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;
1331 /** Compare two items pointing at size_t's of unknown alignment. */
1332 #ifdef MISALIGNED_OK
1333 # define mdb_cmp_clong mdb_cmp_long
1335 # define mdb_cmp_clong mdb_cmp_cint
1339 static SECURITY_DESCRIPTOR mdb_null_sd;
1340 static SECURITY_ATTRIBUTES mdb_all_sa;
1341 static int mdb_sec_inited;
1344 /** Return the library version info. */
1346 mdb_version(int *major, int *minor, int *patch)
1348 if (major) *major = MDB_VERSION_MAJOR;
1349 if (minor) *minor = MDB_VERSION_MINOR;
1350 if (patch) *patch = MDB_VERSION_PATCH;
1351 return MDB_VERSION_STRING;
1354 /** Table of descriptions for LMDB @ref errors */
1355 static char *const mdb_errstr[] = {
1356 "MDB_KEYEXIST: Key/data pair already exists",
1357 "MDB_NOTFOUND: No matching key/data pair found",
1358 "MDB_PAGE_NOTFOUND: Requested page not found",
1359 "MDB_CORRUPTED: Located page was wrong type",
1360 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1361 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1362 "MDB_INVALID: File is not an LMDB file",
1363 "MDB_MAP_FULL: Environment mapsize limit reached",
1364 "MDB_DBS_FULL: Environment maxdbs limit reached",
1365 "MDB_READERS_FULL: Environment maxreaders limit reached",
1366 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1367 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1368 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1369 "MDB_PAGE_FULL: Internal error - page has no more space",
1370 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1371 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1372 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1373 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1374 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1375 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1379 mdb_strerror(int err)
1382 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1383 * This works as long as no function between the call to mdb_strerror
1384 * and the actual use of the message uses more than 4K of stack.
1387 char buf[1024], *ptr = buf;
1391 return ("Successful return: 0");
1393 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1394 i = err - MDB_KEYEXIST;
1395 return mdb_errstr[i];
1399 /* These are the C-runtime error codes we use. The comment indicates
1400 * their numeric value, and the Win32 error they would correspond to
1401 * if the error actually came from a Win32 API. A major mess, we should
1402 * have used LMDB-specific error codes for everything.
1405 case ENOENT: /* 2, FILE_NOT_FOUND */
1406 case EIO: /* 5, ACCESS_DENIED */
1407 case ENOMEM: /* 12, INVALID_ACCESS */
1408 case EACCES: /* 13, INVALID_DATA */
1409 case EBUSY: /* 16, CURRENT_DIRECTORY */
1410 case EINVAL: /* 22, BAD_COMMAND */
1411 case ENOSPC: /* 28, OUT_OF_PAPER */
1412 return strerror(err);
1417 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1418 FORMAT_MESSAGE_IGNORE_INSERTS,
1419 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1422 return strerror(err);
1426 /** assert(3) variant in cursor context */
1427 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1428 /** assert(3) variant in transaction context */
1429 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1430 /** assert(3) variant in environment context */
1431 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1434 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1435 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1438 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1439 const char *func, const char *file, int line)
1442 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1443 file, line, expr_txt, func);
1444 if (env->me_assert_func)
1445 env->me_assert_func(env, buf);
1446 fprintf(stderr, "%s\n", buf);
1450 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1454 /** Return the page number of \b mp which may be sub-page, for debug output */
1456 mdb_dbg_pgno(MDB_page *mp)
1459 COPY_PGNO(ret, mp->mp_pgno);
1463 /** Display a key in hexadecimal and return the address of the result.
1464 * @param[in] key the key to display
1465 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1466 * @return The key in hexadecimal form.
1469 mdb_dkey(MDB_val *key, char *buf)
1472 unsigned char *c = key->mv_data;
1478 if (key->mv_size > DKBUF_MAXKEYSIZE)
1479 return "MDB_MAXKEYSIZE";
1480 /* may want to make this a dynamic check: if the key is mostly
1481 * printable characters, print it as-is instead of converting to hex.
1485 for (i=0; i<key->mv_size; i++)
1486 ptr += sprintf(ptr, "%02x", *c++);
1488 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1494 mdb_leafnode_type(MDB_node *n)
1496 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1497 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1498 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1501 /** Display all the keys in the page. */
1503 mdb_page_list(MDB_page *mp)
1505 pgno_t pgno = mdb_dbg_pgno(mp);
1506 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1508 unsigned int i, nkeys, nsize, total = 0;
1512 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1513 case P_BRANCH: type = "Branch page"; break;
1514 case P_LEAF: type = "Leaf page"; break;
1515 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1516 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1517 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1519 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1520 pgno, mp->mp_pages, state);
1523 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1524 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1527 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1531 nkeys = NUMKEYS(mp);
1532 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1534 for (i=0; i<nkeys; i++) {
1535 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1536 key.mv_size = nsize = mp->mp_pad;
1537 key.mv_data = LEAF2KEY(mp, i, nsize);
1539 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1542 node = NODEPTR(mp, i);
1543 key.mv_size = node->mn_ksize;
1544 key.mv_data = node->mn_data;
1545 nsize = NODESIZE + key.mv_size;
1546 if (IS_BRANCH(mp)) {
1547 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1551 if (F_ISSET(node->mn_flags, F_BIGDATA))
1552 nsize += sizeof(pgno_t);
1554 nsize += NODEDSZ(node);
1556 nsize += sizeof(indx_t);
1557 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1558 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1560 total = EVEN(total);
1562 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1563 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1567 mdb_cursor_chk(MDB_cursor *mc)
1573 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1574 for (i=0; i<mc->mc_top; i++) {
1576 node = NODEPTR(mp, mc->mc_ki[i]);
1577 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1580 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1586 /** Count all the pages in each DB and in the freelist
1587 * and make sure it matches the actual number of pages
1589 * All named DBs must be open for a correct count.
1591 static void mdb_audit(MDB_txn *txn)
1595 MDB_ID freecount, count;
1600 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1601 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1602 freecount += *(MDB_ID *)data.mv_data;
1603 mdb_tassert(txn, rc == MDB_NOTFOUND);
1606 for (i = 0; i<txn->mt_numdbs; i++) {
1608 if (!(txn->mt_dbflags[i] & DB_VALID))
1610 mdb_cursor_init(&mc, txn, i, &mx);
1611 if (txn->mt_dbs[i].md_root == P_INVALID)
1613 count += txn->mt_dbs[i].md_branch_pages +
1614 txn->mt_dbs[i].md_leaf_pages +
1615 txn->mt_dbs[i].md_overflow_pages;
1616 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1617 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1618 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1621 mp = mc.mc_pg[mc.mc_top];
1622 for (j=0; j<NUMKEYS(mp); j++) {
1623 MDB_node *leaf = NODEPTR(mp, j);
1624 if (leaf->mn_flags & F_SUBDATA) {
1626 memcpy(&db, NODEDATA(leaf), sizeof(db));
1627 count += db.md_branch_pages + db.md_leaf_pages +
1628 db.md_overflow_pages;
1632 mdb_tassert(txn, rc == MDB_NOTFOUND);
1635 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1636 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1637 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1643 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1645 return txn->mt_dbxs[dbi].md_cmp(a, b);
1649 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1651 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1652 #if UINT_MAX < SIZE_MAX
1653 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1654 dcmp = mdb_cmp_clong;
1659 /** Allocate memory for a page.
1660 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1663 mdb_page_malloc(MDB_txn *txn, unsigned num)
1665 MDB_env *env = txn->mt_env;
1666 MDB_page *ret = env->me_dpages;
1667 size_t psize = env->me_psize, sz = psize, off;
1668 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1669 * For a single page alloc, we init everything after the page header.
1670 * For multi-page, we init the final page; if the caller needed that
1671 * many pages they will be filling in at least up to the last page.
1675 VGMEMP_ALLOC(env, ret, sz);
1676 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1677 env->me_dpages = ret->mp_next;
1680 psize -= off = PAGEHDRSZ;
1685 if ((ret = malloc(sz)) != NULL) {
1686 VGMEMP_ALLOC(env, ret, sz);
1687 if (!(env->me_flags & MDB_NOMEMINIT)) {
1688 memset((char *)ret + off, 0, psize);
1692 txn->mt_flags |= MDB_TXN_ERROR;
1696 /** Free a single page.
1697 * Saves single pages to a list, for future reuse.
1698 * (This is not used for multi-page overflow pages.)
1701 mdb_page_free(MDB_env *env, MDB_page *mp)
1703 mp->mp_next = env->me_dpages;
1704 VGMEMP_FREE(env, mp);
1705 env->me_dpages = mp;
1708 /** Free a dirty page */
1710 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1712 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1713 mdb_page_free(env, dp);
1715 /* large pages just get freed directly */
1716 VGMEMP_FREE(env, dp);
1721 /** Return all dirty pages to dpage list */
1723 mdb_dlist_free(MDB_txn *txn)
1725 MDB_env *env = txn->mt_env;
1726 MDB_ID2L dl = txn->mt_u.dirty_list;
1727 unsigned i, n = dl[0].mid;
1729 for (i = 1; i <= n; i++) {
1730 mdb_dpage_free(env, dl[i].mptr);
1735 /** Loosen or free a single page.
1736 * Saves single pages to a list for future reuse
1737 * in this same txn. It has been pulled from the freeDB
1738 * and already resides on the dirty list, but has been
1739 * deleted. Use these pages first before pulling again
1742 * If the page wasn't dirtied in this txn, just add it
1743 * to this txn's free list.
