2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2015 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
39 #define _FILE_OFFSET_BITS 64
45 /* We use native NT APIs to setup the memory map, so that we can
46 * let the DB file grow incrementally instead of always preallocating
47 * the full size. These APIs are defined in <wdm.h> and <ntifs.h>
48 * but those headers are meant for driver-level development and
49 * conflict with the regular user-level headers, so we explicitly
50 * declare them here. Using these APIs also means we must link to
51 * ntdll.dll, which is not linked by default in user code.
54 NtCreateSection(OUT PHANDLE sh, IN ACCESS_MASK acc,
55 IN void * oa OPTIONAL,
56 IN PLARGE_INTEGER ms OPTIONAL,
57 IN ULONG pp, IN ULONG aa, IN HANDLE fh OPTIONAL);
59 typedef enum _SECTION_INHERIT {
65 NtMapViewOfSection(IN PHANDLE sh, IN HANDLE ph,
66 IN OUT PVOID *addr, IN ULONG_PTR zbits,
67 IN SIZE_T cs, IN OUT PLARGE_INTEGER off OPTIONAL,
68 IN OUT PSIZE_T vs, IN SECTION_INHERIT ih,
69 IN ULONG at, IN ULONG pp);
74 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
75 * as int64 which is wrong. MSVC doesn't define it at all, so just
79 #define MDB_THR_T DWORD
80 #include <sys/types.h>
83 # include <sys/param.h>
85 # define LITTLE_ENDIAN 1234
86 # define BIG_ENDIAN 4321
87 # define BYTE_ORDER LITTLE_ENDIAN
89 # define SSIZE_MAX INT_MAX
93 #include <sys/types.h>
95 #define MDB_PID_T pid_t
96 #define MDB_THR_T pthread_t
97 #include <sys/param.h>
100 #ifdef HAVE_SYS_FILE_H
101 #include <sys/file.h>
106 #if defined(__mips) && defined(__linux)
107 /* MIPS has cache coherency issues, requires explicit cache control */
108 #include <asm/cachectl.h>
109 extern int cacheflush(char *addr, int nbytes, int cache);
110 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
112 #define CACHEFLUSH(addr, bytes, cache)
115 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
116 /** fdatasync is broken on ext3/ext4fs on older kernels, see
117 * description in #mdb_env_open2 comments. You can safely
118 * define MDB_FDATASYNC_WORKS if this code will only be run
119 * on kernels 3.6 and newer.
121 #define BROKEN_FDATASYNC
127 #include <inttypes.h>
135 typedef SSIZE_T ssize_t;
140 #if defined(__sun) || defined(ANDROID)
141 /* Most platforms have posix_memalign, older may only have memalign */
142 #define HAVE_MEMALIGN 1
146 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
147 #include <netinet/in.h>
148 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
151 #if defined(__APPLE__) || defined (BSD)
152 # if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
153 # define MDB_USE_SYSV_SEM 1
155 # define MDB_FDATASYNC fsync
156 #elif defined(ANDROID)
157 # define MDB_FDATASYNC fsync
162 #ifdef MDB_USE_POSIX_SEM
163 # define MDB_USE_HASH 1
164 #include <semaphore.h>
165 #elif defined(MDB_USE_SYSV_SEM)
168 #ifdef _SEM_SEMUN_UNDEFINED
171 struct semid_ds *buf;
172 unsigned short *array;
174 #endif /* _SEM_SEMUN_UNDEFINED */
176 #define MDB_USE_POSIX_MUTEX 1
177 #endif /* MDB_USE_POSIX_SEM */
180 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
181 + defined(MDB_USE_POSIX_MUTEX) != 1
182 # error "Ambiguous shared-lock implementation"
186 #include <valgrind/memcheck.h>
187 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
188 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
189 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
190 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
191 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
193 #define VGMEMP_CREATE(h,r,z)
194 #define VGMEMP_ALLOC(h,a,s)
195 #define VGMEMP_FREE(h,a)
196 #define VGMEMP_DESTROY(h)
197 #define VGMEMP_DEFINED(a,s)
201 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
202 /* Solaris just defines one or the other */
203 # define LITTLE_ENDIAN 1234
204 # define BIG_ENDIAN 4321
205 # ifdef _LITTLE_ENDIAN
206 # define BYTE_ORDER LITTLE_ENDIAN
208 # define BYTE_ORDER BIG_ENDIAN
211 # define BYTE_ORDER __BYTE_ORDER
215 #ifndef LITTLE_ENDIAN
216 #define LITTLE_ENDIAN __LITTLE_ENDIAN
219 #define BIG_ENDIAN __BIG_ENDIAN
222 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
223 #define MISALIGNED_OK 1
229 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
230 # error "Unknown or unsupported endianness (BYTE_ORDER)"
231 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
232 # error "Two's complement, reasonably sized integer types, please"
236 /** Put infrequently used env functions in separate section */
238 # define ESECT __attribute__ ((section("__TEXT,text_env")))
240 # define ESECT __attribute__ ((section("text_env")))
247 #define CALL_CONV WINAPI
252 /** @defgroup internal LMDB Internals
255 /** @defgroup compat Compatibility Macros
256 * A bunch of macros to minimize the amount of platform-specific ifdefs
257 * needed throughout the rest of the code. When the features this library
258 * needs are similar enough to POSIX to be hidden in a one-or-two line
259 * replacement, this macro approach is used.
263 /** Features under development */
268 /** Wrapper around __func__, which is a C99 feature */
269 #if __STDC_VERSION__ >= 199901L
270 # define mdb_func_ __func__
271 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
272 # define mdb_func_ __FUNCTION__
274 /* If a debug message says <mdb_unknown>(), update the #if statements above */
275 # define mdb_func_ "<mdb_unknown>"
278 /* Internal error codes, not exposed outside liblmdb */
279 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
281 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
282 #elif defined MDB_USE_SYSV_SEM
283 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
284 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
285 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
289 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
291 /** Some platforms define the EOWNERDEAD error code
292 * even though they don't support Robust Mutexes.
293 * Compile with -DMDB_USE_ROBUST=0, or use some other
294 * mechanism like -DMDB_USE_SYSV_SEM instead of
295 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
296 * also Robust, but some systems don't support them
299 #ifndef MDB_USE_ROBUST
300 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
301 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
302 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
303 # define MDB_USE_ROBUST 0
305 # define MDB_USE_ROBUST 1
306 /* glibc < 2.12 only provided _np API */
307 # if defined(__GLIBC__) && GLIBC_VER < 0x02000c
308 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
309 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
310 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
313 #endif /* MDB_USE_ROBUST */
315 #if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
316 #define MDB_ROBUST_SUPPORTED 1
320 #define MDB_USE_HASH 1
321 #define MDB_PIDLOCK 0
322 #define THREAD_RET DWORD
323 #define pthread_t HANDLE
324 #define pthread_mutex_t HANDLE
325 #define pthread_cond_t HANDLE
326 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
327 #define pthread_key_t DWORD
328 #define pthread_self() GetCurrentThreadId()
329 #define pthread_key_create(x,y) \
330 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
331 #define pthread_key_delete(x) TlsFree(x)
332 #define pthread_getspecific(x) TlsGetValue(x)
333 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
334 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
335 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
336 #define pthread_cond_signal(x) SetEvent(*x)
337 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
338 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
339 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
340 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
341 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
342 #define mdb_mutex_consistent(mutex) 0
343 #define getpid() GetCurrentProcessId()
344 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
345 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
346 #define ErrCode() GetLastError()
347 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
348 #define close(fd) (CloseHandle(fd) ? 0 : -1)
349 #define munmap(ptr,len) UnmapViewOfFile(ptr)
350 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
351 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
353 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
357 #define THREAD_RET void *
358 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
359 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
360 #define Z "z" /**< printf format modifier for size_t */
362 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
363 #define MDB_PIDLOCK 1
365 #ifdef MDB_USE_POSIX_SEM
367 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
368 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
369 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
372 mdb_sem_wait(sem_t *sem)
375 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
379 #elif defined MDB_USE_SYSV_SEM
381 typedef struct mdb_mutex {
385 } mdb_mutex_t[1], *mdb_mutexref_t;
387 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
388 #define UNLOCK_MUTEX(mutex) do { \
389 struct sembuf sb = { 0, 1, SEM_UNDO }; \
390 sb.sem_num = (mutex)->semnum; \
391 *(mutex)->locked = 0; \
392 semop((mutex)->semid, &sb, 1); \
396 mdb_sem_wait(mdb_mutexref_t sem)
398 int rc, *locked = sem->locked;
399 struct sembuf sb = { 0, -1, SEM_UNDO };
400 sb.sem_num = sem->semnum;
402 if (!semop(sem->semid, &sb, 1)) {
403 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
407 } while ((rc = errno) == EINTR);
411 #define mdb_mutex_consistent(mutex) 0
413 #else /* MDB_USE_POSIX_MUTEX: */
414 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
415 * local variables keep it (mdb_mutexref_t).
417 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
418 * be the same, or an array[size 1] and a pointer.
421 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
423 /** Lock the reader or writer mutex.
424 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
426 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
427 /** Unlock the reader or writer mutex.
429 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
430 /** Mark mutex-protected data as repaired, after death of previous owner.
432 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
433 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
435 /** Get the error code for the last failed system function.
437 #define ErrCode() errno
439 /** An abstraction for a file handle.
440 * On POSIX systems file handles are small integers. On Windows
441 * they're opaque pointers.
445 /** A value for an invalid file handle.
446 * Mainly used to initialize file variables and signify that they are
449 #define INVALID_HANDLE_VALUE (-1)
451 /** Get the size of a memory page for the system.
452 * This is the basic size that the platform's memory manager uses, and is
453 * fundamental to the use of memory-mapped files.
455 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
468 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
470 #elif defined(MDB_USE_SYSV_SEM)
471 #define MNAME_LEN (sizeof(int))
473 #define MNAME_LEN (sizeof(pthread_mutex_t))
476 #ifdef MDB_USE_SYSV_SEM
477 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
479 #define SYSV_SEM_FLAG 0
484 #ifdef MDB_ROBUST_SUPPORTED
485 /** Lock mutex, handle any error, set rc = result.
486 * Return 0 on success, nonzero (not rc) on error.
488 #define LOCK_MUTEX(rc, env, mutex) \
489 (((rc) = LOCK_MUTEX0(mutex)) && \
490 ((rc) = mdb_mutex_failed(env, mutex, rc)))
491 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
493 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
494 #define mdb_mutex_failed(env, mutex, rc) (rc)
498 /** A flag for opening a file and requesting synchronous data writes.
499 * This is only used when writing a meta page. It's not strictly needed;
500 * we could just do a normal write and then immediately perform a flush.
501 * But if this flag is available it saves us an extra system call.
503 * @note If O_DSYNC is undefined but exists in /usr/include,
504 * preferably set some compiler flag to get the definition.
508 # define MDB_DSYNC O_DSYNC
510 # define MDB_DSYNC O_SYNC
515 /** Function for flushing the data of a file. Define this to fsync
516 * if fdatasync() is not supported.
518 #ifndef MDB_FDATASYNC
519 # define MDB_FDATASYNC fdatasync
523 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
534 /** A page number in the database.
535 * Note that 64 bit page numbers are overkill, since pages themselves
536 * already represent 12-13 bits of addressable memory, and the OS will
537 * always limit applications to a maximum of 63 bits of address space.
539 * @note In the #MDB_node structure, we only store 48 bits of this value,
540 * which thus limits us to only 60 bits of addressable data.
542 typedef MDB_ID pgno_t;
544 /** A transaction ID.
545 * See struct MDB_txn.mt_txnid for details.
547 typedef MDB_ID txnid_t;
549 /** @defgroup debug Debug Macros
553 /** Enable debug output. Needs variable argument macros (a C99 feature).
554 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
555 * read from and written to the database (used for free space management).
561 static int mdb_debug;
562 static txnid_t mdb_debug_start;
564 /** Print a debug message with printf formatting.
565 * Requires double parenthesis around 2 or more args.
567 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
568 # define DPRINTF0(fmt, ...) \
569 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
571 # define DPRINTF(args) ((void) 0)
573 /** Print a debug string.
574 * The string is printed literally, with no format processing.
576 #define DPUTS(arg) DPRINTF(("%s", arg))
577 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
579 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
582 /** @brief The maximum size of a database page.
584 * It is 32k or 64k, since value-PAGEBASE must fit in
585 * #MDB_page.%mp_upper.
587 * LMDB will use database pages < OS pages if needed.
588 * That causes more I/O in write transactions: The OS must
589 * know (read) the whole page before writing a partial page.
591 * Note that we don't currently support Huge pages. On Linux,
592 * regular data files cannot use Huge pages, and in general
593 * Huge pages aren't actually pageable. We rely on the OS
594 * demand-pager to read our data and page it out when memory
595 * pressure from other processes is high. So until OSs have
596 * actual paging support for Huge pages, they're not viable.
598 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
600 /** The minimum number of keys required in a database page.
601 * Setting this to a larger value will place a smaller bound on the
602 * maximum size of a data item. Data items larger than this size will
603 * be pushed into overflow pages instead of being stored directly in
604 * the B-tree node. This value used to default to 4. With a page size
605 * of 4096 bytes that meant that any item larger than 1024 bytes would
606 * go into an overflow page. That also meant that on average 2-3KB of
607 * each overflow page was wasted space. The value cannot be lower than
608 * 2 because then there would no longer be a tree structure. With this
609 * value, items larger than 2KB will go into overflow pages, and on
610 * average only 1KB will be wasted.
612 #define MDB_MINKEYS 2
614 /** A stamp that identifies a file as an LMDB file.
615 * There's nothing special about this value other than that it is easily
616 * recognizable, and it will reflect any byte order mismatches.
618 #define MDB_MAGIC 0xBEEFC0DE
620 /** The version number for a database's datafile format. */
621 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
622 /** The version number for a database's lockfile format. */
623 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
625 /** @brief The max size of a key we can write, or 0 for computed max.
627 * This macro should normally be left alone or set to 0.
628 * Note that a database with big keys or dupsort data cannot be
629 * reliably modified by a liblmdb which uses a smaller max.
630 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
632 * Other values are allowed, for backwards compat. However:
633 * A value bigger than the computed max can break if you do not
634 * know what you are doing, and liblmdb <= 0.9.10 can break when
635 * modifying a DB with keys/dupsort data bigger than its max.
637 * Data items in an #MDB_DUPSORT database are also limited to
638 * this size, since they're actually keys of a sub-DB. Keys and
639 * #MDB_DUPSORT data items must fit on a node in a regular page.
641 #ifndef MDB_MAXKEYSIZE
642 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
645 /** The maximum size of a key we can write to the environment. */
647 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
649 #define ENV_MAXKEY(env) ((env)->me_maxkey)
652 /** @brief The maximum size of a data item.
654 * We only store a 32 bit value for node sizes.
656 #define MAXDATASIZE 0xffffffffUL
659 /** Key size which fits in a #DKBUF.
662 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
665 * This is used for printing a hex dump of a key's contents.
667 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
668 /** Display a key in hex.
670 * Invoke a function to display a key in hex.
672 #define DKEY(x) mdb_dkey(x, kbuf)
678 /** An invalid page number.
679 * Mainly used to denote an empty tree.
681 #define P_INVALID (~(pgno_t)0)
683 /** Test if the flags \b f are set in a flag word \b w. */
684 #define F_ISSET(w, f) (((w) & (f)) == (f))
686 /** Round \b n up to an even number. */
687 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
689 /** Used for offsets within a single page.
690 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
693 typedef uint16_t indx_t;
695 /** Default size of memory map.
696 * This is certainly too small for any actual applications. Apps should always set
697 * the size explicitly using #mdb_env_set_mapsize().
699 #define DEFAULT_MAPSIZE 1048576
701 /** @defgroup readers Reader Lock Table
702 * Readers don't acquire any locks for their data access. Instead, they
703 * simply record their transaction ID in the reader table. The reader
704 * mutex is needed just to find an empty slot in the reader table. The
705 * slot's address is saved in thread-specific data so that subsequent read
706 * transactions started by the same thread need no further locking to proceed.
708 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
710 * No reader table is used if the database is on a read-only filesystem, or
711 * if #MDB_NOLOCK is set.
713 * Since the database uses multi-version concurrency control, readers don't
714 * actually need any locking. This table is used to keep track of which
715 * readers are using data from which old transactions, so that we'll know
716 * when a particular old transaction is no longer in use. Old transactions
717 * that have discarded any data pages can then have those pages reclaimed
718 * for use by a later write transaction.
720 * The lock table is constructed such that reader slots are aligned with the
721 * processor's cache line size. Any slot is only ever used by one thread.
722 * This alignment guarantees that there will be no contention or cache
723 * thrashing as threads update their own slot info, and also eliminates
724 * any need for locking when accessing a slot.
726 * A writer thread will scan every slot in the table to determine the oldest
727 * outstanding reader transaction. Any freed pages older than this will be
728 * reclaimed by the writer. The writer doesn't use any locks when scanning
729 * this table. This means that there's no guarantee that the writer will
730 * see the most up-to-date reader info, but that's not required for correct
731 * operation - all we need is to know the upper bound on the oldest reader,
732 * we don't care at all about the newest reader. So the only consequence of
733 * reading stale information here is that old pages might hang around a
734 * while longer before being reclaimed. That's actually good anyway, because
735 * the longer we delay reclaiming old pages, the more likely it is that a
736 * string of contiguous pages can be found after coalescing old pages from
737 * many old transactions together.
740 /** Number of slots in the reader table.
741 * This value was chosen somewhat arbitrarily. 126 readers plus a
742 * couple mutexes fit exactly into 8KB on my development machine.
743 * Applications should set the table size using #mdb_env_set_maxreaders().
745 #define DEFAULT_READERS 126
747 /** The size of a CPU cache line in bytes. We want our lock structures
748 * aligned to this size to avoid false cache line sharing in the
750 * This value works for most CPUs. For Itanium this should be 128.
756 /** The information we store in a single slot of the reader table.
757 * In addition to a transaction ID, we also record the process and
758 * thread ID that owns a slot, so that we can detect stale information,
759 * e.g. threads or processes that went away without cleaning up.
760 * @note We currently don't check for stale records. We simply re-init
761 * the table when we know that we're the only process opening the
764 typedef struct MDB_rxbody {
765 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
766 * Multiple readers that start at the same time will probably have the
767 * same ID here. Again, it's not important to exclude them from
768 * anything; all we need to know is which version of the DB they
769 * started from so we can avoid overwriting any data used in that
770 * particular version.
772 volatile txnid_t mrb_txnid;
773 /** The process ID of the process owning this reader txn. */
774 volatile MDB_PID_T mrb_pid;
775 /** The thread ID of the thread owning this txn. */
776 volatile MDB_THR_T mrb_tid;
779 /** The actual reader record, with cacheline padding. */
780 typedef struct MDB_reader {
783 /** shorthand for mrb_txnid */
784 #define mr_txnid mru.mrx.mrb_txnid
785 #define mr_pid mru.mrx.mrb_pid
786 #define mr_tid mru.mrx.mrb_tid
787 /** cache line alignment */
788 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
792 /** The header for the reader table.
793 * The table resides in a memory-mapped file. (This is a different file
794 * than is used for the main database.)
796 * For POSIX the actual mutexes reside in the shared memory of this
797 * mapped file. On Windows, mutexes are named objects allocated by the
798 * kernel; we store the mutex names in this mapped file so that other
799 * processes can grab them. This same approach is also used on
800 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
801 * process-shared POSIX mutexes. For these cases where a named object
802 * is used, the object name is derived from a 64 bit FNV hash of the
803 * environment pathname. As such, naming collisions are extremely
804 * unlikely. If a collision occurs, the results are unpredictable.
806 typedef struct MDB_txbody {
807 /** Stamp identifying this as an LMDB file. It must be set
810 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
812 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
813 char mtb_rmname[MNAME_LEN];
814 #elif defined(MDB_USE_SYSV_SEM)
818 /** Mutex protecting access to this table.
819 * This is the reader table lock used with LOCK_MUTEX().
821 mdb_mutex_t mtb_rmutex;
823 /** The ID of the last transaction committed to the database.
824 * This is recorded here only for convenience; the value can always
825 * be determined by reading the main database meta pages.
827 volatile txnid_t mtb_txnid;
828 /** The number of slots that have been used in the reader table.
829 * This always records the maximum count, it is not decremented
830 * when readers release their slots.
832 volatile unsigned mtb_numreaders;
835 /** The actual reader table definition. */
836 typedef struct MDB_txninfo {
839 #define mti_magic mt1.mtb.mtb_magic
840 #define mti_format mt1.mtb.mtb_format
841 #define mti_rmutex mt1.mtb.mtb_rmutex
842 #define mti_rmname mt1.mtb.mtb_rmname
843 #define mti_txnid mt1.mtb.mtb_txnid
844 #define mti_numreaders mt1.mtb.mtb_numreaders
845 #ifdef MDB_USE_SYSV_SEM
846 #define mti_semid mt1.mtb.mtb_semid
847 #define mti_rlocked mt1.mtb.mtb_rlocked
849 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
852 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
853 char mt2_wmname[MNAME_LEN];
854 #define mti_wmname mt2.mt2_wmname
855 #elif defined MDB_USE_SYSV_SEM
857 #define mti_wlocked mt2.mt2_wlocked
859 mdb_mutex_t mt2_wmutex;
860 #define mti_wmutex mt2.mt2_wmutex
862 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
864 MDB_reader mti_readers[1];
867 /** Lockfile format signature: version, features and field layout */
868 #define MDB_LOCK_FORMAT \
870 ((MDB_LOCK_VERSION) \
871 /* Flags which describe functionality */ \
872 + (SYSV_SEM_FLAG << 18) \
873 + (((MDB_PIDLOCK) != 0) << 16)))
876 /** Common header for all page types.
877 * Overflow records occupy a number of contiguous pages with no
878 * headers on any page after the first.
880 typedef struct MDB_page {
881 #define mp_pgno mp_p.p_pgno
882 #define mp_next mp_p.p_next
884 pgno_t p_pgno; /**< page number */
885 struct MDB_page *p_next; /**< for in-memory list of freed pages */
888 /** @defgroup mdb_page Page Flags
890 * Flags for the page headers.
893 #define P_BRANCH 0x01 /**< branch page */
894 #define P_LEAF 0x02 /**< leaf page */
895 #define P_OVERFLOW 0x04 /**< overflow page */
896 #define P_META 0x08 /**< meta page */
897 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
898 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
899 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
900 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
901 #define P_KEEP 0x8000 /**< leave this page alone during spill */
903 uint16_t mp_flags; /**< @ref mdb_page */
904 #define mp_lower mp_pb.pb.pb_lower
905 #define mp_upper mp_pb.pb.pb_upper
906 #define mp_pages mp_pb.pb_pages
909 indx_t pb_lower; /**< lower bound of free space */
910 indx_t pb_upper; /**< upper bound of free space */
912 uint32_t pb_pages; /**< number of overflow pages */
914 indx_t mp_ptrs[1]; /**< dynamic size */
917 /** Size of the page header, excluding dynamic data at the end */
918 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
920 /** Address of first usable data byte in a page, after the header */
921 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
923 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
924 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
926 /** Number of nodes on a page */
927 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
929 /** The amount of space remaining in the page */
930 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
932 /** The percentage of space used in the page, in tenths of a percent. */
933 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
934 ((env)->me_psize - PAGEHDRSZ))
935 /** The minimum page fill factor, in tenths of a percent.
936 * Pages emptier than this are candidates for merging.
938 #define FILL_THRESHOLD 250
940 /** Test if a page is a leaf page */
941 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
942 /** Test if a page is a LEAF2 page */
943 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
944 /** Test if a page is a branch page */
945 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
946 /** Test if a page is an overflow page */
947 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
948 /** Test if a page is a sub page */
949 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
951 /** The number of overflow pages needed to store the given size. */
952 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
954 /** Link in #MDB_txn.%mt_loose_pgs list */
955 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
957 /** Header for a single key/data pair within a page.
958 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
959 * We guarantee 2-byte alignment for 'MDB_node's.
961 typedef struct MDB_node {
962 /** lo and hi are used for data size on leaf nodes and for
963 * child pgno on branch nodes. On 64 bit platforms, flags
964 * is also used for pgno. (Branch nodes have no flags).
965 * They are in host byte order in case that lets some
966 * accesses be optimized into a 32-bit word access.
968 #if BYTE_ORDER == LITTLE_ENDIAN
969 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
971 unsigned short mn_hi, mn_lo;
973 /** @defgroup mdb_node Node Flags
975 * Flags for node headers.
978 #define F_BIGDATA 0x01 /**< data put on overflow page */
979 #define F_SUBDATA 0x02 /**< data is a sub-database */
980 #define F_DUPDATA 0x04 /**< data has duplicates */
982 /** valid flags for #mdb_node_add() */
983 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
986 unsigned short mn_flags; /**< @ref mdb_node */
987 unsigned short mn_ksize; /**< key size */
988 char mn_data[1]; /**< key and data are appended here */
991 /** Size of the node header, excluding dynamic data at the end */
992 #define NODESIZE offsetof(MDB_node, mn_data)
994 /** Bit position of top word in page number, for shifting mn_flags */
995 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
997 /** Size of a node in a branch page with a given key.
998 * This is just the node header plus the key, there is no data.
1000 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
1002 /** Size of a node in a leaf page with a given key and data.
1003 * This is node header plus key plus data size.
1005 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
1007 /** Address of node \b i in page \b p */
1008 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
1010 /** Address of the key for the node */
1011 #define NODEKEY(node) (void *)((node)->mn_data)
1013 /** Address of the data for a node */
1014 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
1016 /** Get the page number pointed to by a branch node */
1017 #define NODEPGNO(node) \
1018 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
1019 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
1020 /** Set the page number in a branch node */
1021 #define SETPGNO(node,pgno) do { \
1022 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
1023 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
1025 /** Get the size of the data in a leaf node */
1026 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
1027 /** Set the size of the data for a leaf node */
1028 #define SETDSZ(node,size) do { \
1029 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
1030 /** The size of a key in a node */
1031 #define NODEKSZ(node) ((node)->mn_ksize)
1033 /** Copy a page number from src to dst */
1034 #ifdef MISALIGNED_OK
1035 #define COPY_PGNO(dst,src) dst = src
1037 #if SIZE_MAX > 4294967295UL
1038 #define COPY_PGNO(dst,src) do { \
1039 unsigned short *s, *d; \
1040 s = (unsigned short *)&(src); \
1041 d = (unsigned short *)&(dst); \
1048 #define COPY_PGNO(dst,src) do { \
1049 unsigned short *s, *d; \
1050 s = (unsigned short *)&(src); \
1051 d = (unsigned short *)&(dst); \
1057 /** The address of a key in a LEAF2 page.
1058 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
1059 * There are no node headers, keys are stored contiguously.
1061 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
1063 /** Set the \b node's key into \b keyptr, if requested. */
1064 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
1065 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
1067 /** Set the \b node's key into \b key. */
1068 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
1070 /** Information about a single database in the environment. */
1071 typedef struct MDB_db {
1072 uint32_t md_pad; /**< also ksize for LEAF2 pages */
1073 uint16_t md_flags; /**< @ref mdb_dbi_open */
1074 uint16_t md_depth; /**< depth of this tree */
1075 pgno_t md_branch_pages; /**< number of internal pages */
1076 pgno_t md_leaf_pages; /**< number of leaf pages */
1077 pgno_t md_overflow_pages; /**< number of overflow pages */
1078 pgno_t md_entries; /**< number of data items */
1079 pgno_t md_root; /**< the root page of this tree */
1082 /** mdb_dbi_open flags */
1083 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1084 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1085 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1086 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1088 /** Handle for the DB used to track free pages. */
1090 /** Handle for the default DB. */
1092 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1095 /** Number of meta pages - also hardcoded elsewhere */
1098 /** Meta page content.
1099 * A meta page is the start point for accessing a database snapshot.
1100 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1102 typedef struct MDB_meta {
1103 /** Stamp identifying this as an LMDB file. It must be set
1106 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1107 uint32_t mm_version;
1109 union { /* always zero since we don't support fixed mapping in MDB_VL32 */
1113 #define mm_address mm_un.mmun_address
1115 void *mm_address; /**< address for fixed mapping */
1117 pgno_t mm_mapsize; /**< size of mmap region */
1118 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1119 /** The size of pages used in this DB */
1120 #define mm_psize mm_dbs[FREE_DBI].md_pad
1121 /** Any persistent environment flags. @ref mdb_env */
1122 #define mm_flags mm_dbs[FREE_DBI].md_flags
1123 pgno_t mm_last_pg; /**< last used page in file */
1124 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1127 /** Buffer for a stack-allocated meta page.
1128 * The members define size and alignment, and silence type
1129 * aliasing warnings. They are not used directly; that could
1130 * mean incorrectly using several union members in parallel.
1132 typedef union MDB_metabuf {
1135 char mm_pad[PAGEHDRSZ];
1140 /** Auxiliary DB info.
1141 * The information here is mostly static/read-only. There is
1142 * only a single copy of this record in the environment.
1144 typedef struct MDB_dbx {
1145 MDB_val md_name; /**< name of the database */
1146 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1147 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1148 MDB_rel_func *md_rel; /**< user relocate function */
1149 void *md_relctx; /**< user-provided context for md_rel */
1152 /** A database transaction.
1153 * Every operation requires a transaction handle.
1156 MDB_txn *mt_parent; /**< parent of a nested txn */
1157 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1159 pgno_t mt_next_pgno; /**< next unallocated page */
1161 pgno_t mt_last_pgno; /**< last written page */
1163 /** The ID of this transaction. IDs are integers incrementing from 1.
1164 * Only committed write transactions increment the ID. If a transaction
1165 * aborts, the ID may be re-used by the next writer.
1168 MDB_env *mt_env; /**< the DB environment */
1169 /** The list of pages that became unused during this transaction.
1171 MDB_IDL mt_free_pgs;
1172 /** The list of loose pages that became unused and may be reused
1173 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1175 MDB_page *mt_loose_pgs;
1176 /* #Number of loose pages (#mt_loose_pgs) */
1178 /** The sorted list of dirty pages we temporarily wrote to disk
1179 * because the dirty list was full. page numbers in here are
1180 * shifted left by 1, deleted slots have the LSB set.
1182 MDB_IDL mt_spill_pgs;
1184 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1185 MDB_ID2L dirty_list;
1186 /** For read txns: This thread/txn's reader table slot, or NULL. */
1189 /** Array of records for each DB known in the environment. */
1191 /** Array of MDB_db records for each known DB */
1193 /** Array of sequence numbers for each DB handle */
1194 unsigned int *mt_dbiseqs;
1195 /** @defgroup mt_dbflag Transaction DB Flags
1199 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1200 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1201 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1202 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1203 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1205 /** In write txns, array of cursors for each DB */
1206 MDB_cursor **mt_cursors;
1207 /** Array of flags for each DB */
1208 unsigned char *mt_dbflags;
1210 /** List of read-only pages (actually chunks) */
1212 /** We map chunks of 16 pages. Even though Windows uses 4KB pages, all
1213 * mappings must begin on 64KB boundaries. So we round off all pgnos to
1214 * a chunk boundary. We do the same on Linux for symmetry, and also to
1215 * reduce the frequency of mmap/munmap calls.
