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; /**< just for writing 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.
1752 mdb_page_malloc(MDB_txn *txn, unsigned num)
1754 MDB_env *env = txn->mt_env;
1755 MDB_page *ret = env->me_dpages;
1756 size_t psize = env->me_psize, sz = psize, off;
1757 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1758 * For a single page alloc, we init everything after the page header.
1759 * For multi-page, we init the final page; if the caller needed that
1760 * many pages they will be filling in at least up to the last page.
1764 VGMEMP_ALLOC(env, ret, sz);
1765 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1766 env->me_dpages = ret->mp_next;
1769 psize -= off = PAGEHDRSZ;
1774 if ((ret = malloc(sz)) != NULL) {
1775 VGMEMP_ALLOC(env, ret, sz);
1776 if (!(env->me_flags & MDB_NOMEMINIT)) {
1777 memset((char *)ret + off, 0, psize);
1781 txn->mt_flags |= MDB_TXN_ERROR;
1785 /** Free a single page.
1786 * Saves single pages to a list, for future reuse.
1787 * (This is not used for multi-page overflow pages.)
1790 mdb_page_free(MDB_env *env, MDB_page *mp)
1792 mp->mp_next = env->me_dpages;
1793 VGMEMP_FREE(env, mp);
1794 env->me_dpages = mp;
1797 /** Free a dirty page */
1799 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1801 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1802 mdb_page_free(env, dp);
1804 /* large pages just get freed directly */
1805 VGMEMP_FREE(env, dp);
1810 /** Return all dirty pages to dpage list */
1812 mdb_dlist_free(MDB_txn *txn)
1814 MDB_env *env = txn->mt_env;
1815 MDB_ID2L dl = txn->mt_u.dirty_list;
1816 unsigned i, n = dl[0].mid;
1818 for (i = 1; i <= n; i++) {
1819 mdb_dpage_free(env, dl[i].mptr);
1824 /** Loosen or free a single page.
1825 * Saves single pages to a list for future reuse
1826 * in this same txn. It has been pulled from the freeDB
1827 * and already resides on the dirty list, but has been
1828 * deleted. Use these pages first before pulling again
1831 * If the page wasn't dirtied in this txn, just add it
1832 * to this txn's free list.
1835 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1838 pgno_t pgno = mp->mp_pgno;
1839 MDB_txn *txn = mc->mc_txn;
1841 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1842 if (txn->mt_parent) {
1843 MDB_ID2 *dl = txn->mt_u.dirty_list;
1844 /* If txn has a parent, make sure the page is in our
1848 unsigned x = mdb_mid2l_search(dl, pgno);
1849 if (x <= dl[0].mid && dl[x].mid == pgno) {
1850 if (mp != dl[x].mptr) { /* bad cursor? */
1851 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1852 txn->mt_flags |= MDB_TXN_ERROR;
1853 return MDB_CORRUPTED;
1860 /* no parent txn, so it's just ours */
1865 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1867 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1868 txn->mt_loose_pgs = mp;
1869 txn->mt_loose_count++;
1870 mp->mp_flags |= P_LOOSE;
1872 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1880 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1881 * @param[in] mc A cursor handle for the current operation.
1882 * @param[in] pflags Flags of the pages to update:
1883 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1884 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1885 * @return 0 on success, non-zero on failure.
1888 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1890 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1891 MDB_txn *txn = mc->mc_txn;
1892 MDB_cursor *m3, *m0 = mc;
1897 int rc = MDB_SUCCESS, level;
1899 /* Mark pages seen by cursors */
1900 if (mc->mc_flags & C_UNTRACK)
1901 mc = NULL; /* will find mc in mt_cursors */
1902 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1903 for (; mc; mc=mc->mc_next) {
1904 if (!(mc->mc_flags & C_INITIALIZED))
1906 for (m3 = mc;; m3 = &mx->mx_cursor) {
1908 for (j=0; j<m3->mc_snum; j++) {
1910 if ((mp->mp_flags & Mask) == pflags)
1911 mp->mp_flags ^= P_KEEP;
1913 mx = m3->mc_xcursor;
1914 /* Proceed to mx if it is at a sub-database */
1915 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1917 if (! (mp && (mp->mp_flags & P_LEAF)))
1919 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1920 if (!(leaf->mn_flags & F_SUBDATA))
1929 /* Mark dirty root pages */
1930 for (i=0; i<txn->mt_numdbs; i++) {
1931 if (txn->mt_dbflags[i] & DB_DIRTY) {
1932 pgno_t pgno = txn->mt_dbs[i].md_root;
1933 if (pgno == P_INVALID)
1935 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
1937 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1938 dp->mp_flags ^= P_KEEP;
1946 static int mdb_page_flush(MDB_txn *txn, int keep);
1948 /** Spill pages from the dirty list back to disk.
1949 * This is intended to prevent running into #MDB_TXN_FULL situations,
1950 * but note that they may still occur in a few cases:
1951 * 1) our estimate of the txn size could be too small. Currently this
1952 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1953 * 2) child txns may run out of space if their parents dirtied a
1954 * lot of pages and never spilled them. TODO: we probably should do
1955 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1956 * the parent's dirty_room is below a given threshold.
1958 * Otherwise, if not using nested txns, it is expected that apps will
1959 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1960 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1961 * If the txn never references them again, they can be left alone.
1962 * If the txn only reads them, they can be used without any fuss.
1963 * If the txn writes them again, they can be dirtied immediately without
1964 * going thru all of the work of #mdb_page_touch(). Such references are
1965 * handled by #mdb_page_unspill().
1967 * Also note, we never spill DB root pages, nor pages of active cursors,
1968 * because we'll need these back again soon anyway. And in nested txns,
1969 * we can't spill a page in a child txn if it was already spilled in a
1970 * parent txn. That would alter the parent txns' data even though
1971 * the child hasn't committed yet, and we'd have no way to undo it if
1972 * the child aborted.
1974 * @param[in] m0 cursor A cursor handle identifying the transaction and
1975 * database for which we are checking space.
1976 * @param[in] key For a put operation, the key being stored.
1977 * @param[in] data For a put operation, the data being stored.
1978 * @return 0 on success, non-zero on failure.
1981 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1983 MDB_txn *txn = m0->mc_txn;
1985 MDB_ID2L dl = txn->mt_u.dirty_list;
1986 unsigned int i, j, need;
1989 if (m0->mc_flags & C_SUB)
1992 /* Estimate how much space this op will take */
1993 i = m0->mc_db->md_depth;
1994 /* Named DBs also dirty the main DB */
1995 if (m0->mc_dbi >= CORE_DBS)
1996 i += txn->mt_dbs[MAIN_DBI].md_depth;
1997 /* For puts, roughly factor in the key+data size */
1999 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2000 i += i; /* double it for good measure */
2003 if (txn->mt_dirty_room > i)
2006 if (!txn->mt_spill_pgs) {
2007 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2008 if (!txn->mt_spill_pgs)
2011 /* purge deleted slots */
2012 MDB_IDL sl = txn->mt_spill_pgs;
2013 unsigned int num = sl[0];
2015 for (i=1; i<=num; i++) {
2022 /* Preserve pages which may soon be dirtied again */
2023 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2026 /* Less aggressive spill - we originally spilled the entire dirty list,
2027 * with a few exceptions for cursor pages and DB root pages. But this
2028 * turns out to be a lot of wasted effort because in a large txn many
2029 * of those pages will need to be used again. So now we spill only 1/8th
2030 * of the dirty pages. Testing revealed this to be a good tradeoff,
2031 * better than 1/2, 1/4, or 1/10.
2033 if (need < MDB_IDL_UM_MAX / 8)
2034 need = MDB_IDL_UM_MAX / 8;
2036 /* Save the page IDs of all the pages we're flushing */
2037 /* flush from the tail forward, this saves a lot of shifting later on. */
2038 for (i=dl[0].mid; i && need; i--) {
2039 MDB_ID pn = dl[i].mid << 1;
2041 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2043 /* Can't spill twice, make sure it's not already in a parent's
2046 if (txn->mt_parent) {
2048 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2049 if (tx2->mt_spill_pgs) {
2050 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2051 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2052 dp->mp_flags |= P_KEEP;
2060 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2064 mdb_midl_sort(txn->mt_spill_pgs);
2066 /* Flush the spilled part of dirty list */
2067 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2070 /* Reset any dirty pages we kept that page_flush didn't see */
2071 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2074 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2078 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2080 mdb_find_oldest(MDB_txn *txn)
2083 txnid_t mr, oldest = txn->mt_txnid - 1;
2084 if (txn->mt_env->me_txns) {
2085 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2086 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2097 /** Add a page to the txn's dirty list */
2099 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2102 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2104 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2105 insert = mdb_mid2l_append;
2107 insert = mdb_mid2l_insert;
2109 mid.mid = mp->mp_pgno;
2111 rc = insert(txn->mt_u.dirty_list, &mid);
2112 mdb_tassert(txn, rc == 0);
2113 txn->mt_dirty_room--;
2116 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2117 * me_pghead and mt_next_pgno.
2119 * If there are free pages available from older transactions, they
2120 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2121 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2122 * and move me_pglast to say which records were consumed. Only this
2123 * function can create me_pghead and move me_pglast/mt_next_pgno.
2124 * @param[in] mc cursor A cursor handle identifying the transaction and
2125 * database for which we are allocating.
2126 * @param[in] num the number of pages to allocate.
2127 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2128 * will always be satisfied by a single contiguous chunk of memory.
2129 * @return 0 on success, non-zero on failure.
2132 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2134 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2135 /* Get at most <Max_retries> more freeDB records once me_pghead
2136 * has enough pages. If not enough, use new pages from the map.
2137 * If <Paranoid> and mc is updating the freeDB, only get new
2138 * records if me_pghead is empty. Then the freelist cannot play
2139 * catch-up with itself by growing while trying to save it.
2141 enum { Paranoid = 1, Max_retries = 500 };
2143 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2145 int rc, retry = num * 60;
2146 MDB_txn *txn = mc->mc_txn;
2147 MDB_env *env = txn->mt_env;
2148 pgno_t pgno, *mop = env->me_pghead;
2149 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2151 txnid_t oldest = 0, last;
2156 /* If there are any loose pages, just use them */
2157 if (num == 1 && txn->mt_loose_pgs) {
2158 np = txn->mt_loose_pgs;
2159 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2160 txn->mt_loose_count--;
2161 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2169 /* If our dirty list is already full, we can't do anything */
2170 if (txn->mt_dirty_room == 0) {
2175 for (op = MDB_FIRST;; op = MDB_NEXT) {
2180 /* Seek a big enough contiguous page range. Prefer
2181 * pages at the tail, just truncating the list.
2187 if (mop[i-n2] == pgno+n2)
2194 if (op == MDB_FIRST) { /* 1st iteration */
2195 /* Prepare to fetch more and coalesce */
2196 last = env->me_pglast;
2197 oldest = env->me_pgoldest;
2198 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2201 key.mv_data = &last; /* will look up last+1 */
2202 key.mv_size = sizeof(last);
2204 if (Paranoid && mc->mc_dbi == FREE_DBI)
2207 if (Paranoid && retry < 0 && mop_len)
2211 /* Do not fetch more if the record will be too recent */
2212 if (oldest <= last) {
2214 oldest = mdb_find_oldest(txn);
2215 env->me_pgoldest = oldest;
2221 rc = mdb_cursor_get(&m2, &key, NULL, op);
2223 if (rc == MDB_NOTFOUND)
2227 last = *(txnid_t*)key.mv_data;
2228 if (oldest <= last) {
2230 oldest = mdb_find_oldest(txn);
2231 env->me_pgoldest = oldest;
2237 np = m2.mc_pg[m2.mc_top];
2238 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2239 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2242 idl = (MDB_ID *) data.mv_data;
2245 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2250 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2252 mop = env->me_pghead;
2254 env->me_pglast = last;
2256 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2257 last, txn->mt_dbs[FREE_DBI].md_root, i));
2259 DPRINTF(("IDL %"Z"u", idl[j]));
2261 /* Merge in descending sorted order */
2262 mdb_midl_xmerge(mop, idl);
2266 /* Use new pages from the map when nothing suitable in the freeDB */
2268 pgno = txn->mt_next_pgno;
2269 if (pgno + num >= env->me_maxpg) {
2270 DPUTS("DB size maxed out");
2276 if (env->me_flags & MDB_WRITEMAP) {
2277 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2279 if (!(np = mdb_page_malloc(txn, num))) {
2285 mop[0] = mop_len -= num;
2286 /* Move any stragglers down */
2287 for (j = i-num; j < mop_len; )
2288 mop[++j] = mop[++i];
2290 txn->mt_next_pgno = pgno + num;
2293 mdb_page_dirty(txn, np);
2299 txn->mt_flags |= MDB_TXN_ERROR;
2303 /** Copy the used portions of a non-overflow page.
2304 * @param[in] dst page to copy into
2305 * @param[in] src page to copy from
2306 * @param[in] psize size of a page
2309 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2311 enum { Align = sizeof(pgno_t) };
2312 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2314 /* If page isn't full, just copy the used portion. Adjust
2315 * alignment so memcpy may copy words instead of bytes.
2317 if ((unused &= -Align) && !IS_LEAF2(src)) {
2318 upper = (upper + PAGEBASE) & -Align;
2319 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2320 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2323 memcpy(dst, src, psize - unused);
2327 /** Pull a page off the txn's spill list, if present.
2328 * If a page being referenced was spilled to disk in this txn, bring
2329 * it back and make it dirty/writable again.
2330 * @param[in] txn the transaction handle.
2331 * @param[in] mp the page being referenced. It must not be dirty.
2332 * @param[out] ret the writable page, if any. ret is unchanged if
2333 * mp wasn't spilled.
2336 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2338 MDB_env *env = txn->mt_env;
2341 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2343 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2344 if (!tx2->mt_spill_pgs)
2346 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2347 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2350 if (txn->mt_dirty_room == 0)
2351 return MDB_TXN_FULL;
2352 if (IS_OVERFLOW(mp))
2356 if (env->me_flags & MDB_WRITEMAP) {
2359 np = mdb_page_malloc(txn, num);
2363 memcpy(np, mp, num * env->me_psize);
2365 mdb_page_copy(np, mp, env->me_psize);
2368 /* If in current txn, this page is no longer spilled.
2369 * If it happens to be the last page, truncate the spill list.
2370 * Otherwise mark it as deleted by setting the LSB.
2372 if (x == txn->mt_spill_pgs[0])
2373 txn->mt_spill_pgs[0]--;
2375 txn->mt_spill_pgs[x] |= 1;
2376 } /* otherwise, if belonging to a parent txn, the
2377 * page remains spilled until child commits
2380 mdb_page_dirty(txn, np);
2381 np->mp_flags |= P_DIRTY;
2389 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2390 * @param[in] mc cursor pointing to the page to be touched
2391 * @return 0 on success, non-zero on failure.
2394 mdb_page_touch(MDB_cursor *mc)
2396 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2397 MDB_txn *txn = mc->mc_txn;
2398 MDB_cursor *m2, *m3;
2402 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2403 if (txn->mt_flags & MDB_TXN_SPILLS) {
2405 rc = mdb_page_unspill(txn, mp, &np);
2411 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2412 (rc = mdb_page_alloc(mc, 1, &np)))
2415 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2416 mp->mp_pgno, pgno));
2417 mdb_cassert(mc, mp->mp_pgno != pgno);
2418 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2419 /* Update the parent page, if any, to point to the new page */
2421 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2422 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2423 SETPGNO(node, pgno);
2425 mc->mc_db->md_root = pgno;
2427 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2428 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2430 /* If txn has a parent, make sure the page is in our
2434 unsigned x = mdb_mid2l_search(dl, pgno);
2435 if (x <= dl[0].mid && dl[x].mid == pgno) {
2436 if (mp != dl[x].mptr) { /* bad cursor? */
2437 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2438 txn->mt_flags |= MDB_TXN_ERROR;
2439 return MDB_CORRUPTED;
2444 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2446 np = mdb_page_malloc(txn, 1);
2451 rc = mdb_mid2l_insert(dl, &mid);
2452 mdb_cassert(mc, rc == 0);
2457 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2459 np->mp_flags |= P_DIRTY;
2462 /* Adjust cursors pointing to mp */
2463 mc->mc_pg[mc->mc_top] = np;
2464 m2 = txn->mt_cursors[mc->mc_dbi];
2465 if (mc->mc_flags & C_SUB) {
2466 for (; m2; m2=m2->mc_next) {
2467 m3 = &m2->mc_xcursor->mx_cursor;
2468 if (m3->mc_snum < mc->mc_snum) continue;
2469 if (m3->mc_pg[mc->mc_top] == mp)
2470 m3->mc_pg[mc->mc_top] = np;
2473 for (; m2; m2=m2->mc_next) {
2474 if (m2->mc_snum < mc->mc_snum) continue;
2475 if (m2 == mc) continue;
2476 if (m2->mc_pg[mc->mc_top] == mp) {
2477 m2->mc_pg[mc->mc_top] = np;
2478 if (XCURSOR_INITED(m2) && IS_LEAF(np))
2479 XCURSOR_REFRESH(m2, np, m2->mc_ki[mc->mc_top]);
2486 txn->mt_flags |= MDB_TXN_ERROR;
2491 mdb_env_sync(MDB_env *env, int force)
2494 if (env->me_flags & MDB_RDONLY)
2496 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2497 if (env->me_flags & MDB_WRITEMAP) {
2498 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2499 ? MS_ASYNC : MS_SYNC;
2500 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2503 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2507 #ifdef BROKEN_FDATASYNC
2508 if (env->me_flags & MDB_FSYNCONLY) {
2509 if (fsync(env->me_fd))
2513 if (MDB_FDATASYNC(env->me_fd))
2520 /** Back up parent txn's cursors, then grab the originals for tracking */
2522 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2524 MDB_cursor *mc, *bk;
2529 for (i = src->mt_numdbs; --i >= 0; ) {
2530 if ((mc = src->mt_cursors[i]) != NULL) {
2531 size = sizeof(MDB_cursor);
2533 size += sizeof(MDB_xcursor);
2534 for (; mc; mc = bk->mc_next) {
2540 mc->mc_db = &dst->mt_dbs[i];
2541 /* Kill pointers into src to reduce abuse: The
2542 * user may not use mc until dst ends. But we need a valid
2543 * txn pointer here for cursor fixups to keep working.
2546 mc->mc_dbflag = &dst->mt_dbflags[i];
2547 if ((mx = mc->mc_xcursor) != NULL) {
2548 *(MDB_xcursor *)(bk+1) = *mx;
2549 mx->mx_cursor.mc_txn = dst;
2551 mc->mc_next = dst->mt_cursors[i];
2552 dst->mt_cursors[i] = mc;
2559 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2560 * @param[in] txn the transaction handle.
2561 * @param[in] merge true to keep changes to parent cursors, false to revert.
2562 * @return 0 on success, non-zero on failure.
2565 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2567 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2571 for (i = txn->mt_numdbs; --i >= 0; ) {
2572 for (mc = cursors[i]; mc; mc = next) {
2574 if ((bk = mc->mc_backup) != NULL) {
2576 /* Commit changes to parent txn */
2577 mc->mc_next = bk->mc_next;
2578 mc->mc_backup = bk->mc_backup;
2579 mc->mc_txn = bk->mc_txn;
2580 mc->mc_db = bk->mc_db;
2581 mc->mc_dbflag = bk->mc_dbflag;
2582 if ((mx = mc->mc_xcursor) != NULL)
2583 mx->mx_cursor.mc_txn = bk->mc_txn;
2585 /* Abort nested txn */
2587 if ((mx = mc->mc_xcursor) != NULL)
2588 *mx = *(MDB_xcursor *)(bk+1);
2592 /* Only malloced cursors are permanently tracked. */
2599 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2605 Pidset = F_SETLK, Pidcheck = F_GETLK
2609 /** Set or check a pid lock. Set returns 0 on success.
2610 * Check returns 0 if the process is certainly dead, nonzero if it may
2611 * be alive (the lock exists or an error happened so we do not know).
2613 * On Windows Pidset is a no-op, we merely check for the existence
2614 * of the process with the given pid. On POSIX we use a single byte
2615 * lock on the lockfile, set at an offset equal to the pid.
2618 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2620 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2623 if (op == Pidcheck) {
2624 h = OpenProcess(env->me_pidquery, FALSE, pid);
2625 /* No documented "no such process" code, but other program use this: */
2627 return ErrCode() != ERROR_INVALID_PARAMETER;
2628 /* A process exists until all handles to it close. Has it exited? */
2629 ret = WaitForSingleObject(h, 0) != 0;
2636 struct flock lock_info;
2637 memset(&lock_info, 0, sizeof(lock_info));
2638 lock_info.l_type = F_WRLCK;
2639 lock_info.l_whence = SEEK_SET;
2640 lock_info.l_start = pid;
2641 lock_info.l_len = 1;
2642 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2643 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2645 } else if ((rc = ErrCode()) == EINTR) {
2653 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2654 * @param[in] txn the transaction handle to initialize
2655 * @return 0 on success, non-zero on failure.
2658 mdb_txn_renew0(MDB_txn *txn)
2660 MDB_env *env = txn->mt_env;
2661 MDB_txninfo *ti = env->me_txns;
2663 unsigned int i, nr, flags = txn->mt_flags;
2665 int rc, new_notls = 0;
2667 if ((flags &= MDB_TXN_RDONLY) != 0) {
2669 meta = mdb_env_pick_meta(env);
2670 txn->mt_txnid = meta->mm_txnid;
2671 txn->mt_u.reader = NULL;
2673 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2674 pthread_getspecific(env->me_txkey);
2676 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2677 return MDB_BAD_RSLOT;
2679 MDB_PID_T pid = env->me_pid;
2680 MDB_THR_T tid = pthread_self();
2681 mdb_mutexref_t rmutex = env->me_rmutex;
2683 if (!env->me_live_reader) {
2684 rc = mdb_reader_pid(env, Pidset, pid);
2687 env->me_live_reader = 1;
2690 if (LOCK_MUTEX(rc, env, rmutex))
2692 nr = ti->mti_numreaders;
2693 for (i=0; i<nr; i++)
2694 if (ti->mti_readers[i].mr_pid == 0)
2696 if (i == env->me_maxreaders) {
2697 UNLOCK_MUTEX(rmutex);
2698 return MDB_READERS_FULL;
2700 r = &ti->mti_readers[i];
2701 /* Claim the reader slot, carefully since other code
2702 * uses the reader table un-mutexed: First reset the
2703 * slot, next publish it in mti_numreaders. After
2704 * that, it is safe for mdb_env_close() to touch it.