1746 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1749 pgno_t pgno = mp->mp_pgno;
1750 MDB_txn *txn = mc->mc_txn;
1752 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1753 if (txn->mt_parent) {
1754 MDB_ID2 *dl = txn->mt_u.dirty_list;
1755 /* If txn has a parent, make sure the page is in our
1759 unsigned x = mdb_mid2l_search(dl, pgno);
1760 if (x <= dl[0].mid && dl[x].mid == pgno) {
1761 if (mp != dl[x].mptr) { /* bad cursor? */
1762 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1763 txn->mt_flags |= MDB_TXN_ERROR;
1764 return MDB_CORRUPTED;
1771 /* no parent txn, so it's just ours */
1776 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1778 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1779 txn->mt_loose_pgs = mp;
1780 txn->mt_loose_count++;
1781 mp->mp_flags |= P_LOOSE;
1783 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1791 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1792 * @param[in] mc A cursor handle for the current operation.
1793 * @param[in] pflags Flags of the pages to update:
1794 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1795 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1796 * @return 0 on success, non-zero on failure.
1799 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1801 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1802 MDB_txn *txn = mc->mc_txn;
1808 int rc = MDB_SUCCESS, level;
1810 /* Mark pages seen by cursors */
1811 if (mc->mc_flags & C_UNTRACK)
1812 mc = NULL; /* will find mc in mt_cursors */
1813 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1814 for (; mc; mc=mc->mc_next) {
1815 if (!(mc->mc_flags & C_INITIALIZED))
1817 for (m3 = mc;; m3 = &mx->mx_cursor) {
1819 for (j=0; j<m3->mc_snum; j++) {
1821 if ((mp->mp_flags & Mask) == pflags)
1822 mp->mp_flags ^= P_KEEP;
1824 mx = m3->mc_xcursor;
1825 /* Proceed to mx if it is at a sub-database */
1826 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1828 if (! (mp && (mp->mp_flags & P_LEAF)))
1830 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1831 if (!(leaf->mn_flags & F_SUBDATA))
1840 /* Mark dirty root pages */
1841 for (i=0; i<txn->mt_numdbs; i++) {
1842 if (txn->mt_dbflags[i] & DB_DIRTY) {
1843 pgno_t pgno = txn->mt_dbs[i].md_root;
1844 if (pgno == P_INVALID)
1846 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1848 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1849 dp->mp_flags ^= P_KEEP;
1857 static int mdb_page_flush(MDB_txn *txn, int keep);
1859 /** Spill pages from the dirty list back to disk.
1860 * This is intended to prevent running into #MDB_TXN_FULL situations,
1861 * but note that they may still occur in a few cases:
1862 * 1) our estimate of the txn size could be too small. Currently this
1863 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1864 * 2) child txns may run out of space if their parents dirtied a
1865 * lot of pages and never spilled them. TODO: we probably should do
1866 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1867 * the parent's dirty_room is below a given threshold.
1869 * Otherwise, if not using nested txns, it is expected that apps will
1870 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1871 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1872 * If the txn never references them again, they can be left alone.
1873 * If the txn only reads them, they can be used without any fuss.
1874 * If the txn writes them again, they can be dirtied immediately without
1875 * going thru all of the work of #mdb_page_touch(). Such references are
1876 * handled by #mdb_page_unspill().
1878 * Also note, we never spill DB root pages, nor pages of active cursors,
1879 * because we'll need these back again soon anyway. And in nested txns,
1880 * we can't spill a page in a child txn if it was already spilled in a
1881 * parent txn. That would alter the parent txns' data even though
1882 * the child hasn't committed yet, and we'd have no way to undo it if
1883 * the child aborted.
1885 * @param[in] m0 cursor A cursor handle identifying the transaction and
1886 * database for which we are checking space.
1887 * @param[in] key For a put operation, the key being stored.
1888 * @param[in] data For a put operation, the data being stored.
1889 * @return 0 on success, non-zero on failure.
1892 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1894 MDB_txn *txn = m0->mc_txn;
1896 MDB_ID2L dl = txn->mt_u.dirty_list;
1897 unsigned int i, j, need;
1900 if (m0->mc_flags & C_SUB)
1903 /* Estimate how much space this op will take */
1904 i = m0->mc_db->md_depth;
1905 /* Named DBs also dirty the main DB */
1906 if (m0->mc_dbi > MAIN_DBI)
1907 i += txn->mt_dbs[MAIN_DBI].md_depth;
1908 /* For puts, roughly factor in the key+data size */
1910 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1911 i += i; /* double it for good measure */
1914 if (txn->mt_dirty_room > i)
1917 if (!txn->mt_spill_pgs) {
1918 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1919 if (!txn->mt_spill_pgs)
1922 /* purge deleted slots */
1923 MDB_IDL sl = txn->mt_spill_pgs;
1924 unsigned int num = sl[0];
1926 for (i=1; i<=num; i++) {
1933 /* Preserve pages which may soon be dirtied again */
1934 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1937 /* Less aggressive spill - we originally spilled the entire dirty list,
1938 * with a few exceptions for cursor pages and DB root pages. But this
1939 * turns out to be a lot of wasted effort because in a large txn many
1940 * of those pages will need to be used again. So now we spill only 1/8th
1941 * of the dirty pages. Testing revealed this to be a good tradeoff,
1942 * better than 1/2, 1/4, or 1/10.
1944 if (need < MDB_IDL_UM_MAX / 8)
1945 need = MDB_IDL_UM_MAX / 8;
1947 /* Save the page IDs of all the pages we're flushing */
1948 /* flush from the tail forward, this saves a lot of shifting later on. */
1949 for (i=dl[0].mid; i && need; i--) {
1950 MDB_ID pn = dl[i].mid << 1;
1952 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1954 /* Can't spill twice, make sure it's not already in a parent's
1957 if (txn->mt_parent) {
1959 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1960 if (tx2->mt_spill_pgs) {
1961 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1962 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1963 dp->mp_flags |= P_KEEP;
1971 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1975 mdb_midl_sort(txn->mt_spill_pgs);
1977 /* Flush the spilled part of dirty list */
1978 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1981 /* Reset any dirty pages we kept that page_flush didn't see */
1982 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1985 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1989 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1991 mdb_find_oldest(MDB_txn *txn)
1994 txnid_t mr, oldest = txn->mt_txnid - 1;
1995 if (txn->mt_env->me_txns) {
1996 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1997 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2008 /** Add a page to the txn's dirty list */
2010 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2013 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2015 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2016 insert = mdb_mid2l_append;
2018 insert = mdb_mid2l_insert;
2020 mid.mid = mp->mp_pgno;
2022 rc = insert(txn->mt_u.dirty_list, &mid);
2023 mdb_tassert(txn, rc == 0);
2024 txn->mt_dirty_room--;
2027 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2028 * me_pghead and mt_next_pgno.
2030 * If there are free pages available from older transactions, they
2031 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2032 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2033 * and move me_pglast to say which records were consumed. Only this
2034 * function can create me_pghead and move me_pglast/mt_next_pgno.
2035 * @param[in] mc cursor A cursor handle identifying the transaction and
2036 * database for which we are allocating.
2037 * @param[in] num the number of pages to allocate.
2038 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2039 * will always be satisfied by a single contiguous chunk of memory.
2040 * @return 0 on success, non-zero on failure.
2043 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2045 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2046 /* Get at most <Max_retries> more freeDB records once me_pghead
2047 * has enough pages. If not enough, use new pages from the map.
2048 * If <Paranoid> and mc is updating the freeDB, only get new
2049 * records if me_pghead is empty. Then the freelist cannot play
2050 * catch-up with itself by growing while trying to save it.
2052 enum { Paranoid = 1, Max_retries = 500 };
2054 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2056 int rc, retry = num * 60;
2057 MDB_txn *txn = mc->mc_txn;
2058 MDB_env *env = txn->mt_env;
2059 pgno_t pgno, *mop = env->me_pghead;
2060 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2062 txnid_t oldest = 0, last;
2067 /* If there are any loose pages, just use them */
2068 if (num == 1 && txn->mt_loose_pgs) {
2069 np = txn->mt_loose_pgs;
2070 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2071 txn->mt_loose_count--;
2072 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2080 /* If our dirty list is already full, we can't do anything */
2081 if (txn->mt_dirty_room == 0) {
2086 for (op = MDB_FIRST;; op = MDB_NEXT) {
2091 /* Seek a big enough contiguous page range. Prefer
2092 * pages at the tail, just truncating the list.