1217 #define MDB_RPAGE_CHUNK 16
1218 #define MDB_TRPAGE_SIZE 4096 /**< size of #mt_rpages array of chunks */
1219 #define MDB_TRPAGE_MAX (MDB_TRPAGE_SIZE-1) /**< maximum chunk index */
1220 unsigned int mt_rpcheck; /**< threshold for reclaiming unref'd chunks */
1222 /** Number of DB records in use, or 0 when the txn is finished.
1223 * This number only ever increments until the txn finishes; we
1224 * don't decrement it when individual DB handles are closed.
1228 /** @defgroup mdb_txn Transaction Flags
1232 /** #mdb_txn_begin() flags */
1233 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1234 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1235 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1236 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1237 /* internal txn flags */
1238 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1239 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1240 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1241 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1242 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1243 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1244 /** most operations on the txn are currently illegal */
1245 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1247 unsigned int mt_flags; /**< @ref mdb_txn */
1248 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1249 * Includes ancestor txns' dirty pages not hidden by other txns'
1250 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1251 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1253 unsigned int mt_dirty_room;
1256 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1257 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1258 * raise this on a 64 bit machine.
1260 #define CURSOR_STACK 32
1264 /** Cursors are used for all DB operations.
1265 * A cursor holds a path of (page pointer, key index) from the DB
1266 * root to a position in the DB, plus other state. #MDB_DUPSORT
1267 * cursors include an xcursor to the current data item. Write txns
1268 * track their cursors and keep them up to date when data moves.
1269 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1270 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1273 /** Next cursor on this DB in this txn */
1274 MDB_cursor *mc_next;
1275 /** Backup of the original cursor if this cursor is a shadow */
1276 MDB_cursor *mc_backup;
1277 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1278 struct MDB_xcursor *mc_xcursor;
1279 /** The transaction that owns this cursor */
1281 /** The database handle this cursor operates on */
1283 /** The database record for this cursor */
1285 /** The database auxiliary record for this cursor */
1287 /** The @ref mt_dbflag for this database */
1288 unsigned char *mc_dbflag;
1289 unsigned short mc_snum; /**< number of pushed pages */
1290 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1291 /** @defgroup mdb_cursor Cursor Flags
1293 * Cursor state flags.
1296 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1297 #define C_EOF 0x02 /**< No more data */
1298 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1299 #define C_DEL 0x08 /**< last op was a cursor_del */
1300 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1302 unsigned int mc_flags; /**< @ref mdb_cursor */
1303 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1304 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1307 /** Context for sorted-dup records.
1308 * We could have gone to a fully recursive design, with arbitrarily
1309 * deep nesting of sub-databases. But for now we only handle these
1310 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1312 typedef struct MDB_xcursor {
1313 /** A sub-cursor for traversing the Dup DB */
1314 MDB_cursor mx_cursor;
1315 /** The database record for this Dup DB */
1317 /** The auxiliary DB record for this Dup DB */
1319 /** The @ref mt_dbflag for this Dup DB */
1320 unsigned char mx_dbflag;
1323 /** State of FreeDB old pages, stored in the MDB_env */
1324 typedef struct MDB_pgstate {
1325 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1326 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1329 /** The database environment. */
1331 HANDLE me_fd; /**< The main data file */
1332 HANDLE me_lfd; /**< The lock file */
1333 HANDLE me_mfd; /**< just for writing the meta pages */
1334 #if defined(MDB_VL32) && defined(_WIN32)
1335 HANDLE me_fmh; /**< File Mapping handle */
1337 /** Failed to update the meta page. Probably an I/O error. */
1338 #define MDB_FATAL_ERROR 0x80000000U
1339 /** Some fields are initialized. */
1340 #define MDB_ENV_ACTIVE 0x20000000U
1341 /** me_txkey is set */
1342 #define MDB_ENV_TXKEY 0x10000000U
1343 /** fdatasync is unreliable */
1344 #define MDB_FSYNCONLY 0x08000000U
1345 uint32_t me_flags; /**< @ref mdb_env */
1346 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1347 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1348 unsigned int me_maxreaders; /**< size of the reader table */
1349 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1350 volatile int me_close_readers;
1351 MDB_dbi me_numdbs; /**< number of DBs opened */
1352 MDB_dbi me_maxdbs; /**< size of the DB table */
1353 MDB_PID_T me_pid; /**< process ID of this env */
1354 char *me_path; /**< path to the DB files */
1355 char *me_map; /**< the memory map of the data file */
1356 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1357 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1358 void *me_pbuf; /**< scratch area for DUPSORT put() */
1359 MDB_txn *me_txn; /**< current write transaction */
1360 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1361 mdb_size_t me_mapsize; /**< size of the data memory map */
1362 off_t me_size; /**< current file size */
1363 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1364 MDB_dbx *me_dbxs; /**< array of static DB info */
1365 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1366 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1367 pthread_key_t me_txkey; /**< thread-key for readers */
1368 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1369 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1370 # define me_pglast me_pgstate.mf_pglast
1371 # define me_pghead me_pgstate.mf_pghead
1372 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1373 /** IDL of pages that became unused in a write txn */
1374 MDB_IDL me_free_pgs;
1375 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1376 MDB_ID2L me_dirty_list;
1377 /** Max number of freelist items that can fit in a single overflow page */
1379 /** Max size of a node on a page */
1380 unsigned int me_nodemax;
1381 #if !(MDB_MAXKEYSIZE)
1382 unsigned int me_maxkey; /**< max size of a key */
1384 int me_live_reader; /**< have liveness lock in reader table */
1386 int me_pidquery; /**< Used in OpenProcess */
1388 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1389 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1390 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1392 mdb_mutex_t me_rmutex;
1393 mdb_mutex_t me_wmutex;
1396 MDB_ID3L me_rpages; /**< like #mt_rpages, but global to env */
1397 mdb_mutex_t me_rpmutex; /**< control access to #me_rpages */
1398 #define MDB_ERPAGE_SIZE 16384
1399 #define MDB_ERPAGE_MAX (MDB_ERPAGE_SIZE-1)
1400 unsigned int me_rpcheck;
1402 void *me_userctx; /**< User-settable context */
1403 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1406 /** Nested transaction */
1407 typedef struct MDB_ntxn {
1408 MDB_txn mnt_txn; /**< the transaction */
1409 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1412 /** max number of pages to commit in one writev() call */
1413 #define MDB_COMMIT_PAGES 64
1414 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1415 #undef MDB_COMMIT_PAGES
1416 #define MDB_COMMIT_PAGES IOV_MAX
1419 /** max bytes to write in one call */
1420 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1422 /** Check \b txn and \b dbi arguments to a function */
1423 #define TXN_DBI_EXIST(txn, dbi, validity) \
1424 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1426 /** Check for misused \b dbi handles */
1427 #define TXN_DBI_CHANGED(txn, dbi) \
1428 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1430 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1431 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1432 static int mdb_page_touch(MDB_cursor *mc);
1434 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1435 "reset-tmp", "fail-begin", "fail-beginchild"}
1437 /* mdb_txn_end operation number, for logging */
1438 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1439 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1441 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1442 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1443 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1444 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1445 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1447 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1448 static int mdb_page_search_root(MDB_cursor *mc,
1449 MDB_val *key, int modify);
1450 #define MDB_PS_MODIFY 1
1451 #define MDB_PS_ROOTONLY 2
1452 #define MDB_PS_FIRST 4
1453 #define MDB_PS_LAST 8
1454 static int mdb_page_search(MDB_cursor *mc,
1455 MDB_val *key, int flags);
1456 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1458 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1459 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1460 pgno_t newpgno, unsigned int nflags);
1462 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1463 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1464 static int mdb_env_write_meta(MDB_txn *txn);
1465 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1466 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1468 static void mdb_env_close0(MDB_env *env, int excl);
1470 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1471 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1472 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1473 static void mdb_node_del(MDB_cursor *mc, int ksize);
1474 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1475 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1476 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1477 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1478 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1480 static int mdb_rebalance(MDB_cursor *mc);
1481 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1483 static void mdb_cursor_pop(MDB_cursor *mc);
1484 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1486 static int mdb_cursor_del0(MDB_cursor *mc);
1487 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1488 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1489 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1490 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1491 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1493 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1494 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1496 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1497 static void mdb_xcursor_init0(MDB_cursor *mc);
1498 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1499 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1501 static int mdb_drop0(MDB_cursor *mc, int subs);
1502 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1503 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1506 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1509 /** Compare two items pointing at size_t's of unknown alignment. */
1510 #ifdef MISALIGNED_OK
1511 # define mdb_cmp_clong mdb_cmp_long
1513 # define mdb_cmp_clong mdb_cmp_cint
1517 static SECURITY_DESCRIPTOR mdb_null_sd;
1518 static SECURITY_ATTRIBUTES mdb_all_sa;
1519 static int mdb_sec_inited;
1521 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize);
1524 /** Return the library version info. */
1526 mdb_version(int *major, int *minor, int *patch)
1528 if (major) *major = MDB_VERSION_MAJOR;
1529 if (minor) *minor = MDB_VERSION_MINOR;
1530 if (patch) *patch = MDB_VERSION_PATCH;
1531 return MDB_VERSION_STRING;
1534 /** Table of descriptions for LMDB @ref errors */
1535 static char *const mdb_errstr[] = {
1536 "MDB_KEYEXIST: Key/data pair already exists",
1537 "MDB_NOTFOUND: No matching key/data pair found",
1538 "MDB_PAGE_NOTFOUND: Requested page not found",
1539 "MDB_CORRUPTED: Located page was wrong type",
1540 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1541 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1542 "MDB_INVALID: File is not an LMDB file",
1543 "MDB_MAP_FULL: Environment mapsize limit reached",
1544 "MDB_DBS_FULL: Environment maxdbs limit reached",
1545 "MDB_READERS_FULL: Environment maxreaders limit reached",
1546 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1547 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1548 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1549 "MDB_PAGE_FULL: Internal error - page has no more space",
1550 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1551 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1552 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1553 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1554 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1555 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1559 mdb_strerror(int err)
1562 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1563 * This works as long as no function between the call to mdb_strerror
1564 * and the actual use of the message uses more than 4K of stack.
1567 char buf[1024], *ptr = buf;
1571 return ("Successful return: 0");
1573 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1574 i = err - MDB_KEYEXIST;
1575 return mdb_errstr[i];
1579 /* These are the C-runtime error codes we use. The comment indicates
1580 * their numeric value, and the Win32 error they would correspond to
1581 * if the error actually came from a Win32 API. A major mess, we should
1582 * have used LMDB-specific error codes for everything.
1585 case ENOENT: /* 2, FILE_NOT_FOUND */
1586 case EIO: /* 5, ACCESS_DENIED */
1587 case ENOMEM: /* 12, INVALID_ACCESS */
1588 case EACCES: /* 13, INVALID_DATA */
1589 case EBUSY: /* 16, CURRENT_DIRECTORY */
1590 case EINVAL: /* 22, BAD_COMMAND */
1591 case ENOSPC: /* 28, OUT_OF_PAPER */
1592 return strerror(err);
1597 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1598 FORMAT_MESSAGE_IGNORE_INSERTS,
1599 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1602 return strerror(err);
1606 /** assert(3) variant in cursor context */
1607 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1608 /** assert(3) variant in transaction context */
1609 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1610 /** assert(3) variant in environment context */
1611 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1614 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1615 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1618 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1619 const char *func, const char *file, int line)
1622 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1623 file, line, expr_txt, func);
1624 if (env->me_assert_func)
1625 env->me_assert_func(env, buf);
1626 fprintf(stderr, "%s\n", buf);
1630 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1634 /** Return the page number of \b mp which may be sub-page, for debug output */
1636 mdb_dbg_pgno(MDB_page *mp)
1639 COPY_PGNO(ret, mp->mp_pgno);
1643 /** Display a key in hexadecimal and return the address of the result.
1644 * @param[in] key the key to display
1645 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1646 * @return The key in hexadecimal form.
1649 mdb_dkey(MDB_val *key, char *buf)
1652 unsigned char *c = key->mv_data;
1658 if (key->mv_size > DKBUF_MAXKEYSIZE)
1659 return "MDB_MAXKEYSIZE";
1660 /* may want to make this a dynamic check: if the key is mostly
1661 * printable characters, print it as-is instead of converting to hex.
1665 for (i=0; i<key->mv_size; i++)
1666 ptr += sprintf(ptr, "%02x", *c++);
1668 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1674 mdb_leafnode_type(MDB_node *n)
1676 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1677 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1678 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1681 /** Display all the keys in the page. */
1683 mdb_page_list(MDB_page *mp)
1685 pgno_t pgno = mdb_dbg_pgno(mp);
1686 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1688 unsigned int i, nkeys, nsize, total = 0;
1692 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1693 case P_BRANCH: type = "Branch page"; break;
1694 case P_LEAF: type = "Leaf page"; break;
1695 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1696 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1697 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1699 fprintf(stderr, "Overflow page %"Y"u pages %u%s\n",
1700 pgno, mp->mp_pages, state);
1703 fprintf(stderr, "Meta-page %"Y"u txnid %"Y"u\n",
1704 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1707 fprintf(stderr, "Bad page %"Y"u flags 0x%u\n", pgno, mp->mp_flags);
1711 nkeys = NUMKEYS(mp);
1712 fprintf(stderr, "%s %"Y"u numkeys %d%s\n", type, pgno, nkeys, state);
1714 for (i=0; i<nkeys; i++) {
1715 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1716 key.mv_size = nsize = mp->mp_pad;
1717 key.mv_data = LEAF2KEY(mp, i, nsize);
1719 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1722 node = NODEPTR(mp, i);
1723 key.mv_size = node->mn_ksize;
1724 key.mv_data = node->mn_data;
1725 nsize = NODESIZE + key.mv_size;
1726 if (IS_BRANCH(mp)) {
1727 fprintf(stderr, "key %d: page %"Y"u, %s\n", i, NODEPGNO(node),
1731 if (F_ISSET(node->mn_flags, F_BIGDATA))
1732 nsize += sizeof(pgno_t);
1734 nsize += NODEDSZ(node);
1736 nsize += sizeof(indx_t);
1737 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1738 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1740 total = EVEN(total);
1742 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1743 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1747 mdb_cursor_chk(MDB_cursor *mc)
1753 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1754 for (i=0; i<mc->mc_top; i++) {
1756 node = NODEPTR(mp, mc->mc_ki[i]);
1757 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1760 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1762 if (mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
1763 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1764 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1765 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1773 /** Count all the pages in each DB and in the freelist
1774 * and make sure it matches the actual number of pages
1776 * All named DBs must be open for a correct count.
1778 static void mdb_audit(MDB_txn *txn)
1782 MDB_ID freecount, count;
1787 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1788 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1789 freecount += *(MDB_ID *)data.mv_data;
1790 mdb_tassert(txn, rc == MDB_NOTFOUND);
1793 for (i = 0; i<txn->mt_numdbs; i++) {
1795 if (!(txn->mt_dbflags[i] & DB_VALID))
1797 mdb_cursor_init(&mc, txn, i, &mx);
1798 if (txn->mt_dbs[i].md_root == P_INVALID)
1800 count += txn->mt_dbs[i].md_branch_pages +
1801 txn->mt_dbs[i].md_leaf_pages +
1802 txn->mt_dbs[i].md_overflow_pages;
1803 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1804 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1805 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1808 mp = mc.mc_pg[mc.mc_top];
1809 for (j=0; j<NUMKEYS(mp); j++) {
1810 MDB_node *leaf = NODEPTR(mp, j);
1811 if (leaf->mn_flags & F_SUBDATA) {
1813 memcpy(&db, NODEDATA(leaf), sizeof(db));
1814 count += db.md_branch_pages + db.md_leaf_pages +
1815 db.md_overflow_pages;
1819 mdb_tassert(txn, rc == MDB_NOTFOUND);
1822 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1823 fprintf(stderr, "audit: %"Y"u freecount: %"Y"u count: %"Y"u total: %"Y"u next_pgno: %"Y"u\n",
1824 txn->mt_txnid, freecount, count+NUM_METAS,
1825 freecount+count+NUM_METAS, txn->mt_next_pgno);
1831 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1833 return txn->mt_dbxs[dbi].md_cmp(a, b);
1837 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1839 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1840 #if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
1841 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(mdb_size_t))
1842 dcmp = mdb_cmp_clong;
1847 /** Allocate memory for a page.
1848 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1851 mdb_page_malloc(MDB_txn *txn, unsigned num)
1853 MDB_env *env = txn->mt_env;
1854 MDB_page *ret = env->me_dpages;
1855 size_t psize = env->me_psize, sz = psize, off;
1856 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1857 * For a single page alloc, we init everything after the page header.
1858 * For multi-page, we init the final page; if the caller needed that
1859 * many pages they will be filling in at least up to the last page.
1863 VGMEMP_ALLOC(env, ret, sz);
1864 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1865 env->me_dpages = ret->mp_next;
1868 psize -= off = PAGEHDRSZ;
1873 if ((ret = malloc(sz)) != NULL) {
1874 VGMEMP_ALLOC(env, ret, sz);
1875 if (!(env->me_flags & MDB_NOMEMINIT)) {
1876 memset((char *)ret + off, 0, psize);
1880 txn->mt_flags |= MDB_TXN_ERROR;
1884 /** Free a single page.
1885 * Saves single pages to a list, for future reuse.
1886 * (This is not used for multi-page overflow pages.)
1889 mdb_page_free(MDB_env *env, MDB_page *mp)
1891 mp->mp_next = env->me_dpages;
1892 VGMEMP_FREE(env, mp);
1893 env->me_dpages = mp;
1896 /** Free a dirty page */
1898 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1900 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1901 mdb_page_free(env, dp);
1903 /* large pages just get freed directly */
1904 VGMEMP_FREE(env, dp);
1909 /** Return all dirty pages to dpage list */
1911 mdb_dlist_free(MDB_txn *txn)
1913 MDB_env *env = txn->mt_env;
1914 MDB_ID2L dl = txn->mt_u.dirty_list;
1915 unsigned i, n = dl[0].mid;
1917 for (i = 1; i <= n; i++) {
1918 mdb_dpage_free(env, dl[i].mptr);
1925 mdb_page_unref(MDB_txn *txn, MDB_page *mp)
1928 MDB_ID3L tl = txn->mt_rpages;
1930 if (mp->mp_flags & (P_SUBP|P_DIRTY))
1932 rem = mp->mp_pgno & (MDB_RPAGE_CHUNK-1);
1933 pgno = mp->mp_pgno ^ rem;
1934 x = mdb_mid3l_search(tl, pgno);
1935 if (x != tl[0].mid && tl[x+1].mid == mp->mp_pgno)
1940 #define MDB_PAGE_UNREF(txn, mp) mdb_page_unref(txn, mp)
1943 mdb_cursor_unref(MDB_cursor *mc)
1946 if (!mc->mc_pg[0] || IS_SUBP(mc->mc_pg[0]))
1948 for (i=0; i<mc->mc_snum; i++)
1949 mdb_page_unref(mc->mc_txn, mc->mc_pg[i]);
1950 mc->mc_snum = mc->mc_top = 0;
1951 mc->mc_pg[0] = NULL;
1952 mc->mc_flags &= ~C_INITIALIZED;
1955 #define MDB_PAGE_UNREF(txn, mp)
1956 #endif /* MDB_VL32 */
1958 /** Loosen or free a single page.
1959 * Saves single pages to a list for future reuse
1960 * in this same txn. It has been pulled from the freeDB
1961 * and already resides on the dirty list, but has been
1962 * deleted. Use these pages first before pulling again
1965 * If the page wasn't dirtied in this txn, just add it
1966 * to this txn's free list.
1969 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1972 pgno_t pgno = mp->mp_pgno;
1973 MDB_txn *txn = mc->mc_txn;
1975 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1976 if (txn->mt_parent) {
1977 MDB_ID2 *dl = txn->mt_u.dirty_list;
1978 /* If txn has a parent, make sure the page is in our
1982 unsigned x = mdb_mid2l_search(dl, pgno);
1983 if (x <= dl[0].mid && dl[x].mid == pgno) {
1984 if (mp != dl[x].mptr) { /* bad cursor? */
1985 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1986 txn->mt_flags |= MDB_TXN_ERROR;
1987 return MDB_CORRUPTED;
1994 /* no parent txn, so it's just ours */
1999 DPRINTF(("loosen db %d page %"Y"u", DDBI(mc),
2001 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
2002 txn->mt_loose_pgs = mp;
2003 txn->mt_loose_count++;
2004 mp->mp_flags |= P_LOOSE;
2006 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
2014 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
2015 * @param[in] mc A cursor handle for the current operation.
2016 * @param[in] pflags Flags of the pages to update:
2017 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
2018 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
2019 * @return 0 on success, non-zero on failure.
2022 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
2024 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
2025 MDB_txn *txn = mc->mc_txn;
2026 MDB_cursor *m3, *m0 = mc;
2031 int rc = MDB_SUCCESS, level;
2033 /* Mark pages seen by cursors */
2034 if (mc->mc_flags & C_UNTRACK)
2035 mc = NULL; /* will find mc in mt_cursors */
2036 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
2037 for (; mc; mc=mc->mc_next) {
2038 if (!(mc->mc_flags & C_INITIALIZED))
2040 for (m3 = mc;; m3 = &mx->mx_cursor) {
2042 for (j=0; j<m3->mc_snum; j++) {
2044 if ((mp->mp_flags & Mask) == pflags)
2045 mp->mp_flags ^= P_KEEP;
2047 mx = m3->mc_xcursor;
2048 /* Proceed to mx if it is at a sub-database */
2049 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
2051 if (! (mp && (mp->mp_flags & P_LEAF)))
2053 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
2054 if (!(leaf->mn_flags & F_SUBDATA))
2063 /* Mark dirty root pages */
2064 for (i=0; i<txn->mt_numdbs; i++) {
2065 if (txn->mt_dbflags[i] & DB_DIRTY) {
2066 pgno_t pgno = txn->mt_dbs[i].md_root;
2067 if (pgno == P_INVALID)
2069 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
2071 if ((dp->mp_flags & Mask) == pflags && level <= 1)
2072 dp->mp_flags ^= P_KEEP;
2080 static int mdb_page_flush(MDB_txn *txn, int keep);
2082 /** Spill pages from the dirty list back to disk.
2083 * This is intended to prevent running into #MDB_TXN_FULL situations,
2084 * but note that they may still occur in a few cases:
2085 * 1) our estimate of the txn size could be too small. Currently this
2086 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
2087 * 2) child txns may run out of space if their parents dirtied a
2088 * lot of pages and never spilled them. TODO: we probably should do
2089 * a preemptive spill during #mdb_txn_begin() of a child txn, if
2090 * the parent's dirty_room is below a given threshold.
2092 * Otherwise, if not using nested txns, it is expected that apps will
2093 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
2094 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
2095 * If the txn never references them again, they can be left alone.
2096 * If the txn only reads them, they can be used without any fuss.
2097 * If the txn writes them again, they can be dirtied immediately without
2098 * going thru all of the work of #mdb_page_touch(). Such references are
2099 * handled by #mdb_page_unspill().
2101 * Also note, we never spill DB root pages, nor pages of active cursors,
2102 * because we'll need these back again soon anyway. And in nested txns,
2103 * we can't spill a page in a child txn if it was already spilled in a
2104 * parent txn. That would alter the parent txns' data even though
2105 * the child hasn't committed yet, and we'd have no way to undo it if
2106 * the child aborted.
2108 * @param[in] m0 cursor A cursor handle identifying the transaction and
2109 * database for which we are checking space.
2110 * @param[in] key For a put operation, the key being stored.
2111 * @param[in] data For a put operation, the data being stored.
2112 * @return 0 on success, non-zero on failure.
2115 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
2117 MDB_txn *txn = m0->mc_txn;
2119 MDB_ID2L dl = txn->mt_u.dirty_list;
2120 unsigned int i, j, need;
2123 if (m0->mc_flags & C_SUB)
2126 /* Estimate how much space this op will take */
2127 i = m0->mc_db->md_depth;
2128 /* Named DBs also dirty the main DB */
2129 if (m0->mc_dbi >= CORE_DBS)
2130 i += txn->mt_dbs[MAIN_DBI].md_depth;
2131 /* For puts, roughly factor in the key+data size */
2133 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2134 i += i; /* double it for good measure */
2137 if (txn->mt_dirty_room > i)
2140 if (!txn->mt_spill_pgs) {
2141 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2142 if (!txn->mt_spill_pgs)
2145 /* purge deleted slots */
2146 MDB_IDL sl = txn->mt_spill_pgs;
2147 unsigned int num = sl[0];
2149 for (i=1; i<=num; i++) {
2156 /* Preserve pages which may soon be dirtied again */
2157 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2160 /* Less aggressive spill - we originally spilled the entire dirty list,
2161 * with a few exceptions for cursor pages and DB root pages. But this
2162 * turns out to be a lot of wasted effort because in a large txn many
2163 * of those pages will need to be used again. So now we spill only 1/8th
2164 * of the dirty pages. Testing revealed this to be a good tradeoff,
2165 * better than 1/2, 1/4, or 1/10.
2167 if (need < MDB_IDL_UM_MAX / 8)
2168 need = MDB_IDL_UM_MAX / 8;
2170 /* Save the page IDs of all the pages we're flushing */
2171 /* flush from the tail forward, this saves a lot of shifting later on. */
2172 for (i=dl[0].mid; i && need; i--) {
2173 MDB_ID pn = dl[i].mid << 1;
2175 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2177 /* Can't spill twice, make sure it's not already in a parent's
2180 if (txn->mt_parent) {
2182 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2183 if (tx2->mt_spill_pgs) {
2184 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2185 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2186 dp->mp_flags |= P_KEEP;
2194 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2198 mdb_midl_sort(txn->mt_spill_pgs);
2200 /* Flush the spilled part of dirty list */
2201 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2204 /* Reset any dirty pages we kept that page_flush didn't see */
2205 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2208 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2212 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2214 mdb_find_oldest(MDB_txn *txn)
2217 txnid_t mr, oldest = txn->mt_txnid - 1;
2218 if (txn->mt_env->me_txns) {
2219 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2220 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2231 /** Add a page to the txn's dirty list */
2233 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2236 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2238 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2239 insert = mdb_mid2l_append;
2241 insert = mdb_mid2l_insert;
2243 mid.mid = mp->mp_pgno;
2245 rc = insert(txn->mt_u.dirty_list, &mid);
2246 mdb_tassert(txn, rc == 0);
2247 txn->mt_dirty_room--;
2250 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2251 * me_pghead and mt_next_pgno.
2253 * If there are free pages available from older transactions, they
2254 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2255 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2256 * and move me_pglast to say which records were consumed. Only this
2257 * function can create me_pghead and move me_pglast/mt_next_pgno.
2258 * @param[in] mc cursor A cursor handle identifying the transaction and
2259 * database for which we are allocating.
2260 * @param[in] num the number of pages to allocate.
2261 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2262 * will always be satisfied by a single contiguous chunk of memory.
2263 * @return 0 on success, non-zero on failure.
2266 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2268 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2269 /* Get at most <Max_retries> more freeDB records once me_pghead
2270 * has enough pages. If not enough, use new pages from the map.
2271 * If <Paranoid> and mc is updating the freeDB, only get new
2272 * records if me_pghead is empty. Then the freelist cannot play
2273 * catch-up with itself by growing while trying to save it.
2275 enum { Paranoid = 1, Max_retries = 500 };
2277 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2279 int rc, retry = num * 60;
2280 MDB_txn *txn = mc->mc_txn;
2281 MDB_env *env = txn->mt_env;
2282 pgno_t pgno, *mop = env->me_pghead;
2283 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2285 txnid_t oldest = 0, last;
2290 /* If there are any loose pages, just use them */
2291 if (num == 1 && txn->mt_loose_pgs) {
2292 np = txn->mt_loose_pgs;
2293 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2294 txn->mt_loose_count--;
2295 DPRINTF(("db %d use loose page %"Y"u", DDBI(mc),
2303 /* If our dirty list is already full, we can't do anything */
2304 if (txn->mt_dirty_room == 0) {
2309 for (op = MDB_FIRST;; op = MDB_NEXT) {
2314 /* Seek a big enough contiguous page range. Prefer
2315 * pages at the tail, just truncating the list.
2321 if (mop[i-n2] == pgno+n2)
2328 if (op == MDB_FIRST) { /* 1st iteration */
2329 /* Prepare to fetch more and coalesce */
2330 last = env->me_pglast;
2331 oldest = env->me_pgoldest;
2332 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2335 key.mv_data = &last; /* will look up last+1 */
2336 key.mv_size = sizeof(last);
2338 if (Paranoid && mc->mc_dbi == FREE_DBI)
2341 if (Paranoid && retry < 0 && mop_len)
2345 /* Do not fetch more if the record will be too recent */
2346 if (oldest <= last) {
2348 oldest = mdb_find_oldest(txn);
2349 env->me_pgoldest = oldest;
2355 rc = mdb_cursor_get(&m2, &key, NULL, op);
2357 if (rc == MDB_NOTFOUND)
2361 last = *(txnid_t*)key.mv_data;
2362 if (oldest <= last) {
2364 oldest = mdb_find_oldest(txn);
2365 env->me_pgoldest = oldest;
2371 np = m2.mc_pg[m2.mc_top];
2372 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2373 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2376 idl = (MDB_ID *) data.mv_data;
2379 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2384 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2386 mop = env->me_pghead;
2388 env->me_pglast = last;
2390 DPRINTF(("IDL read txn %"Y"u root %"Y"u num %u",
2391 last, txn->mt_dbs[FREE_DBI].md_root, i));
2393 DPRINTF(("IDL %"Y"u", idl[j]));
2395 /* Merge in descending sorted order */
2396 mdb_midl_xmerge(mop, idl);
2400 /* Use new pages from the map when nothing suitable in the freeDB */
2402 pgno = txn->mt_next_pgno;
2403 if (pgno + num >= env->me_maxpg) {
2404 DPUTS("DB size maxed out");
2409 if (env->me_flags & MDB_WRITEMAP) {
2411 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
2412 p = VirtualAlloc(p, env->me_psize * num, MEM_COMMIT,
2413 (env->me_flags & MDB_WRITEMAP) ? PAGE_READWRITE:
2416 DPUTS("VirtualAlloc failed");
2424 if (env->me_flags & MDB_WRITEMAP) {
2425 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2427 if (!(np = mdb_page_malloc(txn, num))) {
2433 mop[0] = mop_len -= num;
2434 /* Move any stragglers down */
2435 for (j = i-num; j < mop_len; )
2436 mop[++j] = mop[++i];
2438 txn->mt_next_pgno = pgno + num;
2441 mdb_page_dirty(txn, np);
2447 txn->mt_flags |= MDB_TXN_ERROR;
2451 /** Copy the used portions of a non-overflow page.
