2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2015 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
82 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
83 /** fdatasync is broken on ext3/ext4fs on older kernels, see
84 * description in #mdb_env_open2 comments. You can safely
85 * define MDB_FDATASYNC_WORKS if this code will only be run
86 * on kernels 3.6 and newer.
88 #define BROKEN_FDATASYNC
101 #if defined(__sun) || defined(ANDROID)
102 /* Most platforms have posix_memalign, older may only have memalign */
103 #define HAVE_MEMALIGN 1
107 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
108 #include <netinet/in.h>
109 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
112 #if defined(__APPLE__) || defined (BSD)
113 # define MDB_USE_SYSV_SEM 1
114 # define MDB_FDATASYNC fsync
115 #elif defined(ANDROID)
116 # define MDB_FDATASYNC fsync
121 #ifdef MDB_USE_SYSV_SEM
124 #ifdef _SEM_SEMUN_UNDEFINED
127 struct semid_ds *buf;
128 unsigned short *array;
130 #endif /* _SEM_SEMUN_UNDEFINED */
131 #endif /* MDB_USE_SYSV_SEM */
135 #include <valgrind/memcheck.h>
136 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
137 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
138 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
139 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
140 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
142 #define VGMEMP_CREATE(h,r,z)
143 #define VGMEMP_ALLOC(h,a,s)
144 #define VGMEMP_FREE(h,a)
145 #define VGMEMP_DESTROY(h)
146 #define VGMEMP_DEFINED(a,s)
150 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
151 /* Solaris just defines one or the other */
152 # define LITTLE_ENDIAN 1234
153 # define BIG_ENDIAN 4321
154 # ifdef _LITTLE_ENDIAN
155 # define BYTE_ORDER LITTLE_ENDIAN
157 # define BYTE_ORDER BIG_ENDIAN
160 # define BYTE_ORDER __BYTE_ORDER
164 #ifndef LITTLE_ENDIAN
165 #define LITTLE_ENDIAN __LITTLE_ENDIAN
168 #define BIG_ENDIAN __BIG_ENDIAN
171 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
172 #define MISALIGNED_OK 1
178 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
179 # error "Unknown or unsupported endianness (BYTE_ORDER)"
180 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
181 # error "Two's complement, reasonably sized integer types, please"
185 /** Put infrequently used env functions in separate section */
187 # define ESECT __attribute__ ((section("__TEXT,text_env")))
189 # define ESECT __attribute__ ((section("text_env")))
195 /** @defgroup internal LMDB Internals
198 /** @defgroup compat Compatibility Macros
199 * A bunch of macros to minimize the amount of platform-specific ifdefs
200 * needed throughout the rest of the code. When the features this library
201 * needs are similar enough to POSIX to be hidden in a one-or-two line
202 * replacement, this macro approach is used.
206 /** Features under development */
211 #if defined(_WIN32) || defined(MDB_USE_SYSV_SEM) || defined(EOWNERDEAD)
212 #define MDB_ROBUST_SUPPORTED 1
215 /** Wrapper around __func__, which is a C99 feature */
216 #if __STDC_VERSION__ >= 199901L
217 # define mdb_func_ __func__
218 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
219 # define mdb_func_ __FUNCTION__
221 /* If a debug message says <mdb_unknown>(), update the #if statements above */
222 # define mdb_func_ "<mdb_unknown>"
225 /* Internal error codes, not exposed outside liblmdb */
226 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
228 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
229 #elif defined MDB_USE_SYSV_SEM
230 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
232 #define MDB_OWNERDEAD EOWNERDEAD
236 #define MDB_USE_HASH 1
237 #define MDB_PIDLOCK 0
238 #define THREAD_RET DWORD
239 #define pthread_t HANDLE
240 #define pthread_mutex_t HANDLE
241 #define pthread_cond_t HANDLE
242 typedef HANDLE mdb_mutex_t;
243 #define pthread_key_t DWORD
244 #define pthread_self() GetCurrentThreadId()
245 #define pthread_key_create(x,y) \
246 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
247 #define pthread_key_delete(x) TlsFree(x)
248 #define pthread_getspecific(x) TlsGetValue(x)
249 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
250 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
251 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
252 #define pthread_cond_signal(x) SetEvent(*x)
253 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
254 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
255 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
256 #define MDB_MUTEX(env, rw) ((env)->me_##rw##mutex)
257 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
258 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
259 #define mdb_mutex_consistent(mutex) 0
260 #define getpid() GetCurrentProcessId()
261 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
262 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
263 #define ErrCode() GetLastError()
264 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
265 #define close(fd) (CloseHandle(fd) ? 0 : -1)
266 #define munmap(ptr,len) UnmapViewOfFile(ptr)
267 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
268 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
270 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
274 #define THREAD_RET void *
275 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
276 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
277 #define Z "z" /**< printf format modifier for size_t */
279 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
280 #define MDB_PIDLOCK 1
282 #ifdef MDB_USE_SYSV_SEM
284 typedef struct mdb_mutex {
290 #define MDB_MUTEX(env, rw) (&(env)->me_##rw##mutex)
291 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
292 #define UNLOCK_MUTEX(mutex) do { \
293 struct sembuf sb = { 0, 1, SEM_UNDO }; \
294 sb.sem_num = (mutex)->semnum; \
295 *(mutex)->locked = 0; \
296 semop((mutex)->semid, &sb, 1); \
300 mdb_sem_wait(mdb_mutex_t *sem)
302 int rc, *locked = sem->locked;
303 struct sembuf sb = { 0, -1, SEM_UNDO };
304 sb.sem_num = sem->semnum;
306 if (!semop(sem->semid, &sb, 1)) {
307 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
311 } while ((rc = errno) == EINTR);
315 #define mdb_mutex_consistent(mutex) 0
318 /** Pointer/HANDLE type of shared mutex/semaphore.
320 typedef pthread_mutex_t mdb_mutex_t;
321 /** Mutex for the reader table (rw = r) or write transaction (rw = w).
323 #define MDB_MUTEX(env, rw) (&(env)->me_txns->mti_##rw##mutex)
324 /** Lock the reader or writer mutex.
325 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
327 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
328 /** Unlock the reader or writer mutex.
330 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
331 /** Mark mutex-protected data as repaired, after death of previous owner.
333 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
334 #endif /* MDB_USE_SYSV_SEM */
336 /** Get the error code for the last failed system function.
338 #define ErrCode() errno
340 /** An abstraction for a file handle.
341 * On POSIX systems file handles are small integers. On Windows
342 * they're opaque pointers.
346 /** A value for an invalid file handle.
347 * Mainly used to initialize file variables and signify that they are
350 #define INVALID_HANDLE_VALUE (-1)
352 /** Get the size of a memory page for the system.
353 * This is the basic size that the platform's memory manager uses, and is
354 * fundamental to the use of memory-mapped files.
356 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
361 #elif defined(MDB_USE_SYSV_SEM)
362 #define MNAME_LEN (sizeof(int))
364 #define MNAME_LEN (sizeof(pthread_mutex_t))
367 #ifdef MDB_USE_SYSV_SEM
368 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
370 #define SYSV_SEM_FLAG 0
375 #ifdef MDB_ROBUST_SUPPORTED
376 /** Lock mutex, handle any error, set rc = result.
377 * Return 0 on success, nonzero (not rc) on error.
379 #define LOCK_MUTEX(rc, env, mutex) \
380 (((rc) = LOCK_MUTEX0(mutex)) && \
381 ((rc) = mdb_mutex_failed(env, mutex, rc)))
382 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc);
384 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
385 #define mdb_mutex_failed(env, mutex, rc) (rc)
389 /** A flag for opening a file and requesting synchronous data writes.
390 * This is only used when writing a meta page. It's not strictly needed;
391 * we could just do a normal write and then immediately perform a flush.
392 * But if this flag is available it saves us an extra system call.
394 * @note If O_DSYNC is undefined but exists in /usr/include,
395 * preferably set some compiler flag to get the definition.
396 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
399 # define MDB_DSYNC O_DSYNC
403 /** Function for flushing the data of a file. Define this to fsync
404 * if fdatasync() is not supported.
406 #ifndef MDB_FDATASYNC
407 # define MDB_FDATASYNC fdatasync
411 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
422 /** A page number in the database.
423 * Note that 64 bit page numbers are overkill, since pages themselves
424 * already represent 12-13 bits of addressable memory, and the OS will
425 * always limit applications to a maximum of 63 bits of address space.
427 * @note In the #MDB_node structure, we only store 48 bits of this value,
428 * which thus limits us to only 60 bits of addressable data.
430 typedef MDB_ID pgno_t;
432 /** A transaction ID.
433 * See struct MDB_txn.mt_txnid for details.
435 typedef MDB_ID txnid_t;
437 /** @defgroup debug Debug Macros
441 /** Enable debug output. Needs variable argument macros (a C99 feature).
442 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
443 * read from and written to the database (used for free space management).
449 static int mdb_debug;
450 static txnid_t mdb_debug_start;
452 /** Print a debug message with printf formatting.
453 * Requires double parenthesis around 2 or more args.
455 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
456 # define DPRINTF0(fmt, ...) \
457 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
459 # define DPRINTF(args) ((void) 0)
461 /** Print a debug string.
462 * The string is printed literally, with no format processing.
464 #define DPUTS(arg) DPRINTF(("%s", arg))
465 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
467 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
470 /** @brief The maximum size of a database page.
472 * It is 32k or 64k, since value-PAGEBASE must fit in
473 * #MDB_page.%mp_upper.
475 * LMDB will use database pages < OS pages if needed.
476 * That causes more I/O in write transactions: The OS must
477 * know (read) the whole page before writing a partial page.
479 * Note that we don't currently support Huge pages. On Linux,
480 * regular data files cannot use Huge pages, and in general
481 * Huge pages aren't actually pageable. We rely on the OS
482 * demand-pager to read our data and page it out when memory
483 * pressure from other processes is high. So until OSs have
484 * actual paging support for Huge pages, they're not viable.
486 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
488 /** The minimum number of keys required in a database page.
489 * Setting this to a larger value will place a smaller bound on the
490 * maximum size of a data item. Data items larger than this size will
491 * be pushed into overflow pages instead of being stored directly in
492 * the B-tree node. This value used to default to 4. With a page size
493 * of 4096 bytes that meant that any item larger than 1024 bytes would
494 * go into an overflow page. That also meant that on average 2-3KB of
495 * each overflow page was wasted space. The value cannot be lower than
496 * 2 because then there would no longer be a tree structure. With this
497 * value, items larger than 2KB will go into overflow pages, and on
498 * average only 1KB will be wasted.
500 #define MDB_MINKEYS 2
502 /** A stamp that identifies a file as an LMDB file.
503 * There's nothing special about this value other than that it is easily
504 * recognizable, and it will reflect any byte order mismatches.
506 #define MDB_MAGIC 0xBEEFC0DE
508 /** The version number for a database's datafile format. */
509 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
510 /** The version number for a database's lockfile format. */
511 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
513 /** @brief The max size of a key we can write, or 0 for dynamic max.
515 * Define this as 0 to compute the max from the page size. 511
516 * is default for backwards compat: liblmdb <= 0.9.10 can break
517 * when modifying a DB with keys/dupsort data bigger than its max.
518 * #MDB_DEVEL sets the default to 0.
520 * Data items in an #MDB_DUPSORT database are also limited to
521 * this size, since they're actually keys of a sub-DB. Keys and
522 * #MDB_DUPSORT data items must fit on a node in a regular page.
524 #ifndef MDB_MAXKEYSIZE
525 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
528 /** The maximum size of a key we can write to the environment. */
530 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
532 #define ENV_MAXKEY(env) ((env)->me_maxkey)
535 /** @brief The maximum size of a data item.
537 * We only store a 32 bit value for node sizes.
539 #define MAXDATASIZE 0xffffffffUL
542 /** Key size which fits in a #DKBUF.
545 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
548 * This is used for printing a hex dump of a key's contents.
550 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
551 /** Display a key in hex.
553 * Invoke a function to display a key in hex.
555 #define DKEY(x) mdb_dkey(x, kbuf)
561 /** An invalid page number.
562 * Mainly used to denote an empty tree.
564 #define P_INVALID (~(pgno_t)0)
566 /** Test if the flags \b f are set in a flag word \b w. */
567 #define F_ISSET(w, f) (((w) & (f)) == (f))
569 /** Round \b n up to an even number. */
570 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
572 /** Used for offsets within a single page.
573 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
576 typedef uint16_t indx_t;
578 /** Default size of memory map.
579 * This is certainly too small for any actual applications. Apps should always set
580 * the size explicitly using #mdb_env_set_mapsize().
582 #define DEFAULT_MAPSIZE 1048576
584 /** @defgroup readers Reader Lock Table
585 * Readers don't acquire any locks for their data access. Instead, they
586 * simply record their transaction ID in the reader table. The reader
587 * mutex is needed just to find an empty slot in the reader table. The
588 * slot's address is saved in thread-specific data so that subsequent read
589 * transactions started by the same thread need no further locking to proceed.
591 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
593 * No reader table is used if the database is on a read-only filesystem, or
594 * if #MDB_NOLOCK is set.
596 * Since the database uses multi-version concurrency control, readers don't
597 * actually need any locking. This table is used to keep track of which
598 * readers are using data from which old transactions, so that we'll know
599 * when a particular old transaction is no longer in use. Old transactions
600 * that have discarded any data pages can then have those pages reclaimed
601 * for use by a later write transaction.
603 * The lock table is constructed such that reader slots are aligned with the
604 * processor's cache line size. Any slot is only ever used by one thread.
605 * This alignment guarantees that there will be no contention or cache
606 * thrashing as threads update their own slot info, and also eliminates
607 * any need for locking when accessing a slot.
609 * A writer thread will scan every slot in the table to determine the oldest
610 * outstanding reader transaction. Any freed pages older than this will be
611 * reclaimed by the writer. The writer doesn't use any locks when scanning
612 * this table. This means that there's no guarantee that the writer will
613 * see the most up-to-date reader info, but that's not required for correct
614 * operation - all we need is to know the upper bound on the oldest reader,
615 * we don't care at all about the newest reader. So the only consequence of
616 * reading stale information here is that old pages might hang around a
617 * while longer before being reclaimed. That's actually good anyway, because
618 * the longer we delay reclaiming old pages, the more likely it is that a
619 * string of contiguous pages can be found after coalescing old pages from
620 * many old transactions together.
623 /** Number of slots in the reader table.
624 * This value was chosen somewhat arbitrarily. 126 readers plus a
625 * couple mutexes fit exactly into 8KB on my development machine.
626 * Applications should set the table size using #mdb_env_set_maxreaders().
628 #define DEFAULT_READERS 126
630 /** The size of a CPU cache line in bytes. We want our lock structures
631 * aligned to this size to avoid false cache line sharing in the
633 * This value works for most CPUs. For Itanium this should be 128.
639 /** The information we store in a single slot of the reader table.
640 * In addition to a transaction ID, we also record the process and
641 * thread ID that owns a slot, so that we can detect stale information,
642 * e.g. threads or processes that went away without cleaning up.
643 * @note We currently don't check for stale records. We simply re-init
644 * the table when we know that we're the only process opening the
647 typedef struct MDB_rxbody {
648 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
649 * Multiple readers that start at the same time will probably have the
650 * same ID here. Again, it's not important to exclude them from
651 * anything; all we need to know is which version of the DB they
652 * started from so we can avoid overwriting any data used in that
653 * particular version.
655 volatile txnid_t mrb_txnid;
656 /** The process ID of the process owning this reader txn. */
657 volatile MDB_PID_T mrb_pid;
658 /** The thread ID of the thread owning this txn. */
659 volatile MDB_THR_T mrb_tid;
662 /** The actual reader record, with cacheline padding. */
663 typedef struct MDB_reader {
666 /** shorthand for mrb_txnid */
667 #define mr_txnid mru.mrx.mrb_txnid
668 #define mr_pid mru.mrx.mrb_pid
669 #define mr_tid mru.mrx.mrb_tid
670 /** cache line alignment */
671 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
675 /** The header for the reader table.
676 * The table resides in a memory-mapped file. (This is a different file
677 * than is used for the main database.)
679 * For POSIX the actual mutexes reside in the shared memory of this
680 * mapped file. On Windows, mutexes are named objects allocated by the
681 * kernel; we store the mutex names in this mapped file so that other
682 * processes can grab them. This same approach is also used on
683 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
684 * process-shared POSIX mutexes. For these cases where a named object
685 * is used, the object name is derived from a 64 bit FNV hash of the
686 * environment pathname. As such, naming collisions are extremely
687 * unlikely. If a collision occurs, the results are unpredictable.
689 typedef struct MDB_txbody {
690 /** Stamp identifying this as an LMDB file. It must be set
693 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
696 char mtb_rmname[MNAME_LEN];
697 #elif defined(MDB_USE_SYSV_SEM)
701 /** Mutex protecting access to this table.
702 * This is the #MDB_MUTEX(env,r) reader table lock.
704 pthread_mutex_t mtb_rmutex;
706 /** The ID of the last transaction committed to the database.
707 * This is recorded here only for convenience; the value can always
708 * be determined by reading the main database meta pages.
710 volatile txnid_t mtb_txnid;
711 /** The number of slots that have been used in the reader table.
712 * This always records the maximum count, it is not decremented
713 * when readers release their slots.
715 volatile unsigned mtb_numreaders;
718 /** The actual reader table definition. */
719 typedef struct MDB_txninfo {
722 #define mti_magic mt1.mtb.mtb_magic
723 #define mti_format mt1.mtb.mtb_format
724 #define mti_rmutex mt1.mtb.mtb_rmutex
725 #define mti_rmname mt1.mtb.mtb_rmname
726 #define mti_txnid mt1.mtb.mtb_txnid
727 #define mti_numreaders mt1.mtb.mtb_numreaders
728 #ifdef MDB_USE_SYSV_SEM
729 #define mti_semid mt1.mtb.mtb_semid
730 #define mti_rlocked mt1.mtb.mtb_rlocked
732 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
736 char mt2_wmname[MNAME_LEN];
737 #define mti_wmname mt2.mt2_wmname
738 #elif defined MDB_USE_SYSV_SEM
740 #define mti_wlocked mt2.mt2_wlocked
742 pthread_mutex_t mt2_wmutex;
743 #define mti_wmutex mt2.mt2_wmutex
745 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
747 MDB_reader mti_readers[1];
750 /** Lockfile format signature: version, features and field layout */
751 #define MDB_LOCK_FORMAT \
753 ((MDB_LOCK_VERSION) \
754 /* Flags which describe functionality */ \
755 + (SYSV_SEM_FLAG << 18) \
756 + (((MDB_PIDLOCK) != 0) << 16)))
759 /** Common header for all page types.
760 * Overflow records occupy a number of contiguous pages with no
761 * headers on any page after the first.
763 typedef struct MDB_page {
764 #define mp_pgno mp_p.p_pgno
765 #define mp_next mp_p.p_next
767 pgno_t p_pgno; /**< page number */
768 struct MDB_page *p_next; /**< for in-memory list of freed pages */
771 /** @defgroup mdb_page Page Flags
773 * Flags for the page headers.
776 #define P_BRANCH 0x01 /**< branch page */
777 #define P_LEAF 0x02 /**< leaf page */
778 #define P_OVERFLOW 0x04 /**< overflow page */
779 #define P_META 0x08 /**< meta page */
780 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
781 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
782 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
783 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
784 #define P_KEEP 0x8000 /**< leave this page alone during spill */
786 uint16_t mp_flags; /**< @ref mdb_page */
787 #define mp_lower mp_pb.pb.pb_lower
788 #define mp_upper mp_pb.pb.pb_upper
789 #define mp_pages mp_pb.pb_pages
792 indx_t pb_lower; /**< lower bound of free space */
793 indx_t pb_upper; /**< upper bound of free space */
795 uint32_t pb_pages; /**< number of overflow pages */
797 indx_t mp_ptrs[1]; /**< dynamic size */
800 /** Size of the page header, excluding dynamic data at the end */
801 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
803 /** Address of first usable data byte in a page, after the header */
804 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
806 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
807 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
809 /** Number of nodes on a page */
810 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
812 /** The amount of space remaining in the page */
813 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
815 /** The percentage of space used in the page, in tenths of a percent. */
816 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
817 ((env)->me_psize - PAGEHDRSZ))
818 /** The minimum page fill factor, in tenths of a percent.
819 * Pages emptier than this are candidates for merging.
821 #define FILL_THRESHOLD 250
823 /** Test if a page is a leaf page */
824 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
825 /** Test if a page is a LEAF2 page */
826 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
827 /** Test if a page is a branch page */
828 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
829 /** Test if a page is an overflow page */
830 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
831 /** Test if a page is a sub page */
832 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
834 /** The number of overflow pages needed to store the given size. */
835 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
837 /** Link in #MDB_txn.%mt_loose_pgs list */
838 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
840 /** Header for a single key/data pair within a page.
841 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
842 * We guarantee 2-byte alignment for 'MDB_node's.
844 typedef struct MDB_node {
845 /** lo and hi are used for data size on leaf nodes and for
846 * child pgno on branch nodes. On 64 bit platforms, flags
847 * is also used for pgno. (Branch nodes have no flags).
848 * They are in host byte order in case that lets some
849 * accesses be optimized into a 32-bit word access.
851 #if BYTE_ORDER == LITTLE_ENDIAN
852 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
854 unsigned short mn_hi, mn_lo;
856 /** @defgroup mdb_node Node Flags
858 * Flags for node headers.
861 #define F_BIGDATA 0x01 /**< data put on overflow page */
862 #define F_SUBDATA 0x02 /**< data is a sub-database */
863 #define F_DUPDATA 0x04 /**< data has duplicates */
865 /** valid flags for #mdb_node_add() */
866 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
869 unsigned short mn_flags; /**< @ref mdb_node */
870 unsigned short mn_ksize; /**< key size */
871 char mn_data[1]; /**< key and data are appended here */
874 /** Size of the node header, excluding dynamic data at the end */
875 #define NODESIZE offsetof(MDB_node, mn_data)
877 /** Bit position of top word in page number, for shifting mn_flags */
878 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
880 /** Size of a node in a branch page with a given key.
881 * This is just the node header plus the key, there is no data.
883 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
885 /** Size of a node in a leaf page with a given key and data.
886 * This is node header plus key plus data size.
888 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
890 /** Address of node \b i in page \b p */
891 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
893 /** Address of the key for the node */
894 #define NODEKEY(node) (void *)((node)->mn_data)
896 /** Address of the data for a node */
897 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
899 /** Get the page number pointed to by a branch node */
900 #define NODEPGNO(node) \
901 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
902 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
903 /** Set the page number in a branch node */
904 #define SETPGNO(node,pgno) do { \
905 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
906 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
908 /** Get the size of the data in a leaf node */
909 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
910 /** Set the size of the data for a leaf node */
911 #define SETDSZ(node,size) do { \
912 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
913 /** The size of a key in a node */
914 #define NODEKSZ(node) ((node)->mn_ksize)
916 /** Copy a page number from src to dst */
918 #define COPY_PGNO(dst,src) dst = src
920 #if SIZE_MAX > 4294967295UL
921 #define COPY_PGNO(dst,src) do { \
922 unsigned short *s, *d; \
923 s = (unsigned short *)&(src); \
924 d = (unsigned short *)&(dst); \
931 #define COPY_PGNO(dst,src) do { \
932 unsigned short *s, *d; \
933 s = (unsigned short *)&(src); \
934 d = (unsigned short *)&(dst); \
940 /** The address of a key in a LEAF2 page.
941 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
942 * There are no node headers, keys are stored contiguously.
944 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
946 /** Set the \b node's key into \b keyptr, if requested. */
947 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
948 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
950 /** Set the \b node's key into \b key. */
951 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
953 /** Information about a single database in the environment. */
954 typedef struct MDB_db {
955 uint32_t md_pad; /**< also ksize for LEAF2 pages */
956 uint16_t md_flags; /**< @ref mdb_dbi_open */
957 uint16_t md_depth; /**< depth of this tree */
958 pgno_t md_branch_pages; /**< number of internal pages */
959 pgno_t md_leaf_pages; /**< number of leaf pages */
960 pgno_t md_overflow_pages; /**< number of overflow pages */
961 size_t md_entries; /**< number of data items */
962 pgno_t md_root; /**< the root page of this tree */
965 /** mdb_dbi_open flags */
966 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
967 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
968 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
969 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
971 /** Handle for the DB used to track free pages. */
973 /** Handle for the default DB. */
976 /** Meta page content.
977 * A meta page is the start point for accessing a database snapshot.
978 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
980 typedef struct MDB_meta {
981 /** Stamp identifying this as an LMDB file. It must be set
984 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
986 void *mm_address; /**< address for fixed mapping */
987 size_t mm_mapsize; /**< size of mmap region */
988 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
989 /** The size of pages used in this DB */
990 #define mm_psize mm_dbs[0].md_pad
991 /** Any persistent environment flags. @ref mdb_env */
992 #define mm_flags mm_dbs[0].md_flags
993 pgno_t mm_last_pg; /**< last used page in file */
994 volatile txnid_t mm_txnid; /**< txnid that committed this page */
997 /** Buffer for a stack-allocated meta page.
998 * The members define size and alignment, and silence type
999 * aliasing warnings. They are not used directly; that could
1000 * mean incorrectly using several union members in parallel.
1002 typedef union MDB_metabuf {
1005 char mm_pad[PAGEHDRSZ];
1010 /** Auxiliary DB info.
1011 * The information here is mostly static/read-only. There is
1012 * only a single copy of this record in the environment.
1014 typedef struct MDB_dbx {
1015 MDB_val md_name; /**< name of the database */
1016 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1017 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1018 MDB_rel_func *md_rel; /**< user relocate function */
1019 void *md_relctx; /**< user-provided context for md_rel */
1022 /** A database transaction.
1023 * Every operation requires a transaction handle.
1026 MDB_txn *mt_parent; /**< parent of a nested txn */
1027 MDB_txn *mt_child; /**< nested txn under this txn */
1028 pgno_t mt_next_pgno; /**< next unallocated page */
1029 /** The ID of this transaction. IDs are integers incrementing from 1.
1030 * Only committed write transactions increment the ID. If a transaction
1031 * aborts, the ID may be re-used by the next writer.
1034 MDB_env *mt_env; /**< the DB environment */
1035 /** The list of pages that became unused during this transaction.