2098 if (mop[i-n2] == pgno+n2)
2105 if (op == MDB_FIRST) { /* 1st iteration */
2106 /* Prepare to fetch more and coalesce */
2107 last = env->me_pglast;
2108 oldest = env->me_pgoldest;
2109 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2112 key.mv_data = &last; /* will look up last+1 */
2113 key.mv_size = sizeof(last);
2115 if (Paranoid && mc->mc_dbi == FREE_DBI)
2118 if (Paranoid && retry < 0 && mop_len)
2122 /* Do not fetch more if the record will be too recent */
2123 if (oldest <= last) {
2125 oldest = mdb_find_oldest(txn);
2126 env->me_pgoldest = oldest;
2132 rc = mdb_cursor_get(&m2, &key, NULL, op);
2134 if (rc == MDB_NOTFOUND)
2138 last = *(txnid_t*)key.mv_data;
2139 if (oldest <= last) {
2141 oldest = mdb_find_oldest(txn);
2142 env->me_pgoldest = oldest;
2148 np = m2.mc_pg[m2.mc_top];
2149 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2150 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2153 idl = (MDB_ID *) data.mv_data;
2156 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2161 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2163 mop = env->me_pghead;
2165 env->me_pglast = last;
2167 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2168 last, txn->mt_dbs[FREE_DBI].md_root, i));
2170 DPRINTF(("IDL %"Z"u", idl[j]));
2172 /* Merge in descending sorted order */
2173 mdb_midl_xmerge(mop, idl);
2177 /* Use new pages from the map when nothing suitable in the freeDB */
2179 pgno = txn->mt_next_pgno;
2180 if (pgno + num >= env->me_maxpg) {
2181 DPUTS("DB size maxed out");
2187 if (env->me_flags & MDB_WRITEMAP) {
2188 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2190 if (!(np = mdb_page_malloc(txn, num))) {
2196 mop[0] = mop_len -= num;
2197 /* Move any stragglers down */
2198 for (j = i-num; j < mop_len; )
2199 mop[++j] = mop[++i];
2201 txn->mt_next_pgno = pgno + num;
2204 mdb_page_dirty(txn, np);
2210 txn->mt_flags |= MDB_TXN_ERROR;
2214 /** Copy the used portions of a non-overflow page.
2215 * @param[in] dst page to copy into
2216 * @param[in] src page to copy from
2217 * @param[in] psize size of a page
2220 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2222 enum { Align = sizeof(pgno_t) };
2223 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2225 /* If page isn't full, just copy the used portion. Adjust
2226 * alignment so memcpy may copy words instead of bytes.
2228 if ((unused &= -Align) && !IS_LEAF2(src)) {
2229 upper = (upper + PAGEBASE) & -Align;
2230 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2231 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2234 memcpy(dst, src, psize - unused);
2238 /** Pull a page off the txn's spill list, if present.
2239 * If a page being referenced was spilled to disk in this txn, bring
2240 * it back and make it dirty/writable again.
2241 * @param[in] txn the transaction handle.
2242 * @param[in] mp the page being referenced. It must not be dirty.
2243 * @param[out] ret the writable page, if any. ret is unchanged if
2244 * mp wasn't spilled.
2247 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2249 MDB_env *env = txn->mt_env;
2252 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2254 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2255 if (!tx2->mt_spill_pgs)
2257 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2258 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2261 if (txn->mt_dirty_room == 0)
2262 return MDB_TXN_FULL;
2263 if (IS_OVERFLOW(mp))
2267 if (env->me_flags & MDB_WRITEMAP) {
2270 np = mdb_page_malloc(txn, num);
2274 memcpy(np, mp, num * env->me_psize);
2276 mdb_page_copy(np, mp, env->me_psize);
2279 /* If in current txn, this page is no longer spilled.
2280 * If it happens to be the last page, truncate the spill list.
2281 * Otherwise mark it as deleted by setting the LSB.
2283 if (x == txn->mt_spill_pgs[0])
2284 txn->mt_spill_pgs[0]--;
2286 txn->mt_spill_pgs[x] |= 1;
2287 } /* otherwise, if belonging to a parent txn, the
2288 * page remains spilled until child commits
2291 mdb_page_dirty(txn, np);
2292 np->mp_flags |= P_DIRTY;
2300 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2301 * @param[in] mc cursor pointing to the page to be touched
2302 * @return 0 on success, non-zero on failure.
2305 mdb_page_touch(MDB_cursor *mc)
2307 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2308 MDB_txn *txn = mc->mc_txn;
2309 MDB_cursor *m2, *m3;
2313 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2314 if (txn->mt_flags & MDB_TXN_SPILLS) {
2316 rc = mdb_page_unspill(txn, mp, &np);
2322 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2323 (rc = mdb_page_alloc(mc, 1, &np)))
2326 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2327 mp->mp_pgno, pgno));
2328 mdb_cassert(mc, mp->mp_pgno != pgno);
2329 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2330 /* Update the parent page, if any, to point to the new page */
2332 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2333 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2334 SETPGNO(node, pgno);
2336 mc->mc_db->md_root = pgno;
2338 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2339 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2341 /* If txn has a parent, make sure the page is in our
2345 unsigned x = mdb_mid2l_search(dl, pgno);
2346 if (x <= dl[0].mid && dl[x].mid == pgno) {
2347 if (mp != dl[x].mptr) { /* bad cursor? */
2348 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2349 txn->mt_flags |= MDB_TXN_ERROR;
2350 return MDB_CORRUPTED;
2355 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2357 np = mdb_page_malloc(txn, 1);
2362 rc = mdb_mid2l_insert(dl, &mid);
2363 mdb_cassert(mc, rc == 0);
2368 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2370 np->mp_flags |= P_DIRTY;
2373 /* Adjust cursors pointing to mp */
2374 mc->mc_pg[mc->mc_top] = np;
2375 m2 = txn->mt_cursors[mc->mc_dbi];
2376 if (mc->mc_flags & C_SUB) {
2377 for (; m2; m2=m2->mc_next) {
2378 m3 = &m2->mc_xcursor->mx_cursor;
2379 if (m3->mc_snum < mc->mc_snum) continue;
2380 if (m3->mc_pg[mc->mc_top] == mp)
2381 m3->mc_pg[mc->mc_top] = np;
2384 for (; m2; m2=m2->mc_next) {
2385 if (m2->mc_snum < mc->mc_snum) continue;
2386 if (m2->mc_pg[mc->mc_top] == mp) {
2387 m2->mc_pg[mc->mc_top] = np;
2388 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2390 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2392 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2393 if (!(leaf->mn_flags & F_SUBDATA))
2394 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2402 txn->mt_flags |= MDB_TXN_ERROR;
2407 mdb_env_sync(MDB_env *env, int force)
2410 if (env->me_flags & MDB_RDONLY)
2412 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2413 if (env->me_flags & MDB_WRITEMAP) {
2414 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2415 ? MS_ASYNC : MS_SYNC;
2416 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2419 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2423 #ifdef BROKEN_FDATASYNC
2424 if (env->me_flags & MDB_FSYNCONLY) {
2425 if (fsync(env->me_fd))
2429 if (MDB_FDATASYNC(env->me_fd))
2436 /** Back up parent txn's cursors, then grab the originals for tracking */
2438 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2440 MDB_cursor *mc, *bk;
2445 for (i = src->mt_numdbs; --i >= 0; ) {
2446 if ((mc = src->mt_cursors[i]) != NULL) {
2447 size = sizeof(MDB_cursor);
2449 size += sizeof(MDB_xcursor);
2450 for (; mc; mc = bk->mc_next) {
2456 mc->mc_db = &dst->mt_dbs[i];
2457 /* Kill pointers into src - and dst to reduce abuse: The
2458 * user may not use mc until dst ends. Otherwise we'd...
2460 mc->mc_txn = NULL; /* ...set this to dst */
2461 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2462 if ((mx = mc->mc_xcursor) != NULL) {
2463 *(MDB_xcursor *)(bk+1) = *mx;
2464 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2466 mc->mc_next = dst->mt_cursors[i];
2467 dst->mt_cursors[i] = mc;
2474 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2475 * @param[in] txn the transaction handle.
2476 * @param[in] merge true to keep changes to parent cursors, false to revert.
2477 * @return 0 on success, non-zero on failure.
2480 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2482 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2486 for (i = txn->mt_numdbs; --i >= 0; ) {
2487 for (mc = cursors[i]; mc; mc = next) {
2489 if ((bk = mc->mc_backup) != NULL) {
2491 /* Commit changes to parent txn */
2492 mc->mc_next = bk->mc_next;
2493 mc->mc_backup = bk->mc_backup;
2494 mc->mc_txn = bk->mc_txn;
2495 mc->mc_db = bk->mc_db;
2496 mc->mc_dbflag = bk->mc_dbflag;
2497 if ((mx = mc->mc_xcursor) != NULL)
2498 mx->mx_cursor.mc_txn = bk->mc_txn;
2500 /* Abort nested txn */
2502 if ((mx = mc->mc_xcursor) != NULL)
2503 *mx = *(MDB_xcursor *)(bk+1);
2507 /* Only malloced cursors are permanently tracked. */
2515 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2518 mdb_txn_reset0(MDB_txn *txn, const char *act);
2520 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2526 Pidset = F_SETLK, Pidcheck = F_GETLK
2530 /** Set or check a pid lock. Set returns 0 on success.
2531 * Check returns 0 if the process is certainly dead, nonzero if it may
2532 * be alive (the lock exists or an error happened so we do not know).
2534 * On Windows Pidset is a no-op, we merely check for the existence
2535 * of the process with the given pid. On POSIX we use a single byte
2536 * lock on the lockfile, set at an offset equal to the pid.
2539 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2541 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2544 if (op == Pidcheck) {
2545 h = OpenProcess(env->me_pidquery, FALSE, pid);
2546 /* No documented "no such process" code, but other program use this: */
2548 return ErrCode() != ERROR_INVALID_PARAMETER;
2549 /* A process exists until all handles to it close. Has it exited? */
2550 ret = WaitForSingleObject(h, 0) != 0;
2557 struct flock lock_info;
2558 memset(&lock_info, 0, sizeof(lock_info));
2559 lock_info.l_type = F_WRLCK;
2560 lock_info.l_whence = SEEK_SET;
2561 lock_info.l_start = pid;
2562 lock_info.l_len = 1;
2563 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2564 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2566 } else if ((rc = ErrCode()) == EINTR) {
2574 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2575 * @param[in] txn the transaction handle to initialize
2576 * @return 0 on success, non-zero on failure.
