Wswitch (not all enum values handled) */
4237 fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL);
4239 fd = open(fname->mn_val, which & MDB_O_MASK, mode);
4242 if (fd == INVALID_HANDLE_VALUE)
4246 if (which != MDB_O_RDONLY && which != MDB_O_RDWR) {
4247 /* Set CLOEXEC if we could not pass it to open() */
4248 if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1)
4249 (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
4251 if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) {
4252 /* This may require buffer alignment. There is no portable
4253 * way to ask how much, so we require OS pagesize alignment.
4255 # ifdef F_NOCACHE /* __APPLE__ */
4256 (void) fcntl(fd, F_NOCACHE, 1);
4257 # elif defined O_DIRECT
4258 /* open(...O_DIRECT...) would break on filesystems without
4259 * O_DIRECT support (ITS#7682). Try to set it here instead.
4261 if ((flags = fcntl(fd, F_GETFL)) != -1)
4262 (void) fcntl(fd, F_SETFL, flags | O_DIRECT);
4266 #endif /* !_WIN32 */
4273 #ifdef BROKEN_FDATASYNC
4274 #include <sys/utsname.h>
4275 #include <sys/vfs.h>
4278 /** Further setup required for opening an LMDB environment
4281 mdb_env_open2(MDB_env *env)
4283 unsigned int flags = env->me_flags;
4284 int i, newenv = 0, rc;
4288 /* See if we should use QueryLimited */
4290 if ((rc & 0xff) > 5)
4291 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4293 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4296 #ifdef BROKEN_FDATASYNC
4297 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4298 * https://lkml.org/lkml/2012/9/3/83
4299 * Kernels after 3.6-rc6 are known good.
4300 * https://lkml.org/lkml/2012/9/10/556
4301 * See if the DB is on ext3/ext4, then check for new enough kernel
4302 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4307 fstatfs(env->me_fd, &st);
4308 while (st.f_type == 0xEF53) {
4312 if (uts.release[0] < '3') {
4313 if (!strncmp(uts.release, "2.6.32.", 7)) {
4314 i = atoi(uts.release+7);
4316 break; /* 2.6.32.60 and newer is OK */
4317 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4318 i = atoi(uts.release+7);
4320 break; /* 2.6.34.15 and newer is OK */
4322 } else if (uts.release[0] == '3') {
4323 i = atoi(uts.release+2);
4325 break; /* 3.6 and newer is OK */
4327 i = atoi(uts.release+4);
4329 break; /* 3.5.4 and newer is OK */
4330 } else if (i == 2) {
4331 i = atoi(uts.release+4);
4333 break; /* 3.2.30 and newer is OK */
4335 } else { /* 4.x and newer is OK */
4338 env->me_flags |= MDB_FSYNCONLY;
4344 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4347 DPUTS("new mdbenv");
4349 env->me_psize = env->me_os_psize;
4350 if (env->me_psize > MAX_PAGESIZE)
4351 env->me_psize = MAX_PAGESIZE;
4352 memset(&meta, 0, sizeof(meta));
4353 mdb_env_init_meta0(env, &meta);
4354 meta.mm_mapsize = DEFAULT_MAPSIZE;
4356 env->me_psize = meta.mm_psize;
4359 /* Was a mapsize configured? */
4360 if (!env->me_mapsize) {
4361 env->me_mapsize = meta.mm_mapsize;
4364 /* Make sure mapsize >= committed data size. Even when using
4365 * mm_mapsize, which could be broken in old files (ITS#7789).
4367 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4368 if (env->me_mapsize < minsize)
4369 env->me_mapsize = minsize;
4371 meta.mm_mapsize = env->me_mapsize;
4373 if (newenv && !(flags & MDB_FIXEDMAP)) {
4374 /* mdb_env_map() may grow the datafile. Write the metapages
4375 * first, so the file will be valid if initialization fails.
4376 * Except with FIXEDMAP, since we do not yet know mm_address.
4377 * We could fill in mm_address later, but then a different
4378 * program might end up doing that - one with a memory layout
4379 * and map address which does not suit the main program.
4381 rc = mdb_env_init_meta(env, &meta);
4387 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4392 if (flags & MDB_FIXEDMAP)
4393 meta.mm_address = env->me_map;
4394 i = mdb_env_init_meta(env, &meta);
4395 if (i != MDB_SUCCESS) {
4400 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4401 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4403 #if !(MDB_MAXKEYSIZE)
4404 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4406 env->me_maxpg = env->me_mapsize / env->me_psize;
4410 MDB_meta *meta = mdb_env_pick_meta(env);
4411 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4413 DPRINTF(("opened database version %u, pagesize %u",
4414 meta->mm_version, env->me_psize));
4415 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4416 DPRINTF(("depth: %u", db->md_depth));
4417 DPRINTF(("entries: %"Z"u", db->md_entries));
4418 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4419 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4420 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4421 DPRINTF(("root: %"Z"u", db->md_root));
4429 /** Release a reader thread's slot in the reader lock table.
4430 * This function is called automatically when a thread exits.
4431 * @param[in] ptr This points to the slot in the reader lock table.
4434 mdb_env_reader_dest(void *ptr)
4436 MDB_reader *reader = ptr;
4439 if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */
4441 /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */
4446 /** Junk for arranging thread-specific callbacks on Windows. This is
4447 * necessarily platform and compiler-specific. Windows supports up
4448 * to 1088 keys. Let's assume nobody opens more than 64 environments
4449 * in a single process, for now. They can override this if needed.
4451 #ifndef MAX_TLS_KEYS
4452 #define MAX_TLS_KEYS 64
4454 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4455 static int mdb_tls_nkeys;
4457 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4461 case DLL_PROCESS_ATTACH: break;
4462 case DLL_THREAD_ATTACH: break;
4463 case DLL_THREAD_DETACH:
4464 for (i=0; i<mdb_tls_nkeys; i++) {
4465 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4467 mdb_env_reader_dest(r);
4471 case DLL_PROCESS_DETACH: break;
4476 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4478 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4482 /* Force some symbol references.
4483 * _tls_used forces the linker to create the TLS directory if not already done
4484 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4486 #pragma comment(linker, "/INCLUDE:_tls_used")
4487 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4488 #pragma const_seg(".CRT$XLB")
4489 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4490 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4493 #pragma comment(linker, "/INCLUDE:__tls_used")
4494 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4495 #pragma data_seg(".CRT$XLB")
4496 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4498 #endif /* WIN 32/64 */
4499 #endif /* !__GNUC__ */
4502 /** Downgrade the exclusive lock on the region back to shared */
4504 mdb_env_share_locks(MDB_env *env, int *excl)
4507 MDB_meta *meta = mdb_env_pick_meta(env);
4509 env->me_txns->mti_txnid = meta->mm_txnid;
4514 /* First acquire a shared lock. The Unlock will
4515 * then release the existing exclusive lock.
4517 memset(&ov, 0, sizeof(ov));
4518 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4521 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4527 struct flock lock_info;
4528 /* The shared lock replaces the existing lock */
4529 memset((void *)&lock_info, 0, sizeof(lock_info));
4530 lock_info.l_type = F_RDLCK;
4531 lock_info.l_whence = SEEK_SET;
4532 lock_info.l_start = 0;
4533 lock_info.l_len = 1;
4534 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4535 (rc = ErrCode()) == EINTR) ;
4536 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4543 /** Try to get exclusive lock, otherwise shared.
4544 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4547 mdb_env_excl_lock(MDB_env *env, int *excl)
4551 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4555 memset(&ov, 0, sizeof(ov));
4556 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4563 struct flock lock_info;
4564 memset((void *)&lock_info, 0, sizeof(lock_info));
4565 lock_info.l_type = F_WRLCK;
4566 lock_info.l_whence = SEEK_SET;
4567 lock_info.l_start = 0;
4568 lock_info.l_len = 1;
4569 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4570 (rc = ErrCode()) == EINTR) ;
4574 # ifndef MDB_USE_POSIX_MUTEX
4575 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4578 lock_info.l_type = F_RDLCK;
4579 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4580 (rc = ErrCode()) == EINTR) ;
4590 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4592 * @(#) $Revision: 5.1 $
4593 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4594 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4596 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4600 * Please do not copyright this code. This code is in the public domain.
4602 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4603 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4604 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4605 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4606 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4607 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4608 * PERFORMANCE OF THIS SOFTWARE.
4611 * chongo <Landon Curt Noll> /\oo/\
4612 * http://www.isthe.com/chongo/
4614 * Share and Enjoy! :-)
4617 typedef unsigned long long mdb_hash_t;
4618 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4620 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4621 * @param[in] val value to hash
4622 * @param[in] hval initial value for hash
4623 * @return 64 bit hash
4625 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4626 * hval arg on the first call.
4629 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4631 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4632 unsigned char *end = s + val->mv_size;
4634 * FNV-1a hash each octet of the string
4637 /* xor the bottom with the current octet */
4638 hval ^= (mdb_hash_t)*s++;
4640 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4641 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4642 (hval << 7) + (hval << 8) + (hval << 40);
4644 /* return our new hash value */
4648 /** Hash the string and output the encoded hash.
4649 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4650 * very short name limits. We don't care about the encoding being reversible,
4651 * we just want to preserve as many bits of the input as possible in a
4652 * small printable string.
4653 * @param[in] str string to hash
4654 * @param[out] encbuf an array of 11 chars to hold the hash
4656 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4659 mdb_pack85(unsigned long l, char *out)
4663 for (i=0; i<5; i++) {
4664 *out++ = mdb_a85[l % 85];
4670 mdb_hash_enc(MDB_val *val, char *encbuf)
4672 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4674 mdb_pack85(h, encbuf);
4675 mdb_pack85(h>>32, encbuf+5);
4680 /** Open and/or initialize the lock region for the environment.
4681 * @param[in] env The LMDB environment.
4682 * @param[in] fname Filename + scratch area, from #mdb_fname_init().
4683 * @param[in] mode The Unix permissions for the file, if we create it.
4684 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4685 * @return 0 on success, non-zero on failure.
4688 mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl)
4691 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4693 # define MDB_ERRCODE_ROFS EROFS
4698 rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd);
4700 /* Omit lockfile if read-only env on read-only filesystem */
4701 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4707 if (!(env->me_flags & MDB_NOTLS)) {
4708 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4711 env->me_flags |= MDB_ENV_TXKEY;
4713 /* Windows TLS callbacks need help finding their TLS info. */
4714 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4718 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4722 /* Try to get exclusive lock. If we succeed, then
4723 * nobody is using the lock region and we should initialize it.
4725 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4728 size = GetFileSize(env->me_lfd, NULL);
4730 size = lseek(env->me_lfd, 0, SEEK_END);
4731 if (size == -1) goto fail_errno;
4733 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4734 if (size < rsize && *excl > 0) {
4736 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4737 || !SetEndOfFile(env->me_lfd))
4740 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4744 size = rsize - sizeof(MDB_txninfo);
4745 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4750 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4752 if (!mh) goto fail_errno;
4753 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4755 if (!env->me_txns) goto fail_errno;
4757 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4759 if (m == MAP_FAILED) goto fail_errno;
4765 BY_HANDLE_FILE_INFORMATION stbuf;
4774 if (!mdb_sec_inited) {
4775 InitializeSecurityDescriptor(&mdb_null_sd,
4776 SECURITY_DESCRIPTOR_REVISION);
4777 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4778 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4779 mdb_all_sa.bInheritHandle = FALSE;
4780 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4783 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4784 idbuf.volume = stbuf.dwVolumeSerialNumber;
4785 idbuf.nhigh = stbuf.nFileIndexHigh;
4786 idbuf.nlow = stbuf.nFileIndexLow;
4787 val.mv_data = &idbuf;
4788 val.mv_size = sizeof(idbuf);
4789 mdb_hash_enc(&val, encbuf);
4790 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4791 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4792 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4793 if (!env->me_rmutex) goto fail_errno;
4794 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4795 if (!env->me_wmutex) goto fail_errno;
4796 #elif defined(MDB_USE_POSIX_SEM)
4805 #if defined(__NetBSD__)
4806 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4808 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4809 idbuf.dev = stbuf.st_dev;
4810 idbuf.ino = stbuf.st_ino;
4811 val.mv_data = &idbuf;
4812 val.mv_size = sizeof(idbuf);
4813 mdb_hash_enc(&val, encbuf);
4814 #ifdef MDB_SHORT_SEMNAMES
4815 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4817 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4818 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4819 /* Clean up after a previous run, if needed: Try to
4820 * remove both semaphores before doing anything else.
4822 sem_unlink(env->me_txns->mti_rmname);
4823 sem_unlink(env->me_txns->mti_wmname);
4824 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4825 O_CREAT|O_EXCL, mode, 1);
4826 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4827 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4828 O_CREAT|O_EXCL, mode, 1);
4829 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4830 #else /* MDB_USE_POSIX_MUTEX: */
4831 pthread_mutexattr_t mattr;
4833 /* Solaris needs this before initing a robust mutex. Otherwise
4834 * it may skip the init and return EBUSY "seems someone already
4835 * inited" or EINVAL "it was inited differently".
4837 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
4838 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
4840 if ((rc = pthread_mutexattr_init(&mattr)))
4843 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
4844 #ifdef MDB_ROBUST_SUPPORTED
4845 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
4847 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
4848 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
4849 pthread_mutexattr_destroy(&mattr);
4852 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4854 env->me_txns->mti_magic = MDB_MAGIC;
4855 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4856 env->me_txns->mti_txnid = 0;
4857 env->me_txns->mti_numreaders = 0;
4860 if (env->me_txns->mti_magic != MDB_MAGIC) {
4861 DPUTS("lock region has invalid magic");
4865 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4866 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4867 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4868 rc = MDB_VERSION_MISMATCH;
4872 if (rc && rc != EACCES && rc != EAGAIN) {
4876 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4877 if (!env->me_rmutex) goto fail_errno;
4878 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4879 if (!env->me_wmutex) goto fail_errno;
4880 #elif defined(MDB_USE_POSIX_SEM)
4881 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4882 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4883 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4884 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4895 /** Only a subset of the @ref mdb_env flags can be changed
4896 * at runtime. Changing other flags requires closing the
4897 * environment and re-opening it with the new flags.
4899 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4900 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4901 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4903 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4904 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4908 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4913 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4916 flags |= env->me_flags;
4918 rc = mdb_fname_init(path, flags, &fname);
4922 if (flags & MDB_RDONLY) {
4923 /* silently ignore WRITEMAP when we're only getting read access */
4924 flags &= ~MDB_WRITEMAP;
4926 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4927 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4930 env->me_flags = flags |= MDB_ENV_ACTIVE;
4934 env->me_path = strdup(path);
4935 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4936 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4937 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4938 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4942 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4944 /* For RDONLY, get lockfile after we know datafile exists */
4945 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4946 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
4951 rc = mdb_fopen(env, &fname,
4952 (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR,
4957 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4958 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
4963 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4964 if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) {
4965 /* Synchronous fd for meta writes. Needed even with
4966 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4968 rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd);
4972 DPRINTF(("opened dbenv %p", (void *) env));
4974 rc = mdb_env_share_locks(env, &excl);
4978 if (!(flags & MDB_RDONLY)) {
4980 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4981 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4982 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4983 (txn = calloc(1, size)))
4985 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4986 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4987 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4988 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4990 txn->mt_dbxs = env->me_dbxs;
4991 txn->mt_flags = MDB_TXN_FINISHED;
5001 mdb_env_close0(env, excl);
5003 mdb_fname_destroy(fname);
5007 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5009 mdb_env_close0(MDB_env *env, int excl)
5013 if (!(env->me_flags & MDB_ENV_ACTIVE))
5016 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5018 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5019 free(env->me_dbxs[i].md_name.mv_data);
5024 free(env->me_dbiseqs);
5025 free(env->me_dbflags);
5027 free(env->me_dirty_list);
5029 mdb_midl_free(env->me_free_pgs);
5031 if (env->me_flags & MDB_ENV_TXKEY) {
5032 pthread_key_delete(env->me_txkey);
5034 /* Delete our key from the global list */
5035 for (i=0; i<mdb_tls_nkeys; i++)
5036 if (mdb_tls_keys[i] == env->me_txkey) {
5037 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5045 munmap(env->me_map, env->me_mapsize);
5047 if (env->me_mfd != INVALID_HANDLE_VALUE)
5048 (void) close(env->me_mfd);
5049 if (env->me_fd != INVALID_HANDLE_VALUE)
5050 (void) close(env->me_fd);
5052 MDB_PID_T pid = env->me_pid;
5053 /* Clearing readers is done in this function because
5054 * me_txkey with its destructor must be disabled first.