1037 MDB_IDL mt_free_pgs;
1038 /** The list of loose pages that became unused and may be reused
1039 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1041 MDB_page *mt_loose_pgs;
1042 /* #Number of loose pages (#mt_loose_pgs) */
1044 /** The sorted list of dirty pages we temporarily wrote to disk
1045 * because the dirty list was full. page numbers in here are
1046 * shifted left by 1, deleted slots have the LSB set.
1048 MDB_IDL mt_spill_pgs;
1050 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1051 MDB_ID2L dirty_list;
1052 /** For read txns: This thread/txn's reader table slot, or NULL. */
1055 /** Array of records for each DB known in the environment. */
1057 /** Array of MDB_db records for each known DB */
1059 /** Array of sequence numbers for each DB handle */
1060 unsigned int *mt_dbiseqs;
1061 /** @defgroup mt_dbflag Transaction DB Flags
1065 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1066 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1067 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1068 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1070 /** In write txns, array of cursors for each DB */
1071 MDB_cursor **mt_cursors;
1072 /** Array of flags for each DB */
1073 unsigned char *mt_dbflags;
1074 /** Number of DB records in use. This number only ever increments;
1075 * we don't decrement it when individual DB handles are closed.
1079 /** @defgroup mdb_txn Transaction Flags
1083 /** #mdb_txn_begin() flags */
1084 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1085 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1086 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1087 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1088 /* internal txn flags */
1089 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1090 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1091 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1092 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1094 unsigned int mt_flags; /**< @ref mdb_txn */
1095 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1096 * Includes ancestor txns' dirty pages not hidden by other txns'
1097 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1098 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1100 unsigned int mt_dirty_room;
1103 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1104 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1105 * raise this on a 64 bit machine.
1107 #define CURSOR_STACK 32
1111 /** Cursors are used for all DB operations.
1112 * A cursor holds a path of (page pointer, key index) from the DB
1113 * root to a position in the DB, plus other state. #MDB_DUPSORT
1114 * cursors include an xcursor to the current data item. Write txns
1115 * track their cursors and keep them up to date when data moves.
1116 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1117 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1120 /** Next cursor on this DB in this txn */
1121 MDB_cursor *mc_next;
1122 /** Backup of the original cursor if this cursor is a shadow */
1123 MDB_cursor *mc_backup;
1124 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1125 struct MDB_xcursor *mc_xcursor;
1126 /** The transaction that owns this cursor */
1128 /** The database handle this cursor operates on */
1130 /** The database record for this cursor */
1132 /** The database auxiliary record for this cursor */
1134 /** The @ref mt_dbflag for this database */
1135 unsigned char *mc_dbflag;
1136 unsigned short mc_snum; /**< number of pushed pages */
1137 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1138 /** @defgroup mdb_cursor Cursor Flags
1140 * Cursor state flags.
1143 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1144 #define C_EOF 0x02 /**< No more data */
1145 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1146 #define C_DEL 0x08 /**< last op was a cursor_del */
1147 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1148 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1150 unsigned int mc_flags; /**< @ref mdb_cursor */
1151 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1152 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1155 /** Context for sorted-dup records.
1156 * We could have gone to a fully recursive design, with arbitrarily
1157 * deep nesting of sub-databases. But for now we only handle these
1158 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1160 typedef struct MDB_xcursor {
1161 /** A sub-cursor for traversing the Dup DB */
1162 MDB_cursor mx_cursor;
1163 /** The database record for this Dup DB */
1165 /** The auxiliary DB record for this Dup DB */
1167 /** The @ref mt_dbflag for this Dup DB */
1168 unsigned char mx_dbflag;
1171 /** State of FreeDB old pages, stored in the MDB_env */
1172 typedef struct MDB_pgstate {
1173 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1174 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1177 /** The database environment. */
1179 HANDLE me_fd; /**< The main data file */
1180 HANDLE me_lfd; /**< The lock file */
1181 HANDLE me_mfd; /**< just for writing the meta pages */
1182 /** Failed to update the meta page. Probably an I/O error. */
1183 #define MDB_FATAL_ERROR 0x80000000U
1184 /** Some fields are initialized. */
1185 #define MDB_ENV_ACTIVE 0x20000000U
1186 /** me_txkey is set */
1187 #define MDB_ENV_TXKEY 0x10000000U
1188 /** fdatasync is unreliable */
1189 #define MDB_FSYNCONLY 0x08000000U
1190 uint32_t me_flags; /**< @ref mdb_env */
1191 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1192 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1193 unsigned int me_maxreaders; /**< size of the reader table */
1194 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1195 volatile int me_close_readers;
1196 MDB_dbi me_numdbs; /**< number of DBs opened */
1197 MDB_dbi me_maxdbs; /**< size of the DB table */
1198 MDB_PID_T me_pid; /**< process ID of this env */
1199 char *me_path; /**< path to the DB files */
1200 char *me_map; /**< the memory map of the data file */
1201 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1202 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1203 void *me_pbuf; /**< scratch area for DUPSORT put() */
1204 MDB_txn *me_txn; /**< current write transaction */
1205 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1206 size_t me_mapsize; /**< size of the data memory map */
1207 off_t me_size; /**< current file size */
1208 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1209 MDB_dbx *me_dbxs; /**< array of static DB info */
1210 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1211 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1212 pthread_key_t me_txkey; /**< thread-key for readers */
1213 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1214 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1215 # define me_pglast me_pgstate.mf_pglast
1216 # define me_pghead me_pgstate.mf_pghead
1217 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1218 /** IDL of pages that became unused in a write txn */
1219 MDB_IDL me_free_pgs;
1220 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1221 MDB_ID2L me_dirty_list;
1222 /** Max number of freelist items that can fit in a single overflow page */
1224 /** Max size of a node on a page */
1225 unsigned int me_nodemax;
1226 #if !(MDB_MAXKEYSIZE)
1227 unsigned int me_maxkey; /**< max size of a key */
1229 int me_live_reader; /**< have liveness lock in reader table */
1231 int me_pidquery; /**< Used in OpenProcess */
1233 #if defined(_WIN32) || defined(MDB_USE_SYSV_SEM)
1234 /* Windows mutexes/SysV semaphores do not reside in shared mem */
1235 mdb_mutex_t me_rmutex;
1236 mdb_mutex_t me_wmutex;
1238 void *me_userctx; /**< User-settable context */
1239 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1242 /** Nested transaction */
1243 typedef struct MDB_ntxn {
1244 MDB_txn mnt_txn; /**< the transaction */
1245 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1248 /** max number of pages to commit in one writev() call */
1249 #define MDB_COMMIT_PAGES 64
1250 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1251 #undef MDB_COMMIT_PAGES
1252 #define MDB_COMMIT_PAGES IOV_MAX
1255 /** max bytes to write in one call */
1256 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1258 /** Check \b txn and \b dbi arguments to a function */
1259 #define TXN_DBI_EXIST(txn, dbi) \
1260 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1262 /** Check for misused \b dbi handles */
1263 #define TXN_DBI_CHANGED(txn, dbi) \
1264 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1266 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1267 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1268 static int mdb_page_touch(MDB_cursor *mc);
1270 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1271 static int mdb_page_search_root(MDB_cursor *mc,
1272 MDB_val *key, int modify);
1273 #define MDB_PS_MODIFY 1
1274 #define MDB_PS_ROOTONLY 2
1275 #define MDB_PS_FIRST 4
1276 #define MDB_PS_LAST 8
1277 static int mdb_page_search(MDB_cursor *mc,
1278 MDB_val *key, int flags);
1279 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1281 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1282 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1283 pgno_t newpgno, unsigned int nflags);
1285 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1286 static int mdb_env_pick_meta(const MDB_env *env);
1287 static int mdb_env_write_meta(MDB_txn *txn);
1288 #if !(defined(_WIN32) || defined(MDB_USE_SYSV_SEM)) /* Drop unused excl arg */
1289 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1291 static void mdb_env_close0(MDB_env *env, int excl);
1293 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1294 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1295 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1296 static void mdb_node_del(MDB_cursor *mc, int ksize);
1297 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1298 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1299 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1300 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1301 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1303 static int mdb_rebalance(MDB_cursor *mc);
1304 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1306 static void mdb_cursor_pop(MDB_cursor *mc);
1307 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1309 static int mdb_cursor_del0(MDB_cursor *mc);
1310 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1311 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1312 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1313 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1314 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1316 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1317 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1319 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1320 static void mdb_xcursor_init0(MDB_cursor *mc);
1321 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1323 static int mdb_drop0(MDB_cursor *mc, int subs);
1324 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1325 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1328 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1332 static SECURITY_DESCRIPTOR mdb_null_sd;
1333 static SECURITY_ATTRIBUTES mdb_all_sa;
1334 static int mdb_sec_inited;
1337 /** Return the library version info. */
1339 mdb_version(int *major, int *minor, int *patch)
1341 if (major) *major = MDB_VERSION_MAJOR;
1342 if (minor) *minor = MDB_VERSION_MINOR;
1343 if (patch) *patch = MDB_VERSION_PATCH;
1344 return MDB_VERSION_STRING;
1347 /** Table of descriptions for LMDB @ref errors */
1348 static char *const mdb_errstr[] = {
1349 "MDB_KEYEXIST: Key/data pair already exists",
1350 "MDB_NOTFOUND: No matching key/data pair found",
1351 "MDB_PAGE_NOTFOUND: Requested page not found",
1352 "MDB_CORRUPTED: Located page was wrong type",
1353 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1354 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1355 "MDB_INVALID: File is not an LMDB file",
1356 "MDB_MAP_FULL: Environment mapsize limit reached",
1357 "MDB_DBS_FULL: Environment maxdbs limit reached",
1358 "MDB_READERS_FULL: Environment maxreaders limit reached",
1359 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1360 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1361 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1362 "MDB_PAGE_FULL: Internal error - page has no more space",
1363 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1364 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1365 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1366 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1367 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1368 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1372 mdb_strerror(int err)
1375 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1376 * This works as long as no function between the call to mdb_strerror
1377 * and the actual use of the message uses more than 4K of stack.
1380 char buf[1024], *ptr = buf;
1384 return ("Successful return: 0");
1386 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1387 i = err - MDB_KEYEXIST;
1388 return mdb_errstr[i];
1392 /* These are the C-runtime error codes we use. The comment indicates
1393 * their numeric value, and the Win32 error they would correspond to
1394 * if the error actually came from a Win32 API. A major mess, we should
1395 * have used LMDB-specific error codes for everything.
1398 case ENOENT: /* 2, FILE_NOT_FOUND */
1399 case EIO: /* 5, ACCESS_DENIED */
1400 case ENOMEM: /* 12, INVALID_ACCESS */
1401 case EACCES: /* 13, INVALID_DATA */
1402 case EBUSY: /* 16, CURRENT_DIRECTORY */
1403 case EINVAL: /* 22, BAD_COMMAND */
1404 case ENOSPC: /* 28, OUT_OF_PAPER */
1405 return strerror(err);
1410 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1411 FORMAT_MESSAGE_IGNORE_INSERTS,
1412 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1415 return strerror(err);
1419 /** assert(3) variant in cursor context */
1420 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1421 /** assert(3) variant in transaction context */
1422 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1423 /** assert(3) variant in environment context */
1424 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1427 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1428 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1431 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1432 const char *func, const char *file, int line)
1435 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1436 file, line, expr_txt, func);
1437 if (env->me_assert_func)
1438 env->me_assert_func(env, buf);
1439 fprintf(stderr, "%s\n", buf);
1443 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1447 /** Return the page number of \b mp which may be sub-page, for debug output */
1449 mdb_dbg_pgno(MDB_page *mp)
1452 COPY_PGNO(ret, mp->mp_pgno);
1456 /** Display a key in hexadecimal and return the address of the result.
1457 * @param[in] key the key to display
1458 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1459 * @return The key in hexadecimal form.
1462 mdb_dkey(MDB_val *key, char *buf)
1465 unsigned char *c = key->mv_data;
1471 if (key->mv_size > DKBUF_MAXKEYSIZE)
1472 return "MDB_MAXKEYSIZE";
1473 /* may want to make this a dynamic check: if the key is mostly
1474 * printable characters, print it as-is instead of converting to hex.
1478 for (i=0; i<key->mv_size; i++)
1479 ptr += sprintf(ptr, "%02x", *c++);
1481 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1487 mdb_leafnode_type(MDB_node *n)
1489 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1490 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1491 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1494 /** Display all the keys in the page. */
1496 mdb_page_list(MDB_page *mp)
1498 pgno_t pgno = mdb_dbg_pgno(mp);
1499 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1501 unsigned int i, nkeys, nsize, total = 0;
1505 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1506 case P_BRANCH: type = "Branch page"; break;
1507 case P_LEAF: type = "Leaf page"; break;
1508 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1509 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1510 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1512 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1513 pgno, mp->mp_pages, state);
1516 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1517 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1520 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1524 nkeys = NUMKEYS(mp);
1525 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1527 for (i=0; i<nkeys; i++) {
1528 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1529 key.mv_size = nsize = mp->mp_pad;
1530 key.mv_data = LEAF2KEY(mp, i, nsize);
1532 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1535 node = NODEPTR(mp, i);
1536 key.mv_size = node->mn_ksize;
1537 key.mv_data = node->mn_data;
1538 nsize = NODESIZE + key.mv_size;
1539 if (IS_BRANCH(mp)) {
1540 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1544 if (F_ISSET(node->mn_flags, F_BIGDATA))
1545 nsize += sizeof(pgno_t);
1547 nsize += NODEDSZ(node);
1549 nsize += sizeof(indx_t);
1550 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1551 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1553 total = EVEN(total);
1555 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1556 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1560 mdb_cursor_chk(MDB_cursor *mc)
1566 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1567 for (i=0; i<mc->mc_top; i++) {
1569 node = NODEPTR(mp, mc->mc_ki[i]);
1570 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1573 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1579 /** Count all the pages in each DB and in the freelist
1580 * and make sure it matches the actual number of pages
1582 * All named DBs must be open for a correct count.
1584 static void mdb_audit(MDB_txn *txn)
1588 MDB_ID freecount, count;
1593 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1594 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1595 freecount += *(MDB_ID *)data.mv_data;
1596 mdb_tassert(txn, rc == MDB_NOTFOUND);
1599 for (i = 0; i<txn->mt_numdbs; i++) {
1601 if (!(txn->mt_dbflags[i] & DB_VALID))
1603 mdb_cursor_init(&mc, txn, i, &mx);
1604 if (txn->mt_dbs[i].md_root == P_INVALID)
1606 count += txn->mt_dbs[i].md_branch_pages +
1607 txn->mt_dbs[i].md_leaf_pages +
1608 txn->mt_dbs[i].md_overflow_pages;
1609 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1610 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1611 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1614 mp = mc.mc_pg[mc.mc_top];
1615 for (j=0; j<NUMKEYS(mp); j++) {
1616 MDB_node *leaf = NODEPTR(mp, j);
1617 if (leaf->mn_flags & F_SUBDATA) {
1619 memcpy(&db, NODEDATA(leaf), sizeof(db));
1620 count += db.md_branch_pages + db.md_leaf_pages +
1621 db.md_overflow_pages;
1625 mdb_tassert(txn, rc == MDB_NOTFOUND);
1628 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1629 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1630 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1636 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1638 return txn->mt_dbxs[dbi].md_cmp(a, b);
1642 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1644 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1647 /** Allocate memory for a page.
1648 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1651 mdb_page_malloc(MDB_txn *txn, unsigned num)
1653 MDB_env *env = txn->mt_env;
1654 MDB_page *ret = env->me_dpages;
1655 size_t psize = env->me_psize, sz = psize, off;
1656 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1657 * For a single page alloc, we init everything after the page header.
1658 * For multi-page, we init the final page; if the caller needed that
1659 * many pages they will be filling in at least up to the last page.
1663 VGMEMP_ALLOC(env, ret, sz);
1664 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1665 env->me_dpages = ret->mp_next;
1668 psize -= off = PAGEHDRSZ;
1673 if ((ret = malloc(sz)) != NULL) {
1674 VGMEMP_ALLOC(env, ret, sz);
1675 if (!(env->me_flags & MDB_NOMEMINIT)) {
1676 memset((char *)ret + off, 0, psize);
1680 txn->mt_flags |= MDB_TXN_ERROR;
1684 /** Free a single page.
1685 * Saves single pages to a list, for future reuse.
1686 * (This is not used for multi-page overflow pages.)
1689 mdb_page_free(MDB_env *env, MDB_page *mp)
1691 mp->mp_next = env->me_dpages;
1692 VGMEMP_FREE(env, mp);
1693 env->me_dpages = mp;
1696 /** Free a dirty page */
1698 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1700 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1701 mdb_page_free(env, dp);
1703 /* large pages just get freed directly */
1704 VGMEMP_FREE(env, dp);
1709 /** Return all dirty pages to dpage list */
1711 mdb_dlist_free(MDB_txn *txn)
1713 MDB_env *env = txn->mt_env;
1714 MDB_ID2L dl = txn->mt_u.dirty_list;
1715 unsigned i, n = dl[0].mid;
1717 for (i = 1; i <= n; i++) {
1718 mdb_dpage_free(env, dl[i].mptr);
1723 /** Loosen or free a single page.
1724 * Saves single pages to a list for future reuse
1725 * in this same txn. It has been pulled from the freeDB
1726 * and already resides on the dirty list, but has been
1727 * deleted. Use these pages first before pulling again
1730 * If the page wasn't dirtied in this txn, just add it
1731 * to this txn's free list.
1734 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1737 pgno_t pgno = mp->mp_pgno;
1738 MDB_txn *txn = mc->mc_txn;
1740 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1741 if (txn->mt_parent) {
1742 MDB_ID2 *dl = txn->mt_u.dirty_list;
1743 /* If txn has a parent, make sure the page is in our
1747 unsigned x = mdb_mid2l_search(dl, pgno);
1748 if (x <= dl[0].mid && dl[x].mid == pgno) {
1749 if (mp != dl[x].mptr) { /* bad cursor? */
1750 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1751 txn->mt_flags |= MDB_TXN_ERROR;
1752 return MDB_CORRUPTED;
1759 /* no parent txn, so it's just ours */
1764 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1766 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1767 txn->mt_loose_pgs = mp;
1768 txn->mt_loose_count++;
1769 mp->mp_flags |= P_LOOSE;
1771 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1779 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1780 * @param[in] mc A cursor handle for the current operation.
1781 * @param[in] pflags Flags of the pages to update:
1782 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1783 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1784 * @return 0 on success, non-zero on failure.
1787 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1789 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1790 MDB_txn *txn = mc->mc_txn;
1796 int rc = MDB_SUCCESS, level;
1798 /* Mark pages seen by cursors */
1799 if (mc->mc_flags & C_UNTRACK)
1800 mc = NULL; /* will find mc in mt_cursors */
1801 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1802 for (; mc; mc=mc->mc_next) {
1803 if (!(mc->mc_flags & C_INITIALIZED))
1805 for (m3 = mc;; m3 = &mx->mx_cursor) {
1807 for (j=0; j<m3->mc_snum; j++) {
1809 if ((mp->mp_flags & Mask) == pflags)
1810 mp->mp_flags ^= P_KEEP;
1812 mx = m3->mc_xcursor;
1813 /* Proceed to mx if it is at a sub-database */
1814 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1816 if (! (mp && (mp->mp_flags & P_LEAF)))
1818 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1819 if (!(leaf->mn_flags & F_SUBDATA))
1828 /* Mark dirty root pages */
1829 for (i=0; i<txn->mt_numdbs; i++) {
1830 if (txn->mt_dbflags[i] & DB_DIRTY) {
1831 pgno_t pgno = txn->mt_dbs[i].md_root;
1832 if (pgno == P_INVALID)
1834 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1836 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1837 dp->mp_flags ^= P_KEEP;
1845 static int mdb_page_flush(MDB_txn *txn, int keep);
1847 /** Spill pages from the dirty list back to disk.
1848 * This is intended to prevent running into #MDB_TXN_FULL situations,
1849 * but note that they may still occur in a few cases:
1850 * 1) our estimate of the txn size could be too small. Currently this
1851 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1852 * 2) child txns may run out of space if their parents dirtied a
1853 * lot of pages and never spilled them. TODO: we probably should do
1854 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1855 * the parent's dirty_room is below a given threshold.
1857 * Otherwise, if not using nested txns, it is expected that apps will
1858 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1859 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1860 * If the txn never references them again, they can be left alone.
1861 * If the txn only reads them, they can be used without any fuss.
1862 * If the txn writes them again, they can be dirtied immediately without
1863 * going thru all of the work of #mdb_page_touch(). Such references are
1864 * handled by #mdb_page_unspill().
1866 * Also note, we never spill DB root pages, nor pages of active cursors,
1867 * because we'll need these back again soon anyway. And in nested txns,
1868 * we can't spill a page in a child txn if it was already spilled in a
1869 * parent txn. That would alter the parent txns' data even though
1870 * the child hasn't committed yet, and we'd have no way to undo it if
1871 * the child aborted.
1873 * @param[in] m0 cursor A cursor handle identifying the transaction and
1874 * database for which we are checking space.
1875 * @param[in] key For a put operation, the key being stored.
1876 * @param[in] data For a put operation, the data being stored.
1877 * @return 0 on success, non-zero on failure.
1880 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1882 MDB_txn *txn = m0->mc_txn;
1884 MDB_ID2L dl = txn->mt_u.dirty_list;
1885 unsigned int i, j, need;
1888 if (m0->mc_flags & C_SUB)
1891 /* Estimate how much space this op will take */
1892 i = m0->mc_db->md_depth;
1893 /* Named DBs also dirty the main DB */
1894 if (m0->mc_dbi > MAIN_DBI)
1895 i += txn->mt_dbs[MAIN_DBI].md_depth;
1896 /* For puts, roughly factor in the key+data size */
1898 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1899 i += i; /* double it for good measure */
1902 if (txn->mt_dirty_room > i)
1905 if (!txn->mt_spill_pgs) {
1906 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1907 if (!txn->mt_spill_pgs)
1910 /* purge deleted slots */
1911 MDB_IDL sl = txn->mt_spill_pgs;
1912 unsigned int num = sl[0];
1914 for (i=1; i<=num; i++) {
1921 /* Preserve pages which may soon be dirtied again */
1922 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1925 /* Less aggressive spill - we originally spilled the entire dirty list,
1926 * with a few exceptions for cursor pages and DB root pages. But this
1927 * turns out to be a lot of wasted effort because in a large txn many
1928 * of those pages will need to be used again. So now we spill only 1/8th
1929 * of the dirty pages. Testing revealed this to be a good tradeoff,
1930 * better than 1/2, 1/4, or 1/10.
1932 if (need < MDB_IDL_UM_MAX / 8)
1933 need = MDB_IDL_UM_MAX / 8;
1935 /* Save the page IDs of all the pages we're flushing */
1936 /* flush from the tail forward, this saves a lot of shifting later on. */
1937 for (i=dl[0].mid; i && need; i--) {
1938 MDB_ID pn = dl[i].mid << 1;
1940 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1942 /* Can't spill twice, make sure it's not already in a parent's
1945 if (txn->mt_parent) {
1947 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1948 if (tx2->mt_spill_pgs) {
1949 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1950 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1951 dp->mp_flags |= P_KEEP;
1959 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1963 mdb_midl_sort(txn->mt_spill_pgs);
1965 /* Flush the spilled part of dirty list */
1966 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1969 /* Reset any dirty pages we kept that page_flush didn't see */
1970 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1973 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1977 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1979 mdb_find_oldest(MDB_txn *txn)
1982 txnid_t mr, oldest = txn->mt_txnid - 1;
1983 if (txn->mt_env->me_txns) {
1984 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1985 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1996 /** Add a page to the txn's dirty list */
1998 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2001 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2003 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2004 insert = mdb_mid2l_append;
2006 insert = mdb_mid2l_insert;
2008 mid.mid = mp->mp_pgno;
2010 rc = insert(txn->mt_u.dirty_list, &mid);
2011 mdb_tassert(txn, rc == 0);
2012 txn->mt_dirty_room--;
2015 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2016 * me_pghead and mt_next_pgno.
2018 * If there are free pages available from older transactions, they
2019 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2020 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2021 * and move me_pglast to say which records were consumed. Only this
2022 * function can create me_pghead and move me_pglast/mt_next_pgno.
2023 * @param[in] mc cursor A cursor handle identifying the transaction and
2024 * database for which we are allocating.
2025 * @param[in] num the number of pages to allocate.
2026 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2027 * will always be satisfied by a single contiguous chunk of memory.
2028 * @return 0 on success, non-zero on failure.
2031 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2033 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2034 /* Get at most <Max_retries> more freeDB records once me_pghead
2035 * has enough pages. If not enough, use new pages from the map.
2036 * If <Paranoid> and mc is updating the freeDB, only get new
2037 * records if me_pghead is empty. Then the freelist cannot play
2038 * catch-up with itself by growing while trying to save it.
2040 enum { Paranoid = 1, Max_retries = 500 };
2042 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2044 int rc, retry = num * 60;
2045 MDB_txn *txn = mc->mc_txn;
2046 MDB_env *env = txn->mt_env;
2047 pgno_t pgno, *mop = env->me_pghead;
2048 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2050 txnid_t oldest = 0, last;
2055 /* If there are any loose pages, just use them */
2056 if (num == 1 && txn->mt_loose_pgs) {
2057 np = txn->mt_loose_pgs;
2058 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2059 txn->mt_loose_count--;
2060 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2068 /* If our dirty list is already full, we can't do anything */
2069 if (txn->mt_dirty_room == 0) {
2074 for (op = MDB_FIRST;; op = MDB_NEXT) {
2079 /* Seek a big enough contiguous page range. Prefer
2080 * pages at the tail, just truncating the list.