2579 mdb_txn_renew0(MDB_txn *txn)
2581 MDB_env *env = txn->mt_env;
2582 MDB_txninfo *ti = env->me_txns;
2586 int rc, new_notls = 0;
2588 if (txn->mt_flags & MDB_TXN_RDONLY) {
2589 txn->mt_flags &= MDB_TXN_BEGIN_FLAGS;
2591 txn->mt_numdbs = env->me_numdbs;
2592 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2594 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2595 txn->mt_txnid = meta->mm_txnid;
2596 txn->mt_u.reader = NULL;
2598 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2599 pthread_getspecific(env->me_txkey);
2601 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2602 return MDB_BAD_RSLOT;
2604 MDB_PID_T pid = env->me_pid;
2605 MDB_THR_T tid = pthread_self();
2606 mdb_mutex_t *rmutex = MDB_MUTEX(env, r);
2608 if (!env->me_live_reader) {
2609 rc = mdb_reader_pid(env, Pidset, pid);
2612 env->me_live_reader = 1;
2615 if (LOCK_MUTEX(rc, env, rmutex))
2617 nr = ti->mti_numreaders;
2618 for (i=0; i<nr; i++)
2619 if (ti->mti_readers[i].mr_pid == 0)
2621 if (i == env->me_maxreaders) {
2622 UNLOCK_MUTEX(rmutex);
2623 return MDB_READERS_FULL;
2625 r = &ti->mti_readers[i];
2626 /* Claim the reader slot, carefully since other code
2627 * uses the reader table un-mutexed: First reset the
2628 * slot, next publish it in mti_numreaders. After
2629 * that, it is safe for mdb_env_close() to touch it.
2630 * When it will be closed, we can finally claim it.
2633 r->mr_txnid = (txnid_t)-1;
2636 ti->mti_numreaders = ++nr;
2637 env->me_close_readers = nr;
2639 UNLOCK_MUTEX(rmutex);
2641 new_notls = (env->me_flags & MDB_NOTLS);
2642 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2647 do /* LY: Retry on a race, ITS#7970. */
2648 r->mr_txnid = ti->mti_txnid;
2649 while(r->mr_txnid != ti->mti_txnid);
2650 txn->mt_txnid = r->mr_txnid;
2651 txn->mt_u.reader = r;
2652 meta = env->me_metas[txn->mt_txnid & 1];
2655 /* Not yet touching txn == env->me_txn0, it may be active */
2657 if (LOCK_MUTEX(rc, env, MDB_MUTEX(env, w)))
2659 txn->mt_txnid = ti->mti_txnid;
2660 meta = env->me_metas[txn->mt_txnid & 1];
2662 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2663 txn->mt_txnid = meta->mm_txnid;
2666 txn->mt_numdbs = env->me_numdbs;
2669 if (txn->mt_txnid == mdb_debug_start)
2673 txn->mt_child = NULL;
2674 txn->mt_loose_pgs = NULL;
2675 txn->mt_loose_count = 0;
2676 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2677 txn->mt_u.dirty_list = env->me_dirty_list;
2678 txn->mt_u.dirty_list[0].mid = 0;
2679 txn->mt_free_pgs = env->me_free_pgs;
2680 txn->mt_free_pgs[0] = 0;
2681 txn->mt_spill_pgs = NULL;
2683 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2686 /* Copy the DB info and flags */
2687 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2689 /* Moved to here to avoid a data race in read TXNs */
2690 txn->mt_next_pgno = meta->mm_last_pg+1;
2692 for (i=2; i<txn->mt_numdbs; i++) {
2693 x = env->me_dbflags[i];
2694 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2695 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2697 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2699 if (env->me_maxpg < txn->mt_next_pgno) {
2700 mdb_txn_reset0(txn, "renew0-mapfail");
2702 txn->mt_u.reader->mr_pid = 0;
2703 txn->mt_u.reader = NULL;
2705 return MDB_MAP_RESIZED;
2712 mdb_txn_renew(MDB_txn *txn)
2716 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2719 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2720 DPUTS("environment had fatal error, must shutdown!");
2724 rc = mdb_txn_renew0(txn);
2725 if (rc == MDB_SUCCESS) {
2726 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2727 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2728 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2734 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2738 int rc, size, tsize;
2740 flags &= MDB_TXN_BEGIN_FLAGS;
2741 flags |= env->me_flags & MDB_WRITEMAP;
2743 if (env->me_flags & MDB_FATAL_ERROR) {
2744 DPUTS("environment had fatal error, must shutdown!");
2747 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2750 size = tsize = sizeof(MDB_txn);
2752 /* Nested transactions: Max 1 child, write txns only, no writemap */
2753 flags |= parent->mt_flags;
2754 if (parent->mt_child ||
2755 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2757 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2759 /* Child txns save MDB_pgstate and use own copy of cursors */
2760 size = tsize = sizeof(MDB_ntxn);
2761 size += env->me_maxdbs * sizeof(MDB_cursor *);
2762 } else if (!(flags & MDB_RDONLY)) {
2763 /* Reuse preallocated write txn. However, do not touch it until
2764 * mdb_txn_renew0() succeeds, since it currently may be active.
2769 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2771 if ((txn = calloc(1, size)) == NULL) {
2772 DPRINTF(("calloc: %s", strerror(errno)));
2775 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2776 if (flags & MDB_RDONLY) {
2777 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2778 txn->mt_dbiseqs = env->me_dbiseqs;
2780 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2782 txn->mt_dbiseqs = parent->mt_dbiseqs;
2783 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2785 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2786 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2789 txn->mt_flags = flags;
2794 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2795 if (!txn->mt_u.dirty_list ||
2796 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2798 free(txn->mt_u.dirty_list);
2802 txn->mt_txnid = parent->mt_txnid;
2803 txn->mt_dirty_room = parent->mt_dirty_room;
2804 txn->mt_u.dirty_list[0].mid = 0;
2805 txn->mt_spill_pgs = NULL;
2806 txn->mt_next_pgno = parent->mt_next_pgno;
2807 parent->mt_child = txn;
2808 txn->mt_parent = parent;
2809 txn->mt_numdbs = parent->mt_numdbs;
2810 txn->mt_dbxs = parent->mt_dbxs;
2811 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2812 /* Copy parent's mt_dbflags, but clear DB_NEW */
2813 for (i=0; i<txn->mt_numdbs; i++)
2814 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2816 ntxn = (MDB_ntxn *)txn;
2817 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2818 if (env->me_pghead) {
2819 size = MDB_IDL_SIZEOF(env->me_pghead);
2820 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2822 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2827 rc = mdb_cursor_shadow(parent, txn);
2829 mdb_txn_reset0(txn, "beginchild-fail");
2832 rc = mdb_txn_renew0(txn);
2835 if (txn != env->me_txn0)
2838 txn->mt_flags |= flags; /* for txn==me_txn0, no effect otherwise */
2840 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2841 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2842 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2849 mdb_txn_env(MDB_txn *txn)
2851 if(!txn) return NULL;
2856 mdb_txn_id(MDB_txn *txn)
2859 return txn->mt_txnid;
2862 /** Export or close DBI handles opened in this txn. */
2864 mdb_dbis_update(MDB_txn *txn, int keep)
2867 MDB_dbi n = txn->mt_numdbs;
2868 MDB_env *env = txn->mt_env;
2869 unsigned char *tdbflags = txn->mt_dbflags;
2871 for (i = n; --i >= 2;) {
2872 if (tdbflags[i] & DB_NEW) {
2874 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2876 char *ptr = env->me_dbxs[i].md_name.mv_data;
2878 env->me_dbxs[i].md_name.mv_data = NULL;
2879 env->me_dbxs[i].md_name.mv_size = 0;
2880 env->me_dbflags[i] = 0;
2881 env->me_dbiseqs[i]++;
2887 if (keep && env->me_numdbs < n)
2891 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2892 * May be called twice for readonly txns: First reset it, then abort.
2893 * @param[in] txn the transaction handle to reset
2894 * @param[in] act why the transaction is being reset
2897 mdb_txn_reset0(MDB_txn *txn, const char *act)
2899 MDB_env *env = txn->mt_env;
2901 /* Close any DBI handles opened in this txn */
2902 mdb_dbis_update(txn, 0);
2904 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2905 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2906 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2908 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2909 if (txn->mt_u.reader) {
2910 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2911 if (!(env->me_flags & MDB_NOTLS))
2912 txn->mt_u.reader = NULL; /* txn does not own reader */
2914 txn->mt_numdbs = 0; /* close nothing if called again */
2915 txn->mt_dbxs = NULL; /* mark txn as reset */
2917 pgno_t *pghead = env->me_pghead;
2919 mdb_cursors_close(txn, 0);
2920 if (!(env->me_flags & MDB_WRITEMAP)) {
2921 mdb_dlist_free(txn);
2924 if (!txn->mt_parent) {
2925 if (mdb_midl_shrink(&txn->mt_free_pgs))
2926 env->me_free_pgs = txn->mt_free_pgs;
2928 env->me_pghead = NULL;
2932 /* The writer mutex was locked in mdb_txn_begin. */
2934 UNLOCK_MUTEX(MDB_MUTEX(env, w));
2936 txn->mt_parent->mt_child = NULL;
2937 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2938 mdb_midl_free(txn->mt_free_pgs);
2939 mdb_midl_free(txn->mt_spill_pgs);
2940 free(txn->mt_u.dirty_list);
2943 mdb_midl_free(pghead);
2948 mdb_txn_reset(MDB_txn *txn)
2953 /* This call is only valid for read-only txns */
2954 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2957 mdb_txn_reset0(txn, "reset");
2961 mdb_txn_abort(MDB_txn *txn)
2967 mdb_txn_abort(txn->mt_child);
2969 mdb_txn_reset0(txn, "abort");
2970 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2971 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2972 txn->mt_u.reader->mr_pid = 0;
2974 if (txn != txn->mt_env->me_txn0)
2978 /** Save the freelist as of this transaction to the freeDB.