5056 * We skip the the reader mutex, so we touch only
5057 * data owned by this process (me_close_readers and
5058 * our readers), and clear each reader atomically.
5060 for (i = env->me_close_readers; --i >= 0; )
5061 if (env->me_txns->mti_readers[i].mr_pid == pid)
5062 env->me_txns->mti_readers[i].mr_pid = 0;
5064 if (env->me_rmutex) {
5065 CloseHandle(env->me_rmutex);
5066 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5068 /* Windows automatically destroys the mutexes when
5069 * the last handle closes.
5071 #elif defined(MDB_USE_POSIX_SEM)
5072 if (env->me_rmutex != SEM_FAILED) {
5073 sem_close(env->me_rmutex);
5074 if (env->me_wmutex != SEM_FAILED)
5075 sem_close(env->me_wmutex);
5076 /* If we have the filelock: If we are the
5077 * only remaining user, clean up semaphores.
5080 mdb_env_excl_lock(env, &excl);
5082 sem_unlink(env->me_txns->mti_rmname);
5083 sem_unlink(env->me_txns->mti_wmname);
5087 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5089 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5092 /* Unlock the lockfile. Windows would have unlocked it
5093 * after closing anyway, but not necessarily at once.
5095 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5098 (void) close(env->me_lfd);
5101 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5105 mdb_env_close(MDB_env *env)
5112 VGMEMP_DESTROY(env);
5113 while ((dp = env->me_dpages) != NULL) {
5114 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5115 env->me_dpages = dp->mp_next;
5119 mdb_env_close0(env, 0);
5123 /** Compare two items pointing at aligned size_t's */
5125 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5127 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5128 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5131 /** Compare two items pointing at aligned unsigned int's.
5133 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5134 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5137 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5139 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5140 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5143 /** Compare two items pointing at unsigned ints of unknown alignment.
5144 * Nodes and keys are guaranteed to be 2-byte aligned.
5147 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5149 #if BYTE_ORDER == LITTLE_ENDIAN
5150 unsigned short *u, *c;
5153 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5154 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5157 } while(!x && u > (unsigned short *)a->mv_data);
5160 unsigned short *u, *c, *end;
5163 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5164 u = (unsigned short *)a->mv_data;
5165 c = (unsigned short *)b->mv_data;
5168 } while(!x && u < end);
5173 /** Compare two items lexically */
5175 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5182 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5188 diff = memcmp(a->mv_data, b->mv_data, len);
5189 return diff ? diff : len_diff<0 ? -1 : len_diff;
5192 /** Compare two items in reverse byte order */
5194 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5196 const unsigned char *p1, *p2, *p1_lim;
5200 p1_lim = (const unsigned char *)a->mv_data;
5201 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5202 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5204 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5210 while (p1 > p1_lim) {
5211 diff = *--p1 - *--p2;
5215 return len_diff<0 ? -1 : len_diff;
5218 /** Search for key within a page, using binary search.
5219 * Returns the smallest entry larger or equal to the key.
5220 * If exactp is non-null, stores whether the found entry was an exact match
5221 * in *exactp (1 or 0).
5222 * Updates the cursor index with the index of the found entry.
5223 * If no entry larger or equal to the key is found, returns NULL.
5226 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5228 unsigned int i = 0, nkeys;
5231 MDB_page *mp = mc->mc_pg[mc->mc_top];
5232 MDB_node *node = NULL;
5237 nkeys = NUMKEYS(mp);
5239 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5240 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5243 low = IS_LEAF(mp) ? 0 : 1;
5245 cmp = mc->mc_dbx->md_cmp;
5247 /* Branch pages have no data, so if using integer keys,
5248 * alignment is guaranteed. Use faster mdb_cmp_int.
5250 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5251 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5258 nodekey.mv_size = mc->mc_db->md_pad;
5259 node = NODEPTR(mp, 0); /* fake */
5260 while (low <= high) {
5261 i = (low + high) >> 1;
5262 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5263 rc = cmp(key, &nodekey);
5264 DPRINTF(("found leaf index %u [%s], rc = %i",
5265 i, DKEY(&nodekey), rc));
5274 while (low <= high) {
5275 i = (low + high) >> 1;
5277 node = NODEPTR(mp, i);
5278 nodekey.mv_size = NODEKSZ(node);
5279 nodekey.mv_data = NODEKEY(node);
5281 rc = cmp(key, &nodekey);
5284 DPRINTF(("found leaf index %u [%s], rc = %i",
5285 i, DKEY(&nodekey), rc));
5287 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5288 i, DKEY(&nodekey), NODEPGNO(node), rc));
5299 if (rc > 0) { /* Found entry is less than the key. */
5300 i++; /* Skip to get the smallest entry larger than key. */
5302 node = NODEPTR(mp, i);
5305 *exactp = (rc == 0 && nkeys > 0);
5306 /* store the key index */
5307 mc->mc_ki[mc->mc_top] = i;
5309 /* There is no entry larger or equal to the key. */
5312 /* nodeptr is fake for LEAF2 */
5318 mdb_cursor_adjust(MDB_cursor *mc, func)
5322 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5323 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5330 /** Pop a page off the top of the cursor's stack. */
5332 mdb_cursor_pop(MDB_cursor *mc)
5335 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5336 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5342 mc->mc_flags &= ~C_INITIALIZED;
5347 /** Push a page onto the top of the cursor's stack.
5348 * Set #MDB_TXN_ERROR on failure.
5351 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5353 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5354 DDBI(mc), (void *) mc));
5356 if (mc->mc_snum >= CURSOR_STACK) {
5357 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5358 return MDB_CURSOR_FULL;
5361 mc->mc_top = mc->mc_snum++;
5362 mc->mc_pg[mc->mc_top] = mp;
5363 mc->mc_ki[mc->mc_top] = 0;
5368 /** Find the address of the page corresponding to a given page number.
5369 * Set #MDB_TXN_ERROR on failure.
5370 * @param[in] mc the cursor accessing the page.
5371 * @param[in] pgno the page number for the page to retrieve.
5372 * @param[out] ret address of a pointer where the page's address will be stored.
5373 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5374 * @return 0 on success, non-zero on failure.
5377 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
5379 MDB_txn *txn = mc->mc_txn;
5380 MDB_env *env = txn->mt_env;
5384 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5388 MDB_ID2L dl = tx2->mt_u.dirty_list;
5390 /* Spilled pages were dirtied in this txn and flushed
5391 * because the dirty list got full. Bring this page
5392 * back in from the map (but don't unspill it here,
5393 * leave that unless page_touch happens again).
5395 if (tx2->mt_spill_pgs) {
5396 MDB_ID pn = pgno << 1;
5397 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5398 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5399 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5404 unsigned x = mdb_mid2l_search(dl, pgno);
5405 if (x <= dl[0].mid && dl[x].mid == pgno) {
5411 } while ((tx2 = tx2->mt_parent) != NULL);
5414 if (pgno < txn->mt_next_pgno) {
5416 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5418 DPRINTF(("page %"Z"u not found", pgno));
5419 txn->mt_flags |= MDB_TXN_ERROR;
5420 return MDB_PAGE_NOTFOUND;
5430 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5431 * The cursor is at the root page, set up the rest of it.
5434 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5436 MDB_page *mp = mc->mc_pg[mc->mc_top];
5440 while (IS_BRANCH(mp)) {
5444 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5445 /* Don't assert on branch pages in the FreeDB. We can get here
5446 * while in the process of rebalancing a FreeDB branch page; we must
5447 * let that proceed. ITS#8336
5449 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
5450 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5452 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5454 if (flags & MDB_PS_LAST)
5455 i = NUMKEYS(mp) - 1;
5458 node = mdb_node_search(mc, key, &exact);
5460 i = NUMKEYS(mp) - 1;
5462 i = mc->mc_ki[mc->mc_top];
5464 mdb_cassert(mc, i > 0);
5468 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5471 mdb_cassert(mc, i < NUMKEYS(mp));
5472 node = NODEPTR(mp, i);
5474 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5477 mc->mc_ki[mc->mc_top] = i;
5478 if ((rc = mdb_cursor_push(mc, mp)))
5481 if (flags & MDB_PS_MODIFY) {
5482 if ((rc = mdb_page_touch(mc)) != 0)
5484 mp = mc->mc_pg[mc->mc_top];
5489 DPRINTF(("internal error, index points to a %02X page!?",
5491 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5492 return MDB_CORRUPTED;
5495 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5496 key ? DKEY(key) : "null"));
5497 mc->mc_flags |= C_INITIALIZED;
5498 mc->mc_flags &= ~C_EOF;
5503 /** Search for the lowest key under the current branch page.
5504 * This just bypasses a NUMKEYS check in the current page
5505 * before calling mdb_page_search_root(), because the callers
5506 * are all in situations where the current page is known to
5510 mdb_page_search_lowest(MDB_cursor *mc)
5512 MDB_page *mp = mc->mc_pg[mc->mc_top];
5513 MDB_node *node = NODEPTR(mp, 0);
5516 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5519 mc->mc_ki[mc->mc_top] = 0;
5520 if ((rc = mdb_cursor_push(mc, mp)))
5522 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5525 /** Search for the page a given key should be in.
5526 * Push it and its parent pages on the cursor stack.
5527 * @param[in,out] mc the cursor for this operation.
5528 * @param[in] key the key to search for, or NULL for first/last page.
5529 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5530 * are touched (updated with new page numbers).
5531 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5532 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5533 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5534 * @return 0 on success, non-zero on failure.
5537 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5542 /* Make sure the txn is still viable, then find the root from
5543 * the txn's db table and set it as the root of the cursor's stack.
5545 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5546 DPUTS("transaction may not be used now");
5549 /* Make sure we're using an up-to-date root */
5550 if (*mc->mc_dbflag & DB_STALE) {
5552 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5554 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5555 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5562 MDB_node *leaf = mdb_node_search(&mc2,
5563 &mc->mc_dbx->md_name, &exact);
5565 return MDB_NOTFOUND;
5566 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5567 return MDB_INCOMPATIBLE; /* not a named DB */
5568 rc = mdb_node_read(&mc2, leaf, &data);
5571 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5573 /* The txn may not know this DBI, or another process may
5574 * have dropped and recreated the DB with other flags.
5576 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5577 return MDB_INCOMPATIBLE;
5578 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5580 *mc->mc_dbflag &= ~DB_STALE;
5582 root = mc->mc_db->md_root;
5584 if (root == P_INVALID) { /* Tree is empty. */
5585 DPUTS("tree is empty");
5586 return MDB_NOTFOUND;
5590 mdb_cassert(mc, root > 1);
5591 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5592 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
5598 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5599 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5601 if (flags & MDB_PS_MODIFY) {
5602 if ((rc = mdb_page_touch(mc)))
5606 if (flags & MDB_PS_ROOTONLY)
5609 return mdb_page_search_root(mc, key, flags);
5613 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5615 MDB_txn *txn = mc->mc_txn;
5616 pgno_t pg = mp->mp_pgno;
5617 unsigned x = 0, ovpages = mp->mp_pages;
5618 MDB_env *env = txn->mt_env;
5619 MDB_IDL sl = txn->mt_spill_pgs;
5620 MDB_ID pn = pg << 1;
5623 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5624 /* If the page is dirty or on the spill list we just acquired it,
5625 * so we should give it back to our current free list, if any.
5626 * Otherwise put it onto the list of pages we freed in this txn.
5628 * Won't create me_pghead: me_pglast must be inited along with it.
5629 * Unsupported in nested txns: They would need to hide the page
5630 * range in ancestor txns' dirty and spilled lists.
5632 if (env->me_pghead &&
5634 ((mp->mp_flags & P_DIRTY) ||
5635 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5639 MDB_ID2 *dl, ix, iy;
5640 rc = mdb_midl_need(&env->me_pghead, ovpages);
5643 if (!(mp->mp_flags & P_DIRTY)) {
5644 /* This page is no longer spilled */
5651 /* Remove from dirty list */
5652 dl = txn->mt_u.dirty_list;
5654 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5660 mdb_cassert(mc, x > 1);
5662 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5663 txn->mt_flags |= MDB_TXN_ERROR;
5664 return MDB_CORRUPTED;
5667 txn->mt_dirty_room++;
5668 if (!(env->me_flags & MDB_WRITEMAP))
5669 mdb_dpage_free(env, mp);
5671 /* Insert in me_pghead */
5672 mop = env->me_pghead;
5673 j = mop[0] + ovpages;
5674 for (i = mop[0]; i && mop[i] < pg; i--)
5680 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5684 mc->mc_db->md_overflow_pages -= ovpages;
5688 /** Return the data associated with a given node.
5689 * @param[in] mc The cursor for this operation.
5690 * @param[in] leaf The node being read.
5691 * @param[out] data Updated to point to the node's data.
5692 * @return 0 on success, non-zero on failure.
5695 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
5697 MDB_page *omp; /* overflow page */
5701 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5702 data->mv_size = NODEDSZ(leaf);
5703 data->mv_data = NODEDATA(leaf);
5707 /* Read overflow data.
5709 data->mv_size = NODEDSZ(leaf);
5710 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5711 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
5712 DPRINTF(("read overflow page %"Z"u failed", pgno));
5715 data->mv_data = METADATA(omp);
5721 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5722 MDB_val *key, MDB_val *data)
5729 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5731 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5734 if (txn->mt_flags & MDB_TXN_BLOCKED)
5737 mdb_cursor_init(&mc, txn, dbi, &mx);
5738 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5741 /** Find a sibling for a page.
5742 * Replaces the page at the top of the cursor's stack with the
5743 * specified sibling, if one exists.
5744 * @param[in] mc The cursor for this operation.
5745 * @param[in] move_right Non-zero if the right sibling is requested,
5746 * otherwise the left sibling.
5747 * @return 0 on success, non-zero on failure.