2705 * When it will be closed, we can finally claim it.
2708 r->mr_txnid = (txnid_t)-1;
2711 ti->mti_numreaders = ++nr;
2712 env->me_close_readers = nr;
2714 UNLOCK_MUTEX(rmutex);
2716 new_notls = (env->me_flags & MDB_NOTLS);
2717 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2722 do /* LY: Retry on a race, ITS#7970. */
2723 r->mr_txnid = ti->mti_txnid;
2724 while(r->mr_txnid != ti->mti_txnid);
2725 txn->mt_txnid = r->mr_txnid;
2726 txn->mt_u.reader = r;
2727 meta = env->me_metas[txn->mt_txnid & 1];
2731 /* Not yet touching txn == env->me_txn0, it may be active */
2733 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2735 txn->mt_txnid = ti->mti_txnid;
2736 meta = env->me_metas[txn->mt_txnid & 1];
2738 meta = mdb_env_pick_meta(env);
2739 txn->mt_txnid = meta->mm_txnid;
2743 if (txn->mt_txnid == mdb_debug_start)
2746 txn->mt_child = NULL;
2747 txn->mt_loose_pgs = NULL;
2748 txn->mt_loose_count = 0;
2749 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2750 txn->mt_u.dirty_list = env->me_dirty_list;
2751 txn->mt_u.dirty_list[0].mid = 0;
2752 txn->mt_free_pgs = env->me_free_pgs;
2753 txn->mt_free_pgs[0] = 0;
2754 txn->mt_spill_pgs = NULL;
2756 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2759 /* Copy the DB info and flags */
2760 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2762 /* Moved to here to avoid a data race in read TXNs */
2763 txn->mt_next_pgno = meta->mm_last_pg+1;
2765 txn->mt_flags = flags;
2768 txn->mt_numdbs = env->me_numdbs;
2769 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2770 x = env->me_dbflags[i];
2771 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2772 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2774 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2775 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2777 if (env->me_flags & MDB_FATAL_ERROR) {
2778 DPUTS("environment had fatal error, must shutdown!");
2780 } else if (env->me_maxpg < txn->mt_next_pgno) {
2781 rc = MDB_MAP_RESIZED;
2785 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2790 mdb_txn_renew(MDB_txn *txn)
2794 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2797 rc = mdb_txn_renew0(txn);
2798 if (rc == MDB_SUCCESS) {
2799 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2800 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2801 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2807 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2811 int rc, size, tsize;
2813 flags &= MDB_TXN_BEGIN_FLAGS;
2814 flags |= env->me_flags & MDB_WRITEMAP;
2816 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2820 /* Nested transactions: Max 1 child, write txns only, no writemap */
2821 flags |= parent->mt_flags;
2822 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2823 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2825 /* Child txns save MDB_pgstate and use own copy of cursors */
2826 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2827 size += tsize = sizeof(MDB_ntxn);
2828 } else if (flags & MDB_RDONLY) {
2829 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2830 size += tsize = sizeof(MDB_txn);
2832 /* Reuse preallocated write txn. However, do not touch it until
2833 * mdb_txn_renew0() succeeds, since it currently may be active.
2838 if ((txn = calloc(1, size)) == NULL) {
2839 DPRINTF(("calloc: %s", strerror(errno)));
2842 txn->mt_dbxs = env->me_dbxs; /* static */
2843 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2844 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2845 txn->mt_flags = flags;
2850 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2851 txn->mt_dbiseqs = parent->mt_dbiseqs;
2852 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2853 if (!txn->mt_u.dirty_list ||
2854 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2856 free(txn->mt_u.dirty_list);
2860 txn->mt_txnid = parent->mt_txnid;
2861 txn->mt_dirty_room = parent->mt_dirty_room;
2862 txn->mt_u.dirty_list[0].mid = 0;
2863 txn->mt_spill_pgs = NULL;
2864 txn->mt_next_pgno = parent->mt_next_pgno;
2865 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2866 parent->mt_child = txn;
2867 txn->mt_parent = parent;
2868 txn->mt_numdbs = parent->mt_numdbs;
2869 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2870 /* Copy parent's mt_dbflags, but clear DB_NEW */
2871 for (i=0; i<txn->mt_numdbs; i++)
2872 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2874 ntxn = (MDB_ntxn *)txn;
2875 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2876 if (env->me_pghead) {
2877 size = MDB_IDL_SIZEOF(env->me_pghead);
2878 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2880 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2885 rc = mdb_cursor_shadow(parent, txn);
2887 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2888 } else { /* MDB_RDONLY */
2889 txn->mt_dbiseqs = env->me_dbiseqs;
2891 rc = mdb_txn_renew0(txn);
2894 if (txn != env->me_txn0)
2897 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2899 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2900 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2901 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2908 mdb_txn_env(MDB_txn *txn)
2910 if(!txn) return NULL;
2915 mdb_txn_id(MDB_txn *txn)
2918 return txn->mt_txnid;
2921 /** Export or close DBI handles opened in this txn. */
2923 mdb_dbis_update(MDB_txn *txn, int keep)
2926 MDB_dbi n = txn->mt_numdbs;
2927 MDB_env *env = txn->mt_env;
2928 unsigned char *tdbflags = txn->mt_dbflags;
2930 for (i = n; --i >= CORE_DBS;) {
2931 if (tdbflags[i] & DB_NEW) {
2933 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2935 char *ptr = env->me_dbxs[i].md_name.mv_data;
2937 env->me_dbxs[i].md_name.mv_data = NULL;
2938 env->me_dbxs[i].md_name.mv_size = 0;
2939 env->me_dbflags[i] = 0;
2940 env->me_dbiseqs[i]++;
2946 if (keep && env->me_numdbs < n)
2950 /** End a transaction, except successful commit of a nested transaction.
2951 * May be called twice for readonly txns: First reset it, then abort.
2952 * @param[in] txn the transaction handle to end
2953 * @param[in] mode why and how to end the transaction
2956 mdb_txn_end(MDB_txn *txn, unsigned mode)
2958 MDB_env *env = txn->mt_env;
2960 static const char *const names[] = MDB_END_NAMES;
2963 /* Export or close DBI handles opened in this txn */
2964 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2966 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2967 names[mode & MDB_END_OPMASK],
2968 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2969 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2971 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2972 if (txn->mt_u.reader) {
2973 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2974 if (!(env->me_flags & MDB_NOTLS)) {
2975 txn->mt_u.reader = NULL; /* txn does not own reader */
2976 } else if (mode & MDB_END_SLOT) {
2977 txn->mt_u.reader->mr_pid = 0;
2978 txn->mt_u.reader = NULL;
2979 } /* else txn owns the slot until it does MDB_END_SLOT */
2981 txn->mt_numdbs = 0; /* prevent further DBI activity */
2982 txn->mt_flags |= MDB_TXN_FINISHED;
2984 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2985 pgno_t *pghead = env->me_pghead;
2987 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2988 mdb_cursors_close(txn, 0);
2989 if (!(env->me_flags & MDB_WRITEMAP)) {
2990 mdb_dlist_free(txn);
2994 txn->mt_flags = MDB_TXN_FINISHED;
2996 if (!txn->mt_parent) {
2997 mdb_midl_shrink(&txn->mt_free_pgs);
2998 env->me_free_pgs = txn->mt_free_pgs;
3000 env->me_pghead = NULL;
3004 mode = 0; /* txn == env->me_txn0, do not free() it */
3006 /* The writer mutex was locked in mdb_txn_begin. */
3008 UNLOCK_MUTEX(env->me_wmutex);
3010 txn->mt_parent->mt_child = NULL;
3011 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3012 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3013 mdb_midl_free(txn->mt_free_pgs);
3014 mdb_midl_free(txn->mt_spill_pgs);
3015 free(txn->mt_u.dirty_list);
3018 mdb_midl_free(pghead);
3021 if (mode & MDB_END_FREE)
3026 mdb_txn_reset(MDB_txn *txn)
3031 /* This call is only valid for read-only txns */
3032 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3035 mdb_txn_end(txn, MDB_END_RESET);
3039 mdb_txn_abort(MDB_txn *txn)
3045 mdb_txn_abort(txn->mt_child);
3047 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3050 /** Save the freelist as of this transaction to the freeDB.
3051 * This changes the freelist. Keep trying until it stabilizes.
3054 mdb_freelist_save(MDB_txn *txn)
3056 /* env->me_pghead[] can grow and shrink during this call.
3057 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3058 * Page numbers cannot disappear from txn->mt_free_pgs[].
3061 MDB_env *env = txn->mt_env;
3062 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3063 txnid_t pglast = 0, head_id = 0;
3064 pgno_t freecnt = 0, *free_pgs, *mop;
3065 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3067 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3069 if (env->me_pghead) {
3070 /* Make sure first page of freeDB is touched and on freelist */
3071 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3072 if (rc && rc != MDB_NOTFOUND)
3076 if (!env->me_pghead && txn->mt_loose_pgs) {
3077 /* Put loose page numbers in mt_free_pgs, since
3078 * we may be unable to return them to me_pghead.
3080 MDB_page *mp = txn->mt_loose_pgs;
3081 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3083 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3084 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3085 txn->mt_loose_pgs = NULL;
3086 txn->mt_loose_count = 0;
3089 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3090 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3091 ? SSIZE_MAX : maxfree_1pg;
3094 /* Come back here after each Put() in case freelist changed */
3099 /* If using records from freeDB which we have not yet
3100 * deleted, delete them and any we reserved for me_pghead.
3102 while (pglast < env->me_pglast) {
3103 rc = mdb_cursor_first(&mc, &key, NULL);
3106 pglast = head_id = *(txnid_t *)key.mv_data;
3107 total_room = head_room = 0;
3108 mdb_tassert(txn, pglast <= env->me_pglast);
3109 rc = mdb_cursor_del(&mc, 0);
3114 /* Save the IDL of pages freed by this txn, to a single record */
3115 if (freecnt < txn->mt_free_pgs[0]) {
3117 /* Make sure last page of freeDB is touched and on freelist */
3118 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3119 if (rc && rc != MDB_NOTFOUND)
3122 free_pgs = txn->mt_free_pgs;
3123 /* Write to last page of freeDB */
3124 key.mv_size = sizeof(txn->mt_txnid);
3125 key.mv_data = &txn->mt_txnid;
3127 freecnt = free_pgs[0];
3128 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3129 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3132 /* Retry if mt_free_pgs[] grew during the Put() */
3133 free_pgs = txn->mt_free_pgs;
3134 } while (freecnt < free_pgs[0]);
3135 mdb_midl_sort(free_pgs);
3136 memcpy(data.mv_data, free_pgs, data.mv_size);
3139 unsigned int i = free_pgs[0];
3140 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3141 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3143 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3149 mop = env->me_pghead;
3150 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3152 /* Reserve records for me_pghead[]. Split it if multi-page,
3153 * to avoid searching freeDB for a page range. Use keys in
3154 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3156 if (total_room >= mop_len) {
3157 if (total_room == mop_len || --more < 0)
3159 } else if (head_room >= maxfree_1pg && head_id > 1) {
3160 /* Keep current record (overflow page), add a new one */
3164 /* (Re)write {key = head_id, IDL length = head_room} */
3165 total_room -= head_room;
3166 head_room = mop_len - total_room;
3167 if (head_room > maxfree_1pg && head_id > 1) {
3168 /* Overflow multi-page for part of me_pghead */
3169 head_room /= head_id; /* amortize page sizes */
3170 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3171 } else if (head_room < 0) {
3172 /* Rare case, not bothering to delete this record */
3175 key.mv_size = sizeof(head_id);
3176 key.mv_data = &head_id;
3177 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3178 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3181 /* IDL is initially empty, zero out at least the length */
3182 pgs = (pgno_t *)data.mv_data;
3183 j = head_room > clean_limit ? head_room : 0;
3187 total_room += head_room;
3190 /* Return loose page numbers to me_pghead, though usually none are
3191 * left at this point. The pages themselves remain in dirty_list.
3193 if (txn->mt_loose_pgs) {
3194 MDB_page *mp = txn->mt_loose_pgs;
3195 unsigned count = txn->mt_loose_count;
3197 /* Room for loose pages + temp IDL with same */
3198 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3200 mop = env->me_pghead;
3201 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3202 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3203 loose[ ++count ] = mp->mp_pgno;
3205 mdb_midl_sort(loose);
3206 mdb_midl_xmerge(mop, loose);
3207 txn->mt_loose_pgs = NULL;
3208 txn->mt_loose_count = 0;
3212 /* Fill in the reserved me_pghead records */
3218 rc = mdb_cursor_first(&mc, &key, &data);
3219 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3220 txnid_t id = *(txnid_t *)key.mv_data;
3221 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3224 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3226 if (len > mop_len) {
3228 data.mv_size = (len + 1) * sizeof(MDB_ID);
3230 data.mv_data = mop -= len;
3233 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3235 if (rc || !(mop_len -= len))
3242 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3243 * @param[in] txn the transaction that's being committed
3244 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3245 * @return 0 on success, non-zero on failure.
3248 mdb_page_flush(MDB_txn *txn, int keep)
3250 MDB_env *env = txn->mt_env;
3251 MDB_ID2L dl = txn->mt_u.dirty_list;
3252 unsigned psize = env->me_psize, j;
3253 int i, pagecount = dl[0].mid, rc;
3254 size_t size = 0, pos = 0;
3256 MDB_page *dp = NULL;
3260 struct iovec iov[MDB_COMMIT_PAGES];
3261 ssize_t wpos = 0, wsize = 0, wres;
3262 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3268 if (env->me_flags & MDB_WRITEMAP) {
3269 /* Clear dirty flags */
3270 while (++i <= pagecount) {
3272 /* Don't flush this page yet */
3273 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3274 dp->mp_flags &= ~P_KEEP;
3278 dp->mp_flags &= ~P_DIRTY;
3283 /* Write the pages */
3285 if (++i <= pagecount) {
3287 /* Don't flush this page yet */
3288 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3289 dp->mp_flags &= ~P_KEEP;
3294 /* clear dirty flag */
3295 dp->mp_flags &= ~P_DIRTY;
3298 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3303 /* Windows actually supports scatter/gather I/O, but only on
3304 * unbuffered file handles. Since we're relying on the OS page
3305 * cache for all our data, that's self-defeating. So we just
3306 * write pages one at a time. We use the ov structure to set
3307 * the write offset, to at least save the overhead of a Seek
3310 DPRINTF(("committing page %"Z"u", pgno));
3311 memset(&ov, 0, sizeof(ov));
3312 ov.Offset = pos & 0xffffffff;
3313 ov.OffsetHigh = pos >> 16 >> 16;
3314 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3316 DPRINTF(("WriteFile: %d", rc));
3320 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3321 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3324 /* Write previous page(s) */
3325 #ifdef MDB_USE_PWRITEV
3326 wres = pwritev(env->me_fd, iov, n, wpos);
3329 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3332 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3336 DPRINTF(("lseek: %s", strerror(rc)));
3339 wres = writev(env->me_fd, iov, n);
3342 if (wres != wsize) {
3347 DPRINTF(("Write error: %s", strerror(rc)));
3349 rc = EIO; /* TODO: Use which error code? */
3350 DPUTS("short write, filesystem full?");
3361 DPRINTF(("committing page %"Z"u", pgno));
3362 next_pos = pos + size;
3363 iov[n].iov_len = size;
3364 iov[n].iov_base = (char *)dp;
3370 /* MIPS has cache coherency issues, this is a no-op everywhere else
3371 * Note: for any size >= on-chip cache size, entire on-chip cache is
3374 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3376 for (i = keep; ++i <= pagecount; ) {
3378 /* This is a page we skipped above */
3381 dl[j].mid = dp->mp_pgno;
3384 mdb_dpage_free(env, dp);
3389 txn->mt_dirty_room += i - j;
3395 mdb_txn_commit(MDB_txn *txn)
3398 unsigned int i, end_mode;
3404 /* mdb_txn_end() mode for a commit which writes nothing */
3405 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3407 if (txn->mt_child) {
3408 rc = mdb_txn_commit(txn->mt_child);
3415 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3419 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3420 DPUTS("txn has failed/finished, can't commit");
3422 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3427 if (txn->mt_parent) {
3428 MDB_txn *parent = txn->mt_parent;
3432 unsigned x, y, len, ps_len;
3434 /* Append our free list to parent's */
3435 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3438 mdb_midl_free(txn->mt_free_pgs);
3439 /* Failures after this must either undo the changes
3440 * to the parent or set MDB_TXN_ERROR in the parent.
3443 parent->mt_next_pgno = txn->mt_next_pgno;
3444 parent->mt_flags = txn->mt_flags;
3446 /* Merge our cursors into parent's and close them */
3447 mdb_cursors_close(txn, 1);
3449 /* Update parent's DB table. */
3450 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3451 parent->mt_numdbs = txn->mt_numdbs;
3452 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3453 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3454 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3455 /* preserve parent's DB_NEW status */
3456 x = parent->mt_dbflags[i] & DB_NEW;
3457 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3460 dst = parent->mt_u.dirty_list;
3461 src = txn->mt_u.dirty_list;
3462 /* Remove anything in our dirty list from parent's spill list */
3463 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3465 pspill[0] = (pgno_t)-1;
3466 /* Mark our dirty pages as deleted in parent spill list */
3467 for (i=0, len=src[0].mid; ++i <= len; ) {
3468 MDB_ID pn = src[i].mid << 1;
3469 while (pn > pspill[x])
3471 if (pn == pspill[x]) {
3476 /* Squash deleted pagenums if we deleted any */
3477 for (x=y; ++x <= ps_len; )
3478 if (!(pspill[x] & 1))
3479 pspill[++y] = pspill[x];
3483 /* Remove anything in our spill list from parent's dirty list */
3484 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3485 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3486 MDB_ID pn = txn->mt_spill_pgs[i];
3488 continue; /* deleted spillpg */
3490 y = mdb_mid2l_search(dst, pn);
3491 if (y <= dst[0].mid && dst[y].mid == pn) {
3493 while (y < dst[0].mid) {
3502 /* Find len = length of merging our dirty list with parent's */
3504 dst[0].mid = 0; /* simplify loops */
3505 if (parent->mt_parent) {
3506 len = x + src[0].mid;
3507 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3508 for (i = x; y && i; y--) {
3509 pgno_t yp = src[y].mid;
3510 while (yp < dst[i].mid)
3512 if (yp == dst[i].mid) {
3517 } else { /* Simplify the above for single-ancestor case */
3518 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3520 /* Merge our dirty list with parent's */
3522 for (i = len; y; dst[i--] = src[y--]) {
3523 pgno_t yp = src[y].mid;
3524 while (yp < dst[x].mid)
3525 dst[i--] = dst[x--];
3526 if (yp == dst[x].mid)
3527 free(dst[x--].mptr);
3529 mdb_tassert(txn, i == x);
3531 free(txn->mt_u.dirty_list);
3532 parent->mt_dirty_room = txn->mt_dirty_room;
3533 if (txn->mt_spill_pgs) {
3534 if (parent->mt_spill_pgs) {
3535 /* TODO: Prevent failure here, so parent does not fail */
3536 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3538 parent->mt_flags |= MDB_TXN_ERROR;
3539 mdb_midl_free(txn->mt_spill_pgs);
3540 mdb_midl_sort(parent->mt_spill_pgs);
3542 parent->mt_spill_pgs = txn->mt_spill_pgs;
3546 /* Append our loose page list to parent's */
3547 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3549 *lp = txn->mt_loose_pgs;
3550 parent->mt_loose_count += txn->mt_loose_count;
3552 parent->mt_child = NULL;
3553 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3558 if (txn != env->me_txn) {
3559 DPUTS("attempt to commit unknown transaction");
3564 mdb_cursors_close(txn, 0);
3566 if (!txn->mt_u.dirty_list[0].mid &&
3567 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3570 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3571 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3573 /* Update DB root pointers */
3574 if (txn->mt_numdbs > CORE_DBS) {
3578 data.mv_size = sizeof(MDB_db);
3580 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3581 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3582 if (txn->mt_dbflags[i] & DB_DIRTY) {
3583 if (TXN_DBI_CHANGED(txn, i)) {
3587 data.mv_data = &txn->mt_dbs[i];
3588 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3596 rc = mdb_freelist_save(txn);
3600 mdb_midl_free(env->me_pghead);
3601 env->me_pghead = NULL;
3602 mdb_midl_shrink(&txn->mt_free_pgs);
3608 if ((rc = mdb_page_flush(txn, 0)) ||
3609 (rc = mdb_env_sync(env, 0)) ||
3610 (rc = mdb_env_write_meta(txn)))
3612 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3615 mdb_txn_end(txn, end_mode);
3623 /** Read the environment parameters of a DB environment before
3624 * mapping it into memory.
3625 * @param[in] env the environment handle
3626 * @param[out] meta address of where to store the meta information
3627 * @return 0 on success, non-zero on failure.
3630 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3636 enum { Size = sizeof(pbuf) };
3638 /* We don't know the page size yet, so use a minimum value.
3639 * Read both meta pages so we can use the latest one.
3642 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3646 memset(&ov, 0, sizeof(ov));
3648 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3649 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3652 rc = pread(env->me_fd, &pbuf, Size, off);
3655 if (rc == 0 && off == 0)
3657 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3658 DPRINTF(("read: %s", mdb_strerror(rc)));
3662 p = (MDB_page *)&pbuf;
3664 if (!F_ISSET(p->mp_flags, P_META)) {
3665 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3670 if (m->mm_magic != MDB_MAGIC) {
3671 DPUTS("meta has invalid magic");
3675 if (m->mm_version != MDB_DATA_VERSION) {
3676 DPRINTF(("database is version %u, expected version %u",
3677 m->mm_version, MDB_DATA_VERSION));
3678 return MDB_VERSION_MISMATCH;
3681 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3687 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3689 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3691 meta->mm_magic = MDB_MAGIC;
3692 meta->mm_version = MDB_DATA_VERSION;
3693 meta->mm_mapsize = env->me_mapsize;
3694 meta->mm_psize = env->me_psize;
3695 meta->mm_last_pg = NUM_METAS-1;
3696 meta->mm_flags = env->me_flags & 0xffff;
3697 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3698 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3699 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3702 /** Write the environment parameters of a freshly created DB environment.
3703 * @param[in] env the environment handle
3704 * @param[in] meta the #MDB_meta to write
3705 * @return 0 on success, non-zero on failure.