2086 if (mop[i-n2] == pgno+n2)
2093 if (op == MDB_FIRST) { /* 1st iteration */
2094 /* Prepare to fetch more and coalesce */
2095 last = env->me_pglast;
2096 oldest = env->me_pgoldest;
2097 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2100 key.mv_data = &last; /* will look up last+1 */
2101 key.mv_size = sizeof(last);
2103 if (Paranoid && mc->mc_dbi == FREE_DBI)
2106 if (Paranoid && retry < 0 && mop_len)
2110 /* Do not fetch more if the record will be too recent */
2111 if (oldest <= last) {
2113 oldest = mdb_find_oldest(txn);
2114 env->me_pgoldest = oldest;
2120 rc = mdb_cursor_get(&m2, &key, NULL, op);
2122 if (rc == MDB_NOTFOUND)
2126 last = *(txnid_t*)key.mv_data;
2127 if (oldest <= last) {
2129 oldest = mdb_find_oldest(txn);
2130 env->me_pgoldest = oldest;
2136 np = m2.mc_pg[m2.mc_top];
2137 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2138 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2141 idl = (MDB_ID *) data.mv_data;
2144 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2149 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2151 mop = env->me_pghead;
2153 env->me_pglast = last;
2155 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2156 last, txn->mt_dbs[FREE_DBI].md_root, i));
2158 DPRINTF(("IDL %"Z"u", idl[j]));
2160 /* Merge in descending sorted order */
2161 mdb_midl_xmerge(mop, idl);
2165 /* Use new pages from the map when nothing suitable in the freeDB */
2167 pgno = txn->mt_next_pgno;
2168 if (pgno + num >= env->me_maxpg) {
2169 DPUTS("DB size maxed out");
2175 if (env->me_flags & MDB_WRITEMAP) {
2176 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2178 if (!(np = mdb_page_malloc(txn, num))) {
2184 mop[0] = mop_len -= num;
2185 /* Move any stragglers down */
2186 for (j = i-num; j < mop_len; )
2187 mop[++j] = mop[++i];
2189 txn->mt_next_pgno = pgno + num;
2192 mdb_page_dirty(txn, np);
2198 txn->mt_flags |= MDB_TXN_ERROR;
2202 /** Copy the used portions of a non-overflow page.
2203 * @param[in] dst page to copy into
2204 * @param[in] src page to copy from
2205 * @param[in] psize size of a page
2208 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2210 enum { Align = sizeof(pgno_t) };
2211 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2213 /* If page isn't full, just copy the used portion. Adjust
2214 * alignment so memcpy may copy words instead of bytes.
2216 if ((unused &= -Align) && !IS_LEAF2(src)) {
2217 upper = (upper + PAGEBASE) & -Align;
2218 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2219 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2222 memcpy(dst, src, psize - unused);
2226 /** Pull a page off the txn's spill list, if present.
2227 * If a page being referenced was spilled to disk in this txn, bring
2228 * it back and make it dirty/writable again.
2229 * @param[in] txn the transaction handle.
2230 * @param[in] mp the page being referenced. It must not be dirty.
2231 * @param[out] ret the writable page, if any. ret is unchanged if
2232 * mp wasn't spilled.
2235 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2237 MDB_env *env = txn->mt_env;
2240 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2242 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2243 if (!tx2->mt_spill_pgs)
2245 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2246 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2249 if (txn->mt_dirty_room == 0)
2250 return MDB_TXN_FULL;
2251 if (IS_OVERFLOW(mp))
2255 if (env->me_flags & MDB_WRITEMAP) {
2258 np = mdb_page_malloc(txn, num);
2262 memcpy(np, mp, num * env->me_psize);
2264 mdb_page_copy(np, mp, env->me_psize);
2267 /* If in current txn, this page is no longer spilled.
2268 * If it happens to be the last page, truncate the spill list.
2269 * Otherwise mark it as deleted by setting the LSB.
2271 if (x == txn->mt_spill_pgs[0])
2272 txn->mt_spill_pgs[0]--;
2274 txn->mt_spill_pgs[x] |= 1;
2275 } /* otherwise, if belonging to a parent txn, the
2276 * page remains spilled until child commits
2279 mdb_page_dirty(txn, np);
2280 np->mp_flags |= P_DIRTY;
2288 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2289 * @param[in] mc cursor pointing to the page to be touched
2290 * @return 0 on success, non-zero on failure.
2293 mdb_page_touch(MDB_cursor *mc)
2295 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2296 MDB_txn *txn = mc->mc_txn;
2297 MDB_cursor *m2, *m3;
2301 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2302 if (txn->mt_flags & MDB_TXN_SPILLS) {
2304 rc = mdb_page_unspill(txn, mp, &np);
2310 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2311 (rc = mdb_page_alloc(mc, 1, &np)))
2314 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2315 mp->mp_pgno, pgno));
2316 mdb_cassert(mc, mp->mp_pgno != pgno);
2317 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2318 /* Update the parent page, if any, to point to the new page */
2320 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2321 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2322 SETPGNO(node, pgno);
2324 mc->mc_db->md_root = pgno;
2326 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2327 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2329 /* If txn has a parent, make sure the page is in our
2333 unsigned x = mdb_mid2l_search(dl, pgno);
2334 if (x <= dl[0].mid && dl[x].mid == pgno) {
2335 if (mp != dl[x].mptr) { /* bad cursor? */
2336 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2337 txn->mt_flags |= MDB_TXN_ERROR;
2338 return MDB_CORRUPTED;
2343 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2345 np = mdb_page_malloc(txn, 1);
2350 rc = mdb_mid2l_insert(dl, &mid);
2351 mdb_cassert(mc, rc == 0);
2356 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2358 np->mp_flags |= P_DIRTY;
2361 /* Adjust cursors pointing to mp */
2362 mc->mc_pg[mc->mc_top] = np;
2363 m2 = txn->mt_cursors[mc->mc_dbi];
2364 if (mc->mc_flags & C_SUB) {
2365 for (; m2; m2=m2->mc_next) {
2366 m3 = &m2->mc_xcursor->mx_cursor;
2367 if (m3->mc_snum < mc->mc_snum) continue;
2368 if (m3->mc_pg[mc->mc_top] == mp)
2369 m3->mc_pg[mc->mc_top] = np;
2372 for (; m2; m2=m2->mc_next) {
2373 if (m2->mc_snum < mc->mc_snum) continue;
2374 if (m2->mc_pg[mc->mc_top] == mp) {
2375 m2->mc_pg[mc->mc_top] = np;
2376 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2378 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2380 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2381 if (!(leaf->mn_flags & F_SUBDATA))
2382 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2390 txn->mt_flags |= MDB_TXN_ERROR;
2395 mdb_env_sync(MDB_env *env, int force)
2398 if (env->me_flags & MDB_RDONLY)
2400 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2401 if (env->me_flags & MDB_WRITEMAP) {
2402 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2403 ? MS_ASYNC : MS_SYNC;
2404 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2407 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2411 #ifdef BROKEN_FDATASYNC
2412 if (env->me_flags & MDB_FSYNCONLY) {
2413 if (fsync(env->me_fd))
2417 if (MDB_FDATASYNC(env->me_fd))
2424 /** Back up parent txn's cursors, then grab the originals for tracking */
2426 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2428 MDB_cursor *mc, *bk;
2433 for (i = src->mt_numdbs; --i >= 0; ) {
2434 if ((mc = src->mt_cursors[i]) != NULL) {
2435 size = sizeof(MDB_cursor);
2437 size += sizeof(MDB_xcursor);
2438 for (; mc; mc = bk->mc_next) {
2444 mc->mc_db = &dst->mt_dbs[i];
2445 /* Kill pointers into src - and dst to reduce abuse: The
2446 * user may not use mc until dst ends. Otherwise we'd...
2448 mc->mc_txn = NULL; /* ...set this to dst */
2449 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2450 if ((mx = mc->mc_xcursor) != NULL) {
2451 *(MDB_xcursor *)(bk+1) = *mx;
2452 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2454 mc->mc_next = dst->mt_cursors[i];
2455 dst->mt_cursors[i] = mc;
2462 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2463 * @param[in] txn the transaction handle.
2464 * @param[in] merge true to keep changes to parent cursors, false to revert.
2465 * @return 0 on success, non-zero on failure.
2468 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2470 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2474 for (i = txn->mt_numdbs; --i >= 0; ) {
2475 for (mc = cursors[i]; mc; mc = next) {
2477 if ((bk = mc->mc_backup) != NULL) {
2479 /* Commit changes to parent txn */
2480 mc->mc_next = bk->mc_next;
2481 mc->mc_backup = bk->mc_backup;
2482 mc->mc_txn = bk->mc_txn;
2483 mc->mc_db = bk->mc_db;
2484 mc->mc_dbflag = bk->mc_dbflag;
2485 if ((mx = mc->mc_xcursor) != NULL)
2486 mx->mx_cursor.mc_txn = bk->mc_txn;
2488 /* Abort nested txn */
2490 if ((mx = mc->mc_xcursor) != NULL)
2491 *mx = *(MDB_xcursor *)(bk+1);
2495 /* Only malloced cursors are permanently tracked. */
2503 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2506 mdb_txn_reset0(MDB_txn *txn, const char *act);
2508 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2514 Pidset = F_SETLK, Pidcheck = F_GETLK
2518 /** Set or check a pid lock. Set returns 0 on success.
2519 * Check returns 0 if the process is certainly dead, nonzero if it may
2520 * be alive (the lock exists or an error happened so we do not know).
2522 * On Windows Pidset is a no-op, we merely check for the existence
2523 * of the process with the given pid. On POSIX we use a single byte
2524 * lock on the lockfile, set at an offset equal to the pid.
2527 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2529 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2532 if (op == Pidcheck) {
2533 h = OpenProcess(env->me_pidquery, FALSE, pid);
2534 /* No documented "no such process" code, but other program use this: */
2536 return ErrCode() != ERROR_INVALID_PARAMETER;
2537 /* A process exists until all handles to it close. Has it exited? */
2538 ret = WaitForSingleObject(h, 0) != 0;
2545 struct flock lock_info;
2546 memset(&lock_info, 0, sizeof(lock_info));
2547 lock_info.l_type = F_WRLCK;
2548 lock_info.l_whence = SEEK_SET;
2549 lock_info.l_start = pid;
2550 lock_info.l_len = 1;
2551 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2552 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2554 } else if ((rc = ErrCode()) == EINTR) {
2562 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2563 * @param[in] txn the transaction handle to initialize
2564 * @return 0 on success, non-zero on failure.
2567 mdb_txn_renew0(MDB_txn *txn)
2569 MDB_env *env = txn->mt_env;
2570 MDB_txninfo *ti = env->me_txns;
2574 int rc, new_notls = 0;
2576 if (txn->mt_flags & MDB_TXN_RDONLY) {
2577 txn->mt_flags &= MDB_TXN_BEGIN_FLAGS;
2579 txn->mt_numdbs = env->me_numdbs;
2580 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2582 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2583 txn->mt_txnid = meta->mm_txnid;
2584 txn->mt_u.reader = NULL;
2586 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2587 pthread_getspecific(env->me_txkey);
2589 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2590 return MDB_BAD_RSLOT;
2592 MDB_PID_T pid = env->me_pid;
2593 MDB_THR_T tid = pthread_self();
2594 mdb_mutex_t *rmutex = MDB_MUTEX(env, r);
2596 if (!env->me_live_reader) {
2597 rc = mdb_reader_pid(env, Pidset, pid);
2600 env->me_live_reader = 1;
2603 if (LOCK_MUTEX(rc, env, rmutex))
2605 nr = ti->mti_numreaders;
2606 for (i=0; i<nr; i++)
2607 if (ti->mti_readers[i].mr_pid == 0)
2609 if (i == env->me_maxreaders) {
2610 UNLOCK_MUTEX(rmutex);
2611 return MDB_READERS_FULL;
2613 r = &ti->mti_readers[i];
2614 /* Claim the reader slot, carefully since other code
2615 * uses the reader table un-mutexed: First reset the
2616 * slot, next publish it in mti_numreaders. After
2617 * that, it is safe for mdb_env_close() to touch it.
2618 * When it will be closed, we can finally claim it.
2621 r->mr_txnid = (txnid_t)-1;
2624 ti->mti_numreaders = ++nr;
2625 env->me_close_readers = nr;
2627 UNLOCK_MUTEX(rmutex);
2629 new_notls = (env->me_flags & MDB_NOTLS);
2630 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2635 do /* LY: Retry on a race, ITS#7970. */
2636 r->mr_txnid = ti->mti_txnid;
2637 while(r->mr_txnid != ti->mti_txnid);
2638 txn->mt_txnid = r->mr_txnid;
2639 txn->mt_u.reader = r;
2640 meta = env->me_metas[txn->mt_txnid & 1];
2643 /* Not yet touching txn == env->me_txn0, it may be active */
2645 if (LOCK_MUTEX(rc, env, MDB_MUTEX(env, w)))
2647 txn->mt_txnid = ti->mti_txnid;
2648 meta = env->me_metas[txn->mt_txnid & 1];
2650 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2651 txn->mt_txnid = meta->mm_txnid;
2654 txn->mt_numdbs = env->me_numdbs;
2657 if (txn->mt_txnid == mdb_debug_start)
2661 txn->mt_child = NULL;
2662 txn->mt_loose_pgs = NULL;
2663 txn->mt_loose_count = 0;
2664 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2665 txn->mt_u.dirty_list = env->me_dirty_list;
2666 txn->mt_u.dirty_list[0].mid = 0;
2667 txn->mt_free_pgs = env->me_free_pgs;
2668 txn->mt_free_pgs[0] = 0;
2669 txn->mt_spill_pgs = NULL;
2671 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2674 /* Copy the DB info and flags */
2675 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2677 /* Moved to here to avoid a data race in read TXNs */
2678 txn->mt_next_pgno = meta->mm_last_pg+1;
2680 for (i=2; i<txn->mt_numdbs; i++) {
2681 x = env->me_dbflags[i];
2682 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2683 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2685 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2687 if (env->me_maxpg < txn->mt_next_pgno) {
2688 mdb_txn_reset0(txn, "renew0-mapfail");
2690 txn->mt_u.reader->mr_pid = 0;
2691 txn->mt_u.reader = NULL;
2693 return MDB_MAP_RESIZED;
2700 mdb_txn_renew(MDB_txn *txn)
2704 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2707 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2708 DPUTS("environment had fatal error, must shutdown!");
2712 rc = mdb_txn_renew0(txn);
2713 if (rc == MDB_SUCCESS) {
2714 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2715 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2716 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2722 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2726 int rc, size, tsize;
2728 flags &= MDB_TXN_BEGIN_FLAGS;
2729 flags |= env->me_flags & MDB_WRITEMAP;
2731 if (env->me_flags & MDB_FATAL_ERROR) {
2732 DPUTS("environment had fatal error, must shutdown!");
2735 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2738 size = tsize = sizeof(MDB_txn);
2740 /* Nested transactions: Max 1 child, write txns only, no writemap */
2741 flags |= parent->mt_flags;
2742 if (parent->mt_child ||
2743 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2745 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2747 /* Child txns save MDB_pgstate and use own copy of cursors */
2748 size = tsize = sizeof(MDB_ntxn);
2749 size += env->me_maxdbs * sizeof(MDB_cursor *);
2750 } else if (!(flags & MDB_RDONLY)) {
2751 /* Reuse preallocated write txn. However, do not touch it until
2752 * mdb_txn_renew0() succeeds, since it currently may be active.
2757 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2759 if ((txn = calloc(1, size)) == NULL) {
2760 DPRINTF(("calloc: %s", strerror(errno)));
2763 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2764 if (flags & MDB_RDONLY) {
2765 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2766 txn->mt_dbiseqs = env->me_dbiseqs;
2768 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2770 txn->mt_dbiseqs = parent->mt_dbiseqs;
2771 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2773 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2774 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2777 txn->mt_flags = flags;
2782 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2783 if (!txn->mt_u.dirty_list ||
2784 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2786 free(txn->mt_u.dirty_list);
2790 txn->mt_txnid = parent->mt_txnid;
2791 txn->mt_dirty_room = parent->mt_dirty_room;
2792 txn->mt_u.dirty_list[0].mid = 0;
2793 txn->mt_spill_pgs = NULL;
2794 txn->mt_next_pgno = parent->mt_next_pgno;
2795 parent->mt_child = txn;
2796 txn->mt_parent = parent;
2797 txn->mt_numdbs = parent->mt_numdbs;
2798 txn->mt_dbxs = parent->mt_dbxs;
2799 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2800 /* Copy parent's mt_dbflags, but clear DB_NEW */
2801 for (i=0; i<txn->mt_numdbs; i++)
2802 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2804 ntxn = (MDB_ntxn *)txn;
2805 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2806 if (env->me_pghead) {
2807 size = MDB_IDL_SIZEOF(env->me_pghead);
2808 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2810 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2815 rc = mdb_cursor_shadow(parent, txn);
2817 mdb_txn_reset0(txn, "beginchild-fail");
2820 rc = mdb_txn_renew0(txn);
2823 if (txn != env->me_txn0)
2826 txn->mt_flags |= flags; /* for txn==me_txn0, no effect otherwise */
2828 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2829 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2830 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2837 mdb_txn_env(MDB_txn *txn)
2839 if(!txn) return NULL;
2844 mdb_txn_id(MDB_txn *txn)
2847 return txn->mt_txnid;
2850 /** Export or close DBI handles opened in this txn. */
2852 mdb_dbis_update(MDB_txn *txn, int keep)
2855 MDB_dbi n = txn->mt_numdbs;
2856 MDB_env *env = txn->mt_env;
2857 unsigned char *tdbflags = txn->mt_dbflags;
2859 for (i = n; --i >= 2;) {
2860 if (tdbflags[i] & DB_NEW) {
2862 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2864 char *ptr = env->me_dbxs[i].md_name.mv_data;
2866 env->me_dbxs[i].md_name.mv_data = NULL;
2867 env->me_dbxs[i].md_name.mv_size = 0;
2868 env->me_dbflags[i] = 0;
2869 env->me_dbiseqs[i]++;
2875 if (keep && env->me_numdbs < n)
2879 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2880 * May be called twice for readonly txns: First reset it, then abort.
2881 * @param[in] txn the transaction handle to reset
2882 * @param[in] act why the transaction is being reset
2885 mdb_txn_reset0(MDB_txn *txn, const char *act)
2887 MDB_env *env = txn->mt_env;
2889 /* Close any DBI handles opened in this txn */
2890 mdb_dbis_update(txn, 0);
2892 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2893 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2894 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2896 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2897 if (txn->mt_u.reader) {
2898 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2899 if (!(env->me_flags & MDB_NOTLS))
2900 txn->mt_u.reader = NULL; /* txn does not own reader */
2902 txn->mt_numdbs = 0; /* close nothing if called again */
2903 txn->mt_dbxs = NULL; /* mark txn as reset */
2905 pgno_t *pghead = env->me_pghead;
2907 mdb_cursors_close(txn, 0);
2908 if (!(env->me_flags & MDB_WRITEMAP)) {
2909 mdb_dlist_free(txn);
2912 if (!txn->mt_parent) {
2913 if (mdb_midl_shrink(&txn->mt_free_pgs))
2914 env->me_free_pgs = txn->mt_free_pgs;
2916 env->me_pghead = NULL;
2920 /* The writer mutex was locked in mdb_txn_begin. */
2922 UNLOCK_MUTEX(MDB_MUTEX(env, w));
2924 txn->mt_parent->mt_child = NULL;
2925 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2926 mdb_midl_free(txn->mt_free_pgs);
2927 mdb_midl_free(txn->mt_spill_pgs);
2928 free(txn->mt_u.dirty_list);
2931 mdb_midl_free(pghead);
2936 mdb_txn_reset(MDB_txn *txn)
2941 /* This call is only valid for read-only txns */
2942 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2945 mdb_txn_reset0(txn, "reset");
2949 mdb_txn_abort(MDB_txn *txn)
2955 mdb_txn_abort(txn->mt_child);
2957 mdb_txn_reset0(txn, "abort");
2958 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2959 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2960 txn->mt_u.reader->mr_pid = 0;
2962 if (txn != txn->mt_env->me_txn0)
2966 /** Save the freelist as of this transaction to the freeDB.
2967 * This changes the freelist. Keep trying until it stabilizes.
2970 mdb_freelist_save(MDB_txn *txn)
2972 /* env->me_pghead[] can grow and shrink during this call.
2973 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2974 * Page numbers cannot disappear from txn->mt_free_pgs[].
2977 MDB_env *env = txn->mt_env;
2978 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2979 txnid_t pglast = 0, head_id = 0;
2980 pgno_t freecnt = 0, *free_pgs, *mop;
2981 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2983 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2985 if (env->me_pghead) {
2986 /* Make sure first page of freeDB is touched and on freelist */
2987 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2988 if (rc && rc != MDB_NOTFOUND)
2992 if (!env->me_pghead && txn->mt_loose_pgs) {
2993 /* Put loose page numbers in mt_free_pgs, since
2994 * we may be unable to return them to me_pghead.
2996 MDB_page *mp = txn->mt_loose_pgs;
2997 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2999 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3000 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3001 txn->mt_loose_pgs = NULL;
3002 txn->mt_loose_count = 0;
3005 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3006 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3007 ? SSIZE_MAX : maxfree_1pg;
3010 /* Come back here after each Put() in case freelist changed */
3015 /* If using records from freeDB which we have not yet
3016 * deleted, delete them and any we reserved for me_pghead.
3018 while (pglast < env->me_pglast) {
3019 rc = mdb_cursor_first(&mc, &key, NULL);
3022 pglast = head_id = *(txnid_t *)key.mv_data;
3023 total_room = head_room = 0;
3024 mdb_tassert(txn, pglast <= env->me_pglast);
3025 rc = mdb_cursor_del(&mc, 0);
3030 /* Save the IDL of pages freed by this txn, to a single record */
3031 if (freecnt < txn->mt_free_pgs[0]) {
3033 /* Make sure last page of freeDB is touched and on freelist */
3034 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3035 if (rc && rc != MDB_NOTFOUND)
3038 free_pgs = txn->mt_free_pgs;
3039 /* Write to last page of freeDB */
3040 key.mv_size = sizeof(txn->mt_txnid);
3041 key.mv_data = &txn->mt_txnid;
3043 freecnt = free_pgs[0];
3044 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3045 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3048 /* Retry if mt_free_pgs[] grew during the Put() */
3049 free_pgs = txn->mt_free_pgs;
3050 } while (freecnt < free_pgs[0]);
3051 mdb_midl_sort(free_pgs);
3052 memcpy(data.mv_data, free_pgs, data.mv_size);
3055 unsigned int i = free_pgs[0];
3056 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3057 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3059 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3065 mop = env->me_pghead;
3066 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3068 /* Reserve records for me_pghead[]. Split it if multi-page,
3069 * to avoid searching freeDB for a page range. Use keys in
3070 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3072 if (total_room >= mop_len) {
3073 if (total_room == mop_len || --more < 0)
3075 } else if (head_room >= maxfree_1pg && head_id > 1) {
3076 /* Keep current record (overflow page), add a new one */
3080 /* (Re)write {key = head_id, IDL length = head_room} */
3081 total_room -= head_room;
3082 head_room = mop_len - total_room;
3083 if (head_room > maxfree_1pg && head_id > 1) {
3084 /* Overflow multi-page for part of me_pghead */
3085 head_room /= head_id; /* amortize page sizes */
3086 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3087 } else if (head_room < 0) {
3088 /* Rare case, not bothering to delete this record */
3091 key.mv_size = sizeof(head_id);
3092 key.mv_data = &head_id;
3093 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3094 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3097 /* IDL is initially empty, zero out at least the length */
3098 pgs = (pgno_t *)data.mv_data;
3099 j = head_room > clean_limit ? head_room : 0;
3103 total_room += head_room;
3106 /* Return loose page numbers to me_pghead, though usually none are
3107 * left at this point. The pages themselves remain in dirty_list.
3109 if (txn->mt_loose_pgs) {
3110 MDB_page *mp = txn->mt_loose_pgs;
3111 unsigned count = txn->mt_loose_count;
3113 /* Room for loose pages + temp IDL with same */
3114 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3116 mop = env->me_pghead;
3117 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3118 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3119 loose[ ++count ] = mp->mp_pgno;
3121 mdb_midl_sort(loose);
3122 mdb_midl_xmerge(mop, loose);
3123 txn->mt_loose_pgs = NULL;
3124 txn->mt_loose_count = 0;
3128 /* Fill in the reserved me_pghead records */
3134 rc = mdb_cursor_first(&mc, &key, &data);
3135 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3136 txnid_t id = *(txnid_t *)key.mv_data;
3137 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3140 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3142 if (len > mop_len) {
3144 data.mv_size = (len + 1) * sizeof(MDB_ID);
3146 data.mv_data = mop -= len;
3149 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3151 if (rc || !(mop_len -= len))
3158 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3159 * @param[in] txn the transaction that's being committed
3160 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3161 * @return 0 on success, non-zero on failure.
3164 mdb_page_flush(MDB_txn *txn, int keep)
3166 MDB_env *env = txn->mt_env;
3167 MDB_ID2L dl = txn->mt_u.dirty_list;
3168 unsigned psize = env->me_psize, j;
3169 int i, pagecount = dl[0].mid, rc;
3170 size_t size = 0, pos = 0;
3172 MDB_page *dp = NULL;
3176 struct iovec iov[MDB_COMMIT_PAGES];
3177 ssize_t wpos = 0, wsize = 0, wres;
3178 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3184 if (env->me_flags & MDB_WRITEMAP) {
3185 /* Clear dirty flags */
3186 while (++i <= pagecount) {
3188 /* Don't flush this page yet */
3189 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3190 dp->mp_flags &= ~P_KEEP;
3194 dp->mp_flags &= ~P_DIRTY;
3199 /* Write the pages */
3201 if (++i <= pagecount) {
3203 /* Don't flush this page yet */
3204 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3205 dp->mp_flags &= ~P_KEEP;
3210 /* clear dirty flag */
3211 dp->mp_flags &= ~P_DIRTY;
3214 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3219 /* Windows actually supports scatter/gather I/O, but only on
3220 * unbuffered file handles. Since we're relying on the OS page
3221 * cache for all our data, that's self-defeating. So we just
3222 * write pages one at a time. We use the ov structure to set
3223 * the write offset, to at least save the overhead of a Seek
3226 DPRINTF(("committing page %"Z"u", pgno));
3227 memset(&ov, 0, sizeof(ov));
3228 ov.Offset = pos & 0xffffffff;
3229 ov.OffsetHigh = pos >> 16 >> 16;
3230 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3232 DPRINTF(("WriteFile: %d", rc));
3236 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3237 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3239 /* Write previous page(s) */
3240 #ifdef MDB_USE_PWRITEV
3241 wres = pwritev(env->me_fd, iov, n, wpos);
3244 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3246 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3248 DPRINTF(("lseek: %s", strerror(rc)));
3251 wres = writev(env->me_fd, iov, n);
3254 if (wres != wsize) {
3257 DPRINTF(("Write error: %s", strerror(rc)));
3259 rc = EIO; /* TODO: Use which error code? */
3260 DPUTS("short write, filesystem full?");
3271 DPRINTF(("committing page %"Z"u", pgno));
3272 next_pos = pos + size;
3273 iov[n].iov_len = size;
3274 iov[n].iov_base = (char *)dp;
3280 /* MIPS has cache coherency issues, this is a no-op everywhere else
3281 * Note: for any size >= on-chip cache size, entire on-chip cache is
3284 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3286 for (i = keep; ++i <= pagecount; ) {
3288 /* This is a page we skipped above */
3291 dl[j].mid = dp->mp_pgno;
3294 mdb_dpage_free(env, dp);
3299 txn->mt_dirty_room += i - j;
3305 mdb_txn_commit(MDB_txn *txn)
3311 if (txn == NULL || txn->mt_env == NULL)
3314 if (txn->mt_child) {
3315 rc = mdb_txn_commit(txn->mt_child);
3316 txn->mt_child = NULL;
3323 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3324 mdb_dbis_update(txn, 1);
3325 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3330 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3331 DPUTS("error flag is set, can't commit");
3333 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3338 if (txn->mt_parent) {
3339 MDB_txn *parent = txn->mt_parent;
3343 unsigned x, y, len, ps_len;
3345 /* Append our free list to parent's */
3346 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3349 mdb_midl_free(txn->mt_free_pgs);
3350 /* Failures after this must either undo the changes
3351 * to the parent or set MDB_TXN_ERROR in the parent.