5750 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5756 if (mc->mc_snum < 2) {
5757 return MDB_NOTFOUND; /* root has no siblings */
5761 DPRINTF(("parent page is page %"Z"u, index %u",
5762 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5764 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5765 : (mc->mc_ki[mc->mc_top] == 0)) {
5766 DPRINTF(("no more keys left, moving to %s sibling",
5767 move_right ? "right" : "left"));
5768 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5769 /* undo cursor_pop before returning */
5776 mc->mc_ki[mc->mc_top]++;
5778 mc->mc_ki[mc->mc_top]--;
5779 DPRINTF(("just moving to %s index key %u",
5780 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5782 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5784 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5785 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
5786 /* mc will be inconsistent if caller does mc_snum++ as above */
5787 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5791 mdb_cursor_push(mc, mp);
5793 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5798 /** Move the cursor to the next data item. */
5800 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5806 if ((mc->mc_flags & C_EOF) ||
5807 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
5808 return MDB_NOTFOUND;
5810 if (!(mc->mc_flags & C_INITIALIZED))
5811 return mdb_cursor_first(mc, key, data);
5813 mp = mc->mc_pg[mc->mc_top];
5815 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5816 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5817 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5818 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5819 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5820 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5821 if (rc == MDB_SUCCESS)
5822 MDB_GET_KEY(leaf, key);
5827 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5828 if (op == MDB_NEXT_DUP)
5829 return MDB_NOTFOUND;
5833 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5834 mdb_dbg_pgno(mp), (void *) mc));
5835 if (mc->mc_flags & C_DEL) {
5836 mc->mc_flags ^= C_DEL;
5840 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5841 DPUTS("=====> move to next sibling page");
5842 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5843 mc->mc_flags |= C_EOF;
5846 mp = mc->mc_pg[mc->mc_top];
5847 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5849 mc->mc_ki[mc->mc_top]++;
5852 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5853 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5856 key->mv_size = mc->mc_db->md_pad;
5857 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5861 mdb_cassert(mc, IS_LEAF(mp));
5862 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5864 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5865 mdb_xcursor_init1(mc, leaf);
5868 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5871 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5872 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5873 if (rc != MDB_SUCCESS)
5878 MDB_GET_KEY(leaf, key);
5882 /** Move the cursor to the previous data item. */
5884 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5890 if (!(mc->mc_flags & C_INITIALIZED)) {
5891 rc = mdb_cursor_last(mc, key, data);
5894 mc->mc_ki[mc->mc_top]++;
5897 mp = mc->mc_pg[mc->mc_top];
5899 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5900 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5901 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5902 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5903 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5904 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5905 if (rc == MDB_SUCCESS) {
5906 MDB_GET_KEY(leaf, key);
5907 mc->mc_flags &= ~C_EOF;
5913 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5914 if (op == MDB_PREV_DUP)
5915 return MDB_NOTFOUND;
5919 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5920 mdb_dbg_pgno(mp), (void *) mc));
5922 mc->mc_flags &= ~(C_EOF|C_DEL);
5924 if (mc->mc_ki[mc->mc_top] == 0) {
5925 DPUTS("=====> move to prev sibling page");
5926 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5929 mp = mc->mc_pg[mc->mc_top];
5930 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5931 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5933 mc->mc_ki[mc->mc_top]--;
5935 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5936 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5939 key->mv_size = mc->mc_db->md_pad;
5940 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5944 mdb_cassert(mc, IS_LEAF(mp));
5945 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5947 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5948 mdb_xcursor_init1(mc, leaf);
5951 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5954 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5955 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5956 if (rc != MDB_SUCCESS)
5961 MDB_GET_KEY(leaf, key);
5965 /** Set the cursor on a specific data item. */
5967 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5968 MDB_cursor_op op, int *exactp)
5972 MDB_node *leaf = NULL;
5975 if (key->mv_size == 0)
5976 return MDB_BAD_VALSIZE;
5979 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5981 /* See if we're already on the right page */
5982 if (mc->mc_flags & C_INITIALIZED) {
5985 mp = mc->mc_pg[mc->mc_top];
5987 mc->mc_ki[mc->mc_top] = 0;
5988 return MDB_NOTFOUND;
5990 if (mp->mp_flags & P_LEAF2) {
5991 nodekey.mv_size = mc->mc_db->md_pad;
5992 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5994 leaf = NODEPTR(mp, 0);
5995 MDB_GET_KEY2(leaf, nodekey);
5997 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5999 /* Probably happens rarely, but first node on the page
6000 * was the one we wanted.
6002 mc->mc_ki[mc->mc_top] = 0;
6009 unsigned int nkeys = NUMKEYS(mp);
6011 if (mp->mp_flags & P_LEAF2) {
6012 nodekey.mv_data = LEAF2KEY(mp,
6013 nkeys-1, nodekey.mv_size);
6015 leaf = NODEPTR(mp, nkeys-1);
6016 MDB_GET_KEY2(leaf, nodekey);
6018 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6020 /* last node was the one we wanted */
6021 mc->mc_ki[mc->mc_top] = nkeys-1;
6027 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6028 /* This is definitely the right page, skip search_page */
6029 if (mp->mp_flags & P_LEAF2) {
6030 nodekey.mv_data = LEAF2KEY(mp,
6031 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6033 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6034 MDB_GET_KEY2(leaf, nodekey);
6036 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6038 /* current node was the one we wanted */
6045 mc->mc_flags &= ~C_EOF;
6049 /* If any parents have right-sibs, search.
6050 * Otherwise, there's nothing further.
6052 for (i=0; i<mc->mc_top; i++)
6054 NUMKEYS(mc->mc_pg[i])-1)
6056 if (i == mc->mc_top) {
6057 /* There are no other pages */
6058 mc->mc_ki[mc->mc_top] = nkeys;
6059 return MDB_NOTFOUND;
6063 /* There are no other pages */
6064 mc->mc_ki[mc->mc_top] = 0;
6065 if (op == MDB_SET_RANGE && !exactp) {
6069 return MDB_NOTFOUND;
6075 rc = mdb_page_search(mc, key, 0);
6076 if (rc != MDB_SUCCESS)
6079 mp = mc->mc_pg[mc->mc_top];
6080 mdb_cassert(mc, IS_LEAF(mp));
6083 leaf = mdb_node_search(mc, key, exactp);
6084 if (exactp != NULL && !*exactp) {
6085 /* MDB_SET specified and not an exact match. */
6086 return MDB_NOTFOUND;
6090 DPUTS("===> inexact leaf not found, goto sibling");
6091 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6092 mc->mc_flags |= C_EOF;
6093 return rc; /* no entries matched */
6095 mp = mc->mc_pg[mc->mc_top];
6096 mdb_cassert(mc, IS_LEAF(mp));
6097 leaf = NODEPTR(mp, 0);
6101 mc->mc_flags |= C_INITIALIZED;
6102 mc->mc_flags &= ~C_EOF;
6105 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6106 key->mv_size = mc->mc_db->md_pad;
6107 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6112 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6113 mdb_xcursor_init1(mc, leaf);
6116 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6117 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6118 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6121 if (op == MDB_GET_BOTH) {
6127 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6128 if (rc != MDB_SUCCESS)
6131 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6134 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6136 dcmp = mc->mc_dbx->md_dcmp;
6137 #if UINT_MAX < SIZE_MAX
6138 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6139 dcmp = mdb_cmp_clong;
6141 rc = dcmp(data, &olddata);
6143 if (op == MDB_GET_BOTH || rc > 0)
6144 return MDB_NOTFOUND;
6151 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6152 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6157 /* The key already matches in all other cases */
6158 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6159 MDB_GET_KEY(leaf, key);
6160 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6165 /** Move the cursor to the first item in the database. */
6167 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6173 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6175 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6176 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6177 if (rc != MDB_SUCCESS)
6180 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6182 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6183 mc->mc_flags |= C_INITIALIZED;
6184 mc->mc_flags &= ~C_EOF;
6186 mc->mc_ki[mc->mc_top] = 0;
6188 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6189 key->mv_size = mc->mc_db->md_pad;
6190 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6195 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6196 mdb_xcursor_init1(mc, leaf);
6197 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6201 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6205 MDB_GET_KEY(leaf, key);
6209 /** Move the cursor to the last item in the database. */
6211 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6217 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6219 if (!(mc->mc_flags & C_EOF)) {
6221 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6222 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6223 if (rc != MDB_SUCCESS)
6226 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6229 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6230 mc->mc_flags |= C_INITIALIZED|C_EOF;
6231 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6233 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6234 key->mv_size = mc->mc_db->md_pad;
6235 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6240 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6241 mdb_xcursor_init1(mc, leaf);
6242 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6246 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6251 MDB_GET_KEY(leaf, key);
6256 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6261 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6266 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6270 case MDB_GET_CURRENT:
6271 if (!(mc->mc_flags & C_INITIALIZED)) {
6274 MDB_page *mp = mc->mc_pg[mc->mc_top];
6275 int nkeys = NUMKEYS(mp);
6276 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6277 mc->mc_ki[mc->mc_top] = nkeys;
6283 key->mv_size = mc->mc_db->md_pad;
6284 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6286 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6287 MDB_GET_KEY(leaf, key);
6289 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6290 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6292 rc = mdb_node_read(mc, leaf, data);
6299 case MDB_GET_BOTH_RANGE:
6304 if (mc->mc_xcursor == NULL) {
6305 rc = MDB_INCOMPATIBLE;
6315 rc = mdb_cursor_set(mc, key, data, op,
6316 op == MDB_SET_RANGE ? NULL : &exact);
6319 case MDB_GET_MULTIPLE:
6320 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6324 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6325 rc = MDB_INCOMPATIBLE;
6329 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6330 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6333 case MDB_NEXT_MULTIPLE:
6338 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6339 rc = MDB_INCOMPATIBLE;
6342 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6343 if (rc == MDB_SUCCESS) {
6344 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6347 mx = &mc->mc_xcursor->mx_cursor;
6348 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6350 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6351 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6357 case MDB_PREV_MULTIPLE:
6362 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6363 rc = MDB_INCOMPATIBLE;
6366 if (!(mc->mc_flags & C_INITIALIZED))
6367 rc = mdb_cursor_last(mc, key, data);
6370 if (rc == MDB_SUCCESS) {
6371 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
6372 if (mx->mc_flags & C_INITIALIZED) {
6373 rc = mdb_cursor_sibling(mx, 0);
6374 if (rc == MDB_SUCCESS)
6383 case MDB_NEXT_NODUP:
6384 rc = mdb_cursor_next(mc, key, data, op);
6388 case MDB_PREV_NODUP:
6389 rc = mdb_cursor_prev(mc, key, data, op);
6392 rc = mdb_cursor_first(mc, key, data);
6395 mfunc = mdb_cursor_first;
6397 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6401 if (mc->mc_xcursor == NULL) {
6402 rc = MDB_INCOMPATIBLE;
6406 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6407 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6408 MDB_GET_KEY(leaf, key);
6409 rc = mdb_node_read(mc, leaf, data);
6413 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6417 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6420 rc = mdb_cursor_last(mc, key, data);
6423 mfunc = mdb_cursor_last;
6426 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6431 if (mc->mc_flags & C_DEL)
6432 mc->mc_flags ^= C_DEL;
6437 /** Touch all the pages in the cursor stack. Set mc_top.
6438 * Makes sure all the pages are writable, before attempting a write operation.
6439 * @param[in] mc The cursor to operate on.
6442 mdb_cursor_touch(MDB_cursor *mc)
6444 int rc = MDB_SUCCESS;
6446 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & (DB_DIRTY|DB_DUPDATA))) {
6447 /* Touch DB record of named DB */
6450 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6452 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6453 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6456 *mc->mc_dbflag |= DB_DIRTY;
6461 rc = mdb_page_touch(mc);
6462 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6463 mc->mc_top = mc->mc_snum-1;
6468 /** Do not spill pages to disk if txn is getting full, may fail instead */
6469 #define MDB_NOSPILL 0x8000
6472 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6476 MDB_node *leaf = NULL;
6477 MDB_page *fp, *mp, *sub_root = NULL;
6479 MDB_val xdata, *rdata, dkey, olddata;
6481 int do_sub = 0, insert_key, insert_data;
6482 unsigned int mcount = 0, dcount = 0, nospill;
6485 unsigned int nflags;
6488 if (mc == NULL || key == NULL)
6491 env = mc->mc_txn->mt_env;
6493 /* Check this first so counter will always be zero on any
6496 if (flags & MDB_MULTIPLE) {
6497 dcount = data[1].mv_size;
6498 data[1].mv_size = 0;
6499 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6500 return MDB_INCOMPATIBLE;
6503 nospill = flags & MDB_NOSPILL;
6504 flags &= ~MDB_NOSPILL;
6506 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6507 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6509 if (key->mv_size-1 >= ENV_MAXKEY(env))
6510 return MDB_BAD_VALSIZE;
6512 #if SIZE_MAX > MAXDATASIZE
6513 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6514 return MDB_BAD_VALSIZE;
6516 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6517 return MDB_BAD_VALSIZE;
6520 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6521 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6525 if (flags == MDB_CURRENT) {
6526 if (!(mc->mc_flags & C_INITIALIZED))
6529 } else if (mc->mc_db->md_root == P_INVALID) {
6530 /* new database, cursor has nothing to point to */
6533 mc->mc_flags &= ~C_INITIALIZED;
6538 if (flags & MDB_APPEND) {
6540 rc = mdb_cursor_last(mc, &k2, &d2);
6542 rc = mc->mc_dbx->md_cmp(key, &k2);
6545 mc->mc_ki[mc->mc_top]++;
6547 /* new key is <= last key */
6552 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6554 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6555 DPRINTF(("duplicate key [%s]", DKEY(key)));
6557 return MDB_KEYEXIST;
6559 if (rc && rc != MDB_NOTFOUND)
6563 if (mc->mc_flags & C_DEL)
6564 mc->mc_flags ^= C_DEL;
6566 /* Cursor is positioned, check for room in the dirty list */
6568 if (flags & MDB_MULTIPLE) {
6570 xdata.mv_size = data->mv_size * dcount;
6574 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6578 if (rc == MDB_NO_ROOT) {
6580 /* new database, write a root leaf page */
6581 DPUTS("allocating new root leaf page");
6582 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6585 mdb_cursor_push(mc, np);
6586 mc->mc_db->md_root = np->mp_pgno;
6587 mc->mc_db->md_depth++;
6588 *mc->mc_dbflag |= DB_DIRTY;
6589 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6591 np->mp_flags |= P_LEAF2;
6592 mc->mc_flags |= C_INITIALIZED;
6594 /* make sure all cursor pages are writable */
6595 rc2 = mdb_cursor_touch(mc);
6600 insert_key = insert_data = rc;
6602 /* The key does not exist */
6603 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6604 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6605 LEAFSIZE(key, data) > env->me_nodemax)
6607 /* Too big for a node, insert in sub-DB. Set up an empty
6608 * "old sub-page" for prep_subDB to expand to a full page.
6610 fp_flags = P_LEAF|P_DIRTY;
6612 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6613 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6614 olddata.mv_size = PAGEHDRSZ;
6618 /* there's only a key anyway, so this is a no-op */
6619 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6621 unsigned int ksize = mc->mc_db->md_pad;
6622 if (key->mv_size != ksize)
6623 return MDB_BAD_VALSIZE;
6624 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6625 memcpy(ptr, key->mv_data, ksize);
6627 /* if overwriting slot 0 of leaf, need to
6628 * update branch key if there is a parent page
6630 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6631 unsigned short dtop = 1;
6633 /* slot 0 is always an empty key, find real slot */
6634 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6638 if (mc->mc_ki[mc->mc_top])
6639 rc2 = mdb_update_key(mc, key);
6650 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6651 olddata.mv_size = NODEDSZ(leaf);
6652 olddata.mv_data = NODEDATA(leaf);
6655 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6656 /* Prepare (sub-)page/sub-DB to accept the new item,
6657 * if needed. fp: old sub-page or a header faking
6658 * it. mp: new (sub-)page. offset: growth in page
6659 * size. xdata: node data with new page or DB.