3708 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3716 memset(&ov, 0, sizeof(ov));
3717 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3719 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3722 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3723 len = pwrite(fd, ptr, size, pos); \
3724 if (len == -1 && ErrCode() == EINTR) continue; \
3725 rc = (len >= 0); break; } while(1)
3728 DPUTS("writing new meta page");
3730 psize = env->me_psize;
3732 p = calloc(NUM_METAS, psize);
3737 p->mp_flags = P_META;
3738 *(MDB_meta *)METADATA(p) = *meta;
3740 q = (MDB_page *)((char *)p + psize);
3742 q->mp_flags = P_META;
3743 *(MDB_meta *)METADATA(q) = *meta;
3745 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3748 else if ((unsigned) len == psize * NUM_METAS)
3756 /** Update the environment info to commit a transaction.
3757 * @param[in] txn the transaction that's being committed
3758 * @return 0 on success, non-zero on failure.
3761 mdb_env_write_meta(MDB_txn *txn)
3764 MDB_meta meta, metab, *mp;
3768 int rc, len, toggle;
3777 toggle = txn->mt_txnid & 1;
3778 DPRINTF(("writing meta page %d for root page %"Z"u",
3779 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3782 flags = env->me_flags;
3783 mp = env->me_metas[toggle];
3784 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3785 /* Persist any increases of mapsize config */
3786 if (mapsize < env->me_mapsize)
3787 mapsize = env->me_mapsize;
3789 if (flags & MDB_WRITEMAP) {
3790 mp->mm_mapsize = mapsize;
3791 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3792 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3793 mp->mm_last_pg = txn->mt_next_pgno - 1;
3794 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3795 !(defined(__i386__) || defined(__x86_64__))
3796 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3797 __sync_synchronize();
3799 mp->mm_txnid = txn->mt_txnid;
3800 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3801 unsigned meta_size = env->me_psize;
3802 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3803 ptr = (char *)mp - PAGEHDRSZ;
3804 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3805 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3809 if (MDB_MSYNC(ptr, meta_size, rc)) {
3816 metab.mm_txnid = mp->mm_txnid;
3817 metab.mm_last_pg = mp->mm_last_pg;
3819 meta.mm_mapsize = mapsize;
3820 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3821 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3822 meta.mm_last_pg = txn->mt_next_pgno - 1;
3823 meta.mm_txnid = txn->mt_txnid;
3825 off = offsetof(MDB_meta, mm_mapsize);
3826 ptr = (char *)&meta + off;
3827 len = sizeof(MDB_meta) - off;
3828 off += (char *)mp - env->me_map;
3830 /* Write to the SYNC fd */
3831 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3834 memset(&ov, 0, sizeof(ov));
3836 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3841 rc = pwrite(mfd, ptr, len, off);
3844 rc = rc < 0 ? ErrCode() : EIO;
3849 DPUTS("write failed, disk error?");
3850 /* On a failure, the pagecache still contains the new data.
3851 * Write some old data back, to prevent it from being used.
3852 * Use the non-SYNC fd; we know it will fail anyway.
3854 meta.mm_last_pg = metab.mm_last_pg;
3855 meta.mm_txnid = metab.mm_txnid;
3857 memset(&ov, 0, sizeof(ov));
3859 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3861 r2 = pwrite(env->me_fd, ptr, len, off);
3862 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3865 env->me_flags |= MDB_FATAL_ERROR;
3868 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3869 CACHEFLUSH(env->me_map + off, len, DCACHE);
3871 /* Memory ordering issues are irrelevant; since the entire writer
3872 * is wrapped by wmutex, all of these changes will become visible
3873 * after the wmutex is unlocked. Since the DB is multi-version,
3874 * readers will get consistent data regardless of how fresh or
3875 * how stale their view of these values is.
3878 env->me_txns->mti_txnid = txn->mt_txnid;
3883 /** Check both meta pages to see which one is newer.
3884 * @param[in] env the environment handle
3885 * @return newest #MDB_meta.
3888 mdb_env_pick_meta(const MDB_env *env)
3890 MDB_meta *const *metas = env->me_metas;
3891 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3895 mdb_env_create(MDB_env **env)
3899 e = calloc(1, sizeof(MDB_env));
3903 e->me_maxreaders = DEFAULT_READERS;
3904 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3905 e->me_fd = INVALID_HANDLE_VALUE;
3906 e->me_lfd = INVALID_HANDLE_VALUE;
3907 e->me_mfd = INVALID_HANDLE_VALUE;
3908 #ifdef MDB_USE_POSIX_SEM
3909 e->me_rmutex = SEM_FAILED;
3910 e->me_wmutex = SEM_FAILED;
3912 e->me_pid = getpid();
3913 GET_PAGESIZE(e->me_os_psize);
3914 VGMEMP_CREATE(e,0,0);
3920 mdb_env_map(MDB_env *env, void *addr)
3923 unsigned int flags = env->me_flags;
3927 LONG sizelo, sizehi;
3930 if (flags & MDB_RDONLY) {
3931 /* Don't set explicit map size, use whatever exists */
3936 msize = env->me_mapsize;
3937 sizelo = msize & 0xffffffff;
3938 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3940 /* Windows won't create mappings for zero length files.
3941 * and won't map more than the file size.
3942 * Just set the maxsize right now.
3944 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3945 || !SetEndOfFile(env->me_fd)
3946 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3950 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3951 PAGE_READWRITE : PAGE_READONLY,
3952 sizehi, sizelo, NULL);
3955 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3956 FILE_MAP_WRITE : FILE_MAP_READ,
3958 rc = env->me_map ? 0 : ErrCode();
3963 int prot = PROT_READ;
3964 if (flags & MDB_WRITEMAP) {
3966 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3969 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3971 if (env->me_map == MAP_FAILED) {
3976 if (flags & MDB_NORDAHEAD) {
3977 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3979 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3981 #ifdef POSIX_MADV_RANDOM
3982 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3983 #endif /* POSIX_MADV_RANDOM */
3984 #endif /* MADV_RANDOM */
3988 /* Can happen because the address argument to mmap() is just a
3989 * hint. mmap() can pick another, e.g. if the range is in use.
3990 * The MAP_FIXED flag would prevent that, but then mmap could
3991 * instead unmap existing pages to make room for the new map.
3993 if (addr && env->me_map != addr)
3994 return EBUSY; /* TODO: Make a new MDB_* error code? */
3996 p = (MDB_page *)env->me_map;
3997 env->me_metas[0] = METADATA(p);
3998 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4004 mdb_env_set_mapsize(MDB_env *env, size_t size)
4006 /* If env is already open, caller is responsible for making
4007 * sure there are no active txns.
4015 meta = mdb_env_pick_meta(env);
4017 size = meta->mm_mapsize;
4019 /* Silently round up to minimum if the size is too small */
4020 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4024 munmap(env->me_map, env->me_mapsize);
4025 env->me_mapsize = size;
4026 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4027 rc = mdb_env_map(env, old);
4031 env->me_mapsize = size;
4033 env->me_maxpg = env->me_mapsize / env->me_psize;
4038 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4042 env->me_maxdbs = dbs + CORE_DBS;
4047 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4049 if (env->me_map || readers < 1)
4051 env->me_maxreaders = readers;
4056 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4058 if (!env || !readers)
4060 *readers = env->me_maxreaders;
4065 mdb_fsize(HANDLE fd, size_t *size)
4068 LARGE_INTEGER fsize;
4070 if (!GetFileSizeEx(fd, &fsize))
4073 *size = fsize.QuadPart;
4087 typedef wchar_t mdb_nchar_t;
4088 # define MDB_NAME(str) L##str
4089 # define mdb_name_cpy wcscpy
4091 /** Character type for file names: char on Unix, wchar_t on Windows */
4092 typedef char mdb_nchar_t;
4093 # define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */
4094 # define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */
4097 /** Filename - string of #mdb_nchar_t[] */
4098 typedef struct MDB_name {
4099 int mn_len; /**< Length */
4100 int mn_alloced; /**< True if #mn_val was malloced */
4101 mdb_nchar_t *mn_val; /**< Contents */
4104 /** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */
4105 static const mdb_nchar_t *const mdb_suffixes[2][2] = {
4106 { MDB_NAME("/data.mdb"), MDB_NAME("") },
4107 { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") }
4110 #define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */
4112 /** Set up filename + scratch area for filename suffix, for opening files.
4113 * It should be freed with #mdb_fname_destroy().
4114 * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16.
4116 * @param[in] path Pathname for #mdb_env_open().
4117 * @param[in] envflags Whether a subdir and/or lockfile will be used.
4118 * @param[out] fname Resulting filename, with room for a suffix if necessary.
4121 mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname)
4123 int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK);
4124 fname->mn_alloced = 0;
4126 return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN);
4128 fname->mn_len = strlen(path);
4130 fname->mn_val = (char *) path;
4131 else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) {
4132 fname->mn_alloced = 1;
4133 strcpy(fname->mn_val, path);
4141 /** Destroy \b fname from #mdb_fname_init() */
4142 #define mdb_fname_destroy(fname) \
4143 do { if ((fname).mn_alloced) free((fname).mn_val); } while (0)
4145 #ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */
4146 # define MDB_CLOEXEC O_CLOEXEC
4148 # define MDB_CLOEXEC 0
4151 /** File type, access mode etc. for #mdb_fopen() */
4152 enum mdb_fopen_type {
4154 MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS
4156 /* A comment in mdb_fopen() explains some O_* flag choices. */
4157 MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */
4158 MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */
4159 MDB_O_META = O_RDWR |MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */
4160 MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */
4161 /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits
4162 * distinguish otherwise-equal MDB_O_* constants from each other.
4164 MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY,
4165 MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */
4169 /** Open an LMDB file.
4170 * @param[in] env The LMDB environment.
4171 * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is
4172 * appended if necessary to create the filename, without changing mn_len.
4173 * @param[in] which Determines file type, access mode, etc.
4174 * @param[in] mode The Unix permissions for the file, if we create it.
4175 * @param[out] res Resulting file handle.
4176 * @return 0 on success, non-zero on failure.
4179 mdb_fopen(const MDB_env *env, MDB_name *fname,
4180 enum mdb_fopen_type which, mdb_mode_t mode,
4183 int rc = MDB_SUCCESS;
4186 DWORD acc, share, disp, attrs;
4191 if (fname->mn_alloced) /* modifiable copy */
4192 mdb_name_cpy(fname->mn_val + fname->mn_len,
4193 mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]);
4195 /* The directory must already exist. Usually the file need not.
4196 * MDB_O_META requires the file because we already created it using
4197 * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file.
4199 * With MDB_O_COPY we do not want the OS to cache the writes, since
4200 * the source data is already in the OS cache.
4202 * The lockfile needs FD_CLOEXEC (close file descriptor on exec*())
4203 * to avoid the flock() issues noted under Caveats in lmdb.h.
4204 * Also set it for other filehandles which the user cannot get at
4205 * and close himself, which he may need after fork(). I.e. all but
4206 * me_fd, which programs do use via mdb_env_get_fd().
4210 acc = GENERIC_READ|GENERIC_WRITE;
4211 share = FILE_SHARE_READ|FILE_SHARE_WRITE;
4213 attrs = FILE_ATTRIBUTE_NORMAL;
4215 case MDB_O_RDONLY: /* read-only datafile */
4217 disp = OPEN_EXISTING;
4219 case MDB_O_META: /* for writing metapages */
4220 disp = OPEN_EXISTING;
4221 attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH;
4223 case MDB_O_COPY: /* mdb_env_copy() & co */
4224 acc = GENERIC_WRITE;
4227 attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH;
4229 default: break; /* silence gcc -Wswitch (not all enum values handled) */
4231 fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL);
4233 fd = open(fname->mn_val, which & MDB_O_MASK, mode);
4236 if (fd == INVALID_HANDLE_VALUE)
4240 if (which != MDB_O_RDONLY && which != MDB_O_RDWR) {
4241 /* Set CLOEXEC if we could not pass it to open() */
4242 if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1)
4243 (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
4245 if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) {
4246 /* This may require buffer alignment. There is no portable
4247 * way to ask how much, so we require OS pagesize alignment.
4249 # ifdef F_NOCACHE /* __APPLE__ */
4250 (void) fcntl(fd, F_NOCACHE, 1);
4251 # elif defined O_DIRECT
4252 /* open(...O_DIRECT...) would break on filesystems without
4253 * O_DIRECT support (ITS#7682). Try to set it here instead.
4255 if ((flags = fcntl(fd, F_GETFL)) != -1)
4256 (void) fcntl(fd, F_SETFL, flags | O_DIRECT);
4260 #endif /* !_WIN32 */
4267 #ifdef BROKEN_FDATASYNC
4268 #include <sys/utsname.h>
4269 #include <sys/vfs.h>
4272 /** Further setup required for opening an LMDB environment
4275 mdb_env_open2(MDB_env *env)
4277 unsigned int flags = env->me_flags;
4278 int i, newenv = 0, rc;
4282 /* See if we should use QueryLimited */
4284 if ((rc & 0xff) > 5)
4285 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4287 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4290 #ifdef BROKEN_FDATASYNC
4291 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4292 * https://lkml.org/lkml/2012/9/3/83
4293 * Kernels after 3.6-rc6 are known good.
4294 * https://lkml.org/lkml/2012/9/10/556
4295 * See if the DB is on ext3/ext4, then check for new enough kernel
4296 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4301 fstatfs(env->me_fd, &st);
4302 while (st.f_type == 0xEF53) {
4306 if (uts.release[0] < '3') {
4307 if (!strncmp(uts.release, "2.6.32.", 7)) {
4308 i = atoi(uts.release+7);
4310 break; /* 2.6.32.60 and newer is OK */
4311 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4312 i = atoi(uts.release+7);
4314 break; /* 2.6.34.15 and newer is OK */
4316 } else if (uts.release[0] == '3') {
4317 i = atoi(uts.release+2);
4319 break; /* 3.6 and newer is OK */
4321 i = atoi(uts.release+4);
4323 break; /* 3.5.4 and newer is OK */
4324 } else if (i == 2) {
4325 i = atoi(uts.release+4);
4327 break; /* 3.2.30 and newer is OK */
4329 } else { /* 4.x and newer is OK */
4332 env->me_flags |= MDB_FSYNCONLY;
4338 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4341 DPUTS("new mdbenv");
4343 env->me_psize = env->me_os_psize;
4344 if (env->me_psize > MAX_PAGESIZE)
4345 env->me_psize = MAX_PAGESIZE;
4346 memset(&meta, 0, sizeof(meta));
4347 mdb_env_init_meta0(env, &meta);
4348 meta.mm_mapsize = DEFAULT_MAPSIZE;
4350 env->me_psize = meta.mm_psize;
4353 /* Was a mapsize configured? */
4354 if (!env->me_mapsize) {
4355 env->me_mapsize = meta.mm_mapsize;
4358 /* Make sure mapsize >= committed data size. Even when using
4359 * mm_mapsize, which could be broken in old files (ITS#7789).
4361 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4362 if (env->me_mapsize < minsize)
4363 env->me_mapsize = minsize;
4365 meta.mm_mapsize = env->me_mapsize;
4367 if (newenv && !(flags & MDB_FIXEDMAP)) {
4368 /* mdb_env_map() may grow the datafile. Write the metapages
4369 * first, so the file will be valid if initialization fails.
4370 * Except with FIXEDMAP, since we do not yet know mm_address.
4371 * We could fill in mm_address later, but then a different
4372 * program might end up doing that - one with a memory layout
4373 * and map address which does not suit the main program.
4375 rc = mdb_env_init_meta(env, &meta);
4381 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4386 if (flags & MDB_FIXEDMAP)
4387 meta.mm_address = env->me_map;
4388 i = mdb_env_init_meta(env, &meta);
4389 if (i != MDB_SUCCESS) {
4394 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4395 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4397 #if !(MDB_MAXKEYSIZE)
4398 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4400 env->me_maxpg = env->me_mapsize / env->me_psize;
4404 MDB_meta *meta = mdb_env_pick_meta(env);
4405 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4407 DPRINTF(("opened database version %u, pagesize %u",
4408 meta->mm_version, env->me_psize));
4409 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4410 DPRINTF(("depth: %u", db->md_depth));
4411 DPRINTF(("entries: %"Z"u", db->md_entries));
4412 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4413 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4414 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4415 DPRINTF(("root: %"Z"u", db->md_root));
4423 /** Release a reader thread's slot in the reader lock table.
4424 * This function is called automatically when a thread exits.
4425 * @param[in] ptr This points to the slot in the reader lock table.
4428 mdb_env_reader_dest(void *ptr)
4430 MDB_reader *reader = ptr;
4436 /** Junk for arranging thread-specific callbacks on Windows. This is
4437 * necessarily platform and compiler-specific. Windows supports up
4438 * to 1088 keys. Let's assume nobody opens more than 64 environments
4439 * in a single process, for now. They can override this if needed.
4441 #ifndef MAX_TLS_KEYS
4442 #define MAX_TLS_KEYS 64
4444 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4445 static int mdb_tls_nkeys;
4447 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4451 case DLL_PROCESS_ATTACH: break;
4452 case DLL_THREAD_ATTACH: break;
4453 case DLL_THREAD_DETACH:
4454 for (i=0; i<mdb_tls_nkeys; i++) {
4455 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4457 mdb_env_reader_dest(r);
4461 case DLL_PROCESS_DETACH: break;
4466 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4468 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4472 /* Force some symbol references.
4473 * _tls_used forces the linker to create the TLS directory if not already done
4474 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4476 #pragma comment(linker, "/INCLUDE:_tls_used")
4477 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4478 #pragma const_seg(".CRT$XLB")
4479 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4480 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4483 #pragma comment(linker, "/INCLUDE:__tls_used")
4484 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4485 #pragma data_seg(".CRT$XLB")
4486 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4488 #endif /* WIN 32/64 */
4489 #endif /* !__GNUC__ */
4492 /** Downgrade the exclusive lock on the region back to shared */
4494 mdb_env_share_locks(MDB_env *env, int *excl)
4497 MDB_meta *meta = mdb_env_pick_meta(env);
4499 env->me_txns->mti_txnid = meta->mm_txnid;
4504 /* First acquire a shared lock. The Unlock will
4505 * then release the existing exclusive lock.
4507 memset(&ov, 0, sizeof(ov));
4508 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4511 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4517 struct flock lock_info;
4518 /* The shared lock replaces the existing lock */
4519 memset((void *)&lock_info, 0, sizeof(lock_info));
4520 lock_info.l_type = F_RDLCK;
4521 lock_info.l_whence = SEEK_SET;
4522 lock_info.l_start = 0;
4523 lock_info.l_len = 1;
4524 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4525 (rc = ErrCode()) == EINTR) ;
4526 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4533 /** Try to get exclusive lock, otherwise shared.
4534 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4537 mdb_env_excl_lock(MDB_env *env, int *excl)
4541 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4545 memset(&ov, 0, sizeof(ov));
4546 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4553 struct flock lock_info;
4554 memset((void *)&lock_info, 0, sizeof(lock_info));
4555 lock_info.l_type = F_WRLCK;
4556 lock_info.l_whence = SEEK_SET;
4557 lock_info.l_start = 0;
4558 lock_info.l_len = 1;
4559 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4560 (rc = ErrCode()) == EINTR) ;
4564 # ifndef MDB_USE_POSIX_MUTEX
4565 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4568 lock_info.l_type = F_RDLCK;
4569 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4570 (rc = ErrCode()) == EINTR) ;
4580 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4582 * @(#) $Revision: 5.1 $
4583 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4584 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4586 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4590 * Please do not copyright this code. This code is in the public domain.
4592 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4593 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4594 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4595 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4596 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4597 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4598 * PERFORMANCE OF THIS SOFTWARE.
4601 * chongo <Landon Curt Noll> /\oo/\
4602 * http://www.isthe.com/chongo/
4604 * Share and Enjoy! :-)
4607 typedef unsigned long long mdb_hash_t;
4608 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4610 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4611 * @param[in] val value to hash
4612 * @param[in] hval initial value for hash
4613 * @return 64 bit hash
4615 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4616 * hval arg on the first call.
4619 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4621 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4622 unsigned char *end = s + val->mv_size;
4624 * FNV-1a hash each octet of the string
4627 /* xor the bottom with the current octet */
4628 hval ^= (mdb_hash_t)*s++;
4630 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4631 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4632 (hval << 7) + (hval << 8) + (hval << 40);
4634 /* return our new hash value */
4638 /** Hash the string and output the encoded hash.
4639 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4640 * very short name limits. We don't care about the encoding being reversible,
4641 * we just want to preserve as many bits of the input as possible in a
4642 * small printable string.
4643 * @param[in] str string to hash
4644 * @param[out] encbuf an array of 11 chars to hold the hash
4646 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4649 mdb_pack85(unsigned long l, char *out)
4653 for (i=0; i<5; i++) {
4654 *out++ = mdb_a85[l % 85];
4660 mdb_hash_enc(MDB_val *val, char *encbuf)
4662 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4664 mdb_pack85(h, encbuf);
4665 mdb_pack85(h>>32, encbuf+5);
4670 /** Open and/or initialize the lock region for the environment.
4671 * @param[in] env The LMDB environment.
4672 * @param[in] fname Filename + scratch area, from #mdb_fname_init().
4673 * @param[in] mode The Unix permissions for the file, if we create it.
4674 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4675 * @return 0 on success, non-zero on failure.
4678 mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl)
4681 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4683 # define MDB_ERRCODE_ROFS EROFS
4688 rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd);
4690 /* Omit lockfile if read-only env on read-only filesystem */
4691 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4697 if (!(env->me_flags & MDB_NOTLS)) {
4698 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4701 env->me_flags |= MDB_ENV_TXKEY;
4703 /* Windows TLS callbacks need help finding their TLS info. */
4704 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4708 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4712 /* Try to get exclusive lock. If we succeed, then
4713 * nobody is using the lock region and we should initialize it.