2979 * This changes the freelist. Keep trying until it stabilizes.
2982 mdb_freelist_save(MDB_txn *txn)
2984 /* env->me_pghead[] can grow and shrink during this call.
2985 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2986 * Page numbers cannot disappear from txn->mt_free_pgs[].
2989 MDB_env *env = txn->mt_env;
2990 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2991 txnid_t pglast = 0, head_id = 0;
2992 pgno_t freecnt = 0, *free_pgs, *mop;
2993 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2995 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2997 if (env->me_pghead) {
2998 /* Make sure first page of freeDB is touched and on freelist */
2999 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3000 if (rc && rc != MDB_NOTFOUND)
3004 if (!env->me_pghead && txn->mt_loose_pgs) {
3005 /* Put loose page numbers in mt_free_pgs, since
3006 * we may be unable to return them to me_pghead.
3008 MDB_page *mp = txn->mt_loose_pgs;
3009 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3011 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3012 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3013 txn->mt_loose_pgs = NULL;
3014 txn->mt_loose_count = 0;
3017 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3018 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3019 ? SSIZE_MAX : maxfree_1pg;
3022 /* Come back here after each Put() in case freelist changed */
3027 /* If using records from freeDB which we have not yet
3028 * deleted, delete them and any we reserved for me_pghead.
3030 while (pglast < env->me_pglast) {
3031 rc = mdb_cursor_first(&mc, &key, NULL);
3034 pglast = head_id = *(txnid_t *)key.mv_data;
3035 total_room = head_room = 0;
3036 mdb_tassert(txn, pglast <= env->me_pglast);
3037 rc = mdb_cursor_del(&mc, 0);
3042 /* Save the IDL of pages freed by this txn, to a single record */
3043 if (freecnt < txn->mt_free_pgs[0]) {
3045 /* Make sure last page of freeDB is touched and on freelist */
3046 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3047 if (rc && rc != MDB_NOTFOUND)
3050 free_pgs = txn->mt_free_pgs;
3051 /* Write to last page of freeDB */
3052 key.mv_size = sizeof(txn->mt_txnid);
3053 key.mv_data = &txn->mt_txnid;
3055 freecnt = free_pgs[0];
3056 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3057 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3060 /* Retry if mt_free_pgs[] grew during the Put() */
3061 free_pgs = txn->mt_free_pgs;
3062 } while (freecnt < free_pgs[0]);
3063 mdb_midl_sort(free_pgs);
3064 memcpy(data.mv_data, free_pgs, data.mv_size);
3067 unsigned int i = free_pgs[0];
3068 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3069 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3071 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3077 mop = env->me_pghead;
3078 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3080 /* Reserve records for me_pghead[]. Split it if multi-page,
3081 * to avoid searching freeDB for a page range. Use keys in
3082 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3084 if (total_room >= mop_len) {
3085 if (total_room == mop_len || --more < 0)
3087 } else if (head_room >= maxfree_1pg && head_id > 1) {
3088 /* Keep current record (overflow page), add a new one */
3092 /* (Re)write {key = head_id, IDL length = head_room} */
3093 total_room -= head_room;
3094 head_room = mop_len - total_room;
3095 if (head_room > maxfree_1pg && head_id > 1) {
3096 /* Overflow multi-page for part of me_pghead */
3097 head_room /= head_id; /* amortize page sizes */
3098 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3099 } else if (head_room < 0) {
3100 /* Rare case, not bothering to delete this record */
3103 key.mv_size = sizeof(head_id);
3104 key.mv_data = &head_id;
3105 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3106 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3109 /* IDL is initially empty, zero out at least the length */
3110 pgs = (pgno_t *)data.mv_data;
3111 j = head_room > clean_limit ? head_room : 0;
3115 total_room += head_room;
3118 /* Return loose page numbers to me_pghead, though usually none are
3119 * left at this point. The pages themselves remain in dirty_list.
3121 if (txn->mt_loose_pgs) {
3122 MDB_page *mp = txn->mt_loose_pgs;
3123 unsigned count = txn->mt_loose_count;
3125 /* Room for loose pages + temp IDL with same */
3126 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3128 mop = env->me_pghead;
3129 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3130 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3131 loose[ ++count ] = mp->mp_pgno;
3133 mdb_midl_sort(loose);
3134 mdb_midl_xmerge(mop, loose);
3135 txn->mt_loose_pgs = NULL;
3136 txn->mt_loose_count = 0;
3140 /* Fill in the reserved me_pghead records */
3146 rc = mdb_cursor_first(&mc, &key, &data);
3147 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3148 txnid_t id = *(txnid_t *)key.mv_data;
3149 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3152 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3154 if (len > mop_len) {
3156 data.mv_size = (len + 1) * sizeof(MDB_ID);
3158 data.mv_data = mop -= len;
3161 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3163 if (rc || !(mop_len -= len))
3170 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3171 * @param[in] txn the transaction that's being committed
3172 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3173 * @return 0 on success, non-zero on failure.
3176 mdb_page_flush(MDB_txn *txn, int keep)
3178 MDB_env *env = txn->mt_env;
3179 MDB_ID2L dl = txn->mt_u.dirty_list;
3180 unsigned psize = env->me_psize, j;
3181 int i, pagecount = dl[0].mid, rc;
3182 size_t size = 0, pos = 0;
3184 MDB_page *dp = NULL;
3188 struct iovec iov[MDB_COMMIT_PAGES];
3189 ssize_t wpos = 0, wsize = 0, wres;
3190 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3196 if (env->me_flags & MDB_WRITEMAP) {
3197 /* Clear dirty flags */
3198 while (++i <= pagecount) {
3200 /* Don't flush this page yet */
3201 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3202 dp->mp_flags &= ~P_KEEP;
3206 dp->mp_flags &= ~P_DIRTY;
3211 /* Write the pages */
3213 if (++i <= pagecount) {
3215 /* Don't flush this page yet */
3216 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3217 dp->mp_flags &= ~P_KEEP;
3222 /* clear dirty flag */
3223 dp->mp_flags &= ~P_DIRTY;
3226 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3231 /* Windows actually supports scatter/gather I/O, but only on
3232 * unbuffered file handles. Since we're relying on the OS page
3233 * cache for all our data, that's self-defeating. So we just
3234 * write pages one at a time. We use the ov structure to set
3235 * the write offset, to at least save the overhead of a Seek
3238 DPRINTF(("committing page %"Z"u", pgno));
3239 memset(&ov, 0, sizeof(ov));
3240 ov.Offset = pos & 0xffffffff;
3241 ov.OffsetHigh = pos >> 16 >> 16;
3242 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3244 DPRINTF(("WriteFile: %d", rc));
3248 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3249 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3252 /* Write previous page(s) */
3253 #ifdef MDB_USE_PWRITEV
3254 wres = pwritev(env->me_fd, iov, n, wpos);
3257 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3260 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3264 DPRINTF(("lseek: %s", strerror(rc)));
3267 wres = writev(env->me_fd, iov, n);
3270 if (wres != wsize) {
3275 DPRINTF(("Write error: %s", strerror(rc)));
3277 rc = EIO; /* TODO: Use which error code? */
3278 DPUTS("short write, filesystem full?");
3289 DPRINTF(("committing page %"Z"u", pgno));
3290 next_pos = pos + size;
3291 iov[n].iov_len = size;
3292 iov[n].iov_base = (char *)dp;
3298 /* MIPS has cache coherency issues, this is a no-op everywhere else
3299 * Note: for any size >= on-chip cache size, entire on-chip cache is
3302 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3304 for (i = keep; ++i <= pagecount; ) {
3306 /* This is a page we skipped above */
3309 dl[j].mid = dp->mp_pgno;
3312 mdb_dpage_free(env, dp);
3317 txn->mt_dirty_room += i - j;
3323 mdb_txn_commit(MDB_txn *txn)
3329 if (txn == NULL || txn->mt_env == NULL)
3332 if (txn->mt_child) {
3333 rc = mdb_txn_commit(txn->mt_child);
3334 txn->mt_child = NULL;
3341 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3342 mdb_dbis_update(txn, 1);
3343 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3348 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3349 DPUTS("error flag is set, can't commit");
3351 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3356 if (txn->mt_parent) {
3357 MDB_txn *parent = txn->mt_parent;
3361 unsigned x, y, len, ps_len;
3363 /* Append our free list to parent's */
3364 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3367 mdb_midl_free(txn->mt_free_pgs);
3368 /* Failures after this must either undo the changes
3369 * to the parent or set MDB_TXN_ERROR in the parent.