6661 unsigned i, offset = 0;
6662 mp = fp = xdata.mv_data = env->me_pbuf;
6663 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6665 /* Was a single item before, must convert now */
6666 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6668 /* Just overwrite the current item */
6669 if (flags == MDB_CURRENT)
6671 dcmp = mc->mc_dbx->md_dcmp;
6672 #if UINT_MAX < SIZE_MAX
6673 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6674 dcmp = mdb_cmp_clong;
6676 /* does data match? */
6677 if (!dcmp(data, &olddata)) {
6678 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
6679 return MDB_KEYEXIST;
6684 /* Back up original data item */
6685 dkey.mv_size = olddata.mv_size;
6686 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6688 /* Make sub-page header for the dup items, with dummy body */
6689 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6690 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6691 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6692 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6693 fp->mp_flags |= P_LEAF2;
6694 fp->mp_pad = data->mv_size;
6695 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6697 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6698 (dkey.mv_size & 1) + (data->mv_size & 1);
6700 fp->mp_upper = xdata.mv_size - PAGEBASE;
6701 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6702 } else if (leaf->mn_flags & F_SUBDATA) {
6703 /* Data is on sub-DB, just store it */
6704 flags |= F_DUPDATA|F_SUBDATA;
6707 /* Data is on sub-page */
6708 fp = olddata.mv_data;
6711 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6712 offset = EVEN(NODESIZE + sizeof(indx_t) +
6716 offset = fp->mp_pad;
6717 if (SIZELEFT(fp) < offset) {
6718 offset *= 4; /* space for 4 more */
6721 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6723 fp->mp_flags |= P_DIRTY;
6724 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6725 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6729 xdata.mv_size = olddata.mv_size + offset;
6732 fp_flags = fp->mp_flags;
6733 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6734 /* Too big for a sub-page, convert to sub-DB */
6735 fp_flags &= ~P_SUBP;
6737 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6738 fp_flags |= P_LEAF2;
6739 dummy.md_pad = fp->mp_pad;
6740 dummy.md_flags = MDB_DUPFIXED;
6741 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6742 dummy.md_flags |= MDB_INTEGERKEY;
6748 dummy.md_branch_pages = 0;
6749 dummy.md_leaf_pages = 1;
6750 dummy.md_overflow_pages = 0;
6751 dummy.md_entries = NUMKEYS(fp);
6752 xdata.mv_size = sizeof(MDB_db);
6753 xdata.mv_data = &dummy;
6754 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6756 offset = env->me_psize - olddata.mv_size;
6757 flags |= F_DUPDATA|F_SUBDATA;
6758 dummy.md_root = mp->mp_pgno;
6762 mp->mp_flags = fp_flags | P_DIRTY;
6763 mp->mp_pad = fp->mp_pad;
6764 mp->mp_lower = fp->mp_lower;
6765 mp->mp_upper = fp->mp_upper + offset;
6766 if (fp_flags & P_LEAF2) {
6767 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6769 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6770 olddata.mv_size - fp->mp_upper - PAGEBASE);
6771 for (i=0; i<NUMKEYS(fp); i++)
6772 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6780 mdb_node_del(mc, 0);
6784 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6785 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6786 return MDB_INCOMPATIBLE;
6787 /* overflow page overwrites need special handling */
6788 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6791 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6793 memcpy(&pg, olddata.mv_data, sizeof(pg));
6794 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
6796 ovpages = omp->mp_pages;
6798 /* Is the ov page large enough? */
6799 if (ovpages >= dpages) {
6800 if (!(omp->mp_flags & P_DIRTY) &&
6801 (level || (env->me_flags & MDB_WRITEMAP)))
6803 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6806 level = 0; /* dirty in this txn or clean */
6809 if (omp->mp_flags & P_DIRTY) {
6810 /* yes, overwrite it. Note in this case we don't
6811 * bother to try shrinking the page if the new data
6812 * is smaller than the overflow threshold.
6815 /* It is writable only in a parent txn */
6816 size_t sz = (size_t) env->me_psize * ovpages, off;
6817 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6823 /* Note - this page is already counted in parent's dirty_room */
6824 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6825 mdb_cassert(mc, rc2 == 0);
6826 /* Currently we make the page look as with put() in the
6827 * parent txn, in case the user peeks at MDB_RESERVEd
6828 * or unused parts. Some users treat ovpages specially.
6830 if (!(flags & MDB_RESERVE)) {
6831 /* Skip the part where LMDB will put *data.
6832 * Copy end of page, adjusting alignment so
6833 * compiler may copy words instead of bytes.
6835 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6836 memcpy((size_t *)((char *)np + off),
6837 (size_t *)((char *)omp + off), sz - off);
6840 memcpy(np, omp, sz); /* Copy beginning of page */
6843 SETDSZ(leaf, data->mv_size);
6844 if (F_ISSET(flags, MDB_RESERVE))
6845 data->mv_data = METADATA(omp);
6847 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6851 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6853 } else if (data->mv_size == olddata.mv_size) {
6854 /* same size, just replace it. Note that we could
6855 * also reuse this node if the new data is smaller,
6856 * but instead we opt to shrink the node in that case.
6858 if (F_ISSET(flags, MDB_RESERVE))
6859 data->mv_data = olddata.mv_data;
6860 else if (!(mc->mc_flags & C_SUB))
6861 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6863 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6868 mdb_node_del(mc, 0);
6874 nflags = flags & NODE_ADD_FLAGS;
6875 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6876 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6877 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6878 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6880 nflags |= MDB_SPLIT_REPLACE;
6881 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6883 /* There is room already in this leaf page. */
6884 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6886 /* Adjust other cursors pointing to mp */
6887 MDB_cursor *m2, *m3;
6888 MDB_dbi dbi = mc->mc_dbi;
6889 unsigned i = mc->mc_top;
6890 MDB_page *mp = mc->mc_pg[i];
6892 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6893 if (mc->mc_flags & C_SUB)
6894 m3 = &m2->mc_xcursor->mx_cursor;
6897 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
6898 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
6901 if (XCURSOR_INITED(m3))
6902 XCURSOR_REFRESH(m3, mp, m3->mc_ki[i]);
6907 if (rc == MDB_SUCCESS) {
6908 /* Now store the actual data in the child DB. Note that we're
6909 * storing the user data in the keys field, so there are strict
6910 * size limits on dupdata. The actual data fields of the child
6911 * DB are all zero size.
6914 int xflags, new_dupdata;
6919 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6920 if (flags & MDB_CURRENT) {
6921 xflags = MDB_CURRENT|MDB_NOSPILL;
6923 mdb_xcursor_init1(mc, leaf);
6924 xflags = (flags & MDB_NODUPDATA) ?
6925 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6928 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6929 new_dupdata = (int)dkey.mv_size;
6930 /* converted, write the original data first */
6932 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6935 /* we've done our job */
6938 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6939 /* Adjust other cursors pointing to mp */
6941 MDB_xcursor *mx = mc->mc_xcursor;
6942 unsigned i = mc->mc_top;
6943 MDB_page *mp = mc->mc_pg[i];
6944 int nkeys = NUMKEYS(mp);
6946 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6947 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6948 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6949 if (m2->mc_pg[i] == mp) {
6950 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6951 mdb_xcursor_init2(m2, mx, new_dupdata);
6952 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
6953 XCURSOR_REFRESH(m2, mp, m2->mc_ki[i]);
6958 ecount = mc->mc_xcursor->mx_db.md_entries;
6959 if (flags & MDB_APPENDDUP)
6960 xflags |= MDB_APPEND;
6961 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6962 if (flags & F_SUBDATA) {
6963 void *db = NODEDATA(leaf);
6964 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6966 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6968 /* Increment count unless we just replaced an existing item. */
6970 mc->mc_db->md_entries++;
6972 /* Invalidate txn if we created an empty sub-DB */
6975 /* If we succeeded and the key didn't exist before,
6976 * make sure the cursor is marked valid.
6978 mc->mc_flags |= C_INITIALIZED;
6980 if (flags & MDB_MULTIPLE) {
6983 /* let caller know how many succeeded, if any */
6984 data[1].mv_size = mcount;
6985 if (mcount < dcount) {
6986 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6987 insert_key = insert_data = 0;
6994 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6997 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7002 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7008 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7009 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7011 if (!(mc->mc_flags & C_INITIALIZED))
7014 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7015 return MDB_NOTFOUND;
7017 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7020 rc = mdb_cursor_touch(mc);
7024 mp = mc->mc_pg[mc->mc_top];
7027 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7029 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7030 if (flags & MDB_NODUPDATA) {
7031 /* mdb_cursor_del0() will subtract the final entry */
7032 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7033 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7035 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7036 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7038 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7041 /* If sub-DB still has entries, we're done */
7042 if (mc->mc_xcursor->mx_db.md_entries) {
7043 if (leaf->mn_flags & F_SUBDATA) {
7044 /* update subDB info */
7045 void *db = NODEDATA(leaf);
7046 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7049 /* shrink fake page */
7050 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7051 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7052 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7053 /* fix other sub-DB cursors pointed at fake pages on this page */
7054 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7055 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7056 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7057 if (m2->mc_pg[mc->mc_top] == mp) {
7058 MDB_node *n2 = leaf;
7059 if (m2->mc_ki[mc->mc_top] != mc->mc_ki[mc->mc_top]) {
7060 n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7061 if (n2->mn_flags & F_SUBDATA) continue;
7063 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7067 mc->mc_db->md_entries--;
7070 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7072 /* otherwise fall thru and delete the sub-DB */
7075 if (leaf->mn_flags & F_SUBDATA) {
7076 /* add all the child DB's pages to the free list */
7077 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7082 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7083 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7084 rc = MDB_INCOMPATIBLE;
7088 /* add overflow pages to free list */
7089 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7093 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7094 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7095 (rc = mdb_ovpage_free(mc, omp)))
7100 return mdb_cursor_del0(mc);
7103 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7107 /** Allocate and initialize new pages for a database.
7108 * Set #MDB_TXN_ERROR on failure.
7109 * @param[in] mc a cursor on the database being added to.
7110 * @param[in] flags flags defining what type of page is being allocated.
7111 * @param[in] num the number of pages to allocate. This is usually 1,
7112 * unless allocating overflow pages for a large record.
7113 * @param[out] mp Address of a page, or NULL on failure.
7114 * @return 0 on success, non-zero on failure.
7117 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7122 if ((rc = mdb_page_alloc(mc, num, &np)))
7124 DPRINTF(("allocated new mpage %"Z"u, page size %u",
7125 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7126 np->mp_flags = flags | P_DIRTY;
7127 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7128 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7131 mc->mc_db->md_branch_pages++;
7132 else if (IS_LEAF(np))
7133 mc->mc_db->md_leaf_pages++;
7134 else if (IS_OVERFLOW(np)) {
7135 mc->mc_db->md_overflow_pages += num;
7143 /** Calculate the size of a leaf node.
7144 * The size depends on the environment's page size; if a data item
7145 * is too large it will be put onto an overflow page and the node
7146 * size will only include the key and not the data. Sizes are always
7147 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7148 * of the #MDB_node headers.
7149 * @param[in] env The environment handle.
7150 * @param[in] key The key for the node.
7151 * @param[in] data The data for the node.
7152 * @return The number of bytes needed to store the node.
7155 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7159 sz = LEAFSIZE(key, data);
7160 if (sz > env->me_nodemax) {
7161 /* put on overflow page */
7162 sz -= data->mv_size - sizeof(pgno_t);
7165 return EVEN(sz + sizeof(indx_t));
7168 /** Calculate the size of a branch node.
7169 * The size should depend on the environment's page size but since
7170 * we currently don't support spilling large keys onto overflow
7171 * pages, it's simply the size of the #MDB_node header plus the
7172 * size of the key. Sizes are always rounded up to an even number
7173 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7174 * @param[in] env The environment handle.
7175 * @param[in] key The key for the node.
7176 * @return The number of bytes needed to store the node.
7179 mdb_branch_size(MDB_env *env, MDB_val *key)
7184 if (sz > env->me_nodemax) {
7185 /* put on overflow page */
7186 /* not implemented */
7187 /* sz -= key->size - sizeof(pgno_t); */
7190 return sz + sizeof(indx_t);
7193 /** Add a node to the page pointed to by the cursor.
7194 * Set #MDB_TXN_ERROR on failure.
7195 * @param[in] mc The cursor for this operation.
7196 * @param[in] indx The index on the page where the new node should be added.
7197 * @param[in] key The key for the new node.
7198 * @param[in] data The data for the new node, if any.
7199 * @param[in] pgno The page number, if adding a branch node.
7200 * @param[in] flags Flags for the node.
7201 * @return 0 on success, non-zero on failure. Possible errors are:
7203 * <li>ENOMEM - failed to allocate overflow pages for the node.
7204 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7205 * should never happen since all callers already calculate the
7206 * page's free space before calling this function.
7210 mdb_node_add(MDB_cursor *mc, indx_t indx,
7211 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7214 size_t node_size = NODESIZE;
7218 MDB_page *mp = mc->mc_pg[mc->mc_top];
7219 MDB_page *ofp = NULL; /* overflow page */
7223 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7225 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7226 IS_LEAF(mp) ? "leaf" : "branch",
7227 IS_SUBP(mp) ? "sub-" : "",
7228 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7229 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7232 /* Move higher keys up one slot. */
7233 int ksize = mc->mc_db->md_pad, dif;
7234 char *ptr = LEAF2KEY(mp, indx, ksize);
7235 dif = NUMKEYS(mp) - indx;
7237 memmove(ptr+ksize, ptr, dif*ksize);
7238 /* insert new key */
7239 memcpy(ptr, key->mv_data, ksize);
7241 /* Just using these for counting */
7242 mp->mp_lower += sizeof(indx_t);
7243 mp->mp_upper -= ksize - sizeof(indx_t);
7247 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7249 node_size += key->mv_size;
7251 mdb_cassert(mc, key && data);
7252 if (F_ISSET(flags, F_BIGDATA)) {
7253 /* Data already on overflow page. */
7254 node_size += sizeof(pgno_t);
7255 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7256 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7258 /* Put data on overflow page. */
7259 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7260 data->mv_size, node_size+data->mv_size));
7261 node_size = EVEN(node_size + sizeof(pgno_t));
7262 if ((ssize_t)node_size > room)
7264 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7266 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7270 node_size += data->mv_size;
7273 node_size = EVEN(node_size);
7274 if ((ssize_t)node_size > room)
7278 /* Move higher pointers up one slot. */
7279 for (i = NUMKEYS(mp); i > indx; i--)
7280 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7282 /* Adjust free space offsets. */
7283 ofs = mp->mp_upper - node_size;
7284 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7285 mp->mp_ptrs[indx] = ofs;
7287 mp->mp_lower += sizeof(indx_t);
7289 /* Write the node data. */
7290 node = NODEPTR(mp, indx);
7291 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7292 node->mn_flags = flags;
7294 SETDSZ(node,data->mv_size);
7299 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7302 ndata = NODEDATA(node);
7304 if (F_ISSET(flags, F_BIGDATA))
7305 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7306 else if (F_ISSET(flags, MDB_RESERVE))
7307 data->mv_data = ndata;
7309 memcpy(ndata, data->mv_data, data->mv_size);
7311 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7312 ndata = METADATA(ofp);
7313 if (F_ISSET(flags, MDB_RESERVE))
7314 data->mv_data = ndata;
7316 memcpy(ndata, data->mv_data, data->mv_size);
7323 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7324 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7325 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7326 DPRINTF(("node size = %"Z"u", node_size));
7327 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7328 return MDB_PAGE_FULL;
7331 /** Delete the specified node from a page.
7332 * @param[in] mc Cursor pointing to the node to delete.
7333 * @param[in] ksize The size of a node. Only used if the page is
7334 * part of a #MDB_DUPFIXED database.