4715 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4718 size = GetFileSize(env->me_lfd, NULL);
4720 size = lseek(env->me_lfd, 0, SEEK_END);
4721 if (size == -1) goto fail_errno;
4723 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4724 if (size < rsize && *excl > 0) {
4726 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4727 || !SetEndOfFile(env->me_lfd))
4730 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4734 size = rsize - sizeof(MDB_txninfo);
4735 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4740 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4742 if (!mh) goto fail_errno;
4743 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4745 if (!env->me_txns) goto fail_errno;
4747 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4749 if (m == MAP_FAILED) goto fail_errno;
4755 BY_HANDLE_FILE_INFORMATION stbuf;
4764 if (!mdb_sec_inited) {
4765 InitializeSecurityDescriptor(&mdb_null_sd,
4766 SECURITY_DESCRIPTOR_REVISION);
4767 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4768 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4769 mdb_all_sa.bInheritHandle = FALSE;
4770 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4773 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4774 idbuf.volume = stbuf.dwVolumeSerialNumber;
4775 idbuf.nhigh = stbuf.nFileIndexHigh;
4776 idbuf.nlow = stbuf.nFileIndexLow;
4777 val.mv_data = &idbuf;
4778 val.mv_size = sizeof(idbuf);
4779 mdb_hash_enc(&val, encbuf);
4780 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4781 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4782 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4783 if (!env->me_rmutex) goto fail_errno;
4784 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4785 if (!env->me_wmutex) goto fail_errno;
4786 #elif defined(MDB_USE_POSIX_SEM)
4795 #if defined(__NetBSD__)
4796 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4798 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4799 idbuf.dev = stbuf.st_dev;
4800 idbuf.ino = stbuf.st_ino;
4801 val.mv_data = &idbuf;
4802 val.mv_size = sizeof(idbuf);
4803 mdb_hash_enc(&val, encbuf);
4804 #ifdef MDB_SHORT_SEMNAMES
4805 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4807 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4808 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4809 /* Clean up after a previous run, if needed: Try to
4810 * remove both semaphores before doing anything else.
4812 sem_unlink(env->me_txns->mti_rmname);
4813 sem_unlink(env->me_txns->mti_wmname);
4814 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4815 O_CREAT|O_EXCL, mode, 1);
4816 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4817 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4818 O_CREAT|O_EXCL, mode, 1);
4819 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4820 #else /* MDB_USE_POSIX_MUTEX: */
4821 pthread_mutexattr_t mattr;
4823 /* Solaris needs this before initing a robust mutex. Otherwise
4824 * it may skip the init and return EBUSY "seems someone already
4825 * inited" or EINVAL "it was inited differently".
4827 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
4828 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
4830 if ((rc = pthread_mutexattr_init(&mattr)))
4833 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
4834 #ifdef MDB_ROBUST_SUPPORTED
4835 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
4837 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
4838 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
4839 pthread_mutexattr_destroy(&mattr);
4842 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4844 env->me_txns->mti_magic = MDB_MAGIC;
4845 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4846 env->me_txns->mti_txnid = 0;
4847 env->me_txns->mti_numreaders = 0;
4850 if (env->me_txns->mti_magic != MDB_MAGIC) {
4851 DPUTS("lock region has invalid magic");
4855 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4856 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4857 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4858 rc = MDB_VERSION_MISMATCH;
4862 if (rc && rc != EACCES && rc != EAGAIN) {
4866 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4867 if (!env->me_rmutex) goto fail_errno;
4868 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4869 if (!env->me_wmutex) goto fail_errno;
4870 #elif defined(MDB_USE_POSIX_SEM)
4871 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4872 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4873 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4874 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4885 /** Only a subset of the @ref mdb_env flags can be changed
4886 * at runtime. Changing other flags requires closing the
4887 * environment and re-opening it with the new flags.
4889 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4890 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4891 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4893 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4894 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4898 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4903 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4906 flags |= env->me_flags;
4908 rc = mdb_fname_init(path, flags, &fname);
4912 if (flags & MDB_RDONLY) {
4913 /* silently ignore WRITEMAP when we're only getting read access */
4914 flags &= ~MDB_WRITEMAP;
4916 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4917 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4920 env->me_flags = flags |= MDB_ENV_ACTIVE;
4924 env->me_path = strdup(path);
4925 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4926 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4927 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4928 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4932 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4934 /* For RDONLY, get lockfile after we know datafile exists */
4935 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4936 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
4941 rc = mdb_fopen(env, &fname,
4942 (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR,
4947 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4948 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
4953 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4954 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4955 env->me_mfd = env->me_fd;
4957 /* Synchronous fd for meta writes. Needed even with
4958 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4960 rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd);
4964 DPRINTF(("opened dbenv %p", (void *) env));
4966 rc = mdb_env_share_locks(env, &excl);
4970 if (!(flags & MDB_RDONLY)) {
4972 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4973 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4974 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4975 (txn = calloc(1, size)))
4977 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4978 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4979 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4980 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4982 txn->mt_dbxs = env->me_dbxs;
4983 txn->mt_flags = MDB_TXN_FINISHED;
4993 mdb_env_close0(env, excl);
4995 mdb_fname_destroy(fname);
4999 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5001 mdb_env_close0(MDB_env *env, int excl)
5005 if (!(env->me_flags & MDB_ENV_ACTIVE))
5008 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5010 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5011 free(env->me_dbxs[i].md_name.mv_data);
5016 free(env->me_dbiseqs);
5017 free(env->me_dbflags);
5019 free(env->me_dirty_list);
5021 mdb_midl_free(env->me_free_pgs);
5023 if (env->me_flags & MDB_ENV_TXKEY) {
5024 pthread_key_delete(env->me_txkey);
5026 /* Delete our key from the global list */
5027 for (i=0; i<mdb_tls_nkeys; i++)
5028 if (mdb_tls_keys[i] == env->me_txkey) {
5029 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5037 munmap(env->me_map, env->me_mapsize);
5039 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
5040 (void) close(env->me_mfd);
5041 if (env->me_fd != INVALID_HANDLE_VALUE)
5042 (void) close(env->me_fd);
5044 MDB_PID_T pid = env->me_pid;
5045 /* Clearing readers is done in this function because
5046 * me_txkey with its destructor must be disabled first.
5048 * We skip the the reader mutex, so we touch only
5049 * data owned by this process (me_close_readers and
5050 * our readers), and clear each reader atomically.
5052 for (i = env->me_close_readers; --i >= 0; )
5053 if (env->me_txns->mti_readers[i].mr_pid == pid)
5054 env->me_txns->mti_readers[i].mr_pid = 0;
5056 if (env->me_rmutex) {
5057 CloseHandle(env->me_rmutex);
5058 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5060 /* Windows automatically destroys the mutexes when
5061 * the last handle closes.
5063 #elif defined(MDB_USE_POSIX_SEM)
5064 if (env->me_rmutex != SEM_FAILED) {
5065 sem_close(env->me_rmutex);
5066 if (env->me_wmutex != SEM_FAILED)
5067 sem_close(env->me_wmutex);
5068 /* If we have the filelock: If we are the
5069 * only remaining user, clean up semaphores.
5072 mdb_env_excl_lock(env, &excl);
5074 sem_unlink(env->me_txns->mti_rmname);
5075 sem_unlink(env->me_txns->mti_wmname);
5079 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5081 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5084 /* Unlock the lockfile. Windows would have unlocked it
5085 * after closing anyway, but not necessarily at once.
5087 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5090 (void) close(env->me_lfd);
5093 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5097 mdb_env_close(MDB_env *env)
5104 VGMEMP_DESTROY(env);
5105 while ((dp = env->me_dpages) != NULL) {
5106 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5107 env->me_dpages = dp->mp_next;
5111 mdb_env_close0(env, 0);
5115 /** Compare two items pointing at aligned size_t's */
5117 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5119 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5120 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5123 /** Compare two items pointing at aligned unsigned int's.
5125 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5126 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5129 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5131 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5132 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5135 /** Compare two items pointing at unsigned ints of unknown alignment.
5136 * Nodes and keys are guaranteed to be 2-byte aligned.
5139 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5141 #if BYTE_ORDER == LITTLE_ENDIAN
5142 unsigned short *u, *c;
5145 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5146 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5149 } while(!x && u > (unsigned short *)a->mv_data);
5152 unsigned short *u, *c, *end;
5155 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5156 u = (unsigned short *)a->mv_data;
5157 c = (unsigned short *)b->mv_data;
5160 } while(!x && u < end);
5165 /** Compare two items lexically */
5167 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5174 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5180 diff = memcmp(a->mv_data, b->mv_data, len);
5181 return diff ? diff : len_diff<0 ? -1 : len_diff;
5184 /** Compare two items in reverse byte order */
5186 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5188 const unsigned char *p1, *p2, *p1_lim;
5192 p1_lim = (const unsigned char *)a->mv_data;
5193 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5194 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5196 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5202 while (p1 > p1_lim) {
5203 diff = *--p1 - *--p2;
5207 return len_diff<0 ? -1 : len_diff;
5210 /** Search for key within a page, using binary search.
5211 * Returns the smallest entry larger or equal to the key.
5212 * If exactp is non-null, stores whether the found entry was an exact match
5213 * in *exactp (1 or 0).
5214 * Updates the cursor index with the index of the found entry.
5215 * If no entry larger or equal to the key is found, returns NULL.
5218 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5220 unsigned int i = 0, nkeys;
5223 MDB_page *mp = mc->mc_pg[mc->mc_top];
5224 MDB_node *node = NULL;
5229 nkeys = NUMKEYS(mp);
5231 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5232 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5235 low = IS_LEAF(mp) ? 0 : 1;
5237 cmp = mc->mc_dbx->md_cmp;
5239 /* Branch pages have no data, so if using integer keys,
5240 * alignment is guaranteed. Use faster mdb_cmp_int.
5242 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5243 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5250 nodekey.mv_size = mc->mc_db->md_pad;
5251 node = NODEPTR(mp, 0); /* fake */
5252 while (low <= high) {
5253 i = (low + high) >> 1;
5254 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5255 rc = cmp(key, &nodekey);
5256 DPRINTF(("found leaf index %u [%s], rc = %i",
5257 i, DKEY(&nodekey), rc));
5266 while (low <= high) {
5267 i = (low + high) >> 1;
5269 node = NODEPTR(mp, i);
5270 nodekey.mv_size = NODEKSZ(node);
5271 nodekey.mv_data = NODEKEY(node);
5273 rc = cmp(key, &nodekey);
5276 DPRINTF(("found leaf index %u [%s], rc = %i",
5277 i, DKEY(&nodekey), rc));
5279 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5280 i, DKEY(&nodekey), NODEPGNO(node), rc));
5291 if (rc > 0) { /* Found entry is less than the key. */
5292 i++; /* Skip to get the smallest entry larger than key. */
5294 node = NODEPTR(mp, i);
5297 *exactp = (rc == 0 && nkeys > 0);
5298 /* store the key index */
5299 mc->mc_ki[mc->mc_top] = i;
5301 /* There is no entry larger or equal to the key. */
5304 /* nodeptr is fake for LEAF2 */
5310 mdb_cursor_adjust(MDB_cursor *mc, func)
5314 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5315 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5322 /** Pop a page off the top of the cursor's stack. */
5324 mdb_cursor_pop(MDB_cursor *mc)
5327 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5328 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5334 mc->mc_flags &= ~C_INITIALIZED;
5339 /** Push a page onto the top of the cursor's stack. */
5341 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5343 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5344 DDBI(mc), (void *) mc));
5346 if (mc->mc_snum >= CURSOR_STACK) {
5347 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5348 return MDB_CURSOR_FULL;
5351 mc->mc_top = mc->mc_snum++;
5352 mc->mc_pg[mc->mc_top] = mp;
5353 mc->mc_ki[mc->mc_top] = 0;
5358 /** Find the address of the page corresponding to a given page number.
5359 * @param[in] mc the cursor accessing the page.
5360 * @param[in] pgno the page number for the page to retrieve.
5361 * @param[out] ret address of a pointer where the page's address will be stored.
5362 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5363 * @return 0 on success, non-zero on failure.
5366 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
5368 MDB_txn *txn = mc->mc_txn;
5369 MDB_env *env = txn->mt_env;
5373 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5377 MDB_ID2L dl = tx2->mt_u.dirty_list;
5379 /* Spilled pages were dirtied in this txn and flushed
5380 * because the dirty list got full. Bring this page
5381 * back in from the map (but don't unspill it here,
5382 * leave that unless page_touch happens again).
5384 if (tx2->mt_spill_pgs) {
5385 MDB_ID pn = pgno << 1;
5386 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5387 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5388 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5393 unsigned x = mdb_mid2l_search(dl, pgno);
5394 if (x <= dl[0].mid && dl[x].mid == pgno) {
5400 } while ((tx2 = tx2->mt_parent) != NULL);
5403 if (pgno < txn->mt_next_pgno) {
5405 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5407 DPRINTF(("page %"Z"u not found", pgno));
5408 txn->mt_flags |= MDB_TXN_ERROR;
5409 return MDB_PAGE_NOTFOUND;
5419 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5420 * The cursor is at the root page, set up the rest of it.
5423 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5425 MDB_page *mp = mc->mc_pg[mc->mc_top];
5429 while (IS_BRANCH(mp)) {
5433 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5434 /* Don't assert on branch pages in the FreeDB. We can get here
5435 * while in the process of rebalancing a FreeDB branch page; we must
5436 * let that proceed. ITS#8336
5438 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
5439 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5441 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5443 if (flags & MDB_PS_LAST)
5444 i = NUMKEYS(mp) - 1;
5447 node = mdb_node_search(mc, key, &exact);
5449 i = NUMKEYS(mp) - 1;
5451 i = mc->mc_ki[mc->mc_top];
5453 mdb_cassert(mc, i > 0);
5457 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5460 mdb_cassert(mc, i < NUMKEYS(mp));
5461 node = NODEPTR(mp, i);
5463 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5466 mc->mc_ki[mc->mc_top] = i;
5467 if ((rc = mdb_cursor_push(mc, mp)))
5470 if (flags & MDB_PS_MODIFY) {
5471 if ((rc = mdb_page_touch(mc)) != 0)
5473 mp = mc->mc_pg[mc->mc_top];
5478 DPRINTF(("internal error, index points to a %02X page!?",
5480 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5481 return MDB_CORRUPTED;
5484 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5485 key ? DKEY(key) : "null"));
5486 mc->mc_flags |= C_INITIALIZED;
5487 mc->mc_flags &= ~C_EOF;
5492 /** Search for the lowest key under the current branch page.
5493 * This just bypasses a NUMKEYS check in the current page
5494 * before calling mdb_page_search_root(), because the callers
5495 * are all in situations where the current page is known to
5499 mdb_page_search_lowest(MDB_cursor *mc)
5501 MDB_page *mp = mc->mc_pg[mc->mc_top];
5502 MDB_node *node = NODEPTR(mp, 0);
5505 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5508 mc->mc_ki[mc->mc_top] = 0;
5509 if ((rc = mdb_cursor_push(mc, mp)))
5511 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5514 /** Search for the page a given key should be in.
5515 * Push it and its parent pages on the cursor stack.
5516 * @param[in,out] mc the cursor for this operation.
5517 * @param[in] key the key to search for, or NULL for first/last page.
5518 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5519 * are touched (updated with new page numbers).
5520 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5521 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5522 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5523 * @return 0 on success, non-zero on failure.
5526 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5531 /* Make sure the txn is still viable, then find the root from
5532 * the txn's db table and set it as the root of the cursor's stack.
5534 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5535 DPUTS("transaction may not be used now");
5538 /* Make sure we're using an up-to-date root */
5539 if (*mc->mc_dbflag & DB_STALE) {
5541 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5543 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5544 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5551 MDB_node *leaf = mdb_node_search(&mc2,
5552 &mc->mc_dbx->md_name, &exact);
5554 return MDB_NOTFOUND;
5555 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5556 return MDB_INCOMPATIBLE; /* not a named DB */
5557 rc = mdb_node_read(&mc2, leaf, &data);
5560 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5562 /* The txn may not know this DBI, or another process may
5563 * have dropped and recreated the DB with other flags.
5565 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5566 return MDB_INCOMPATIBLE;
5567 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5569 *mc->mc_dbflag &= ~DB_STALE;
5571 root = mc->mc_db->md_root;
5573 if (root == P_INVALID) { /* Tree is empty. */
5574 DPUTS("tree is empty");
5575 return MDB_NOTFOUND;
5579 mdb_cassert(mc, root > 1);
5580 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5581 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
5587 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5588 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5590 if (flags & MDB_PS_MODIFY) {
5591 if ((rc = mdb_page_touch(mc)))
5595 if (flags & MDB_PS_ROOTONLY)
5598 return mdb_page_search_root(mc, key, flags);
5602 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5604 MDB_txn *txn = mc->mc_txn;
5605 pgno_t pg = mp->mp_pgno;
5606 unsigned x = 0, ovpages = mp->mp_pages;
5607 MDB_env *env = txn->mt_env;
5608 MDB_IDL sl = txn->mt_spill_pgs;
5609 MDB_ID pn = pg << 1;
5612 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5613 /* If the page is dirty or on the spill list we just acquired it,
5614 * so we should give it back to our current free list, if any.
5615 * Otherwise put it onto the list of pages we freed in this txn.
5617 * Won't create me_pghead: me_pglast must be inited along with it.
5618 * Unsupported in nested txns: They would need to hide the page
5619 * range in ancestor txns' dirty and spilled lists.
5621 if (env->me_pghead &&
5623 ((mp->mp_flags & P_DIRTY) ||
5624 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5628 MDB_ID2 *dl, ix, iy;
5629 rc = mdb_midl_need(&env->me_pghead, ovpages);
5632 if (!(mp->mp_flags & P_DIRTY)) {
5633 /* This page is no longer spilled */
5640 /* Remove from dirty list */
5641 dl = txn->mt_u.dirty_list;
5643 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5649 mdb_cassert(mc, x > 1);
5651 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5652 txn->mt_flags |= MDB_TXN_ERROR;
5653 return MDB_CORRUPTED;
5656 txn->mt_dirty_room++;
5657 if (!(env->me_flags & MDB_WRITEMAP))
5658 mdb_dpage_free(env, mp);
5660 /* Insert in me_pghead */
5661 mop = env->me_pghead;
5662 j = mop[0] + ovpages;
5663 for (i = mop[0]; i && mop[i] < pg; i--)
5669 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5673 mc->mc_db->md_overflow_pages -= ovpages;
5677 /** Return the data associated with a given node.
5678 * @param[in] mc The cursor for this operation.
5679 * @param[in] leaf The node being read.
5680 * @param[out] data Updated to point to the node's data.
5681 * @return 0 on success, non-zero on failure.
5684 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
5686 MDB_page *omp; /* overflow page */
5690 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5691 data->mv_size = NODEDSZ(leaf);
5692 data->mv_data = NODEDATA(leaf);
5696 /* Read overflow data.
5698 data->mv_size = NODEDSZ(leaf);
5699 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5700 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
5701 DPRINTF(("read overflow page %"Z"u failed", pgno));
5704 data->mv_data = METADATA(omp);
5710 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5711 MDB_val *key, MDB_val *data)
5718 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5720 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5723 if (txn->mt_flags & MDB_TXN_BLOCKED)
5726 mdb_cursor_init(&mc, txn, dbi, &mx);
5727 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5730 /** Find a sibling for a page.
5731 * Replaces the page at the top of the cursor's stack with the
5732 * specified sibling, if one exists.
5733 * @param[in] mc The cursor for this operation.
5734 * @param[in] move_right Non-zero if the right sibling is requested,
5735 * otherwise the left sibling.
5736 * @return 0 on success, non-zero on failure.
5739 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5745 if (mc->mc_snum < 2) {
5746 return MDB_NOTFOUND; /* root has no siblings */
5750 DPRINTF(("parent page is page %"Z"u, index %u",
5751 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5753 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5754 : (mc->mc_ki[mc->mc_top] == 0)) {
5755 DPRINTF(("no more keys left, moving to %s sibling",
5756 move_right ? "right" : "left"));
5757 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5758 /* undo cursor_pop before returning */
5765 mc->mc_ki[mc->mc_top]++;
5767 mc->mc_ki[mc->mc_top]--;
5768 DPRINTF(("just moving to %s index key %u",
5769 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5771 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5773 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5774 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
5775 /* mc will be inconsistent if caller does mc_snum++ as above */
5776 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5780 mdb_cursor_push(mc, mp);
5782 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5787 /** Move the cursor to the next data item. */
5789 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5795 if ((mc->mc_flags & C_EOF) ||
5796 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
5797 return MDB_NOTFOUND;
5799 if (!(mc->mc_flags & C_INITIALIZED))
5800 return mdb_cursor_first(mc, key, data);
5802 mp = mc->mc_pg[mc->mc_top];
5804 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5805 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5806 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5807 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5808 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5809 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5810 if (rc == MDB_SUCCESS)
5811 MDB_GET_KEY(leaf, key);
5816 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5817 if (op == MDB_NEXT_DUP)
5818 return MDB_NOTFOUND;
5822 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5823 mdb_dbg_pgno(mp), (void *) mc));
5824 if (mc->mc_flags & C_DEL) {
5825 mc->mc_flags ^= C_DEL;
5829 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5830 DPUTS("=====> move to next sibling page");
5831 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5832 mc->mc_flags |= C_EOF;
5835 mp = mc->mc_pg[mc->mc_top];
5836 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5838 mc->mc_ki[mc->mc_top]++;
5841 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5842 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5845 key->mv_size = mc->mc_db->md_pad;
5846 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5850 mdb_cassert(mc, IS_LEAF(mp));
5851 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5853 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5854 mdb_xcursor_init1(mc, leaf);
5857 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5860 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5861 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5862 if (rc != MDB_SUCCESS)
5867 MDB_GET_KEY(leaf, key);
5871 /** Move the cursor to the previous data item. */
5873 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5879 if (!(mc->mc_flags & C_INITIALIZED)) {
5880 rc = mdb_cursor_last(mc, key, data);
5883 mc->mc_ki[mc->mc_top]++;
5886 mp = mc->mc_pg[mc->mc_top];
5888 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5889 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5890 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5891 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5892 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5893 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5894 if (rc == MDB_SUCCESS) {
5895 MDB_GET_KEY(leaf, key);
5896 mc->mc_flags &= ~C_EOF;
5902 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5903 if (op == MDB_PREV_DUP)
5904 return MDB_NOTFOUND;
5908 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5909 mdb_dbg_pgno(mp), (void *) mc));
5911 mc->mc_flags &= ~(C_EOF|C_DEL);
5913 if (mc->mc_ki[mc->mc_top] == 0) {
5914 DPUTS("=====> move to prev sibling page");
5915 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5918 mp = mc->mc_pg[mc->mc_top];
5919 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5920 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5922 mc->mc_ki[mc->mc_top]--;
5924 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5925 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5928 key->mv_size = mc->mc_db->md_pad;
5929 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5933 mdb_cassert(mc, IS_LEAF(mp));
5934 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5936 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5937 mdb_xcursor_init1(mc, leaf);
5940 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5943 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5944 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5945 if (rc != MDB_SUCCESS)
5950 MDB_GET_KEY(leaf, key);
5954 /** Set the cursor on a specific data item. */
5956 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5957 MDB_cursor_op op, int *exactp)
5961 MDB_node *leaf = NULL;
5964 if (key->mv_size == 0)
5965 return MDB_BAD_VALSIZE;
5968 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5970 /* See if we're already on the right page */
5971 if (mc->mc_flags & C_INITIALIZED) {
5974 mp = mc->mc_pg[mc->mc_top];
5976 mc->mc_ki[mc->mc_top] = 0;
5977 return MDB_NOTFOUND;
5979 if (mp->mp_flags & P_LEAF2) {
5980 nodekey.mv_size = mc->mc_db->md_pad;
5981 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5983 leaf = NODEPTR(mp, 0);
5984 MDB_GET_KEY2(leaf, nodekey);
5986 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5988 /* Probably happens rarely, but first node on the page
5989 * was the one we wanted.