3354 parent->mt_next_pgno = txn->mt_next_pgno;
3355 parent->mt_flags = txn->mt_flags;
3357 /* Merge our cursors into parent's and close them */
3358 mdb_cursors_close(txn, 1);
3360 /* Update parent's DB table. */
3361 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3362 parent->mt_numdbs = txn->mt_numdbs;
3363 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3364 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3365 for (i=2; i<txn->mt_numdbs; i++) {
3366 /* preserve parent's DB_NEW status */
3367 x = parent->mt_dbflags[i] & DB_NEW;
3368 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3371 dst = parent->mt_u.dirty_list;
3372 src = txn->mt_u.dirty_list;
3373 /* Remove anything in our dirty list from parent's spill list */
3374 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3376 pspill[0] = (pgno_t)-1;
3377 /* Mark our dirty pages as deleted in parent spill list */
3378 for (i=0, len=src[0].mid; ++i <= len; ) {
3379 MDB_ID pn = src[i].mid << 1;
3380 while (pn > pspill[x])
3382 if (pn == pspill[x]) {
3387 /* Squash deleted pagenums if we deleted any */
3388 for (x=y; ++x <= ps_len; )
3389 if (!(pspill[x] & 1))
3390 pspill[++y] = pspill[x];
3394 /* Find len = length of merging our dirty list with parent's */
3396 dst[0].mid = 0; /* simplify loops */
3397 if (parent->mt_parent) {
3398 len = x + src[0].mid;
3399 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3400 for (i = x; y && i; y--) {
3401 pgno_t yp = src[y].mid;
3402 while (yp < dst[i].mid)
3404 if (yp == dst[i].mid) {
3409 } else { /* Simplify the above for single-ancestor case */
3410 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3412 /* Merge our dirty list with parent's */
3414 for (i = len; y; dst[i--] = src[y--]) {
3415 pgno_t yp = src[y].mid;
3416 while (yp < dst[x].mid)
3417 dst[i--] = dst[x--];
3418 if (yp == dst[x].mid)
3419 free(dst[x--].mptr);
3421 mdb_tassert(txn, i == x);
3423 free(txn->mt_u.dirty_list);
3424 parent->mt_dirty_room = txn->mt_dirty_room;
3425 if (txn->mt_spill_pgs) {
3426 if (parent->mt_spill_pgs) {
3427 /* TODO: Prevent failure here, so parent does not fail */
3428 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3430 parent->mt_flags |= MDB_TXN_ERROR;
3431 mdb_midl_free(txn->mt_spill_pgs);
3432 mdb_midl_sort(parent->mt_spill_pgs);
3434 parent->mt_spill_pgs = txn->mt_spill_pgs;
3438 /* Append our loose page list to parent's */
3439 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3441 *lp = txn->mt_loose_pgs;
3442 parent->mt_loose_count += txn->mt_loose_count;
3444 parent->mt_child = NULL;
3445 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3450 if (txn != env->me_txn) {
3451 DPUTS("attempt to commit unknown transaction");
3456 mdb_cursors_close(txn, 0);
3458 if (!txn->mt_u.dirty_list[0].mid &&
3459 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3462 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3463 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3465 /* Update DB root pointers */
3466 if (txn->mt_numdbs > 2) {
3470 data.mv_size = sizeof(MDB_db);
3472 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3473 for (i = 2; i < txn->mt_numdbs; i++) {
3474 if (txn->mt_dbflags[i] & DB_DIRTY) {
3475 if (TXN_DBI_CHANGED(txn, i)) {
3479 data.mv_data = &txn->mt_dbs[i];
3480 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3487 rc = mdb_freelist_save(txn);
3491 mdb_midl_free(env->me_pghead);
3492 env->me_pghead = NULL;
3493 if (mdb_midl_shrink(&txn->mt_free_pgs))
3494 env->me_free_pgs = txn->mt_free_pgs;
3500 if ((rc = mdb_page_flush(txn, 0)))
3502 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3503 (rc = mdb_env_sync(env, 0)))
3505 if ((rc = mdb_env_write_meta(txn)))
3508 /* Free P_LOOSE pages left behind in dirty_list */
3509 if (!(env->me_flags & MDB_WRITEMAP))
3510 mdb_dlist_free(txn);
3515 mdb_dbis_update(txn, 1);
3518 UNLOCK_MUTEX(MDB_MUTEX(env, w));
3519 if (txn != env->me_txn0)
3529 /** Read the environment parameters of a DB environment before
3530 * mapping it into memory.
3531 * @param[in] env the environment handle
3532 * @param[out] meta address of where to store the meta information
3533 * @return 0 on success, non-zero on failure.
3536 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3542 enum { Size = sizeof(pbuf) };
3544 /* We don't know the page size yet, so use a minimum value.
3545 * Read both meta pages so we can use the latest one.
3548 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3552 memset(&ov, 0, sizeof(ov));
3554 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3555 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3558 rc = pread(env->me_fd, &pbuf, Size, off);
3561 if (rc == 0 && off == 0)
3563 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3564 DPRINTF(("read: %s", mdb_strerror(rc)));
3568 p = (MDB_page *)&pbuf;
3570 if (!F_ISSET(p->mp_flags, P_META)) {
3571 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3576 if (m->mm_magic != MDB_MAGIC) {
3577 DPUTS("meta has invalid magic");
3581 if (m->mm_version != MDB_DATA_VERSION) {
3582 DPRINTF(("database is version %u, expected version %u",
3583 m->mm_version, MDB_DATA_VERSION));
3584 return MDB_VERSION_MISMATCH;
3587 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3593 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3595 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3597 meta->mm_magic = MDB_MAGIC;
3598 meta->mm_version = MDB_DATA_VERSION;
3599 meta->mm_mapsize = env->me_mapsize;
3600 meta->mm_psize = env->me_psize;
3601 meta->mm_last_pg = 1;
3602 meta->mm_flags = env->me_flags & 0xffff;
3603 meta->mm_flags |= MDB_INTEGERKEY;
3604 meta->mm_dbs[0].md_root = P_INVALID;
3605 meta->mm_dbs[1].md_root = P_INVALID;
3608 /** Write the environment parameters of a freshly created DB environment.
3609 * @param[in] env the environment handle
3610 * @param[in] meta the #MDB_meta to write
3611 * @return 0 on success, non-zero on failure.
3614 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3622 memset(&ov, 0, sizeof(ov));
3623 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3625 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3628 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3629 len = pwrite(fd, ptr, size, pos); \
3630 rc = (len >= 0); } while(0)
3633 DPUTS("writing new meta page");
3635 psize = env->me_psize;
3637 p = calloc(2, psize);
3639 p->mp_flags = P_META;
3640 *(MDB_meta *)METADATA(p) = *meta;
3642 q = (MDB_page *)((char *)p + psize);
3644 q->mp_flags = P_META;
3645 *(MDB_meta *)METADATA(q) = *meta;
3647 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3650 else if ((unsigned) len == psize * 2)
3658 /** Update the environment info to commit a transaction.
3659 * @param[in] txn the transaction that's being committed
3660 * @return 0 on success, non-zero on failure.
3663 mdb_env_write_meta(MDB_txn *txn)
3666 MDB_meta meta, metab, *mp;
3670 int rc, len, toggle;
3679 toggle = txn->mt_txnid & 1;
3680 DPRINTF(("writing meta page %d for root page %"Z"u",
3681 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3684 flags = txn->mt_flags & env->me_flags;
3685 mp = env->me_metas[toggle];
3686 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3687 /* Persist any increases of mapsize config */
3688 if (mapsize < env->me_mapsize)
3689 mapsize = env->me_mapsize;
3691 if (flags & MDB_WRITEMAP) {
3692 mp->mm_mapsize = mapsize;
3693 mp->mm_dbs[0] = txn->mt_dbs[0];
3694 mp->mm_dbs[1] = txn->mt_dbs[1];
3695 mp->mm_last_pg = txn->mt_next_pgno - 1;
3696 #if !(defined(_MSC_VER) || defined(__i386__) || defined(__x86_64__))
3697 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3698 __sync_synchronize();
3700 mp->mm_txnid = txn->mt_txnid;
3701 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3702 unsigned meta_size = env->me_psize;
3703 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3706 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3707 if (meta_size < env->me_os_psize)
3708 meta_size += meta_size;
3713 if (MDB_MSYNC(ptr, meta_size, rc)) {
3720 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3721 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3723 meta.mm_mapsize = mapsize;
3724 meta.mm_dbs[0] = txn->mt_dbs[0];
3725 meta.mm_dbs[1] = txn->mt_dbs[1];
3726 meta.mm_last_pg = txn->mt_next_pgno - 1;
3727 meta.mm_txnid = txn->mt_txnid;
3729 off = offsetof(MDB_meta, mm_mapsize);
3730 ptr = (char *)&meta + off;
3731 len = sizeof(MDB_meta) - off;
3733 off += env->me_psize;
3736 /* Write to the SYNC fd */
3737 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3740 memset(&ov, 0, sizeof(ov));
3742 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3746 rc = pwrite(mfd, ptr, len, off);
3749 rc = rc < 0 ? ErrCode() : EIO;
3750 DPUTS("write failed, disk error?");
3751 /* On a failure, the pagecache still contains the new data.
3752 * Write some old data back, to prevent it from being used.
3753 * Use the non-SYNC fd; we know it will fail anyway.
3755 meta.mm_last_pg = metab.mm_last_pg;
3756 meta.mm_txnid = metab.mm_txnid;
3758 memset(&ov, 0, sizeof(ov));
3760 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3762 r2 = pwrite(env->me_fd, ptr, len, off);
3763 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3766 env->me_flags |= MDB_FATAL_ERROR;
3769 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3770 CACHEFLUSH(env->me_map + off, len, DCACHE);
3772 /* Memory ordering issues are irrelevant; since the entire writer
3773 * is wrapped by wmutex, all of these changes will become visible
3774 * after the wmutex is unlocked. Since the DB is multi-version,
3775 * readers will get consistent data regardless of how fresh or
3776 * how stale their view of these values is.
3779 env->me_txns->mti_txnid = txn->mt_txnid;
3784 /** Check both meta pages to see which one is newer.
3785 * @param[in] env the environment handle
3786 * @return meta toggle (0 or 1).
3789 mdb_env_pick_meta(const MDB_env *env)
3791 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3795 mdb_env_create(MDB_env **env)
3799 e = calloc(1, sizeof(MDB_env));
3803 e->me_maxreaders = DEFAULT_READERS;
3804 e->me_maxdbs = e->me_numdbs = 2;
3805 e->me_fd = INVALID_HANDLE_VALUE;
3806 e->me_lfd = INVALID_HANDLE_VALUE;
3807 e->me_mfd = INVALID_HANDLE_VALUE;
3808 #ifdef MDB_USE_SYSV_SEM
3809 e->me_rmutex.semid = -1;
3810 e->me_wmutex.semid = -1;
3812 e->me_pid = getpid();
3813 GET_PAGESIZE(e->me_os_psize);
3814 VGMEMP_CREATE(e,0,0);
3820 mdb_env_map(MDB_env *env, void *addr)
3823 unsigned int flags = env->me_flags;
3827 LONG sizelo, sizehi;
3830 if (flags & MDB_RDONLY) {
3831 /* Don't set explicit map size, use whatever exists */
3836 msize = env->me_mapsize;
3837 sizelo = msize & 0xffffffff;
3838 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3840 /* Windows won't create mappings for zero length files.
3841 * and won't map more than the file size.
3842 * Just set the maxsize right now.
3844 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3845 || !SetEndOfFile(env->me_fd)
3846 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3850 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3851 PAGE_READWRITE : PAGE_READONLY,
3852 sizehi, sizelo, NULL);
3855 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3856 FILE_MAP_WRITE : FILE_MAP_READ,
3858 rc = env->me_map ? 0 : ErrCode();
3863 int prot = PROT_READ;
3864 if (flags & MDB_WRITEMAP) {
3866 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3869 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3871 if (env->me_map == MAP_FAILED) {
3876 if (flags & MDB_NORDAHEAD) {
3877 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3879 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3881 #ifdef POSIX_MADV_RANDOM
3882 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3883 #endif /* POSIX_MADV_RANDOM */
3884 #endif /* MADV_RANDOM */
3888 /* Can happen because the address argument to mmap() is just a
3889 * hint. mmap() can pick another, e.g. if the range is in use.
3890 * The MAP_FIXED flag would prevent that, but then mmap could
3891 * instead unmap existing pages to make room for the new map.
3893 if (addr && env->me_map != addr)
3894 return EBUSY; /* TODO: Make a new MDB_* error code? */
3896 p = (MDB_page *)env->me_map;
3897 env->me_metas[0] = METADATA(p);
3898 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3904 mdb_env_set_mapsize(MDB_env *env, size_t size)
3906 /* If env is already open, caller is responsible for making
3907 * sure there are no active txns.
3915 meta = env->me_metas[mdb_env_pick_meta(env)];
3917 size = meta->mm_mapsize;
3919 /* Silently round up to minimum if the size is too small */
3920 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3924 munmap(env->me_map, env->me_mapsize);
3925 env->me_mapsize = size;
3926 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3927 rc = mdb_env_map(env, old);
3931 env->me_mapsize = size;
3933 env->me_maxpg = env->me_mapsize / env->me_psize;
3938 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3942 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3947 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3949 if (env->me_map || readers < 1)
3951 env->me_maxreaders = readers;
3956 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3958 if (!env || !readers)
3960 *readers = env->me_maxreaders;
3965 mdb_fsize(HANDLE fd, size_t *size)
3968 LARGE_INTEGER fsize;
3970 if (!GetFileSizeEx(fd, &fsize))
3973 *size = fsize.QuadPart;
3985 #ifdef BROKEN_FDATASYNC
3986 #include <sys/utsname.h>
3987 #include <sys/vfs.h>
3990 /** Further setup required for opening an LMDB environment
3993 mdb_env_open2(MDB_env *env)
3995 unsigned int flags = env->me_flags;
3996 int i, newenv = 0, rc;
4000 /* See if we should use QueryLimited */
4002 if ((rc & 0xff) > 5)
4003 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4005 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4008 #ifdef BROKEN_FDATASYNC
4009 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4010 * https://lkml.org/lkml/2012/9/3/83
4011 * Kernels after 3.6-rc6 are known good.
4012 * https://lkml.org/lkml/2012/9/10/556
4013 * See if the DB is on ext3/ext4, then check for new enough kernel
4014 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4019 fstatfs(env->me_fd, &st);
4020 while (st.f_type == 0xEF53) {
4024 if (uts.release[0] < '3') {
4025 if (!strncmp(uts.release, "2.6.32.", 7)) {
4026 i = atoi(uts.release+7);
4028 break; /* 2.6.32.60 and newer is OK */
4029 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4030 i = atoi(uts.release+7);
4032 break; /* 2.6.34.15 and newer is OK */
4034 } else if (uts.release[0] == '3') {
4035 i = atoi(uts.release+2);
4037 break; /* 3.6 and newer is OK */
4039 i = atoi(uts.release+4);
4041 break; /* 3.5.4 and newer is OK */
4042 } else if (i == 2) {
4043 i = atoi(uts.release+4);
4045 break; /* 3.2.30 and newer is OK */
4047 } else { /* 4.x and newer is OK */
4050 env->me_flags |= MDB_FSYNCONLY;
4056 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4059 DPUTS("new mdbenv");
4061 env->me_psize = env->me_os_psize;
4062 if (env->me_psize > MAX_PAGESIZE)
4063 env->me_psize = MAX_PAGESIZE;
4064 memset(&meta, 0, sizeof(meta));
4065 mdb_env_init_meta0(env, &meta);
4066 meta.mm_mapsize = DEFAULT_MAPSIZE;
4068 env->me_psize = meta.mm_psize;
4071 /* Was a mapsize configured? */
4072 if (!env->me_mapsize) {
4073 env->me_mapsize = meta.mm_mapsize;
4076 /* Make sure mapsize >= committed data size. Even when using
4077 * mm_mapsize, which could be broken in old files (ITS#7789).
4079 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4080 if (env->me_mapsize < minsize)
4081 env->me_mapsize = minsize;
4083 meta.mm_mapsize = env->me_mapsize;
4085 if (newenv && !(flags & MDB_FIXEDMAP)) {
4086 /* mdb_env_map() may grow the datafile. Write the metapages
4087 * first, so the file will be valid if initialization fails.
4088 * Except with FIXEDMAP, since we do not yet know mm_address.
4089 * We could fill in mm_address later, but then a different
4090 * program might end up doing that - one with a memory layout
4091 * and map address which does not suit the main program.
4093 rc = mdb_env_init_meta(env, &meta);
4099 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4104 if (flags & MDB_FIXEDMAP)
4105 meta.mm_address = env->me_map;
4106 i = mdb_env_init_meta(env, &meta);
4107 if (i != MDB_SUCCESS) {
4112 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4113 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4115 #if !(MDB_MAXKEYSIZE)
4116 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4118 env->me_maxpg = env->me_mapsize / env->me_psize;
4122 int toggle = mdb_env_pick_meta(env);
4123 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4125 DPRINTF(("opened database version %u, pagesize %u",
4126 env->me_metas[0]->mm_version, env->me_psize));
4127 DPRINTF(("using meta page %d", toggle));
4128 DPRINTF(("depth: %u", db->md_depth));
4129 DPRINTF(("entries: %"Z"u", db->md_entries));
4130 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4131 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4132 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4133 DPRINTF(("root: %"Z"u", db->md_root));
4141 /** Release a reader thread's slot in the reader lock table.
4142 * This function is called automatically when a thread exits.
4143 * @param[in] ptr This points to the slot in the reader lock table.
4146 mdb_env_reader_dest(void *ptr)
4148 MDB_reader *reader = ptr;
4154 /** Junk for arranging thread-specific callbacks on Windows. This is
4155 * necessarily platform and compiler-specific. Windows supports up
4156 * to 1088 keys. Let's assume nobody opens more than 64 environments
4157 * in a single process, for now. They can override this if needed.
4159 #ifndef MAX_TLS_KEYS
4160 #define MAX_TLS_KEYS 64
4162 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4163 static int mdb_tls_nkeys;
4165 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4169 case DLL_PROCESS_ATTACH: break;
4170 case DLL_THREAD_ATTACH: break;
4171 case DLL_THREAD_DETACH:
4172 for (i=0; i<mdb_tls_nkeys; i++) {
4173 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4175 mdb_env_reader_dest(r);
4179 case DLL_PROCESS_DETACH: break;
4184 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4186 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4190 /* Force some symbol references.
4191 * _tls_used forces the linker to create the TLS directory if not already done
4192 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4194 #pragma comment(linker, "/INCLUDE:_tls_used")
4195 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4196 #pragma const_seg(".CRT$XLB")
4197 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4198 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4201 #pragma comment(linker, "/INCLUDE:__tls_used")
4202 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4203 #pragma data_seg(".CRT$XLB")
4204 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4206 #endif /* WIN 32/64 */
4207 #endif /* !__GNUC__ */
4210 /** Downgrade the exclusive lock on the region back to shared */
4212 mdb_env_share_locks(MDB_env *env, int *excl)
4214 int rc = 0, toggle = mdb_env_pick_meta(env);
4216 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4221 /* First acquire a shared lock. The Unlock will
4222 * then release the existing exclusive lock.
4224 memset(&ov, 0, sizeof(ov));
4225 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4228 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4234 struct flock lock_info;
4235 /* The shared lock replaces the existing lock */
4236 memset((void *)&lock_info, 0, sizeof(lock_info));
4237 lock_info.l_type = F_RDLCK;
4238 lock_info.l_whence = SEEK_SET;
4239 lock_info.l_start = 0;
4240 lock_info.l_len = 1;
4241 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4242 (rc = ErrCode()) == EINTR) ;
4243 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4250 /** Try to get exclusive lock, otherwise shared.
4251 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4254 mdb_env_excl_lock(MDB_env *env, int *excl)
4258 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4262 memset(&ov, 0, sizeof(ov));
4263 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4270 struct flock lock_info;
4271 memset((void *)&lock_info, 0, sizeof(lock_info));
4272 lock_info.l_type = F_WRLCK;
4273 lock_info.l_whence = SEEK_SET;
4274 lock_info.l_start = 0;
4275 lock_info.l_len = 1;
4276 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4277 (rc = ErrCode()) == EINTR) ;
4281 # ifdef MDB_USE_SYSV_SEM
4282 if (*excl < 0) /* always true when !MDB_USE_SYSV_SEM */
4285 lock_info.l_type = F_RDLCK;
4286 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4287 (rc = ErrCode()) == EINTR) ;
4297 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4299 * @(#) $Revision: 5.1 $
4300 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4301 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4303 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4307 * Please do not copyright this code. This code is in the public domain.
4309 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4310 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4311 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4312 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4313 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4314 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4315 * PERFORMANCE OF THIS SOFTWARE.
4318 * chongo <Landon Curt Noll> /\oo/\
4319 * http://www.isthe.com/chongo/
4321 * Share and Enjoy! :-)
4324 typedef unsigned long long mdb_hash_t;
4325 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4327 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4328 * @param[in] val value to hash
4329 * @param[in] hval initial value for hash
4330 * @return 64 bit hash
4332 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4333 * hval arg on the first call.
4336 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4338 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4339 unsigned char *end = s + val->mv_size;
4341 * FNV-1a hash each octet of the string
4344 /* xor the bottom with the current octet */
4345 hval ^= (mdb_hash_t)*s++;
4347 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4348 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4349 (hval << 7) + (hval << 8) + (hval << 40);
4351 /* return our new hash value */
4355 /** Hash the string and output the encoded hash.
4356 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4357 * very short name limits. We don't care about the encoding being reversible,
4358 * we just want to preserve as many bits of the input as possible in a
4359 * small printable string.
4360 * @param[in] str string to hash
4361 * @param[out] encbuf an array of 11 chars to hold the hash
4363 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4366 mdb_pack85(unsigned long l, char *out)
4370 for (i=0; i<5; i++) {
4371 *out++ = mdb_a85[l % 85];
4377 mdb_hash_enc(MDB_val *val, char *encbuf)
4379 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4381 mdb_pack85(h, encbuf);
4382 mdb_pack85(h>>32, encbuf+5);
4387 /** Open and/or initialize the lock region for the environment.
4388 * @param[in] env The LMDB environment.
4389 * @param[in] lpath The pathname of the file used for the lock region.
4390 * @param[in] mode The Unix permissions for the file, if we create it.
4391 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4392 * @return 0 on success, non-zero on failure.
4395 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4398 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4400 # define MDB_ERRCODE_ROFS EROFS
4401 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4402 # define MDB_CLOEXEC O_CLOEXEC
4405 # define MDB_CLOEXEC 0
4408 #ifdef MDB_USE_SYSV_SEM
4416 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4417 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4418 FILE_ATTRIBUTE_NORMAL, NULL);
4420 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4422 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4424 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4429 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4430 /* Lose record locks when exec*() */
4431 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4432 fcntl(env->me_lfd, F_SETFD, fdflags);
4435 if (!(env->me_flags & MDB_NOTLS)) {
4436 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4439 env->me_flags |= MDB_ENV_TXKEY;
4441 /* Windows TLS callbacks need help finding their TLS info. */
4442 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4446 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4450 /* Try to get exclusive lock. If we succeed, then
4451 * nobody is using the lock region and we should initialize it.