7337 mdb_node_del(MDB_cursor *mc, int ksize)
7339 MDB_page *mp = mc->mc_pg[mc->mc_top];
7340 indx_t indx = mc->mc_ki[mc->mc_top];
7342 indx_t i, j, numkeys, ptr;
7346 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7347 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7348 numkeys = NUMKEYS(mp);
7349 mdb_cassert(mc, indx < numkeys);
7352 int x = numkeys - 1 - indx;
7353 base = LEAF2KEY(mp, indx, ksize);
7355 memmove(base, base + ksize, x * ksize);
7356 mp->mp_lower -= sizeof(indx_t);
7357 mp->mp_upper += ksize - sizeof(indx_t);
7361 node = NODEPTR(mp, indx);
7362 sz = NODESIZE + node->mn_ksize;
7364 if (F_ISSET(node->mn_flags, F_BIGDATA))
7365 sz += sizeof(pgno_t);
7367 sz += NODEDSZ(node);
7371 ptr = mp->mp_ptrs[indx];
7372 for (i = j = 0; i < numkeys; i++) {
7374 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7375 if (mp->mp_ptrs[i] < ptr)
7376 mp->mp_ptrs[j] += sz;
7381 base = (char *)mp + mp->mp_upper + PAGEBASE;
7382 memmove(base + sz, base, ptr - mp->mp_upper);
7384 mp->mp_lower -= sizeof(indx_t);
7388 /** Compact the main page after deleting a node on a subpage.
7389 * @param[in] mp The main page to operate on.
7390 * @param[in] indx The index of the subpage on the main page.
7393 mdb_node_shrink(MDB_page *mp, indx_t indx)
7398 indx_t delta, nsize, len, ptr;
7401 node = NODEPTR(mp, indx);
7402 sp = (MDB_page *)NODEDATA(node);
7403 delta = SIZELEFT(sp);
7404 nsize = NODEDSZ(node) - delta;
7406 /* Prepare to shift upward, set len = length(subpage part to shift) */
7410 return; /* do not make the node uneven-sized */
7412 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7413 for (i = NUMKEYS(sp); --i >= 0; )
7414 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7417 sp->mp_upper = sp->mp_lower;
7418 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7419 SETDSZ(node, nsize);
7421 /* Shift <lower nodes...initial part of subpage> upward */
7422 base = (char *)mp + mp->mp_upper + PAGEBASE;
7423 memmove(base + delta, base, (char *)sp + len - base);
7425 ptr = mp->mp_ptrs[indx];
7426 for (i = NUMKEYS(mp); --i >= 0; ) {
7427 if (mp->mp_ptrs[i] <= ptr)
7428 mp->mp_ptrs[i] += delta;
7430 mp->mp_upper += delta;
7433 /** Initial setup of a sorted-dups cursor.
7434 * Sorted duplicates are implemented as a sub-database for the given key.
7435 * The duplicate data items are actually keys of the sub-database.
7436 * Operations on the duplicate data items are performed using a sub-cursor
7437 * initialized when the sub-database is first accessed. This function does
7438 * the preliminary setup of the sub-cursor, filling in the fields that
7439 * depend only on the parent DB.
7440 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7443 mdb_xcursor_init0(MDB_cursor *mc)
7445 MDB_xcursor *mx = mc->mc_xcursor;
7447 mx->mx_cursor.mc_xcursor = NULL;
7448 mx->mx_cursor.mc_txn = mc->mc_txn;
7449 mx->mx_cursor.mc_db = &mx->mx_db;
7450 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7451 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7452 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7453 mx->mx_cursor.mc_snum = 0;
7454 mx->mx_cursor.mc_top = 0;
7455 mx->mx_cursor.mc_flags = C_SUB;
7456 mx->mx_dbx.md_name.mv_size = 0;
7457 mx->mx_dbx.md_name.mv_data = NULL;
7458 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7459 mx->mx_dbx.md_dcmp = NULL;
7460 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7463 /** Final setup of a sorted-dups cursor.
7464 * Sets up the fields that depend on the data from the main cursor.
7465 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7466 * @param[in] node The data containing the #MDB_db record for the
7467 * sorted-dup database.
7470 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7472 MDB_xcursor *mx = mc->mc_xcursor;
7474 if (node->mn_flags & F_SUBDATA) {
7475 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7476 mx->mx_cursor.mc_pg[0] = 0;
7477 mx->mx_cursor.mc_snum = 0;
7478 mx->mx_cursor.mc_top = 0;
7479 mx->mx_cursor.mc_flags = C_SUB;
7481 MDB_page *fp = NODEDATA(node);
7482 mx->mx_db.md_pad = 0;
7483 mx->mx_db.md_flags = 0;
7484 mx->mx_db.md_depth = 1;
7485 mx->mx_db.md_branch_pages = 0;
7486 mx->mx_db.md_leaf_pages = 1;
7487 mx->mx_db.md_overflow_pages = 0;
7488 mx->mx_db.md_entries = NUMKEYS(fp);
7489 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7490 mx->mx_cursor.mc_snum = 1;
7491 mx->mx_cursor.mc_top = 0;
7492 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7493 mx->mx_cursor.mc_pg[0] = fp;
7494 mx->mx_cursor.mc_ki[0] = 0;
7495 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7496 mx->mx_db.md_flags = MDB_DUPFIXED;
7497 mx->mx_db.md_pad = fp->mp_pad;
7498 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7499 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7502 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7503 mx->mx_db.md_root));
7504 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7505 #if UINT_MAX < SIZE_MAX
7506 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7507 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7512 /** Fixup a sorted-dups cursor due to underlying update.
7513 * Sets up some fields that depend on the data from the main cursor.
7514 * Almost the same as init1, but skips initialization steps if the
7515 * xcursor had already been used.
7516 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7517 * @param[in] src_mx The xcursor of an up-to-date cursor.
7518 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7521 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7523 MDB_xcursor *mx = mc->mc_xcursor;
7526 mx->mx_cursor.mc_snum = 1;
7527 mx->mx_cursor.mc_top = 0;
7528 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7529 mx->mx_cursor.mc_ki[0] = 0;
7530 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7531 #if UINT_MAX < SIZE_MAX
7532 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7534 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7537 mx->mx_db = src_mx->mx_db;
7538 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7539 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7540 mx->mx_db.md_root));
7543 /** Initialize a cursor for a given transaction and database. */
7545 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7548 mc->mc_backup = NULL;
7551 mc->mc_db = &txn->mt_dbs[dbi];
7552 mc->mc_dbx = &txn->mt_dbxs[dbi];
7553 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7559 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7560 mdb_tassert(txn, mx != NULL);
7561 mc->mc_xcursor = mx;
7562 mdb_xcursor_init0(mc);
7564 mc->mc_xcursor = NULL;
7566 if (*mc->mc_dbflag & DB_STALE) {
7567 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7572 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7575 size_t size = sizeof(MDB_cursor);
7577 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7580 if (txn->mt_flags & MDB_TXN_BLOCKED)
7583 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7586 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7587 size += sizeof(MDB_xcursor);
7589 if ((mc = malloc(size)) != NULL) {
7590 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7591 if (txn->mt_cursors) {
7592 mc->mc_next = txn->mt_cursors[dbi];
7593 txn->mt_cursors[dbi] = mc;
7594 mc->mc_flags |= C_UNTRACK;
7606 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7608 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7611 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7614 if (txn->mt_flags & MDB_TXN_BLOCKED)
7617 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7621 /* Return the count of duplicate data items for the current key */
7623 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7627 if (mc == NULL || countp == NULL)
7630 if (mc->mc_xcursor == NULL)
7631 return MDB_INCOMPATIBLE;
7633 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7636 if (!(mc->mc_flags & C_INITIALIZED))
7639 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7640 return MDB_NOTFOUND;
7642 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7643 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7646 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7649 *countp = mc->mc_xcursor->mx_db.md_entries;
7655 mdb_cursor_close(MDB_cursor *mc)
7657 if (mc && !mc->mc_backup) {
7658 /* remove from txn, if tracked */
7659 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7660 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7661 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7663 *prev = mc->mc_next;
7670 mdb_cursor_txn(MDB_cursor *mc)
7672 if (!mc) return NULL;
7677 mdb_cursor_dbi(MDB_cursor *mc)
7682 /** Replace the key for a branch node with a new key.
7683 * Set #MDB_TXN_ERROR on failure.
7684 * @param[in] mc Cursor pointing to the node to operate on.
7685 * @param[in] key The new key to use.
7686 * @return 0 on success, non-zero on failure.
7689 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7695 int delta, ksize, oksize;
7696 indx_t ptr, i, numkeys, indx;
7699 indx = mc->mc_ki[mc->mc_top];
7700 mp = mc->mc_pg[mc->mc_top];
7701 node = NODEPTR(mp, indx);
7702 ptr = mp->mp_ptrs[indx];
7706 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7707 k2.mv_data = NODEKEY(node);
7708 k2.mv_size = node->mn_ksize;
7709 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7711 mdb_dkey(&k2, kbuf2),
7717 /* Sizes must be 2-byte aligned. */
7718 ksize = EVEN(key->mv_size);
7719 oksize = EVEN(node->mn_ksize);
7720 delta = ksize - oksize;
7722 /* Shift node contents if EVEN(key length) changed. */
7724 if (delta > 0 && SIZELEFT(mp) < delta) {
7726 /* not enough space left, do a delete and split */
7727 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7728 pgno = NODEPGNO(node);
7729 mdb_node_del(mc, 0);
7730 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7733 numkeys = NUMKEYS(mp);
7734 for (i = 0; i < numkeys; i++) {
7735 if (mp->mp_ptrs[i] <= ptr)
7736 mp->mp_ptrs[i] -= delta;
7739 base = (char *)mp + mp->mp_upper + PAGEBASE;
7740 len = ptr - mp->mp_upper + NODESIZE;
7741 memmove(base - delta, base, len);
7742 mp->mp_upper -= delta;
7744 node = NODEPTR(mp, indx);
7747 /* But even if no shift was needed, update ksize */
7748 if (node->mn_ksize != key->mv_size)
7749 node->mn_ksize = key->mv_size;
7752 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7758 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7760 /** Perform \b act while tracking temporary cursor \b mn */
7761 #define WITH_CURSOR_TRACKING(mn, act) do { \
7762 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
7763 if ((mn).mc_flags & C_SUB) { \
7764 dummy.mc_flags = C_INITIALIZED; \
7765 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
7770 tracked->mc_next = *tp; \
7773 *tp = tracked->mc_next; \
7776 /** Move a node from csrc to cdst.
7779 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
7786 unsigned short flags;
7790 /* Mark src and dst as dirty. */
7791 if ((rc = mdb_page_touch(csrc)) ||
7792 (rc = mdb_page_touch(cdst)))
7795 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7796 key.mv_size = csrc->mc_db->md_pad;
7797 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7799 data.mv_data = NULL;
7803 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7804 mdb_cassert(csrc, !((size_t)srcnode & 1));
7805 srcpg = NODEPGNO(srcnode);
7806 flags = srcnode->mn_flags;
7807 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7808 unsigned int snum = csrc->mc_snum;
7810 /* must find the lowest key below src */
7811 rc = mdb_page_search_lowest(csrc);
7814 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7815 key.mv_size = csrc->mc_db->md_pad;
7816 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7818 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7819 key.mv_size = NODEKSZ(s2);
7820 key.mv_data = NODEKEY(s2);
7822 csrc->mc_snum = snum--;
7823 csrc->mc_top = snum;
7825 key.mv_size = NODEKSZ(srcnode);
7826 key.mv_data = NODEKEY(srcnode);
7828 data.mv_size = NODEDSZ(srcnode);
7829 data.mv_data = NODEDATA(srcnode);
7831 mn.mc_xcursor = NULL;
7832 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7833 unsigned int snum = cdst->mc_snum;
7836 /* must find the lowest key below dst */
7837 mdb_cursor_copy(cdst, &mn);
7838 rc = mdb_page_search_lowest(&mn);
7841 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7842 bkey.mv_size = mn.mc_db->md_pad;
7843 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7845 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7846 bkey.mv_size = NODEKSZ(s2);
7847 bkey.mv_data = NODEKEY(s2);
7849 mn.mc_snum = snum--;
7852 rc = mdb_update_key(&mn, &bkey);
7857 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7858 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7859 csrc->mc_ki[csrc->mc_top],
7861 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7862 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7864 /* Add the node to the destination page.
7866 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7867 if (rc != MDB_SUCCESS)
7870 /* Delete the node from the source page.
7872 mdb_node_del(csrc, key.mv_size);
7875 /* Adjust other cursors pointing to mp */
7876 MDB_cursor *m2, *m3;
7877 MDB_dbi dbi = csrc->mc_dbi;
7878 MDB_page *mpd, *mps;
7880 mps = csrc->mc_pg[csrc->mc_top];
7881 /* If we're adding on the left, bump others up */
7883 mpd = cdst->mc_pg[csrc->mc_top];
7884 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7885 if (csrc->mc_flags & C_SUB)
7886 m3 = &m2->mc_xcursor->mx_cursor;
7889 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7892 m3->mc_pg[csrc->mc_top] == mpd &&
7893 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7894 m3->mc_ki[csrc->mc_top]++;
7897 m3->mc_pg[csrc->mc_top] == mps &&
7898 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7899 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7900 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7901 m3->mc_ki[csrc->mc_top-1]++;
7903 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7904 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7907 /* Adding on the right, bump others down */
7909 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7910 if (csrc->mc_flags & C_SUB)
7911 m3 = &m2->mc_xcursor->mx_cursor;
7914 if (m3 == csrc) continue;
7915 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7917 if (m3->mc_pg[csrc->mc_top] == mps) {
7918 if (!m3->mc_ki[csrc->mc_top]) {
7919 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7920 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7921 m3->mc_ki[csrc->mc_top-1]--;
7923 m3->mc_ki[csrc->mc_top]--;
7925 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7926 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7932 /* Update the parent separators.
7934 if (csrc->mc_ki[csrc->mc_top] == 0) {
7935 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7936 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7937 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7939 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7940 key.mv_size = NODEKSZ(srcnode);
7941 key.mv_data = NODEKEY(srcnode);
7943 DPRINTF(("update separator for source page %"Z"u to [%s]",
7944 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7945 mdb_cursor_copy(csrc, &mn);
7948 /* We want mdb_rebalance to find mn when doing fixups */
7949 WITH_CURSOR_TRACKING(mn,
7950 rc = mdb_update_key(&mn, &key));
7954 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7956 indx_t ix = csrc->mc_ki[csrc->mc_top];
7957 nullkey.mv_size = 0;
7958 csrc->mc_ki[csrc->mc_top] = 0;
7959 rc = mdb_update_key(csrc, &nullkey);
7960 csrc->mc_ki[csrc->mc_top] = ix;
7961 mdb_cassert(csrc, rc == MDB_SUCCESS);
7965 if (cdst->mc_ki[cdst->mc_top] == 0) {
7966 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7967 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7968 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7970 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7971 key.mv_size = NODEKSZ(srcnode);
7972 key.mv_data = NODEKEY(srcnode);
7974 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7975 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7976 mdb_cursor_copy(cdst, &mn);
7979 /* We want mdb_rebalance to find mn when doing fixups */
7980 WITH_CURSOR_TRACKING(mn,
7981 rc = mdb_update_key(&mn, &key));
7985 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7987 indx_t ix = cdst->mc_ki[cdst->mc_top];
7988 nullkey.mv_size = 0;
7989 cdst->mc_ki[cdst->mc_top] = 0;
7990 rc = mdb_update_key(cdst, &nullkey);
7991 cdst->mc_ki[cdst->mc_top] = ix;
7992 mdb_cassert(cdst, rc == MDB_SUCCESS);
7999 /** Merge one page into another.