2452 * @param[in] dst page to copy into
2453 * @param[in] src page to copy from
2454 * @param[in] psize size of a page
2457 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2459 enum { Align = sizeof(pgno_t) };
2460 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2462 /* If page isn't full, just copy the used portion. Adjust
2463 * alignment so memcpy may copy words instead of bytes.
2465 if ((unused &= -Align) && !IS_LEAF2(src)) {
2466 upper = (upper + PAGEBASE) & -Align;
2467 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2468 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2471 memcpy(dst, src, psize - unused);
2475 /** Pull a page off the txn's spill list, if present.
2476 * If a page being referenced was spilled to disk in this txn, bring
2477 * it back and make it dirty/writable again.
2478 * @param[in] txn the transaction handle.
2479 * @param[in] mp the page being referenced. It must not be dirty.
2480 * @param[out] ret the writable page, if any. ret is unchanged if
2481 * mp wasn't spilled.
2484 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2486 MDB_env *env = txn->mt_env;
2489 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2491 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2492 if (!tx2->mt_spill_pgs)
2494 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2495 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2498 if (txn->mt_dirty_room == 0)
2499 return MDB_TXN_FULL;
2500 if (IS_OVERFLOW(mp))
2504 if (env->me_flags & MDB_WRITEMAP) {
2507 np = mdb_page_malloc(txn, num);
2511 memcpy(np, mp, num * env->me_psize);
2513 mdb_page_copy(np, mp, env->me_psize);
2516 /* If in current txn, this page is no longer spilled.
2517 * If it happens to be the last page, truncate the spill list.
2518 * Otherwise mark it as deleted by setting the LSB.
2520 if (x == txn->mt_spill_pgs[0])
2521 txn->mt_spill_pgs[0]--;
2523 txn->mt_spill_pgs[x] |= 1;
2524 } /* otherwise, if belonging to a parent txn, the
2525 * page remains spilled until child commits
2528 mdb_page_dirty(txn, np);
2529 np->mp_flags |= P_DIRTY;
2537 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2538 * @param[in] mc cursor pointing to the page to be touched
2539 * @return 0 on success, non-zero on failure.
2542 mdb_page_touch(MDB_cursor *mc)
2544 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2545 MDB_txn *txn = mc->mc_txn;
2546 MDB_cursor *m2, *m3;
2550 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2551 if (txn->mt_flags & MDB_TXN_SPILLS) {
2553 rc = mdb_page_unspill(txn, mp, &np);
2559 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2560 (rc = mdb_page_alloc(mc, 1, &np)))
2563 DPRINTF(("touched db %d page %"Y"u -> %"Y"u", DDBI(mc),
2564 mp->mp_pgno, pgno));
2565 mdb_cassert(mc, mp->mp_pgno != pgno);
2566 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2567 /* Update the parent page, if any, to point to the new page */
2569 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2570 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2571 SETPGNO(node, pgno);
2573 mc->mc_db->md_root = pgno;
2575 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2576 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2578 /* If txn has a parent, make sure the page is in our
2582 unsigned x = mdb_mid2l_search(dl, pgno);
2583 if (x <= dl[0].mid && dl[x].mid == pgno) {
2584 if (mp != dl[x].mptr) { /* bad cursor? */
2585 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2586 txn->mt_flags |= MDB_TXN_ERROR;
2587 return MDB_CORRUPTED;
2592 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2594 np = mdb_page_malloc(txn, 1);
2599 rc = mdb_mid2l_insert(dl, &mid);
2600 mdb_cassert(mc, rc == 0);
2605 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2607 np->mp_flags |= P_DIRTY;
2610 /* Adjust cursors pointing to mp */
2611 mc->mc_pg[mc->mc_top] = np;
2612 m2 = txn->mt_cursors[mc->mc_dbi];
2613 if (mc->mc_flags & C_SUB) {
2614 for (; m2; m2=m2->mc_next) {
2615 m3 = &m2->mc_xcursor->mx_cursor;
2616 if (m3->mc_snum < mc->mc_snum) continue;
2617 if (m3->mc_pg[mc->mc_top] == mp)
2618 m3->mc_pg[mc->mc_top] = np;
2621 for (; m2; m2=m2->mc_next) {
2622 if (m2->mc_snum < mc->mc_snum) continue;
2623 if (m2 == mc) continue;
2624 if (m2->mc_pg[mc->mc_top] == mp) {
2625 m2->mc_pg[mc->mc_top] = np;
2626 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2628 (m2->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
2630 MDB_node *leaf = NODEPTR(np, m2->mc_ki[mc->mc_top]);
2631 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
2632 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2637 MDB_PAGE_UNREF(mc->mc_txn, mp);
2641 txn->mt_flags |= MDB_TXN_ERROR;
2646 mdb_env_sync0(MDB_env *env, int force, pgno_t numpgs)
2649 if (env->me_flags & MDB_RDONLY)
2651 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2652 if (env->me_flags & MDB_WRITEMAP) {
2653 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2654 ? MS_ASYNC : MS_SYNC;
2655 if (MDB_MSYNC(env->me_map, env->me_psize * numpgs, flags))
2658 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2662 #ifdef BROKEN_FDATASYNC
2663 if (env->me_flags & MDB_FSYNCONLY) {
2664 if (fsync(env->me_fd))
2668 if (MDB_FDATASYNC(env->me_fd))
2676 mdb_env_sync(MDB_env *env, int force)
2678 MDB_meta *m = mdb_env_pick_meta(env);
2679 return mdb_env_sync0(env, force, m->mm_last_pg+1);
2682 /** Back up parent txn's cursors, then grab the originals for tracking */
2684 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2686 MDB_cursor *mc, *bk;
2691 for (i = src->mt_numdbs; --i >= 0; ) {
2692 if ((mc = src->mt_cursors[i]) != NULL) {
2693 size = sizeof(MDB_cursor);
2695 size += sizeof(MDB_xcursor);
2696 for (; mc; mc = bk->mc_next) {
2702 mc->mc_db = &dst->mt_dbs[i];
2703 /* Kill pointers into src to reduce abuse: The
2704 * user may not use mc until dst ends. But we need a valid
2705 * txn pointer here for cursor fixups to keep working.
2708 mc->mc_dbflag = &dst->mt_dbflags[i];
2709 if ((mx = mc->mc_xcursor) != NULL) {
2710 *(MDB_xcursor *)(bk+1) = *mx;
2711 mx->mx_cursor.mc_txn = dst;
2713 mc->mc_next = dst->mt_cursors[i];
2714 dst->mt_cursors[i] = mc;
2721 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2722 * @param[in] txn the transaction handle.
2723 * @param[in] merge true to keep changes to parent cursors, false to revert.
2724 * @return 0 on success, non-zero on failure.
2727 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2729 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2733 for (i = txn->mt_numdbs; --i >= 0; ) {
2734 for (mc = cursors[i]; mc; mc = next) {
2736 if ((bk = mc->mc_backup) != NULL) {
2738 /* Commit changes to parent txn */
2739 mc->mc_next = bk->mc_next;
2740 mc->mc_backup = bk->mc_backup;
2741 mc->mc_txn = bk->mc_txn;
2742 mc->mc_db = bk->mc_db;
2743 mc->mc_dbflag = bk->mc_dbflag;
2744 if ((mx = mc->mc_xcursor) != NULL)
2745 mx->mx_cursor.mc_txn = bk->mc_txn;
2747 /* Abort nested txn */
2749 if ((mx = mc->mc_xcursor) != NULL)
2750 *mx = *(MDB_xcursor *)(bk+1);
2754 /* Only malloced cursors are permanently tracked. */
2761 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2767 Pidset = F_SETLK, Pidcheck = F_GETLK
2771 /** Set or check a pid lock. Set returns 0 on success.
2772 * Check returns 0 if the process is certainly dead, nonzero if it may
2773 * be alive (the lock exists or an error happened so we do not know).
2775 * On Windows Pidset is a no-op, we merely check for the existence
2776 * of the process with the given pid. On POSIX we use a single byte
2777 * lock on the lockfile, set at an offset equal to the pid.
2780 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2782 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2785 if (op == Pidcheck) {
2786 h = OpenProcess(env->me_pidquery, FALSE, pid);
2787 /* No documented "no such process" code, but other program use this: */
2789 return ErrCode() != ERROR_INVALID_PARAMETER;
2790 /* A process exists until all handles to it close. Has it exited? */
2791 ret = WaitForSingleObject(h, 0) != 0;
2798 struct flock lock_info;
2799 memset(&lock_info, 0, sizeof(lock_info));
2800 lock_info.l_type = F_WRLCK;
2801 lock_info.l_whence = SEEK_SET;
2802 lock_info.l_start = pid;
2803 lock_info.l_len = 1;
2804 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2805 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2807 } else if ((rc = ErrCode()) == EINTR) {
2815 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2816 * @param[in] txn the transaction handle to initialize
2817 * @return 0 on success, non-zero on failure.
2820 mdb_txn_renew0(MDB_txn *txn)
2822 MDB_env *env = txn->mt_env;
2823 MDB_txninfo *ti = env->me_txns;
2825 unsigned int i, nr, flags = txn->mt_flags;
2827 int rc, new_notls = 0;
2829 if ((flags &= MDB_TXN_RDONLY) != 0) {
2831 meta = mdb_env_pick_meta(env);
2832 txn->mt_txnid = meta->mm_txnid;
2833 txn->mt_u.reader = NULL;
2835 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2836 pthread_getspecific(env->me_txkey);
2838 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2839 return MDB_BAD_RSLOT;
2841 MDB_PID_T pid = env->me_pid;
2842 MDB_THR_T tid = pthread_self();
2843 mdb_mutexref_t rmutex = env->me_rmutex;
2845 if (!env->me_live_reader) {
2846 rc = mdb_reader_pid(env, Pidset, pid);
2849 env->me_live_reader = 1;
2852 if (LOCK_MUTEX(rc, env, rmutex))
2854 nr = ti->mti_numreaders;
2855 for (i=0; i<nr; i++)
2856 if (ti->mti_readers[i].mr_pid == 0)
2858 if (i == env->me_maxreaders) {
2859 UNLOCK_MUTEX(rmutex);
2860 return MDB_READERS_FULL;
2862 r = &ti->mti_readers[i];
2863 /* Claim the reader slot, carefully since other code
2864 * uses the reader table un-mutexed: First reset the
2865 * slot, next publish it in mti_numreaders. After
2866 * that, it is safe for mdb_env_close() to touch it.
2867 * When it will be closed, we can finally claim it.
2870 r->mr_txnid = (txnid_t)-1;
2873 ti->mti_numreaders = ++nr;
2874 env->me_close_readers = nr;
2876 UNLOCK_MUTEX(rmutex);
2878 new_notls = (env->me_flags & MDB_NOTLS);
2879 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2884 do /* LY: Retry on a race, ITS#7970. */
2885 r->mr_txnid = ti->mti_txnid;
2886 while(r->mr_txnid != ti->mti_txnid);
2887 txn->mt_txnid = r->mr_txnid;
2888 txn->mt_u.reader = r;
2889 meta = env->me_metas[txn->mt_txnid & 1];
2893 /* Not yet touching txn == env->me_txn0, it may be active */
2895 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2897 txn->mt_txnid = ti->mti_txnid;
2898 meta = env->me_metas[txn->mt_txnid & 1];
2900 meta = mdb_env_pick_meta(env);
2901 txn->mt_txnid = meta->mm_txnid;
2905 if (txn->mt_txnid == mdb_debug_start)
2908 txn->mt_child = NULL;
2909 txn->mt_loose_pgs = NULL;
2910 txn->mt_loose_count = 0;
2911 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2912 txn->mt_u.dirty_list = env->me_dirty_list;
2913 txn->mt_u.dirty_list[0].mid = 0;
2914 txn->mt_free_pgs = env->me_free_pgs;
2915 txn->mt_free_pgs[0] = 0;
2916 txn->mt_spill_pgs = NULL;
2918 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2921 /* Copy the DB info and flags */
2922 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2924 /* Moved to here to avoid a data race in read TXNs */
2925 txn->mt_next_pgno = meta->mm_last_pg+1;
2927 txn->mt_last_pgno = txn->mt_next_pgno - 1;
2930 txn->mt_flags = flags;
2933 txn->mt_numdbs = env->me_numdbs;
2934 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2935 x = env->me_dbflags[i];
2936 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2937 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2939 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2940 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2942 if (env->me_flags & MDB_FATAL_ERROR) {
2943 DPUTS("environment had fatal error, must shutdown!");
2945 } else if (env->me_maxpg < txn->mt_next_pgno) {
2946 rc = MDB_MAP_RESIZED;
2950 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2955 mdb_txn_renew(MDB_txn *txn)
2959 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2962 rc = mdb_txn_renew0(txn);
2963 if (rc == MDB_SUCCESS) {
2964 DPRINTF(("renew txn %"Y"u%c %p on mdbenv %p, root page %"Y"u",
2965 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2966 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2972 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2976 int rc, size, tsize;
2978 flags &= MDB_TXN_BEGIN_FLAGS;
2979 flags |= env->me_flags & MDB_WRITEMAP;
2981 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2985 /* Nested transactions: Max 1 child, write txns only, no writemap */
2986 flags |= parent->mt_flags;
2987 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2988 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2990 /* Child txns save MDB_pgstate and use own copy of cursors */
2991 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2992 size += tsize = sizeof(MDB_ntxn);
2993 } else if (flags & MDB_RDONLY) {
2994 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2995 size += tsize = sizeof(MDB_txn);
2997 /* Reuse preallocated write txn. However, do not touch it until
2998 * mdb_txn_renew0() succeeds, since it currently may be active.
3003 if ((txn = calloc(1, size)) == NULL) {
3004 DPRINTF(("calloc: %s", strerror(errno)));
3009 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
3010 if (!txn->mt_rpages) {
3014 txn->mt_rpages[0].mid = 0;
3015 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
3018 txn->mt_dbxs = env->me_dbxs; /* static */
3019 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
3020 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
3021 txn->mt_flags = flags;
3026 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
3027 txn->mt_dbiseqs = parent->mt_dbiseqs;
3028 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
3029 if (!txn->mt_u.dirty_list ||
3030 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
3032 free(txn->mt_u.dirty_list);
3036 txn->mt_txnid = parent->mt_txnid;
3037 txn->mt_dirty_room = parent->mt_dirty_room;
3038 txn->mt_u.dirty_list[0].mid = 0;
3039 txn->mt_spill_pgs = NULL;
3040 txn->mt_next_pgno = parent->mt_next_pgno;
3041 parent->mt_flags |= MDB_TXN_HAS_CHILD;
3042 parent->mt_child = txn;
3043 txn->mt_parent = parent;
3044 txn->mt_numdbs = parent->mt_numdbs;
3046 txn->mt_rpages = parent->mt_rpages;
3048 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3049 /* Copy parent's mt_dbflags, but clear DB_NEW */
3050 for (i=0; i<txn->mt_numdbs; i++)
3051 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
3053 ntxn = (MDB_ntxn *)txn;
3054 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
3055 if (env->me_pghead) {
3056 size = MDB_IDL_SIZEOF(env->me_pghead);
3057 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
3059 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
3064 rc = mdb_cursor_shadow(parent, txn);
3066 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
3067 } else { /* MDB_RDONLY */
3068 txn->mt_dbiseqs = env->me_dbiseqs;
3070 rc = mdb_txn_renew0(txn);
3073 if (txn != env->me_txn0) {
3075 free(txn->mt_rpages);
3080 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
3082 DPRINTF(("begin txn %"Y"u%c %p on mdbenv %p, root page %"Y"u",
3083 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
3084 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
3091 mdb_txn_env(MDB_txn *txn)
3093 if(!txn) return NULL;
3098 mdb_txn_id(MDB_txn *txn)
3101 return txn->mt_txnid;
3104 /** Export or close DBI handles opened in this txn. */
3106 mdb_dbis_update(MDB_txn *txn, int keep)
3109 MDB_dbi n = txn->mt_numdbs;
3110 MDB_env *env = txn->mt_env;
3111 unsigned char *tdbflags = txn->mt_dbflags;
3113 for (i = n; --i >= CORE_DBS;) {
3114 if (tdbflags[i] & DB_NEW) {
3116 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
3118 char *ptr = env->me_dbxs[i].md_name.mv_data;
3120 env->me_dbxs[i].md_name.mv_data = NULL;
3121 env->me_dbxs[i].md_name.mv_size = 0;
3122 env->me_dbflags[i] = 0;
3123 env->me_dbiseqs[i]++;
3129 if (keep && env->me_numdbs < n)
3133 /** End a transaction, except successful commit of a nested transaction.
3134 * May be called twice for readonly txns: First reset it, then abort.
3135 * @param[in] txn the transaction handle to end
3136 * @param[in] mode why and how to end the transaction
3139 mdb_txn_end(MDB_txn *txn, unsigned mode)
3141 MDB_env *env = txn->mt_env;
3143 static const char *const names[] = MDB_END_NAMES;
3146 /* Export or close DBI handles opened in this txn */
3147 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
3149 DPRINTF(("%s txn %"Y"u%c %p on mdbenv %p, root page %"Y"u",
3150 names[mode & MDB_END_OPMASK],
3151 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3152 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
3154 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3155 if (txn->mt_u.reader) {
3156 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
3157 if (!(env->me_flags & MDB_NOTLS)) {
3158 txn->mt_u.reader = NULL; /* txn does not own reader */
3159 } else if (mode & MDB_END_SLOT) {
3160 txn->mt_u.reader->mr_pid = 0;
3161 txn->mt_u.reader = NULL;
3162 } /* else txn owns the slot until it does MDB_END_SLOT */
3164 txn->mt_numdbs = 0; /* prevent further DBI activity */
3165 txn->mt_flags |= MDB_TXN_FINISHED;
3167 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
3168 pgno_t *pghead = env->me_pghead;
3170 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
3171 mdb_cursors_close(txn, 0);
3172 if (!(env->me_flags & MDB_WRITEMAP)) {
3173 mdb_dlist_free(txn);
3177 txn->mt_flags = MDB_TXN_FINISHED;
3179 if (!txn->mt_parent) {
3180 mdb_midl_shrink(&txn->mt_free_pgs);
3181 env->me_free_pgs = txn->mt_free_pgs;
3183 env->me_pghead = NULL;
3187 mode = 0; /* txn == env->me_txn0, do not free() it */
3189 /* The writer mutex was locked in mdb_txn_begin. */
3191 UNLOCK_MUTEX(env->me_wmutex);
3193 txn->mt_parent->mt_child = NULL;
3194 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3195 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3196 mdb_midl_free(txn->mt_free_pgs);
3197 mdb_midl_free(txn->mt_spill_pgs);
3198 free(txn->mt_u.dirty_list);
3201 mdb_midl_free(pghead);
3204 if (!txn->mt_parent) {
3205 MDB_ID3L el = env->me_rpages, tl = txn->mt_rpages;
3206 unsigned i, x, n = tl[0].mid;
3207 LOCK_MUTEX0(env->me_rpmutex);
3208 for (i = 1; i <= n; i++) {
3209 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
3210 /* tmp overflow pages that we didn't share in env */
3211 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3213 x = mdb_mid3l_search(el, tl[i].mid);
3214 if (tl[i].mptr == el[x].mptr) {
3217 /* another tmp overflow page */
3218 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3222 UNLOCK_MUTEX(env->me_rpmutex);
3224 if (mode & MDB_END_FREE)
3228 if (mode & MDB_END_FREE)
3233 mdb_txn_reset(MDB_txn *txn)
3238 /* This call is only valid for read-only txns */
3239 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3242 mdb_txn_end(txn, MDB_END_RESET);
3246 mdb_txn_abort(MDB_txn *txn)
3252 mdb_txn_abort(txn->mt_child);
3254 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3257 /** Save the freelist as of this transaction to the freeDB.
3258 * This changes the freelist. Keep trying until it stabilizes.
3261 mdb_freelist_save(MDB_txn *txn)
3263 /* env->me_pghead[] can grow and shrink during this call.
3264 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3265 * Page numbers cannot disappear from txn->mt_free_pgs[].
3268 MDB_env *env = txn->mt_env;
3269 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3270 txnid_t pglast = 0, head_id = 0;
3271 pgno_t freecnt = 0, *free_pgs, *mop;
3272 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3274 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3276 if (env->me_pghead) {
3277 /* Make sure first page of freeDB is touched and on freelist */
3278 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3279 if (rc && rc != MDB_NOTFOUND)
3283 if (!env->me_pghead && txn->mt_loose_pgs) {
3284 /* Put loose page numbers in mt_free_pgs, since
3285 * we may be unable to return them to me_pghead.
3287 MDB_page *mp = txn->mt_loose_pgs;
3288 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3290 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3291 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3292 txn->mt_loose_pgs = NULL;
3293 txn->mt_loose_count = 0;
3296 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3297 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3298 ? SSIZE_MAX : maxfree_1pg;
3301 /* Come back here after each Put() in case freelist changed */
3306 /* If using records from freeDB which we have not yet
3307 * deleted, delete them and any we reserved for me_pghead.
3309 while (pglast < env->me_pglast) {
3310 rc = mdb_cursor_first(&mc, &key, NULL);
3313 pglast = head_id = *(txnid_t *)key.mv_data;
3314 total_room = head_room = 0;
3315 mdb_tassert(txn, pglast <= env->me_pglast);
3316 rc = mdb_cursor_del(&mc, 0);
3321 /* Save the IDL of pages freed by this txn, to a single record */
3322 if (freecnt < txn->mt_free_pgs[0]) {
3324 /* Make sure last page of freeDB is touched and on freelist */
3325 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3326 if (rc && rc != MDB_NOTFOUND)
3329 free_pgs = txn->mt_free_pgs;
3330 /* Write to last page of freeDB */
3331 key.mv_size = sizeof(txn->mt_txnid);
3332 key.mv_data = &txn->mt_txnid;
3334 freecnt = free_pgs[0];
3335 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3336 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3339 /* Retry if mt_free_pgs[] grew during the Put() */
3340 free_pgs = txn->mt_free_pgs;
3341 } while (freecnt < free_pgs[0]);
3342 mdb_midl_sort(free_pgs);
3343 memcpy(data.mv_data, free_pgs, data.mv_size);
3346 unsigned int i = free_pgs[0];
3347 DPRINTF(("IDL write txn %"Y"u root %"Y"u num %u",
3348 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3350 DPRINTF(("IDL %"Y"u", free_pgs[i]));
3356 mop = env->me_pghead;
3357 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3359 /* Reserve records for me_pghead[]. Split it if multi-page,
3360 * to avoid searching freeDB for a page range. Use keys in
3361 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3363 if (total_room >= mop_len) {
3364 if (total_room == mop_len || --more < 0)
3366 } else if (head_room >= maxfree_1pg && head_id > 1) {
3367 /* Keep current record (overflow page), add a new one */
3371 /* (Re)write {key = head_id, IDL length = head_room} */
3372 total_room -= head_room;
3373 head_room = mop_len - total_room;
3374 if (head_room > maxfree_1pg && head_id > 1) {
3375 /* Overflow multi-page for part of me_pghead */
3376 head_room /= head_id; /* amortize page sizes */
3377 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3378 } else if (head_room < 0) {
3379 /* Rare case, not bothering to delete this record */
3382 key.mv_size = sizeof(head_id);
3383 key.mv_data = &head_id;
3384 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3385 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3388 /* IDL is initially empty, zero out at least the length */
3389 pgs = (pgno_t *)data.mv_data;
3390 j = head_room > clean_limit ? head_room : 0;
3394 total_room += head_room;
3397 /* Return loose page numbers to me_pghead, though usually none are
3398 * left at this point. The pages themselves remain in dirty_list.
3400 if (txn->mt_loose_pgs) {
3401 MDB_page *mp = txn->mt_loose_pgs;
3402 unsigned count = txn->mt_loose_count;
3404 /* Room for loose pages + temp IDL with same */
3405 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3407 mop = env->me_pghead;
3408 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3409 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3410 loose[ ++count ] = mp->mp_pgno;
3412 mdb_midl_sort(loose);
3413 mdb_midl_xmerge(mop, loose);
3414 txn->mt_loose_pgs = NULL;
3415 txn->mt_loose_count = 0;
3419 /* Fill in the reserved me_pghead records */
3425 rc = mdb_cursor_first(&mc, &key, &data);
3426 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3427 txnid_t id = *(txnid_t *)key.mv_data;
3428 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3431 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3433 if (len > mop_len) {
3435 data.mv_size = (len + 1) * sizeof(MDB_ID);
3437 data.mv_data = mop -= len;
3440 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3442 if (rc || !(mop_len -= len))
3449 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3450 * @param[in] txn the transaction that's being committed
3451 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3452 * @return 0 on success, non-zero on failure.
3455 mdb_page_flush(MDB_txn *txn, int keep)
3457 MDB_env *env = txn->mt_env;
3458 MDB_ID2L dl = txn->mt_u.dirty_list;
3459 unsigned psize = env->me_psize, j;
3460 int i, pagecount = dl[0].mid, rc;
3464 MDB_page *dp = NULL;
3468 struct iovec iov[MDB_COMMIT_PAGES];
3469 ssize_t wsize = 0, wres;
3470 off_t wpos = 0, next_pos = 1; /* impossible pos, so pos != next_pos */
3476 if (env->me_flags & MDB_WRITEMAP) {
3477 /* Clear dirty flags */
3478 while (++i <= pagecount) {
3480 /* Don't flush this page yet */
3481 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3482 dp->mp_flags &= ~P_KEEP;
3486 dp->mp_flags &= ~P_DIRTY;
3491 /* Write the pages */
3493 if (++i <= pagecount) {
3495 /* Don't flush this page yet */
3496 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3497 dp->mp_flags &= ~P_KEEP;
3502 /* clear dirty flag */
3503 dp->mp_flags &= ~P_DIRTY;
3506 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3511 /* Windows actually supports scatter/gather I/O, but only on
3512 * unbuffered file handles. Since we're relying on the OS page
3513 * cache for all our data, that's self-defeating. So we just
3514 * write pages one at a time. We use the ov structure to set
3515 * the write offset, to at least save the overhead of a Seek
3518 DPRINTF(("committing page %"Z"u", pgno));
3519 memset(&ov, 0, sizeof(ov));
3520 ov.Offset = pos & 0xffffffff;
3521 ov.OffsetHigh = pos >> 16 >> 16;
3522 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3524 DPRINTF(("WriteFile: %d", rc));
3528 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3529 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3532 /* Write previous page(s) */
3533 #ifdef MDB_USE_PWRITEV
3534 wres = pwritev(env->me_fd, iov, n, wpos);
3537 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3540 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3544 DPRINTF(("lseek: %s", strerror(rc)));
3547 wres = writev(env->me_fd, iov, n);
3550 if (wres != wsize) {
3555 DPRINTF(("Write error: %s", strerror(rc)));
3557 rc = EIO; /* TODO: Use which error code? */
3558 DPUTS("short write, filesystem full?");
3569 DPRINTF(("committing page %"Y"u", pgno));
3570 next_pos = pos + size;
3571 iov[n].iov_len = size;
3572 iov[n].iov_base = (char *)dp;
3578 if (pgno > txn->mt_last_pgno)
3579 txn->mt_last_pgno = pgno;
3582 /* MIPS has cache coherency issues, this is a no-op everywhere else
3583 * Note: for any size >= on-chip cache size, entire on-chip cache is
3586 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3588 for (i = keep; ++i <= pagecount; ) {
3590 /* This is a page we skipped above */
3593 dl[j].mid = dp->mp_pgno;
3596 mdb_dpage_free(env, dp);
3601 txn->mt_dirty_room += i - j;
3607 mdb_txn_commit(MDB_txn *txn)
3610 unsigned int i, end_mode;
3616 /* mdb_txn_end() mode for a commit which writes nothing */
3617 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3619 if (txn->mt_child) {
3620 rc = mdb_txn_commit(txn->mt_child);
3627 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3631 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3632 DPUTS("txn has failed/finished, can't commit");
3634 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3639 if (txn->mt_parent) {
3640 MDB_txn *parent = txn->mt_parent;
3644 unsigned x, y, len, ps_len;
3646 /* Append our free list to parent's */
3647 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3650 mdb_midl_free(txn->mt_free_pgs);
3651 /* Failures after this must either undo the changes
3652 * to the parent or set MDB_TXN_ERROR in the parent.