4453 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4456 size = GetFileSize(env->me_lfd, NULL);
4458 size = lseek(env->me_lfd, 0, SEEK_END);
4459 if (size == -1) goto fail_errno;
4461 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4462 if (size < rsize && *excl > 0) {
4464 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4465 || !SetEndOfFile(env->me_lfd))
4468 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4472 size = rsize - sizeof(MDB_txninfo);
4473 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4478 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4480 if (!mh) goto fail_errno;
4481 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4483 if (!env->me_txns) goto fail_errno;
4485 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4487 if (m == MAP_FAILED) goto fail_errno;
4493 BY_HANDLE_FILE_INFORMATION stbuf;
4502 if (!mdb_sec_inited) {
4503 InitializeSecurityDescriptor(&mdb_null_sd,
4504 SECURITY_DESCRIPTOR_REVISION);
4505 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4506 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4507 mdb_all_sa.bInheritHandle = FALSE;
4508 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4511 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4512 idbuf.volume = stbuf.dwVolumeSerialNumber;
4513 idbuf.nhigh = stbuf.nFileIndexHigh;
4514 idbuf.nlow = stbuf.nFileIndexLow;
4515 val.mv_data = &idbuf;
4516 val.mv_size = sizeof(idbuf);
4517 mdb_hash_enc(&val, encbuf);
4518 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4519 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4520 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4521 if (!env->me_rmutex) goto fail_errno;
4522 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4523 if (!env->me_wmutex) goto fail_errno;
4524 #elif defined(MDB_USE_SYSV_SEM)
4525 unsigned short vals[2] = {1, 1};
4526 semid = semget(IPC_PRIVATE, 2, mode);
4530 if (semctl(semid, 0, SETALL, semu) < 0)
4532 env->me_txns->mti_semid = semid;
4533 #else /* MDB_USE_SYSV_SEM */
4534 pthread_mutexattr_t mattr;
4536 if ((rc = pthread_mutexattr_init(&mattr))
4537 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4538 #ifdef MDB_ROBUST_SUPPORTED
4539 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4541 || (rc = pthread_mutex_init(&env->me_txns->mti_rmutex, &mattr))
4542 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4544 pthread_mutexattr_destroy(&mattr);
4545 #endif /* _WIN32 || MDB_USE_SYSV_SEM */
4547 env->me_txns->mti_magic = MDB_MAGIC;
4548 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4549 env->me_txns->mti_txnid = 0;
4550 env->me_txns->mti_numreaders = 0;
4553 #ifdef MDB_USE_SYSV_SEM
4554 struct semid_ds buf;
4556 if (env->me_txns->mti_magic != MDB_MAGIC) {
4557 DPUTS("lock region has invalid magic");
4561 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4562 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4563 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4564 rc = MDB_VERSION_MISMATCH;
4568 if (rc && rc != EACCES && rc != EAGAIN) {
4572 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4573 if (!env->me_rmutex) goto fail_errno;
4574 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4575 if (!env->me_wmutex) goto fail_errno;
4576 #elif defined(MDB_USE_SYSV_SEM)
4577 semid = env->me_txns->mti_semid;
4579 /* check for read access */
4580 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4582 /* check for write access */
4583 if (semctl(semid, 0, IPC_SET, semu) < 0)
4587 #ifdef MDB_USE_SYSV_SEM
4588 env->me_rmutex.semid = semid;
4589 env->me_wmutex.semid = semid;
4590 env->me_rmutex.semnum = 0;
4591 env->me_wmutex.semnum = 1;
4592 env->me_rmutex.locked = &env->me_txns->mti_rlocked;
4593 env->me_wmutex.locked = &env->me_txns->mti_wlocked;
4604 /** The name of the lock file in the DB environment */
4605 #define LOCKNAME "/lock.mdb"
4606 /** The name of the data file in the DB environment */
4607 #define DATANAME "/data.mdb"
4608 /** The suffix of the lock file when no subdir is used */
4609 #define LOCKSUFF "-lock"
4610 /** Only a subset of the @ref mdb_env flags can be changed
4611 * at runtime. Changing other flags requires closing the
4612 * environment and re-opening it with the new flags.
4614 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4615 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4616 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4618 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4619 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4623 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4625 int oflags, rc, len, excl = -1;
4626 char *lpath, *dpath;
4628 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4632 if (flags & MDB_NOSUBDIR) {
4633 rc = len + sizeof(LOCKSUFF) + len + 1;
4635 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4640 if (flags & MDB_NOSUBDIR) {
4641 dpath = lpath + len + sizeof(LOCKSUFF);
4642 sprintf(lpath, "%s" LOCKSUFF, path);
4643 strcpy(dpath, path);
4645 dpath = lpath + len + sizeof(LOCKNAME);
4646 sprintf(lpath, "%s" LOCKNAME, path);
4647 sprintf(dpath, "%s" DATANAME, path);
4651 flags |= env->me_flags;
4652 if (flags & MDB_RDONLY) {
4653 /* silently ignore WRITEMAP when we're only getting read access */
4654 flags &= ~MDB_WRITEMAP;
4656 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4657 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4660 env->me_flags = flags |= MDB_ENV_ACTIVE;
4664 env->me_path = strdup(path);
4665 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4666 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4667 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4668 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4673 /* For RDONLY, get lockfile after we know datafile exists */
4674 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4675 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4681 if (F_ISSET(flags, MDB_RDONLY)) {
4682 oflags = GENERIC_READ;
4683 len = OPEN_EXISTING;
4685 oflags = GENERIC_READ|GENERIC_WRITE;
4688 mode = FILE_ATTRIBUTE_NORMAL;
4689 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4690 NULL, len, mode, NULL);
4692 if (F_ISSET(flags, MDB_RDONLY))
4695 oflags = O_RDWR | O_CREAT;
4697 env->me_fd = open(dpath, oflags, mode);
4699 if (env->me_fd == INVALID_HANDLE_VALUE) {
4704 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4705 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4710 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4711 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4712 env->me_mfd = env->me_fd;
4714 /* Synchronous fd for meta writes. Needed even with
4715 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4718 len = OPEN_EXISTING;
4719 env->me_mfd = CreateFile(dpath, oflags,
4720 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4721 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4724 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4726 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4731 DPRINTF(("opened dbenv %p", (void *) env));
4733 rc = mdb_env_share_locks(env, &excl);
4737 if (!(flags & MDB_RDONLY)) {
4739 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4740 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4741 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4742 (txn = calloc(1, size)))
4744 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4745 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4746 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4747 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4749 txn->mt_dbxs = env->me_dbxs;
4759 mdb_env_close0(env, excl);
4765 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4767 mdb_env_close0(MDB_env *env, int excl)
4771 if (!(env->me_flags & MDB_ENV_ACTIVE))
4774 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4776 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4777 free(env->me_dbxs[i].md_name.mv_data);
4782 free(env->me_dbiseqs);
4783 free(env->me_dbflags);
4785 free(env->me_dirty_list);
4787 mdb_midl_free(env->me_free_pgs);
4789 if (env->me_flags & MDB_ENV_TXKEY) {
4790 pthread_key_delete(env->me_txkey);
4792 /* Delete our key from the global list */
4793 for (i=0; i<mdb_tls_nkeys; i++)
4794 if (mdb_tls_keys[i] == env->me_txkey) {
4795 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4803 munmap(env->me_map, env->me_mapsize);
4805 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4806 (void) close(env->me_mfd);
4807 if (env->me_fd != INVALID_HANDLE_VALUE)
4808 (void) close(env->me_fd);
4810 MDB_PID_T pid = env->me_pid;
4811 /* Clearing readers is done in this function because
4812 * me_txkey with its destructor must be disabled first.
4814 * We skip the the reader mutex, so we touch only
4815 * data owned by this process (me_close_readers and
4816 * our readers), and clear each reader atomically.
4818 for (i = env->me_close_readers; --i >= 0; )
4819 if (env->me_txns->mti_readers[i].mr_pid == pid)
4820 env->me_txns->mti_readers[i].mr_pid = 0;
4822 if (env->me_rmutex) {
4823 CloseHandle(env->me_rmutex);
4824 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4826 /* Windows automatically destroys the mutexes when
4827 * the last handle closes.
4829 #elif defined(MDB_USE_SYSV_SEM)
4830 if (env->me_rmutex.semid != -1) {
4831 /* If we have the filelock: If we are the
4832 * only remaining user, clean up semaphores.
4835 mdb_env_excl_lock(env, &excl);
4837 semctl(env->me_rmutex.semid, 0, IPC_RMID);
4840 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4842 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4845 /* Unlock the lockfile. Windows would have unlocked it
4846 * after closing anyway, but not necessarily at once.
4848 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4851 (void) close(env->me_lfd);
4854 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4858 mdb_env_close(MDB_env *env)
4865 VGMEMP_DESTROY(env);
4866 while ((dp = env->me_dpages) != NULL) {
4867 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4868 env->me_dpages = dp->mp_next;
4872 mdb_env_close0(env, 0);
4876 /** Compare two items pointing at aligned size_t's */
4878 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4880 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4881 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4884 /** Compare two items pointing at aligned unsigned int's */
4886 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4888 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4889 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4892 /** Compare two items pointing at unsigned ints of unknown alignment.
4893 * Nodes and keys are guaranteed to be 2-byte aligned.
4896 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4898 #if BYTE_ORDER == LITTLE_ENDIAN
4899 unsigned short *u, *c;
4902 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4903 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4906 } while(!x && u > (unsigned short *)a->mv_data);
4909 unsigned short *u, *c, *end;
4912 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4913 u = (unsigned short *)a->mv_data;
4914 c = (unsigned short *)b->mv_data;
4917 } while(!x && u < end);
4922 /** Compare two items pointing at size_t's of unknown alignment. */
4923 #ifdef MISALIGNED_OK
4924 # define mdb_cmp_clong mdb_cmp_long
4926 # define mdb_cmp_clong mdb_cmp_cint
4929 /** Compare two items lexically */
4931 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4938 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4944 diff = memcmp(a->mv_data, b->mv_data, len);
4945 return diff ? diff : len_diff<0 ? -1 : len_diff;
4948 /** Compare two items in reverse byte order */
4950 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4952 const unsigned char *p1, *p2, *p1_lim;
4956 p1_lim = (const unsigned char *)a->mv_data;
4957 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4958 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4960 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4966 while (p1 > p1_lim) {
4967 diff = *--p1 - *--p2;
4971 return len_diff<0 ? -1 : len_diff;
4974 /** Search for key within a page, using binary search.
4975 * Returns the smallest entry larger or equal to the key.
4976 * If exactp is non-null, stores whether the found entry was an exact match
4977 * in *exactp (1 or 0).
4978 * Updates the cursor index with the index of the found entry.
4979 * If no entry larger or equal to the key is found, returns NULL.
4982 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4984 unsigned int i = 0, nkeys;
4987 MDB_page *mp = mc->mc_pg[mc->mc_top];
4988 MDB_node *node = NULL;
4993 nkeys = NUMKEYS(mp);
4995 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4996 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4999 low = IS_LEAF(mp) ? 0 : 1;
5001 cmp = mc->mc_dbx->md_cmp;
5003 /* Branch pages have no data, so if using integer keys,
5004 * alignment is guaranteed. Use faster mdb_cmp_int.
5006 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5007 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5014 nodekey.mv_size = mc->mc_db->md_pad;
5015 node = NODEPTR(mp, 0); /* fake */
5016 while (low <= high) {
5017 i = (low + high) >> 1;
5018 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5019 rc = cmp(key, &nodekey);
5020 DPRINTF(("found leaf index %u [%s], rc = %i",
5021 i, DKEY(&nodekey), rc));
5030 while (low <= high) {
5031 i = (low + high) >> 1;
5033 node = NODEPTR(mp, i);
5034 nodekey.mv_size = NODEKSZ(node);
5035 nodekey.mv_data = NODEKEY(node);
5037 rc = cmp(key, &nodekey);
5040 DPRINTF(("found leaf index %u [%s], rc = %i",
5041 i, DKEY(&nodekey), rc));
5043 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5044 i, DKEY(&nodekey), NODEPGNO(node), rc));
5055 if (rc > 0) { /* Found entry is less than the key. */
5056 i++; /* Skip to get the smallest entry larger than key. */
5058 node = NODEPTR(mp, i);
5061 *exactp = (rc == 0 && nkeys > 0);
5062 /* store the key index */
5063 mc->mc_ki[mc->mc_top] = i;
5065 /* There is no entry larger or equal to the key. */
5068 /* nodeptr is fake for LEAF2 */
5074 mdb_cursor_adjust(MDB_cursor *mc, func)
5078 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5079 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5086 /** Pop a page off the top of the cursor's stack. */
5088 mdb_cursor_pop(MDB_cursor *mc)
5092 MDB_page *top = mc->mc_pg[mc->mc_top];
5098 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5099 DDBI(mc), (void *) mc));
5103 /** Push a page onto the top of the cursor's stack. */
5105 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5107 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5108 DDBI(mc), (void *) mc));
5110 if (mc->mc_snum >= CURSOR_STACK) {
5111 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5112 return MDB_CURSOR_FULL;
5115 mc->mc_top = mc->mc_snum++;
5116 mc->mc_pg[mc->mc_top] = mp;
5117 mc->mc_ki[mc->mc_top] = 0;
5122 /** Find the address of the page corresponding to a given page number.
5123 * @param[in] txn the transaction for this access.
5124 * @param[in] pgno the page number for the page to retrieve.
5125 * @param[out] ret address of a pointer where the page's address will be stored.
5126 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5127 * @return 0 on success, non-zero on failure.
5130 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5132 MDB_env *env = txn->mt_env;
5136 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5140 MDB_ID2L dl = tx2->mt_u.dirty_list;
5142 /* Spilled pages were dirtied in this txn and flushed
5143 * because the dirty list got full. Bring this page
5144 * back in from the map (but don't unspill it here,
5145 * leave that unless page_touch happens again).
5147 if (tx2->mt_spill_pgs) {
5148 MDB_ID pn = pgno << 1;
5149 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5150 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5151 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5156 unsigned x = mdb_mid2l_search(dl, pgno);
5157 if (x <= dl[0].mid && dl[x].mid == pgno) {
5163 } while ((tx2 = tx2->mt_parent) != NULL);
5166 if (pgno < txn->mt_next_pgno) {
5168 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5170 DPRINTF(("page %"Z"u not found", pgno));
5171 txn->mt_flags |= MDB_TXN_ERROR;
5172 return MDB_PAGE_NOTFOUND;
5182 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5183 * The cursor is at the root page, set up the rest of it.
5186 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5188 MDB_page *mp = mc->mc_pg[mc->mc_top];
5192 while (IS_BRANCH(mp)) {
5196 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5197 mdb_cassert(mc, NUMKEYS(mp) > 1);
5198 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5200 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5202 if (flags & MDB_PS_LAST)
5203 i = NUMKEYS(mp) - 1;
5206 node = mdb_node_search(mc, key, &exact);
5208 i = NUMKEYS(mp) - 1;
5210 i = mc->mc_ki[mc->mc_top];
5212 mdb_cassert(mc, i > 0);
5216 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5219 mdb_cassert(mc, i < NUMKEYS(mp));
5220 node = NODEPTR(mp, i);
5222 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5225 mc->mc_ki[mc->mc_top] = i;
5226 if ((rc = mdb_cursor_push(mc, mp)))
5229 if (flags & MDB_PS_MODIFY) {
5230 if ((rc = mdb_page_touch(mc)) != 0)
5232 mp = mc->mc_pg[mc->mc_top];
5237 DPRINTF(("internal error, index points to a %02X page!?",
5239 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5240 return MDB_CORRUPTED;
5243 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5244 key ? DKEY(key) : "null"));
5245 mc->mc_flags |= C_INITIALIZED;
5246 mc->mc_flags &= ~C_EOF;
5251 /** Search for the lowest key under the current branch page.
5252 * This just bypasses a NUMKEYS check in the current page
5253 * before calling mdb_page_search_root(), because the callers
5254 * are all in situations where the current page is known to
5258 mdb_page_search_lowest(MDB_cursor *mc)
5260 MDB_page *mp = mc->mc_pg[mc->mc_top];
5261 MDB_node *node = NODEPTR(mp, 0);
5264 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5267 mc->mc_ki[mc->mc_top] = 0;
5268 if ((rc = mdb_cursor_push(mc, mp)))
5270 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5273 /** Search for the page a given key should be in.
5274 * Push it and its parent pages on the cursor stack.
5275 * @param[in,out] mc the cursor for this operation.
5276 * @param[in] key the key to search for, or NULL for first/last page.
5277 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5278 * are touched (updated with new page numbers).
5279 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5280 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5281 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5282 * @return 0 on success, non-zero on failure.
5285 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5290 /* Make sure the txn is still viable, then find the root from
5291 * the txn's db table and set it as the root of the cursor's stack.
5293 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5294 DPUTS("transaction has failed, must abort");
5297 /* Make sure we're using an up-to-date root */
5298 if (*mc->mc_dbflag & DB_STALE) {
5300 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5302 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5303 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5310 MDB_node *leaf = mdb_node_search(&mc2,
5311 &mc->mc_dbx->md_name, &exact);
5313 return MDB_NOTFOUND;
5314 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5317 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5319 /* The txn may not know this DBI, or another process may
5320 * have dropped and recreated the DB with other flags.
5322 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5323 return MDB_INCOMPATIBLE;
5324 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5326 *mc->mc_dbflag &= ~DB_STALE;
5328 root = mc->mc_db->md_root;
5330 if (root == P_INVALID) { /* Tree is empty. */
5331 DPUTS("tree is empty");
5332 return MDB_NOTFOUND;
5336 mdb_cassert(mc, root > 1);
5337 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5338 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5344 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5345 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5347 if (flags & MDB_PS_MODIFY) {
5348 if ((rc = mdb_page_touch(mc)))
5352 if (flags & MDB_PS_ROOTONLY)
5355 return mdb_page_search_root(mc, key, flags);
5359 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5361 MDB_txn *txn = mc->mc_txn;
5362 pgno_t pg = mp->mp_pgno;
5363 unsigned x = 0, ovpages = mp->mp_pages;
5364 MDB_env *env = txn->mt_env;
5365 MDB_IDL sl = txn->mt_spill_pgs;
5366 MDB_ID pn = pg << 1;
5369 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5370 /* If the page is dirty or on the spill list we just acquired it,
5371 * so we should give it back to our current free list, if any.
5372 * Otherwise put it onto the list of pages we freed in this txn.
5374 * Won't create me_pghead: me_pglast must be inited along with it.
5375 * Unsupported in nested txns: They would need to hide the page
5376 * range in ancestor txns' dirty and spilled lists.
5378 if (env->me_pghead &&
5380 ((mp->mp_flags & P_DIRTY) ||
5381 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5385 MDB_ID2 *dl, ix, iy;
5386 rc = mdb_midl_need(&env->me_pghead, ovpages);
5389 if (!(mp->mp_flags & P_DIRTY)) {
5390 /* This page is no longer spilled */
5397 /* Remove from dirty list */
5398 dl = txn->mt_u.dirty_list;
5400 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5406 mdb_cassert(mc, x > 1);
5408 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5409 txn->mt_flags |= MDB_TXN_ERROR;
5410 return MDB_CORRUPTED;
5413 if (!(env->me_flags & MDB_WRITEMAP))
5414 mdb_dpage_free(env, mp);
5416 /* Insert in me_pghead */
5417 mop = env->me_pghead;
5418 j = mop[0] + ovpages;
5419 for (i = mop[0]; i && mop[i] < pg; i--)
5425 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5429 mc->mc_db->md_overflow_pages -= ovpages;
5433 /** Return the data associated with a given node.
5434 * @param[in] txn The transaction for this operation.
5435 * @param[in] leaf The node being read.
5436 * @param[out] data Updated to point to the node's data.
5437 * @return 0 on success, non-zero on failure.
5440 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5442 MDB_page *omp; /* overflow page */
5446 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5447 data->mv_size = NODEDSZ(leaf);
5448 data->mv_data = NODEDATA(leaf);
5452 /* Read overflow data.
5454 data->mv_size = NODEDSZ(leaf);
5455 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5456 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5457 DPRINTF(("read overflow page %"Z"u failed", pgno));
5460 data->mv_data = METADATA(omp);
5466 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5467 MDB_val *key, MDB_val *data)
5474 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5476 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5479 if (txn->mt_flags & MDB_TXN_ERROR)
5482 mdb_cursor_init(&mc, txn, dbi, &mx);
5483 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5486 /** Find a sibling for a page.
5487 * Replaces the page at the top of the cursor's stack with the
5488 * specified sibling, if one exists.
5489 * @param[in] mc The cursor for this operation.
5490 * @param[in] move_right Non-zero if the right sibling is requested,
5491 * otherwise the left sibling.
5492 * @return 0 on success, non-zero on failure.
5495 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5501 if (mc->mc_snum < 2) {
5502 return MDB_NOTFOUND; /* root has no siblings */
5506 DPRINTF(("parent page is page %"Z"u, index %u",
5507 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5509 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5510 : (mc->mc_ki[mc->mc_top] == 0)) {
5511 DPRINTF(("no more keys left, moving to %s sibling",
5512 move_right ? "right" : "left"));
5513 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5514 /* undo cursor_pop before returning */
5521 mc->mc_ki[mc->mc_top]++;
5523 mc->mc_ki[mc->mc_top]--;
5524 DPRINTF(("just moving to %s index key %u",
5525 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5527 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5529 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5530 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5531 /* mc will be inconsistent if caller does mc_snum++ as above */
5532 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5536 mdb_cursor_push(mc, mp);
5538 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5543 /** Move the cursor to the next data item. */
5545 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5551 if (mc->mc_flags & C_EOF) {
5552 return MDB_NOTFOUND;
5555 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5557 mp = mc->mc_pg[mc->mc_top];
5559 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5560 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5561 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5562 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5563 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5564 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5565 if (rc == MDB_SUCCESS)
5566 MDB_GET_KEY(leaf, key);
5571 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5572 if (op == MDB_NEXT_DUP)
5573 return MDB_NOTFOUND;
5577 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5578 mdb_dbg_pgno(mp), (void *) mc));
5579 if (mc->mc_flags & C_DEL)
5582 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5583 DPUTS("=====> move to next sibling page");
5584 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5585 mc->mc_flags |= C_EOF;
5588 mp = mc->mc_pg[mc->mc_top];
5589 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5591 mc->mc_ki[mc->mc_top]++;
5594 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5595 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5598 key->mv_size = mc->mc_db->md_pad;
5599 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5603 mdb_cassert(mc, IS_LEAF(mp));
5604 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5606 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5607 mdb_xcursor_init1(mc, leaf);
5610 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5613 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5614 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5615 if (rc != MDB_SUCCESS)
5620 MDB_GET_KEY(leaf, key);
5624 /** Move the cursor to the previous data item. */
5626 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5632 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5634 mp = mc->mc_pg[mc->mc_top];
5636 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5637 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5638 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5639 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5640 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5641 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5642 if (rc == MDB_SUCCESS) {
5643 MDB_GET_KEY(leaf, key);
5644 mc->mc_flags &= ~C_EOF;
5650 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5651 if (op == MDB_PREV_DUP)
5652 return MDB_NOTFOUND;
5656 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5657 mdb_dbg_pgno(mp), (void *) mc));
5659 if (mc->mc_ki[mc->mc_top] == 0) {
5660 DPUTS("=====> move to prev sibling page");
5661 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5664 mp = mc->mc_pg[mc->mc_top];
5665 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5666 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5668 mc->mc_ki[mc->mc_top]--;
5670 mc->mc_flags &= ~C_EOF;
5672 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5673 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5676 key->mv_size = mc->mc_db->md_pad;
5677 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5681 mdb_cassert(mc, IS_LEAF(mp));
5682 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5684 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5685 mdb_xcursor_init1(mc, leaf);
5688 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5691 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5692 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5693 if (rc != MDB_SUCCESS)
5698 MDB_GET_KEY(leaf, key);
5702 /** Set the cursor on a specific data item. */
5704 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5705 MDB_cursor_op op, int *exactp)
5709 MDB_node *leaf = NULL;
5712 if (key->mv_size == 0)
5713 return MDB_BAD_VALSIZE;
5716 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5718 /* See if we're already on the right page */
5719 if (mc->mc_flags & C_INITIALIZED) {
5722 mp = mc->mc_pg[mc->mc_top];
5724 mc->mc_ki[mc->mc_top] = 0;
5725 return MDB_NOTFOUND;
5727 if (mp->mp_flags & P_LEAF2) {
5728 nodekey.mv_size = mc->mc_db->md_pad;
5729 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5731 leaf = NODEPTR(mp, 0);
5732 MDB_GET_KEY2(leaf, nodekey);
5734 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5736 /* Probably happens rarely, but first node on the page
5737 * was the one we wanted.
5739 mc->mc_ki[mc->mc_top] = 0;
5746 unsigned int nkeys = NUMKEYS(mp);
5748 if (mp->mp_flags & P_LEAF2) {
5749 nodekey.mv_data = LEAF2KEY(mp,
5750 nkeys-1, nodekey.mv_size);
5752 leaf = NODEPTR(mp, nkeys-1);
5753 MDB_GET_KEY2(leaf, nodekey);
5755 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5757 /* last node was the one we wanted */
5758 mc->mc_ki[mc->mc_top] = nkeys-1;
5764 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5765 /* This is definitely the right page, skip search_page */
5766 if (mp->mp_flags & P_LEAF2) {
5767 nodekey.mv_data = LEAF2KEY(mp,
5768 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5770 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5771 MDB_GET_KEY2(leaf, nodekey);
5773 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5775 /* current node was the one we wanted */
5785 /* If any parents have right-sibs, search.
5786 * Otherwise, there's nothing further.