8000 * The nodes from the page pointed to by \b csrc will
8001 * be copied to the page pointed to by \b cdst and then
8002 * the \b csrc page will be freed.
8003 * @param[in] csrc Cursor pointing to the source page.
8004 * @param[in] cdst Cursor pointing to the destination page.
8005 * @return 0 on success, non-zero on failure.
8008 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8010 MDB_page *psrc, *pdst;
8017 psrc = csrc->mc_pg[csrc->mc_top];
8018 pdst = cdst->mc_pg[cdst->mc_top];
8020 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
8022 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8023 mdb_cassert(csrc, cdst->mc_snum > 1);
8025 /* Mark dst as dirty. */
8026 if ((rc = mdb_page_touch(cdst)))
8029 /* get dst page again now that we've touched it. */
8030 pdst = cdst->mc_pg[cdst->mc_top];
8032 /* Move all nodes from src to dst.
8034 j = nkeys = NUMKEYS(pdst);
8035 if (IS_LEAF2(psrc)) {
8036 key.mv_size = csrc->mc_db->md_pad;
8037 key.mv_data = METADATA(psrc);
8038 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8039 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8040 if (rc != MDB_SUCCESS)
8042 key.mv_data = (char *)key.mv_data + key.mv_size;
8045 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8046 srcnode = NODEPTR(psrc, i);
8047 if (i == 0 && IS_BRANCH(psrc)) {
8050 mdb_cursor_copy(csrc, &mn);
8051 mn.mc_xcursor = NULL;
8052 /* must find the lowest key below src */
8053 rc = mdb_page_search_lowest(&mn);
8056 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8057 key.mv_size = mn.mc_db->md_pad;
8058 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8060 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8061 key.mv_size = NODEKSZ(s2);
8062 key.mv_data = NODEKEY(s2);
8065 key.mv_size = srcnode->mn_ksize;
8066 key.mv_data = NODEKEY(srcnode);
8069 data.mv_size = NODEDSZ(srcnode);
8070 data.mv_data = NODEDATA(srcnode);
8071 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8072 if (rc != MDB_SUCCESS)
8077 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
8078 pdst->mp_pgno, NUMKEYS(pdst),
8079 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8081 /* Unlink the src page from parent and add to free list.
8084 mdb_node_del(csrc, 0);
8085 if (csrc->mc_ki[csrc->mc_top] == 0) {
8087 rc = mdb_update_key(csrc, &key);
8095 psrc = csrc->mc_pg[csrc->mc_top];
8096 /* If not operating on FreeDB, allow this page to be reused
8097 * in this txn. Otherwise just add to free list.
8099 rc = mdb_page_loose(csrc, psrc);
8103 csrc->mc_db->md_leaf_pages--;
8105 csrc->mc_db->md_branch_pages--;
8107 /* Adjust other cursors pointing to mp */
8108 MDB_cursor *m2, *m3;
8109 MDB_dbi dbi = csrc->mc_dbi;
8110 unsigned int top = csrc->mc_top;
8112 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8113 if (csrc->mc_flags & C_SUB)
8114 m3 = &m2->mc_xcursor->mx_cursor;
8117 if (m3 == csrc) continue;
8118 if (m3->mc_snum < csrc->mc_snum) continue;
8119 if (m3->mc_pg[top] == psrc) {
8120 m3->mc_pg[top] = pdst;
8121 m3->mc_ki[top] += nkeys;
8122 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8123 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8124 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8127 if (XCURSOR_INITED(m3) && IS_LEAF(psrc))
8128 XCURSOR_REFRESH(m3, m3->mc_pg[top], m3->mc_ki[top]);
8132 unsigned int snum = cdst->mc_snum;
8133 uint16_t depth = cdst->mc_db->md_depth;
8134 mdb_cursor_pop(cdst);
8135 rc = mdb_rebalance(cdst);
8136 /* Did the tree height change? */
8137 if (depth != cdst->mc_db->md_depth)
8138 snum += cdst->mc_db->md_depth - depth;
8139 cdst->mc_snum = snum;
8140 cdst->mc_top = snum-1;
8145 /** Copy the contents of a cursor.
8146 * @param[in] csrc The cursor to copy from.
8147 * @param[out] cdst The cursor to copy to.
8150 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8154 cdst->mc_txn = csrc->mc_txn;
8155 cdst->mc_dbi = csrc->mc_dbi;
8156 cdst->mc_db = csrc->mc_db;
8157 cdst->mc_dbx = csrc->mc_dbx;
8158 cdst->mc_snum = csrc->mc_snum;
8159 cdst->mc_top = csrc->mc_top;
8160 cdst->mc_flags = csrc->mc_flags;
8162 for (i=0; i<csrc->mc_snum; i++) {
8163 cdst->mc_pg[i] = csrc->mc_pg[i];
8164 cdst->mc_ki[i] = csrc->mc_ki[i];
8168 /** Rebalance the tree after a delete operation.
8169 * @param[in] mc Cursor pointing to the page where rebalancing
8171 * @return 0 on success, non-zero on failure.
8174 mdb_rebalance(MDB_cursor *mc)
8178 unsigned int ptop, minkeys, thresh;
8182 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8187 thresh = FILL_THRESHOLD;
8189 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
8190 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8191 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8192 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8194 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8195 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8196 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
8197 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8201 if (mc->mc_snum < 2) {
8202 MDB_page *mp = mc->mc_pg[0];
8204 DPUTS("Can't rebalance a subpage, ignoring");
8207 if (NUMKEYS(mp) == 0) {
8208 DPUTS("tree is completely empty");
8209 mc->mc_db->md_root = P_INVALID;
8210 mc->mc_db->md_depth = 0;
8211 mc->mc_db->md_leaf_pages = 0;
8212 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8215 /* Adjust cursors pointing to mp */
8218 mc->mc_flags &= ~C_INITIALIZED;
8220 MDB_cursor *m2, *m3;
8221 MDB_dbi dbi = mc->mc_dbi;
8223 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8224 if (mc->mc_flags & C_SUB)
8225 m3 = &m2->mc_xcursor->mx_cursor;
8228 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8230 if (m3->mc_pg[0] == mp) {
8233 m3->mc_flags &= ~C_INITIALIZED;
8237 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8239 DPUTS("collapsing root page!");
8240 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8243 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8244 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
8247 mc->mc_db->md_depth--;
8248 mc->mc_db->md_branch_pages--;
8249 mc->mc_ki[0] = mc->mc_ki[1];
8250 for (i = 1; i<mc->mc_db->md_depth; i++) {
8251 mc->mc_pg[i] = mc->mc_pg[i+1];
8252 mc->mc_ki[i] = mc->mc_ki[i+1];
8255 /* Adjust other cursors pointing to mp */
8256 MDB_cursor *m2, *m3;
8257 MDB_dbi dbi = mc->mc_dbi;
8259 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8260 if (mc->mc_flags & C_SUB)
8261 m3 = &m2->mc_xcursor->mx_cursor;
8264 if (m3 == mc) continue;
8265 if (!(m3->mc_flags & C_INITIALIZED))
8267 if (m3->mc_pg[0] == mp) {
8268 for (i=0; i<mc->mc_db->md_depth; i++) {
8269 m3->mc_pg[i] = m3->mc_pg[i+1];
8270 m3->mc_ki[i] = m3->mc_ki[i+1];
8278 DPUTS("root page doesn't need rebalancing");
8282 /* The parent (branch page) must have at least 2 pointers,
8283 * otherwise the tree is invalid.
8285 ptop = mc->mc_top-1;
8286 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8288 /* Leaf page fill factor is below the threshold.
8289 * Try to move keys from left or right neighbor, or
8290 * merge with a neighbor page.
8295 mdb_cursor_copy(mc, &mn);
8296 mn.mc_xcursor = NULL;
8298 oldki = mc->mc_ki[mc->mc_top];
8299 if (mc->mc_ki[ptop] == 0) {
8300 /* We're the leftmost leaf in our parent.
8302 DPUTS("reading right neighbor");
8304 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8305 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8308 mn.mc_ki[mn.mc_top] = 0;
8309 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8312 /* There is at least one neighbor to the left.
8314 DPUTS("reading left neighbor");
8316 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8317 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8320 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8321 mc->mc_ki[mc->mc_top] = 0;
8325 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8326 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8327 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8329 /* If the neighbor page is above threshold and has enough keys,
8330 * move one key from it. Otherwise we should try to merge them.
8331 * (A branch page must never have less than 2 keys.)
8333 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8334 rc = mdb_node_move(&mn, mc, fromleft);
8336 /* if we inserted on left, bump position up */
8341 rc = mdb_page_merge(&mn, mc);
8343 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8344 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8345 /* We want mdb_rebalance to find mn when doing fixups */
8346 WITH_CURSOR_TRACKING(mn,
8347 rc = mdb_page_merge(mc, &mn));
8348 mdb_cursor_copy(&mn, mc);
8350 mc->mc_flags &= ~C_EOF;
8352 mc->mc_ki[mc->mc_top] = oldki;
8356 /** Complete a delete operation started by #mdb_cursor_del(). */
8358 mdb_cursor_del0(MDB_cursor *mc)
8364 MDB_cursor *m2, *m3;
8365 MDB_dbi dbi = mc->mc_dbi;
8367 ki = mc->mc_ki[mc->mc_top];
8368 mp = mc->mc_pg[mc->mc_top];
8369 mdb_node_del(mc, mc->mc_db->md_pad);
8370 mc->mc_db->md_entries--;
8372 /* Adjust other cursors pointing to mp */
8373 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8374 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8375 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8377 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8379 if (m3->mc_pg[mc->mc_top] == mp) {
8380 if (m3->mc_ki[mc->mc_top] == ki) {
8381 m3->mc_flags |= C_DEL;
8382 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8383 /* Sub-cursor referred into dataset which is gone */
8384 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
8387 } else if (m3->mc_ki[mc->mc_top] > ki) {
8388 m3->mc_ki[mc->mc_top]--;
8390 if (XCURSOR_INITED(m3))
8391 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8395 rc = mdb_rebalance(mc);
8397 if (rc == MDB_SUCCESS) {
8398 /* DB is totally empty now, just bail out.
8399 * Other cursors adjustments were already done
8400 * by mdb_rebalance and aren't needed here.
8405 mp = mc->mc_pg[mc->mc_top];
8406 nkeys = NUMKEYS(mp);
8408 /* Adjust other cursors pointing to mp */
8409 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8410 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8411 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8413 if (m3->mc_snum < mc->mc_snum)
8415 if (m3->mc_pg[mc->mc_top] == mp) {
8416 /* if m3 points past last node in page, find next sibling */
8417 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
8418 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8419 rc = mdb_cursor_sibling(m3, 1);
8420 if (rc == MDB_NOTFOUND) {
8421 m3->mc_flags |= C_EOF;
8426 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8427 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
8428 /* If this node is a fake page, it needs to be reinited
8429 * because its data has moved. But just reset mc_pg[0]
8430 * if the xcursor is already live.
8432 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) {
8433 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)
8434 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8436 mdb_xcursor_init1(m3, node);
8442 mc->mc_flags |= C_DEL;
8446 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8451 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8452 MDB_val *key, MDB_val *data)
8454 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8457 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8458 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8460 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8461 /* must ignore any data */
8465 return mdb_del0(txn, dbi, key, data, 0);
8469 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8470 MDB_val *key, MDB_val *data, unsigned flags)
8475 MDB_val rdata, *xdata;
8479 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8481 mdb_cursor_init(&mc, txn, dbi, &mx);
8490 flags |= MDB_NODUPDATA;
8492 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8494 /* let mdb_page_split know about this cursor if needed:
8495 * delete will trigger a rebalance; if it needs to move
8496 * a node from one page to another, it will have to
8497 * update the parent's separator key(s). If the new sepkey
8498 * is larger than the current one, the parent page may
8499 * run out of space, triggering a split. We need this
8500 * cursor to be consistent until the end of the rebalance.
8502 mc.mc_flags |= C_UNTRACK;
8503 mc.mc_next = txn->mt_cursors[dbi];
8504 txn->mt_cursors[dbi] = &mc;
8505 rc = mdb_cursor_del(&mc, flags);
8506 txn->mt_cursors[dbi] = mc.mc_next;
8511 /** Split a page and insert a new node.
8512 * Set #MDB_TXN_ERROR on failure.
8513 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8514 * The cursor will be updated to point to the actual page and index where
8515 * the node got inserted after the split.
8516 * @param[in] newkey The key for the newly inserted node.
8517 * @param[in] newdata The data for the newly inserted node.
8518 * @param[in] newpgno The page number, if the new node is a branch node.
8519 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8520 * @return 0 on success, non-zero on failure.
8523 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8524 unsigned int nflags)
8527 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8530 int i, j, split_indx, nkeys, pmax;
8531 MDB_env *env = mc->mc_txn->mt_env;
8533 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8534 MDB_page *copy = NULL;
8535 MDB_page *mp, *rp, *pp;
8540 mp = mc->mc_pg[mc->mc_top];
8541 newindx = mc->mc_ki[mc->mc_top];
8542 nkeys = NUMKEYS(mp);
8544 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8545 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8546 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8548 /* Create a right sibling. */
8549 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8551 rp->mp_pad = mp->mp_pad;
8552 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8554 /* Usually when splitting the root page, the cursor
8555 * height is 1. But when called from mdb_update_key,
8556 * the cursor height may be greater because it walks
8557 * up the stack while finding the branch slot to update.
8559 if (mc->mc_top < 1) {
8560 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8562 /* shift current top to make room for new parent */
8563 for (i=mc->mc_snum; i>0; i--) {
8564 mc->mc_pg[i] = mc->mc_pg[i-1];
8565 mc->mc_ki[i] = mc->mc_ki[i-1];
8569 mc->mc_db->md_root = pp->mp_pgno;
8570 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8571 new_root = mc->mc_db->md_depth++;
8573 /* Add left (implicit) pointer. */
8574 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8575 /* undo the pre-push */
8576 mc->mc_pg[0] = mc->mc_pg[1];
8577 mc->mc_ki[0] = mc->mc_ki[1];
8578 mc->mc_db->md_root = mp->mp_pgno;
8579 mc->mc_db->md_depth--;
8586 ptop = mc->mc_top-1;
8587 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8590 mdb_cursor_copy(mc, &mn);
8591 mn.mc_xcursor = NULL;
8592 mn.mc_pg[mn.mc_top] = rp;
8593 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8595 if (nflags & MDB_APPEND) {
8596 mn.mc_ki[mn.mc_top] = 0;
8598 split_indx = newindx;
8602 split_indx = (nkeys+1) / 2;
8607 unsigned int lsize, rsize, ksize;
8608 /* Move half of the keys to the right sibling */
8609 x = mc->mc_ki[mc->mc_top] - split_indx;
8610 ksize = mc->mc_db->md_pad;
8611 split = LEAF2KEY(mp, split_indx, ksize);
8612 rsize = (nkeys - split_indx) * ksize;
8613 lsize = (nkeys - split_indx) * sizeof(indx_t);
8614 mp->mp_lower -= lsize;
8615 rp->mp_lower += lsize;
8616 mp->mp_upper += rsize - lsize;
8617 rp->mp_upper -= rsize - lsize;
8618 sepkey.mv_size = ksize;
8619 if (newindx == split_indx) {
8620 sepkey.mv_data = newkey->mv_data;
8622 sepkey.mv_data = split;
8625 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8626 memcpy(rp->mp_ptrs, split, rsize);
8627 sepkey.mv_data = rp->mp_ptrs;
8628 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8629 memcpy(ins, newkey->mv_data, ksize);
8630 mp->mp_lower += sizeof(indx_t);
8631 mp->mp_upper -= ksize - sizeof(indx_t);
8634 memcpy(rp->mp_ptrs, split, x * ksize);
8635 ins = LEAF2KEY(rp, x, ksize);
8636 memcpy(ins, newkey->mv_data, ksize);
8637 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8638 rp->mp_lower += sizeof(indx_t);
8639 rp->mp_upper -= ksize - sizeof(indx_t);
8640 mc->mc_ki[mc->mc_top] = x;
8643 int psize, nsize, k;
8644 /* Maximum free space in an empty page */
8645 pmax = env->me_psize - PAGEHDRSZ;
8647 nsize = mdb_leaf_size(env, newkey, newdata);
8649 nsize = mdb_branch_size(env, newkey);
8650 nsize = EVEN(nsize);
8652 /* grab a page to hold a temporary copy */
8653 copy = mdb_page_malloc(mc->mc_txn, 1);
8658 copy->mp_pgno = mp->mp_pgno;
8659 copy->mp_flags = mp->mp_flags;
8660 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8661 copy->mp_upper = env->me_psize - PAGEBASE;
8663 /* prepare to insert */
8664 for (i=0, j=0; i<nkeys; i++) {
8666 copy->mp_ptrs[j++] = 0;
8668 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8671 /* When items are relatively large the split point needs
8672 * to be checked, because being off-by-one will make the
8673 * difference between success or failure in mdb_node_add.