5991 mc->mc_ki[mc->mc_top] = 0;
5998 unsigned int nkeys = NUMKEYS(mp);
6000 if (mp->mp_flags & P_LEAF2) {
6001 nodekey.mv_data = LEAF2KEY(mp,
6002 nkeys-1, nodekey.mv_size);
6004 leaf = NODEPTR(mp, nkeys-1);
6005 MDB_GET_KEY2(leaf, nodekey);
6007 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6009 /* last node was the one we wanted */
6010 mc->mc_ki[mc->mc_top] = nkeys-1;
6016 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6017 /* This is definitely the right page, skip search_page */
6018 if (mp->mp_flags & P_LEAF2) {
6019 nodekey.mv_data = LEAF2KEY(mp,
6020 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6022 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6023 MDB_GET_KEY2(leaf, nodekey);
6025 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6027 /* current node was the one we wanted */
6034 mc->mc_flags &= ~C_EOF;
6038 /* If any parents have right-sibs, search.
6039 * Otherwise, there's nothing further.
6041 for (i=0; i<mc->mc_top; i++)
6043 NUMKEYS(mc->mc_pg[i])-1)
6045 if (i == mc->mc_top) {
6046 /* There are no other pages */
6047 mc->mc_ki[mc->mc_top] = nkeys;
6048 return MDB_NOTFOUND;
6052 /* There are no other pages */
6053 mc->mc_ki[mc->mc_top] = 0;
6054 if (op == MDB_SET_RANGE && !exactp) {
6058 return MDB_NOTFOUND;
6064 rc = mdb_page_search(mc, key, 0);
6065 if (rc != MDB_SUCCESS)
6068 mp = mc->mc_pg[mc->mc_top];
6069 mdb_cassert(mc, IS_LEAF(mp));
6072 leaf = mdb_node_search(mc, key, exactp);
6073 if (exactp != NULL && !*exactp) {
6074 /* MDB_SET specified and not an exact match. */
6075 return MDB_NOTFOUND;
6079 DPUTS("===> inexact leaf not found, goto sibling");
6080 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6081 mc->mc_flags |= C_EOF;
6082 return rc; /* no entries matched */
6084 mp = mc->mc_pg[mc->mc_top];
6085 mdb_cassert(mc, IS_LEAF(mp));
6086 leaf = NODEPTR(mp, 0);
6090 mc->mc_flags |= C_INITIALIZED;
6091 mc->mc_flags &= ~C_EOF;
6094 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6095 key->mv_size = mc->mc_db->md_pad;
6096 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6101 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6102 mdb_xcursor_init1(mc, leaf);
6105 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6106 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6107 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6110 if (op == MDB_GET_BOTH) {
6116 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6117 if (rc != MDB_SUCCESS)
6120 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6123 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6125 dcmp = mc->mc_dbx->md_dcmp;
6126 #if UINT_MAX < SIZE_MAX
6127 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6128 dcmp = mdb_cmp_clong;
6130 rc = dcmp(data, &olddata);
6132 if (op == MDB_GET_BOTH || rc > 0)
6133 return MDB_NOTFOUND;
6140 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6141 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6146 /* The key already matches in all other cases */
6147 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6148 MDB_GET_KEY(leaf, key);
6149 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6154 /** Move the cursor to the first item in the database. */
6156 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6162 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6164 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6165 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6166 if (rc != MDB_SUCCESS)
6169 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6171 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6172 mc->mc_flags |= C_INITIALIZED;
6173 mc->mc_flags &= ~C_EOF;
6175 mc->mc_ki[mc->mc_top] = 0;
6177 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6178 key->mv_size = mc->mc_db->md_pad;
6179 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6184 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6185 mdb_xcursor_init1(mc, leaf);
6186 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6190 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6194 MDB_GET_KEY(leaf, key);
6198 /** Move the cursor to the last item in the database. */
6200 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6206 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6208 if (!(mc->mc_flags & C_EOF)) {
6210 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6211 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6212 if (rc != MDB_SUCCESS)
6215 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6218 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6219 mc->mc_flags |= C_INITIALIZED|C_EOF;
6220 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6222 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6223 key->mv_size = mc->mc_db->md_pad;
6224 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6229 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6230 mdb_xcursor_init1(mc, leaf);
6231 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6235 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6240 MDB_GET_KEY(leaf, key);
6245 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6250 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6255 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6259 case MDB_GET_CURRENT:
6260 if (!(mc->mc_flags & C_INITIALIZED)) {
6263 MDB_page *mp = mc->mc_pg[mc->mc_top];
6264 int nkeys = NUMKEYS(mp);
6265 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6266 mc->mc_ki[mc->mc_top] = nkeys;
6272 key->mv_size = mc->mc_db->md_pad;
6273 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6275 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6276 MDB_GET_KEY(leaf, key);
6278 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6279 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6281 rc = mdb_node_read(mc, leaf, data);
6288 case MDB_GET_BOTH_RANGE:
6293 if (mc->mc_xcursor == NULL) {
6294 rc = MDB_INCOMPATIBLE;
6304 rc = mdb_cursor_set(mc, key, data, op,
6305 op == MDB_SET_RANGE ? NULL : &exact);
6308 case MDB_GET_MULTIPLE:
6309 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6313 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6314 rc = MDB_INCOMPATIBLE;
6318 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6319 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6322 case MDB_NEXT_MULTIPLE:
6327 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6328 rc = MDB_INCOMPATIBLE;
6331 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6332 if (rc == MDB_SUCCESS) {
6333 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6336 mx = &mc->mc_xcursor->mx_cursor;
6337 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6339 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6340 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6346 case MDB_PREV_MULTIPLE:
6351 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6352 rc = MDB_INCOMPATIBLE;
6355 if (!(mc->mc_flags & C_INITIALIZED))
6356 rc = mdb_cursor_last(mc, key, data);
6359 if (rc == MDB_SUCCESS) {
6360 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
6361 if (mx->mc_flags & C_INITIALIZED) {
6362 rc = mdb_cursor_sibling(mx, 0);
6363 if (rc == MDB_SUCCESS)
6372 case MDB_NEXT_NODUP:
6373 rc = mdb_cursor_next(mc, key, data, op);
6377 case MDB_PREV_NODUP:
6378 rc = mdb_cursor_prev(mc, key, data, op);
6381 rc = mdb_cursor_first(mc, key, data);
6384 mfunc = mdb_cursor_first;
6386 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6390 if (mc->mc_xcursor == NULL) {
6391 rc = MDB_INCOMPATIBLE;
6395 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6396 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6397 MDB_GET_KEY(leaf, key);
6398 rc = mdb_node_read(mc, leaf, data);
6402 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6406 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6409 rc = mdb_cursor_last(mc, key, data);
6412 mfunc = mdb_cursor_last;
6415 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6420 if (mc->mc_flags & C_DEL)
6421 mc->mc_flags ^= C_DEL;
6426 /** Touch all the pages in the cursor stack. Set mc_top.
6427 * Makes sure all the pages are writable, before attempting a write operation.
6428 * @param[in] mc The cursor to operate on.
6431 mdb_cursor_touch(MDB_cursor *mc)
6433 int rc = MDB_SUCCESS;
6435 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & (DB_DIRTY|DB_DUPDATA))) {
6436 /* Touch DB record of named DB */
6439 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6441 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6442 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6445 *mc->mc_dbflag |= DB_DIRTY;
6450 rc = mdb_page_touch(mc);
6451 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6452 mc->mc_top = mc->mc_snum-1;
6457 /** Do not spill pages to disk if txn is getting full, may fail instead */
6458 #define MDB_NOSPILL 0x8000
6461 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6465 MDB_node *leaf = NULL;
6466 MDB_page *fp, *mp, *sub_root = NULL;
6468 MDB_val xdata, *rdata, dkey, olddata;
6470 int do_sub = 0, insert_key, insert_data;
6471 unsigned int mcount = 0, dcount = 0, nospill;
6474 unsigned int nflags;
6477 if (mc == NULL || key == NULL)
6480 env = mc->mc_txn->mt_env;
6482 /* Check this first so counter will always be zero on any
6485 if (flags & MDB_MULTIPLE) {
6486 dcount = data[1].mv_size;
6487 data[1].mv_size = 0;
6488 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6489 return MDB_INCOMPATIBLE;
6492 nospill = flags & MDB_NOSPILL;
6493 flags &= ~MDB_NOSPILL;
6495 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6496 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6498 if (key->mv_size-1 >= ENV_MAXKEY(env))
6499 return MDB_BAD_VALSIZE;
6501 #if SIZE_MAX > MAXDATASIZE
6502 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6503 return MDB_BAD_VALSIZE;
6505 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6506 return MDB_BAD_VALSIZE;
6509 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6510 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6514 if (flags == MDB_CURRENT) {
6515 if (!(mc->mc_flags & C_INITIALIZED))
6518 } else if (mc->mc_db->md_root == P_INVALID) {
6519 /* new database, cursor has nothing to point to */
6522 mc->mc_flags &= ~C_INITIALIZED;
6527 if (flags & MDB_APPEND) {
6529 rc = mdb_cursor_last(mc, &k2, &d2);
6531 rc = mc->mc_dbx->md_cmp(key, &k2);
6534 mc->mc_ki[mc->mc_top]++;
6536 /* new key is <= last key */
6541 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6543 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6544 DPRINTF(("duplicate key [%s]", DKEY(key)));
6546 return MDB_KEYEXIST;
6548 if (rc && rc != MDB_NOTFOUND)
6552 if (mc->mc_flags & C_DEL)
6553 mc->mc_flags ^= C_DEL;
6555 /* Cursor is positioned, check for room in the dirty list */
6557 if (flags & MDB_MULTIPLE) {
6559 xdata.mv_size = data->mv_size * dcount;
6563 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6567 if (rc == MDB_NO_ROOT) {
6569 /* new database, write a root leaf page */
6570 DPUTS("allocating new root leaf page");
6571 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6574 mdb_cursor_push(mc, np);
6575 mc->mc_db->md_root = np->mp_pgno;
6576 mc->mc_db->md_depth++;
6577 *mc->mc_dbflag |= DB_DIRTY;
6578 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6580 np->mp_flags |= P_LEAF2;
6581 mc->mc_flags |= C_INITIALIZED;
6583 /* make sure all cursor pages are writable */
6584 rc2 = mdb_cursor_touch(mc);
6589 insert_key = insert_data = rc;
6591 /* The key does not exist */
6592 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6593 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6594 LEAFSIZE(key, data) > env->me_nodemax)
6596 /* Too big for a node, insert in sub-DB. Set up an empty
6597 * "old sub-page" for prep_subDB to expand to a full page.
6599 fp_flags = P_LEAF|P_DIRTY;
6601 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6602 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6603 olddata.mv_size = PAGEHDRSZ;
6607 /* there's only a key anyway, so this is a no-op */
6608 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6610 unsigned int ksize = mc->mc_db->md_pad;
6611 if (key->mv_size != ksize)
6612 return MDB_BAD_VALSIZE;
6613 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6614 memcpy(ptr, key->mv_data, ksize);
6616 /* if overwriting slot 0 of leaf, need to
6617 * update branch key if there is a parent page
6619 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6620 unsigned short dtop = 1;
6622 /* slot 0 is always an empty key, find real slot */
6623 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6627 if (mc->mc_ki[mc->mc_top])
6628 rc2 = mdb_update_key(mc, key);
6639 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6640 olddata.mv_size = NODEDSZ(leaf);
6641 olddata.mv_data = NODEDATA(leaf);
6644 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6645 /* Prepare (sub-)page/sub-DB to accept the new item,
6646 * if needed. fp: old sub-page or a header faking
6647 * it. mp: new (sub-)page. offset: growth in page
6648 * size. xdata: node data with new page or DB.
6650 unsigned i, offset = 0;
6651 mp = fp = xdata.mv_data = env->me_pbuf;
6652 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6654 /* Was a single item before, must convert now */
6655 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6657 /* Just overwrite the current item */
6658 if (flags == MDB_CURRENT)
6660 dcmp = mc->mc_dbx->md_dcmp;
6661 #if UINT_MAX < SIZE_MAX
6662 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6663 dcmp = mdb_cmp_clong;
6665 /* does data match? */
6666 if (!dcmp(data, &olddata)) {
6667 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
6668 return MDB_KEYEXIST;
6673 /* Back up original data item */
6674 dkey.mv_size = olddata.mv_size;
6675 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6677 /* Make sub-page header for the dup items, with dummy body */
6678 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6679 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6680 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6681 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6682 fp->mp_flags |= P_LEAF2;
6683 fp->mp_pad = data->mv_size;
6684 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6686 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6687 (dkey.mv_size & 1) + (data->mv_size & 1);
6689 fp->mp_upper = xdata.mv_size - PAGEBASE;
6690 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6691 } else if (leaf->mn_flags & F_SUBDATA) {
6692 /* Data is on sub-DB, just store it */
6693 flags |= F_DUPDATA|F_SUBDATA;
6696 /* Data is on sub-page */
6697 fp = olddata.mv_data;
6700 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6701 offset = EVEN(NODESIZE + sizeof(indx_t) +
6705 offset = fp->mp_pad;
6706 if (SIZELEFT(fp) < offset) {
6707 offset *= 4; /* space for 4 more */
6710 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6712 fp->mp_flags |= P_DIRTY;
6713 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6714 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6718 xdata.mv_size = olddata.mv_size + offset;
6721 fp_flags = fp->mp_flags;
6722 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6723 /* Too big for a sub-page, convert to sub-DB */
6724 fp_flags &= ~P_SUBP;
6726 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6727 fp_flags |= P_LEAF2;
6728 dummy.md_pad = fp->mp_pad;
6729 dummy.md_flags = MDB_DUPFIXED;
6730 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6731 dummy.md_flags |= MDB_INTEGERKEY;
6737 dummy.md_branch_pages = 0;
6738 dummy.md_leaf_pages = 1;
6739 dummy.md_overflow_pages = 0;
6740 dummy.md_entries = NUMKEYS(fp);
6741 xdata.mv_size = sizeof(MDB_db);
6742 xdata.mv_data = &dummy;
6743 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6745 offset = env->me_psize - olddata.mv_size;
6746 flags |= F_DUPDATA|F_SUBDATA;
6747 dummy.md_root = mp->mp_pgno;
6751 mp->mp_flags = fp_flags | P_DIRTY;
6752 mp->mp_pad = fp->mp_pad;
6753 mp->mp_lower = fp->mp_lower;
6754 mp->mp_upper = fp->mp_upper + offset;
6755 if (fp_flags & P_LEAF2) {
6756 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6758 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6759 olddata.mv_size - fp->mp_upper - PAGEBASE);
6760 for (i=0; i<NUMKEYS(fp); i++)
6761 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6769 mdb_node_del(mc, 0);
6773 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6774 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6775 return MDB_INCOMPATIBLE;
6776 /* overflow page overwrites need special handling */
6777 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6780 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6782 memcpy(&pg, olddata.mv_data, sizeof(pg));
6783 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
6785 ovpages = omp->mp_pages;
6787 /* Is the ov page large enough? */
6788 if (ovpages >= dpages) {
6789 if (!(omp->mp_flags & P_DIRTY) &&
6790 (level || (env->me_flags & MDB_WRITEMAP)))
6792 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6795 level = 0; /* dirty in this txn or clean */
6798 if (omp->mp_flags & P_DIRTY) {
6799 /* yes, overwrite it. Note in this case we don't
6800 * bother to try shrinking the page if the new data
6801 * is smaller than the overflow threshold.
6804 /* It is writable only in a parent txn */
6805 size_t sz = (size_t) env->me_psize * ovpages, off;
6806 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6812 /* Note - this page is already counted in parent's dirty_room */
6813 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6814 mdb_cassert(mc, rc2 == 0);
6815 /* Currently we make the page look as with put() in the
6816 * parent txn, in case the user peeks at MDB_RESERVEd
6817 * or unused parts. Some users treat ovpages specially.
6819 if (!(flags & MDB_RESERVE)) {
6820 /* Skip the part where LMDB will put *data.
6821 * Copy end of page, adjusting alignment so
6822 * compiler may copy words instead of bytes.
6824 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6825 memcpy((size_t *)((char *)np + off),
6826 (size_t *)((char *)omp + off), sz - off);
6829 memcpy(np, omp, sz); /* Copy beginning of page */
6832 SETDSZ(leaf, data->mv_size);
6833 if (F_ISSET(flags, MDB_RESERVE))
6834 data->mv_data = METADATA(omp);
6836 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6840 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6842 } else if (data->mv_size == olddata.mv_size) {
6843 /* same size, just replace it. Note that we could
6844 * also reuse this node if the new data is smaller,
6845 * but instead we opt to shrink the node in that case.
6847 if (F_ISSET(flags, MDB_RESERVE))
6848 data->mv_data = olddata.mv_data;
6849 else if (!(mc->mc_flags & C_SUB))
6850 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6852 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6857 mdb_node_del(mc, 0);
6863 nflags = flags & NODE_ADD_FLAGS;
6864 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6865 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6866 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6867 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6869 nflags |= MDB_SPLIT_REPLACE;
6870 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6872 /* There is room already in this leaf page. */
6873 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6875 /* Adjust other cursors pointing to mp */
6876 MDB_cursor *m2, *m3;
6877 MDB_dbi dbi = mc->mc_dbi;
6878 unsigned i = mc->mc_top;
6879 MDB_page *mp = mc->mc_pg[i];
6881 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6882 if (mc->mc_flags & C_SUB)
6883 m3 = &m2->mc_xcursor->mx_cursor;
6886 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
6887 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
6890 if (XCURSOR_INITED(m3))
6891 XCURSOR_REFRESH(m3, mp, m3->mc_ki[i]);
6896 if (rc == MDB_SUCCESS) {
6897 /* Now store the actual data in the child DB. Note that we're
6898 * storing the user data in the keys field, so there are strict
6899 * size limits on dupdata. The actual data fields of the child
6900 * DB are all zero size.
6903 int xflags, new_dupdata;
6908 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6909 if (flags & MDB_CURRENT) {
6910 xflags = MDB_CURRENT|MDB_NOSPILL;
6912 mdb_xcursor_init1(mc, leaf);
6913 xflags = (flags & MDB_NODUPDATA) ?
6914 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6917 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6918 new_dupdata = (int)dkey.mv_size;
6919 /* converted, write the original data first */
6921 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6924 /* we've done our job */
6927 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6928 /* Adjust other cursors pointing to mp */
6930 MDB_xcursor *mx = mc->mc_xcursor;
6931 unsigned i = mc->mc_top;
6932 MDB_page *mp = mc->mc_pg[i];
6933 int nkeys = NUMKEYS(mp);
6935 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6936 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6937 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6938 if (m2->mc_pg[i] == mp) {
6939 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6940 mdb_xcursor_init2(m2, mx, new_dupdata);
6941 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
6942 XCURSOR_REFRESH(m2, mp, m2->mc_ki[i]);
6947 ecount = mc->mc_xcursor->mx_db.md_entries;
6948 if (flags & MDB_APPENDDUP)
6949 xflags |= MDB_APPEND;
6950 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6951 if (flags & F_SUBDATA) {
6952 void *db = NODEDATA(leaf);
6953 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6955 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6957 /* Increment count unless we just replaced an existing item. */
6959 mc->mc_db->md_entries++;
6961 /* Invalidate txn if we created an empty sub-DB */
6964 /* If we succeeded and the key didn't exist before,
6965 * make sure the cursor is marked valid.
6967 mc->mc_flags |= C_INITIALIZED;
6969 if (flags & MDB_MULTIPLE) {
6972 /* let caller know how many succeeded, if any */
6973 data[1].mv_size = mcount;
6974 if (mcount < dcount) {
6975 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6976 insert_key = insert_data = 0;
6983 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6986 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6991 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6997 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6998 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7000 if (!(mc->mc_flags & C_INITIALIZED))
7003 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7004 return MDB_NOTFOUND;
7006 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7009 rc = mdb_cursor_touch(mc);
7013 mp = mc->mc_pg[mc->mc_top];
7016 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7018 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7019 if (flags & MDB_NODUPDATA) {
7020 /* mdb_cursor_del0() will subtract the final entry */
7021 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7022 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7024 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7025 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7027 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7030 /* If sub-DB still has entries, we're done */
7031 if (mc->mc_xcursor->mx_db.md_entries) {
7032 if (leaf->mn_flags & F_SUBDATA) {
7033 /* update subDB info */
7034 void *db = NODEDATA(leaf);
7035 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7038 /* shrink fake page */
7039 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7040 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7041 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7042 /* fix other sub-DB cursors pointed at fake pages on this page */
7043 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7044 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7045 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7046 if (m2->mc_pg[mc->mc_top] == mp) {
7047 MDB_node *n2 = leaf;
7048 if (m2->mc_ki[mc->mc_top] != mc->mc_ki[mc->mc_top]) {
7049 n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7050 if (n2->mn_flags & F_SUBDATA) continue;
7052 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7056 mc->mc_db->md_entries--;
7059 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7061 /* otherwise fall thru and delete the sub-DB */
7064 if (leaf->mn_flags & F_SUBDATA) {
7065 /* add all the child DB's pages to the free list */
7066 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7071 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7072 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7073 rc = MDB_INCOMPATIBLE;
7077 /* add overflow pages to free list */
7078 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7082 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7083 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7084 (rc = mdb_ovpage_free(mc, omp)))
7089 return mdb_cursor_del0(mc);
7092 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7096 /** Allocate and initialize new pages for a database.
7097 * @param[in] mc a cursor on the database being added to.
7098 * @param[in] flags flags defining what type of page is being allocated.