5788 for (i=0; i<mc->mc_top; i++)
5790 NUMKEYS(mc->mc_pg[i])-1)
5792 if (i == mc->mc_top) {
5793 /* There are no other pages */
5794 mc->mc_ki[mc->mc_top] = nkeys;
5795 return MDB_NOTFOUND;
5799 /* There are no other pages */
5800 mc->mc_ki[mc->mc_top] = 0;
5801 if (op == MDB_SET_RANGE && !exactp) {
5805 return MDB_NOTFOUND;
5809 rc = mdb_page_search(mc, key, 0);
5810 if (rc != MDB_SUCCESS)
5813 mp = mc->mc_pg[mc->mc_top];
5814 mdb_cassert(mc, IS_LEAF(mp));
5817 leaf = mdb_node_search(mc, key, exactp);
5818 if (exactp != NULL && !*exactp) {
5819 /* MDB_SET specified and not an exact match. */
5820 return MDB_NOTFOUND;
5824 DPUTS("===> inexact leaf not found, goto sibling");
5825 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5826 return rc; /* no entries matched */
5827 mp = mc->mc_pg[mc->mc_top];
5828 mdb_cassert(mc, IS_LEAF(mp));
5829 leaf = NODEPTR(mp, 0);
5833 mc->mc_flags |= C_INITIALIZED;
5834 mc->mc_flags &= ~C_EOF;
5837 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5838 key->mv_size = mc->mc_db->md_pad;
5839 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5844 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5845 mdb_xcursor_init1(mc, leaf);
5848 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5849 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5850 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5853 if (op == MDB_GET_BOTH) {
5859 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5860 if (rc != MDB_SUCCESS)
5863 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5865 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5867 rc = mc->mc_dbx->md_dcmp(data, &d2);
5869 if (op == MDB_GET_BOTH || rc > 0)
5870 return MDB_NOTFOUND;
5877 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5878 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5883 /* The key already matches in all other cases */
5884 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5885 MDB_GET_KEY(leaf, key);
5886 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5891 /** Move the cursor to the first item in the database. */
5893 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5899 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5901 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5902 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5903 if (rc != MDB_SUCCESS)
5906 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5908 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5909 mc->mc_flags |= C_INITIALIZED;
5910 mc->mc_flags &= ~C_EOF;
5912 mc->mc_ki[mc->mc_top] = 0;
5914 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5915 key->mv_size = mc->mc_db->md_pad;
5916 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5921 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5922 mdb_xcursor_init1(mc, leaf);
5923 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5927 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5931 MDB_GET_KEY(leaf, key);
5935 /** Move the cursor to the last item in the database. */
5937 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5943 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5945 if (!(mc->mc_flags & C_EOF)) {
5947 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5948 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5949 if (rc != MDB_SUCCESS)
5952 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5955 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5956 mc->mc_flags |= C_INITIALIZED|C_EOF;
5957 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5959 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5960 key->mv_size = mc->mc_db->md_pad;
5961 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5966 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5967 mdb_xcursor_init1(mc, leaf);
5968 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5972 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5977 MDB_GET_KEY(leaf, key);
5982 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5987 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5992 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5996 case MDB_GET_CURRENT:
5997 if (!(mc->mc_flags & C_INITIALIZED)) {
6000 MDB_page *mp = mc->mc_pg[mc->mc_top];
6001 int nkeys = NUMKEYS(mp);
6002 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6003 mc->mc_ki[mc->mc_top] = nkeys;
6009 key->mv_size = mc->mc_db->md_pad;
6010 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6012 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6013 MDB_GET_KEY(leaf, key);
6015 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6016 if (mc->mc_flags & C_DEL)
6017 mdb_xcursor_init1(mc, leaf);
6018 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6020 rc = mdb_node_read(mc->mc_txn, leaf, data);
6027 case MDB_GET_BOTH_RANGE:
6032 if (mc->mc_xcursor == NULL) {
6033 rc = MDB_INCOMPATIBLE;
6043 rc = mdb_cursor_set(mc, key, data, op,
6044 op == MDB_SET_RANGE ? NULL : &exact);
6047 case MDB_GET_MULTIPLE:
6048 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6052 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6053 rc = MDB_INCOMPATIBLE;
6057 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6058 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6061 case MDB_NEXT_MULTIPLE:
6066 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6067 rc = MDB_INCOMPATIBLE;
6070 if (!(mc->mc_flags & C_INITIALIZED))
6071 rc = mdb_cursor_first(mc, key, data);
6073 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6074 if (rc == MDB_SUCCESS) {
6075 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6078 mx = &mc->mc_xcursor->mx_cursor;
6079 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6081 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6082 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6090 case MDB_NEXT_NODUP:
6091 if (!(mc->mc_flags & C_INITIALIZED))
6092 rc = mdb_cursor_first(mc, key, data);
6094 rc = mdb_cursor_next(mc, key, data, op);
6098 case MDB_PREV_NODUP:
6099 if (!(mc->mc_flags & C_INITIALIZED)) {
6100 rc = mdb_cursor_last(mc, key, data);
6103 mc->mc_flags |= C_INITIALIZED;
6104 mc->mc_ki[mc->mc_top]++;
6106 rc = mdb_cursor_prev(mc, key, data, op);
6109 rc = mdb_cursor_first(mc, key, data);
6112 mfunc = mdb_cursor_first;
6114 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6118 if (mc->mc_xcursor == NULL) {
6119 rc = MDB_INCOMPATIBLE;
6123 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6124 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6125 MDB_GET_KEY(leaf, key);
6126 rc = mdb_node_read(mc->mc_txn, leaf, data);
6130 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6134 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6137 rc = mdb_cursor_last(mc, key, data);
6140 mfunc = mdb_cursor_last;
6143 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6148 if (mc->mc_flags & C_DEL)
6149 mc->mc_flags ^= C_DEL;
6154 /** Touch all the pages in the cursor stack. Set mc_top.
6155 * Makes sure all the pages are writable, before attempting a write operation.
6156 * @param[in] mc The cursor to operate on.
6159 mdb_cursor_touch(MDB_cursor *mc)
6161 int rc = MDB_SUCCESS;
6163 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6166 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6168 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6169 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6172 *mc->mc_dbflag |= DB_DIRTY;
6177 rc = mdb_page_touch(mc);
6178 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6179 mc->mc_top = mc->mc_snum-1;
6184 /** Do not spill pages to disk if txn is getting full, may fail instead */
6185 #define MDB_NOSPILL 0x8000
6188 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6192 MDB_node *leaf = NULL;
6195 MDB_val xdata, *rdata, dkey, olddata;
6197 int do_sub = 0, insert_key, insert_data;
6198 unsigned int mcount = 0, dcount = 0, nospill;
6201 unsigned int nflags;
6204 if (mc == NULL || key == NULL)
6207 env = mc->mc_txn->mt_env;
6209 /* Check this first so counter will always be zero on any
6212 if (flags & MDB_MULTIPLE) {
6213 dcount = data[1].mv_size;
6214 data[1].mv_size = 0;
6215 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6216 return MDB_INCOMPATIBLE;
6219 nospill = flags & MDB_NOSPILL;
6220 flags &= ~MDB_NOSPILL;
6222 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6223 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6225 if (key->mv_size-1 >= ENV_MAXKEY(env))
6226 return MDB_BAD_VALSIZE;
6228 #if SIZE_MAX > MAXDATASIZE
6229 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6230 return MDB_BAD_VALSIZE;
6232 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6233 return MDB_BAD_VALSIZE;
6236 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6237 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6241 if (flags == MDB_CURRENT) {
6242 if (!(mc->mc_flags & C_INITIALIZED))
6245 } else if (mc->mc_db->md_root == P_INVALID) {
6246 /* new database, cursor has nothing to point to */
6249 mc->mc_flags &= ~C_INITIALIZED;
6254 if (flags & MDB_APPEND) {
6256 rc = mdb_cursor_last(mc, &k2, &d2);
6258 rc = mc->mc_dbx->md_cmp(key, &k2);
6261 mc->mc_ki[mc->mc_top]++;
6263 /* new key is <= last key */
6268 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6270 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6271 DPRINTF(("duplicate key [%s]", DKEY(key)));
6273 return MDB_KEYEXIST;
6275 if (rc && rc != MDB_NOTFOUND)
6279 if (mc->mc_flags & C_DEL)
6280 mc->mc_flags ^= C_DEL;
6282 /* Cursor is positioned, check for room in the dirty list */
6284 if (flags & MDB_MULTIPLE) {
6286 xdata.mv_size = data->mv_size * dcount;
6290 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6294 if (rc == MDB_NO_ROOT) {
6296 /* new database, write a root leaf page */
6297 DPUTS("allocating new root leaf page");
6298 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6301 mdb_cursor_push(mc, np);
6302 mc->mc_db->md_root = np->mp_pgno;
6303 mc->mc_db->md_depth++;
6304 *mc->mc_dbflag |= DB_DIRTY;
6305 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6307 np->mp_flags |= P_LEAF2;
6308 mc->mc_flags |= C_INITIALIZED;
6310 /* make sure all cursor pages are writable */
6311 rc2 = mdb_cursor_touch(mc);
6316 insert_key = insert_data = rc;
6318 /* The key does not exist */
6319 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6320 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6321 LEAFSIZE(key, data) > env->me_nodemax)
6323 /* Too big for a node, insert in sub-DB. Set up an empty
6324 * "old sub-page" for prep_subDB to expand to a full page.
6326 fp_flags = P_LEAF|P_DIRTY;
6328 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6329 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6330 olddata.mv_size = PAGEHDRSZ;
6334 /* there's only a key anyway, so this is a no-op */
6335 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6337 unsigned int ksize = mc->mc_db->md_pad;
6338 if (key->mv_size != ksize)
6339 return MDB_BAD_VALSIZE;
6340 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6341 memcpy(ptr, key->mv_data, ksize);
6343 /* if overwriting slot 0 of leaf, need to
6344 * update branch key if there is a parent page
6346 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6347 unsigned short top = mc->mc_top;
6349 /* slot 0 is always an empty key, find real slot */
6350 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6352 if (mc->mc_ki[mc->mc_top])
6353 rc2 = mdb_update_key(mc, key);
6364 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6365 olddata.mv_size = NODEDSZ(leaf);
6366 olddata.mv_data = NODEDATA(leaf);
6369 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6370 /* Prepare (sub-)page/sub-DB to accept the new item,
6371 * if needed. fp: old sub-page or a header faking
6372 * it. mp: new (sub-)page. offset: growth in page
6373 * size. xdata: node data with new page or DB.
6375 unsigned i, offset = 0;
6376 mp = fp = xdata.mv_data = env->me_pbuf;
6377 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6379 /* Was a single item before, must convert now */
6380 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6381 /* Just overwrite the current item */
6382 if (flags == MDB_CURRENT)
6385 #if UINT_MAX < SIZE_MAX
6386 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6387 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6389 /* does data match? */
6390 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6391 if (flags & MDB_NODUPDATA)
6392 return MDB_KEYEXIST;
6397 /* Back up original data item */
6398 dkey.mv_size = olddata.mv_size;
6399 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6401 /* Make sub-page header for the dup items, with dummy body */
6402 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6403 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6404 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6405 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6406 fp->mp_flags |= P_LEAF2;
6407 fp->mp_pad = data->mv_size;
6408 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6410 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6411 (dkey.mv_size & 1) + (data->mv_size & 1);
6413 fp->mp_upper = xdata.mv_size - PAGEBASE;
6414 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6415 } else if (leaf->mn_flags & F_SUBDATA) {
6416 /* Data is on sub-DB, just store it */
6417 flags |= F_DUPDATA|F_SUBDATA;
6420 /* Data is on sub-page */
6421 fp = olddata.mv_data;
6424 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6425 offset = EVEN(NODESIZE + sizeof(indx_t) +
6429 offset = fp->mp_pad;
6430 if (SIZELEFT(fp) < offset) {
6431 offset *= 4; /* space for 4 more */
6434 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6436 fp->mp_flags |= P_DIRTY;
6437 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6438 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6442 xdata.mv_size = olddata.mv_size + offset;
6445 fp_flags = fp->mp_flags;
6446 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6447 /* Too big for a sub-page, convert to sub-DB */
6448 fp_flags &= ~P_SUBP;
6450 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6451 fp_flags |= P_LEAF2;
6452 dummy.md_pad = fp->mp_pad;
6453 dummy.md_flags = MDB_DUPFIXED;
6454 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6455 dummy.md_flags |= MDB_INTEGERKEY;
6461 dummy.md_branch_pages = 0;
6462 dummy.md_leaf_pages = 1;
6463 dummy.md_overflow_pages = 0;
6464 dummy.md_entries = NUMKEYS(fp);
6465 xdata.mv_size = sizeof(MDB_db);
6466 xdata.mv_data = &dummy;
6467 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6469 offset = env->me_psize - olddata.mv_size;
6470 flags |= F_DUPDATA|F_SUBDATA;
6471 dummy.md_root = mp->mp_pgno;
6474 mp->mp_flags = fp_flags | P_DIRTY;
6475 mp->mp_pad = fp->mp_pad;
6476 mp->mp_lower = fp->mp_lower;
6477 mp->mp_upper = fp->mp_upper + offset;
6478 if (fp_flags & P_LEAF2) {
6479 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6481 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6482 olddata.mv_size - fp->mp_upper - PAGEBASE);
6483 for (i=0; i<NUMKEYS(fp); i++)
6484 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6492 mdb_node_del(mc, 0);
6496 /* overflow page overwrites need special handling */
6497 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6500 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6502 memcpy(&pg, olddata.mv_data, sizeof(pg));
6503 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6505 ovpages = omp->mp_pages;
6507 /* Is the ov page large enough? */
6508 if (ovpages >= dpages) {
6509 if (!(omp->mp_flags & P_DIRTY) &&
6510 (level || (env->me_flags & MDB_WRITEMAP)))
6512 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6515 level = 0; /* dirty in this txn or clean */
6518 if (omp->mp_flags & P_DIRTY) {
6519 /* yes, overwrite it. Note in this case we don't
6520 * bother to try shrinking the page if the new data
6521 * is smaller than the overflow threshold.
6524 /* It is writable only in a parent txn */
6525 size_t sz = (size_t) env->me_psize * ovpages, off;
6526 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6532 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6533 mdb_cassert(mc, rc2 == 0);
6534 if (!(flags & MDB_RESERVE)) {
6535 /* Copy end of page, adjusting alignment so
6536 * compiler may copy words instead of bytes.
6538 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6539 memcpy((size_t *)((char *)np + off),
6540 (size_t *)((char *)omp + off), sz - off);
6543 memcpy(np, omp, sz); /* Copy beginning of page */
6546 SETDSZ(leaf, data->mv_size);
6547 if (F_ISSET(flags, MDB_RESERVE))
6548 data->mv_data = METADATA(omp);
6550 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6554 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6556 } else if (data->mv_size == olddata.mv_size) {
6557 /* same size, just replace it. Note that we could
6558 * also reuse this node if the new data is smaller,
6559 * but instead we opt to shrink the node in that case.
6561 if (F_ISSET(flags, MDB_RESERVE))
6562 data->mv_data = olddata.mv_data;
6563 else if (!(mc->mc_flags & C_SUB))
6564 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6566 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6571 mdb_node_del(mc, 0);
6577 nflags = flags & NODE_ADD_FLAGS;
6578 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6579 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6580 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6581 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6583 nflags |= MDB_SPLIT_REPLACE;
6584 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6586 /* There is room already in this leaf page. */
6587 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6588 if (rc == 0 && insert_key) {
6589 /* Adjust other cursors pointing to mp */
6590 MDB_cursor *m2, *m3;
6591 MDB_dbi dbi = mc->mc_dbi;
6592 unsigned i = mc->mc_top;
6593 MDB_page *mp = mc->mc_pg[i];
6595 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6596 if (mc->mc_flags & C_SUB)
6597 m3 = &m2->mc_xcursor->mx_cursor;
6600 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6601 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6608 if (rc == MDB_SUCCESS) {
6609 /* Now store the actual data in the child DB. Note that we're
6610 * storing the user data in the keys field, so there are strict
6611 * size limits on dupdata. The actual data fields of the child
6612 * DB are all zero size.
6620 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6621 if (flags & MDB_CURRENT) {
6622 xflags = MDB_CURRENT|MDB_NOSPILL;
6624 mdb_xcursor_init1(mc, leaf);
6625 xflags = (flags & MDB_NODUPDATA) ?
6626 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6628 /* converted, write the original data first */
6630 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6634 /* Adjust other cursors pointing to mp */
6636 unsigned i = mc->mc_top;
6637 MDB_page *mp = mc->mc_pg[i];
6639 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6640 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6641 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6642 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6643 mdb_xcursor_init1(m2, leaf);
6647 /* we've done our job */
6650 ecount = mc->mc_xcursor->mx_db.md_entries;
6651 if (flags & MDB_APPENDDUP)
6652 xflags |= MDB_APPEND;
6653 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6654 if (flags & F_SUBDATA) {
6655 void *db = NODEDATA(leaf);
6656 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6658 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6660 /* Increment count unless we just replaced an existing item. */
6662 mc->mc_db->md_entries++;
6664 /* Invalidate txn if we created an empty sub-DB */
6667 /* If we succeeded and the key didn't exist before,
6668 * make sure the cursor is marked valid.
6670 mc->mc_flags |= C_INITIALIZED;
6672 if (flags & MDB_MULTIPLE) {
6675 /* let caller know how many succeeded, if any */
6676 data[1].mv_size = mcount;
6677 if (mcount < dcount) {
6678 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6679 insert_key = insert_data = 0;
6686 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6689 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6694 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6700 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6701 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6703 if (!(mc->mc_flags & C_INITIALIZED))
6706 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6707 return MDB_NOTFOUND;
6709 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6712 rc = mdb_cursor_touch(mc);
6716 mp = mc->mc_pg[mc->mc_top];
6719 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6721 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6722 if (flags & MDB_NODUPDATA) {
6723 /* mdb_cursor_del0() will subtract the final entry */
6724 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6726 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6727 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6729 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6732 /* If sub-DB still has entries, we're done */
6733 if (mc->mc_xcursor->mx_db.md_entries) {
6734 if (leaf->mn_flags & F_SUBDATA) {
6735 /* update subDB info */
6736 void *db = NODEDATA(leaf);
6737 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6740 /* shrink fake page */
6741 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6742 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6743 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6744 /* fix other sub-DB cursors pointed at this fake page */
6745 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6746 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6747 if (m2->mc_pg[mc->mc_top] == mp &&
6748 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6749 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6752 mc->mc_db->md_entries--;
6753 mc->mc_flags |= C_DEL;
6756 /* otherwise fall thru and delete the sub-DB */
6759 if (leaf->mn_flags & F_SUBDATA) {
6760 /* add all the child DB's pages to the free list */
6761 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6767 /* add overflow pages to free list */
6768 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6772 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6773 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6774 (rc = mdb_ovpage_free(mc, omp)))
6779 return mdb_cursor_del0(mc);
6782 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6786 /** Allocate and initialize new pages for a database.
6787 * @param[in] mc a cursor on the database being added to.
6788 * @param[in] flags flags defining what type of page is being allocated.
6789 * @param[in] num the number of pages to allocate. This is usually 1,
6790 * unless allocating overflow pages for a large record.
6791 * @param[out] mp Address of a page, or NULL on failure.
6792 * @return 0 on success, non-zero on failure.
6795 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6800 if ((rc = mdb_page_alloc(mc, num, &np)))
6802 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6803 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6804 np->mp_flags = flags | P_DIRTY;
6805 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6806 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6809 mc->mc_db->md_branch_pages++;
6810 else if (IS_LEAF(np))
6811 mc->mc_db->md_leaf_pages++;
6812 else if (IS_OVERFLOW(np)) {
6813 mc->mc_db->md_overflow_pages += num;
6821 /** Calculate the size of a leaf node.
6822 * The size depends on the environment's page size; if a data item
6823 * is too large it will be put onto an overflow page and the node
6824 * size will only include the key and not the data. Sizes are always
6825 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6826 * of the #MDB_node headers.
6827 * @param[in] env The environment handle.
6828 * @param[in] key The key for the node.
6829 * @param[in] data The data for the node.
6830 * @return The number of bytes needed to store the node.
6833 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6837 sz = LEAFSIZE(key, data);
6838 if (sz > env->me_nodemax) {
6839 /* put on overflow page */
6840 sz -= data->mv_size - sizeof(pgno_t);
6843 return EVEN(sz + sizeof(indx_t));
6846 /** Calculate the size of a branch node.
6847 * The size should depend on the environment's page size but since
6848 * we currently don't support spilling large keys onto overflow
6849 * pages, it's simply the size of the #MDB_node header plus the
6850 * size of the key. Sizes are always rounded up to an even number
6851 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6852 * @param[in] env The environment handle.
6853 * @param[in] key The key for the node.
6854 * @return The number of bytes needed to store the node.
6857 mdb_branch_size(MDB_env *env, MDB_val *key)
6862 if (sz > env->me_nodemax) {
6863 /* put on overflow page */
6864 /* not implemented */
6865 /* sz -= key->size - sizeof(pgno_t); */
6868 return sz + sizeof(indx_t);
6871 /** Add a node to the page pointed to by the cursor.
6872 * @param[in] mc The cursor for this operation.
6873 * @param[in] indx The index on the page where the new node should be added.
6874 * @param[in] key The key for the new node.
6875 * @param[in] data The data for the new node, if any.
6876 * @param[in] pgno The page number, if adding a branch node.
6877 * @param[in] flags Flags for the node.
6878 * @return 0 on success, non-zero on failure. Possible errors are:
6880 * <li>ENOMEM - failed to allocate overflow pages for the node.
6881 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6882 * should never happen since all callers already calculate the
6883 * page's free space before calling this function.
6887 mdb_node_add(MDB_cursor *mc, indx_t indx,
6888 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6891 size_t node_size = NODESIZE;
6895 MDB_page *mp = mc->mc_pg[mc->mc_top];
6896 MDB_page *ofp = NULL; /* overflow page */
6899 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6901 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6902 IS_LEAF(mp) ? "leaf" : "branch",
6903 IS_SUBP(mp) ? "sub-" : "",
6904 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6905 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6908 /* Move higher keys up one slot. */
6909 int ksize = mc->mc_db->md_pad, dif;
6910 char *ptr = LEAF2KEY(mp, indx, ksize);
6911 dif = NUMKEYS(mp) - indx;
6913 memmove(ptr+ksize, ptr, dif*ksize);
6914 /* insert new key */
6915 memcpy(ptr, key->mv_data, ksize);
6917 /* Just using these for counting */
6918 mp->mp_lower += sizeof(indx_t);
6919 mp->mp_upper -= ksize - sizeof(indx_t);
6923 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6925 node_size += key->mv_size;
6927 mdb_cassert(mc, data);
6928 if (F_ISSET(flags, F_BIGDATA)) {
6929 /* Data already on overflow page. */
6930 node_size += sizeof(pgno_t);
6931 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6932 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6934 /* Put data on overflow page. */
6935 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6936 data->mv_size, node_size+data->mv_size));
6937 node_size = EVEN(node_size + sizeof(pgno_t));
6938 if ((ssize_t)node_size > room)
6940 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6942 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6946 node_size += data->mv_size;
6949 node_size = EVEN(node_size);
6950 if ((ssize_t)node_size > room)
6954 /* Move higher pointers up one slot. */
6955 for (i = NUMKEYS(mp); i > indx; i--)
6956 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6958 /* Adjust free space offsets. */
6959 ofs = mp->mp_upper - node_size;
6960 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6961 mp->mp_ptrs[indx] = ofs;
6963 mp->mp_lower += sizeof(indx_t);
6965 /* Write the node data. */
6966 node = NODEPTR(mp, indx);
6967 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6968 node->mn_flags = flags;
6970 SETDSZ(node,data->mv_size);
6975 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6978 mdb_cassert(mc, key);
6980 if (F_ISSET(flags, F_BIGDATA))
6981 memcpy(node->mn_data + key->mv_size, data->mv_data,
6983 else if (F_ISSET(flags, MDB_RESERVE))
6984 data->mv_data = node->mn_data + key->mv_size;
6986 memcpy(node->mn_data + key->mv_size, data->mv_data,
6989 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6991 if (F_ISSET(flags, MDB_RESERVE))
6992 data->mv_data = METADATA(ofp);
6994 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
7001 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7002 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7003 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7004 DPRINTF(("node size = %"Z"u", node_size));
7005 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7006 return MDB_PAGE_FULL;
7009 /** Delete the specified node from a page.
7010 * @param[in] mc Cursor pointing to the node to delete.
7011 * @param[in] ksize The size of a node. Only used if the page is
7012 * part of a #MDB_DUPFIXED database.
7015 mdb_node_del(MDB_cursor *mc, int ksize)
7017 MDB_page *mp = mc->mc_pg[mc->mc_top];
7018 indx_t indx = mc->mc_ki[mc->mc_top];
7020 indx_t i, j, numkeys, ptr;
7024 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7025 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7026 numkeys = NUMKEYS(mp);
7027 mdb_cassert(mc, indx < numkeys);
7030 int x = numkeys - 1 - indx;
7031 base = LEAF2KEY(mp, indx, ksize);
7033 memmove(base, base + ksize, x * ksize);
7034 mp->mp_lower -= sizeof(indx_t);
7035 mp->mp_upper += ksize - sizeof(indx_t);
7039 node = NODEPTR(mp, indx);
7040 sz = NODESIZE + node->mn_ksize;
7042 if (F_ISSET(node->mn_flags, F_BIGDATA))
7043 sz += sizeof(pgno_t);
7045 sz += NODEDSZ(node);
7049 ptr = mp->mp_ptrs[indx];
7050 for (i = j = 0; i < numkeys; i++) {
7052 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7053 if (mp->mp_ptrs[i] < ptr)
7054 mp->mp_ptrs[j] += sz;
7059 base = (char *)mp + mp->mp_upper + PAGEBASE;
7060 memmove(base + sz, base, ptr - mp->mp_upper);
7062 mp->mp_lower -= sizeof(indx_t);
7066 /** Compact the main page after deleting a node on a subpage.
7067 * @param[in] mp The main page to operate on.
7068 * @param[in] indx The index of the subpage on the main page.
7071 mdb_node_shrink(MDB_page *mp, indx_t indx)
7077 indx_t i, numkeys, ptr;
7079 node = NODEPTR(mp, indx);
7080 sp = (MDB_page *)NODEDATA(node);
7081 delta = SIZELEFT(sp);
7082 xp = (MDB_page *)((char *)sp + delta);
7084 /* shift subpage upward */
7086 nsize = NUMKEYS(sp) * sp->mp_pad;
7088 return; /* do not make the node uneven-sized */
7089 memmove(METADATA(xp), METADATA(sp), nsize);
7092 numkeys = NUMKEYS(sp);
7093 for (i=numkeys-1; i>=0; i--)
7094 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7096 xp->mp_upper = sp->mp_lower;
7097 xp->mp_lower = sp->mp_lower;
7098 xp->mp_flags = sp->mp_flags;
7099 xp->mp_pad = sp->mp_pad;
7100 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7102 nsize = NODEDSZ(node) - delta;
7103 SETDSZ(node, nsize);
7105 /* shift lower nodes upward */
7106 ptr = mp->mp_ptrs[indx];
7107 numkeys = NUMKEYS(mp);
7108 for (i = 0; i < numkeys; i++) {
7109 if (mp->mp_ptrs[i] <= ptr)
7110 mp->mp_ptrs[i] += delta;
7113 base = (char *)mp + mp->mp_upper + PAGEBASE;
7114 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7115 mp->mp_upper += delta;
7118 /** Initial setup of a sorted-dups cursor.
7119 * Sorted duplicates are implemented as a sub-database for the given key.
7120 * The duplicate data items are actually keys of the sub-database.
7121 * Operations on the duplicate data items are performed using a sub-cursor
7122 * initialized when the sub-database is first accessed. This function does
7123 * the preliminary setup of the sub-cursor, filling in the fields that
7124 * depend only on the parent DB.