3655 parent->mt_next_pgno = txn->mt_next_pgno;
3656 parent->mt_flags = txn->mt_flags;
3658 /* Merge our cursors into parent's and close them */
3659 mdb_cursors_close(txn, 1);
3661 /* Update parent's DB table. */
3662 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3663 parent->mt_numdbs = txn->mt_numdbs;
3664 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3665 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3666 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3667 /* preserve parent's DB_NEW status */
3668 x = parent->mt_dbflags[i] & DB_NEW;
3669 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3672 dst = parent->mt_u.dirty_list;
3673 src = txn->mt_u.dirty_list;
3674 /* Remove anything in our dirty list from parent's spill list */
3675 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3677 pspill[0] = (pgno_t)-1;
3678 /* Mark our dirty pages as deleted in parent spill list */
3679 for (i=0, len=src[0].mid; ++i <= len; ) {
3680 MDB_ID pn = src[i].mid << 1;
3681 while (pn > pspill[x])
3683 if (pn == pspill[x]) {
3688 /* Squash deleted pagenums if we deleted any */
3689 for (x=y; ++x <= ps_len; )
3690 if (!(pspill[x] & 1))
3691 pspill[++y] = pspill[x];
3695 /* Remove anything in our spill list from parent's dirty list */
3696 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3697 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3698 MDB_ID pn = txn->mt_spill_pgs[i];
3700 continue; /* deleted spillpg */
3702 y = mdb_mid2l_search(dst, pn);
3703 if (y <= dst[0].mid && dst[y].mid == pn) {
3705 while (y < dst[0].mid) {
3714 /* Find len = length of merging our dirty list with parent's */
3716 dst[0].mid = 0; /* simplify loops */
3717 if (parent->mt_parent) {
3718 len = x + src[0].mid;
3719 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3720 for (i = x; y && i; y--) {
3721 pgno_t yp = src[y].mid;
3722 while (yp < dst[i].mid)
3724 if (yp == dst[i].mid) {
3729 } else { /* Simplify the above for single-ancestor case */
3730 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3732 /* Merge our dirty list with parent's */
3734 for (i = len; y; dst[i--] = src[y--]) {
3735 pgno_t yp = src[y].mid;
3736 while (yp < dst[x].mid)
3737 dst[i--] = dst[x--];
3738 if (yp == dst[x].mid)
3739 free(dst[x--].mptr);
3741 mdb_tassert(txn, i == x);
3743 free(txn->mt_u.dirty_list);
3744 parent->mt_dirty_room = txn->mt_dirty_room;
3745 if (txn->mt_spill_pgs) {
3746 if (parent->mt_spill_pgs) {
3747 /* TODO: Prevent failure here, so parent does not fail */
3748 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3750 parent->mt_flags |= MDB_TXN_ERROR;
3751 mdb_midl_free(txn->mt_spill_pgs);
3752 mdb_midl_sort(parent->mt_spill_pgs);
3754 parent->mt_spill_pgs = txn->mt_spill_pgs;
3758 /* Append our loose page list to parent's */
3759 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3761 *lp = txn->mt_loose_pgs;
3762 parent->mt_loose_count += txn->mt_loose_count;
3764 parent->mt_child = NULL;
3765 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3770 if (txn != env->me_txn) {
3771 DPUTS("attempt to commit unknown transaction");
3776 mdb_cursors_close(txn, 0);
3778 if (!txn->mt_u.dirty_list[0].mid &&
3779 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3782 DPRINTF(("committing txn %"Y"u %p on mdbenv %p, root page %"Y"u",
3783 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3785 /* Update DB root pointers */
3786 if (txn->mt_numdbs > CORE_DBS) {
3790 data.mv_size = sizeof(MDB_db);
3792 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3793 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3794 if (txn->mt_dbflags[i] & DB_DIRTY) {
3795 if (TXN_DBI_CHANGED(txn, i)) {
3799 data.mv_data = &txn->mt_dbs[i];
3800 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3808 rc = mdb_freelist_save(txn);
3812 mdb_midl_free(env->me_pghead);
3813 env->me_pghead = NULL;
3814 mdb_midl_shrink(&txn->mt_free_pgs);
3820 if ((rc = mdb_page_flush(txn, 0)))
3822 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3823 (rc = mdb_env_sync0(env, 0, txn->mt_next_pgno)))
3825 if ((rc = mdb_env_write_meta(txn)))
3827 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3830 mdb_txn_end(txn, end_mode);
3838 /** Read the environment parameters of a DB environment before
3839 * mapping it into memory.
3840 * @param[in] env the environment handle
3841 * @param[out] meta address of where to store the meta information
3842 * @return 0 on success, non-zero on failure.
3845 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3851 enum { Size = sizeof(pbuf) };
3853 /* We don't know the page size yet, so use a minimum value.
3854 * Read both meta pages so we can use the latest one.
3857 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3861 memset(&ov, 0, sizeof(ov));
3863 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3864 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3867 rc = pread(env->me_fd, &pbuf, Size, off);
3870 if (rc == 0 && off == 0)
3872 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3873 DPRINTF(("read: %s", mdb_strerror(rc)));
3877 p = (MDB_page *)&pbuf;
3879 if (!F_ISSET(p->mp_flags, P_META)) {
3880 DPRINTF(("page %"Y"u not a meta page", p->mp_pgno));
3885 if (m->mm_magic != MDB_MAGIC) {
3886 DPUTS("meta has invalid magic");
3890 if (m->mm_version != MDB_DATA_VERSION) {
3891 DPRINTF(("database is version %u, expected version %u",
3892 m->mm_version, MDB_DATA_VERSION));
3893 return MDB_VERSION_MISMATCH;
3896 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3902 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3904 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3906 meta->mm_magic = MDB_MAGIC;
3907 meta->mm_version = MDB_DATA_VERSION;
3908 meta->mm_mapsize = env->me_mapsize;
3909 meta->mm_psize = env->me_psize;
3910 meta->mm_last_pg = NUM_METAS-1;
3911 meta->mm_flags = env->me_flags & 0xffff;
3912 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3913 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3914 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3917 /** Write the environment parameters of a freshly created DB environment.
3918 * @param[in] env the environment handle
3919 * @param[in] meta the #MDB_meta to write
3920 * @return 0 on success, non-zero on failure.
3923 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3931 memset(&ov, 0, sizeof(ov));
3932 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3934 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3937 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3938 len = pwrite(fd, ptr, size, pos); \
3939 if (len == -1 && ErrCode() == EINTR) continue; \
3940 rc = (len >= 0); break; } while(1)
3943 DPUTS("writing new meta page");
3945 psize = env->me_psize;
3947 p = calloc(NUM_METAS, psize);
3951 p->mp_flags = P_META;
3952 *(MDB_meta *)METADATA(p) = *meta;
3954 q = (MDB_page *)((char *)p + psize);
3956 q->mp_flags = P_META;
3957 *(MDB_meta *)METADATA(q) = *meta;
3959 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3962 else if ((unsigned) len == psize * NUM_METAS)
3970 /** Update the environment info to commit a transaction.
3971 * @param[in] txn the transaction that's being committed
3972 * @return 0 on success, non-zero on failure.
3975 mdb_env_write_meta(MDB_txn *txn)
3978 MDB_meta meta, metab, *mp;
3982 int rc, len, toggle;
3991 toggle = txn->mt_txnid & 1;
3992 DPRINTF(("writing meta page %d for root page %"Y"u",
3993 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3996 flags = txn->mt_flags | env->me_flags;
3997 mp = env->me_metas[toggle];
3998 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3999 /* Persist any increases of mapsize config */
4000 if (mapsize < env->me_mapsize)
4001 mapsize = env->me_mapsize;
4003 if (flags & MDB_WRITEMAP) {
4004 mp->mm_mapsize = mapsize;
4005 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4006 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4007 mp->mm_last_pg = txn->mt_next_pgno - 1;
4008 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
4009 !(defined(__i386__) || defined(__x86_64__))
4010 /* LY: issue a memory barrier, if not x86. ITS#7969 */
4011 __sync_synchronize();
4013 mp->mm_txnid = txn->mt_txnid;
4014 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
4015 unsigned meta_size = env->me_psize;
4016 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
4017 ptr = (char *)mp - PAGEHDRSZ;
4018 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
4019 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
4023 if (MDB_MSYNC(ptr, meta_size, rc)) {
4030 metab.mm_txnid = mp->mm_txnid;
4031 metab.mm_last_pg = mp->mm_last_pg;
4033 meta.mm_mapsize = mapsize;
4034 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4035 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4036 meta.mm_last_pg = txn->mt_next_pgno - 1;
4037 meta.mm_txnid = txn->mt_txnid;
4039 off = offsetof(MDB_meta, mm_mapsize);
4040 ptr = (char *)&meta + off;
4041 len = sizeof(MDB_meta) - off;
4042 off += (char *)mp - env->me_map;
4044 /* Write to the SYNC fd */
4045 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
4048 memset(&ov, 0, sizeof(ov));
4050 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
4055 rc = pwrite(mfd, ptr, len, off);
4058 rc = rc < 0 ? ErrCode() : EIO;
4063 DPUTS("write failed, disk error?");
4064 /* On a failure, the pagecache still contains the new data.
4065 * Write some old data back, to prevent it from being used.
4066 * Use the non-SYNC fd; we know it will fail anyway.
4068 meta.mm_last_pg = metab.mm_last_pg;
4069 meta.mm_txnid = metab.mm_txnid;
4071 memset(&ov, 0, sizeof(ov));
4073 WriteFile(env->me_fd, ptr, len, NULL, &ov);
4075 r2 = pwrite(env->me_fd, ptr, len, off);
4076 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
4079 env->me_flags |= MDB_FATAL_ERROR;
4082 /* MIPS has cache coherency issues, this is a no-op everywhere else */
4083 CACHEFLUSH(env->me_map + off, len, DCACHE);
4085 /* Memory ordering issues are irrelevant; since the entire writer
4086 * is wrapped by wmutex, all of these changes will become visible
4087 * after the wmutex is unlocked. Since the DB is multi-version,
4088 * readers will get consistent data regardless of how fresh or
4089 * how stale their view of these values is.
4092 env->me_txns->mti_txnid = txn->mt_txnid;
4097 /** Check both meta pages to see which one is newer.
4098 * @param[in] env the environment handle
4099 * @return newest #MDB_meta.
4102 mdb_env_pick_meta(const MDB_env *env)
4104 MDB_meta *const *metas = env->me_metas;
4105 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
4109 mdb_env_create(MDB_env **env)
4113 e = calloc(1, sizeof(MDB_env));
4117 e->me_maxreaders = DEFAULT_READERS;
4118 e->me_maxdbs = e->me_numdbs = CORE_DBS;
4119 e->me_fd = INVALID_HANDLE_VALUE;
4120 e->me_lfd = INVALID_HANDLE_VALUE;
4121 e->me_mfd = INVALID_HANDLE_VALUE;
4122 #ifdef MDB_USE_POSIX_SEM
4123 e->me_rmutex = SEM_FAILED;
4124 e->me_wmutex = SEM_FAILED;
4125 #elif defined MDB_USE_SYSV_SEM
4126 e->me_rmutex->semid = -1;
4127 e->me_wmutex->semid = -1;
4129 e->me_pid = getpid();
4130 GET_PAGESIZE(e->me_os_psize);
4131 VGMEMP_CREATE(e,0,0);
4137 mdb_env_map(MDB_env *env, void *addr)
4140 unsigned int flags = env->me_flags;
4143 int access = SECTION_MAP_READ;
4147 ULONG pageprot = PAGE_READONLY, secprot, alloctype;
4149 if (flags & MDB_WRITEMAP) {
4150 access |= SECTION_MAP_WRITE;
4151 pageprot = PAGE_READWRITE;
4153 if (flags & MDB_RDONLY) {
4154 secprot = PAGE_READONLY;
4158 secprot = PAGE_READWRITE;
4159 msize = env->me_mapsize;
4160 alloctype = MEM_RESERVE;
4163 rc = NtCreateSection(&mh, access, NULL, NULL, secprot, SEC_RESERVE, env->me_fd);
4168 msize = NUM_METAS * env->me_psize;
4170 rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
4182 env->me_map = mmap(addr, NUM_METAS * env->me_psize, PROT_READ, MAP_SHARED,
4184 if (env->me_map == MAP_FAILED) {
4189 int prot = PROT_READ;
4190 if (flags & MDB_WRITEMAP) {
4192 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
4195 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
4197 if (env->me_map == MAP_FAILED) {
4202 if (flags & MDB_NORDAHEAD) {
4203 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
4205 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
4207 #ifdef POSIX_MADV_RANDOM
4208 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
4209 #endif /* POSIX_MADV_RANDOM */
4210 #endif /* MADV_RANDOM */
4214 /* Can happen because the address argument to mmap() is just a
4215 * hint. mmap() can pick another, e.g. if the range is in use.
4216 * The MAP_FIXED flag would prevent that, but then mmap could
4217 * instead unmap existing pages to make room for the new map.
4219 if (addr && env->me_map != addr)
4220 return EBUSY; /* TODO: Make a new MDB_* error code? */
4223 p = (MDB_page *)env->me_map;
4224 env->me_metas[0] = METADATA(p);
4225 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4231 mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
4233 /* If env is already open, caller is responsible for making
4234 * sure there are no active txns.
4244 meta = mdb_env_pick_meta(env);
4246 size = meta->mm_mapsize;
4248 /* Silently round up to minimum if the size is too small */
4249 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4254 /* For MDB_VL32 this bit is a noop since we dynamically remap
4255 * chunks of the DB anyway.
4257 munmap(env->me_map, env->me_mapsize);
4258 env->me_mapsize = size;
4259 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4260 rc = mdb_env_map(env, old);
4263 #endif /* !MDB_VL32 */
4265 env->me_mapsize = size;
4267 env->me_maxpg = env->me_mapsize / env->me_psize;
4272 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4276 env->me_maxdbs = dbs + CORE_DBS;
4281 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4283 if (env->me_map || readers < 1)
4285 env->me_maxreaders = readers;
4290 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4292 if (!env || !readers)
4294 *readers = env->me_maxreaders;
4299 mdb_fsize(HANDLE fd, mdb_size_t *size)
4302 LARGE_INTEGER fsize;
4304 if (!GetFileSizeEx(fd, &fsize))
4307 *size = fsize.QuadPart;
4319 #ifdef BROKEN_FDATASYNC
4320 #include <sys/utsname.h>
4321 #include <sys/vfs.h>
4324 /** Further setup required for opening an LMDB environment
4327 mdb_env_open2(MDB_env *env)
4329 unsigned int flags = env->me_flags;
4330 int i, newenv = 0, rc;
4334 /* See if we should use QueryLimited */
4336 if ((rc & 0xff) > 5)
4337 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4339 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4342 #ifdef BROKEN_FDATASYNC
4343 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4344 * https://lkml.org/lkml/2012/9/3/83
4345 * Kernels after 3.6-rc6 are known good.
4346 * https://lkml.org/lkml/2012/9/10/556
4347 * See if the DB is on ext3/ext4, then check for new enough kernel
4348 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4353 fstatfs(env->me_fd, &st);
4354 while (st.f_type == 0xEF53) {
4358 if (uts.release[0] < '3') {
4359 if (!strncmp(uts.release, "2.6.32.", 7)) {
4360 i = atoi(uts.release+7);
4362 break; /* 2.6.32.60 and newer is OK */
4363 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4364 i = atoi(uts.release+7);
4366 break; /* 2.6.34.15 and newer is OK */
4368 } else if (uts.release[0] == '3') {
4369 i = atoi(uts.release+2);
4371 break; /* 3.6 and newer is OK */
4373 i = atoi(uts.release+4);
4375 break; /* 3.5.4 and newer is OK */
4376 } else if (i == 2) {
4377 i = atoi(uts.release+4);
4379 break; /* 3.2.30 and newer is OK */
4381 } else { /* 4.x and newer is OK */
4384 env->me_flags |= MDB_FSYNCONLY;
4390 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4393 DPUTS("new mdbenv");
4395 env->me_psize = env->me_os_psize;
4396 if (env->me_psize > MAX_PAGESIZE)
4397 env->me_psize = MAX_PAGESIZE;
4398 memset(&meta, 0, sizeof(meta));
4399 mdb_env_init_meta0(env, &meta);
4400 meta.mm_mapsize = DEFAULT_MAPSIZE;
4402 env->me_psize = meta.mm_psize;
4405 /* Was a mapsize configured? */
4406 if (!env->me_mapsize) {
4407 env->me_mapsize = meta.mm_mapsize;
4410 /* Make sure mapsize >= committed data size. Even when using
4411 * mm_mapsize, which could be broken in old files (ITS#7789).
4413 mdb_size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4414 if (env->me_mapsize < minsize)
4415 env->me_mapsize = minsize;
4417 meta.mm_mapsize = env->me_mapsize;
4419 if (newenv && !(flags & MDB_FIXEDMAP)) {
4420 /* mdb_env_map() may grow the datafile. Write the metapages
4421 * first, so the file will be valid if initialization fails.
4422 * Except with FIXEDMAP, since we do not yet know mm_address.
4423 * We could fill in mm_address later, but then a different
4424 * program might end up doing that - one with a memory layout
4425 * and map address which does not suit the main program.
4427 rc = mdb_env_init_meta(env, &meta);
4433 /* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
4437 rc = WriteFile(env->me_fd, &dummy, 1, &len, NULL);
4445 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4450 if (flags & MDB_FIXEDMAP)
4451 meta.mm_address = env->me_map;
4452 i = mdb_env_init_meta(env, &meta);
4453 if (i != MDB_SUCCESS) {
4458 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4459 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4461 #if !(MDB_MAXKEYSIZE)
4462 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4464 env->me_maxpg = env->me_mapsize / env->me_psize;
4468 MDB_meta *meta = mdb_env_pick_meta(env);
4469 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4471 DPRINTF(("opened database version %u, pagesize %u",
4472 meta->mm_version, env->me_psize));
4473 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4474 DPRINTF(("depth: %u", db->md_depth));
4475 DPRINTF(("entries: %"Y"u", db->md_entries));
4476 DPRINTF(("branch pages: %"Y"u", db->md_branch_pages));
4477 DPRINTF(("leaf pages: %"Y"u", db->md_leaf_pages));
4478 DPRINTF(("overflow pages: %"Y"u", db->md_overflow_pages));
4479 DPRINTF(("root: %"Y"u", db->md_root));
4487 /** Release a reader thread's slot in the reader lock table.
4488 * This function is called automatically when a thread exits.
4489 * @param[in] ptr This points to the slot in the reader lock table.
4492 mdb_env_reader_dest(void *ptr)
4494 MDB_reader *reader = ptr;
4500 /** Junk for arranging thread-specific callbacks on Windows. This is
4501 * necessarily platform and compiler-specific. Windows supports up
4502 * to 1088 keys. Let's assume nobody opens more than 64 environments
4503 * in a single process, for now. They can override this if needed.
4505 #ifndef MAX_TLS_KEYS
4506 #define MAX_TLS_KEYS 64
4508 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4509 static int mdb_tls_nkeys;
4511 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4515 case DLL_PROCESS_ATTACH: break;
4516 case DLL_THREAD_ATTACH: break;
4517 case DLL_THREAD_DETACH:
4518 for (i=0; i<mdb_tls_nkeys; i++) {
4519 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4521 mdb_env_reader_dest(r);
4525 case DLL_PROCESS_DETACH: break;
4530 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4532 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4536 /* Force some symbol references.
4537 * _tls_used forces the linker to create the TLS directory if not already done
4538 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4540 #pragma comment(linker, "/INCLUDE:_tls_used")
4541 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4542 #pragma const_seg(".CRT$XLB")
4543 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4544 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4547 #pragma comment(linker, "/INCLUDE:__tls_used")
4548 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4549 #pragma data_seg(".CRT$XLB")
4550 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4552 #endif /* WIN 32/64 */
4553 #endif /* !__GNUC__ */
4556 /** Downgrade the exclusive lock on the region back to shared */
4558 mdb_env_share_locks(MDB_env *env, int *excl)
4561 MDB_meta *meta = mdb_env_pick_meta(env);
4563 env->me_txns->mti_txnid = meta->mm_txnid;
4568 /* First acquire a shared lock. The Unlock will
4569 * then release the existing exclusive lock.
4571 memset(&ov, 0, sizeof(ov));
4572 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4575 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4581 struct flock lock_info;
4582 /* The shared lock replaces the existing lock */
4583 memset((void *)&lock_info, 0, sizeof(lock_info));
4584 lock_info.l_type = F_RDLCK;
4585 lock_info.l_whence = SEEK_SET;
4586 lock_info.l_start = 0;
4587 lock_info.l_len = 1;
4588 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4589 (rc = ErrCode()) == EINTR) ;
4590 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4597 /** Try to get exclusive lock, otherwise shared.
4598 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4601 mdb_env_excl_lock(MDB_env *env, int *excl)
4605 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4609 memset(&ov, 0, sizeof(ov));
4610 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4617 struct flock lock_info;
4618 memset((void *)&lock_info, 0, sizeof(lock_info));
4619 lock_info.l_type = F_WRLCK;
4620 lock_info.l_whence = SEEK_SET;
4621 lock_info.l_start = 0;
4622 lock_info.l_len = 1;
4623 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4624 (rc = ErrCode()) == EINTR) ;
4628 # ifndef MDB_USE_POSIX_MUTEX
4629 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4632 lock_info.l_type = F_RDLCK;
4633 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4634 (rc = ErrCode()) == EINTR) ;
4644 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4646 * @(#) $Revision: 5.1 $
4647 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4648 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4650 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4654 * Please do not copyright this code. This code is in the public domain.
4656 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4657 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4658 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4659 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4660 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4661 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4662 * PERFORMANCE OF THIS SOFTWARE.
4665 * chongo <Landon Curt Noll> /\oo/\
4666 * http://www.isthe.com/chongo/
4668 * Share and Enjoy! :-)
4671 typedef unsigned long long mdb_hash_t;
4672 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4674 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4675 * @param[in] val value to hash
4676 * @param[in] hval initial value for hash
4677 * @return 64 bit hash
4679 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4680 * hval arg on the first call.
4683 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4685 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4686 unsigned char *end = s + val->mv_size;
4688 * FNV-1a hash each octet of the string
4691 /* xor the bottom with the current octet */
4692 hval ^= (mdb_hash_t)*s++;
4694 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4695 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4696 (hval << 7) + (hval << 8) + (hval << 40);
4698 /* return our new hash value */
4702 /** Hash the string and output the encoded hash.
4703 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4704 * very short name limits. We don't care about the encoding being reversible,
4705 * we just want to preserve as many bits of the input as possible in a
4706 * small printable string.
4707 * @param[in] str string to hash
4708 * @param[out] encbuf an array of 11 chars to hold the hash
4710 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4713 mdb_pack85(unsigned long l, char *out)
4717 for (i=0; i<5; i++) {
4718 *out++ = mdb_a85[l % 85];
4724 mdb_hash_enc(MDB_val *val, char *encbuf)
4726 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4728 mdb_pack85(h, encbuf);
4729 mdb_pack85(h>>32, encbuf+5);
4734 /** Open and/or initialize the lock region for the environment.
4735 * @param[in] env The LMDB environment.
4736 * @param[in] lpath The pathname of the file used for the lock region.
4737 * @param[in] mode The Unix permissions for the file, if we create it.
4738 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4739 * @return 0 on success, non-zero on failure.
4742 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4745 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4747 # define MDB_ERRCODE_ROFS EROFS
4748 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4749 # define MDB_CLOEXEC O_CLOEXEC
4752 # define MDB_CLOEXEC 0
4755 #ifdef MDB_USE_SYSV_SEM
4764 rc = utf8_to_utf16(lpath, -1, &wlpath, NULL);
4767 env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
4768 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4769 FILE_ATTRIBUTE_NORMAL, NULL);
4772 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4774 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4776 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4781 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4782 /* Lose record locks when exec*() */
4783 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4784 fcntl(env->me_lfd, F_SETFD, fdflags);
4787 if (!(env->me_flags & MDB_NOTLS)) {
4788 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4791 env->me_flags |= MDB_ENV_TXKEY;
4793 /* Windows TLS callbacks need help finding their TLS info. */
4794 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4798 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4802 /* Try to get exclusive lock. If we succeed, then
4803 * nobody is using the lock region and we should initialize it.
4805 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4808 size = GetFileSize(env->me_lfd, NULL);
4810 size = lseek(env->me_lfd, 0, SEEK_END);
4811 if (size == -1) goto fail_errno;
4813 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4814 if (size < rsize && *excl > 0) {
4816 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4817 || !SetEndOfFile(env->me_lfd))
4820 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4824 size = rsize - sizeof(MDB_txninfo);
4825 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4830 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4832 if (!mh) goto fail_errno;
4833 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4835 if (!env->me_txns) goto fail_errno;
4837 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4839 if (m == MAP_FAILED) goto fail_errno;
4845 BY_HANDLE_FILE_INFORMATION stbuf;
4854 if (!mdb_sec_inited) {
4855 InitializeSecurityDescriptor(&mdb_null_sd,
4856 SECURITY_DESCRIPTOR_REVISION);
4857 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4858 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4859 mdb_all_sa.bInheritHandle = FALSE;
4860 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4863 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4864 idbuf.volume = stbuf.dwVolumeSerialNumber;
4865 idbuf.nhigh = stbuf.nFileIndexHigh;
4866 idbuf.nlow = stbuf.nFileIndexLow;
4867 val.mv_data = &idbuf;
4868 val.mv_size = sizeof(idbuf);
4869 mdb_hash_enc(&val, encbuf);
4870 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4871 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4872 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4873 if (!env->me_rmutex) goto fail_errno;
4874 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4875 if (!env->me_wmutex) goto fail_errno;
4876 #elif defined(MDB_USE_POSIX_SEM)
4885 #if defined(__NetBSD__)
4886 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4888 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4889 idbuf.dev = stbuf.st_dev;
4890 idbuf.ino = stbuf.st_ino;
4891 val.mv_data = &idbuf;
4892 val.mv_size = sizeof(idbuf);
4893 mdb_hash_enc(&val, encbuf);
4894 #ifdef MDB_SHORT_SEMNAMES
4895 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4897 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4898 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4899 /* Clean up after a previous run, if needed: Try to
4900 * remove both semaphores before doing anything else.
4902 sem_unlink(env->me_txns->mti_rmname);
4903 sem_unlink(env->me_txns->mti_wmname);
4904 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4905 O_CREAT|O_EXCL, mode, 1);
4906 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4907 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4908 O_CREAT|O_EXCL, mode, 1);
4909 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4910 #elif defined(MDB_USE_SYSV_SEM)
4911 unsigned short vals[2] = {1, 1};
4912 key_t key = ftok(lpath, 'M');
4915 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4919 if (semctl(semid, 0, SETALL, semu) < 0)
4921 env->me_txns->mti_semid = semid;
4922 #else /* MDB_USE_POSIX_MUTEX: */
4923 pthread_mutexattr_t mattr;
4925 if ((rc = pthread_mutexattr_init(&mattr))
4926 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4927 #ifdef MDB_ROBUST_SUPPORTED
4928 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4930 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4931 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4933 pthread_mutexattr_destroy(&mattr);
4934 #endif /* _WIN32 || ... */
4936 env->me_txns->mti_magic = MDB_MAGIC;
4937 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4938 env->me_txns->mti_txnid = 0;
4939 env->me_txns->mti_numreaders = 0;
4942 #ifdef MDB_USE_SYSV_SEM
4943 struct semid_ds buf;
4945 if (env->me_txns->mti_magic != MDB_MAGIC) {
4946 DPUTS("lock region has invalid magic");
4950 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4951 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4952 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4953 rc = MDB_VERSION_MISMATCH;
4957 if (rc && rc != EACCES && rc != EAGAIN) {
4961 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4962 if (!env->me_rmutex) goto fail_errno;
4963 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4964 if (!env->me_wmutex) goto fail_errno;
4965 #elif defined(MDB_USE_POSIX_SEM)
4966 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4967 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4968 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4969 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4970 #elif defined(MDB_USE_SYSV_SEM)
4971 semid = env->me_txns->mti_semid;
4973 /* check for read access */
4974 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4976 /* check for write access */
4977 if (semctl(semid, 0, IPC_SET, semu) < 0)
4981 #ifdef MDB_USE_SYSV_SEM
4982 env->me_rmutex->semid = semid;
4983 env->me_wmutex->semid = semid;
4984 env->me_rmutex->semnum = 0;
4985 env->me_wmutex->semnum = 1;
4986 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
4987 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
4991 env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
4993 pthread_mutex_init(env->me_rpmutex, NULL);
5005 /** The name of the lock file in the DB environment */
5006 #define LOCKNAME "/lock.mdb"
5007 /** The name of the data file in the DB environment */
5008 #define DATANAME "/data.mdb"
5009 /** The suffix of the lock file when no subdir is used */
5010 #define LOCKSUFF "-lock"
5011 /** Only a subset of the @ref mdb_env flags can be changed
5012 * at runtime. Changing other flags requires closing the
5013 * environment and re-opening it with the new flags.
5015 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
5016 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
5017 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
5019 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
5020 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
5024 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
5026 int oflags, rc, len, excl = -1;
5027 char *lpath, *dpath;
5032 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
5036 if (flags & MDB_WRITEMAP) {
5037 /* silently ignore WRITEMAP in 32 bit mode */
5038 flags ^= MDB_WRITEMAP;
5040 if (flags & MDB_FIXEDMAP) {
5041 /* cannot support FIXEDMAP */
5047 if (flags & MDB_NOSUBDIR) {
5048 rc = len + sizeof(LOCKSUFF) + len + 1;
5050 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
5055 if (flags & MDB_NOSUBDIR) {
5056 dpath = lpath + len + sizeof(LOCKSUFF);
5057 sprintf(lpath, "%s" LOCKSUFF, path);
5058 strcpy(dpath, path);
5060 dpath = lpath + len + sizeof(LOCKNAME);
5061 sprintf(lpath, "%s" LOCKNAME, path);
5062 sprintf(dpath, "%s" DATANAME, path);
5066 flags |= env->me_flags;
5067 if (flags & MDB_RDONLY) {
5068 /* silently ignore WRITEMAP when we're only getting read access */
5069 flags &= ~MDB_WRITEMAP;
5071 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
5072 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
5077 env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
5078 if (!env->me_rpages) {
5082 env->me_rpages[0].mid = 0;
5083 env->me_rpcheck = MDB_ERPAGE_SIZE/2;
5086 env->me_flags = flags |= MDB_ENV_ACTIVE;
5090 env->me_path = strdup(path);
5091 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
5092 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
5093 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
5094 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
5098 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
5100 /* For RDONLY, get lockfile after we know datafile exists */
5101 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
5102 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
5108 if (F_ISSET(flags, MDB_RDONLY)) {
5109 oflags = GENERIC_READ;
5110 len = OPEN_EXISTING;
5112 oflags = GENERIC_READ|GENERIC_WRITE;
5115 mode = FILE_ATTRIBUTE_NORMAL;
5116 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
5119 env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
5120 NULL, len, mode, NULL);
5123 if (F_ISSET(flags, MDB_RDONLY))
5126 oflags = O_RDWR | O_CREAT;
5128 env->me_fd = open(dpath, oflags, mode);
5130 if (env->me_fd == INVALID_HANDLE_VALUE) {
5135 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
5136 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
5141 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
5142 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
5143 env->me_mfd = env->me_fd;
5145 /* Synchronous fd for meta writes. Needed even with
5146 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
5149 len = OPEN_EXISTING;
5150 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
5153 env->me_mfd = CreateFileW(wpath, oflags,
5154 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
5155 mode | FILE_FLAG_WRITE_THROUGH, NULL);
5159 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
5161 if (env->me_mfd == INVALID_HANDLE_VALUE) {
5166 DPRINTF(("opened dbenv %p", (void *) env));
5168 rc = mdb_env_share_locks(env, &excl);
5172 if (!(flags & MDB_RDONLY)) {
5174 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
5175 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
5176 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
5177 (txn = calloc(1, size)))
5179 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
5180 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
5181 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
5182 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
5185 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
5186 if (!txn->mt_rpages) {
5191 txn->mt_rpages[0].mid = 0;
5192 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
5194 txn->mt_dbxs = env->me_dbxs;
5195 txn->mt_flags = MDB_TXN_FINISHED;
5205 mdb_env_close0(env, excl);
5211 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5213 mdb_env_close0(MDB_env *env, int excl)
5217 if (!(env->me_flags & MDB_ENV_ACTIVE))
5220 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5222 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5223 free(env->me_dbxs[i].md_name.mv_data);
5228 free(env->me_dbiseqs);
5229 free(env->me_dbflags);
5231 free(env->me_dirty_list);
5235 for (x=1; x<=env->me_rpages[0].mid; x++)
5236 munmap(env->me_rpages[x].mptr, env->me_rpages[x].mcnt * env->me_psize);
5238 free(env->me_rpages);
5240 mdb_midl_free(env->me_free_pgs);
5242 if (env->me_flags & MDB_ENV_TXKEY) {
5243 pthread_key_delete(env->me_txkey);
5245 /* Delete our key from the global list */
5246 for (i=0; i<mdb_tls_nkeys; i++)
5247 if (mdb_tls_keys[i] == env->me_txkey) {
5248 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5257 munmap(env->me_map, NUM_METAS*env->me_psize);
5259 munmap(env->me_map, env->me_mapsize);
5262 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
5263 (void) close(env->me_mfd);
5264 if (env->me_fd != INVALID_HANDLE_VALUE)
5265 (void) close(env->me_fd);
5267 MDB_PID_T pid = env->me_pid;
5268 /* Clearing readers is done in this function because
5269 * me_txkey with its destructor must be disabled first.