3372 parent->mt_next_pgno = txn->mt_next_pgno;
3373 parent->mt_flags = txn->mt_flags;
3375 /* Merge our cursors into parent's and close them */
3376 mdb_cursors_close(txn, 1);
3378 /* Update parent's DB table. */
3379 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3380 parent->mt_numdbs = txn->mt_numdbs;
3381 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3382 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3383 for (i=2; i<txn->mt_numdbs; i++) {
3384 /* preserve parent's DB_NEW status */
3385 x = parent->mt_dbflags[i] & DB_NEW;
3386 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3389 dst = parent->mt_u.dirty_list;
3390 src = txn->mt_u.dirty_list;
3391 /* Remove anything in our dirty list from parent's spill list */
3392 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3394 pspill[0] = (pgno_t)-1;
3395 /* Mark our dirty pages as deleted in parent spill list */
3396 for (i=0, len=src[0].mid; ++i <= len; ) {
3397 MDB_ID pn = src[i].mid << 1;
3398 while (pn > pspill[x])
3400 if (pn == pspill[x]) {
3405 /* Squash deleted pagenums if we deleted any */
3406 for (x=y; ++x <= ps_len; )
3407 if (!(pspill[x] & 1))
3408 pspill[++y] = pspill[x];
3412 /* Find len = length of merging our dirty list with parent's */
3414 dst[0].mid = 0; /* simplify loops */
3415 if (parent->mt_parent) {
3416 len = x + src[0].mid;
3417 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3418 for (i = x; y && i; y--) {
3419 pgno_t yp = src[y].mid;
3420 while (yp < dst[i].mid)
3422 if (yp == dst[i].mid) {
3427 } else { /* Simplify the above for single-ancestor case */
3428 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3430 /* Merge our dirty list with parent's */
3432 for (i = len; y; dst[i--] = src[y--]) {
3433 pgno_t yp = src[y].mid;
3434 while (yp < dst[x].mid)
3435 dst[i--] = dst[x--];
3436 if (yp == dst[x].mid)
3437 free(dst[x--].mptr);
3439 mdb_tassert(txn, i == x);
3441 free(txn->mt_u.dirty_list);
3442 parent->mt_dirty_room = txn->mt_dirty_room;
3443 if (txn->mt_spill_pgs) {
3444 if (parent->mt_spill_pgs) {
3445 /* TODO: Prevent failure here, so parent does not fail */
3446 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3448 parent->mt_flags |= MDB_TXN_ERROR;
3449 mdb_midl_free(txn->mt_spill_pgs);
3450 mdb_midl_sort(parent->mt_spill_pgs);
3452 parent->mt_spill_pgs = txn->mt_spill_pgs;
3456 /* Append our loose page list to parent's */
3457 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3459 *lp = txn->mt_loose_pgs;
3460 parent->mt_loose_count += txn->mt_loose_count;
3462 parent->mt_child = NULL;
3463 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3468 if (txn != env->me_txn) {
3469 DPUTS("attempt to commit unknown transaction");
3474 mdb_cursors_close(txn, 0);
3476 if (!txn->mt_u.dirty_list[0].mid &&
3477 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3480 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3481 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3483 /* Update DB root pointers */
3484 if (txn->mt_numdbs > 2) {
3488 data.mv_size = sizeof(MDB_db);
3490 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3491 for (i = 2; i < txn->mt_numdbs; i++) {
3492 if (txn->mt_dbflags[i] & DB_DIRTY) {
3493 if (TXN_DBI_CHANGED(txn, i)) {
3497 data.mv_data = &txn->mt_dbs[i];
3498 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3505 rc = mdb_freelist_save(txn);
3509 mdb_midl_free(env->me_pghead);
3510 env->me_pghead = NULL;
3511 if (mdb_midl_shrink(&txn->mt_free_pgs))
3512 env->me_free_pgs = txn->mt_free_pgs;
3518 if ((rc = mdb_page_flush(txn, 0)))
3520 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3521 (rc = mdb_env_sync(env, 0)))
3523 if ((rc = mdb_env_write_meta(txn)))
3526 /* Free P_LOOSE pages left behind in dirty_list */
3527 if (!(env->me_flags & MDB_WRITEMAP))
3528 mdb_dlist_free(txn);
3533 mdb_dbis_update(txn, 1);
3536 UNLOCK_MUTEX(MDB_MUTEX(env, w));
3537 if (txn != env->me_txn0)
3547 /** Read the environment parameters of a DB environment before
3548 * mapping it into memory.
3549 * @param[in] env the environment handle
3550 * @param[out] meta address of where to store the meta information
3551 * @return 0 on success, non-zero on failure.
3554 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3560 enum { Size = sizeof(pbuf) };
3562 /* We don't know the page size yet, so use a minimum value.
3563 * Read both meta pages so we can use the latest one.
3566 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3570 memset(&ov, 0, sizeof(ov));
3572 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3573 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3576 rc = pread(env->me_fd, &pbuf, Size, off);
3579 if (rc == 0 && off == 0)
3581 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3582 DPRINTF(("read: %s", mdb_strerror(rc)));
3586 p = (MDB_page *)&pbuf;
3588 if (!F_ISSET(p->mp_flags, P_META)) {
3589 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3594 if (m->mm_magic != MDB_MAGIC) {
3595 DPUTS("meta has invalid magic");
3599 if (m->mm_version != MDB_DATA_VERSION) {
3600 DPRINTF(("database is version %u, expected version %u",
3601 m->mm_version, MDB_DATA_VERSION));
3602 return MDB_VERSION_MISMATCH;
3605 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3611 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3613 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3615 meta->mm_magic = MDB_MAGIC;
3616 meta->mm_version = MDB_DATA_VERSION;
3617 meta->mm_mapsize = env->me_mapsize;
3618 meta->mm_psize = env->me_psize;
3619 meta->mm_last_pg = 1;
3620 meta->mm_flags = env->me_flags & 0xffff;
3621 meta->mm_flags |= MDB_INTEGERKEY;
3622 meta->mm_dbs[0].md_root = P_INVALID;
3623 meta->mm_dbs[1].md_root = P_INVALID;
3626 /** Write the environment parameters of a freshly created DB environment.
3627 * @param[in] env the environment handle
3628 * @param[in] meta the #MDB_meta to write
3629 * @return 0 on success, non-zero on failure.
3632 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3640 memset(&ov, 0, sizeof(ov));
3641 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3643 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3646 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3647 len = pwrite(fd, ptr, size, pos); \
3648 if (len == -1 && ErrCode() == EINTR) continue; \
3649 rc = (len >= 0); break; } while(1)
3652 DPUTS("writing new meta page");
3654 psize = env->me_psize;
3656 p = calloc(2, psize);
3658 p->mp_flags = P_META;
3659 *(MDB_meta *)METADATA(p) = *meta;
3661 q = (MDB_page *)((char *)p + psize);
3663 q->mp_flags = P_META;
3664 *(MDB_meta *)METADATA(q) = *meta;
3666 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3669 else if ((unsigned) len == psize * 2)
3677 /** Update the environment info to commit a transaction.
3678 * @param[in] txn the transaction that's being committed
3679 * @return 0 on success, non-zero on failure.
3682 mdb_env_write_meta(MDB_txn *txn)
3685 MDB_meta meta, metab, *mp;
3689 int rc, len, toggle;
3698 toggle = txn->mt_txnid & 1;
3699 DPRINTF(("writing meta page %d for root page %"Z"u",
3700 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3703 flags = txn->mt_flags & env->me_flags;
3704 mp = env->me_metas[toggle];
3705 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3706 /* Persist any increases of mapsize config */
3707 if (mapsize < env->me_mapsize)
3708 mapsize = env->me_mapsize;
3710 if (flags & MDB_WRITEMAP) {
3711 mp->mm_mapsize = mapsize;
3712 mp->mm_dbs[0] = txn->mt_dbs[0];
3713 mp->mm_dbs[1] = txn->mt_dbs[1];
3714 mp->mm_last_pg = txn->mt_next_pgno - 1;
3715 #if !(defined(_MSC_VER) || defined(__i386__) || defined(__x86_64__))
3716 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3717 __sync_synchronize();
3719 mp->mm_txnid = txn->mt_txnid;
3720 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3721 unsigned meta_size = env->me_psize;
3722 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3725 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3726 if (meta_size < env->me_os_psize)
3727 meta_size += meta_size;
3732 if (MDB_MSYNC(ptr, meta_size, rc)) {
3739 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3740 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3742 meta.mm_mapsize = mapsize;
3743 meta.mm_dbs[0] = txn->mt_dbs[0];
3744 meta.mm_dbs[1] = txn->mt_dbs[1];
3745 meta.mm_last_pg = txn->mt_next_pgno - 1;
3746 meta.mm_txnid = txn->mt_txnid;
3748 off = offsetof(MDB_meta, mm_mapsize);
3749 ptr = (char *)&meta + off;
3750 len = sizeof(MDB_meta) - off;
3752 off += env->me_psize;
3755 /* Write to the SYNC fd */
3756 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3760 memset(&ov, 0, sizeof(ov));
3762 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3766 rc = pwrite(mfd, ptr, len, off);
3769 rc = rc < 0 ? ErrCode() : EIO;
3772 DPUTS("write failed, disk error?");
3773 /* On a failure, the pagecache still contains the new data.
3774 * Write some old data back, to prevent it from being used.
3775 * Use the non-SYNC fd; we know it will fail anyway.