7099 * @param[in] num the number of pages to allocate. This is usually 1,
7100 * unless allocating overflow pages for a large record.
7101 * @param[out] mp Address of a page, or NULL on failure.
7102 * @return 0 on success, non-zero on failure.
7105 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7110 if ((rc = mdb_page_alloc(mc, num, &np)))
7112 DPRINTF(("allocated new mpage %"Z"u, page size %u",
7113 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7114 np->mp_flags = flags | P_DIRTY;
7115 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7116 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7119 mc->mc_db->md_branch_pages++;
7120 else if (IS_LEAF(np))
7121 mc->mc_db->md_leaf_pages++;
7122 else if (IS_OVERFLOW(np)) {
7123 mc->mc_db->md_overflow_pages += num;
7131 /** Calculate the size of a leaf node.
7132 * The size depends on the environment's page size; if a data item
7133 * is too large it will be put onto an overflow page and the node
7134 * size will only include the key and not the data. Sizes are always
7135 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7136 * of the #MDB_node headers.
7137 * @param[in] env The environment handle.
7138 * @param[in] key The key for the node.
7139 * @param[in] data The data for the node.
7140 * @return The number of bytes needed to store the node.
7143 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7147 sz = LEAFSIZE(key, data);
7148 if (sz > env->me_nodemax) {
7149 /* put on overflow page */
7150 sz -= data->mv_size - sizeof(pgno_t);
7153 return EVEN(sz + sizeof(indx_t));
7156 /** Calculate the size of a branch node.
7157 * The size should depend on the environment's page size but since
7158 * we currently don't support spilling large keys onto overflow
7159 * pages, it's simply the size of the #MDB_node header plus the
7160 * size of the key. Sizes are always rounded up to an even number
7161 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7162 * @param[in] env The environment handle.
7163 * @param[in] key The key for the node.
7164 * @return The number of bytes needed to store the node.
7167 mdb_branch_size(MDB_env *env, MDB_val *key)
7172 if (sz > env->me_nodemax) {
7173 /* put on overflow page */
7174 /* not implemented */
7175 /* sz -= key->size - sizeof(pgno_t); */
7178 return sz + sizeof(indx_t);
7181 /** Add a node to the page pointed to by the cursor.
7182 * @param[in] mc The cursor for this operation.
7183 * @param[in] indx The index on the page where the new node should be added.
7184 * @param[in] key The key for the new node.
7185 * @param[in] data The data for the new node, if any.
7186 * @param[in] pgno The page number, if adding a branch node.
7187 * @param[in] flags Flags for the node.
7188 * @return 0 on success, non-zero on failure. Possible errors are:
7190 * <li>ENOMEM - failed to allocate overflow pages for the node.
7191 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7192 * should never happen since all callers already calculate the
7193 * page's free space before calling this function.
7197 mdb_node_add(MDB_cursor *mc, indx_t indx,
7198 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7201 size_t node_size = NODESIZE;
7205 MDB_page *mp = mc->mc_pg[mc->mc_top];
7206 MDB_page *ofp = NULL; /* overflow page */
7210 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7212 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7213 IS_LEAF(mp) ? "leaf" : "branch",
7214 IS_SUBP(mp) ? "sub-" : "",
7215 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7216 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7219 /* Move higher keys up one slot. */
7220 int ksize = mc->mc_db->md_pad, dif;
7221 char *ptr = LEAF2KEY(mp, indx, ksize);
7222 dif = NUMKEYS(mp) - indx;
7224 memmove(ptr+ksize, ptr, dif*ksize);
7225 /* insert new key */
7226 memcpy(ptr, key->mv_data, ksize);
7228 /* Just using these for counting */
7229 mp->mp_lower += sizeof(indx_t);
7230 mp->mp_upper -= ksize - sizeof(indx_t);
7234 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7236 node_size += key->mv_size;
7238 mdb_cassert(mc, key && data);
7239 if (F_ISSET(flags, F_BIGDATA)) {
7240 /* Data already on overflow page. */
7241 node_size += sizeof(pgno_t);
7242 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7243 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7245 /* Put data on overflow page. */
7246 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7247 data->mv_size, node_size+data->mv_size));
7248 node_size = EVEN(node_size + sizeof(pgno_t));
7249 if ((ssize_t)node_size > room)
7251 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7253 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7257 node_size += data->mv_size;
7260 node_size = EVEN(node_size);
7261 if ((ssize_t)node_size > room)
7265 /* Move higher pointers up one slot. */
7266 for (i = NUMKEYS(mp); i > indx; i--)
7267 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7269 /* Adjust free space offsets. */
7270 ofs = mp->mp_upper - node_size;
7271 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7272 mp->mp_ptrs[indx] = ofs;
7274 mp->mp_lower += sizeof(indx_t);
7276 /* Write the node data. */
7277 node = NODEPTR(mp, indx);
7278 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7279 node->mn_flags = flags;
7281 SETDSZ(node,data->mv_size);
7286 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7289 ndata = NODEDATA(node);
7291 if (F_ISSET(flags, F_BIGDATA))
7292 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7293 else if (F_ISSET(flags, MDB_RESERVE))
7294 data->mv_data = ndata;
7296 memcpy(ndata, data->mv_data, data->mv_size);
7298 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7299 ndata = METADATA(ofp);
7300 if (F_ISSET(flags, MDB_RESERVE))
7301 data->mv_data = ndata;
7303 memcpy(ndata, data->mv_data, data->mv_size);
7310 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7311 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7312 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7313 DPRINTF(("node size = %"Z"u", node_size));
7314 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7315 return MDB_PAGE_FULL;
7318 /** Delete the specified node from a page.
7319 * @param[in] mc Cursor pointing to the node to delete.
7320 * @param[in] ksize The size of a node. Only used if the page is
7321 * part of a #MDB_DUPFIXED database.
7324 mdb_node_del(MDB_cursor *mc, int ksize)
7326 MDB_page *mp = mc->mc_pg[mc->mc_top];
7327 indx_t indx = mc->mc_ki[mc->mc_top];
7329 indx_t i, j, numkeys, ptr;
7333 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7334 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7335 numkeys = NUMKEYS(mp);
7336 mdb_cassert(mc, indx < numkeys);
7339 int x = numkeys - 1 - indx;
7340 base = LEAF2KEY(mp, indx, ksize);
7342 memmove(base, base + ksize, x * ksize);
7343 mp->mp_lower -= sizeof(indx_t);
7344 mp->mp_upper += ksize - sizeof(indx_t);
7348 node = NODEPTR(mp, indx);
7349 sz = NODESIZE + node->mn_ksize;
7351 if (F_ISSET(node->mn_flags, F_BIGDATA))
7352 sz += sizeof(pgno_t);
7354 sz += NODEDSZ(node);
7358 ptr = mp->mp_ptrs[indx];
7359 for (i = j = 0; i < numkeys; i++) {
7361 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7362 if (mp->mp_ptrs[i] < ptr)
7363 mp->mp_ptrs[j] += sz;
7368 base = (char *)mp + mp->mp_upper + PAGEBASE;
7369 memmove(base + sz, base, ptr - mp->mp_upper);
7371 mp->mp_lower -= sizeof(indx_t);
7375 /** Compact the main page after deleting a node on a subpage.
7376 * @param[in] mp The main page to operate on.
7377 * @param[in] indx The index of the subpage on the main page.
7380 mdb_node_shrink(MDB_page *mp, indx_t indx)
7385 indx_t delta, nsize, len, ptr;
7388 node = NODEPTR(mp, indx);
7389 sp = (MDB_page *)NODEDATA(node);
7390 delta = SIZELEFT(sp);
7391 nsize = NODEDSZ(node) - delta;
7393 /* Prepare to shift upward, set len = length(subpage part to shift) */
7397 return; /* do not make the node uneven-sized */
7399 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7400 for (i = NUMKEYS(sp); --i >= 0; )
7401 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7404 sp->mp_upper = sp->mp_lower;
7405 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7406 SETDSZ(node, nsize);
7408 /* Shift <lower nodes...initial part of subpage> upward */
7409 base = (char *)mp + mp->mp_upper + PAGEBASE;
7410 memmove(base + delta, base, (char *)sp + len - base);
7412 ptr = mp->mp_ptrs[indx];
7413 for (i = NUMKEYS(mp); --i >= 0; ) {
7414 if (mp->mp_ptrs[i] <= ptr)
7415 mp->mp_ptrs[i] += delta;
7417 mp->mp_upper += delta;
7420 /** Initial setup of a sorted-dups cursor.
7421 * Sorted duplicates are implemented as a sub-database for the given key.
7422 * The duplicate data items are actually keys of the sub-database.
7423 * Operations on the duplicate data items are performed using a sub-cursor
7424 * initialized when the sub-database is first accessed. This function does
7425 * the preliminary setup of the sub-cursor, filling in the fields that
7426 * depend only on the parent DB.
7427 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7430 mdb_xcursor_init0(MDB_cursor *mc)
7432 MDB_xcursor *mx = mc->mc_xcursor;
7434 mx->mx_cursor.mc_xcursor = NULL;
7435 mx->mx_cursor.mc_txn = mc->mc_txn;
7436 mx->mx_cursor.mc_db = &mx->mx_db;
7437 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7438 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7439 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7440 mx->mx_cursor.mc_snum = 0;
7441 mx->mx_cursor.mc_top = 0;
7442 mx->mx_cursor.mc_flags = C_SUB;
7443 mx->mx_dbx.md_name.mv_size = 0;
7444 mx->mx_dbx.md_name.mv_data = NULL;
7445 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7446 mx->mx_dbx.md_dcmp = NULL;
7447 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7450 /** Final setup of a sorted-dups cursor.
7451 * Sets up the fields that depend on the data from the main cursor.
7452 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7453 * @param[in] node The data containing the #MDB_db record for the
7454 * sorted-dup database.
7457 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7459 MDB_xcursor *mx = mc->mc_xcursor;
7461 if (node->mn_flags & F_SUBDATA) {
7462 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7463 mx->mx_cursor.mc_pg[0] = 0;
7464 mx->mx_cursor.mc_snum = 0;
7465 mx->mx_cursor.mc_top = 0;
7466 mx->mx_cursor.mc_flags = C_SUB;
7468 MDB_page *fp = NODEDATA(node);
7469 mx->mx_db.md_pad = 0;
7470 mx->mx_db.md_flags = 0;
7471 mx->mx_db.md_depth = 1;
7472 mx->mx_db.md_branch_pages = 0;
7473 mx->mx_db.md_leaf_pages = 1;
7474 mx->mx_db.md_overflow_pages = 0;
7475 mx->mx_db.md_entries = NUMKEYS(fp);
7476 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7477 mx->mx_cursor.mc_snum = 1;
7478 mx->mx_cursor.mc_top = 0;
7479 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7480 mx->mx_cursor.mc_pg[0] = fp;
7481 mx->mx_cursor.mc_ki[0] = 0;
7482 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7483 mx->mx_db.md_flags = MDB_DUPFIXED;
7484 mx->mx_db.md_pad = fp->mp_pad;
7485 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7486 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7489 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7490 mx->mx_db.md_root));
7491 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7492 #if UINT_MAX < SIZE_MAX
7493 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7494 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7499 /** Fixup a sorted-dups cursor due to underlying update.
7500 * Sets up some fields that depend on the data from the main cursor.
7501 * Almost the same as init1, but skips initialization steps if the
7502 * xcursor had already been used.
7503 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7504 * @param[in] src_mx The xcursor of an up-to-date cursor.
7505 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7508 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7510 MDB_xcursor *mx = mc->mc_xcursor;
7513 mx->mx_cursor.mc_snum = 1;
7514 mx->mx_cursor.mc_top = 0;
7515 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7516 mx->mx_cursor.mc_ki[0] = 0;
7517 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7518 #if UINT_MAX < SIZE_MAX
7519 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7521 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7524 mx->mx_db = src_mx->mx_db;
7525 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7526 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7527 mx->mx_db.md_root));
7530 /** Initialize a cursor for a given transaction and database. */
7532 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7535 mc->mc_backup = NULL;
7538 mc->mc_db = &txn->mt_dbs[dbi];
7539 mc->mc_dbx = &txn->mt_dbxs[dbi];
7540 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7546 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7547 mdb_tassert(txn, mx != NULL);
7548 mc->mc_xcursor = mx;
7549 mdb_xcursor_init0(mc);
7551 mc->mc_xcursor = NULL;
7553 if (*mc->mc_dbflag & DB_STALE) {
7554 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7559 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7562 size_t size = sizeof(MDB_cursor);
7564 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7567 if (txn->mt_flags & MDB_TXN_BLOCKED)
7570 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7573 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7574 size += sizeof(MDB_xcursor);
7576 if ((mc = malloc(size)) != NULL) {
7577 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7578 if (txn->mt_cursors) {
7579 mc->mc_next = txn->mt_cursors[dbi];
7580 txn->mt_cursors[dbi] = mc;
7581 mc->mc_flags |= C_UNTRACK;
7593 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7595 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7598 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7601 if (txn->mt_flags & MDB_TXN_BLOCKED)
7604 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7608 /* Return the count of duplicate data items for the current key */
7610 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7614 if (mc == NULL || countp == NULL)
7617 if (mc->mc_xcursor == NULL)
7618 return MDB_INCOMPATIBLE;
7620 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7623 if (!(mc->mc_flags & C_INITIALIZED))
7626 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7627 return MDB_NOTFOUND;
7629 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7630 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7633 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7636 *countp = mc->mc_xcursor->mx_db.md_entries;
7642 mdb_cursor_close(MDB_cursor *mc)
7644 if (mc && !mc->mc_backup) {
7645 /* remove from txn, if tracked */
7646 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7647 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7648 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7650 *prev = mc->mc_next;
7657 mdb_cursor_txn(MDB_cursor *mc)
7659 if (!mc) return NULL;
7664 mdb_cursor_dbi(MDB_cursor *mc)
7669 /** Replace the key for a branch node with a new key.
7670 * @param[in] mc Cursor pointing to the node to operate on.
7671 * @param[in] key The new key to use.
7672 * @return 0 on success, non-zero on failure.
7675 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7681 int delta, ksize, oksize;
7682 indx_t ptr, i, numkeys, indx;
7685 indx = mc->mc_ki[mc->mc_top];
7686 mp = mc->mc_pg[mc->mc_top];
7687 node = NODEPTR(mp, indx);
7688 ptr = mp->mp_ptrs[indx];
7692 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7693 k2.mv_data = NODEKEY(node);
7694 k2.mv_size = node->mn_ksize;
7695 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7697 mdb_dkey(&k2, kbuf2),
7703 /* Sizes must be 2-byte aligned. */
7704 ksize = EVEN(key->mv_size);
7705 oksize = EVEN(node->mn_ksize);
7706 delta = ksize - oksize;
7708 /* Shift node contents if EVEN(key length) changed. */
7710 if (delta > 0 && SIZELEFT(mp) < delta) {
7712 /* not enough space left, do a delete and split */
7713 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7714 pgno = NODEPGNO(node);
7715 mdb_node_del(mc, 0);
7716 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7719 numkeys = NUMKEYS(mp);
7720 for (i = 0; i < numkeys; i++) {
7721 if (mp->mp_ptrs[i] <= ptr)
7722 mp->mp_ptrs[i] -= delta;
7725 base = (char *)mp + mp->mp_upper + PAGEBASE;
7726 len = ptr - mp->mp_upper + NODESIZE;
7727 memmove(base - delta, base, len);
7728 mp->mp_upper -= delta;
7730 node = NODEPTR(mp, indx);
7733 /* But even if no shift was needed, update ksize */
7734 if (node->mn_ksize != key->mv_size)
7735 node->mn_ksize = key->mv_size;
7738 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7744 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7746 /** Perform \b act while tracking temporary cursor \b mn */
7747 #define WITH_CURSOR_TRACKING(mn, act) do { \
7748 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
7749 if ((mn).mc_flags & C_SUB) { \
7750 dummy.mc_flags = C_INITIALIZED; \
7751 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
7756 tracked->mc_next = *tp; \
7759 *tp = tracked->mc_next; \
7762 /** Move a node from csrc to cdst.
7765 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
7772 unsigned short flags;
7776 /* Mark src and dst as dirty. */
7777 if ((rc = mdb_page_touch(csrc)) ||
7778 (rc = mdb_page_touch(cdst)))
7781 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7782 key.mv_size = csrc->mc_db->md_pad;
7783 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7785 data.mv_data = NULL;
7789 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7790 mdb_cassert(csrc, !((size_t)srcnode & 1));
7791 srcpg = NODEPGNO(srcnode);
7792 flags = srcnode->mn_flags;
7793 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7794 unsigned int snum = csrc->mc_snum;
7796 /* must find the lowest key below src */
7797 rc = mdb_page_search_lowest(csrc);
7800 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7801 key.mv_size = csrc->mc_db->md_pad;
7802 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7804 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7805 key.mv_size = NODEKSZ(s2);
7806 key.mv_data = NODEKEY(s2);
7808 csrc->mc_snum = snum--;
7809 csrc->mc_top = snum;
7811 key.mv_size = NODEKSZ(srcnode);
7812 key.mv_data = NODEKEY(srcnode);
7814 data.mv_size = NODEDSZ(srcnode);
7815 data.mv_data = NODEDATA(srcnode);
7817 mn.mc_xcursor = NULL;
7818 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7819 unsigned int snum = cdst->mc_snum;
7822 /* must find the lowest key below dst */
7823 mdb_cursor_copy(cdst, &mn);
7824 rc = mdb_page_search_lowest(&mn);
7827 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7828 bkey.mv_size = mn.mc_db->md_pad;
7829 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7831 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7832 bkey.mv_size = NODEKSZ(s2);
7833 bkey.mv_data = NODEKEY(s2);
7835 mn.mc_snum = snum--;
7838 rc = mdb_update_key(&mn, &bkey);
7843 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7844 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7845 csrc->mc_ki[csrc->mc_top],
7847 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7848 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7850 /* Add the node to the destination page.
7852 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7853 if (rc != MDB_SUCCESS)
7856 /* Delete the node from the source page.
7858 mdb_node_del(csrc, key.mv_size);
7861 /* Adjust other cursors pointing to mp */
7862 MDB_cursor *m2, *m3;
7863 MDB_dbi dbi = csrc->mc_dbi;
7864 MDB_page *mpd, *mps;
7866 mps = csrc->mc_pg[csrc->mc_top];
7867 /* If we're adding on the left, bump others up */
7869 mpd = cdst->mc_pg[csrc->mc_top];
7870 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7871 if (csrc->mc_flags & C_SUB)
7872 m3 = &m2->mc_xcursor->mx_cursor;
7875 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7878 m3->mc_pg[csrc->mc_top] == mpd &&
7879 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7880 m3->mc_ki[csrc->mc_top]++;
7883 m3->mc_pg[csrc->mc_top] == mps &&
7884 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7885 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7886 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7887 m3->mc_ki[csrc->mc_top-1]++;
7889 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7890 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7893 /* Adding on the right, bump others down */
7895 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7896 if (csrc->mc_flags & C_SUB)
7897 m3 = &m2->mc_xcursor->mx_cursor;
7900 if (m3 == csrc) continue;
7901 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7903 if (m3->mc_pg[csrc->mc_top] == mps) {
7904 if (!m3->mc_ki[csrc->mc_top]) {
7905 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7906 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7907 m3->mc_ki[csrc->mc_top-1]--;
7909 m3->mc_ki[csrc->mc_top]--;
7911 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7912 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7918 /* Update the parent separators.
7920 if (csrc->mc_ki[csrc->mc_top] == 0) {
7921 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7922 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7923 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7925 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7926 key.mv_size = NODEKSZ(srcnode);
7927 key.mv_data = NODEKEY(srcnode);
7929 DPRINTF(("update separator for source page %"Z"u to [%s]",
7930 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7931 mdb_cursor_copy(csrc, &mn);
7934 /* We want mdb_rebalance to find mn when doing fixups */
7935 WITH_CURSOR_TRACKING(mn,
7936 rc = mdb_update_key(&mn, &key));
7940 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7942 indx_t ix = csrc->mc_ki[csrc->mc_top];
7943 nullkey.mv_size = 0;
7944 csrc->mc_ki[csrc->mc_top] = 0;
7945 rc = mdb_update_key(csrc, &nullkey);
7946 csrc->mc_ki[csrc->mc_top] = ix;
7947 mdb_cassert(csrc, rc == MDB_SUCCESS);
7951 if (cdst->mc_ki[cdst->mc_top] == 0) {
7952 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7953 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7954 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7956 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7957 key.mv_size = NODEKSZ(srcnode);
7958 key.mv_data = NODEKEY(srcnode);
7960 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7961 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7962 mdb_cursor_copy(cdst, &mn);
7965 /* We want mdb_rebalance to find mn when doing fixups */
7966 WITH_CURSOR_TRACKING(mn,
7967 rc = mdb_update_key(&mn, &key));
7971 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7973 indx_t ix = cdst->mc_ki[cdst->mc_top];
7974 nullkey.mv_size = 0;
7975 cdst->mc_ki[cdst->mc_top] = 0;
7976 rc = mdb_update_key(cdst, &nullkey);
7977 cdst->mc_ki[cdst->mc_top] = ix;
7978 mdb_cassert(cdst, rc == MDB_SUCCESS);
7985 /** Merge one page into another.
7986 * The nodes from the page pointed to by \b csrc will
7987 * be copied to the page pointed to by \b cdst and then
7988 * the \b csrc page will be freed.
7989 * @param[in] csrc Cursor pointing to the source page.
7990 * @param[in] cdst Cursor pointing to the destination page.
7991 * @return 0 on success, non-zero on failure.
7994 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7996 MDB_page *psrc, *pdst;
8003 psrc = csrc->mc_pg[csrc->mc_top];
8004 pdst = cdst->mc_pg[cdst->mc_top];
8006 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
8008 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8009 mdb_cassert(csrc, cdst->mc_snum > 1);
8011 /* Mark dst as dirty. */
8012 if ((rc = mdb_page_touch(cdst)))
8015 /* get dst page again now that we've touched it. */
8016 pdst = cdst->mc_pg[cdst->mc_top];
8018 /* Move all nodes from src to dst.