5271 * We skip the the reader mutex, so we touch only
5272 * data owned by this process (me_close_readers and
5273 * our readers), and clear each reader atomically.
5275 for (i = env->me_close_readers; --i >= 0; )
5276 if (env->me_txns->mti_readers[i].mr_pid == pid)
5277 env->me_txns->mti_readers[i].mr_pid = 0;
5279 if (env->me_rmutex) {
5280 CloseHandle(env->me_rmutex);
5281 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5283 /* Windows automatically destroys the mutexes when
5284 * the last handle closes.
5286 #elif defined(MDB_USE_POSIX_SEM)
5287 if (env->me_rmutex != SEM_FAILED) {
5288 sem_close(env->me_rmutex);
5289 if (env->me_wmutex != SEM_FAILED)
5290 sem_close(env->me_wmutex);
5291 /* If we have the filelock: If we are the
5292 * only remaining user, clean up semaphores.
5295 mdb_env_excl_lock(env, &excl);
5297 sem_unlink(env->me_txns->mti_rmname);
5298 sem_unlink(env->me_txns->mti_wmname);
5301 #elif defined(MDB_USE_SYSV_SEM)
5302 if (env->me_rmutex->semid != -1) {
5303 /* If we have the filelock: If we are the
5304 * only remaining user, clean up semaphores.
5307 mdb_env_excl_lock(env, &excl);
5309 semctl(env->me_rmutex->semid, 0, IPC_RMID);
5312 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5314 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5317 /* Unlock the lockfile. Windows would have unlocked it
5318 * after closing anyway, but not necessarily at once.
5320 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5323 (void) close(env->me_lfd);
5327 if (env->me_rpmutex) CloseHandle(env->me_rpmutex);
5329 pthread_mutex_destroy(env->me_rpmutex);
5333 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5337 mdb_env_close(MDB_env *env)
5344 VGMEMP_DESTROY(env);
5345 while ((dp = env->me_dpages) != NULL) {
5346 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5347 env->me_dpages = dp->mp_next;
5351 mdb_env_close0(env, 0);
5355 /** Compare two items pointing at aligned mdb_size_t's */
5357 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5359 return (*(mdb_size_t *)a->mv_data < *(mdb_size_t *)b->mv_data) ? -1 :
5360 *(mdb_size_t *)a->mv_data > *(mdb_size_t *)b->mv_data;
5363 /** Compare two items pointing at aligned unsigned int's.
5365 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5366 * but #mdb_cmp_clong() is called instead if the data type is mdb_size_t.
5369 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5371 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5372 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5375 /** Compare two items pointing at unsigned ints of unknown alignment.
5376 * Nodes and keys are guaranteed to be 2-byte aligned.
5379 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5381 #if BYTE_ORDER == LITTLE_ENDIAN
5382 unsigned short *u, *c;
5385 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5386 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5389 } while(!x && u > (unsigned short *)a->mv_data);
5392 unsigned short *u, *c, *end;
5395 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5396 u = (unsigned short *)a->mv_data;
5397 c = (unsigned short *)b->mv_data;
5400 } while(!x && u < end);
5405 /** Compare two items lexically */
5407 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5414 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5420 diff = memcmp(a->mv_data, b->mv_data, len);
5421 return diff ? diff : len_diff<0 ? -1 : len_diff;
5424 /** Compare two items in reverse byte order */
5426 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5428 const unsigned char *p1, *p2, *p1_lim;
5432 p1_lim = (const unsigned char *)a->mv_data;
5433 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5434 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5436 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5442 while (p1 > p1_lim) {
5443 diff = *--p1 - *--p2;
5447 return len_diff<0 ? -1 : len_diff;
5450 /** Search for key within a page, using binary search.
5451 * Returns the smallest entry larger or equal to the key.
5452 * If exactp is non-null, stores whether the found entry was an exact match
5453 * in *exactp (1 or 0).
5454 * Updates the cursor index with the index of the found entry.
5455 * If no entry larger or equal to the key is found, returns NULL.
5458 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5460 unsigned int i = 0, nkeys;
5463 MDB_page *mp = mc->mc_pg[mc->mc_top];
5464 MDB_node *node = NULL;
5469 nkeys = NUMKEYS(mp);
5471 DPRINTF(("searching %u keys in %s %spage %"Y"u",
5472 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5475 low = IS_LEAF(mp) ? 0 : 1;
5477 cmp = mc->mc_dbx->md_cmp;
5479 /* Branch pages have no data, so if using integer keys,
5480 * alignment is guaranteed. Use faster mdb_cmp_int.
5482 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5483 if (NODEPTR(mp, 1)->mn_ksize == sizeof(mdb_size_t))
5490 nodekey.mv_size = mc->mc_db->md_pad;
5491 node = NODEPTR(mp, 0); /* fake */
5492 while (low <= high) {
5493 i = (low + high) >> 1;
5494 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5495 rc = cmp(key, &nodekey);
5496 DPRINTF(("found leaf index %u [%s], rc = %i",
5497 i, DKEY(&nodekey), rc));
5506 while (low <= high) {
5507 i = (low + high) >> 1;
5509 node = NODEPTR(mp, i);
5510 nodekey.mv_size = NODEKSZ(node);
5511 nodekey.mv_data = NODEKEY(node);
5513 rc = cmp(key, &nodekey);
5516 DPRINTF(("found leaf index %u [%s], rc = %i",
5517 i, DKEY(&nodekey), rc));
5519 DPRINTF(("found branch index %u [%s -> %"Y"u], rc = %i",
5520 i, DKEY(&nodekey), NODEPGNO(node), rc));
5531 if (rc > 0) { /* Found entry is less than the key. */
5532 i++; /* Skip to get the smallest entry larger than key. */
5534 node = NODEPTR(mp, i);
5537 *exactp = (rc == 0 && nkeys > 0);
5538 /* store the key index */
5539 mc->mc_ki[mc->mc_top] = i;
5541 /* There is no entry larger or equal to the key. */
5544 /* nodeptr is fake for LEAF2 */
5550 mdb_cursor_adjust(MDB_cursor *mc, func)
5554 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5555 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5562 /** Pop a page off the top of the cursor's stack. */
5564 mdb_cursor_pop(MDB_cursor *mc)
5567 DPRINTF(("popping page %"Y"u off db %d cursor %p",
5568 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5574 mc->mc_flags &= ~C_INITIALIZED;
5579 /** Push a page onto the top of the cursor's stack. */
5581 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5583 DPRINTF(("pushing page %"Y"u on db %d cursor %p", mp->mp_pgno,
5584 DDBI(mc), (void *) mc));
5586 if (mc->mc_snum >= CURSOR_STACK) {
5587 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5588 return MDB_CURSOR_FULL;
5591 mc->mc_top = mc->mc_snum++;
5592 mc->mc_pg[mc->mc_top] = mp;
5593 mc->mc_ki[mc->mc_top] = 0;
5599 /** Map a read-only page.
5600 * There are two levels of tracking in use, a per-txn list and a per-env list.
5601 * ref'ing and unref'ing the per-txn list is faster since it requires no
5602 * locking. Pages are cached in the per-env list for global reuse, and a lock
5603 * is required. Pages are not immediately unmapped when their refcnt goes to
5604 * zero; they hang around in case they will be reused again soon.
5606 * When the per-txn list gets full, all pages with refcnt=0 are purged from the
5607 * list and their refcnts in the per-env list are decremented.
5609 * When the per-env list gets full, all pages with refcnt=0 are purged from the
5610 * list and their pages are unmapped.
5612 * @note "full" means the list has reached its respective rpcheck threshold.
5613 * This threshold slowly raises if no pages could be purged on a given check,
5614 * and returns to its original value when enough pages were purged.
5616 * If purging doesn't free any slots, filling the per-txn list will return
5617 * MDB_TXN_FULL, and filling the per-env list returns MDB_MAP_FULL.
5619 * Reference tracking in a txn is imperfect, pages can linger with non-zero
5620 * refcnt even without active references. It was deemed to be too invasive
5621 * to add unrefs in every required location. However, all pages are unref'd
5622 * at the end of the transaction. This guarantees that no stale references
5623 * linger in the per-env list.
5625 * Usually we map chunks of 16 pages at a time, but if an overflow page begins
5626 * at the tail of the chunk we extend the chunk to include the entire overflow
5627 * page. Unfortunately, pages can be turned into overflow pages after their
5628 * chunk was already mapped. In that case we must remap the chunk if the
5629 * overflow page is referenced. If the chunk's refcnt is 0 we can just remap
5630 * it, otherwise we temporarily map a new chunk just for the overflow page.
5632 * @note this chunk handling means we cannot guarantee that a data item
5633 * returned from the DB will stay alive for the duration of the transaction:
5634 * We unref pages as soon as a cursor moves away from the page
5635 * A subsequent op may cause a purge, which may unmap any unref'd chunks
5636 * The caller must copy the data if it must be used later in the same txn.
5638 * Also - our reference counting revolves around cursors, but overflow pages
5639 * aren't ever pointed to by a cursor. As such, they always remain referenced
5640 * in a txn until it ends.
5642 * @param[in] txn the transaction for this access.
5643 * @param[in] pgno the page number for the page to retrieve.
5644 * @param[out] ret address of a pointer where the page's address will be stored.
5645 * @return 0 on success, non-zero on failure.
5648 mdb_rpage_get(MDB_txn *txn, pgno_t pg0, MDB_page **ret)
5650 MDB_env *env = txn->mt_env;
5652 MDB_ID3L tl = txn->mt_rpages;
5653 MDB_ID3L el = env->me_rpages;
5657 int rc, retries = 1;
5661 #define SET_OFF(off,val) off.QuadPart = val
5662 #define MAP(rc,env,addr,len,off) \
5664 rc = NtMapViewOfSection(env->me_fmh, GetCurrentProcess(), &addr, 0, \
5665 len, &off, &len, ViewUnmap, MEM_RESERVE, PAGE_READONLY)
5669 #define SET_OFF(off,val) off = val
5670 #define MAP(rc,env,addr,len,off) \
5671 addr = mmap(NULL, len, PROT_READ, MAP_SHARED, env->me_fd, off); \
5672 rc = (addr == MAP_FAILED) ? errno : 0
5675 /* remember the offset of the actual page number, so we can
5676 * return the correct pointer at the end.
5678 rem = pg0 & (MDB_RPAGE_CHUNK-1);
5682 x = mdb_mid3l_search(tl, pgno);
5683 if (x <= tl[0].mid && tl[x].mid == pgno) {
5684 if (x != tl[0].mid && tl[x+1].mid == pg0)
5686 /* check for overflow size */
5687 p = (MDB_page *)((char *)tl[x].mptr + rem * env->me_psize);
5688 if (IS_OVERFLOW(p) && p->mp_pages + rem > tl[x].mcnt) {
5689 id3.mcnt = p->mp_pages + rem;
5690 len = id3.mcnt * env->me_psize;
5691 SET_OFF(off, pgno * env->me_psize);
5692 MAP(rc, env, id3.mptr, len, off);
5695 /* check for local-only page */
5697 mdb_tassert(txn, tl[x].mid != pg0);
5698 /* hope there's room to insert this locally.
5699 * setting mid here tells later code to just insert
5700 * this id3 instead of searching for a match.
5705 /* ignore the mapping we got from env, use new one */
5706 tl[x].mptr = id3.mptr;
5707 tl[x].mcnt = id3.mcnt;
5708 /* if no active ref, see if we can replace in env */
5711 LOCK_MUTEX0(env->me_rpmutex);
5712 i = mdb_mid3l_search(el, tl[x].mid);
5713 if (el[i].mref == 1) {
5714 /* just us, replace it */
5715 munmap(el[i].mptr, el[i].mcnt * env->me_psize);
5716 el[i].mptr = tl[x].mptr;
5717 el[i].mcnt = tl[x].mcnt;
5719 /* there are others, remove ourself */
5722 UNLOCK_MUTEX(env->me_rpmutex);
5726 id3.mptr = tl[x].mptr;
5727 id3.mcnt = tl[x].mcnt;
5733 if (tl[0].mid >= MDB_TRPAGE_MAX - txn->mt_rpcheck) {
5735 /* purge unref'd pages from our list and unref in env */
5736 LOCK_MUTEX0(env->me_rpmutex);
5739 for (i=1; i<tl[0].mid; i++) {
5742 /* tmp overflow pages don't go to env */
5743 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
5744 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
5747 x = mdb_mid3l_search(el, tl[i].mid);
5751 UNLOCK_MUTEX(env->me_rpmutex);
5753 /* we didn't find any unref'd chunks.
5754 * if we're out of room, fail.
5756 if (tl[0].mid >= MDB_TRPAGE_MAX)
5757 return MDB_TXN_FULL;
5758 /* otherwise, raise threshold for next time around
5761 txn->mt_rpcheck /= 2;
5763 /* we found some unused; consolidate the list */
5764 for (i=y+1; i<= tl[0].mid; i++)
5768 /* decrease the check threshold toward its original value */
5769 if (!txn->mt_rpcheck)
5770 txn->mt_rpcheck = 1;
5771 while (txn->mt_rpcheck < tl[0].mid && txn->mt_rpcheck < MDB_TRPAGE_SIZE/2)
5772 txn->mt_rpcheck *= 2;
5775 if (tl[0].mid < MDB_TRPAGE_SIZE) {
5779 len = env->me_psize * MDB_RPAGE_CHUNK;
5781 id3.mcnt = MDB_RPAGE_CHUNK;
5783 /* search for page in env */
5784 LOCK_MUTEX0(env->me_rpmutex);
5785 x = mdb_mid3l_search(el, pgno);
5786 if (x <= el[0].mid && el[x].mid == pgno) {
5787 id3.mptr = el[x].mptr;
5788 id3.mcnt = el[x].mcnt;
5789 /* check for overflow size */
5790 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5791 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
5792 id3.mcnt = p->mp_pages + rem;
5793 len = id3.mcnt * env->me_psize;
5794 SET_OFF(off, pgno * env->me_psize);
5795 MAP(rc, env, id3.mptr, len, off);
5799 munmap(el[x].mptr, el[x].mcnt);
5800 el[x].mptr = id3.mptr;
5801 el[x].mcnt = id3.mcnt;
5804 UNLOCK_MUTEX(env->me_rpmutex);
5809 UNLOCK_MUTEX(env->me_rpmutex);
5812 if (el[0].mid >= MDB_ERPAGE_MAX - env->me_rpcheck) {
5813 /* purge unref'd pages */
5815 for (i=1; i<el[0].mid; i++) {
5818 munmap(el[i].mptr, env->me_psize * el[i].mcnt);
5823 /* see if we can unref some local pages */
5828 if (el[0].mid >= MDB_ERPAGE_MAX) {
5829 UNLOCK_MUTEX(env->me_rpmutex);
5830 return MDB_MAP_FULL;
5832 env->me_rpcheck /= 2;
5834 for (i=y+1; i<= el[0].mid; i++)
5838 if (!env->me_rpcheck)
5839 env->me_rpcheck = 1;
5840 while (env->me_rpcheck < el[0].mid && env->me_rpcheck < MDB_ERPAGE_SIZE/2)
5841 env->me_rpcheck *= 2;
5844 SET_OFF(off, pgno * env->me_psize);
5845 MAP(rc, env, id3.mptr, len, off);
5848 UNLOCK_MUTEX(env->me_rpmutex);
5851 /* If this page is far enough from the end of the env, scan for
5852 * any overflow pages that would spill onto another block.
5853 * Note we must compare against mt_last_pgno, the last written
5854 * page in the environment. Not mt_next_pgno, which increases
5855 * for every newly allocated (but not yet written) page. If
5856 * we scanned beyond the last written page we'd get a bus error.
5858 if (pgno + MDB_RPAGE_CHUNK <= txn->mt_last_pgno) {
5860 char *cp = (char *)id3.mptr + rem * env->me_psize;
5861 for (i=rem; i<MDB_RPAGE_CHUNK;) {
5863 if (IS_OVERFLOW(p)) {
5864 int nop = p->mp_pages;
5865 if (nop + i > MDB_RPAGE_CHUNK) {
5866 munmap(id3.mptr, len);
5868 len = id3.mcnt * env->me_psize;
5869 MAP(rc, env, id3.mptr, len, off);
5875 cp += nop * env->me_psize;
5878 cp += env->me_psize;
5882 mdb_mid3l_insert(el, &id3);
5883 UNLOCK_MUTEX(env->me_rpmutex);
5885 mdb_mid3l_insert(tl, &id3);
5887 return MDB_TXN_FULL;
5890 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5891 #if MDB_DEBUG /* we don't need this check any more */
5892 if (IS_OVERFLOW(p)) {
5893 mdb_tassert(txn, p->mp_pages + rem <= id3.mcnt);
5901 /** Find the address of the page corresponding to a given page number.
5902 * @param[in] mc the cursor accessing the page.
5903 * @param[in] pgno the page number for the page to retrieve.
5904 * @param[out] ret address of a pointer where the page's address will be stored.
5905 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5906 * @return 0 on success, non-zero on failure.
5909 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
5911 MDB_txn *txn = mc->mc_txn;
5913 MDB_env *env = txn->mt_env;
5918 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5922 MDB_ID2L dl = tx2->mt_u.dirty_list;
5924 /* Spilled pages were dirtied in this txn and flushed
5925 * because the dirty list got full. Bring this page
5926 * back in from the map (but don't unspill it here,
5927 * leave that unless page_touch happens again).
5929 if (tx2->mt_spill_pgs) {
5930 MDB_ID pn = pgno << 1;
5931 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5932 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5934 int rc = mdb_rpage_get(txn, pgno, &p);
5938 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5944 unsigned x = mdb_mid2l_search(dl, pgno);
5945 if (x <= dl[0].mid && dl[x].mid == pgno) {
5951 } while ((tx2 = tx2->mt_parent) != NULL);
5954 if (pgno < txn->mt_next_pgno) {
5958 int rc = mdb_rpage_get(txn, pgno, &p);
5963 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5966 DPRINTF(("page %"Y"u not found", pgno));
5967 txn->mt_flags |= MDB_TXN_ERROR;
5968 return MDB_PAGE_NOTFOUND;
5978 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5979 * The cursor is at the root page, set up the rest of it.
5982 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5984 MDB_page *mp = mc->mc_pg[mc->mc_top];
5988 while (IS_BRANCH(mp)) {
5992 DPRINTF(("branch page %"Y"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5993 /* Don't assert on branch pages in the FreeDB. We can get here
5994 * while in the process of rebalancing a FreeDB branch page; we must
5995 * let that proceed. ITS#8336
5997 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
5998 DPRINTF(("found index 0 to page %"Y"u", NODEPGNO(NODEPTR(mp, 0))));
6000 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
6002 if (flags & MDB_PS_LAST)
6003 i = NUMKEYS(mp) - 1;
6006 node = mdb_node_search(mc, key, &exact);
6008 i = NUMKEYS(mp) - 1;
6010 i = mc->mc_ki[mc->mc_top];
6012 mdb_cassert(mc, i > 0);
6016 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
6019 mdb_cassert(mc, i < NUMKEYS(mp));
6020 node = NODEPTR(mp, i);
6022 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6025 mc->mc_ki[mc->mc_top] = i;
6026 if ((rc = mdb_cursor_push(mc, mp)))
6029 if (flags & MDB_PS_MODIFY) {
6030 if ((rc = mdb_page_touch(mc)) != 0)
6032 mp = mc->mc_pg[mc->mc_top];
6037 DPRINTF(("internal error, index points to a %02X page!?",
6039 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6040 return MDB_CORRUPTED;
6043 DPRINTF(("found leaf page %"Y"u for key [%s]", mp->mp_pgno,
6044 key ? DKEY(key) : "null"));
6045 mc->mc_flags |= C_INITIALIZED;
6046 mc->mc_flags &= ~C_EOF;
6051 /** Search for the lowest key under the current branch page.
6052 * This just bypasses a NUMKEYS check in the current page
6053 * before calling mdb_page_search_root(), because the callers
6054 * are all in situations where the current page is known to
6058 mdb_page_search_lowest(MDB_cursor *mc)
6060 MDB_page *mp = mc->mc_pg[mc->mc_top];
6061 MDB_node *node = NODEPTR(mp, 0);
6064 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6067 mc->mc_ki[mc->mc_top] = 0;
6068 if ((rc = mdb_cursor_push(mc, mp)))
6070 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
6073 /** Search for the page a given key should be in.
6074 * Push it and its parent pages on the cursor stack.
6075 * @param[in,out] mc the cursor for this operation.
6076 * @param[in] key the key to search for, or NULL for first/last page.
6077 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
6078 * are touched (updated with new page numbers).
6079 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
6080 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
6081 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
6082 * @return 0 on success, non-zero on failure.
6085 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
6090 /* Make sure the txn is still viable, then find the root from
6091 * the txn's db table and set it as the root of the cursor's stack.
6093 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
6094 DPUTS("transaction may not be used now");
6097 /* Make sure we're using an up-to-date root */
6098 if (*mc->mc_dbflag & DB_STALE) {
6100 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6102 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
6103 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
6110 MDB_node *leaf = mdb_node_search(&mc2,
6111 &mc->mc_dbx->md_name, &exact);
6113 return MDB_NOTFOUND;
6114 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
6115 return MDB_INCOMPATIBLE; /* not a named DB */
6116 rc = mdb_node_read(&mc2, leaf, &data);
6119 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
6121 /* The txn may not know this DBI, or another process may
6122 * have dropped and recreated the DB with other flags.
6124 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
6125 return MDB_INCOMPATIBLE;
6126 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
6128 *mc->mc_dbflag &= ~DB_STALE;
6130 root = mc->mc_db->md_root;
6132 if (root == P_INVALID) { /* Tree is empty. */
6133 DPUTS("tree is empty");
6134 return MDB_NOTFOUND;
6138 mdb_cassert(mc, root > 1);
6139 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) {
6142 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[0]);
6144 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
6151 for (i=1; i<mc->mc_snum; i++)
6152 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[i]);
6158 DPRINTF(("db %d root page %"Y"u has flags 0x%X",
6159 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
6161 if (flags & MDB_PS_MODIFY) {
6162 if ((rc = mdb_page_touch(mc)))
6166 if (flags & MDB_PS_ROOTONLY)
6169 return mdb_page_search_root(mc, key, flags);
6173 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
6175 MDB_txn *txn = mc->mc_txn;
6176 pgno_t pg = mp->mp_pgno;
6177 unsigned x = 0, ovpages = mp->mp_pages;
6178 MDB_env *env = txn->mt_env;
6179 MDB_IDL sl = txn->mt_spill_pgs;
6180 MDB_ID pn = pg << 1;
6183 DPRINTF(("free ov page %"Y"u (%d)", pg, ovpages));
6184 /* If the page is dirty or on the spill list we just acquired it,
6185 * so we should give it back to our current free list, if any.
6186 * Otherwise put it onto the list of pages we freed in this txn.
6188 * Won't create me_pghead: me_pglast must be inited along with it.
6189 * Unsupported in nested txns: They would need to hide the page
6190 * range in ancestor txns' dirty and spilled lists.
6192 if (env->me_pghead &&
6194 ((mp->mp_flags & P_DIRTY) ||
6195 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
6199 MDB_ID2 *dl, ix, iy;
6200 rc = mdb_midl_need(&env->me_pghead, ovpages);
6203 if (!(mp->mp_flags & P_DIRTY)) {
6204 /* This page is no longer spilled */
6211 /* Remove from dirty list */
6212 dl = txn->mt_u.dirty_list;
6214 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
6220 mdb_cassert(mc, x > 1);
6222 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
6223 txn->mt_flags |= MDB_TXN_ERROR;
6224 return MDB_CORRUPTED;
6227 txn->mt_dirty_room++;
6228 if (!(env->me_flags & MDB_WRITEMAP))
6229 mdb_dpage_free(env, mp);
6231 /* Insert in me_pghead */
6232 mop = env->me_pghead;
6233 j = mop[0] + ovpages;
6234 for (i = mop[0]; i && mop[i] < pg; i--)
6240 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
6244 mc->mc_db->md_overflow_pages -= ovpages;
6248 /** Return the data associated with a given node.
6249 * @param[in] mc The cursor for this operation.
6250 * @param[in] leaf The node being read.
6251 * @param[out] data Updated to point to the node's data.
6252 * @return 0 on success, non-zero on failure.
6255 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
6257 MDB_page *omp; /* overflow page */
6261 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6262 data->mv_size = NODEDSZ(leaf);
6263 data->mv_data = NODEDATA(leaf);
6267 /* Read overflow data.
6269 data->mv_size = NODEDSZ(leaf);
6270 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
6271 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
6272 DPRINTF(("read overflow page %"Y"u failed", pgno));
6275 data->mv_data = METADATA(omp);
6281 mdb_get(MDB_txn *txn, MDB_dbi dbi,
6282 MDB_val *key, MDB_val *data)
6289 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
6291 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
6294 if (txn->mt_flags & MDB_TXN_BLOCKED)
6297 mdb_cursor_init(&mc, txn, dbi, &mx);
6298 rc = mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
6301 /* unref all the pages - caller must copy the data
6302 * before doing anything else
6304 mdb_cursor_unref(&mc);
6310 /** Find a sibling for a page.
6311 * Replaces the page at the top of the cursor's stack with the
6312 * specified sibling, if one exists.
6313 * @param[in] mc The cursor for this operation.
6314 * @param[in] move_right Non-zero if the right sibling is requested,
6315 * otherwise the left sibling.
6316 * @return 0 on success, non-zero on failure.
6319 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
6328 if (mc->mc_snum < 2) {
6329 return MDB_NOTFOUND; /* root has no siblings */
6333 op = mc->mc_pg[mc->mc_top];
6336 DPRINTF(("parent page is page %"Y"u, index %u",
6337 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
6339 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6340 : (mc->mc_ki[mc->mc_top] == 0)) {
6341 DPRINTF(("no more keys left, moving to %s sibling",
6342 move_right ? "right" : "left"));
6343 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
6344 /* undo cursor_pop before returning */
6351 mc->mc_ki[mc->mc_top]++;
6353 mc->mc_ki[mc->mc_top]--;
6354 DPRINTF(("just moving to %s index key %u",
6355 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
6357 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
6359 MDB_PAGE_UNREF(mc->mc_txn, op);
6361 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6362 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
6363 /* mc will be inconsistent if caller does mc_snum++ as above */
6364 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
6368 mdb_cursor_push(mc, mp);
6370 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
6375 /** Move the cursor to the next data item. */
6377 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6383 if (mc->mc_flags & C_EOF) {
6384 return MDB_NOTFOUND;
6387 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
6389 mp = mc->mc_pg[mc->mc_top];
6391 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6392 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6393 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6394 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
6395 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
6396 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
6397 if (rc == MDB_SUCCESS)
6398 MDB_GET_KEY(leaf, key);
6404 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6405 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
6410 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6411 if (op == MDB_NEXT_DUP)
6412 return MDB_NOTFOUND;
6416 DPRINTF(("cursor_next: top page is %"Y"u in cursor %p",
6417 mdb_dbg_pgno(mp), (void *) mc));
6418 if (mc->mc_flags & C_DEL)
6421 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
6422 DPUTS("=====> move to next sibling page");
6423 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6424 mc->mc_flags |= C_EOF;
6427 mp = mc->mc_pg[mc->mc_top];
6428 DPRINTF(("next page is %"Y"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6430 mc->mc_ki[mc->mc_top]++;
6433 DPRINTF(("==> cursor points to page %"Y"u with %u keys, key index %u",
6434 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6437 key->mv_size = mc->mc_db->md_pad;
6438 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6442 mdb_cassert(mc, IS_LEAF(mp));
6443 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6445 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6446 mdb_xcursor_init1(mc, leaf);
6449 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6452 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6453 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6454 if (rc != MDB_SUCCESS)
6459 MDB_GET_KEY(leaf, key);
6463 /** Move the cursor to the previous data item. */
6465 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6471 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
6473 mp = mc->mc_pg[mc->mc_top];
6475 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6476 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6477 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6478 if (op == MDB_PREV || op == MDB_PREV_DUP) {
6479 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
6480 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
6481 if (rc == MDB_SUCCESS) {
6482 MDB_GET_KEY(leaf, key);
6483 mc->mc_flags &= ~C_EOF;
6490 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6491 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
6496 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6497 if (op == MDB_PREV_DUP)
6498 return MDB_NOTFOUND;
6502 DPRINTF(("cursor_prev: top page is %"Y"u in cursor %p",
6503 mdb_dbg_pgno(mp), (void *) mc));
6505 if (mc->mc_ki[mc->mc_top] == 0) {
6506 DPUTS("=====> move to prev sibling page");
6507 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
6510 mp = mc->mc_pg[mc->mc_top];
6511 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
6512 DPRINTF(("prev page is %"Y"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6514 mc->mc_ki[mc->mc_top]--;
6516 mc->mc_flags &= ~C_EOF;
6518 DPRINTF(("==> cursor points to page %"Y"u with %u keys, key index %u",
6519 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6522 key->mv_size = mc->mc_db->md_pad;
6523 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6527 mdb_cassert(mc, IS_LEAF(mp));
6528 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6530 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6531 mdb_xcursor_init1(mc, leaf);
6534 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6537 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6538 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6539 if (rc != MDB_SUCCESS)
6544 MDB_GET_KEY(leaf, key);
6548 /** Set the cursor on a specific data item. */
6550 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6551 MDB_cursor_op op, int *exactp)
6555 MDB_node *leaf = NULL;
6558 if (key->mv_size == 0)
6559 return MDB_BAD_VALSIZE;
6562 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6564 /* See if we're already on the right page */
6565 if (mc->mc_flags & C_INITIALIZED) {
6568 mp = mc->mc_pg[mc->mc_top];
6570 mc->mc_ki[mc->mc_top] = 0;
6571 return MDB_NOTFOUND;
6573 if (mp->mp_flags & P_LEAF2) {
6574 nodekey.mv_size = mc->mc_db->md_pad;
6575 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
6577 leaf = NODEPTR(mp, 0);
6578 MDB_GET_KEY2(leaf, nodekey);
6580 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6582 /* Probably happens rarely, but first node on the page
6583 * was the one we wanted.