3777 meta.mm_last_pg = metab.mm_last_pg;
3778 meta.mm_txnid = metab.mm_txnid;
3780 memset(&ov, 0, sizeof(ov));
3782 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3784 r2 = pwrite(env->me_fd, ptr, len, off);
3785 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3788 env->me_flags |= MDB_FATAL_ERROR;
3791 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3792 CACHEFLUSH(env->me_map + off, len, DCACHE);
3794 /* Memory ordering issues are irrelevant; since the entire writer
3795 * is wrapped by wmutex, all of these changes will become visible
3796 * after the wmutex is unlocked. Since the DB is multi-version,
3797 * readers will get consistent data regardless of how fresh or
3798 * how stale their view of these values is.
3801 env->me_txns->mti_txnid = txn->mt_txnid;
3806 /** Check both meta pages to see which one is newer.
3807 * @param[in] env the environment handle
3808 * @return meta toggle (0 or 1).
3811 mdb_env_pick_meta(const MDB_env *env)
3813 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3817 mdb_env_create(MDB_env **env)
3821 e = calloc(1, sizeof(MDB_env));
3825 e->me_maxreaders = DEFAULT_READERS;
3826 e->me_maxdbs = e->me_numdbs = 2;
3827 e->me_fd = INVALID_HANDLE_VALUE;
3828 e->me_lfd = INVALID_HANDLE_VALUE;
3829 e->me_mfd = INVALID_HANDLE_VALUE;
3830 #ifdef MDB_USE_SYSV_SEM
3831 e->me_rmutex.semid = -1;
3832 e->me_wmutex.semid = -1;
3834 e->me_pid = getpid();
3835 GET_PAGESIZE(e->me_os_psize);
3836 VGMEMP_CREATE(e,0,0);
3842 mdb_env_map(MDB_env *env, void *addr)
3845 unsigned int flags = env->me_flags;
3849 LONG sizelo, sizehi;
3852 if (flags & MDB_RDONLY) {
3853 /* Don't set explicit map size, use whatever exists */
3858 msize = env->me_mapsize;
3859 sizelo = msize & 0xffffffff;
3860 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3862 /* Windows won't create mappings for zero length files.
3863 * and won't map more than the file size.
3864 * Just set the maxsize right now.
3866 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3867 || !SetEndOfFile(env->me_fd)
3868 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3872 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3873 PAGE_READWRITE : PAGE_READONLY,
3874 sizehi, sizelo, NULL);
3877 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3878 FILE_MAP_WRITE : FILE_MAP_READ,
3880 rc = env->me_map ? 0 : ErrCode();
3885 int prot = PROT_READ;
3886 if (flags & MDB_WRITEMAP) {
3888 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3891 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3893 if (env->me_map == MAP_FAILED) {
3898 if (flags & MDB_NORDAHEAD) {
3899 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3901 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3903 #ifdef POSIX_MADV_RANDOM
3904 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3905 #endif /* POSIX_MADV_RANDOM */
3906 #endif /* MADV_RANDOM */
3910 /* Can happen because the address argument to mmap() is just a
3911 * hint. mmap() can pick another, e.g. if the range is in use.
3912 * The MAP_FIXED flag would prevent that, but then mmap could
3913 * instead unmap existing pages to make room for the new map.
3915 if (addr && env->me_map != addr)
3916 return EBUSY; /* TODO: Make a new MDB_* error code? */
3918 p = (MDB_page *)env->me_map;
3919 env->me_metas[0] = METADATA(p);
3920 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3926 mdb_env_set_mapsize(MDB_env *env, size_t size)
3928 /* If env is already open, caller is responsible for making
3929 * sure there are no active txns.
3937 meta = env->me_metas[mdb_env_pick_meta(env)];
3939 size = meta->mm_mapsize;
3941 /* Silently round up to minimum if the size is too small */
3942 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3946 munmap(env->me_map, env->me_mapsize);
3947 env->me_mapsize = size;
3948 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3949 rc = mdb_env_map(env, old);
3953 env->me_mapsize = size;
3955 env->me_maxpg = env->me_mapsize / env->me_psize;
3960 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3964 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3969 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3971 if (env->me_map || readers < 1)
3973 env->me_maxreaders = readers;
3978 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3980 if (!env || !readers)
3982 *readers = env->me_maxreaders;
3987 mdb_fsize(HANDLE fd, size_t *size)
3990 LARGE_INTEGER fsize;
3992 if (!GetFileSizeEx(fd, &fsize))
3995 *size = fsize.QuadPart;
4007 #ifdef BROKEN_FDATASYNC
4008 #include <sys/utsname.h>
4009 #include <sys/vfs.h>
4012 /** Further setup required for opening an LMDB environment
4015 mdb_env_open2(MDB_env *env)
4017 unsigned int flags = env->me_flags;
4018 int i, newenv = 0, rc;
4022 /* See if we should use QueryLimited */
4024 if ((rc & 0xff) > 5)
4025 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4027 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4030 #ifdef BROKEN_FDATASYNC
4031 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4032 * https://lkml.org/lkml/2012/9/3/83
4033 * Kernels after 3.6-rc6 are known good.
4034 * https://lkml.org/lkml/2012/9/10/556
4035 * See if the DB is on ext3/ext4, then check for new enough kernel
4036 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4041 fstatfs(env->me_fd, &st);
4042 while (st.f_type == 0xEF53) {
4046 if (uts.release[0] < '3') {
4047 if (!strncmp(uts.release, "2.6.32.", 7)) {
4048 i = atoi(uts.release+7);
4050 break; /* 2.6.32.60 and newer is OK */
4051 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4052 i = atoi(uts.release+7);
4054 break; /* 2.6.34.15 and newer is OK */
4056 } else if (uts.release[0] == '3') {
4057 i = atoi(uts.release+2);
4059 break; /* 3.6 and newer is OK */
4061 i = atoi(uts.release+4);
4063 break; /* 3.5.4 and newer is OK */
4064 } else if (i == 2) {
4065 i = atoi(uts.release+4);
4067 break; /* 3.2.30 and newer is OK */
4069 } else { /* 4.x and newer is OK */
4072 env->me_flags |= MDB_FSYNCONLY;
4078 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4081 DPUTS("new mdbenv");
4083 env->me_psize = env->me_os_psize;
4084 if (env->me_psize > MAX_PAGESIZE)
4085 env->me_psize = MAX_PAGESIZE;
4086 memset(&meta, 0, sizeof(meta));
4087 mdb_env_init_meta0(env, &meta);
4088 meta.mm_mapsize = DEFAULT_MAPSIZE;
4090 env->me_psize = meta.mm_psize;
4093 /* Was a mapsize configured? */
4094 if (!env->me_mapsize) {
4095 env->me_mapsize = meta.mm_mapsize;
4098 /* Make sure mapsize >= committed data size. Even when using
4099 * mm_mapsize, which could be broken in old files (ITS#7789).
4101 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4102 if (env->me_mapsize < minsize)
4103 env->me_mapsize = minsize;
4105 meta.mm_mapsize = env->me_mapsize;
4107 if (newenv && !(flags & MDB_FIXEDMAP)) {
4108 /* mdb_env_map() may grow the datafile. Write the metapages
4109 * first, so the file will be valid if initialization fails.
4110 * Except with FIXEDMAP, since we do not yet know mm_address.
4111 * We could fill in mm_address later, but then a different
4112 * program might end up doing that - one with a memory layout
4113 * and map address which does not suit the main program.
4115 rc = mdb_env_init_meta(env, &meta);
4121 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4126 if (flags & MDB_FIXEDMAP)
4127 meta.mm_address = env->me_map;
4128 i = mdb_env_init_meta(env, &meta);
4129 if (i != MDB_SUCCESS) {
4134 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4135 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4137 #if !(MDB_MAXKEYSIZE)
4138 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4140 env->me_maxpg = env->me_mapsize / env->me_psize;
4144 int toggle = mdb_env_pick_meta(env);
4145 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4147 DPRINTF(("opened database version %u, pagesize %u",
4148 env->me_metas[0]->mm_version, env->me_psize));
4149 DPRINTF(("using meta page %d", toggle));
4150 DPRINTF(("depth: %u", db->md_depth));
4151 DPRINTF(("entries: %"Z"u", db->md_entries));
4152 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4153 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4154 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4155 DPRINTF(("root: %"Z"u", db->md_root));
4163 /** Release a reader thread's slot in the reader lock table.
4164 * This function is called automatically when a thread exits.
4165 * @param[in] ptr This points to the slot in the reader lock table.
4168 mdb_env_reader_dest(void *ptr)
4170 MDB_reader *reader = ptr;
4176 /** Junk for arranging thread-specific callbacks on Windows. This is
4177 * necessarily platform and compiler-specific. Windows supports up
4178 * to 1088 keys. Let's assume nobody opens more than 64 environments
4179 * in a single process, for now. They can override this if needed.
4181 #ifndef MAX_TLS_KEYS
4182 #define MAX_TLS_KEYS 64
4184 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4185 static int mdb_tls_nkeys;
4187 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4191 case DLL_PROCESS_ATTACH: break;
4192 case DLL_THREAD_ATTACH: break;
4193 case DLL_THREAD_DETACH:
4194 for (i=0; i<mdb_tls_nkeys; i++) {
4195 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4197 mdb_env_reader_dest(r);
4201 case DLL_PROCESS_DETACH: break;
4206 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4208 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4212 /* Force some symbol references.