8020 j = nkeys = NUMKEYS(pdst);
8021 if (IS_LEAF2(psrc)) {
8022 key.mv_size = csrc->mc_db->md_pad;
8023 key.mv_data = METADATA(psrc);
8024 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8025 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8026 if (rc != MDB_SUCCESS)
8028 key.mv_data = (char *)key.mv_data + key.mv_size;
8031 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8032 srcnode = NODEPTR(psrc, i);
8033 if (i == 0 && IS_BRANCH(psrc)) {
8036 mdb_cursor_copy(csrc, &mn);
8037 mn.mc_xcursor = NULL;
8038 /* must find the lowest key below src */
8039 rc = mdb_page_search_lowest(&mn);
8042 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8043 key.mv_size = mn.mc_db->md_pad;
8044 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8046 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8047 key.mv_size = NODEKSZ(s2);
8048 key.mv_data = NODEKEY(s2);
8051 key.mv_size = srcnode->mn_ksize;
8052 key.mv_data = NODEKEY(srcnode);
8055 data.mv_size = NODEDSZ(srcnode);
8056 data.mv_data = NODEDATA(srcnode);
8057 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8058 if (rc != MDB_SUCCESS)
8063 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
8064 pdst->mp_pgno, NUMKEYS(pdst),
8065 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8067 /* Unlink the src page from parent and add to free list.
8070 mdb_node_del(csrc, 0);
8071 if (csrc->mc_ki[csrc->mc_top] == 0) {
8073 rc = mdb_update_key(csrc, &key);
8081 psrc = csrc->mc_pg[csrc->mc_top];
8082 /* If not operating on FreeDB, allow this page to be reused
8083 * in this txn. Otherwise just add to free list.
8085 rc = mdb_page_loose(csrc, psrc);
8089 csrc->mc_db->md_leaf_pages--;
8091 csrc->mc_db->md_branch_pages--;
8093 /* Adjust other cursors pointing to mp */
8094 MDB_cursor *m2, *m3;
8095 MDB_dbi dbi = csrc->mc_dbi;
8096 unsigned int top = csrc->mc_top;
8098 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8099 if (csrc->mc_flags & C_SUB)
8100 m3 = &m2->mc_xcursor->mx_cursor;
8103 if (m3 == csrc) continue;
8104 if (m3->mc_snum < csrc->mc_snum) continue;
8105 if (m3->mc_pg[top] == psrc) {
8106 m3->mc_pg[top] = pdst;
8107 m3->mc_ki[top] += nkeys;
8108 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8109 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8110 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8113 if (XCURSOR_INITED(m3) && IS_LEAF(psrc))
8114 XCURSOR_REFRESH(m3, m3->mc_pg[top], m3->mc_ki[top]);
8118 unsigned int snum = cdst->mc_snum;
8119 uint16_t depth = cdst->mc_db->md_depth;
8120 mdb_cursor_pop(cdst);
8121 rc = mdb_rebalance(cdst);
8122 /* Did the tree height change? */
8123 if (depth != cdst->mc_db->md_depth)
8124 snum += cdst->mc_db->md_depth - depth;
8125 cdst->mc_snum = snum;
8126 cdst->mc_top = snum-1;
8131 /** Copy the contents of a cursor.
8132 * @param[in] csrc The cursor to copy from.
8133 * @param[out] cdst The cursor to copy to.
8136 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8140 cdst->mc_txn = csrc->mc_txn;
8141 cdst->mc_dbi = csrc->mc_dbi;
8142 cdst->mc_db = csrc->mc_db;
8143 cdst->mc_dbx = csrc->mc_dbx;
8144 cdst->mc_snum = csrc->mc_snum;
8145 cdst->mc_top = csrc->mc_top;
8146 cdst->mc_flags = csrc->mc_flags;
8148 for (i=0; i<csrc->mc_snum; i++) {
8149 cdst->mc_pg[i] = csrc->mc_pg[i];
8150 cdst->mc_ki[i] = csrc->mc_ki[i];
8154 /** Rebalance the tree after a delete operation.
8155 * @param[in] mc Cursor pointing to the page where rebalancing
8157 * @return 0 on success, non-zero on failure.
8160 mdb_rebalance(MDB_cursor *mc)
8164 unsigned int ptop, minkeys, thresh;
8168 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8173 thresh = FILL_THRESHOLD;
8175 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
8176 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8177 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8178 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8180 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8181 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8182 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
8183 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8187 if (mc->mc_snum < 2) {
8188 MDB_page *mp = mc->mc_pg[0];
8190 DPUTS("Can't rebalance a subpage, ignoring");
8193 if (NUMKEYS(mp) == 0) {
8194 DPUTS("tree is completely empty");
8195 mc->mc_db->md_root = P_INVALID;
8196 mc->mc_db->md_depth = 0;
8197 mc->mc_db->md_leaf_pages = 0;
8198 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8201 /* Adjust cursors pointing to mp */
8204 mc->mc_flags &= ~C_INITIALIZED;
8206 MDB_cursor *m2, *m3;
8207 MDB_dbi dbi = mc->mc_dbi;
8209 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8210 if (mc->mc_flags & C_SUB)
8211 m3 = &m2->mc_xcursor->mx_cursor;
8214 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8216 if (m3->mc_pg[0] == mp) {
8219 m3->mc_flags &= ~C_INITIALIZED;
8223 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8225 DPUTS("collapsing root page!");
8226 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8229 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8230 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
8233 mc->mc_db->md_depth--;
8234 mc->mc_db->md_branch_pages--;
8235 mc->mc_ki[0] = mc->mc_ki[1];
8236 for (i = 1; i<mc->mc_db->md_depth; i++) {
8237 mc->mc_pg[i] = mc->mc_pg[i+1];
8238 mc->mc_ki[i] = mc->mc_ki[i+1];
8241 /* Adjust other cursors pointing to mp */
8242 MDB_cursor *m2, *m3;
8243 MDB_dbi dbi = mc->mc_dbi;
8245 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8246 if (mc->mc_flags & C_SUB)
8247 m3 = &m2->mc_xcursor->mx_cursor;
8250 if (m3 == mc) continue;
8251 if (!(m3->mc_flags & C_INITIALIZED))
8253 if (m3->mc_pg[0] == mp) {
8254 for (i=0; i<mc->mc_db->md_depth; i++) {
8255 m3->mc_pg[i] = m3->mc_pg[i+1];
8256 m3->mc_ki[i] = m3->mc_ki[i+1];
8264 DPUTS("root page doesn't need rebalancing");
8268 /* The parent (branch page) must have at least 2 pointers,
8269 * otherwise the tree is invalid.
8271 ptop = mc->mc_top-1;
8272 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8274 /* Leaf page fill factor is below the threshold.
8275 * Try to move keys from left or right neighbor, or
8276 * merge with a neighbor page.
8281 mdb_cursor_copy(mc, &mn);
8282 mn.mc_xcursor = NULL;
8284 oldki = mc->mc_ki[mc->mc_top];
8285 if (mc->mc_ki[ptop] == 0) {
8286 /* We're the leftmost leaf in our parent.
8288 DPUTS("reading right neighbor");
8290 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8291 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8294 mn.mc_ki[mn.mc_top] = 0;
8295 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8298 /* There is at least one neighbor to the left.
8300 DPUTS("reading left neighbor");
8302 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8303 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8306 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8307 mc->mc_ki[mc->mc_top] = 0;
8311 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8312 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8313 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8315 /* If the neighbor page is above threshold and has enough keys,
8316 * move one key from it. Otherwise we should try to merge them.
8317 * (A branch page must never have less than 2 keys.)
8319 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8320 rc = mdb_node_move(&mn, mc, fromleft);
8322 /* if we inserted on left, bump position up */
8327 rc = mdb_page_merge(&mn, mc);
8329 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8330 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8331 /* We want mdb_rebalance to find mn when doing fixups */
8332 WITH_CURSOR_TRACKING(mn,
8333 rc = mdb_page_merge(mc, &mn));
8334 mdb_cursor_copy(&mn, mc);
8336 mc->mc_flags &= ~C_EOF;
8338 mc->mc_ki[mc->mc_top] = oldki;
8342 /** Complete a delete operation started by #mdb_cursor_del(). */
8344 mdb_cursor_del0(MDB_cursor *mc)
8350 MDB_cursor *m2, *m3;
8351 MDB_dbi dbi = mc->mc_dbi;
8353 ki = mc->mc_ki[mc->mc_top];
8354 mp = mc->mc_pg[mc->mc_top];
8355 mdb_node_del(mc, mc->mc_db->md_pad);
8356 mc->mc_db->md_entries--;
8358 /* Adjust other cursors pointing to mp */
8359 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8360 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8361 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8363 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8365 if (m3->mc_pg[mc->mc_top] == mp) {
8366 if (m3->mc_ki[mc->mc_top] == ki) {
8367 m3->mc_flags |= C_DEL;
8368 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8369 /* Sub-cursor referred into dataset which is gone */
8370 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
8373 } else if (m3->mc_ki[mc->mc_top] > ki) {
8374 m3->mc_ki[mc->mc_top]--;
8376 if (XCURSOR_INITED(m3))
8377 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8381 rc = mdb_rebalance(mc);
8383 if (rc == MDB_SUCCESS) {
8384 /* DB is totally empty now, just bail out.
8385 * Other cursors adjustments were already done
8386 * by mdb_rebalance and aren't needed here.
8391 mp = mc->mc_pg[mc->mc_top];
8392 nkeys = NUMKEYS(mp);
8394 /* Adjust other cursors pointing to mp */
8395 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8396 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8397 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8399 if (m3->mc_snum < mc->mc_snum)
8401 if (m3->mc_pg[mc->mc_top] == mp) {
8402 /* if m3 points past last node in page, find next sibling */
8403 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
8404 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8405 rc = mdb_cursor_sibling(m3, 1);
8406 if (rc == MDB_NOTFOUND) {
8407 m3->mc_flags |= C_EOF;
8412 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8413 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
8414 /* If this node is a fake page, it needs to be reinited
8415 * because its data has moved. But just reset mc_pg[0]
8416 * if the xcursor is already live.
8418 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) {
8419 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)
8420 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8422 mdb_xcursor_init1(m3, node);
8428 mc->mc_flags |= C_DEL;
8432 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8437 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8438 MDB_val *key, MDB_val *data)
8440 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8443 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8444 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8446 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8447 /* must ignore any data */
8451 return mdb_del0(txn, dbi, key, data, 0);
8455 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8456 MDB_val *key, MDB_val *data, unsigned flags)
8461 MDB_val rdata, *xdata;
8465 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8467 mdb_cursor_init(&mc, txn, dbi, &mx);
8476 flags |= MDB_NODUPDATA;
8478 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8480 /* let mdb_page_split know about this cursor if needed:
8481 * delete will trigger a rebalance; if it needs to move
8482 * a node from one page to another, it will have to
8483 * update the parent's separator key(s). If the new sepkey
8484 * is larger than the current one, the parent page may
8485 * run out of space, triggering a split. We need this
8486 * cursor to be consistent until the end of the rebalance.
8488 mc.mc_flags |= C_UNTRACK;
8489 mc.mc_next = txn->mt_cursors[dbi];
8490 txn->mt_cursors[dbi] = &mc;
8491 rc = mdb_cursor_del(&mc, flags);
8492 txn->mt_cursors[dbi] = mc.mc_next;
8497 /** Split a page and insert a new node.
8498 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8499 * The cursor will be updated to point to the actual page and index where
8500 * the node got inserted after the split.
8501 * @param[in] newkey The key for the newly inserted node.
8502 * @param[in] newdata The data for the newly inserted node.
8503 * @param[in] newpgno The page number, if the new node is a branch node.
8504 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8505 * @return 0 on success, non-zero on failure.
8508 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8509 unsigned int nflags)
8512 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8515 int i, j, split_indx, nkeys, pmax;
8516 MDB_env *env = mc->mc_txn->mt_env;
8518 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8519 MDB_page *copy = NULL;
8520 MDB_page *mp, *rp, *pp;
8525 mp = mc->mc_pg[mc->mc_top];
8526 newindx = mc->mc_ki[mc->mc_top];
8527 nkeys = NUMKEYS(mp);
8529 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8530 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8531 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8533 /* Create a right sibling. */
8534 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8536 rp->mp_pad = mp->mp_pad;
8537 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8539 /* Usually when splitting the root page, the cursor
8540 * height is 1. But when called from mdb_update_key,
8541 * the cursor height may be greater because it walks
8542 * up the stack while finding the branch slot to update.
8544 if (mc->mc_top < 1) {
8545 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8547 /* shift current top to make room for new parent */
8548 for (i=mc->mc_snum; i>0; i--) {
8549 mc->mc_pg[i] = mc->mc_pg[i-1];
8550 mc->mc_ki[i] = mc->mc_ki[i-1];
8554 mc->mc_db->md_root = pp->mp_pgno;
8555 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8556 new_root = mc->mc_db->md_depth++;
8558 /* Add left (implicit) pointer. */
8559 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8560 /* undo the pre-push */
8561 mc->mc_pg[0] = mc->mc_pg[1];
8562 mc->mc_ki[0] = mc->mc_ki[1];
8563 mc->mc_db->md_root = mp->mp_pgno;
8564 mc->mc_db->md_depth--;
8571 ptop = mc->mc_top-1;
8572 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8575 mdb_cursor_copy(mc, &mn);
8576 mn.mc_xcursor = NULL;
8577 mn.mc_pg[mn.mc_top] = rp;
8578 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8580 if (nflags & MDB_APPEND) {
8581 mn.mc_ki[mn.mc_top] = 0;
8583 split_indx = newindx;
8587 split_indx = (nkeys+1) / 2;
8592 unsigned int lsize, rsize, ksize;
8593 /* Move half of the keys to the right sibling */
8594 x = mc->mc_ki[mc->mc_top] - split_indx;
8595 ksize = mc->mc_db->md_pad;
8596 split = LEAF2KEY(mp, split_indx, ksize);
8597 rsize = (nkeys - split_indx) * ksize;
8598 lsize = (nkeys - split_indx) * sizeof(indx_t);
8599 mp->mp_lower -= lsize;
8600 rp->mp_lower += lsize;
8601 mp->mp_upper += rsize - lsize;
8602 rp->mp_upper -= rsize - lsize;
8603 sepkey.mv_size = ksize;
8604 if (newindx == split_indx) {
8605 sepkey.mv_data = newkey->mv_data;
8607 sepkey.mv_data = split;
8610 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8611 memcpy(rp->mp_ptrs, split, rsize);
8612 sepkey.mv_data = rp->mp_ptrs;
8613 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8614 memcpy(ins, newkey->mv_data, ksize);
8615 mp->mp_lower += sizeof(indx_t);
8616 mp->mp_upper -= ksize - sizeof(indx_t);
8619 memcpy(rp->mp_ptrs, split, x * ksize);
8620 ins = LEAF2KEY(rp, x, ksize);
8621 memcpy(ins, newkey->mv_data, ksize);
8622 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8623 rp->mp_lower += sizeof(indx_t);
8624 rp->mp_upper -= ksize - sizeof(indx_t);
8625 mc->mc_ki[mc->mc_top] = x;
8628 int psize, nsize, k;
8629 /* Maximum free space in an empty page */
8630 pmax = env->me_psize - PAGEHDRSZ;
8632 nsize = mdb_leaf_size(env, newkey, newdata);
8634 nsize = mdb_branch_size(env, newkey);
8635 nsize = EVEN(nsize);
8637 /* grab a page to hold a temporary copy */
8638 copy = mdb_page_malloc(mc->mc_txn, 1);
8643 copy->mp_pgno = mp->mp_pgno;
8644 copy->mp_flags = mp->mp_flags;
8645 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8646 copy->mp_upper = env->me_psize - PAGEBASE;
8648 /* prepare to insert */
8649 for (i=0, j=0; i<nkeys; i++) {
8651 copy->mp_ptrs[j++] = 0;
8653 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8656 /* When items are relatively large the split point needs
8657 * to be checked, because being off-by-one will make the
8658 * difference between success or failure in mdb_node_add.
8660 * It's also relevant if a page happens to be laid out
8661 * such that one half of its nodes are all "small" and
8662 * the other half of its nodes are "large." If the new
8663 * item is also "large" and falls on the half with
8664 * "large" nodes, it also may not fit.
8666 * As a final tweak, if the new item goes on the last
8667 * spot on the page (and thus, onto the new page), bias
8668 * the split so the new page is emptier than the old page.
8669 * This yields better packing during sequential inserts.
8671 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8672 /* Find split point */
8674 if (newindx <= split_indx || newindx >= nkeys) {
8676 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8681 for (; i!=k; i+=j) {
8686 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8687 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8689 if (F_ISSET(node->mn_flags, F_BIGDATA))
8690 psize += sizeof(pgno_t);
8692 psize += NODEDSZ(node);
8694 psize = EVEN(psize);
8696 if (psize > pmax || i == k-j) {
8697 split_indx = i + (j<0);
8702 if (split_indx == newindx) {
8703 sepkey.mv_size = newkey->mv_size;
8704 sepkey.mv_data = newkey->mv_data;
8706 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8707 sepkey.mv_size = node->mn_ksize;
8708 sepkey.mv_data = NODEKEY(node);
8713 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8715 /* Copy separator key to the parent.
8717 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8718 int snum = mc->mc_snum;
8722 /* We want other splits to find mn when doing fixups */
8723 WITH_CURSOR_TRACKING(mn,
8724 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
8729 if (mc->mc_snum > snum) {
8732 /* Right page might now have changed parent.
8733 * Check if left page also changed parent.
8735 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8736 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8737 for (i=0; i<ptop; i++) {
8738 mc->mc_pg[i] = mn.mc_pg[i];
8739 mc->mc_ki[i] = mn.mc_ki[i];
8741 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8742 if (mn.mc_ki[ptop]) {
8743 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8745 /* find right page's left sibling */
8746 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8747 mdb_cursor_sibling(mc, 0);
8752 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8755 if (rc != MDB_SUCCESS) {
8758 if (nflags & MDB_APPEND) {
8759 mc->mc_pg[mc->mc_top] = rp;
8760 mc->mc_ki[mc->mc_top] = 0;
8761 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8764 for (i=0; i<mc->mc_top; i++)
8765 mc->mc_ki[i] = mn.mc_ki[i];
8766 } else if (!IS_LEAF2(mp)) {
8768 mc->mc_pg[mc->mc_top] = rp;
8773 rkey.mv_data = newkey->mv_data;
8774 rkey.mv_size = newkey->mv_size;
8780 /* Update index for the new key. */
8781 mc->mc_ki[mc->mc_top] = j;
8783 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8784 rkey.mv_data = NODEKEY(node);
8785 rkey.mv_size = node->mn_ksize;
8787 xdata.mv_data = NODEDATA(node);
8788 xdata.mv_size = NODEDSZ(node);
8791 pgno = NODEPGNO(node);
8792 flags = node->mn_flags;
8795 if (!IS_LEAF(mp) && j == 0) {
8796 /* First branch index doesn't need key data. */
8800 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8806 mc->mc_pg[mc->mc_top] = copy;
8811 } while (i != split_indx);
8813 nkeys = NUMKEYS(copy);
8814 for (i=0; i<nkeys; i++)
8815 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8816 mp->mp_lower = copy->mp_lower;
8817 mp->mp_upper = copy->mp_upper;
8818 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8819 env->me_psize - copy->mp_upper - PAGEBASE);
8821 /* reset back to original page */
8822 if (newindx < split_indx) {
8823 mc->mc_pg[mc->mc_top] = mp;
8825 mc->mc_pg[mc->mc_top] = rp;
8827 /* Make sure mc_ki is still valid.
8829 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8830 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8831 for (i=0; i<=ptop; i++) {
8832 mc->mc_pg[i] = mn.mc_pg[i];
8833 mc->mc_ki[i] = mn.mc_ki[i];
8837 if (nflags & MDB_RESERVE) {
8838 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8839 if (!(node->mn_flags & F_BIGDATA))
8840 newdata->mv_data = NODEDATA(node);
8843 if (newindx >= split_indx) {
8844 mc->mc_pg[mc->mc_top] = rp;
8846 /* Make sure mc_ki is still valid.
8848 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8849 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8850 for (i=0; i<=ptop; i++) {
8851 mc->mc_pg[i] = mn.mc_pg[i];
8852 mc->mc_ki[i] = mn.mc_ki[i];
8859 /* Adjust other cursors pointing to mp */
8860 MDB_cursor *m2, *m3;
8861 MDB_dbi dbi = mc->mc_dbi;
8862 nkeys = NUMKEYS(mp);
8864 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8865 if (mc->mc_flags & C_SUB)
8866 m3 = &m2->mc_xcursor->mx_cursor;
8871 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8875 /* sub cursors may be on different DB */
8876 if (m3->mc_pg[0] != mp)
8879 for (k=new_root; k>=0; k--) {
8880 m3->mc_ki[k+1] = m3->mc_ki[k];
8881 m3->mc_pg[k+1] = m3->mc_pg[k];
8883 if (m3->mc_ki[0] >= nkeys) {
8888 m3->mc_pg[0] = mc->mc_pg[0];
8892 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8893 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8894 m3->mc_ki[mc->mc_top]++;
8895 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8896 m3->mc_pg[mc->mc_top] = rp;
8897 m3->mc_ki[mc->mc_top] -= nkeys;
8898 for (i=0; i<mc->mc_top; i++) {
8899 m3->mc_ki[i] = mn.mc_ki[i];
8900 m3->mc_pg[i] = mn.mc_pg[i];
8903 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8904 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8907 if (XCURSOR_INITED(m3) && IS_LEAF(mp))
8908 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8911 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8914 if (copy) /* tmp page */
8915 mdb_page_free(env, copy);
8917 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8922 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8923 MDB_val *key, MDB_val *data, unsigned int flags)
8929 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8932 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8935 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8936 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8938 mdb_cursor_init(&mc, txn, dbi, &mx);
8939 mc.mc_next = txn->mt_cursors[dbi];
8940 txn->mt_cursors[dbi] = &mc;
8941 rc = mdb_cursor_put(&mc, key, data, flags);
8942 txn->mt_cursors[dbi] = mc.mc_next;
8947 #define MDB_WBUF (1024*1024)
8949 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
8951 /** State needed for a double-buffering compacting copy. */
8952 typedef struct mdb_copy {
8955 pthread_mutex_t mc_mutex;
8956 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
8961 pgno_t mc_next_pgno;
8963 int mc_toggle; /**< Buffer number in provider */
8964 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
8965 /** Error code. Never cleared if set. Both threads can set nonzero
8966 * to fail the copy. Not mutex-protected, LMDB expects atomic int.