6585 mc->mc_ki[mc->mc_top] = 0;
6592 unsigned int nkeys = NUMKEYS(mp);
6594 if (mp->mp_flags & P_LEAF2) {
6595 nodekey.mv_data = LEAF2KEY(mp,
6596 nkeys-1, nodekey.mv_size);
6598 leaf = NODEPTR(mp, nkeys-1);
6599 MDB_GET_KEY2(leaf, nodekey);
6601 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6603 /* last node was the one we wanted */
6604 mc->mc_ki[mc->mc_top] = nkeys-1;
6610 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6611 /* This is definitely the right page, skip search_page */
6612 if (mp->mp_flags & P_LEAF2) {
6613 nodekey.mv_data = LEAF2KEY(mp,
6614 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6616 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6617 MDB_GET_KEY2(leaf, nodekey);
6619 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6621 /* current node was the one we wanted */
6631 /* If any parents have right-sibs, search.
6632 * Otherwise, there's nothing further.
6634 for (i=0; i<mc->mc_top; i++)
6636 NUMKEYS(mc->mc_pg[i])-1)
6638 if (i == mc->mc_top) {
6639 /* There are no other pages */
6640 mc->mc_ki[mc->mc_top] = nkeys;
6641 return MDB_NOTFOUND;
6645 /* There are no other pages */
6646 mc->mc_ki[mc->mc_top] = 0;
6647 if (op == MDB_SET_RANGE && !exactp) {
6651 return MDB_NOTFOUND;
6657 rc = mdb_page_search(mc, key, 0);
6658 if (rc != MDB_SUCCESS)
6661 mp = mc->mc_pg[mc->mc_top];
6662 mdb_cassert(mc, IS_LEAF(mp));
6665 leaf = mdb_node_search(mc, key, exactp);
6666 if (exactp != NULL && !*exactp) {
6667 /* MDB_SET specified and not an exact match. */
6668 return MDB_NOTFOUND;
6672 DPUTS("===> inexact leaf not found, goto sibling");
6673 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6674 mc->mc_flags |= C_EOF;
6675 return rc; /* no entries matched */
6677 mp = mc->mc_pg[mc->mc_top];
6678 mdb_cassert(mc, IS_LEAF(mp));
6679 leaf = NODEPTR(mp, 0);
6683 mc->mc_flags |= C_INITIALIZED;
6684 mc->mc_flags &= ~C_EOF;
6687 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6688 key->mv_size = mc->mc_db->md_pad;
6689 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6695 if (mc->mc_xcursor && mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6696 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
6699 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6700 mdb_xcursor_init1(mc, leaf);
6703 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6704 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6705 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6708 if (op == MDB_GET_BOTH) {
6714 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6715 if (rc != MDB_SUCCESS)
6718 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6721 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6723 dcmp = mc->mc_dbx->md_dcmp;
6724 #if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
6725 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(mdb_size_t))
6726 dcmp = mdb_cmp_clong;
6728 rc = dcmp(data, &olddata);
6730 if (op == MDB_GET_BOTH || rc > 0)
6731 return MDB_NOTFOUND;
6738 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6739 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6744 /* The key already matches in all other cases */
6745 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6746 MDB_GET_KEY(leaf, key);
6747 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6752 /** Move the cursor to the first item in the database. */
6754 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6759 if (mc->mc_xcursor) {
6761 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6762 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
6765 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6768 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6769 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6770 if (rc != MDB_SUCCESS)
6773 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6775 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6776 mc->mc_flags |= C_INITIALIZED;
6777 mc->mc_flags &= ~C_EOF;
6779 mc->mc_ki[mc->mc_top] = 0;
6781 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6782 key->mv_size = mc->mc_db->md_pad;
6783 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6788 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6789 mdb_xcursor_init1(mc, leaf);
6790 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6794 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6798 MDB_GET_KEY(leaf, key);
6802 /** Move the cursor to the last item in the database. */
6804 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6809 if (mc->mc_xcursor) {
6811 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6812 mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
6815 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6818 if (!(mc->mc_flags & C_EOF)) {
6820 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6821 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6822 if (rc != MDB_SUCCESS)
6825 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6828 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6829 mc->mc_flags |= C_INITIALIZED|C_EOF;
6830 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6832 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6833 key->mv_size = mc->mc_db->md_pad;
6834 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6839 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6840 mdb_xcursor_init1(mc, leaf);
6841 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6845 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6850 MDB_GET_KEY(leaf, key);
6855 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6860 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6865 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6869 case MDB_GET_CURRENT:
6870 if (!(mc->mc_flags & C_INITIALIZED)) {
6873 MDB_page *mp = mc->mc_pg[mc->mc_top];
6874 int nkeys = NUMKEYS(mp);
6875 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6876 mc->mc_ki[mc->mc_top] = nkeys;
6882 key->mv_size = mc->mc_db->md_pad;
6883 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6885 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6886 MDB_GET_KEY(leaf, key);
6888 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6889 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6891 rc = mdb_node_read(mc, leaf, data);
6898 case MDB_GET_BOTH_RANGE:
6903 if (mc->mc_xcursor == NULL) {
6904 rc = MDB_INCOMPATIBLE;
6914 rc = mdb_cursor_set(mc, key, data, op,
6915 op == MDB_SET_RANGE ? NULL : &exact);
6918 case MDB_GET_MULTIPLE:
6919 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6923 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6924 rc = MDB_INCOMPATIBLE;
6928 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6929 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6932 case MDB_NEXT_MULTIPLE:
6937 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6938 rc = MDB_INCOMPATIBLE;
6941 if (!(mc->mc_flags & C_INITIALIZED))
6942 rc = mdb_cursor_first(mc, key, data);
6944 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6945 if (rc == MDB_SUCCESS) {
6946 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6949 mx = &mc->mc_xcursor->mx_cursor;
6950 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6952 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6953 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6961 case MDB_NEXT_NODUP:
6962 if (!(mc->mc_flags & C_INITIALIZED))
6963 rc = mdb_cursor_first(mc, key, data);
6965 rc = mdb_cursor_next(mc, key, data, op);
6969 case MDB_PREV_NODUP:
6970 if (!(mc->mc_flags & C_INITIALIZED)) {
6971 rc = mdb_cursor_last(mc, key, data);
6974 mc->mc_flags |= C_INITIALIZED;
6975 mc->mc_ki[mc->mc_top]++;
6977 rc = mdb_cursor_prev(mc, key, data, op);
6980 rc = mdb_cursor_first(mc, key, data);
6983 mfunc = mdb_cursor_first;
6985 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6989 if (mc->mc_xcursor == NULL) {
6990 rc = MDB_INCOMPATIBLE;
6994 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6995 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6996 MDB_GET_KEY(leaf, key);
6997 rc = mdb_node_read(mc, leaf, data);
7001 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7005 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
7008 rc = mdb_cursor_last(mc, key, data);
7011 mfunc = mdb_cursor_last;
7014 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
7019 if (mc->mc_flags & C_DEL)
7020 mc->mc_flags ^= C_DEL;
7025 /** Touch all the pages in the cursor stack. Set mc_top.
7026 * Makes sure all the pages are writable, before attempting a write operation.
7027 * @param[in] mc The cursor to operate on.
7030 mdb_cursor_touch(MDB_cursor *mc)
7032 int rc = MDB_SUCCESS;
7034 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
7037 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
7039 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
7040 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
7043 *mc->mc_dbflag |= DB_DIRTY;
7048 rc = mdb_page_touch(mc);
7049 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
7050 mc->mc_top = mc->mc_snum-1;
7055 /** Do not spill pages to disk if txn is getting full, may fail instead */
7056 #define MDB_NOSPILL 0x8000
7059 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
7063 MDB_node *leaf = NULL;
7064 MDB_page *fp, *mp, *sub_root = NULL;
7066 MDB_val xdata, *rdata, dkey, olddata;
7068 int do_sub = 0, insert_key, insert_data;
7069 unsigned int mcount = 0, dcount = 0, nospill;
7072 unsigned int nflags;
7075 if (mc == NULL || key == NULL)
7078 env = mc->mc_txn->mt_env;
7080 /* Check this first so counter will always be zero on any
7083 if (flags & MDB_MULTIPLE) {
7084 dcount = data[1].mv_size;
7085 data[1].mv_size = 0;
7086 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
7087 return MDB_INCOMPATIBLE;
7090 nospill = flags & MDB_NOSPILL;
7091 flags &= ~MDB_NOSPILL;
7093 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7094 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7096 if (key->mv_size-1 >= ENV_MAXKEY(env))
7097 return MDB_BAD_VALSIZE;
7099 #if SIZE_MAX > MAXDATASIZE
7100 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
7101 return MDB_BAD_VALSIZE;
7103 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
7104 return MDB_BAD_VALSIZE;
7107 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
7108 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
7112 if (flags == MDB_CURRENT) {
7113 if (!(mc->mc_flags & C_INITIALIZED))
7116 } else if (mc->mc_db->md_root == P_INVALID) {
7117 /* new database, cursor has nothing to point to */
7120 mc->mc_flags &= ~C_INITIALIZED;
7125 if (flags & MDB_APPEND) {
7127 rc = mdb_cursor_last(mc, &k2, &d2);
7129 rc = mc->mc_dbx->md_cmp(key, &k2);
7132 mc->mc_ki[mc->mc_top]++;
7134 /* new key is <= last key */
7139 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
7141 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
7142 DPRINTF(("duplicate key [%s]", DKEY(key)));
7144 return MDB_KEYEXIST;
7146 if (rc && rc != MDB_NOTFOUND)
7150 if (mc->mc_flags & C_DEL)
7151 mc->mc_flags ^= C_DEL;
7153 /* Cursor is positioned, check for room in the dirty list */
7155 if (flags & MDB_MULTIPLE) {
7157 xdata.mv_size = data->mv_size * dcount;
7161 if ((rc2 = mdb_page_spill(mc, key, rdata)))
7165 if (rc == MDB_NO_ROOT) {
7167 /* new database, write a root leaf page */
7168 DPUTS("allocating new root leaf page");
7169 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
7172 mdb_cursor_push(mc, np);
7173 mc->mc_db->md_root = np->mp_pgno;
7174 mc->mc_db->md_depth++;
7175 *mc->mc_dbflag |= DB_DIRTY;
7176 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
7178 np->mp_flags |= P_LEAF2;
7179 mc->mc_flags |= C_INITIALIZED;
7181 /* make sure all cursor pages are writable */
7182 rc2 = mdb_cursor_touch(mc);
7187 insert_key = insert_data = rc;
7189 /* The key does not exist */
7190 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
7191 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
7192 LEAFSIZE(key, data) > env->me_nodemax)
7194 /* Too big for a node, insert in sub-DB. Set up an empty
7195 * "old sub-page" for prep_subDB to expand to a full page.
7197 fp_flags = P_LEAF|P_DIRTY;
7199 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
7200 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
7201 olddata.mv_size = PAGEHDRSZ;
7205 /* there's only a key anyway, so this is a no-op */
7206 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
7208 unsigned int ksize = mc->mc_db->md_pad;
7209 if (key->mv_size != ksize)
7210 return MDB_BAD_VALSIZE;
7211 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
7212 memcpy(ptr, key->mv_data, ksize);
7214 /* if overwriting slot 0 of leaf, need to
7215 * update branch key if there is a parent page
7217 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7218 unsigned short dtop = 1;
7220 /* slot 0 is always an empty key, find real slot */
7221 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7225 if (mc->mc_ki[mc->mc_top])
7226 rc2 = mdb_update_key(mc, key);
7237 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7238 olddata.mv_size = NODEDSZ(leaf);
7239 olddata.mv_data = NODEDATA(leaf);
7242 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
7243 /* Prepare (sub-)page/sub-DB to accept the new item,
7244 * if needed. fp: old sub-page or a header faking
7245 * it. mp: new (sub-)page. offset: growth in page
7246 * size. xdata: node data with new page or DB.
7248 unsigned i, offset = 0;
7249 mp = fp = xdata.mv_data = env->me_pbuf;
7250 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
7252 /* Was a single item before, must convert now */
7253 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7255 /* Just overwrite the current item */
7256 if (flags == MDB_CURRENT)
7258 dcmp = mc->mc_dbx->md_dcmp;
7259 #if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
7260 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(mdb_size_t))
7261 dcmp = mdb_cmp_clong;
7263 /* does data match? */
7264 if (!dcmp(data, &olddata)) {
7265 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
7266 return MDB_KEYEXIST;
7271 /* Back up original data item */
7272 dkey.mv_size = olddata.mv_size;
7273 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
7275 /* Make sub-page header for the dup items, with dummy body */
7276 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
7277 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
7278 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
7279 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7280 fp->mp_flags |= P_LEAF2;
7281 fp->mp_pad = data->mv_size;
7282 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
7284 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
7285 (dkey.mv_size & 1) + (data->mv_size & 1);
7287 fp->mp_upper = xdata.mv_size - PAGEBASE;
7288 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
7289 } else if (leaf->mn_flags & F_SUBDATA) {
7290 /* Data is on sub-DB, just store it */
7291 flags |= F_DUPDATA|F_SUBDATA;
7294 /* Data is on sub-page */
7295 fp = olddata.mv_data;
7298 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7299 offset = EVEN(NODESIZE + sizeof(indx_t) +
7303 offset = fp->mp_pad;
7304 if (SIZELEFT(fp) < offset) {
7305 offset *= 4; /* space for 4 more */
7308 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
7310 fp->mp_flags |= P_DIRTY;
7311 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
7312 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
7316 xdata.mv_size = olddata.mv_size + offset;
7319 fp_flags = fp->mp_flags;
7320 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
7321 /* Too big for a sub-page, convert to sub-DB */
7322 fp_flags &= ~P_SUBP;
7324 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7325 fp_flags |= P_LEAF2;
7326 dummy.md_pad = fp->mp_pad;
7327 dummy.md_flags = MDB_DUPFIXED;
7328 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7329 dummy.md_flags |= MDB_INTEGERKEY;
7335 dummy.md_branch_pages = 0;
7336 dummy.md_leaf_pages = 1;
7337 dummy.md_overflow_pages = 0;
7338 dummy.md_entries = NUMKEYS(fp);
7339 xdata.mv_size = sizeof(MDB_db);
7340 xdata.mv_data = &dummy;
7341 if ((rc = mdb_page_alloc(mc, 1, &mp)))
7343 offset = env->me_psize - olddata.mv_size;
7344 flags |= F_DUPDATA|F_SUBDATA;
7345 dummy.md_root = mp->mp_pgno;
7349 mp->mp_flags = fp_flags | P_DIRTY;
7350 mp->mp_pad = fp->mp_pad;
7351 mp->mp_lower = fp->mp_lower;
7352 mp->mp_upper = fp->mp_upper + offset;
7353 if (fp_flags & P_LEAF2) {
7354 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
7356 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
7357 olddata.mv_size - fp->mp_upper - PAGEBASE);
7358 for (i=0; i<NUMKEYS(fp); i++)
7359 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
7367 mdb_node_del(mc, 0);
7371 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
7372 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
7373 return MDB_INCOMPATIBLE;
7374 /* overflow page overwrites need special handling */
7375 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7378 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
7380 memcpy(&pg, olddata.mv_data, sizeof(pg));
7381 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
7383 ovpages = omp->mp_pages;
7385 /* Is the ov page large enough? */
7386 if (ovpages >= dpages) {
7387 if (!(omp->mp_flags & P_DIRTY) &&
7388 (level || (env->me_flags & MDB_WRITEMAP)))
7390 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
7393 level = 0; /* dirty in this txn or clean */
7396 if (omp->mp_flags & P_DIRTY) {
7397 /* yes, overwrite it. Note in this case we don't
7398 * bother to try shrinking the page if the new data
7399 * is smaller than the overflow threshold.
7402 /* It is writable only in a parent txn */
7403 size_t sz = (size_t) env->me_psize * ovpages, off;
7404 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
7410 /* Note - this page is already counted in parent's dirty_room */
7411 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
7412 mdb_cassert(mc, rc2 == 0);
7413 if (!(flags & MDB_RESERVE)) {
7414 /* Copy end of page, adjusting alignment so
7415 * compiler may copy words instead of bytes.
7417 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
7418 memcpy((size_t *)((char *)np + off),
7419 (size_t *)((char *)omp + off), sz - off);
7422 memcpy(np, omp, sz); /* Copy beginning of page */
7425 SETDSZ(leaf, data->mv_size);
7426 if (F_ISSET(flags, MDB_RESERVE))
7427 data->mv_data = METADATA(omp);
7429 memcpy(METADATA(omp), data->mv_data, data->mv_size);
7433 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
7435 } else if (data->mv_size == olddata.mv_size) {
7436 /* same size, just replace it. Note that we could
7437 * also reuse this node if the new data is smaller,
7438 * but instead we opt to shrink the node in that case.
7440 if (F_ISSET(flags, MDB_RESERVE))
7441 data->mv_data = olddata.mv_data;
7442 else if (!(mc->mc_flags & C_SUB))
7443 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
7445 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
7450 mdb_node_del(mc, 0);
7456 nflags = flags & NODE_ADD_FLAGS;
7457 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
7458 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
7459 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
7460 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
7462 nflags |= MDB_SPLIT_REPLACE;
7463 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
7465 /* There is room already in this leaf page. */
7466 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
7468 /* Adjust other cursors pointing to mp */
7469 MDB_cursor *m2, *m3;
7470 MDB_dbi dbi = mc->mc_dbi;
7471 unsigned i = mc->mc_top;
7472 MDB_page *mp = mc->mc_pg[i];
7474 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7475 if (mc->mc_flags & C_SUB)
7476 m3 = &m2->mc_xcursor->mx_cursor;
7479 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
7480 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
7483 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7484 MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
7485 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7486 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7492 if (rc == MDB_SUCCESS) {
7493 /* Now store the actual data in the child DB. Note that we're
7494 * storing the user data in the keys field, so there are strict
7495 * size limits on dupdata. The actual data fields of the child
7496 * DB are all zero size.
7499 int xflags, new_dupdata;
7504 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7505 if (flags & MDB_CURRENT) {
7506 xflags = MDB_CURRENT|MDB_NOSPILL;
7508 mdb_xcursor_init1(mc, leaf);
7509 xflags = (flags & MDB_NODUPDATA) ?
7510 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
7513 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
7514 new_dupdata = (int)dkey.mv_size;
7515 /* converted, write the original data first */
7517 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
7520 /* we've done our job */
7523 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
7524 /* Adjust other cursors pointing to mp */
7526 MDB_xcursor *mx = mc->mc_xcursor;
7527 unsigned i = mc->mc_top;
7528 MDB_page *mp = mc->mc_pg[i];
7529 int nkeys = NUMKEYS(mp);
7531 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7532 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7533 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7534 if (m2->mc_pg[i] == mp) {
7535 if (m2->mc_ki[i] == mc->mc_ki[i]) {
7536 mdb_xcursor_init2(m2, mx, new_dupdata);
7537 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
7538 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
7539 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7540 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7545 ecount = mc->mc_xcursor->mx_db.md_entries;
7546 if (flags & MDB_APPENDDUP)
7547 xflags |= MDB_APPEND;
7548 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
7549 if (flags & F_SUBDATA) {
7550 void *db = NODEDATA(leaf);
7551 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7553 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
7555 /* Increment count unless we just replaced an existing item. */
7557 mc->mc_db->md_entries++;
7559 /* Invalidate txn if we created an empty sub-DB */
7562 /* If we succeeded and the key didn't exist before,
7563 * make sure the cursor is marked valid.
7565 mc->mc_flags |= C_INITIALIZED;
7567 if (flags & MDB_MULTIPLE) {
7570 /* let caller know how many succeeded, if any */
7571 data[1].mv_size = mcount;
7572 if (mcount < dcount) {
7573 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
7574 insert_key = insert_data = 0;
7581 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
7584 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7589 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7595 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7596 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7598 if (!(mc->mc_flags & C_INITIALIZED))
7601 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7602 return MDB_NOTFOUND;
7604 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7607 rc = mdb_cursor_touch(mc);
7611 mp = mc->mc_pg[mc->mc_top];
7614 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7616 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7617 if (flags & MDB_NODUPDATA) {
7618 /* mdb_cursor_del0() will subtract the final entry */
7619 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7620 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7622 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7623 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7625 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7628 /* If sub-DB still has entries, we're done */
7629 if (mc->mc_xcursor->mx_db.md_entries) {
7630 if (leaf->mn_flags & F_SUBDATA) {
7631 /* update subDB info */
7632 void *db = NODEDATA(leaf);
7633 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7636 /* shrink fake page */
7637 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7638 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7639 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7640 /* fix other sub-DB cursors pointed at fake pages on this page */
7641 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7642 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7643 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7644 if (m2->mc_pg[mc->mc_top] == mp) {
7645 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
7646 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7648 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7649 if (!(n2->mn_flags & F_SUBDATA))
7650 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7655 mc->mc_db->md_entries--;
7658 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7660 /* otherwise fall thru and delete the sub-DB */
7663 if (leaf->mn_flags & F_SUBDATA) {
7664 /* add all the child DB's pages to the free list */
7665 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7670 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7671 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7672 rc = MDB_INCOMPATIBLE;
7676 /* add overflow pages to free list */
7677 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7681 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7682 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7683 (rc = mdb_ovpage_free(mc, omp)))
7688 return mdb_cursor_del0(mc);
7691 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7695 /** Allocate and initialize new pages for a database.
7696 * @param[in] mc a cursor on the database being added to.
7697 * @param[in] flags flags defining what type of page is being allocated.
7698 * @param[in] num the number of pages to allocate. This is usually 1,
7699 * unless allocating overflow pages for a large record.
7700 * @param[out] mp Address of a page, or NULL on failure.
7701 * @return 0 on success, non-zero on failure.
7704 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7709 if ((rc = mdb_page_alloc(mc, num, &np)))
7711 DPRINTF(("allocated new mpage %"Y"u, page size %u",
7712 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7713 np->mp_flags = flags | P_DIRTY;
7714 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7715 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7718 mc->mc_db->md_branch_pages++;
7719 else if (IS_LEAF(np))
7720 mc->mc_db->md_leaf_pages++;
7721 else if (IS_OVERFLOW(np)) {
7722 mc->mc_db->md_overflow_pages += num;
7730 /** Calculate the size of a leaf node.
7731 * The size depends on the environment's page size; if a data item
7732 * is too large it will be put onto an overflow page and the node
7733 * size will only include the key and not the data. Sizes are always
7734 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7735 * of the #MDB_node headers.
7736 * @param[in] env The environment handle.
7737 * @param[in] key The key for the node.
7738 * @param[in] data The data for the node.
7739 * @return The number of bytes needed to store the node.
7742 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7746 sz = LEAFSIZE(key, data);
7747 if (sz > env->me_nodemax) {
7748 /* put on overflow page */
7749 sz -= data->mv_size - sizeof(pgno_t);
7752 return EVEN(sz + sizeof(indx_t));
7755 /** Calculate the size of a branch node.
7756 * The size should depend on the environment's page size but since
7757 * we currently don't support spilling large keys onto overflow
7758 * pages, it's simply the size of the #MDB_node header plus the
7759 * size of the key. Sizes are always rounded up to an even number
7760 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7761 * @param[in] env The environment handle.
7762 * @param[in] key The key for the node.
7763 * @return The number of bytes needed to store the node.
7766 mdb_branch_size(MDB_env *env, MDB_val *key)
7771 if (sz > env->me_nodemax) {
7772 /* put on overflow page */
7773 /* not implemented */
7774 /* sz -= key->size - sizeof(pgno_t); */
7777 return sz + sizeof(indx_t);
7780 /** Add a node to the page pointed to by the cursor.
7781 * @param[in] mc The cursor for this operation.
7782 * @param[in] indx The index on the page where the new node should be added.
7783 * @param[in] key The key for the new node.
7784 * @param[in] data The data for the new node, if any.
7785 * @param[in] pgno The page number, if adding a branch node.
7786 * @param[in] flags Flags for the node.
7787 * @return 0 on success, non-zero on failure. Possible errors are:
7789 * <li>ENOMEM - failed to allocate overflow pages for the node.
7790 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7791 * should never happen since all callers already calculate the
7792 * page's free space before calling this function.
7796 mdb_node_add(MDB_cursor *mc, indx_t indx,
7797 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7800 size_t node_size = NODESIZE;
7804 MDB_page *mp = mc->mc_pg[mc->mc_top];
7805 MDB_page *ofp = NULL; /* overflow page */
7809 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7811 DPRINTF(("add to %s %spage %"Y"u index %i, data size %"Z"u key size %"Z"u [%s]",
7812 IS_LEAF(mp) ? "leaf" : "branch",
7813 IS_SUBP(mp) ? "sub-" : "",
7814 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7815 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7818 /* Move higher keys up one slot. */
7819 int ksize = mc->mc_db->md_pad, dif;
7820 char *ptr = LEAF2KEY(mp, indx, ksize);
7821 dif = NUMKEYS(mp) - indx;
7823 memmove(ptr+ksize, ptr, dif*ksize);
7824 /* insert new key */
7825 memcpy(ptr, key->mv_data, ksize);
7827 /* Just using these for counting */
7828 mp->mp_lower += sizeof(indx_t);
7829 mp->mp_upper -= ksize - sizeof(indx_t);
7833 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7835 node_size += key->mv_size;
7837 mdb_cassert(mc, key && data);
7838 if (F_ISSET(flags, F_BIGDATA)) {
7839 /* Data already on overflow page. */
7840 node_size += sizeof(pgno_t);
7841 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7842 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7844 /* Put data on overflow page. */
7845 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7846 data->mv_size, node_size+data->mv_size));
7847 node_size = EVEN(node_size + sizeof(pgno_t));
7848 if ((ssize_t)node_size > room)
7850 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7852 DPRINTF(("allocated overflow page %"Y"u", ofp->mp_pgno));
7856 node_size += data->mv_size;
7859 node_size = EVEN(node_size);
7860 if ((ssize_t)node_size > room)
7864 /* Move higher pointers up one slot. */
7865 for (i = NUMKEYS(mp); i > indx; i--)
7866 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7868 /* Adjust free space offsets. */
7869 ofs = mp->mp_upper - node_size;
7870 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7871 mp->mp_ptrs[indx] = ofs;
7873 mp->mp_lower += sizeof(indx_t);
7875 /* Write the node data. */
7876 node = NODEPTR(mp, indx);
7877 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7878 node->mn_flags = flags;
7880 SETDSZ(node,data->mv_size);
7885 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7888 ndata = NODEDATA(node);
7890 if (F_ISSET(flags, F_BIGDATA))
7891 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7892 else if (F_ISSET(flags, MDB_RESERVE))
7893 data->mv_data = ndata;
7895 memcpy(ndata, data->mv_data, data->mv_size);
7897 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7898 ndata = METADATA(ofp);
7899 if (F_ISSET(flags, MDB_RESERVE))
7900 data->mv_data = ndata;
7902 memcpy(ndata, data->mv_data, data->mv_size);
7909 DPRINTF(("not enough room in page %"Y"u, got %u ptrs",
7910 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7911 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7912 DPRINTF(("node size = %"Z"u", node_size));
7913 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7914 return MDB_PAGE_FULL;
7917 /** Delete the specified node from a page.
7918 * @param[in] mc Cursor pointing to the node to delete.
7919 * @param[in] ksize The size of a node. Only used if the page is
7920 * part of a #MDB_DUPFIXED database.
7923 mdb_node_del(MDB_cursor *mc, int ksize)
7925 MDB_page *mp = mc->mc_pg[mc->mc_top];
7926 indx_t indx = mc->mc_ki[mc->mc_top];
7928 indx_t i, j, numkeys, ptr;
7932 DPRINTF(("delete node %u on %s page %"Y"u", indx,
7933 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7934 numkeys = NUMKEYS(mp);
7935 mdb_cassert(mc, indx < numkeys);
7938 int x = numkeys - 1 - indx;
7939 base = LEAF2KEY(mp, indx, ksize);
7941 memmove(base, base + ksize, x * ksize);
7942 mp->mp_lower -= sizeof(indx_t);
7943 mp->mp_upper += ksize - sizeof(indx_t);
7947 node = NODEPTR(mp, indx);
7948 sz = NODESIZE + node->mn_ksize;
7950 if (F_ISSET(node->mn_flags, F_BIGDATA))
7951 sz += sizeof(pgno_t);
7953 sz += NODEDSZ(node);
7957 ptr = mp->mp_ptrs[indx];
7958 for (i = j = 0; i < numkeys; i++) {
7960 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7961 if (mp->mp_ptrs[i] < ptr)
7962 mp->mp_ptrs[j] += sz;
7967 base = (char *)mp + mp->mp_upper + PAGEBASE;
7968 memmove(base + sz, base, ptr - mp->mp_upper);
7970 mp->mp_lower -= sizeof(indx_t);
7974 /** Compact the main page after deleting a node on a subpage.
7975 * @param[in] mp The main page to operate on.
7976 * @param[in] indx The index of the subpage on the main page.
7979 mdb_node_shrink(MDB_page *mp, indx_t indx)
7984 indx_t delta, nsize, len, ptr;
7987 node = NODEPTR(mp, indx);
7988 sp = (MDB_page *)NODEDATA(node);
7989 delta = SIZELEFT(sp);
7990 nsize = NODEDSZ(node) - delta;
7992 /* Prepare to shift upward, set len = length(subpage part to shift) */
7996 return; /* do not make the node uneven-sized */
7998 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7999 for (i = NUMKEYS(sp); --i >= 0; )
8000 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
8003 sp->mp_upper = sp->mp_lower;
8004 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
8005 SETDSZ(node, nsize);
8007 /* Shift <lower nodes...initial part of subpage> upward */
8008 base = (char *)mp + mp->mp_upper + PAGEBASE;
8009 memmove(base + delta, base, (char *)sp + len - base);
8011 ptr = mp->mp_ptrs[indx];
8012 for (i = NUMKEYS(mp); --i >= 0; ) {
8013 if (mp->mp_ptrs[i] <= ptr)
8014 mp->mp_ptrs[i] += delta;
8016 mp->mp_upper += delta;
8019 /** Initial setup of a sorted-dups cursor.
8020 * Sorted duplicates are implemented as a sub-database for the given key.
8021 * The duplicate data items are actually keys of the sub-database.
8022 * Operations on the duplicate data items are performed using a sub-cursor
8023 * initialized when the sub-database is first accessed. This function does
8024 * the preliminary setup of the sub-cursor, filling in the fields that
8025 * depend only on the parent DB.