8675 * It's also relevant if a page happens to be laid out
8676 * such that one half of its nodes are all "small" and
8677 * the other half of its nodes are "large." If the new
8678 * item is also "large" and falls on the half with
8679 * "large" nodes, it also may not fit.
8681 * As a final tweak, if the new item goes on the last
8682 * spot on the page (and thus, onto the new page), bias
8683 * the split so the new page is emptier than the old page.
8684 * This yields better packing during sequential inserts.
8686 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8687 /* Find split point */
8689 if (newindx <= split_indx || newindx >= nkeys) {
8691 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8696 for (; i!=k; i+=j) {
8701 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8702 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8704 if (F_ISSET(node->mn_flags, F_BIGDATA))
8705 psize += sizeof(pgno_t);
8707 psize += NODEDSZ(node);
8709 psize = EVEN(psize);
8711 if (psize > pmax || i == k-j) {
8712 split_indx = i + (j<0);
8717 if (split_indx == newindx) {
8718 sepkey.mv_size = newkey->mv_size;
8719 sepkey.mv_data = newkey->mv_data;
8721 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8722 sepkey.mv_size = node->mn_ksize;
8723 sepkey.mv_data = NODEKEY(node);
8728 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8730 /* Copy separator key to the parent.
8732 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8733 int snum = mc->mc_snum;
8737 /* We want other splits to find mn when doing fixups */
8738 WITH_CURSOR_TRACKING(mn,
8739 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
8744 if (mc->mc_snum > snum) {
8747 /* Right page might now have changed parent.
8748 * Check if left page also changed parent.
8750 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8751 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8752 for (i=0; i<ptop; i++) {
8753 mc->mc_pg[i] = mn.mc_pg[i];
8754 mc->mc_ki[i] = mn.mc_ki[i];
8756 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8757 if (mn.mc_ki[ptop]) {
8758 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8760 /* find right page's left sibling */
8761 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8762 mdb_cursor_sibling(mc, 0);
8767 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8770 if (rc != MDB_SUCCESS) {
8773 if (nflags & MDB_APPEND) {
8774 mc->mc_pg[mc->mc_top] = rp;
8775 mc->mc_ki[mc->mc_top] = 0;
8776 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8779 for (i=0; i<mc->mc_top; i++)
8780 mc->mc_ki[i] = mn.mc_ki[i];
8781 } else if (!IS_LEAF2(mp)) {
8783 mc->mc_pg[mc->mc_top] = rp;
8788 rkey.mv_data = newkey->mv_data;
8789 rkey.mv_size = newkey->mv_size;
8795 /* Update index for the new key. */
8796 mc->mc_ki[mc->mc_top] = j;
8798 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8799 rkey.mv_data = NODEKEY(node);
8800 rkey.mv_size = node->mn_ksize;
8802 xdata.mv_data = NODEDATA(node);
8803 xdata.mv_size = NODEDSZ(node);
8806 pgno = NODEPGNO(node);
8807 flags = node->mn_flags;
8810 if (!IS_LEAF(mp) && j == 0) {
8811 /* First branch index doesn't need key data. */
8815 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8821 mc->mc_pg[mc->mc_top] = copy;
8826 } while (i != split_indx);
8828 nkeys = NUMKEYS(copy);
8829 for (i=0; i<nkeys; i++)
8830 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8831 mp->mp_lower = copy->mp_lower;
8832 mp->mp_upper = copy->mp_upper;
8833 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8834 env->me_psize - copy->mp_upper - PAGEBASE);
8836 /* reset back to original page */
8837 if (newindx < split_indx) {
8838 mc->mc_pg[mc->mc_top] = mp;
8840 mc->mc_pg[mc->mc_top] = rp;
8842 /* Make sure mc_ki is still valid.
8844 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8845 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8846 for (i=0; i<=ptop; i++) {
8847 mc->mc_pg[i] = mn.mc_pg[i];
8848 mc->mc_ki[i] = mn.mc_ki[i];
8852 if (nflags & MDB_RESERVE) {
8853 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8854 if (!(node->mn_flags & F_BIGDATA))
8855 newdata->mv_data = NODEDATA(node);
8858 if (newindx >= split_indx) {
8859 mc->mc_pg[mc->mc_top] = rp;
8861 /* Make sure mc_ki is still valid.
8863 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8864 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8865 for (i=0; i<=ptop; i++) {
8866 mc->mc_pg[i] = mn.mc_pg[i];
8867 mc->mc_ki[i] = mn.mc_ki[i];
8874 /* Adjust other cursors pointing to mp */
8875 MDB_cursor *m2, *m3;
8876 MDB_dbi dbi = mc->mc_dbi;
8877 nkeys = NUMKEYS(mp);
8879 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8880 if (mc->mc_flags & C_SUB)
8881 m3 = &m2->mc_xcursor->mx_cursor;
8886 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8890 /* sub cursors may be on different DB */
8891 if (m3->mc_pg[0] != mp)
8894 for (k=new_root; k>=0; k--) {
8895 m3->mc_ki[k+1] = m3->mc_ki[k];
8896 m3->mc_pg[k+1] = m3->mc_pg[k];
8898 if (m3->mc_ki[0] >= nkeys) {
8903 m3->mc_pg[0] = mc->mc_pg[0];
8907 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8908 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8909 m3->mc_ki[mc->mc_top]++;
8910 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8911 m3->mc_pg[mc->mc_top] = rp;
8912 m3->mc_ki[mc->mc_top] -= nkeys;
8913 for (i=0; i<mc->mc_top; i++) {
8914 m3->mc_ki[i] = mn.mc_ki[i];
8915 m3->mc_pg[i] = mn.mc_pg[i];
8918 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8919 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8922 if (XCURSOR_INITED(m3) && IS_LEAF(mp))
8923 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8926 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8929 if (copy) /* tmp page */
8930 mdb_page_free(env, copy);
8932 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8937 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8938 MDB_val *key, MDB_val *data, unsigned int flags)
8944 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8947 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8950 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8951 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8953 mdb_cursor_init(&mc, txn, dbi, &mx);
8954 mc.mc_next = txn->mt_cursors[dbi];
8955 txn->mt_cursors[dbi] = &mc;
8956 rc = mdb_cursor_put(&mc, key, data, flags);
8957 txn->mt_cursors[dbi] = mc.mc_next;
8962 #define MDB_WBUF (1024*1024)
8964 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
8966 /** State needed for a double-buffering compacting copy. */
8967 typedef struct mdb_copy {
8970 pthread_mutex_t mc_mutex;
8971 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
8976 pgno_t mc_next_pgno;
8978 int mc_toggle; /**< Buffer number in provider */
8979 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
8980 /** Error code. Never cleared if set. Both threads can set nonzero
8981 * to fail the copy. Not mutex-protected, LMDB expects atomic int.
8983 volatile int mc_error;
8986 /** Dedicated writer thread for compacting copy. */
8987 static THREAD_RET ESECT CALL_CONV
8988 mdb_env_copythr(void *arg)
8992 int toggle = 0, wsize, rc;
8995 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8998 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9001 pthread_mutex_lock(&my->mc_mutex);
9004 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9005 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
9007 wsize = my->mc_wlen[toggle];
9008 ptr = my->mc_wbuf[toggle];
9011 while (wsize > 0 && !my->mc_error) {
9012 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9016 } else if (len > 0) {
9029 /* If there's an overflow page tail, write it too */
9030 if (my->mc_olen[toggle]) {
9031 wsize = my->mc_olen[toggle];
9032 ptr = my->mc_over[toggle];
9033 my->mc_olen[toggle] = 0;
9036 my->mc_wlen[toggle] = 0;
9038 /* Return the empty buffer to provider */
9040 pthread_cond_signal(&my->mc_cond);
9042 pthread_mutex_unlock(&my->mc_mutex);
9043 return (THREAD_RET)0;
9047 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9049 * @param[in] my control structure.
9050 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9053 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9055 pthread_mutex_lock(&my->mc_mutex);
9056 my->mc_new += adjust;
9057 pthread_cond_signal(&my->mc_cond);
9058 while (my->mc_new & 2) /* both buffers in use */
9059 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9060 pthread_mutex_unlock(&my->mc_mutex);
9062 my->mc_toggle ^= (adjust & 1);
9063 /* Both threads reset mc_wlen, to be safe from threading errors */
9064 my->mc_wlen[my->mc_toggle] = 0;
9065 return my->mc_error;
9068 /** Depth-first tree traversal for compacting copy.
9069 * @param[in] my control structure.
9070 * @param[in,out] pg database root.
9071 * @param[in] flags includes #F_DUPDATA if it is a sorted-duplicate sub-DB.
9074 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9076 MDB_cursor mc = {0};
9078 MDB_page *mo, *mp, *leaf;
9083 /* Empty DB, nothing to do */
9084 if (*pg == P_INVALID)
9088 mc.mc_txn = my->mc_txn;
9090 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9093 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9097 /* Make cursor pages writable */
9098 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9102 for (i=0; i<mc.mc_top; i++) {
9103 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9104 mc.mc_pg[i] = (MDB_page *)ptr;
9105 ptr += my->mc_env->me_psize;
9108 /* This is writable space for a leaf page. Usually not needed. */
9109 leaf = (MDB_page *)ptr;
9111 toggle = my->mc_toggle;
9112 while (mc.mc_snum > 0) {
9114 mp = mc.mc_pg[mc.mc_top];
9118 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9119 for (i=0; i<n; i++) {
9120 ni = NODEPTR(mp, i);
9121 if (ni->mn_flags & F_BIGDATA) {
9125 /* Need writable leaf */
9127 mc.mc_pg[mc.mc_top] = leaf;
9128 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9130 ni = NODEPTR(mp, i);
9133 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9134 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
9135 rc = mdb_page_get(&mc, pg, &omp, NULL);
9138 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9139 rc = mdb_env_cthr_toggle(my, 1);
9142 toggle = my->mc_toggle;
9144 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9145 memcpy(mo, omp, my->mc_env->me_psize);
9146 mo->mp_pgno = my->mc_next_pgno;
9147 my->mc_next_pgno += omp->mp_pages;
9148 my->mc_wlen[toggle] += my->mc_env->me_psize;
9149 if (omp->mp_pages > 1) {
9150 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9151 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9152 rc = mdb_env_cthr_toggle(my, 1);
9155 toggle = my->mc_toggle;
9157 } else if (ni->mn_flags & F_SUBDATA) {
9160 /* Need writable leaf */
9162 mc.mc_pg[mc.mc_top] = leaf;
9163 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9165 ni = NODEPTR(mp, i);
9168 memcpy(&db, NODEDATA(ni), sizeof(db));
9169 my->mc_toggle = toggle;
9170 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9173 toggle = my->mc_toggle;
9174 memcpy(NODEDATA(ni), &db, sizeof(db));
9179 mc.mc_ki[mc.mc_top]++;
9180 if (mc.mc_ki[mc.mc_top] < n) {
9183 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9185 rc = mdb_page_get(&mc, pg, &mp, NULL);
9190 mc.mc_ki[mc.mc_top] = 0;
9191 if (IS_BRANCH(mp)) {
9192 /* Whenever we advance to a sibling branch page,
9193 * we must proceed all the way down to its first leaf.
9195 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9198 mc.mc_pg[mc.mc_top] = mp;
9202 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9203 rc = mdb_env_cthr_toggle(my, 1);
9206 toggle = my->mc_toggle;
9208 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9209 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9210 mo->mp_pgno = my->mc_next_pgno++;
9211 my->mc_wlen[toggle] += my->mc_env->me_psize;
9213 /* Update parent if there is one */
9214 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9215 SETPGNO(ni, mo->mp_pgno);
9216 mdb_cursor_pop(&mc);
9218 /* Otherwise we're done */
9228 /** Copy environment with compaction. */
9230 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9235 MDB_txn *txn = NULL;
9237 pgno_t root, new_root;
9238 int rc = MDB_SUCCESS;
9241 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
9242 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
9246 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9247 if (my.mc_wbuf[0] == NULL) {
9248 /* _aligned_malloc() sets errno, but we use Windows error codes */
9249 rc = ERROR_NOT_ENOUGH_MEMORY;
9253 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
9255 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
9257 #ifdef HAVE_MEMALIGN
9258 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9259 if (my.mc_wbuf[0] == NULL) {
9266 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
9272 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9273 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9274 my.mc_next_pgno = NUM_METAS;
9277 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
9281 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9285 mp = (MDB_page *)my.mc_wbuf[0];
9286 memset(mp, 0, NUM_METAS * env->me_psize);
9288 mp->mp_flags = P_META;
9289 mm = (MDB_meta *)METADATA(mp);
9290 mdb_env_init_meta0(env, mm);
9291 mm->mm_address = env->me_metas[0]->mm_address;
9293 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9295 mp->mp_flags = P_META;
9296 *(MDB_meta *)METADATA(mp) = *mm;
9297 mm = (MDB_meta *)METADATA(mp);
9299 /* Set metapage 1 with current main DB */
9300 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
9301 if (root != P_INVALID) {
9302 /* Count free pages + freeDB pages. Subtract from last_pg
9303 * to find the new last_pg, which also becomes the new root.
9305 MDB_ID freecount = 0;
9308 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9309 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9310 freecount += *(MDB_ID *)data.mv_data;
9311 if (rc != MDB_NOTFOUND)
9313 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9314 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9315 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9317 new_root = txn->mt_next_pgno - 1 - freecount;
9318 mm->mm_last_pg = new_root;
9319 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9320 mm->mm_dbs[MAIN_DBI].md_root = new_root;
9322 /* When the DB is empty, handle it specially to
9323 * fix any breakage like page leaks from ITS#8174.
9325 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
9327 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
9328 mm->mm_txnid = 1; /* use metapage 1 */
9331 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9333 rc = mdb_env_cwalk(&my, &root, 0);
9334 if (rc == MDB_SUCCESS && root != new_root) {
9335 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
9341 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
9342 rc = THREAD_FINISH(thr);
9347 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
9348 if (my.mc_cond) CloseHandle(my.mc_cond);
9349 if (my.mc_mutex) CloseHandle(my.mc_mutex);
9351 free(my.mc_wbuf[0]);
9352 pthread_cond_destroy(&my.mc_cond);
9354 pthread_mutex_destroy(&my.mc_mutex);
9356 return rc ? rc : my.mc_error;
9359 /** Copy environment as-is. */
9361 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9363 MDB_txn *txn = NULL;
9364 mdb_mutexref_t wmutex = NULL;
9370 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9374 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9377 /* Do the lock/unlock of the reader mutex before starting the
9378 * write txn. Otherwise other read txns could block writers.