8968 volatile int mc_error;
8971 /** Dedicated writer thread for compacting copy. */
8972 static THREAD_RET ESECT CALL_CONV
8973 mdb_env_copythr(void *arg)
8977 int toggle = 0, wsize, rc;
8980 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8983 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8986 pthread_mutex_lock(&my->mc_mutex);
8989 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8990 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
8992 wsize = my->mc_wlen[toggle];
8993 ptr = my->mc_wbuf[toggle];
8996 while (wsize > 0 && !my->mc_error) {
8997 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9001 } else if (len > 0) {
9014 /* If there's an overflow page tail, write it too */
9015 if (my->mc_olen[toggle]) {
9016 wsize = my->mc_olen[toggle];
9017 ptr = my->mc_over[toggle];
9018 my->mc_olen[toggle] = 0;
9021 my->mc_wlen[toggle] = 0;
9023 /* Return the empty buffer to provider */
9025 pthread_cond_signal(&my->mc_cond);
9027 pthread_mutex_unlock(&my->mc_mutex);
9028 return (THREAD_RET)0;
9032 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9034 * @param[in] my control structure.
9035 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9038 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9040 pthread_mutex_lock(&my->mc_mutex);
9041 my->mc_new += adjust;
9042 pthread_cond_signal(&my->mc_cond);
9043 while (my->mc_new & 2) /* both buffers in use */
9044 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9045 pthread_mutex_unlock(&my->mc_mutex);
9047 my->mc_toggle ^= (adjust & 1);
9048 /* Both threads reset mc_wlen, to be safe from threading errors */
9049 my->mc_wlen[my->mc_toggle] = 0;
9050 return my->mc_error;
9053 /** Depth-first tree traversal for compacting copy.
9054 * @param[in] my control structure.
9055 * @param[in,out] pg database root.
9056 * @param[in] flags includes #F_DUPDATA if it is a sorted-duplicate sub-DB.
9059 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9061 MDB_cursor mc = {0};
9063 MDB_page *mo, *mp, *leaf;
9068 /* Empty DB, nothing to do */
9069 if (*pg == P_INVALID)
9073 mc.mc_txn = my->mc_txn;
9075 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9078 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9082 /* Make cursor pages writable */
9083 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9087 for (i=0; i<mc.mc_top; i++) {
9088 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9089 mc.mc_pg[i] = (MDB_page *)ptr;
9090 ptr += my->mc_env->me_psize;
9093 /* This is writable space for a leaf page. Usually not needed. */
9094 leaf = (MDB_page *)ptr;
9096 toggle = my->mc_toggle;
9097 while (mc.mc_snum > 0) {
9099 mp = mc.mc_pg[mc.mc_top];
9103 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9104 for (i=0; i<n; i++) {
9105 ni = NODEPTR(mp, i);
9106 if (ni->mn_flags & F_BIGDATA) {
9110 /* Need writable leaf */
9112 mc.mc_pg[mc.mc_top] = leaf;
9113 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9115 ni = NODEPTR(mp, i);
9118 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9119 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
9120 rc = mdb_page_get(&mc, pg, &omp, NULL);
9123 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9124 rc = mdb_env_cthr_toggle(my, 1);
9127 toggle = my->mc_toggle;
9129 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9130 memcpy(mo, omp, my->mc_env->me_psize);
9131 mo->mp_pgno = my->mc_next_pgno;
9132 my->mc_next_pgno += omp->mp_pages;
9133 my->mc_wlen[toggle] += my->mc_env->me_psize;
9134 if (omp->mp_pages > 1) {
9135 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9136 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9137 rc = mdb_env_cthr_toggle(my, 1);
9140 toggle = my->mc_toggle;
9142 } else if (ni->mn_flags & F_SUBDATA) {
9145 /* Need writable leaf */
9147 mc.mc_pg[mc.mc_top] = leaf;
9148 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9150 ni = NODEPTR(mp, i);
9153 memcpy(&db, NODEDATA(ni), sizeof(db));
9154 my->mc_toggle = toggle;
9155 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9158 toggle = my->mc_toggle;
9159 memcpy(NODEDATA(ni), &db, sizeof(db));
9164 mc.mc_ki[mc.mc_top]++;
9165 if (mc.mc_ki[mc.mc_top] < n) {
9168 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9170 rc = mdb_page_get(&mc, pg, &mp, NULL);
9175 mc.mc_ki[mc.mc_top] = 0;
9176 if (IS_BRANCH(mp)) {
9177 /* Whenever we advance to a sibling branch page,
9178 * we must proceed all the way down to its first leaf.
9180 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9183 mc.mc_pg[mc.mc_top] = mp;
9187 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9188 rc = mdb_env_cthr_toggle(my, 1);
9191 toggle = my->mc_toggle;
9193 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9194 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9195 mo->mp_pgno = my->mc_next_pgno++;
9196 my->mc_wlen[toggle] += my->mc_env->me_psize;
9198 /* Update parent if there is one */
9199 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9200 SETPGNO(ni, mo->mp_pgno);
9201 mdb_cursor_pop(&mc);
9203 /* Otherwise we're done */
9213 /** Copy environment with compaction. */
9215 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9220 MDB_txn *txn = NULL;
9222 pgno_t root, new_root;
9223 int rc = MDB_SUCCESS;
9226 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
9227 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
9231 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9232 if (my.mc_wbuf[0] == NULL) {
9233 /* _aligned_malloc() sets errno, but we use Windows error codes */
9234 rc = ERROR_NOT_ENOUGH_MEMORY;
9238 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
9240 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
9242 #ifdef HAVE_MEMALIGN
9243 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9244 if (my.mc_wbuf[0] == NULL) {
9251 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
9257 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9258 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9259 my.mc_next_pgno = NUM_METAS;
9262 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
9266 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9270 mp = (MDB_page *)my.mc_wbuf[0];
9271 memset(mp, 0, NUM_METAS * env->me_psize);
9273 mp->mp_flags = P_META;
9274 mm = (MDB_meta *)METADATA(mp);
9275 mdb_env_init_meta0(env, mm);
9276 mm->mm_address = env->me_metas[0]->mm_address;
9278 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9280 mp->mp_flags = P_META;
9281 *(MDB_meta *)METADATA(mp) = *mm;
9282 mm = (MDB_meta *)METADATA(mp);
9284 /* Set metapage 1 with current main DB */
9285 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
9286 if (root != P_INVALID) {
9287 /* Count free pages + freeDB pages. Subtract from last_pg
9288 * to find the new last_pg, which also becomes the new root.
9290 MDB_ID freecount = 0;
9293 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9294 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9295 freecount += *(MDB_ID *)data.mv_data;
9296 if (rc != MDB_NOTFOUND)
9298 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9299 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9300 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9302 new_root = txn->mt_next_pgno - 1 - freecount;
9303 mm->mm_last_pg = new_root;
9304 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9305 mm->mm_dbs[MAIN_DBI].md_root = new_root;
9307 /* When the DB is empty, handle it specially to
9308 * fix any breakage like page leaks from ITS#8174.
9310 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
9312 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
9313 mm->mm_txnid = 1; /* use metapage 1 */
9316 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9318 rc = mdb_env_cwalk(&my, &root, 0);
9319 if (rc == MDB_SUCCESS && root != new_root) {
9320 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
9326 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
9327 rc = THREAD_FINISH(thr);
9332 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
9333 if (my.mc_cond) CloseHandle(my.mc_cond);
9334 if (my.mc_mutex) CloseHandle(my.mc_mutex);
9336 free(my.mc_wbuf[0]);
9337 pthread_cond_destroy(&my.mc_cond);
9339 pthread_mutex_destroy(&my.mc_mutex);
9341 return rc ? rc : my.mc_error;
9344 /** Copy environment as-is. */
9346 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9348 MDB_txn *txn = NULL;
9349 mdb_mutexref_t wmutex = NULL;
9355 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9359 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9362 /* Do the lock/unlock of the reader mutex before starting the
9363 * write txn. Otherwise other read txns could block writers.
9365 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9370 /* We must start the actual read txn after blocking writers */
9371 mdb_txn_end(txn, MDB_END_RESET_TMP);
9373 /* Temporarily block writers until we snapshot the meta pages */
9374 wmutex = env->me_wmutex;
9375 if (LOCK_MUTEX(rc, env, wmutex))
9378 rc = mdb_txn_renew0(txn);
9380 UNLOCK_MUTEX(wmutex);
9385 wsize = env->me_psize * NUM_METAS;
9389 DO_WRITE(rc, fd, ptr, w2, len);
9393 } else if (len > 0) {
9399 /* Non-blocking or async handles are not supported */
9405 UNLOCK_MUTEX(wmutex);
9410 w3 = txn->mt_next_pgno * env->me_psize;
9413 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9420 if (wsize > MAX_WRITE)
9424 DO_WRITE(rc, fd, ptr, w2, len);
9428 } else if (len > 0) {
9445 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9447 if (flags & MDB_CP_COMPACT)
9448 return mdb_env_copyfd1(env, fd);
9450 return mdb_env_copyfd0(env, fd);
9454 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9456 return mdb_env_copyfd2(env, fd, 0);
9460 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9464 HANDLE newfd = INVALID_HANDLE_VALUE;
9466 rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname);
9467 if (rc == MDB_SUCCESS) {
9468 rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd);
9469 mdb_fname_destroy(fname);
9471 if (rc == MDB_SUCCESS) {
9472 rc = mdb_env_copyfd2(env, newfd, flags);
9473 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9480 mdb_env_copy(MDB_env *env, const char *path)
9482 return mdb_env_copy2(env, path, 0);
9486 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9488 if (flag & ~CHANGEABLE)
9491 env->me_flags |= flag;
9493 env->me_flags &= ~flag;
9498 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9503 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9508 mdb_env_set_userctx(MDB_env *env, void *ctx)
9512 env->me_userctx = ctx;
9517 mdb_env_get_userctx(MDB_env *env)
9519 return env ? env->me_userctx : NULL;
9523 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9528 env->me_assert_func = func;
9534 mdb_env_get_path(MDB_env *env, const char **arg)
9539 *arg = env->me_path;
9544 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9553 /** Common code for #mdb_stat() and #mdb_env_stat().
9554 * @param[in] env the environment to operate in.
9555 * @param[in] db the #MDB_db record containing the stats to return.
9556 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9557 * @return 0, this function always succeeds.
9560 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9562 arg->ms_psize = env->me_psize;
9563 arg->ms_depth = db->md_depth;
9564 arg->ms_branch_pages = db->md_branch_pages;
9565 arg->ms_leaf_pages = db->md_leaf_pages;
9566 arg->ms_overflow_pages = db->md_overflow_pages;
9567 arg->ms_entries = db->md_entries;
9573 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9577 if (env == NULL || arg == NULL)
9580 meta = mdb_env_pick_meta(env);
9582 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9586 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9590 if (env == NULL || arg == NULL)
9593 meta = mdb_env_pick_meta(env);
9594 arg->me_mapaddr = meta->mm_address;
9595 arg->me_last_pgno = meta->mm_last_pg;
9596 arg->me_last_txnid = meta->mm_txnid;
9598 arg->me_mapsize = env->me_mapsize;
9599 arg->me_maxreaders = env->me_maxreaders;
9600 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9604 /** Set the default comparison functions for a database.
9605 * Called immediately after a database is opened to set the defaults.
9606 * The user can then override them with #mdb_set_compare() or
9607 * #mdb_set_dupsort().
9608 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9609 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9612 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9614 uint16_t f = txn->mt_dbs[dbi].md_flags;
9616 txn->mt_dbxs[dbi].md_cmp =
9617 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9618 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9620 txn->mt_dbxs[dbi].md_dcmp =
9621 !(f & MDB_DUPSORT) ? 0 :
9622 ((f & MDB_INTEGERDUP)
9623 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9624 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9627 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9633 int rc, dbflag, exact;
9634 unsigned int unused = 0, seq;
9638 if (flags & ~VALID_FLAGS)
9640 if (txn->mt_flags & MDB_TXN_BLOCKED)
9646 if (flags & PERSISTENT_FLAGS) {
9647 uint16_t f2 = flags & PERSISTENT_FLAGS;
9648 /* make sure flag changes get committed */
9649 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9650 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9651 txn->mt_flags |= MDB_TXN_DIRTY;
9654 mdb_default_cmp(txn, MAIN_DBI);
9658 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9659 mdb_default_cmp(txn, MAIN_DBI);
9662 /* Is the DB already open? */
9664 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9665 if (!txn->mt_dbxs[i].md_name.mv_size) {
9666 /* Remember this free slot */
9667 if (!unused) unused = i;
9670 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9671 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9677 /* If no free slot and max hit, fail */
9678 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9679 return MDB_DBS_FULL;
9681 /* Cannot mix named databases with some mainDB flags */
9682 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9683 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9685 /* Find the DB info */
9686 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9689 key.mv_data = (void *)name;
9690 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9691 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9692 if (rc == MDB_SUCCESS) {
9693 /* make sure this is actually a DB */
9694 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9695 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9696 return MDB_INCOMPATIBLE;
9697 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
9701 /* Done here so we cannot fail after creating a new DB */
9702 if ((namedup = strdup(name)) == NULL)
9706 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
9707 data.mv_size = sizeof(MDB_db);
9708 data.mv_data = &dummy;
9709 memset(&dummy, 0, sizeof(dummy));
9710 dummy.md_root = P_INVALID;
9711 dummy.md_flags = flags & PERSISTENT_FLAGS;
9712 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9719 /* Got info, register DBI in this txn */
9720 unsigned int slot = unused ? unused : txn->mt_numdbs;
9721 txn->mt_dbxs[slot].md_name.mv_data = namedup;
9722 txn->mt_dbxs[slot].md_name.mv_size = len;
9723 txn->mt_dbxs[slot].md_rel = NULL;
9724 txn->mt_dbflags[slot] = dbflag;
9725 /* txn-> and env-> are the same in read txns, use
9726 * tmp variable to avoid undefined assignment
9728 seq = ++txn->mt_env->me_dbiseqs[slot];
9729 txn->mt_dbiseqs[slot] = seq;
9731 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9733 mdb_default_cmp(txn, slot);
9743 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9745 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9748 if (txn->mt_flags & MDB_TXN_BLOCKED)
9751 if (txn->mt_dbflags[dbi] & DB_STALE) {
9754 /* Stale, must read the DB's root. cursor_init does it for us. */
9755 mdb_cursor_init(&mc, txn, dbi, &mx);
9757 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9760 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9763 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9765 ptr = env->me_dbxs[dbi].md_name.mv_data;
9766 /* If there was no name, this was already closed */
9768 env->me_dbxs[dbi].md_name.mv_data = NULL;
9769 env->me_dbxs[dbi].md_name.mv_size = 0;
9770 env->me_dbflags[dbi] = 0;
9771 env->me_dbiseqs[dbi]++;
9776 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9778 /* We could return the flags for the FREE_DBI too but what's the point? */
9779 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9781 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9785 /** Add all the DB's pages to the free list.
9786 * @param[in] mc Cursor on the DB to free.
9787 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9788 * @return 0 on success, non-zero on failure.
9791 mdb_drop0(MDB_cursor *mc, int subs)
9795 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9796 if (rc == MDB_SUCCESS) {
9797 MDB_txn *txn = mc->mc_txn;
9802 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9803 * This also avoids any P_LEAF2 pages, which have no nodes.
9804 * Also if the DB doesn't have sub-DBs and has no overflow
9805 * pages, omit scanning leaves.
9807 if ((mc->mc_flags & C_SUB) ||
9808 (!subs && !mc->mc_db->md_overflow_pages))
9811 mdb_cursor_copy(mc, &mx);
9812 while (mc->mc_snum > 0) {
9813 MDB_page *mp = mc->mc_pg[mc->mc_top];
9814 unsigned n = NUMKEYS(mp);
9816 for (i=0; i<n; i++) {
9817 ni = NODEPTR(mp, i);
9818 if (ni->mn_flags & F_BIGDATA) {
9821 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9822 rc = mdb_page_get(mc, pg, &omp, NULL);
9825 mdb_cassert(mc, IS_OVERFLOW(omp));
9826 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9830 mc->mc_db->md_overflow_pages -= omp->mp_pages;
9831 if (!mc->mc_db->md_overflow_pages && !subs)
9833 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9834 mdb_xcursor_init1(mc, ni);
9835 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9840 if (!subs && !mc->mc_db->md_overflow_pages)
9843 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9845 for (i=0; i<n; i++) {
9847 ni = NODEPTR(mp, i);
9850 mdb_midl_xappend(txn->mt_free_pgs, pg);
9855 mc->mc_ki[mc->mc_top] = i;
9856 rc = mdb_cursor_sibling(mc, 1);
9858 if (rc != MDB_NOTFOUND)
9860 /* no more siblings, go back to beginning
9861 * of previous level.
9866 for (i=1; i<mc->mc_snum; i++) {
9868 mc->mc_pg[i] = mx.mc_pg[i];
9873 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9876 txn->mt_flags |= MDB_TXN_ERROR;
9877 } else if (rc == MDB_NOTFOUND) {
9880 mc->mc_flags &= ~C_INITIALIZED;
9884 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9886 MDB_cursor *mc, *m2;
9889 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9892 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9895 if (TXN_DBI_CHANGED(txn, dbi))
9898 rc = mdb_cursor_open(txn, dbi, &mc);
9902 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9903 /* Invalidate the dropped DB's cursors */
9904 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9905 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9909 /* Can't delete the main DB */
9910 if (del && dbi >= CORE_DBS) {
9911 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9913 txn->mt_dbflags[dbi] = DB_STALE;
9914 mdb_dbi_close(txn->mt_env, dbi);
9916 txn->mt_flags |= MDB_TXN_ERROR;
9919 /* reset the DB record, mark it dirty */
9920 txn->mt_dbflags[dbi] |= DB_DIRTY;
9921 txn->mt_dbs[dbi].md_depth = 0;
9922 txn->mt_dbs[dbi].md_branch_pages = 0;
9923 txn->mt_dbs[dbi].md_leaf_pages = 0;
9924 txn->mt_dbs[dbi].md_overflow_pages = 0;
9925 txn->mt_dbs[dbi].md_entries = 0;
9926 txn->mt_dbs[dbi].md_root = P_INVALID;
9928 txn->mt_flags |= MDB_TXN_DIRTY;
9931 mdb_cursor_close(mc);
9935 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9937 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9940 txn->mt_dbxs[dbi].md_cmp = cmp;
9944 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9946 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9949 txn->mt_dbxs[dbi].md_dcmp = cmp;
9953 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9955 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9958 txn->mt_dbxs[dbi].md_rel = rel;
9962 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9964 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9967 txn->mt_dbxs[dbi].md_relctx = ctx;
9972 mdb_env_get_maxkeysize(MDB_env *env)
9974 return ENV_MAXKEY(env);
9978 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9980 unsigned int i, rdrs;
9983 int rc = 0, first = 1;
9987 if (!env->me_txns) {
9988 return func("(no reader locks)\n", ctx);
9990 rdrs = env->me_txns->mti_numreaders;
9991 mr = env->me_txns->mti_readers;
9992 for (i=0; i<rdrs; i++) {
9994 txnid_t txnid = mr[i].mr_txnid;
9995 sprintf(buf, txnid == (txnid_t)-1 ?
9996 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9997 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10000 rc = func(" pid thread txnid\n", ctx);
10004 rc = func(buf, ctx);
10010 rc = func("(no active readers)\n", ctx);
10015 /** Insert pid into list if not already present.
10016 * return -1 if already present.
10019 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10021 /* binary search of pid in list */
10023 unsigned cursor = 1;
10025 unsigned n = ids[0];
10028 unsigned pivot = n >> 1;
10029 cursor = base + pivot + 1;
10030 val = pid - ids[cursor];
10035 } else if ( val > 0 ) {
10040 /* found, so it's a duplicate */
10049 for (n = ids[0]; n > cursor; n--)
10056 mdb_reader_check(MDB_env *env, int *dead)
10062 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10065 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
10067 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10069 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10070 unsigned int i, j, rdrs;
10072 MDB_PID_T *pids, pid;
10073 int rc = MDB_SUCCESS, count = 0;
10075 rdrs = env->me_txns->mti_numreaders;
10076 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10080 mr = env->me_txns->mti_readers;
10081 for (i=0; i<rdrs; i++) {
10082 pid = mr[i].mr_pid;
10083 if (pid && pid != env->me_pid) {
10084 if (mdb_pid_insert(pids, pid) == 0) {
10085 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10086 /* Stale reader found */
10089 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10090 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10092 rdrs = 0; /* the above checked all readers */
10094 /* Recheck, a new process may have reused pid */
10095 if (mdb_reader_pid(env, Pidcheck, pid))
10099 for (; j<rdrs; j++)
10100 if (mr[j].mr_pid == pid) {
10101 DPRINTF(("clear stale reader pid %u txn %"Z"d",
10102 (unsigned) pid, mr[j].mr_txnid));
10107 UNLOCK_MUTEX(rmutex);
10118 #ifdef MDB_ROBUST_SUPPORTED
10119 /** Handle #LOCK_MUTEX0() failure.
10120 * Try to repair the lock file if the mutex owner died.
10121 * @param[in] env the environment handle
10122 * @param[in] mutex LOCK_MUTEX0() mutex
10123 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10124 * @return 0 on success with the mutex locked, or an error code on failure.
10127 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10132 if (rc == MDB_OWNERDEAD) {
10133 /* We own the mutex. Clean up after dead previous owner. */
10135 rlocked = (mutex == env->me_rmutex);
10137 /* Keep mti_txnid updated, otherwise next writer can
10138 * overwrite data which latest meta page refers to.
10140 meta = mdb_env_pick_meta(env);
10141 env->me_txns->mti_txnid = meta->mm_txnid;
10142 /* env is hosed if the dead thread was ours */
10144 env->me_flags |= MDB_FATAL_ERROR;
10145 env->me_txn = NULL;
10149 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10150 (rc ? "this process' env is hosed" : "recovering")));
10151 rc2 = mdb_reader_check0(env, rlocked, NULL);
10153 rc2 = mdb_mutex_consistent(mutex);
10154 if (rc || (rc = rc2)) {
10155 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10156 UNLOCK_MUTEX(mutex);
10162 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10167 #endif /* MDB_ROBUST_SUPPORTED */
10169 #if defined(_WIN32)
10170 /** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */
10172 utf8_to_utf16(const char *src, MDB_name *dst, int xtra)
10175 wchar_t *result = NULL;
10176 for (;;) { /* malloc result, then fill it in */
10177 need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need);
10184 result = malloc(sizeof(wchar_t) * (need + xtra));
10189 dst->mn_alloced = 1;
10190 dst->mn_len = need - 1;
10191 dst->mn_val = result;
10192 return MDB_SUCCESS;
10195 #endif /* defined(_WIN32) */