8026 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8029 mdb_xcursor_init0(MDB_cursor *mc)
8031 MDB_xcursor *mx = mc->mc_xcursor;
8033 mx->mx_cursor.mc_xcursor = NULL;
8034 mx->mx_cursor.mc_txn = mc->mc_txn;
8035 mx->mx_cursor.mc_db = &mx->mx_db;
8036 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
8037 mx->mx_cursor.mc_dbi = mc->mc_dbi;
8038 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
8039 mx->mx_cursor.mc_snum = 0;
8040 mx->mx_cursor.mc_top = 0;
8041 mx->mx_cursor.mc_flags = C_SUB;
8042 mx->mx_dbx.md_name.mv_size = 0;
8043 mx->mx_dbx.md_name.mv_data = NULL;
8044 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
8045 mx->mx_dbx.md_dcmp = NULL;
8046 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
8049 /** Final setup of a sorted-dups cursor.
8050 * Sets up the fields that depend on the data from the main cursor.
8051 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8052 * @param[in] node The data containing the #MDB_db record for the
8053 * sorted-dup database.
8056 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
8058 MDB_xcursor *mx = mc->mc_xcursor;
8060 if (node->mn_flags & F_SUBDATA) {
8061 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
8062 mx->mx_cursor.mc_pg[0] = 0;
8063 mx->mx_cursor.mc_snum = 0;
8064 mx->mx_cursor.mc_top = 0;
8065 mx->mx_cursor.mc_flags = C_SUB;
8067 MDB_page *fp = NODEDATA(node);
8068 mx->mx_db.md_pad = 0;
8069 mx->mx_db.md_flags = 0;
8070 mx->mx_db.md_depth = 1;
8071 mx->mx_db.md_branch_pages = 0;
8072 mx->mx_db.md_leaf_pages = 1;
8073 mx->mx_db.md_overflow_pages = 0;
8074 mx->mx_db.md_entries = NUMKEYS(fp);
8075 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
8076 mx->mx_cursor.mc_snum = 1;
8077 mx->mx_cursor.mc_top = 0;
8078 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
8079 mx->mx_cursor.mc_pg[0] = fp;
8080 mx->mx_cursor.mc_ki[0] = 0;
8081 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
8082 mx->mx_db.md_flags = MDB_DUPFIXED;
8083 mx->mx_db.md_pad = fp->mp_pad;
8084 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
8085 mx->mx_db.md_flags |= MDB_INTEGERKEY;
8088 DPRINTF(("Sub-db -%u root page %"Y"u", mx->mx_cursor.mc_dbi,
8089 mx->mx_db.md_root));
8090 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8091 #if UINT_MAX < SIZE_MAX || defined(MDB_VL32)
8092 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(mdb_size_t))
8093 mx->mx_dbx.md_cmp = mdb_cmp_clong;
8098 /** Fixup a sorted-dups cursor due to underlying update.
8099 * Sets up some fields that depend on the data from the main cursor.
8100 * Almost the same as init1, but skips initialization steps if the
8101 * xcursor had already been used.
8102 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
8103 * @param[in] src_mx The xcursor of an up-to-date cursor.
8104 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
8107 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
8109 MDB_xcursor *mx = mc->mc_xcursor;
8112 mx->mx_cursor.mc_snum = 1;
8113 mx->mx_cursor.mc_top = 0;
8114 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8115 mx->mx_cursor.mc_ki[0] = 0;
8116 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8117 #if UINT_MAX < SIZE_MAX
8118 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
8120 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
8123 mx->mx_db = src_mx->mx_db;
8124 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
8125 DPRINTF(("Sub-db -%u root page %"Y"u", mx->mx_cursor.mc_dbi,
8126 mx->mx_db.md_root));
8129 /** Initialize a cursor for a given transaction and database. */
8131 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
8134 mc->mc_backup = NULL;
8137 mc->mc_db = &txn->mt_dbs[dbi];
8138 mc->mc_dbx = &txn->mt_dbxs[dbi];
8139 mc->mc_dbflag = &txn->mt_dbflags[dbi];
8145 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
8146 mdb_tassert(txn, mx != NULL);
8147 mc->mc_xcursor = mx;
8148 mdb_xcursor_init0(mc);
8150 mc->mc_xcursor = NULL;
8152 if (*mc->mc_dbflag & DB_STALE) {
8153 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
8158 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
8161 size_t size = sizeof(MDB_cursor);
8163 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
8166 if (txn->mt_flags & MDB_TXN_BLOCKED)
8169 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8172 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
8173 size += sizeof(MDB_xcursor);
8175 if ((mc = malloc(size)) != NULL) {
8176 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
8177 if (txn->mt_cursors) {
8178 mc->mc_next = txn->mt_cursors[dbi];
8179 txn->mt_cursors[dbi] = mc;
8180 mc->mc_flags |= C_UNTRACK;
8192 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
8194 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
8197 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
8200 if (txn->mt_flags & MDB_TXN_BLOCKED)
8203 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
8207 /* Return the count of duplicate data items for the current key */
8209 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
8213 if (mc == NULL || countp == NULL)
8216 if (mc->mc_xcursor == NULL)
8217 return MDB_INCOMPATIBLE;
8219 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
8222 if (!(mc->mc_flags & C_INITIALIZED))
8225 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
8226 return MDB_NOTFOUND;
8228 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8229 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
8232 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
8235 *countp = mc->mc_xcursor->mx_db.md_entries;
8241 mdb_cursor_close(MDB_cursor *mc)
8243 if (mc && !mc->mc_backup) {
8244 /* remove from txn, if tracked */
8245 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
8246 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
8247 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
8249 *prev = mc->mc_next;
8256 mdb_cursor_txn(MDB_cursor *mc)
8258 if (!mc) return NULL;
8263 mdb_cursor_dbi(MDB_cursor *mc)
8268 /** Replace the key for a branch node with a new key.
8269 * @param[in] mc Cursor pointing to the node to operate on.
8270 * @param[in] key The new key to use.
8271 * @return 0 on success, non-zero on failure.
8274 mdb_update_key(MDB_cursor *mc, MDB_val *key)
8280 int delta, ksize, oksize;
8281 indx_t ptr, i, numkeys, indx;
8284 indx = mc->mc_ki[mc->mc_top];
8285 mp = mc->mc_pg[mc->mc_top];
8286 node = NODEPTR(mp, indx);
8287 ptr = mp->mp_ptrs[indx];
8291 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
8292 k2.mv_data = NODEKEY(node);
8293 k2.mv_size = node->mn_ksize;
8294 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Y"u",
8296 mdb_dkey(&k2, kbuf2),
8302 /* Sizes must be 2-byte aligned. */
8303 ksize = EVEN(key->mv_size);
8304 oksize = EVEN(node->mn_ksize);
8305 delta = ksize - oksize;
8307 /* Shift node contents if EVEN(key length) changed. */
8309 if (delta > 0 && SIZELEFT(mp) < delta) {
8311 /* not enough space left, do a delete and split */
8312 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
8313 pgno = NODEPGNO(node);
8314 mdb_node_del(mc, 0);
8315 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
8318 numkeys = NUMKEYS(mp);
8319 for (i = 0; i < numkeys; i++) {
8320 if (mp->mp_ptrs[i] <= ptr)
8321 mp->mp_ptrs[i] -= delta;
8324 base = (char *)mp + mp->mp_upper + PAGEBASE;
8325 len = ptr - mp->mp_upper + NODESIZE;
8326 memmove(base - delta, base, len);
8327 mp->mp_upper -= delta;
8329 node = NODEPTR(mp, indx);
8332 /* But even if no shift was needed, update ksize */
8333 if (node->mn_ksize != key->mv_size)
8334 node->mn_ksize = key->mv_size;
8337 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8343 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
8345 /** Perform \b act while tracking temporary cursor \b mn */
8346 #define WITH_CURSOR_TRACKING(mn, act) do { \
8347 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
8348 if ((mn).mc_flags & C_SUB) { \
8349 dummy.mc_flags = C_INITIALIZED; \
8350 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
8355 tracked->mc_next = *tp; \
8358 *tp = tracked->mc_next; \
8361 /** Move a node from csrc to cdst.
8364 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
8371 unsigned short flags;
8375 /* Mark src and dst as dirty. */
8376 if ((rc = mdb_page_touch(csrc)) ||
8377 (rc = mdb_page_touch(cdst)))
8380 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8381 key.mv_size = csrc->mc_db->md_pad;
8382 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
8384 data.mv_data = NULL;
8388 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
8389 mdb_cassert(csrc, !((size_t)srcnode & 1));
8390 srcpg = NODEPGNO(srcnode);
8391 flags = srcnode->mn_flags;
8392 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8393 unsigned int snum = csrc->mc_snum;
8395 /* must find the lowest key below src */
8396 rc = mdb_page_search_lowest(csrc);
8399 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8400 key.mv_size = csrc->mc_db->md_pad;
8401 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8403 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8404 key.mv_size = NODEKSZ(s2);
8405 key.mv_data = NODEKEY(s2);
8407 csrc->mc_snum = snum--;
8408 csrc->mc_top = snum;
8410 key.mv_size = NODEKSZ(srcnode);
8411 key.mv_data = NODEKEY(srcnode);
8413 data.mv_size = NODEDSZ(srcnode);
8414 data.mv_data = NODEDATA(srcnode);
8416 mn.mc_xcursor = NULL;
8417 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
8418 unsigned int snum = cdst->mc_snum;
8421 /* must find the lowest key below dst */
8422 mdb_cursor_copy(cdst, &mn);
8423 rc = mdb_page_search_lowest(&mn);
8426 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8427 bkey.mv_size = mn.mc_db->md_pad;
8428 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
8430 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8431 bkey.mv_size = NODEKSZ(s2);
8432 bkey.mv_data = NODEKEY(s2);
8434 mn.mc_snum = snum--;
8437 rc = mdb_update_key(&mn, &bkey);
8442 DPRINTF(("moving %s node %u [%s] on page %"Y"u to node %u on page %"Y"u",
8443 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
8444 csrc->mc_ki[csrc->mc_top],
8446 csrc->mc_pg[csrc->mc_top]->mp_pgno,
8447 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
8449 /* Add the node to the destination page.
8451 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
8452 if (rc != MDB_SUCCESS)
8455 /* Delete the node from the source page.
8457 mdb_node_del(csrc, key.mv_size);
8460 /* Adjust other cursors pointing to mp */
8461 MDB_cursor *m2, *m3;
8462 MDB_dbi dbi = csrc->mc_dbi;
8463 MDB_page *mpd, *mps;
8465 mps = csrc->mc_pg[csrc->mc_top];
8466 /* If we're adding on the left, bump others up */
8468 mpd = cdst->mc_pg[csrc->mc_top];
8469 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8470 if (csrc->mc_flags & C_SUB)
8471 m3 = &m2->mc_xcursor->mx_cursor;
8474 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8477 m3->mc_pg[csrc->mc_top] == mpd &&
8478 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
8479 m3->mc_ki[csrc->mc_top]++;
8482 m3->mc_pg[csrc->mc_top] == mps &&
8483 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
8484 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8485 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8486 m3->mc_ki[csrc->mc_top-1]++;
8488 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8490 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8491 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8492 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8496 /* Adding on the right, bump others down */
8498 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8499 if (csrc->mc_flags & C_SUB)
8500 m3 = &m2->mc_xcursor->mx_cursor;
8503 if (m3 == csrc) continue;
8504 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8506 if (m3->mc_pg[csrc->mc_top] == mps) {
8507 if (!m3->mc_ki[csrc->mc_top]) {
8508 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8509 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8510 m3->mc_ki[csrc->mc_top-1]--;
8512 m3->mc_ki[csrc->mc_top]--;
8514 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8516 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8517 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8518 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8525 /* Update the parent separators.
8527 if (csrc->mc_ki[csrc->mc_top] == 0) {
8528 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
8529 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8530 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8532 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8533 key.mv_size = NODEKSZ(srcnode);
8534 key.mv_data = NODEKEY(srcnode);
8536 DPRINTF(("update separator for source page %"Y"u to [%s]",
8537 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
8538 mdb_cursor_copy(csrc, &mn);
8541 /* We want mdb_rebalance to find mn when doing fixups */
8542 WITH_CURSOR_TRACKING(mn,
8543 rc = mdb_update_key(&mn, &key));
8547 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8549 indx_t ix = csrc->mc_ki[csrc->mc_top];
8550 nullkey.mv_size = 0;
8551 csrc->mc_ki[csrc->mc_top] = 0;
8552 rc = mdb_update_key(csrc, &nullkey);
8553 csrc->mc_ki[csrc->mc_top] = ix;
8554 mdb_cassert(csrc, rc == MDB_SUCCESS);
8558 if (cdst->mc_ki[cdst->mc_top] == 0) {
8559 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
8560 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8561 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
8563 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
8564 key.mv_size = NODEKSZ(srcnode);
8565 key.mv_data = NODEKEY(srcnode);
8567 DPRINTF(("update separator for destination page %"Y"u to [%s]",
8568 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
8569 mdb_cursor_copy(cdst, &mn);
8572 /* We want mdb_rebalance to find mn when doing fixups */
8573 WITH_CURSOR_TRACKING(mn,
8574 rc = mdb_update_key(&mn, &key));
8578 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
8580 indx_t ix = cdst->mc_ki[cdst->mc_top];
8581 nullkey.mv_size = 0;
8582 cdst->mc_ki[cdst->mc_top] = 0;
8583 rc = mdb_update_key(cdst, &nullkey);
8584 cdst->mc_ki[cdst->mc_top] = ix;
8585 mdb_cassert(cdst, rc == MDB_SUCCESS);
8592 /** Merge one page into another.
8593 * The nodes from the page pointed to by \b csrc will
8594 * be copied to the page pointed to by \b cdst and then
8595 * the \b csrc page will be freed.
8596 * @param[in] csrc Cursor pointing to the source page.
8597 * @param[in] cdst Cursor pointing to the destination page.
8598 * @return 0 on success, non-zero on failure.
8601 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8603 MDB_page *psrc, *pdst;
8610 psrc = csrc->mc_pg[csrc->mc_top];
8611 pdst = cdst->mc_pg[cdst->mc_top];
8613 DPRINTF(("merging page %"Y"u into %"Y"u", psrc->mp_pgno, pdst->mp_pgno));
8615 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8616 mdb_cassert(csrc, cdst->mc_snum > 1);
8618 /* Mark dst as dirty. */
8619 if ((rc = mdb_page_touch(cdst)))
8622 /* get dst page again now that we've touched it. */
8623 pdst = cdst->mc_pg[cdst->mc_top];
8625 /* Move all nodes from src to dst.
8627 j = nkeys = NUMKEYS(pdst);
8628 if (IS_LEAF2(psrc)) {
8629 key.mv_size = csrc->mc_db->md_pad;
8630 key.mv_data = METADATA(psrc);
8631 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8632 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8633 if (rc != MDB_SUCCESS)
8635 key.mv_data = (char *)key.mv_data + key.mv_size;
8638 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8639 srcnode = NODEPTR(psrc, i);
8640 if (i == 0 && IS_BRANCH(psrc)) {
8643 mdb_cursor_copy(csrc, &mn);
8644 mn.mc_xcursor = NULL;
8645 /* must find the lowest key below src */
8646 rc = mdb_page_search_lowest(&mn);
8649 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8650 key.mv_size = mn.mc_db->md_pad;
8651 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8653 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8654 key.mv_size = NODEKSZ(s2);
8655 key.mv_data = NODEKEY(s2);
8658 key.mv_size = srcnode->mn_ksize;
8659 key.mv_data = NODEKEY(srcnode);
8662 data.mv_size = NODEDSZ(srcnode);
8663 data.mv_data = NODEDATA(srcnode);
8664 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8665 if (rc != MDB_SUCCESS)
8670 DPRINTF(("dst page %"Y"u now has %u keys (%.1f%% filled)",
8671 pdst->mp_pgno, NUMKEYS(pdst),
8672 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8674 /* Unlink the src page from parent and add to free list.
8677 mdb_node_del(csrc, 0);
8678 if (csrc->mc_ki[csrc->mc_top] == 0) {
8680 rc = mdb_update_key(csrc, &key);
8688 psrc = csrc->mc_pg[csrc->mc_top];
8689 /* If not operating on FreeDB, allow this page to be reused
8690 * in this txn. Otherwise just add to free list.
8692 rc = mdb_page_loose(csrc, psrc);
8696 csrc->mc_db->md_leaf_pages--;
8698 csrc->mc_db->md_branch_pages--;
8700 /* Adjust other cursors pointing to mp */
8701 MDB_cursor *m2, *m3;
8702 MDB_dbi dbi = csrc->mc_dbi;
8703 unsigned int top = csrc->mc_top;
8705 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8706 if (csrc->mc_flags & C_SUB)
8707 m3 = &m2->mc_xcursor->mx_cursor;
8710 if (m3 == csrc) continue;
8711 if (m3->mc_snum < csrc->mc_snum) continue;
8712 if (m3->mc_pg[top] == psrc) {
8713 m3->mc_pg[top] = pdst;
8714 m3->mc_ki[top] += nkeys;
8715 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8716 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8717 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8720 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8722 MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
8723 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8724 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8729 unsigned int snum = cdst->mc_snum;
8730 uint16_t depth = cdst->mc_db->md_depth;
8731 mdb_cursor_pop(cdst);
8732 rc = mdb_rebalance(cdst);
8733 /* Did the tree height change? */
8734 if (depth != cdst->mc_db->md_depth)
8735 snum += cdst->mc_db->md_depth - depth;
8736 cdst->mc_snum = snum;
8737 cdst->mc_top = snum-1;
8742 /** Copy the contents of a cursor.
8743 * @param[in] csrc The cursor to copy from.
8744 * @param[out] cdst The cursor to copy to.
8747 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8751 cdst->mc_txn = csrc->mc_txn;
8752 cdst->mc_dbi = csrc->mc_dbi;
8753 cdst->mc_db = csrc->mc_db;
8754 cdst->mc_dbx = csrc->mc_dbx;
8755 cdst->mc_snum = csrc->mc_snum;
8756 cdst->mc_top = csrc->mc_top;
8757 cdst->mc_flags = csrc->mc_flags;
8759 for (i=0; i<csrc->mc_snum; i++) {
8760 cdst->mc_pg[i] = csrc->mc_pg[i];
8761 cdst->mc_ki[i] = csrc->mc_ki[i];
8765 /** Rebalance the tree after a delete operation.
8766 * @param[in] mc Cursor pointing to the page where rebalancing
8768 * @return 0 on success, non-zero on failure.
8771 mdb_rebalance(MDB_cursor *mc)
8775 unsigned int ptop, minkeys, thresh;
8779 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8784 thresh = FILL_THRESHOLD;
8786 DPRINTF(("rebalancing %s page %"Y"u (has %u keys, %.1f%% full)",
8787 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8788 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8789 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8791 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8792 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8793 DPRINTF(("no need to rebalance page %"Y"u, above fill threshold",
8794 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8798 if (mc->mc_snum < 2) {
8799 MDB_page *mp = mc->mc_pg[0];
8801 DPUTS("Can't rebalance a subpage, ignoring");
8804 if (NUMKEYS(mp) == 0) {
8805 DPUTS("tree is completely empty");
8806 mc->mc_db->md_root = P_INVALID;
8807 mc->mc_db->md_depth = 0;
8808 mc->mc_db->md_leaf_pages = 0;
8809 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8812 /* Adjust cursors pointing to mp */
8815 mc->mc_flags &= ~C_INITIALIZED;
8817 MDB_cursor *m2, *m3;
8818 MDB_dbi dbi = mc->mc_dbi;
8820 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8821 if (mc->mc_flags & C_SUB)
8822 m3 = &m2->mc_xcursor->mx_cursor;
8825 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8827 if (m3->mc_pg[0] == mp) {
8830 m3->mc_flags &= ~C_INITIALIZED;
8834 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8836 DPUTS("collapsing root page!");
8837 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8840 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8841 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
8844 mc->mc_db->md_depth--;
8845 mc->mc_db->md_branch_pages--;
8846 mc->mc_ki[0] = mc->mc_ki[1];
8847 for (i = 1; i<mc->mc_db->md_depth; i++) {
8848 mc->mc_pg[i] = mc->mc_pg[i+1];
8849 mc->mc_ki[i] = mc->mc_ki[i+1];
8852 /* Adjust other cursors pointing to mp */
8853 MDB_cursor *m2, *m3;
8854 MDB_dbi dbi = mc->mc_dbi;
8856 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8857 if (mc->mc_flags & C_SUB)
8858 m3 = &m2->mc_xcursor->mx_cursor;
8861 if (m3 == mc) continue;
8862 if (!(m3->mc_flags & C_INITIALIZED))
8864 if (m3->mc_pg[0] == mp) {
8865 for (i=0; i<mc->mc_db->md_depth; i++) {
8866 m3->mc_pg[i] = m3->mc_pg[i+1];
8867 m3->mc_ki[i] = m3->mc_ki[i+1];
8875 DPUTS("root page doesn't need rebalancing");
8879 /* The parent (branch page) must have at least 2 pointers,
8880 * otherwise the tree is invalid.
8882 ptop = mc->mc_top-1;
8883 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8885 /* Leaf page fill factor is below the threshold.
8886 * Try to move keys from left or right neighbor, or
8887 * merge with a neighbor page.
8892 mdb_cursor_copy(mc, &mn);
8893 mn.mc_xcursor = NULL;
8895 oldki = mc->mc_ki[mc->mc_top];
8896 if (mc->mc_ki[ptop] == 0) {
8897 /* We're the leftmost leaf in our parent.
8899 DPUTS("reading right neighbor");
8901 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8902 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8905 mn.mc_ki[mn.mc_top] = 0;
8906 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8909 /* There is at least one neighbor to the left.
8911 DPUTS("reading left neighbor");
8913 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8914 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8917 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8918 mc->mc_ki[mc->mc_top] = 0;
8922 DPRINTF(("found neighbor page %"Y"u (%u keys, %.1f%% full)",
8923 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8924 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8926 /* If the neighbor page is above threshold and has enough keys,
8927 * move one key from it. Otherwise we should try to merge them.
8928 * (A branch page must never have less than 2 keys.)
8930 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8931 rc = mdb_node_move(&mn, mc, fromleft);
8933 /* if we inserted on left, bump position up */
8938 rc = mdb_page_merge(&mn, mc);
8940 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8941 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8942 /* We want mdb_rebalance to find mn when doing fixups */
8943 WITH_CURSOR_TRACKING(mn,
8944 rc = mdb_page_merge(mc, &mn));
8945 mdb_cursor_copy(&mn, mc);
8947 mc->mc_flags &= ~C_EOF;
8949 mc->mc_ki[mc->mc_top] = oldki;
8953 /** Complete a delete operation started by #mdb_cursor_del(). */
8955 mdb_cursor_del0(MDB_cursor *mc)
8961 MDB_cursor *m2, *m3;
8962 MDB_dbi dbi = mc->mc_dbi;
8964 ki = mc->mc_ki[mc->mc_top];
8965 mp = mc->mc_pg[mc->mc_top];
8966 mdb_node_del(mc, mc->mc_db->md_pad);
8967 mc->mc_db->md_entries--;
8969 /* Adjust other cursors pointing to mp */
8970 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8971 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8972 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8974 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8976 if (m3->mc_pg[mc->mc_top] == mp) {
8977 if (m3->mc_ki[mc->mc_top] == ki) {
8978 m3->mc_flags |= C_DEL;
8979 if (mc->mc_db->md_flags & MDB_DUPSORT)
8980 m3->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
8981 } else if (m3->mc_ki[mc->mc_top] > ki) {
8982 m3->mc_ki[mc->mc_top]--;
8984 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
8985 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8986 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8987 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8992 rc = mdb_rebalance(mc);
8994 if (rc == MDB_SUCCESS) {
8995 /* DB is totally empty now, just bail out.
8996 * Other cursors adjustments were already done
8997 * by mdb_rebalance and aren't needed here.
9002 mp = mc->mc_pg[mc->mc_top];
9003 nkeys = NUMKEYS(mp);
9005 /* Adjust other cursors pointing to mp */
9006 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
9007 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9008 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9010 if (m3->mc_snum < mc->mc_snum)
9012 if (m3->mc_pg[mc->mc_top] == mp) {
9013 /* if m3 points past last node in page, find next sibling */
9014 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9015 rc = mdb_cursor_sibling(m3, 1);
9016 if (rc == MDB_NOTFOUND) {
9017 m3->mc_flags |= C_EOF;
9023 mc->mc_flags |= C_DEL;
9027 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9032 mdb_del(MDB_txn *txn, MDB_dbi dbi,
9033 MDB_val *key, MDB_val *data)
9035 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9038 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9039 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9041 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
9042 /* must ignore any data */
9046 return mdb_del0(txn, dbi, key, data, 0);
9050 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
9051 MDB_val *key, MDB_val *data, unsigned flags)
9056 MDB_val rdata, *xdata;
9060 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
9062 mdb_cursor_init(&mc, txn, dbi, &mx);
9071 flags |= MDB_NODUPDATA;
9073 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
9075 /* let mdb_page_split know about this cursor if needed:
9076 * delete will trigger a rebalance; if it needs to move
9077 * a node from one page to another, it will have to
9078 * update the parent's separator key(s). If the new sepkey
9079 * is larger than the current one, the parent page may
9080 * run out of space, triggering a split. We need this
9081 * cursor to be consistent until the end of the rebalance.
9083 mc.mc_flags |= C_UNTRACK;
9084 mc.mc_next = txn->mt_cursors[dbi];
9085 txn->mt_cursors[dbi] = &mc;
9086 rc = mdb_cursor_del(&mc, flags);
9087 txn->mt_cursors[dbi] = mc.mc_next;
9092 /** Split a page and insert a new node.
9093 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
9094 * The cursor will be updated to point to the actual page and index where
9095 * the node got inserted after the split.
9096 * @param[in] newkey The key for the newly inserted node.
9097 * @param[in] newdata The data for the newly inserted node.
9098 * @param[in] newpgno The page number, if the new node is a branch node.
9099 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
9100 * @return 0 on success, non-zero on failure.
9103 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
9104 unsigned int nflags)
9107 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
9110 int i, j, split_indx, nkeys, pmax;
9111 MDB_env *env = mc->mc_txn->mt_env;
9113 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
9114 MDB_page *copy = NULL;
9115 MDB_page *mp, *rp, *pp;
9120 mp = mc->mc_pg[mc->mc_top];
9121 newindx = mc->mc_ki[mc->mc_top];
9122 nkeys = NUMKEYS(mp);
9124 DPRINTF(("-----> splitting %s page %"Y"u and adding [%s] at index %i/%i",
9125 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
9126 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
9128 /* Create a right sibling. */
9129 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
9131 rp->mp_pad = mp->mp_pad;
9132 DPRINTF(("new right sibling: page %"Y"u", rp->mp_pgno));
9134 /* Usually when splitting the root page, the cursor
9135 * height is 1. But when called from mdb_update_key,
9136 * the cursor height may be greater because it walks
9137 * up the stack while finding the branch slot to update.
9139 if (mc->mc_top < 1) {
9140 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
9142 /* shift current top to make room for new parent */
9143 for (i=mc->mc_snum; i>0; i--) {
9144 mc->mc_pg[i] = mc->mc_pg[i-1];
9145 mc->mc_ki[i] = mc->mc_ki[i-1];
9149 mc->mc_db->md_root = pp->mp_pgno;
9150 DPRINTF(("root split! new root = %"Y"u", pp->mp_pgno));
9151 new_root = mc->mc_db->md_depth++;
9153 /* Add left (implicit) pointer. */
9154 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
9155 /* undo the pre-push */
9156 mc->mc_pg[0] = mc->mc_pg[1];
9157 mc->mc_ki[0] = mc->mc_ki[1];
9158 mc->mc_db->md_root = mp->mp_pgno;
9159 mc->mc_db->md_depth--;
9166 ptop = mc->mc_top-1;
9167 DPRINTF(("parent branch page is %"Y"u", mc->mc_pg[ptop]->mp_pgno));
9170 mdb_cursor_copy(mc, &mn);
9171 mn.mc_xcursor = NULL;
9172 mn.mc_pg[mn.mc_top] = rp;
9173 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
9175 if (nflags & MDB_APPEND) {
9176 mn.mc_ki[mn.mc_top] = 0;
9178 split_indx = newindx;
9182 split_indx = (nkeys+1) / 2;
9187 unsigned int lsize, rsize, ksize;
9188 /* Move half of the keys to the right sibling */
9189 x = mc->mc_ki[mc->mc_top] - split_indx;
9190 ksize = mc->mc_db->md_pad;
9191 split = LEAF2KEY(mp, split_indx, ksize);
9192 rsize = (nkeys - split_indx) * ksize;
9193 lsize = (nkeys - split_indx) * sizeof(indx_t);
9194 mp->mp_lower -= lsize;
9195 rp->mp_lower += lsize;
9196 mp->mp_upper += rsize - lsize;
9197 rp->mp_upper -= rsize - lsize;
9198 sepkey.mv_size = ksize;
9199 if (newindx == split_indx) {
9200 sepkey.mv_data = newkey->mv_data;
9202 sepkey.mv_data = split;
9205 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
9206 memcpy(rp->mp_ptrs, split, rsize);
9207 sepkey.mv_data = rp->mp_ptrs;
9208 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
9209 memcpy(ins, newkey->mv_data, ksize);
9210 mp->mp_lower += sizeof(indx_t);
9211 mp->mp_upper -= ksize - sizeof(indx_t);
9214 memcpy(rp->mp_ptrs, split, x * ksize);
9215 ins = LEAF2KEY(rp, x, ksize);
9216 memcpy(ins, newkey->mv_data, ksize);
9217 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
9218 rp->mp_lower += sizeof(indx_t);
9219 rp->mp_upper -= ksize - sizeof(indx_t);
9220 mc->mc_ki[mc->mc_top] = x;
9223 int psize, nsize, k;
9224 /* Maximum free space in an empty page */
9225 pmax = env->me_psize - PAGEHDRSZ;
9227 nsize = mdb_leaf_size(env, newkey, newdata);
9229 nsize = mdb_branch_size(env, newkey);
9230 nsize = EVEN(nsize);
9232 /* grab a page to hold a temporary copy */
9233 copy = mdb_page_malloc(mc->mc_txn, 1);
9238 copy->mp_pgno = mp->mp_pgno;
9239 copy->mp_flags = mp->mp_flags;
9240 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
9241 copy->mp_upper = env->me_psize - PAGEBASE;
9243 /* prepare to insert */
9244 for (i=0, j=0; i<nkeys; i++) {
9246 copy->mp_ptrs[j++] = 0;
9248 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
9251 /* When items are relatively large the split point needs
9252 * to be checked, because being off-by-one will make the
9253 * difference between success or failure in mdb_node_add.
9255 * It's also relevant if a page happens to be laid out
9256 * such that one half of its nodes are all "small" and
9257 * the other half of its nodes are "large." If the new
9258 * item is also "large" and falls on the half with
9259 * "large" nodes, it also may not fit.
9261 * As a final tweak, if the new item goes on the last
9262 * spot on the page (and thus, onto the new page), bias
9263 * the split so the new page is emptier than the old page.