9380 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9385 /* We must start the actual read txn after blocking writers */
9386 mdb_txn_end(txn, MDB_END_RESET_TMP);
9388 /* Temporarily block writers until we snapshot the meta pages */
9389 wmutex = env->me_wmutex;
9390 if (LOCK_MUTEX(rc, env, wmutex))
9393 rc = mdb_txn_renew0(txn);
9395 UNLOCK_MUTEX(wmutex);
9400 wsize = env->me_psize * NUM_METAS;
9404 DO_WRITE(rc, fd, ptr, w2, len);
9408 } else if (len > 0) {
9414 /* Non-blocking or async handles are not supported */
9420 UNLOCK_MUTEX(wmutex);
9425 w3 = txn->mt_next_pgno * env->me_psize;
9428 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9435 if (wsize > MAX_WRITE)
9439 DO_WRITE(rc, fd, ptr, w2, len);
9443 } else if (len > 0) {
9460 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9462 if (flags & MDB_CP_COMPACT)
9463 return mdb_env_copyfd1(env, fd);
9465 return mdb_env_copyfd0(env, fd);
9469 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9471 return mdb_env_copyfd2(env, fd, 0);
9475 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9479 HANDLE newfd = INVALID_HANDLE_VALUE;
9481 rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname);
9482 if (rc == MDB_SUCCESS) {
9483 rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd);
9484 mdb_fname_destroy(fname);
9486 if (rc == MDB_SUCCESS) {
9487 rc = mdb_env_copyfd2(env, newfd, flags);
9488 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9495 mdb_env_copy(MDB_env *env, const char *path)
9497 return mdb_env_copy2(env, path, 0);
9501 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9503 if (flag & ~CHANGEABLE)
9506 env->me_flags |= flag;
9508 env->me_flags &= ~flag;
9513 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9518 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9523 mdb_env_set_userctx(MDB_env *env, void *ctx)
9527 env->me_userctx = ctx;
9532 mdb_env_get_userctx(MDB_env *env)
9534 return env ? env->me_userctx : NULL;
9538 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9543 env->me_assert_func = func;
9549 mdb_env_get_path(MDB_env *env, const char **arg)
9554 *arg = env->me_path;
9559 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9568 /** Common code for #mdb_stat() and #mdb_env_stat().
9569 * @param[in] env the environment to operate in.
9570 * @param[in] db the #MDB_db record containing the stats to return.
9571 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9572 * @return 0, this function always succeeds.
9575 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9577 arg->ms_psize = env->me_psize;
9578 arg->ms_depth = db->md_depth;
9579 arg->ms_branch_pages = db->md_branch_pages;
9580 arg->ms_leaf_pages = db->md_leaf_pages;
9581 arg->ms_overflow_pages = db->md_overflow_pages;
9582 arg->ms_entries = db->md_entries;
9588 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9592 if (env == NULL || arg == NULL)
9595 meta = mdb_env_pick_meta(env);
9597 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9601 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9605 if (env == NULL || arg == NULL)
9608 meta = mdb_env_pick_meta(env);
9609 arg->me_mapaddr = meta->mm_address;
9610 arg->me_last_pgno = meta->mm_last_pg;
9611 arg->me_last_txnid = meta->mm_txnid;
9613 arg->me_mapsize = env->me_mapsize;
9614 arg->me_maxreaders = env->me_maxreaders;
9615 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9619 /** Set the default comparison functions for a database.
9620 * Called immediately after a database is opened to set the defaults.
9621 * The user can then override them with #mdb_set_compare() or
9622 * #mdb_set_dupsort().
9623 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9624 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9627 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9629 uint16_t f = txn->mt_dbs[dbi].md_flags;
9631 txn->mt_dbxs[dbi].md_cmp =
9632 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9633 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9635 txn->mt_dbxs[dbi].md_dcmp =
9636 !(f & MDB_DUPSORT) ? 0 :
9637 ((f & MDB_INTEGERDUP)
9638 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9639 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9642 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9648 int rc, dbflag, exact;
9649 unsigned int unused = 0, seq;
9653 if (flags & ~VALID_FLAGS)
9655 if (txn->mt_flags & MDB_TXN_BLOCKED)
9661 if (flags & PERSISTENT_FLAGS) {
9662 uint16_t f2 = flags & PERSISTENT_FLAGS;
9663 /* make sure flag changes get committed */
9664 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9665 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9666 txn->mt_flags |= MDB_TXN_DIRTY;
9669 mdb_default_cmp(txn, MAIN_DBI);
9673 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9674 mdb_default_cmp(txn, MAIN_DBI);
9677 /* Is the DB already open? */
9679 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9680 if (!txn->mt_dbxs[i].md_name.mv_size) {
9681 /* Remember this free slot */
9682 if (!unused) unused = i;
9685 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9686 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9692 /* If no free slot and max hit, fail */
9693 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9694 return MDB_DBS_FULL;
9696 /* Cannot mix named databases with some mainDB flags */
9697 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9698 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9700 /* Find the DB info */
9701 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9704 key.mv_data = (void *)name;
9705 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9706 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9707 if (rc == MDB_SUCCESS) {
9708 /* make sure this is actually a DB */
9709 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9710 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9711 return MDB_INCOMPATIBLE;
9712 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
9716 /* Done here so we cannot fail after creating a new DB */
9717 if ((namedup = strdup(name)) == NULL)
9721 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
9722 data.mv_size = sizeof(MDB_db);
9723 data.mv_data = &dummy;
9724 memset(&dummy, 0, sizeof(dummy));
9725 dummy.md_root = P_INVALID;
9726 dummy.md_flags = flags & PERSISTENT_FLAGS;
9727 WITH_CURSOR_TRACKING(mc,
9728 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA));
9735 /* Got info, register DBI in this txn */
9736 unsigned int slot = unused ? unused : txn->mt_numdbs;
9737 txn->mt_dbxs[slot].md_name.mv_data = namedup;
9738 txn->mt_dbxs[slot].md_name.mv_size = len;
9739 txn->mt_dbxs[slot].md_rel = NULL;
9740 txn->mt_dbflags[slot] = dbflag;
9741 /* txn-> and env-> are the same in read txns, use
9742 * tmp variable to avoid undefined assignment
9744 seq = ++txn->mt_env->me_dbiseqs[slot];
9745 txn->mt_dbiseqs[slot] = seq;
9747 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9749 mdb_default_cmp(txn, slot);
9759 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9761 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9764 if (txn->mt_flags & MDB_TXN_BLOCKED)
9767 if (txn->mt_dbflags[dbi] & DB_STALE) {
9770 /* Stale, must read the DB's root. cursor_init does it for us. */
9771 mdb_cursor_init(&mc, txn, dbi, &mx);
9773 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9776 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9779 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9781 ptr = env->me_dbxs[dbi].md_name.mv_data;
9782 /* If there was no name, this was already closed */
9784 env->me_dbxs[dbi].md_name.mv_data = NULL;
9785 env->me_dbxs[dbi].md_name.mv_size = 0;
9786 env->me_dbflags[dbi] = 0;
9787 env->me_dbiseqs[dbi]++;
9792 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9794 /* We could return the flags for the FREE_DBI too but what's the point? */
9795 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9797 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9801 /** Add all the DB's pages to the free list.
9802 * @param[in] mc Cursor on the DB to free.
9803 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9804 * @return 0 on success, non-zero on failure.
9807 mdb_drop0(MDB_cursor *mc, int subs)
9811 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9812 if (rc == MDB_SUCCESS) {
9813 MDB_txn *txn = mc->mc_txn;
9818 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9819 * This also avoids any P_LEAF2 pages, which have no nodes.
9820 * Also if the DB doesn't have sub-DBs and has no overflow
9821 * pages, omit scanning leaves.
9823 if ((mc->mc_flags & C_SUB) ||
9824 (!subs && !mc->mc_db->md_overflow_pages))
9827 mdb_cursor_copy(mc, &mx);
9828 while (mc->mc_snum > 0) {
9829 MDB_page *mp = mc->mc_pg[mc->mc_top];
9830 unsigned n = NUMKEYS(mp);
9832 for (i=0; i<n; i++) {
9833 ni = NODEPTR(mp, i);
9834 if (ni->mn_flags & F_BIGDATA) {
9837 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9838 rc = mdb_page_get(mc, pg, &omp, NULL);
9841 mdb_cassert(mc, IS_OVERFLOW(omp));
9842 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9846 mc->mc_db->md_overflow_pages -= omp->mp_pages;
9847 if (!mc->mc_db->md_overflow_pages && !subs)
9849 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9850 mdb_xcursor_init1(mc, ni);
9851 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9856 if (!subs && !mc->mc_db->md_overflow_pages)
9859 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9861 for (i=0; i<n; i++) {
9863 ni = NODEPTR(mp, i);
9866 mdb_midl_xappend(txn->mt_free_pgs, pg);
9871 mc->mc_ki[mc->mc_top] = i;
9872 rc = mdb_cursor_sibling(mc, 1);
9874 if (rc != MDB_NOTFOUND)
9876 /* no more siblings, go back to beginning
9877 * of previous level.
9882 for (i=1; i<mc->mc_snum; i++) {
9884 mc->mc_pg[i] = mx.mc_pg[i];
9889 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9892 txn->mt_flags |= MDB_TXN_ERROR;
9893 } else if (rc == MDB_NOTFOUND) {
9896 mc->mc_flags &= ~C_INITIALIZED;
9900 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9902 MDB_cursor *mc, *m2;
9905 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9908 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9911 if (TXN_DBI_CHANGED(txn, dbi))
9914 rc = mdb_cursor_open(txn, dbi, &mc);
9918 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9919 /* Invalidate the dropped DB's cursors */
9920 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9921 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9925 /* Can't delete the main DB */
9926 if (del && dbi >= CORE_DBS) {
9927 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9929 txn->mt_dbflags[dbi] = DB_STALE;
9930 mdb_dbi_close(txn->mt_env, dbi);
9932 txn->mt_flags |= MDB_TXN_ERROR;
9935 /* reset the DB record, mark it dirty */
9936 txn->mt_dbflags[dbi] |= DB_DIRTY;
9937 txn->mt_dbs[dbi].md_depth = 0;
9938 txn->mt_dbs[dbi].md_branch_pages = 0;
9939 txn->mt_dbs[dbi].md_leaf_pages = 0;
9940 txn->mt_dbs[dbi].md_overflow_pages = 0;
9941 txn->mt_dbs[dbi].md_entries = 0;
9942 txn->mt_dbs[dbi].md_root = P_INVALID;
9944 txn->mt_flags |= MDB_TXN_DIRTY;
9947 mdb_cursor_close(mc);
9951 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9953 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9956 txn->mt_dbxs[dbi].md_cmp = cmp;
9960 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9962 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9965 txn->mt_dbxs[dbi].md_dcmp = cmp;
9969 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9971 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9974 txn->mt_dbxs[dbi].md_rel = rel;
9978 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9980 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9983 txn->mt_dbxs[dbi].md_relctx = ctx;
9988 mdb_env_get_maxkeysize(MDB_env *env)
9990 return ENV_MAXKEY(env);
9994 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9996 unsigned int i, rdrs;
9999 int rc = 0, first = 1;
10003 if (!env->me_txns) {
10004 return func("(no reader locks)\n", ctx);
10006 rdrs = env->me_txns->mti_numreaders;
10007 mr = env->me_txns->mti_readers;
10008 for (i=0; i<rdrs; i++) {
10009 if (mr[i].mr_pid) {
10010 txnid_t txnid = mr[i].mr_txnid;
10011 sprintf(buf, txnid == (txnid_t)-1 ?
10012 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
10013 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10016 rc = func(" pid thread txnid\n", ctx);
10020 rc = func(buf, ctx);
10026 rc = func("(no active readers)\n", ctx);
10031 /** Insert pid into list if not already present.
10032 * return -1 if already present.
10035 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10037 /* binary search of pid in list */
10039 unsigned cursor = 1;
10041 unsigned n = ids[0];
10044 unsigned pivot = n >> 1;
10045 cursor = base + pivot + 1;
10046 val = pid - ids[cursor];
10051 } else if ( val > 0 ) {
10056 /* found, so it's a duplicate */
10065 for (n = ids[0]; n > cursor; n--)
10072 mdb_reader_check(MDB_env *env, int *dead)
10078 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10081 /** As #mdb_reader_check(). \b rlocked is set if caller locked #me_rmutex. */
10083 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10085 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10086 unsigned int i, j, rdrs;
10088 MDB_PID_T *pids, pid;
10089 int rc = MDB_SUCCESS, count = 0;
10091 rdrs = env->me_txns->mti_numreaders;
10092 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10096 mr = env->me_txns->mti_readers;
10097 for (i=0; i<rdrs; i++) {
10098 pid = mr[i].mr_pid;
10099 if (pid && pid != env->me_pid) {
10100 if (mdb_pid_insert(pids, pid) == 0) {
10101 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10102 /* Stale reader found */
10105 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10106 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10108 rdrs = 0; /* the above checked all readers */
10110 /* Recheck, a new process may have reused pid */
10111 if (mdb_reader_pid(env, Pidcheck, pid))
10115 for (; j<rdrs; j++)
10116 if (mr[j].mr_pid == pid) {
10117 DPRINTF(("clear stale reader pid %u txn %"Z"d",
10118 (unsigned) pid, mr[j].mr_txnid));
10123 UNLOCK_MUTEX(rmutex);
10134 #ifdef MDB_ROBUST_SUPPORTED
10135 /** Handle #LOCK_MUTEX0() failure.
10136 * Try to repair the lock file if the mutex owner died.
10137 * @param[in] env the environment handle
10138 * @param[in] mutex LOCK_MUTEX0() mutex
10139 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10140 * @return 0 on success with the mutex locked, or an error code on failure.
10143 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10148 if (rc == MDB_OWNERDEAD) {
10149 /* We own the mutex. Clean up after dead previous owner. */
10151 rlocked = (mutex == env->me_rmutex);
10153 /* Keep mti_txnid updated, otherwise next writer can
10154 * overwrite data which latest meta page refers to.
10156 meta = mdb_env_pick_meta(env);
10157 env->me_txns->mti_txnid = meta->mm_txnid;
10158 /* env is hosed if the dead thread was ours */
10160 env->me_flags |= MDB_FATAL_ERROR;
10161 env->me_txn = NULL;
10165 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10166 (rc ? "this process' env is hosed" : "recovering")));
10167 rc2 = mdb_reader_check0(env, rlocked, NULL);
10169 rc2 = mdb_mutex_consistent(mutex);
10170 if (rc || (rc = rc2)) {
10171 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10172 UNLOCK_MUTEX(mutex);
10178 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10183 #endif /* MDB_ROBUST_SUPPORTED */
10185 #if defined(_WIN32)
10186 /** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */
10188 utf8_to_utf16(const char *src, MDB_name *dst, int xtra)
10191 wchar_t *result = NULL;
10192 for (;;) { /* malloc result, then fill it in */
10193 need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need);
10200 result = malloc(sizeof(wchar_t) * (need + xtra));
10205 dst->mn_alloced = 1;
10206 dst->mn_len = need - 1;
10207 dst->mn_val = result;
10208 return MDB_SUCCESS;
10211 #endif /* defined(_WIN32) */