7125 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7128 mdb_xcursor_init0(MDB_cursor *mc)
7130 MDB_xcursor *mx = mc->mc_xcursor;
7132 mx->mx_cursor.mc_xcursor = NULL;
7133 mx->mx_cursor.mc_txn = mc->mc_txn;
7134 mx->mx_cursor.mc_db = &mx->mx_db;
7135 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7136 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7137 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7138 mx->mx_cursor.mc_snum = 0;
7139 mx->mx_cursor.mc_top = 0;
7140 mx->mx_cursor.mc_flags = C_SUB;
7141 mx->mx_dbx.md_name.mv_size = 0;
7142 mx->mx_dbx.md_name.mv_data = NULL;
7143 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7144 mx->mx_dbx.md_dcmp = NULL;
7145 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7148 /** Final setup of a sorted-dups cursor.
7149 * Sets up the fields that depend on the data from the main cursor.
7150 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7151 * @param[in] node The data containing the #MDB_db record for the
7152 * sorted-dup database.
7155 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7157 MDB_xcursor *mx = mc->mc_xcursor;
7159 if (node->mn_flags & F_SUBDATA) {
7160 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7161 mx->mx_cursor.mc_pg[0] = 0;
7162 mx->mx_cursor.mc_snum = 0;
7163 mx->mx_cursor.mc_top = 0;
7164 mx->mx_cursor.mc_flags = C_SUB;
7166 MDB_page *fp = NODEDATA(node);
7167 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7168 mx->mx_db.md_flags = 0;
7169 mx->mx_db.md_depth = 1;
7170 mx->mx_db.md_branch_pages = 0;
7171 mx->mx_db.md_leaf_pages = 1;
7172 mx->mx_db.md_overflow_pages = 0;
7173 mx->mx_db.md_entries = NUMKEYS(fp);
7174 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7175 mx->mx_cursor.mc_snum = 1;
7176 mx->mx_cursor.mc_top = 0;
7177 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7178 mx->mx_cursor.mc_pg[0] = fp;
7179 mx->mx_cursor.mc_ki[0] = 0;
7180 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7181 mx->mx_db.md_flags = MDB_DUPFIXED;
7182 mx->mx_db.md_pad = fp->mp_pad;
7183 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7184 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7187 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7188 mx->mx_db.md_root));
7189 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7190 #if UINT_MAX < SIZE_MAX
7191 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7192 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7196 /** Initialize a cursor for a given transaction and database. */
7198 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7201 mc->mc_backup = NULL;
7204 mc->mc_db = &txn->mt_dbs[dbi];
7205 mc->mc_dbx = &txn->mt_dbxs[dbi];
7206 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7212 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7213 mdb_tassert(txn, mx != NULL);
7214 mc->mc_xcursor = mx;
7215 mdb_xcursor_init0(mc);
7217 mc->mc_xcursor = NULL;
7219 if (*mc->mc_dbflag & DB_STALE) {
7220 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7225 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7228 size_t size = sizeof(MDB_cursor);
7230 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7233 if (txn->mt_flags & MDB_TXN_ERROR)
7236 /* Allow read access to the freelist */
7237 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7240 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7241 size += sizeof(MDB_xcursor);
7243 if ((mc = malloc(size)) != NULL) {
7244 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7245 if (txn->mt_cursors) {
7246 mc->mc_next = txn->mt_cursors[dbi];
7247 txn->mt_cursors[dbi] = mc;
7248 mc->mc_flags |= C_UNTRACK;
7260 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7262 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7265 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7268 if (txn->mt_flags & MDB_TXN_ERROR)
7271 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7275 /* Return the count of duplicate data items for the current key */
7277 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7281 if (mc == NULL || countp == NULL)
7284 if (mc->mc_xcursor == NULL)
7285 return MDB_INCOMPATIBLE;
7287 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7290 if (!(mc->mc_flags & C_INITIALIZED))
7293 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7294 return MDB_NOTFOUND;
7296 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7297 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7300 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7303 *countp = mc->mc_xcursor->mx_db.md_entries;
7309 mdb_cursor_close(MDB_cursor *mc)
7311 if (mc && !mc->mc_backup) {
7312 /* remove from txn, if tracked */
7313 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7314 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7315 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7317 *prev = mc->mc_next;
7324 mdb_cursor_txn(MDB_cursor *mc)
7326 if (!mc) return NULL;
7331 mdb_cursor_dbi(MDB_cursor *mc)
7336 /** Replace the key for a branch node with a new key.
7337 * @param[in] mc Cursor pointing to the node to operate on.
7338 * @param[in] key The new key to use.
7339 * @return 0 on success, non-zero on failure.
7342 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7348 int delta, ksize, oksize;
7349 indx_t ptr, i, numkeys, indx;
7352 indx = mc->mc_ki[mc->mc_top];
7353 mp = mc->mc_pg[mc->mc_top];
7354 node = NODEPTR(mp, indx);
7355 ptr = mp->mp_ptrs[indx];
7359 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7360 k2.mv_data = NODEKEY(node);
7361 k2.mv_size = node->mn_ksize;
7362 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7364 mdb_dkey(&k2, kbuf2),
7370 /* Sizes must be 2-byte aligned. */
7371 ksize = EVEN(key->mv_size);
7372 oksize = EVEN(node->mn_ksize);
7373 delta = ksize - oksize;
7375 /* Shift node contents if EVEN(key length) changed. */
7377 if (delta > 0 && SIZELEFT(mp) < delta) {
7379 /* not enough space left, do a delete and split */
7380 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7381 pgno = NODEPGNO(node);
7382 mdb_node_del(mc, 0);
7383 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7386 numkeys = NUMKEYS(mp);
7387 for (i = 0; i < numkeys; i++) {
7388 if (mp->mp_ptrs[i] <= ptr)
7389 mp->mp_ptrs[i] -= delta;
7392 base = (char *)mp + mp->mp_upper + PAGEBASE;
7393 len = ptr - mp->mp_upper + NODESIZE;
7394 memmove(base - delta, base, len);
7395 mp->mp_upper -= delta;
7397 node = NODEPTR(mp, indx);
7400 /* But even if no shift was needed, update ksize */
7401 if (node->mn_ksize != key->mv_size)
7402 node->mn_ksize = key->mv_size;
7405 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7411 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7413 /** Move a node from csrc to cdst.
7416 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7423 unsigned short flags;
7427 /* Mark src and dst as dirty. */
7428 if ((rc = mdb_page_touch(csrc)) ||
7429 (rc = mdb_page_touch(cdst)))
7432 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7433 key.mv_size = csrc->mc_db->md_pad;
7434 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7436 data.mv_data = NULL;
7440 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7441 mdb_cassert(csrc, !((size_t)srcnode & 1));
7442 srcpg = NODEPGNO(srcnode);
7443 flags = srcnode->mn_flags;
7444 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7445 unsigned int snum = csrc->mc_snum;
7447 /* must find the lowest key below src */
7448 rc = mdb_page_search_lowest(csrc);
7451 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7452 key.mv_size = csrc->mc_db->md_pad;
7453 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7455 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7456 key.mv_size = NODEKSZ(s2);
7457 key.mv_data = NODEKEY(s2);
7459 csrc->mc_snum = snum--;
7460 csrc->mc_top = snum;
7462 key.mv_size = NODEKSZ(srcnode);
7463 key.mv_data = NODEKEY(srcnode);
7465 data.mv_size = NODEDSZ(srcnode);
7466 data.mv_data = NODEDATA(srcnode);
7468 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7469 unsigned int snum = cdst->mc_snum;
7472 /* must find the lowest key below dst */
7473 mdb_cursor_copy(cdst, &mn);
7474 rc = mdb_page_search_lowest(&mn);
7477 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7478 bkey.mv_size = mn.mc_db->md_pad;
7479 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7481 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7482 bkey.mv_size = NODEKSZ(s2);
7483 bkey.mv_data = NODEKEY(s2);
7485 mn.mc_snum = snum--;
7488 rc = mdb_update_key(&mn, &bkey);
7493 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7494 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7495 csrc->mc_ki[csrc->mc_top],
7497 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7498 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7500 /* Add the node to the destination page.
7502 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7503 if (rc != MDB_SUCCESS)
7506 /* Delete the node from the source page.
7508 mdb_node_del(csrc, key.mv_size);
7511 /* Adjust other cursors pointing to mp */
7512 MDB_cursor *m2, *m3;
7513 MDB_dbi dbi = csrc->mc_dbi;
7514 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7516 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7517 if (csrc->mc_flags & C_SUB)
7518 m3 = &m2->mc_xcursor->mx_cursor;
7521 if (m3 == csrc) continue;
7522 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7523 csrc->mc_ki[csrc->mc_top]) {
7524 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7525 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7530 /* Update the parent separators.
7532 if (csrc->mc_ki[csrc->mc_top] == 0) {
7533 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7534 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7535 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7537 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7538 key.mv_size = NODEKSZ(srcnode);
7539 key.mv_data = NODEKEY(srcnode);
7541 DPRINTF(("update separator for source page %"Z"u to [%s]",
7542 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7543 mdb_cursor_copy(csrc, &mn);
7546 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7549 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7551 indx_t ix = csrc->mc_ki[csrc->mc_top];
7552 nullkey.mv_size = 0;
7553 csrc->mc_ki[csrc->mc_top] = 0;
7554 rc = mdb_update_key(csrc, &nullkey);
7555 csrc->mc_ki[csrc->mc_top] = ix;
7556 mdb_cassert(csrc, rc == MDB_SUCCESS);
7560 if (cdst->mc_ki[cdst->mc_top] == 0) {
7561 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7562 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7563 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7565 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7566 key.mv_size = NODEKSZ(srcnode);
7567 key.mv_data = NODEKEY(srcnode);
7569 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7570 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7571 mdb_cursor_copy(cdst, &mn);
7574 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7577 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7579 indx_t ix = cdst->mc_ki[cdst->mc_top];
7580 nullkey.mv_size = 0;
7581 cdst->mc_ki[cdst->mc_top] = 0;
7582 rc = mdb_update_key(cdst, &nullkey);
7583 cdst->mc_ki[cdst->mc_top] = ix;
7584 mdb_cassert(cdst, rc == MDB_SUCCESS);
7591 /** Merge one page into another.
7592 * The nodes from the page pointed to by \b csrc will
7593 * be copied to the page pointed to by \b cdst and then
7594 * the \b csrc page will be freed.
7595 * @param[in] csrc Cursor pointing to the source page.
7596 * @param[in] cdst Cursor pointing to the destination page.
7597 * @return 0 on success, non-zero on failure.
7600 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7602 MDB_page *psrc, *pdst;
7609 psrc = csrc->mc_pg[csrc->mc_top];
7610 pdst = cdst->mc_pg[cdst->mc_top];
7612 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7614 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7615 mdb_cassert(csrc, cdst->mc_snum > 1);
7617 /* Mark dst as dirty. */
7618 if ((rc = mdb_page_touch(cdst)))
7621 /* Move all nodes from src to dst.
7623 j = nkeys = NUMKEYS(pdst);
7624 if (IS_LEAF2(psrc)) {
7625 key.mv_size = csrc->mc_db->md_pad;
7626 key.mv_data = METADATA(psrc);
7627 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7628 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7629 if (rc != MDB_SUCCESS)
7631 key.mv_data = (char *)key.mv_data + key.mv_size;
7634 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7635 srcnode = NODEPTR(psrc, i);
7636 if (i == 0 && IS_BRANCH(psrc)) {
7639 mdb_cursor_copy(csrc, &mn);
7640 /* must find the lowest key below src */
7641 rc = mdb_page_search_lowest(&mn);
7644 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7645 key.mv_size = mn.mc_db->md_pad;
7646 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7648 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7649 key.mv_size = NODEKSZ(s2);
7650 key.mv_data = NODEKEY(s2);
7653 key.mv_size = srcnode->mn_ksize;
7654 key.mv_data = NODEKEY(srcnode);
7657 data.mv_size = NODEDSZ(srcnode);
7658 data.mv_data = NODEDATA(srcnode);
7659 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7660 if (rc != MDB_SUCCESS)
7665 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7666 pdst->mp_pgno, NUMKEYS(pdst),
7667 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7669 /* Unlink the src page from parent and add to free list.
7672 mdb_node_del(csrc, 0);
7673 if (csrc->mc_ki[csrc->mc_top] == 0) {
7675 rc = mdb_update_key(csrc, &key);
7683 psrc = csrc->mc_pg[csrc->mc_top];
7684 /* If not operating on FreeDB, allow this page to be reused
7685 * in this txn. Otherwise just add to free list.
7687 rc = mdb_page_loose(csrc, psrc);
7691 csrc->mc_db->md_leaf_pages--;
7693 csrc->mc_db->md_branch_pages--;
7695 /* Adjust other cursors pointing to mp */
7696 MDB_cursor *m2, *m3;
7697 MDB_dbi dbi = csrc->mc_dbi;
7699 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7700 if (csrc->mc_flags & C_SUB)
7701 m3 = &m2->mc_xcursor->mx_cursor;
7704 if (m3 == csrc) continue;
7705 if (m3->mc_snum < csrc->mc_snum) continue;
7706 if (m3->mc_pg[csrc->mc_top] == psrc) {
7707 m3->mc_pg[csrc->mc_top] = pdst;
7708 m3->mc_ki[csrc->mc_top] += nkeys;
7713 unsigned int snum = cdst->mc_snum;
7714 uint16_t depth = cdst->mc_db->md_depth;
7715 mdb_cursor_pop(cdst);
7716 rc = mdb_rebalance(cdst);
7717 /* Did the tree shrink? */
7718 if (depth > cdst->mc_db->md_depth)
7720 cdst->mc_snum = snum;
7721 cdst->mc_top = snum-1;
7726 /** Copy the contents of a cursor.
7727 * @param[in] csrc The cursor to copy from.
7728 * @param[out] cdst The cursor to copy to.
7731 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7735 cdst->mc_txn = csrc->mc_txn;
7736 cdst->mc_dbi = csrc->mc_dbi;
7737 cdst->mc_db = csrc->mc_db;
7738 cdst->mc_dbx = csrc->mc_dbx;
7739 cdst->mc_snum = csrc->mc_snum;
7740 cdst->mc_top = csrc->mc_top;
7741 cdst->mc_flags = csrc->mc_flags;
7743 for (i=0; i<csrc->mc_snum; i++) {
7744 cdst->mc_pg[i] = csrc->mc_pg[i];
7745 cdst->mc_ki[i] = csrc->mc_ki[i];
7749 /** Rebalance the tree after a delete operation.
7750 * @param[in] mc Cursor pointing to the page where rebalancing
7752 * @return 0 on success, non-zero on failure.
7755 mdb_rebalance(MDB_cursor *mc)
7759 unsigned int ptop, minkeys;
7763 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7764 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7765 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7766 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7767 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7769 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7770 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7771 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7772 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7776 if (mc->mc_snum < 2) {
7777 MDB_page *mp = mc->mc_pg[0];
7779 DPUTS("Can't rebalance a subpage, ignoring");
7782 if (NUMKEYS(mp) == 0) {
7783 DPUTS("tree is completely empty");
7784 mc->mc_db->md_root = P_INVALID;
7785 mc->mc_db->md_depth = 0;
7786 mc->mc_db->md_leaf_pages = 0;
7787 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7790 /* Adjust cursors pointing to mp */
7793 mc->mc_flags &= ~C_INITIALIZED;
7795 MDB_cursor *m2, *m3;
7796 MDB_dbi dbi = mc->mc_dbi;
7798 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7799 if (mc->mc_flags & C_SUB)
7800 m3 = &m2->mc_xcursor->mx_cursor;
7803 if (m3->mc_snum < mc->mc_snum) continue;
7804 if (m3->mc_pg[0] == mp) {
7807 m3->mc_flags &= ~C_INITIALIZED;
7811 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7813 DPUTS("collapsing root page!");
7814 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7817 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7818 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7821 mc->mc_db->md_depth--;
7822 mc->mc_db->md_branch_pages--;
7823 mc->mc_ki[0] = mc->mc_ki[1];
7824 for (i = 1; i<mc->mc_db->md_depth; i++) {
7825 mc->mc_pg[i] = mc->mc_pg[i+1];
7826 mc->mc_ki[i] = mc->mc_ki[i+1];
7829 /* Adjust other cursors pointing to mp */
7830 MDB_cursor *m2, *m3;
7831 MDB_dbi dbi = mc->mc_dbi;
7833 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7834 if (mc->mc_flags & C_SUB)
7835 m3 = &m2->mc_xcursor->mx_cursor;
7838 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7839 if (m3->mc_pg[0] == mp) {
7842 for (i=0; i<m3->mc_snum; i++) {
7843 m3->mc_pg[i] = m3->mc_pg[i+1];
7844 m3->mc_ki[i] = m3->mc_ki[i+1];
7850 DPUTS("root page doesn't need rebalancing");
7854 /* The parent (branch page) must have at least 2 pointers,
7855 * otherwise the tree is invalid.
7857 ptop = mc->mc_top-1;
7858 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7860 /* Leaf page fill factor is below the threshold.
7861 * Try to move keys from left or right neighbor, or
7862 * merge with a neighbor page.
7867 mdb_cursor_copy(mc, &mn);
7868 mn.mc_xcursor = NULL;
7870 oldki = mc->mc_ki[mc->mc_top];
7871 if (mc->mc_ki[ptop] == 0) {
7872 /* We're the leftmost leaf in our parent.
7874 DPUTS("reading right neighbor");
7876 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7877 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7880 mn.mc_ki[mn.mc_top] = 0;
7881 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7883 /* There is at least one neighbor to the left.
7885 DPUTS("reading left neighbor");
7887 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7888 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7891 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7892 mc->mc_ki[mc->mc_top] = 0;
7895 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7896 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7897 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7899 /* If the neighbor page is above threshold and has enough keys,
7900 * move one key from it. Otherwise we should try to merge them.
7901 * (A branch page must never have less than 2 keys.)
7903 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7904 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7905 rc = mdb_node_move(&mn, mc);
7906 if (mc->mc_ki[ptop]) {
7910 if (mc->mc_ki[ptop] == 0) {
7911 rc = mdb_page_merge(&mn, mc);
7913 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7914 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7915 rc = mdb_page_merge(mc, &mn);
7916 mdb_cursor_copy(&mn, mc);
7918 mc->mc_flags &= ~C_EOF;
7920 mc->mc_ki[mc->mc_top] = oldki;
7924 /** Complete a delete operation started by #mdb_cursor_del(). */
7926 mdb_cursor_del0(MDB_cursor *mc)
7933 ki = mc->mc_ki[mc->mc_top];
7934 mdb_node_del(mc, mc->mc_db->md_pad);
7935 mc->mc_db->md_entries--;
7936 rc = mdb_rebalance(mc);
7938 if (rc == MDB_SUCCESS) {
7939 MDB_cursor *m2, *m3;
7940 MDB_dbi dbi = mc->mc_dbi;
7942 mp = mc->mc_pg[mc->mc_top];
7943 nkeys = NUMKEYS(mp);
7945 /* if mc points past last node in page, find next sibling */
7946 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7947 rc = mdb_cursor_sibling(mc, 1);
7948 if (rc == MDB_NOTFOUND) {
7949 mc->mc_flags |= C_EOF;
7954 /* Adjust other cursors pointing to mp */
7955 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7956 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7957 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7959 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7961 if (m3->mc_pg[mc->mc_top] == mp) {
7962 if (m3->mc_ki[mc->mc_top] >= ki) {
7963 m3->mc_flags |= C_DEL;
7964 if (m3->mc_ki[mc->mc_top] > ki)
7965 m3->mc_ki[mc->mc_top]--;
7966 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7967 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7969 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7970 rc = mdb_cursor_sibling(m3, 1);
7971 if (rc == MDB_NOTFOUND) {
7972 m3->mc_flags |= C_EOF;
7978 mc->mc_flags |= C_DEL;
7982 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7987 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7988 MDB_val *key, MDB_val *data)
7990 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7993 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7994 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7996 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7997 /* must ignore any data */
8001 return mdb_del0(txn, dbi, key, data, 0);
8005 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8006 MDB_val *key, MDB_val *data, unsigned flags)
8011 MDB_val rdata, *xdata;
8015 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8017 mdb_cursor_init(&mc, txn, dbi, &mx);
8026 flags |= MDB_NODUPDATA;
8028 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8030 /* let mdb_page_split know about this cursor if needed:
8031 * delete will trigger a rebalance; if it needs to move
8032 * a node from one page to another, it will have to
8033 * update the parent's separator key(s). If the new sepkey
8034 * is larger than the current one, the parent page may
8035 * run out of space, triggering a split. We need this
8036 * cursor to be consistent until the end of the rebalance.
8038 mc.mc_flags |= C_UNTRACK;
8039 mc.mc_next = txn->mt_cursors[dbi];
8040 txn->mt_cursors[dbi] = &mc;
8041 rc = mdb_cursor_del(&mc, flags);
8042 txn->mt_cursors[dbi] = mc.mc_next;
8047 /** Split a page and insert a new node.
8048 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8049 * The cursor will be updated to point to the actual page and index where
8050 * the node got inserted after the split.
8051 * @param[in] newkey The key for the newly inserted node.
8052 * @param[in] newdata The data for the newly inserted node.
8053 * @param[in] newpgno The page number, if the new node is a branch node.
8054 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8055 * @return 0 on success, non-zero on failure.
8058 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8059 unsigned int nflags)
8062 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8065 int i, j, split_indx, nkeys, pmax;
8066 MDB_env *env = mc->mc_txn->mt_env;
8068 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8069 MDB_page *copy = NULL;
8070 MDB_page *mp, *rp, *pp;
8075 mp = mc->mc_pg[mc->mc_top];
8076 newindx = mc->mc_ki[mc->mc_top];
8077 nkeys = NUMKEYS(mp);
8079 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8080 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8081 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8083 /* Create a right sibling. */
8084 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8086 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8088 if (mc->mc_snum < 2) {
8089 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8091 /* shift current top to make room for new parent */
8092 mc->mc_pg[1] = mc->mc_pg[0];
8093 mc->mc_ki[1] = mc->mc_ki[0];
8096 mc->mc_db->md_root = pp->mp_pgno;
8097 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8098 mc->mc_db->md_depth++;
8101 /* Add left (implicit) pointer. */
8102 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8103 /* undo the pre-push */
8104 mc->mc_pg[0] = mc->mc_pg[1];
8105 mc->mc_ki[0] = mc->mc_ki[1];
8106 mc->mc_db->md_root = mp->mp_pgno;
8107 mc->mc_db->md_depth--;
8114 ptop = mc->mc_top-1;
8115 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8118 mc->mc_flags |= C_SPLITTING;
8119 mdb_cursor_copy(mc, &mn);
8120 mn.mc_pg[mn.mc_top] = rp;
8121 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8123 if (nflags & MDB_APPEND) {
8124 mn.mc_ki[mn.mc_top] = 0;
8126 split_indx = newindx;
8130 split_indx = (nkeys+1) / 2;
8135 unsigned int lsize, rsize, ksize;
8136 /* Move half of the keys to the right sibling */
8137 x = mc->mc_ki[mc->mc_top] - split_indx;
8138 ksize = mc->mc_db->md_pad;
8139 split = LEAF2KEY(mp, split_indx, ksize);
8140 rsize = (nkeys - split_indx) * ksize;
8141 lsize = (nkeys - split_indx) * sizeof(indx_t);
8142 mp->mp_lower -= lsize;
8143 rp->mp_lower += lsize;
8144 mp->mp_upper += rsize - lsize;
8145 rp->mp_upper -= rsize - lsize;
8146 sepkey.mv_size = ksize;
8147 if (newindx == split_indx) {
8148 sepkey.mv_data = newkey->mv_data;
8150 sepkey.mv_data = split;
8153 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8154 memcpy(rp->mp_ptrs, split, rsize);
8155 sepkey.mv_data = rp->mp_ptrs;
8156 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8157 memcpy(ins, newkey->mv_data, ksize);
8158 mp->mp_lower += sizeof(indx_t);
8159 mp->mp_upper -= ksize - sizeof(indx_t);
8162 memcpy(rp->mp_ptrs, split, x * ksize);
8163 ins = LEAF2KEY(rp, x, ksize);
8164 memcpy(ins, newkey->mv_data, ksize);
8165 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8166 rp->mp_lower += sizeof(indx_t);
8167 rp->mp_upper -= ksize - sizeof(indx_t);
8168 mc->mc_ki[mc->mc_top] = x;
8169 mc->mc_pg[mc->mc_top] = rp;
8172 int psize, nsize, k;
8173 /* Maximum free space in an empty page */
8174 pmax = env->me_psize - PAGEHDRSZ;
8176 nsize = mdb_leaf_size(env, newkey, newdata);
8178 nsize = mdb_branch_size(env, newkey);
8179 nsize = EVEN(nsize);
8181 /* grab a page to hold a temporary copy */
8182 copy = mdb_page_malloc(mc->mc_txn, 1);
8187 copy->mp_pgno = mp->mp_pgno;
8188 copy->mp_flags = mp->mp_flags;
8189 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8190 copy->mp_upper = env->me_psize - PAGEBASE;
8192 /* prepare to insert */
8193 for (i=0, j=0; i<nkeys; i++) {
8195 copy->mp_ptrs[j++] = 0;
8197 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8200 /* When items are relatively large the split point needs
8201 * to be checked, because being off-by-one will make the
8202 * difference between success or failure in mdb_node_add.
8204 * It's also relevant if a page happens to be laid out
8205 * such that one half of its nodes are all "small" and
8206 * the other half of its nodes are "large." If the new
8207 * item is also "large" and falls on the half with
8208 * "large" nodes, it also may not fit.
8210 * As a final tweak, if the new item goes on the last
8211 * spot on the page (and thus, onto the new page), bias
8212 * the split so the new page is emptier than the old page.
8213 * This yields better packing during sequential inserts.
8215 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8216 /* Find split point */
8218 if (newindx <= split_indx || newindx >= nkeys) {
8220 k = newindx >= nkeys ? nkeys : split_indx+2;
8225 for (; i!=k; i+=j) {
8230 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8231 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8233 if (F_ISSET(node->mn_flags, F_BIGDATA))
8234 psize += sizeof(pgno_t);
8236 psize += NODEDSZ(node);
8238 psize = EVEN(psize);
8240 if (psize > pmax || i == k-j) {
8241 split_indx = i + (j<0);
8246 if (split_indx == newindx) {
8247 sepkey.mv_size = newkey->mv_size;
8248 sepkey.mv_data = newkey->mv_data;
8250 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8251 sepkey.mv_size = node->mn_ksize;
8252 sepkey.mv_data = NODEKEY(node);
8257 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8259 /* Copy separator key to the parent.