9264 * This yields better packing during sequential inserts.
9266 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
9267 /* Find split point */
9269 if (newindx <= split_indx || newindx >= nkeys) {
9271 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
9276 for (; i!=k; i+=j) {
9281 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9282 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
9284 if (F_ISSET(node->mn_flags, F_BIGDATA))
9285 psize += sizeof(pgno_t);
9287 psize += NODEDSZ(node);
9289 psize = EVEN(psize);
9291 if (psize > pmax || i == k-j) {
9292 split_indx = i + (j<0);
9297 if (split_indx == newindx) {
9298 sepkey.mv_size = newkey->mv_size;
9299 sepkey.mv_data = newkey->mv_data;
9301 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
9302 sepkey.mv_size = node->mn_ksize;
9303 sepkey.mv_data = NODEKEY(node);
9308 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
9310 /* Copy separator key to the parent.
9312 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
9313 int snum = mc->mc_snum;
9317 /* We want other splits to find mn when doing fixups */
9318 WITH_CURSOR_TRACKING(mn,
9319 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
9324 if (mc->mc_snum > snum) {
9327 /* Right page might now have changed parent.
9328 * Check if left page also changed parent.
9330 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9331 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9332 for (i=0; i<ptop; i++) {
9333 mc->mc_pg[i] = mn.mc_pg[i];
9334 mc->mc_ki[i] = mn.mc_ki[i];
9336 mc->mc_pg[ptop] = mn.mc_pg[ptop];
9337 if (mn.mc_ki[ptop]) {
9338 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
9340 /* find right page's left sibling */
9341 mc->mc_ki[ptop] = mn.mc_ki[ptop];
9342 mdb_cursor_sibling(mc, 0);
9347 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
9350 if (rc != MDB_SUCCESS) {
9353 if (nflags & MDB_APPEND) {
9354 mc->mc_pg[mc->mc_top] = rp;
9355 mc->mc_ki[mc->mc_top] = 0;
9356 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
9359 for (i=0; i<mc->mc_top; i++)
9360 mc->mc_ki[i] = mn.mc_ki[i];
9361 } else if (!IS_LEAF2(mp)) {
9363 mc->mc_pg[mc->mc_top] = rp;
9368 rkey.mv_data = newkey->mv_data;
9369 rkey.mv_size = newkey->mv_size;
9375 /* Update index for the new key. */
9376 mc->mc_ki[mc->mc_top] = j;
9378 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9379 rkey.mv_data = NODEKEY(node);
9380 rkey.mv_size = node->mn_ksize;
9382 xdata.mv_data = NODEDATA(node);
9383 xdata.mv_size = NODEDSZ(node);
9386 pgno = NODEPGNO(node);
9387 flags = node->mn_flags;
9390 if (!IS_LEAF(mp) && j == 0) {
9391 /* First branch index doesn't need key data. */
9395 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
9401 mc->mc_pg[mc->mc_top] = copy;
9406 } while (i != split_indx);
9408 nkeys = NUMKEYS(copy);
9409 for (i=0; i<nkeys; i++)
9410 mp->mp_ptrs[i] = copy->mp_ptrs[i];
9411 mp->mp_lower = copy->mp_lower;
9412 mp->mp_upper = copy->mp_upper;
9413 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
9414 env->me_psize - copy->mp_upper - PAGEBASE);
9416 /* reset back to original page */
9417 if (newindx < split_indx) {
9418 mc->mc_pg[mc->mc_top] = mp;
9420 mc->mc_pg[mc->mc_top] = rp;
9422 /* Make sure mc_ki is still valid.
9424 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9425 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9426 for (i=0; i<=ptop; i++) {
9427 mc->mc_pg[i] = mn.mc_pg[i];
9428 mc->mc_ki[i] = mn.mc_ki[i];
9432 if (nflags & MDB_RESERVE) {
9433 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
9434 if (!(node->mn_flags & F_BIGDATA))
9435 newdata->mv_data = NODEDATA(node);
9438 if (newindx >= split_indx) {
9439 mc->mc_pg[mc->mc_top] = rp;
9441 /* Make sure mc_ki is still valid.
9443 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9444 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9445 for (i=0; i<=ptop; i++) {
9446 mc->mc_pg[i] = mn.mc_pg[i];
9447 mc->mc_ki[i] = mn.mc_ki[i];
9454 /* Adjust other cursors pointing to mp */
9455 MDB_cursor *m2, *m3;
9456 MDB_dbi dbi = mc->mc_dbi;
9457 nkeys = NUMKEYS(mp);
9459 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9460 if (mc->mc_flags & C_SUB)
9461 m3 = &m2->mc_xcursor->mx_cursor;
9466 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9470 /* sub cursors may be on different DB */
9471 if (m3->mc_pg[0] != mp)
9474 for (k=new_root; k>=0; k--) {
9475 m3->mc_ki[k+1] = m3->mc_ki[k];
9476 m3->mc_pg[k+1] = m3->mc_pg[k];
9478 if (m3->mc_ki[0] >= nkeys) {
9483 m3->mc_pg[0] = mc->mc_pg[0];
9487 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
9488 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
9489 m3->mc_ki[mc->mc_top]++;
9490 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9491 m3->mc_pg[mc->mc_top] = rp;
9492 m3->mc_ki[mc->mc_top] -= nkeys;
9493 for (i=0; i<mc->mc_top; i++) {
9494 m3->mc_ki[i] = mn.mc_ki[i];
9495 m3->mc_pg[i] = mn.mc_pg[i];
9498 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
9499 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
9502 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
9504 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9505 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
9506 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9510 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
9513 if (copy) /* tmp page */
9514 mdb_page_free(env, copy);
9516 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9521 mdb_put(MDB_txn *txn, MDB_dbi dbi,
9522 MDB_val *key, MDB_val *data, unsigned int flags)
9528 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9531 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
9534 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9535 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9537 mdb_cursor_init(&mc, txn, dbi, &mx);
9538 mc.mc_next = txn->mt_cursors[dbi];
9539 txn->mt_cursors[dbi] = &mc;
9540 rc = mdb_cursor_put(&mc, key, data, flags);
9541 txn->mt_cursors[dbi] = mc.mc_next;
9546 #define MDB_WBUF (1024*1024)
9549 /** State needed for a compacting copy. */
9550 typedef struct mdb_copy {
9551 pthread_mutex_t mc_mutex;
9552 pthread_cond_t mc_cond;
9559 pgno_t mc_next_pgno;
9562 volatile int mc_new;
9567 /** Dedicated writer thread for compacting copy. */
9568 static THREAD_RET ESECT CALL_CONV
9569 mdb_env_copythr(void *arg)
9573 int toggle = 0, wsize, rc;
9576 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9579 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9582 pthread_mutex_lock(&my->mc_mutex);
9584 pthread_cond_signal(&my->mc_cond);
9587 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9588 if (my->mc_new < 0) {
9593 wsize = my->mc_wlen[toggle];
9594 ptr = my->mc_wbuf[toggle];
9597 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9601 } else if (len > 0) {
9615 /* If there's an overflow page tail, write it too */
9616 if (my->mc_olen[toggle]) {
9617 wsize = my->mc_olen[toggle];
9618 ptr = my->mc_over[toggle];
9619 my->mc_olen[toggle] = 0;
9622 my->mc_wlen[toggle] = 0;
9624 pthread_cond_signal(&my->mc_cond);
9626 pthread_cond_signal(&my->mc_cond);
9627 pthread_mutex_unlock(&my->mc_mutex);
9628 return (THREAD_RET)0;
9632 /** Tell the writer thread there's a buffer ready to write */
9634 mdb_env_cthr_toggle(mdb_copy *my, int st)
9636 int toggle = my->mc_toggle ^ 1;
9637 pthread_mutex_lock(&my->mc_mutex);
9638 if (my->mc_status) {
9639 pthread_mutex_unlock(&my->mc_mutex);
9640 return my->mc_status;
9642 while (my->mc_new == 1)
9643 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9645 my->mc_toggle = toggle;
9646 pthread_cond_signal(&my->mc_cond);
9647 pthread_mutex_unlock(&my->mc_mutex);
9651 /** Depth-first tree traversal for compacting copy. */
9653 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9657 MDB_page *mo, *mp, *leaf;
9662 /* Empty DB, nothing to do */
9663 if (*pg == P_INVALID)
9668 mc.mc_txn = my->mc_txn;
9670 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9673 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9677 /* Make cursor pages writable */
9678 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9682 for (i=0; i<mc.mc_top; i++) {
9683 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9684 mc.mc_pg[i] = (MDB_page *)ptr;
9685 ptr += my->mc_env->me_psize;
9688 /* This is writable space for a leaf page. Usually not needed. */
9689 leaf = (MDB_page *)ptr;
9691 toggle = my->mc_toggle;
9692 while (mc.mc_snum > 0) {
9694 mp = mc.mc_pg[mc.mc_top];
9698 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9699 for (i=0; i<n; i++) {
9700 ni = NODEPTR(mp, i);
9701 if (ni->mn_flags & F_BIGDATA) {
9705 /* Need writable leaf */
9707 mc.mc_pg[mc.mc_top] = leaf;
9708 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9710 ni = NODEPTR(mp, i);
9713 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9714 rc = mdb_page_get(&mc, pg, &omp, NULL);
9717 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9718 rc = mdb_env_cthr_toggle(my, 1);
9721 toggle = my->mc_toggle;
9723 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9724 memcpy(mo, omp, my->mc_env->me_psize);
9725 mo->mp_pgno = my->mc_next_pgno;
9726 my->mc_next_pgno += omp->mp_pages;
9727 my->mc_wlen[toggle] += my->mc_env->me_psize;
9728 if (omp->mp_pages > 1) {
9729 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9730 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9731 rc = mdb_env_cthr_toggle(my, 1);
9734 toggle = my->mc_toggle;
9736 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
9737 } else if (ni->mn_flags & F_SUBDATA) {
9740 /* Need writable leaf */
9742 mc.mc_pg[mc.mc_top] = leaf;
9743 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9745 ni = NODEPTR(mp, i);
9748 memcpy(&db, NODEDATA(ni), sizeof(db));
9749 my->mc_toggle = toggle;
9750 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9753 toggle = my->mc_toggle;
9754 memcpy(NODEDATA(ni), &db, sizeof(db));
9759 mc.mc_ki[mc.mc_top]++;
9760 if (mc.mc_ki[mc.mc_top] < n) {
9763 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9765 rc = mdb_page_get(&mc, pg, &mp, NULL);
9770 mc.mc_ki[mc.mc_top] = 0;
9771 if (IS_BRANCH(mp)) {
9772 /* Whenever we advance to a sibling branch page,
9773 * we must proceed all the way down to its first leaf.
9775 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9778 mc.mc_pg[mc.mc_top] = mp;
9782 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9783 rc = mdb_env_cthr_toggle(my, 1);
9786 toggle = my->mc_toggle;
9788 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9789 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9790 mo->mp_pgno = my->mc_next_pgno++;
9791 my->mc_wlen[toggle] += my->mc_env->me_psize;
9793 /* Update parent if there is one */
9794 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9795 SETPGNO(ni, mo->mp_pgno);
9796 mdb_cursor_pop(&mc);
9798 /* Otherwise we're done */
9808 /** Copy environment with compaction. */
9810 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9815 MDB_txn *txn = NULL;
9820 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
9821 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
9822 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9823 if (my.mc_wbuf[0] == NULL)
9826 pthread_mutex_init(&my.mc_mutex, NULL);
9827 pthread_cond_init(&my.mc_cond, NULL);
9828 #ifdef HAVE_MEMALIGN
9829 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9830 if (my.mc_wbuf[0] == NULL)
9833 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
9838 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9839 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9844 my.mc_next_pgno = NUM_METAS;
9850 THREAD_CREATE(thr, mdb_env_copythr, &my);
9852 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9856 mp = (MDB_page *)my.mc_wbuf[0];
9857 memset(mp, 0, NUM_METAS * env->me_psize);
9859 mp->mp_flags = P_META;
9860 mm = (MDB_meta *)METADATA(mp);
9861 mdb_env_init_meta0(env, mm);
9862 mm->mm_address = env->me_metas[0]->mm_address;
9864 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9866 mp->mp_flags = P_META;
9867 *(MDB_meta *)METADATA(mp) = *mm;
9868 mm = (MDB_meta *)METADATA(mp);
9870 /* Count the number of free pages, subtract from lastpg to find
9871 * number of active pages
9874 MDB_ID freecount = 0;
9877 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9878 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9879 freecount += *(MDB_ID *)data.mv_data;
9880 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9881 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9882 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9884 /* Set metapage 1 */
9885 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
9886 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9887 if (mm->mm_last_pg > NUM_METAS-1) {
9888 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
9891 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
9894 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9896 pthread_mutex_lock(&my.mc_mutex);
9898 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9899 pthread_mutex_unlock(&my.mc_mutex);
9900 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
9901 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
9902 rc = mdb_env_cthr_toggle(&my, 1);
9903 mdb_env_cthr_toggle(&my, -1);
9904 pthread_mutex_lock(&my.mc_mutex);
9906 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9907 pthread_mutex_unlock(&my.mc_mutex);
9912 CloseHandle(my.mc_cond);
9913 CloseHandle(my.mc_mutex);
9914 _aligned_free(my.mc_wbuf[0]);
9916 pthread_cond_destroy(&my.mc_cond);
9917 pthread_mutex_destroy(&my.mc_mutex);
9918 free(my.mc_wbuf[0]);
9923 /** Copy environment as-is. */
9925 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9927 MDB_txn *txn = NULL;
9928 mdb_mutexref_t wmutex = NULL;
9934 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9938 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9941 /* Do the lock/unlock of the reader mutex before starting the
9942 * write txn. Otherwise other read txns could block writers.
9944 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9949 /* We must start the actual read txn after blocking writers */
9950 mdb_txn_end(txn, MDB_END_RESET_TMP);
9952 /* Temporarily block writers until we snapshot the meta pages */
9953 wmutex = env->me_wmutex;
9954 if (LOCK_MUTEX(rc, env, wmutex))
9957 rc = mdb_txn_renew0(txn);
9959 UNLOCK_MUTEX(wmutex);
9964 wsize = env->me_psize * NUM_METAS;
9968 DO_WRITE(rc, fd, ptr, w2, len);
9972 } else if (len > 0) {
9978 /* Non-blocking or async handles are not supported */
9984 UNLOCK_MUTEX(wmutex);
9989 w2 = txn->mt_next_pgno * env->me_psize;
9991 mdb_size_t fsize = 0;
9992 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9999 if (wsize > MAX_WRITE)
10003 DO_WRITE(rc, fd, ptr, w2, len);
10007 } else if (len > 0) {
10019 mdb_txn_abort(txn);
10024 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
10026 if (flags & MDB_CP_COMPACT)
10027 return mdb_env_copyfd1(env, fd);
10029 return mdb_env_copyfd0(env, fd);
10033 mdb_env_copyfd(MDB_env *env, HANDLE fd)
10035 return mdb_env_copyfd2(env, fd, 0);
10039 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
10043 HANDLE newfd = INVALID_HANDLE_VALUE;
10048 if (env->me_flags & MDB_NOSUBDIR) {
10049 lpath = (char *)path;
10051 len = strlen(path);
10052 len += sizeof(DATANAME);
10053 lpath = malloc(len);
10056 sprintf(lpath, "%s" DATANAME, path);
10059 /* The destination path must exist, but the destination file must not.
10060 * We don't want the OS to cache the writes, since the source data is
10061 * already in the OS cache.
10064 rc = utf8_to_utf16(lpath, -1, &wpath, NULL);
10067 newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
10068 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
10071 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
10073 if (newfd == INVALID_HANDLE_VALUE) {
10078 if (env->me_psize >= env->me_os_psize) {
10080 /* Set O_DIRECT if the file system supports it */
10081 if ((rc = fcntl(newfd, F_GETFL)) != -1)
10082 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
10084 #ifdef F_NOCACHE /* __APPLE__ */
10085 rc = fcntl(newfd, F_NOCACHE, 1);
10093 rc = mdb_env_copyfd2(env, newfd, flags);
10096 if (!(env->me_flags & MDB_NOSUBDIR))
10098 if (newfd != INVALID_HANDLE_VALUE)
10099 if (close(newfd) < 0 && rc == MDB_SUCCESS)
10106 mdb_env_copy(MDB_env *env, const char *path)
10108 return mdb_env_copy2(env, path, 0);
10112 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
10114 if (flag & ~CHANGEABLE)
10117 env->me_flags |= flag;
10119 env->me_flags &= ~flag;
10120 return MDB_SUCCESS;
10124 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
10129 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
10130 return MDB_SUCCESS;
10134 mdb_env_set_userctx(MDB_env *env, void *ctx)
10138 env->me_userctx = ctx;
10139 return MDB_SUCCESS;
10143 mdb_env_get_userctx(MDB_env *env)
10145 return env ? env->me_userctx : NULL;
10149 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
10154 env->me_assert_func = func;
10156 return MDB_SUCCESS;
10160 mdb_env_get_path(MDB_env *env, const char **arg)
10165 *arg = env->me_path;
10166 return MDB_SUCCESS;
10170 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
10176 return MDB_SUCCESS;
10179 /** Common code for #mdb_stat() and #mdb_env_stat().
10180 * @param[in] env the environment to operate in.
10181 * @param[in] db the #MDB_db record containing the stats to return.
10182 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
10183 * @return 0, this function always succeeds.
10186 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
10188 arg->ms_psize = env->me_psize;
10189 arg->ms_depth = db->md_depth;
10190 arg->ms_branch_pages = db->md_branch_pages;
10191 arg->ms_leaf_pages = db->md_leaf_pages;
10192 arg->ms_overflow_pages = db->md_overflow_pages;
10193 arg->ms_entries = db->md_entries;
10195 return MDB_SUCCESS;
10199 mdb_env_stat(MDB_env *env, MDB_stat *arg)
10203 if (env == NULL || arg == NULL)
10206 meta = mdb_env_pick_meta(env);
10208 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
10212 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
10216 if (env == NULL || arg == NULL)
10219 meta = mdb_env_pick_meta(env);
10220 arg->me_mapaddr = meta->mm_address;
10221 arg->me_last_pgno = meta->mm_last_pg;
10222 arg->me_last_txnid = meta->mm_txnid;
10224 arg->me_mapsize = env->me_mapsize;
10225 arg->me_maxreaders = env->me_maxreaders;
10226 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
10227 return MDB_SUCCESS;
10230 /** Set the default comparison functions for a database.
10231 * Called immediately after a database is opened to set the defaults.
10232 * The user can then override them with #mdb_set_compare() or
10233 * #mdb_set_dupsort().
10234 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
10235 * @param[in] dbi A database handle returned by #mdb_dbi_open()
10238 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
10240 uint16_t f = txn->mt_dbs[dbi].md_flags;
10242 txn->mt_dbxs[dbi].md_cmp =
10243 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
10244 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
10246 txn->mt_dbxs[dbi].md_dcmp =
10247 !(f & MDB_DUPSORT) ? 0 :
10248 ((f & MDB_INTEGERDUP)
10249 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
10250 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
10253 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
10259 int rc, dbflag, exact;
10260 unsigned int unused = 0, seq;
10264 if (flags & ~VALID_FLAGS)
10266 if (txn->mt_flags & MDB_TXN_BLOCKED)
10267 return MDB_BAD_TXN;
10272 if (flags & PERSISTENT_FLAGS) {
10273 uint16_t f2 = flags & PERSISTENT_FLAGS;
10274 /* make sure flag changes get committed */
10275 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
10276 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
10277 txn->mt_flags |= MDB_TXN_DIRTY;
10280 mdb_default_cmp(txn, MAIN_DBI);
10281 return MDB_SUCCESS;
10284 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
10285 mdb_default_cmp(txn, MAIN_DBI);
10288 /* Is the DB already open? */
10289 len = strlen(name);
10290 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
10291 if (!txn->mt_dbxs[i].md_name.mv_size) {
10292 /* Remember this free slot */
10293 if (!unused) unused = i;
10296 if (len == txn->mt_dbxs[i].md_name.mv_size &&
10297 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
10299 return MDB_SUCCESS;
10303 /* If no free slot and max hit, fail */
10304 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
10305 return MDB_DBS_FULL;
10307 /* Cannot mix named databases with some mainDB flags */
10308 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
10309 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
10311 /* Find the DB info */
10312 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
10315 key.mv_data = (void *)name;
10316 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
10317 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
10318 if (rc == MDB_SUCCESS) {
10319 /* make sure this is actually a DB */
10320 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
10321 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
10322 return MDB_INCOMPATIBLE;
10323 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
10327 /* Done here so we cannot fail after creating a new DB */
10328 if ((namedup = strdup(name)) == NULL)
10332 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
10333 data.mv_size = sizeof(MDB_db);
10334 data.mv_data = &dummy;
10335 memset(&dummy, 0, sizeof(dummy));
10336 dummy.md_root = P_INVALID;
10337 dummy.md_flags = flags & PERSISTENT_FLAGS;
10338 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
10339 dbflag |= DB_DIRTY;
10345 /* Got info, register DBI in this txn */
10346 unsigned int slot = unused ? unused : txn->mt_numdbs;
10347 txn->mt_dbxs[slot].md_name.mv_data = namedup;
10348 txn->mt_dbxs[slot].md_name.mv_size = len;
10349 txn->mt_dbxs[slot].md_rel = NULL;
10350 txn->mt_dbflags[slot] = dbflag;
10351 /* txn-> and env-> are the same in read txns, use
10352 * tmp variable to avoid undefined assignment
10354 seq = ++txn->mt_env->me_dbiseqs[slot];
10355 txn->mt_dbiseqs[slot] = seq;
10357 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
10359 mdb_default_cmp(txn, slot);
10369 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
10371 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
10374 if (txn->mt_flags & MDB_TXN_BLOCKED)
10375 return MDB_BAD_TXN;
10377 if (txn->mt_dbflags[dbi] & DB_STALE) {
10380 /* Stale, must read the DB's root. cursor_init does it for us. */
10381 mdb_cursor_init(&mc, txn, dbi, &mx);
10383 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
10386 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
10389 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
10391 ptr = env->me_dbxs[dbi].md_name.mv_data;
10392 /* If there was no name, this was already closed */
10394 env->me_dbxs[dbi].md_name.mv_data = NULL;
10395 env->me_dbxs[dbi].md_name.mv_size = 0;
10396 env->me_dbflags[dbi] = 0;
10397 env->me_dbiseqs[dbi]++;
10402 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
10404 /* We could return the flags for the FREE_DBI too but what's the point? */
10405 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10407 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
10408 return MDB_SUCCESS;
10411 /** Add all the DB's pages to the free list.
10412 * @param[in] mc Cursor on the DB to free.
10413 * @param[in] subs non-Zero to check for sub-DBs in this DB.
10414 * @return 0 on success, non-zero on failure.
10417 mdb_drop0(MDB_cursor *mc, int subs)
10421 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
10422 if (rc == MDB_SUCCESS) {
10423 MDB_txn *txn = mc->mc_txn;
10428 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
10429 * This also avoids any P_LEAF2 pages, which have no nodes.
10431 if (mc->mc_flags & C_SUB)
10432 mdb_cursor_pop(mc);
10434 mdb_cursor_copy(mc, &mx);
10436 /* bump refcount for mx's pages */
10437 for (i=0; i<mc->mc_snum; i++)
10438 mdb_page_get(&mx, mc->mc_pg[i]->mp_pgno, &mx.mc_pg[i], NULL);
10440 while (mc->mc_snum > 0) {
10441 MDB_page *mp = mc->mc_pg[mc->mc_top];
10442 unsigned n = NUMKEYS(mp);
10444 for (i=0; i<n; i++) {
10445 ni = NODEPTR(mp, i);
10446 if (ni->mn_flags & F_BIGDATA) {
10449 memcpy(&pg, NODEDATA(ni), sizeof(pg));
10450 rc = mdb_page_get(mc, pg, &omp, NULL);
10453 mdb_cassert(mc, IS_OVERFLOW(omp));
10454 rc = mdb_midl_append_range(&txn->mt_free_pgs,
10455 pg, omp->mp_pages);
10458 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
10459 mdb_xcursor_init1(mc, ni);
10460 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
10466 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
10468 for (i=0; i<n; i++) {
10470 ni = NODEPTR(mp, i);
10473 mdb_midl_xappend(txn->mt_free_pgs, pg);
10478 mc->mc_ki[mc->mc_top] = i;
10479 rc = mdb_cursor_sibling(mc, 1);
10481 if (rc != MDB_NOTFOUND)
10483 /* no more siblings, go back to beginning
10484 * of previous level.
10486 mdb_cursor_pop(mc);
10488 for (i=1; i<mc->mc_snum; i++) {
10490 mc->mc_pg[i] = mx.mc_pg[i];
10495 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
10498 txn->mt_flags |= MDB_TXN_ERROR;
10500 /* drop refcount for mx's pages */
10501 mdb_cursor_unref(&mx);
10503 } else if (rc == MDB_NOTFOUND) {
10506 mc->mc_flags &= ~C_INITIALIZED;
10510 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
10512 MDB_cursor *mc, *m2;
10515 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10518 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
10521 if (TXN_DBI_CHANGED(txn, dbi))
10522 return MDB_BAD_DBI;
10524 rc = mdb_cursor_open(txn, dbi, &mc);
10528 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
10529 /* Invalidate the dropped DB's cursors */
10530 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
10531 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
10535 /* Can't delete the main DB */
10536 if (del && dbi >= CORE_DBS) {
10537 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
10539 txn->mt_dbflags[dbi] = DB_STALE;
10540 mdb_dbi_close(txn->mt_env, dbi);
10542 txn->mt_flags |= MDB_TXN_ERROR;
10545 /* reset the DB record, mark it dirty */
10546 txn->mt_dbflags[dbi] |= DB_DIRTY;
10547 txn->mt_dbs[dbi].md_depth = 0;
10548 txn->mt_dbs[dbi].md_branch_pages = 0;
10549 txn->mt_dbs[dbi].md_leaf_pages = 0;
10550 txn->mt_dbs[dbi].md_overflow_pages = 0;
10551 txn->mt_dbs[dbi].md_entries = 0;
10552 txn->mt_dbs[dbi].md_root = P_INVALID;
10554 txn->mt_flags |= MDB_TXN_DIRTY;
10557 mdb_cursor_close(mc);
10561 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10563 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10566 txn->mt_dbxs[dbi].md_cmp = cmp;
10567 return MDB_SUCCESS;
10570 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10572 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10575 txn->mt_dbxs[dbi].md_dcmp = cmp;
10576 return MDB_SUCCESS;
10579 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
10581 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10584 txn->mt_dbxs[dbi].md_rel = rel;
10585 return MDB_SUCCESS;
10588 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
10590 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10593 txn->mt_dbxs[dbi].md_relctx = ctx;
10594 return MDB_SUCCESS;
10598 mdb_env_get_maxkeysize(MDB_env *env)
10600 return ENV_MAXKEY(env);
10604 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
10606 unsigned int i, rdrs;
10609 int rc = 0, first = 1;
10613 if (!env->me_txns) {
10614 return func("(no reader locks)\n", ctx);
10616 rdrs = env->me_txns->mti_numreaders;
10617 mr = env->me_txns->mti_readers;
10618 for (i=0; i<rdrs; i++) {
10619 if (mr[i].mr_pid) {
10620 txnid_t txnid = mr[i].mr_txnid;
10621 sprintf(buf, txnid == (txnid_t)-1 ?
10622 "%10d %"Z"x -\n" : "%10d %"Z"x %"Y"u\n",
10623 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10626 rc = func(" pid thread txnid\n", ctx);
10630 rc = func(buf, ctx);
10636 rc = func("(no active readers)\n", ctx);
10641 /** Insert pid into list if not already present.
10642 * return -1 if already present.
10645 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10647 /* binary search of pid in list */
10649 unsigned cursor = 1;
10651 unsigned n = ids[0];
10654 unsigned pivot = n >> 1;
10655 cursor = base + pivot + 1;
10656 val = pid - ids[cursor];
10661 } else if ( val > 0 ) {
10666 /* found, so it's a duplicate */
10675 for (n = ids[0]; n > cursor; n--)
10682 mdb_reader_check(MDB_env *env, int *dead)
10688 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10691 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
10693 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10695 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10696 unsigned int i, j, rdrs;
10698 MDB_PID_T *pids, pid;
10699 int rc = MDB_SUCCESS, count = 0;
10701 rdrs = env->me_txns->mti_numreaders;
10702 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10706 mr = env->me_txns->mti_readers;
10707 for (i=0; i<rdrs; i++) {
10708 pid = mr[i].mr_pid;
10709 if (pid && pid != env->me_pid) {
10710 if (mdb_pid_insert(pids, pid) == 0) {
10711 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10712 /* Stale reader found */
10715 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10716 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10718 rdrs = 0; /* the above checked all readers */
10720 /* Recheck, a new process may have reused pid */
10721 if (mdb_reader_pid(env, Pidcheck, pid))
10725 for (; j<rdrs; j++)
10726 if (mr[j].mr_pid == pid) {
10727 DPRINTF(("clear stale reader pid %u txn %"Y"d",
10728 (unsigned) pid, mr[j].mr_txnid));
10733 UNLOCK_MUTEX(rmutex);
10744 #ifdef MDB_ROBUST_SUPPORTED
10745 /** Handle #LOCK_MUTEX0() failure.
10746 * Try to repair the lock file if the mutex owner died.
10747 * @param[in] env the environment handle
10748 * @param[in] mutex LOCK_MUTEX0() mutex
10749 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10750 * @return 0 on success with the mutex locked, or an error code on failure.
10753 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10758 if (rc == MDB_OWNERDEAD) {
10759 /* We own the mutex. Clean up after dead previous owner. */
10761 rlocked = (mutex == env->me_rmutex);
10763 /* Keep mti_txnid updated, otherwise next writer can
10764 * overwrite data which latest meta page refers to.
10766 meta = mdb_env_pick_meta(env);
10767 env->me_txns->mti_txnid = meta->mm_txnid;
10768 /* env is hosed if the dead thread was ours */
10770 env->me_flags |= MDB_FATAL_ERROR;
10771 env->me_txn = NULL;
10775 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10776 (rc ? "this process' env is hosed" : "recovering")));
10777 rc2 = mdb_reader_check0(env, rlocked, NULL);
10779 rc2 = mdb_mutex_consistent(mutex);
10780 if (rc || (rc = rc2)) {
10781 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10782 UNLOCK_MUTEX(mutex);
10788 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10793 #endif /* MDB_ROBUST_SUPPORTED */
10796 #if defined(_WIN32)
10797 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
10801 need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
10802 if (need == 0xFFFD)
10806 result = malloc(sizeof(wchar_t) * need);
10809 MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
10815 #endif /* defined(_WIN32) */