8261 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8265 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8270 if (mn.mc_snum == mc->mc_snum) {
8271 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8272 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8273 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8274 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8279 /* Right page might now have changed parent.
8280 * Check if left page also changed parent.
8282 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8283 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8284 for (i=0; i<ptop; i++) {
8285 mc->mc_pg[i] = mn.mc_pg[i];
8286 mc->mc_ki[i] = mn.mc_ki[i];
8288 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8289 if (mn.mc_ki[ptop]) {
8290 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8292 /* find right page's left sibling */
8293 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8294 mdb_cursor_sibling(mc, 0);
8299 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8302 mc->mc_flags ^= C_SPLITTING;
8303 if (rc != MDB_SUCCESS) {
8306 if (nflags & MDB_APPEND) {
8307 mc->mc_pg[mc->mc_top] = rp;
8308 mc->mc_ki[mc->mc_top] = 0;
8309 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8312 for (i=0; i<mc->mc_top; i++)
8313 mc->mc_ki[i] = mn.mc_ki[i];
8314 } else if (!IS_LEAF2(mp)) {
8316 mc->mc_pg[mc->mc_top] = rp;
8321 rkey.mv_data = newkey->mv_data;
8322 rkey.mv_size = newkey->mv_size;
8328 /* Update index for the new key. */
8329 mc->mc_ki[mc->mc_top] = j;
8331 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8332 rkey.mv_data = NODEKEY(node);
8333 rkey.mv_size = node->mn_ksize;
8335 xdata.mv_data = NODEDATA(node);
8336 xdata.mv_size = NODEDSZ(node);
8339 pgno = NODEPGNO(node);
8340 flags = node->mn_flags;
8343 if (!IS_LEAF(mp) && j == 0) {
8344 /* First branch index doesn't need key data. */
8348 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8354 mc->mc_pg[mc->mc_top] = copy;
8359 } while (i != split_indx);
8361 nkeys = NUMKEYS(copy);
8362 for (i=0; i<nkeys; i++)
8363 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8364 mp->mp_lower = copy->mp_lower;
8365 mp->mp_upper = copy->mp_upper;
8366 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8367 env->me_psize - copy->mp_upper - PAGEBASE);
8369 /* reset back to original page */
8370 if (newindx < split_indx) {
8371 mc->mc_pg[mc->mc_top] = mp;
8372 if (nflags & MDB_RESERVE) {
8373 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8374 if (!(node->mn_flags & F_BIGDATA))
8375 newdata->mv_data = NODEDATA(node);
8378 mc->mc_pg[mc->mc_top] = rp;
8380 /* Make sure mc_ki is still valid.
8382 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8383 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8384 for (i=0; i<=ptop; i++) {
8385 mc->mc_pg[i] = mn.mc_pg[i];
8386 mc->mc_ki[i] = mn.mc_ki[i];
8393 /* Adjust other cursors pointing to mp */
8394 MDB_cursor *m2, *m3;
8395 MDB_dbi dbi = mc->mc_dbi;
8396 int fixup = NUMKEYS(mp);
8398 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8399 if (mc->mc_flags & C_SUB)
8400 m3 = &m2->mc_xcursor->mx_cursor;
8405 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8407 if (m3->mc_flags & C_SPLITTING)
8412 for (k=m3->mc_top; k>=0; k--) {
8413 m3->mc_ki[k+1] = m3->mc_ki[k];
8414 m3->mc_pg[k+1] = m3->mc_pg[k];
8416 if (m3->mc_ki[0] >= split_indx) {
8421 m3->mc_pg[0] = mc->mc_pg[0];
8425 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8426 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8427 m3->mc_ki[mc->mc_top]++;
8428 if (m3->mc_ki[mc->mc_top] >= fixup) {
8429 m3->mc_pg[mc->mc_top] = rp;
8430 m3->mc_ki[mc->mc_top] -= fixup;
8431 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8433 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8434 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8439 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8442 if (copy) /* tmp page */
8443 mdb_page_free(env, copy);
8445 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8450 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8451 MDB_val *key, MDB_val *data, unsigned int flags)
8456 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8459 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8462 mdb_cursor_init(&mc, txn, dbi, &mx);
8463 return mdb_cursor_put(&mc, key, data, flags);
8467 #define MDB_WBUF (1024*1024)
8470 /** State needed for a compacting copy. */
8471 typedef struct mdb_copy {
8472 pthread_mutex_t mc_mutex;
8473 pthread_cond_t mc_cond;
8480 pgno_t mc_next_pgno;
8483 volatile int mc_new;
8488 /** Dedicated writer thread for compacting copy. */
8489 static THREAD_RET ESECT
8490 mdb_env_copythr(void *arg)
8494 int toggle = 0, wsize, rc;
8497 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8500 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8503 pthread_mutex_lock(&my->mc_mutex);
8505 pthread_cond_signal(&my->mc_cond);
8508 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8509 if (my->mc_new < 0) {
8514 wsize = my->mc_wlen[toggle];
8515 ptr = my->mc_wbuf[toggle];
8518 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8522 } else if (len > 0) {
8536 /* If there's an overflow page tail, write it too */
8537 if (my->mc_olen[toggle]) {
8538 wsize = my->mc_olen[toggle];
8539 ptr = my->mc_over[toggle];
8540 my->mc_olen[toggle] = 0;
8543 my->mc_wlen[toggle] = 0;
8545 pthread_cond_signal(&my->mc_cond);
8547 pthread_cond_signal(&my->mc_cond);
8548 pthread_mutex_unlock(&my->mc_mutex);
8549 return (THREAD_RET)0;
8553 /** Tell the writer thread there's a buffer ready to write */
8555 mdb_env_cthr_toggle(mdb_copy *my, int st)
8557 int toggle = my->mc_toggle ^ 1;
8558 pthread_mutex_lock(&my->mc_mutex);
8559 if (my->mc_status) {
8560 pthread_mutex_unlock(&my->mc_mutex);
8561 return my->mc_status;
8563 while (my->mc_new == 1)
8564 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8566 my->mc_toggle = toggle;
8567 pthread_cond_signal(&my->mc_cond);
8568 pthread_mutex_unlock(&my->mc_mutex);
8572 /** Depth-first tree traversal for compacting copy. */
8574 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8577 MDB_txn *txn = my->mc_txn;
8579 MDB_page *mo, *mp, *leaf;
8584 /* Empty DB, nothing to do */
8585 if (*pg == P_INVALID)
8592 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8595 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8599 /* Make cursor pages writable */
8600 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8604 for (i=0; i<mc.mc_top; i++) {
8605 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8606 mc.mc_pg[i] = (MDB_page *)ptr;
8607 ptr += my->mc_env->me_psize;
8610 /* This is writable space for a leaf page. Usually not needed. */
8611 leaf = (MDB_page *)ptr;
8613 toggle = my->mc_toggle;
8614 while (mc.mc_snum > 0) {
8616 mp = mc.mc_pg[mc.mc_top];
8620 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8621 for (i=0; i<n; i++) {
8622 ni = NODEPTR(mp, i);
8623 if (ni->mn_flags & F_BIGDATA) {
8627 /* Need writable leaf */
8629 mc.mc_pg[mc.mc_top] = leaf;
8630 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8632 ni = NODEPTR(mp, i);
8635 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8636 rc = mdb_page_get(txn, pg, &omp, NULL);
8639 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8640 rc = mdb_env_cthr_toggle(my, 1);
8643 toggle = my->mc_toggle;
8645 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8646 memcpy(mo, omp, my->mc_env->me_psize);
8647 mo->mp_pgno = my->mc_next_pgno;
8648 my->mc_next_pgno += omp->mp_pages;
8649 my->mc_wlen[toggle] += my->mc_env->me_psize;
8650 if (omp->mp_pages > 1) {
8651 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8652 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8653 rc = mdb_env_cthr_toggle(my, 1);
8656 toggle = my->mc_toggle;
8658 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8659 } else if (ni->mn_flags & F_SUBDATA) {
8662 /* Need writable leaf */
8664 mc.mc_pg[mc.mc_top] = leaf;
8665 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8667 ni = NODEPTR(mp, i);
8670 memcpy(&db, NODEDATA(ni), sizeof(db));
8671 my->mc_toggle = toggle;
8672 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8675 toggle = my->mc_toggle;
8676 memcpy(NODEDATA(ni), &db, sizeof(db));
8681 mc.mc_ki[mc.mc_top]++;
8682 if (mc.mc_ki[mc.mc_top] < n) {
8685 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8687 rc = mdb_page_get(txn, pg, &mp, NULL);
8692 mc.mc_ki[mc.mc_top] = 0;
8693 if (IS_BRANCH(mp)) {
8694 /* Whenever we advance to a sibling branch page,
8695 * we must proceed all the way down to its first leaf.
8697 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8700 mc.mc_pg[mc.mc_top] = mp;
8704 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8705 rc = mdb_env_cthr_toggle(my, 1);
8708 toggle = my->mc_toggle;
8710 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8711 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8712 mo->mp_pgno = my->mc_next_pgno++;
8713 my->mc_wlen[toggle] += my->mc_env->me_psize;
8715 /* Update parent if there is one */
8716 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8717 SETPGNO(ni, mo->mp_pgno);
8718 mdb_cursor_pop(&mc);
8720 /* Otherwise we're done */
8730 /** Copy environment with compaction. */
8732 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8737 MDB_txn *txn = NULL;
8742 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8743 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8744 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8745 if (my.mc_wbuf[0] == NULL)
8748 pthread_mutex_init(&my.mc_mutex, NULL);
8749 pthread_cond_init(&my.mc_cond, NULL);
8750 #ifdef HAVE_MEMALIGN
8751 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8752 if (my.mc_wbuf[0] == NULL)
8755 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8760 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8761 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8766 my.mc_next_pgno = 2;
8772 THREAD_CREATE(thr, mdb_env_copythr, &my);
8774 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8778 mp = (MDB_page *)my.mc_wbuf[0];
8779 memset(mp, 0, 2*env->me_psize);
8781 mp->mp_flags = P_META;
8782 mm = (MDB_meta *)METADATA(mp);
8783 mdb_env_init_meta0(env, mm);
8784 mm->mm_address = env->me_metas[0]->mm_address;
8786 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8788 mp->mp_flags = P_META;
8789 *(MDB_meta *)METADATA(mp) = *mm;
8790 mm = (MDB_meta *)METADATA(mp);
8792 /* Count the number of free pages, subtract from lastpg to find
8793 * number of active pages
8796 MDB_ID freecount = 0;
8799 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8800 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8801 freecount += *(MDB_ID *)data.mv_data;
8802 freecount += txn->mt_dbs[0].md_branch_pages +
8803 txn->mt_dbs[0].md_leaf_pages +
8804 txn->mt_dbs[0].md_overflow_pages;
8806 /* Set metapage 1 */
8807 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8808 mm->mm_dbs[1] = txn->mt_dbs[1];
8809 if (mm->mm_last_pg > 1) {
8810 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8813 mm->mm_dbs[1].md_root = P_INVALID;
8816 my.mc_wlen[0] = env->me_psize * 2;
8818 pthread_mutex_lock(&my.mc_mutex);
8820 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8821 pthread_mutex_unlock(&my.mc_mutex);
8822 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8823 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8824 rc = mdb_env_cthr_toggle(&my, 1);
8825 mdb_env_cthr_toggle(&my, -1);
8826 pthread_mutex_lock(&my.mc_mutex);
8828 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8829 pthread_mutex_unlock(&my.mc_mutex);
8834 CloseHandle(my.mc_cond);
8835 CloseHandle(my.mc_mutex);
8836 _aligned_free(my.mc_wbuf[0]);
8838 pthread_cond_destroy(&my.mc_cond);
8839 pthread_mutex_destroy(&my.mc_mutex);
8840 free(my.mc_wbuf[0]);
8845 /** Copy environment as-is. */
8847 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8849 MDB_txn *txn = NULL;
8850 mdb_mutex_t *wmutex = NULL;
8856 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8860 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8863 /* Do the lock/unlock of the reader mutex before starting the
8864 * write txn. Otherwise other read txns could block writers.
8866 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8871 /* We must start the actual read txn after blocking writers */
8872 mdb_txn_reset0(txn, "reset-stage1");
8874 /* Temporarily block writers until we snapshot the meta pages */
8875 wmutex = MDB_MUTEX(env, w);
8876 if (LOCK_MUTEX(rc, env, wmutex))
8879 rc = mdb_txn_renew0(txn);
8881 UNLOCK_MUTEX(wmutex);
8886 wsize = env->me_psize * 2;
8890 DO_WRITE(rc, fd, ptr, w2, len);
8894 } else if (len > 0) {
8900 /* Non-blocking or async handles are not supported */
8906 UNLOCK_MUTEX(wmutex);
8911 w2 = txn->mt_next_pgno * env->me_psize;
8914 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8921 if (wsize > MAX_WRITE)
8925 DO_WRITE(rc, fd, ptr, w2, len);
8929 } else if (len > 0) {
8946 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8948 if (flags & MDB_CP_COMPACT)
8949 return mdb_env_copyfd1(env, fd);
8951 return mdb_env_copyfd0(env, fd);
8955 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8957 return mdb_env_copyfd2(env, fd, 0);
8961 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8965 HANDLE newfd = INVALID_HANDLE_VALUE;
8967 if (env->me_flags & MDB_NOSUBDIR) {
8968 lpath = (char *)path;
8971 len += sizeof(DATANAME);
8972 lpath = malloc(len);
8975 sprintf(lpath, "%s" DATANAME, path);
8978 /* The destination path must exist, but the destination file must not.
8979 * We don't want the OS to cache the writes, since the source data is
8980 * already in the OS cache.
8983 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8984 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8986 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8988 if (newfd == INVALID_HANDLE_VALUE) {
8993 if (env->me_psize >= env->me_os_psize) {
8995 /* Set O_DIRECT if the file system supports it */
8996 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8997 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8999 #ifdef F_NOCACHE /* __APPLE__ */
9000 rc = fcntl(newfd, F_NOCACHE, 1);
9008 rc = mdb_env_copyfd2(env, newfd, flags);
9011 if (!(env->me_flags & MDB_NOSUBDIR))
9013 if (newfd != INVALID_HANDLE_VALUE)
9014 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9021 mdb_env_copy(MDB_env *env, const char *path)
9023 return mdb_env_copy2(env, path, 0);
9027 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9029 if ((flag & CHANGEABLE) != flag)
9032 env->me_flags |= flag;
9034 env->me_flags &= ~flag;
9039 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9044 *arg = env->me_flags;
9049 mdb_env_set_userctx(MDB_env *env, void *ctx)
9053 env->me_userctx = ctx;
9058 mdb_env_get_userctx(MDB_env *env)
9060 return env ? env->me_userctx : NULL;
9064 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9069 env->me_assert_func = func;
9075 mdb_env_get_path(MDB_env *env, const char **arg)
9080 *arg = env->me_path;
9085 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9094 /** Common code for #mdb_stat() and #mdb_env_stat().
9095 * @param[in] env the environment to operate in.
9096 * @param[in] db the #MDB_db record containing the stats to return.
9097 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9098 * @return 0, this function always succeeds.
9101 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9103 arg->ms_psize = env->me_psize;
9104 arg->ms_depth = db->md_depth;
9105 arg->ms_branch_pages = db->md_branch_pages;
9106 arg->ms_leaf_pages = db->md_leaf_pages;
9107 arg->ms_overflow_pages = db->md_overflow_pages;
9108 arg->ms_entries = db->md_entries;
9114 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9118 if (env == NULL || arg == NULL)
9121 toggle = mdb_env_pick_meta(env);
9123 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9127 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9131 if (env == NULL || arg == NULL)
9134 toggle = mdb_env_pick_meta(env);
9135 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9136 arg->me_mapsize = env->me_mapsize;
9137 arg->me_maxreaders = env->me_maxreaders;
9138 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9140 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9141 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9145 /** Set the default comparison functions for a database.
9146 * Called immediately after a database is opened to set the defaults.
9147 * The user can then override them with #mdb_set_compare() or
9148 * #mdb_set_dupsort().
9149 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9150 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9153 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9155 uint16_t f = txn->mt_dbs[dbi].md_flags;
9157 txn->mt_dbxs[dbi].md_cmp =
9158 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9159 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9161 txn->mt_dbxs[dbi].md_dcmp =
9162 !(f & MDB_DUPSORT) ? 0 :
9163 ((f & MDB_INTEGERDUP)
9164 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9165 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9168 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9174 int rc, dbflag, exact;
9175 unsigned int unused = 0, seq;
9178 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9179 mdb_default_cmp(txn, FREE_DBI);
9182 if ((flags & VALID_FLAGS) != flags)
9184 if (txn->mt_flags & MDB_TXN_ERROR)
9190 if (flags & PERSISTENT_FLAGS) {
9191 uint16_t f2 = flags & PERSISTENT_FLAGS;
9192 /* make sure flag changes get committed */
9193 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9194 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9195 txn->mt_flags |= MDB_TXN_DIRTY;
9198 mdb_default_cmp(txn, MAIN_DBI);
9202 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9203 mdb_default_cmp(txn, MAIN_DBI);
9206 /* Is the DB already open? */
9208 for (i=2; i<txn->mt_numdbs; i++) {
9209 if (!txn->mt_dbxs[i].md_name.mv_size) {
9210 /* Remember this free slot */
9211 if (!unused) unused = i;
9214 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9215 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9221 /* If no free slot and max hit, fail */
9222 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9223 return MDB_DBS_FULL;
9225 /* Cannot mix named databases with some mainDB flags */
9226 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9227 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9229 /* Find the DB info */
9230 dbflag = DB_NEW|DB_VALID;
9233 key.mv_data = (void *)name;
9234 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9235 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9236 if (rc == MDB_SUCCESS) {
9237 /* make sure this is actually a DB */
9238 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9239 if (!(node->mn_flags & F_SUBDATA))
9240 return MDB_INCOMPATIBLE;
9241 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9242 /* Create if requested */
9243 data.mv_size = sizeof(MDB_db);
9244 data.mv_data = &dummy;
9245 memset(&dummy, 0, sizeof(dummy));
9246 dummy.md_root = P_INVALID;
9247 dummy.md_flags = flags & PERSISTENT_FLAGS;
9248 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9252 /* OK, got info, add to table */
9253 if (rc == MDB_SUCCESS) {
9254 unsigned int slot = unused ? unused : txn->mt_numdbs;
9255 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9256 txn->mt_dbxs[slot].md_name.mv_size = len;
9257 txn->mt_dbxs[slot].md_rel = NULL;
9258 txn->mt_dbflags[slot] = dbflag;
9259 /* txn-> and env-> are the same in read txns, use
9260 * tmp variable to avoid undefined assignment
9262 seq = ++txn->mt_env->me_dbiseqs[slot];
9263 txn->mt_dbiseqs[slot] = seq;
9265 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9267 mdb_default_cmp(txn, slot);
9276 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9278 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9281 if (txn->mt_flags & MDB_TXN_ERROR)
9284 if (txn->mt_dbflags[dbi] & DB_STALE) {
9287 /* Stale, must read the DB's root. cursor_init does it for us. */
9288 mdb_cursor_init(&mc, txn, dbi, &mx);
9290 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9293 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9296 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9298 ptr = env->me_dbxs[dbi].md_name.mv_data;
9299 /* If there was no name, this was already closed */
9301 env->me_dbxs[dbi].md_name.mv_data = NULL;
9302 env->me_dbxs[dbi].md_name.mv_size = 0;
9303 env->me_dbflags[dbi] = 0;
9304 env->me_dbiseqs[dbi]++;
9309 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9311 /* We could return the flags for the FREE_DBI too but what's the point? */
9312 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9314 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9318 /** Add all the DB's pages to the free list.
9319 * @param[in] mc Cursor on the DB to free.
9320 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9321 * @return 0 on success, non-zero on failure.
9324 mdb_drop0(MDB_cursor *mc, int subs)
9328 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9329 if (rc == MDB_SUCCESS) {
9330 MDB_txn *txn = mc->mc_txn;
9335 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9336 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9339 mdb_cursor_copy(mc, &mx);
9340 while (mc->mc_snum > 0) {
9341 MDB_page *mp = mc->mc_pg[mc->mc_top];
9342 unsigned n = NUMKEYS(mp);
9344 for (i=0; i<n; i++) {
9345 ni = NODEPTR(mp, i);
9346 if (ni->mn_flags & F_BIGDATA) {
9349 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9350 rc = mdb_page_get(txn, pg, &omp, NULL);
9353 mdb_cassert(mc, IS_OVERFLOW(omp));
9354 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9358 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9359 mdb_xcursor_init1(mc, ni);
9360 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9366 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9368 for (i=0; i<n; i++) {
9370 ni = NODEPTR(mp, i);
9373 mdb_midl_xappend(txn->mt_free_pgs, pg);
9378 mc->mc_ki[mc->mc_top] = i;
9379 rc = mdb_cursor_sibling(mc, 1);
9381 if (rc != MDB_NOTFOUND)
9383 /* no more siblings, go back to beginning
9384 * of previous level.
9388 for (i=1; i<mc->mc_snum; i++) {
9390 mc->mc_pg[i] = mx.mc_pg[i];
9395 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9398 txn->mt_flags |= MDB_TXN_ERROR;
9399 } else if (rc == MDB_NOTFOUND) {
9405 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9407 MDB_cursor *mc, *m2;
9410 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9413 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9416 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9419 rc = mdb_cursor_open(txn, dbi, &mc);
9423 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9424 /* Invalidate the dropped DB's cursors */
9425 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9426 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9430 /* Can't delete the main DB */
9431 if (del && dbi > MAIN_DBI) {
9432 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9434 txn->mt_dbflags[dbi] = DB_STALE;
9435 mdb_dbi_close(txn->mt_env, dbi);
9437 txn->mt_flags |= MDB_TXN_ERROR;
9440 /* reset the DB record, mark it dirty */
9441 txn->mt_dbflags[dbi] |= DB_DIRTY;
9442 txn->mt_dbs[dbi].md_depth = 0;
9443 txn->mt_dbs[dbi].md_branch_pages = 0;
9444 txn->mt_dbs[dbi].md_leaf_pages = 0;
9445 txn->mt_dbs[dbi].md_overflow_pages = 0;
9446 txn->mt_dbs[dbi].md_entries = 0;
9447 txn->mt_dbs[dbi].md_root = P_INVALID;
9449 txn->mt_flags |= MDB_TXN_DIRTY;
9452 mdb_cursor_close(mc);
9456 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9458 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9461 txn->mt_dbxs[dbi].md_cmp = cmp;
9465 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9467 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9470 txn->mt_dbxs[dbi].md_dcmp = cmp;
9474 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9476 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9479 txn->mt_dbxs[dbi].md_rel = rel;
9483 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9485 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9488 txn->mt_dbxs[dbi].md_relctx = ctx;
9493 mdb_env_get_maxkeysize(MDB_env *env)
9495 return ENV_MAXKEY(env);
9499 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9501 unsigned int i, rdrs;
9504 int rc = 0, first = 1;
9508 if (!env->me_txns) {
9509 return func("(no reader locks)\n", ctx);
9511 rdrs = env->me_txns->mti_numreaders;
9512 mr = env->me_txns->mti_readers;
9513 for (i=0; i<rdrs; i++) {
9515 txnid_t txnid = mr[i].mr_txnid;
9516 sprintf(buf, txnid == (txnid_t)-1 ?
9517 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9518 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9521 rc = func(" pid thread txnid\n", ctx);
9525 rc = func(buf, ctx);
9531 rc = func("(no active readers)\n", ctx);
9536 /** Insert pid into list if not already present.
9537 * return -1 if already present.
9540 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9542 /* binary search of pid in list */
9544 unsigned cursor = 1;
9546 unsigned n = ids[0];
9549 unsigned pivot = n >> 1;
9550 cursor = base + pivot + 1;
9551 val = pid - ids[cursor];
9556 } else if ( val > 0 ) {
9561 /* found, so it's a duplicate */
9570 for (n = ids[0]; n > cursor; n--)
9577 mdb_reader_check(MDB_env *env, int *dead)
9583 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9586 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9587 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9589 mdb_mutex_t *rmutex = rlocked ? NULL : MDB_MUTEX(env, r);
9590 unsigned int i, j, rdrs;
9592 MDB_PID_T *pids, pid;
9593 int rc = MDB_SUCCESS, count = 0;
9595 rdrs = env->me_txns->mti_numreaders;
9596 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9600 mr = env->me_txns->mti_readers;
9601 for (i=0; i<rdrs; i++) {
9603 if (pid && pid != env->me_pid) {
9604 if (mdb_pid_insert(pids, pid) == 0) {
9605 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9606 /* Stale reader found */
9609 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9610 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9612 rdrs = 0; /* the above checked all readers */
9614 /* Recheck, a new process may have reused pid */
9615 if (mdb_reader_pid(env, Pidcheck, pid))
9620 if (mr[j].mr_pid == pid) {
9621 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9622 (unsigned) pid, mr[j].mr_txnid));
9627 UNLOCK_MUTEX(rmutex);
9638 #ifdef MDB_ROBUST_SUPPORTED
9639 /** Handle #LOCK_MUTEX0() failure.
9640 * Try to repair the lock file if the mutex owner died.
9641 * @param[in] env the environment handle
9642 * @param[in] mutex LOCK_MUTEX0() mutex
9643 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9644 * @return 0 on success with the mutex locked, or an error code on failure.
9646 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc)
9648 int toggle, rlocked, rc2;
9650 if (rc == MDB_OWNERDEAD) {
9651 /* We own the mutex. Clean up after dead previous owner. */
9653 rlocked = (mutex == MDB_MUTEX(env, r));
9655 /* Keep mti_txnid updated, otherwise next writer can
9656 * overwrite data which latest meta page refers to.
9658 toggle = mdb_env_pick_meta(env);
9659 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9660 /* env is hosed if the dead thread was ours */
9662 env->me_flags |= MDB_FATAL_ERROR;
9667 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9668 (rc ? "this process' env is hosed" : "recovering")));
9669 rc2 = mdb_reader_check0(env, rlocked, NULL);
9671 rc2 = mdb_mutex_consistent(mutex);
9672 if (rc || (rc = rc2)) {
9673 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9674 UNLOCK_MUTEX(mutex);
9680 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9685 #endif /* MDB_ROBUST_SUPPORTED */