4213 * _tls_used forces the linker to create the TLS directory if not already done
4214 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4216 #pragma comment(linker, "/INCLUDE:_tls_used")
4217 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4218 #pragma const_seg(".CRT$XLB")
4219 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4220 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4223 #pragma comment(linker, "/INCLUDE:__tls_used")
4224 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4225 #pragma data_seg(".CRT$XLB")
4226 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4228 #endif /* WIN 32/64 */
4229 #endif /* !__GNUC__ */
4232 /** Downgrade the exclusive lock on the region back to shared */
4234 mdb_env_share_locks(MDB_env *env, int *excl)
4236 int rc = 0, toggle = mdb_env_pick_meta(env);
4238 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4243 /* First acquire a shared lock. The Unlock will
4244 * then release the existing exclusive lock.
4246 memset(&ov, 0, sizeof(ov));
4247 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4250 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4256 struct flock lock_info;
4257 /* The shared lock replaces the existing lock */
4258 memset((void *)&lock_info, 0, sizeof(lock_info));
4259 lock_info.l_type = F_RDLCK;
4260 lock_info.l_whence = SEEK_SET;
4261 lock_info.l_start = 0;
4262 lock_info.l_len = 1;
4263 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4264 (rc = ErrCode()) == EINTR) ;
4265 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4272 /** Try to get exclusive lock, otherwise shared.
4273 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4276 mdb_env_excl_lock(MDB_env *env, int *excl)
4280 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4284 memset(&ov, 0, sizeof(ov));
4285 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4292 struct flock lock_info;
4293 memset((void *)&lock_info, 0, sizeof(lock_info));
4294 lock_info.l_type = F_WRLCK;
4295 lock_info.l_whence = SEEK_SET;
4296 lock_info.l_start = 0;
4297 lock_info.l_len = 1;
4298 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4299 (rc = ErrCode()) == EINTR) ;
4303 # ifdef MDB_USE_SYSV_SEM
4304 if (*excl < 0) /* always true when !MDB_USE_SYSV_SEM */
4307 lock_info.l_type = F_RDLCK;
4308 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4309 (rc = ErrCode()) == EINTR) ;
4319 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4321 * @(#) $Revision: 5.1 $
4322 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4323 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4325 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4329 * Please do not copyright this code. This code is in the public domain.
4331 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4332 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4333 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4334 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4335 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4336 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4337 * PERFORMANCE OF THIS SOFTWARE.
4340 * chongo <Landon Curt Noll> /\oo/\
4341 * http://www.isthe.com/chongo/
4343 * Share and Enjoy! :-)
4346 typedef unsigned long long mdb_hash_t;
4347 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4349 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4350 * @param[in] val value to hash
4351 * @param[in] hval initial value for hash
4352 * @return 64 bit hash
4354 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4355 * hval arg on the first call.
4358 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4360 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4361 unsigned char *end = s + val->mv_size;
4363 * FNV-1a hash each octet of the string
4366 /* xor the bottom with the current octet */
4367 hval ^= (mdb_hash_t)*s++;
4369 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4370 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4371 (hval << 7) + (hval << 8) + (hval << 40);
4373 /* return our new hash value */
4377 /** Hash the string and output the encoded hash.
4378 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4379 * very short name limits. We don't care about the encoding being reversible,
4380 * we just want to preserve as many bits of the input as possible in a
4381 * small printable string.
4382 * @param[in] str string to hash
4383 * @param[out] encbuf an array of 11 chars to hold the hash
4385 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4388 mdb_pack85(unsigned long l, char *out)
4392 for (i=0; i<5; i++) {
4393 *out++ = mdb_a85[l % 85];
4399 mdb_hash_enc(MDB_val *val, char *encbuf)
4401 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4403 mdb_pack85(h, encbuf);
4404 mdb_pack85(h>>32, encbuf+5);
4409 /** Open and/or initialize the lock region for the environment.
4410 * @param[in] env The LMDB environment.
4411 * @param[in] lpath The pathname of the file used for the lock region.
4412 * @param[in] mode The Unix permissions for the file, if we create it.
4413 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4414 * @return 0 on success, non-zero on failure.
4417 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4420 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4422 # define MDB_ERRCODE_ROFS EROFS
4423 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4424 # define MDB_CLOEXEC O_CLOEXEC
4427 # define MDB_CLOEXEC 0
4430 #ifdef MDB_USE_SYSV_SEM
4438 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4439 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4440 FILE_ATTRIBUTE_NORMAL, NULL);
4442 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4444 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4446 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4451 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4452 /* Lose record locks when exec*() */
4453 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4454 fcntl(env->me_lfd, F_SETFD, fdflags);
4457 if (!(env->me_flags & MDB_NOTLS)) {
4458 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4461 env->me_flags |= MDB_ENV_TXKEY;
4463 /* Windows TLS callbacks need help finding their TLS info. */
4464 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4468 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4472 /* Try to get exclusive lock. If we succeed, then
4473 * nobody is using the lock region and we should initialize it.
4475 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4478 size = GetFileSize(env->me_lfd, NULL);
4480 size = lseek(env->me_lfd, 0, SEEK_END);
4481 if (size == -1) goto fail_errno;
4483 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4484 if (size < rsize && *excl > 0) {
4486 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4487 || !SetEndOfFile(env->me_lfd))
4490 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4494 size = rsize - sizeof(MDB_txninfo);
4495 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4500 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4502 if (!mh) goto fail_errno;
4503 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4505 if (!env->me_txns) goto fail_errno;
4507 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4509 if (m == MAP_FAILED) goto fail_errno;
4515 BY_HANDLE_FILE_INFORMATION stbuf;
4524 if (!mdb_sec_inited) {
4525 InitializeSecurityDescriptor(&mdb_null_sd,
4526 SECURITY_DESCRIPTOR_REVISION);
4527 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4528 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4529 mdb_all_sa.bInheritHandle = FALSE;
4530 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4533 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4534 idbuf.volume = stbuf.dwVolumeSerialNumber;
4535 idbuf.nhigh = stbuf.nFileIndexHigh;
4536 idbuf.nlow = stbuf.nFileIndexLow;
4537 val.mv_data = &idbuf;
4538 val.mv_size = sizeof(idbuf);
4539 mdb_hash_enc(&val, encbuf);
4540 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4541 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4542 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4543 if (!env->me_rmutex) goto fail_errno;
4544 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4545 if (!env->me_wmutex) goto fail_errno;
4546 #elif defined(MDB_USE_SYSV_SEM)
4547 unsigned short vals[2] = {1, 1};
4548 semid = semget(IPC_PRIVATE, 2, mode);
4552 if (semctl(semid, 0, SETALL, semu) < 0)
4554 env->me_txns->mti_semid = semid;
4555 #else /* MDB_USE_SYSV_SEM */
4556 pthread_mutexattr_t mattr;
4558 if ((rc = pthread_mutexattr_init(&mattr))
4559 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4560 #ifdef MDB_ROBUST_SUPPORTED
4561 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4563 || (rc = pthread_mutex_init(&env->me_txns->mti_rmutex, &mattr))
4564 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4566 pthread_mutexattr_destroy(&mattr);
4567 #endif /* _WIN32 || MDB_USE_SYSV_SEM */
4569 env->me_txns->mti_magic = MDB_MAGIC;
4570 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4571 env->me_txns->mti_txnid = 0;
4572 env->me_txns->mti_numreaders = 0;
4575 #ifdef MDB_USE_SYSV_SEM
4576 struct semid_ds buf;
4578 if (env->me_txns->mti_magic != MDB_MAGIC) {
4579 DPUTS("lock region has invalid magic");
4583 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4584 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4585 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4586 rc = MDB_VERSION_MISMATCH;
4590 if (rc && rc != EACCES && rc != EAGAIN) {
4594 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4595 if (!env->me_rmutex) goto fail_errno;
4596 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4597 if (!env->me_wmutex) goto fail_errno;
4598 #elif defined(MDB_USE_SYSV_SEM)
4599 semid = env->me_txns->mti_semid;
4601 /* check for read access */
4602 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4604 /* check for write access */
4605 if (semctl(semid, 0, IPC_SET, semu) < 0)
4609 #ifdef MDB_USE_SYSV_SEM
4610 env->me_rmutex.semid = semid;
4611 env->me_wmutex.semid = semid;
4612 env->me_rmutex.semnum = 0;
4613 env->me_wmutex.semnum = 1;
4614 env->me_rmutex.locked = &env->me_txns->mti_rlocked;
4615 env->me_wmutex.locked = &env->me_txns->mti_wlocked;
4626 /** The name of the lock file in the DB environment */
4627 #define LOCKNAME "/lock.mdb"
4628 /** The name of the data file in the DB environment */
4629 #define DATANAME "/data.mdb"
4630 /** The suffix of the lock file when no subdir is used */
4631 #define LOCKSUFF "-lock"
4632 /** Only a subset of the @ref mdb_env flags can be changed
4633 * at runtime. Changing other flags requires closing the
4634 * environment and re-opening it with the new flags.
4636 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4637 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4638 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4640 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4641 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4645 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4647 int oflags, rc, len, excl = -1;
4648 char *lpath, *dpath;
4650 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4654 if (flags & MDB_NOSUBDIR) {
4655 rc = len + sizeof(LOCKSUFF) + len + 1;
4657 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4662 if (flags & MDB_NOSUBDIR) {
4663 dpath = lpath + len + sizeof(LOCKSUFF);
4664 sprintf(lpath, "%s" LOCKSUFF, path);
4665 strcpy(dpath, path);
4667 dpath = lpath + len + sizeof(LOCKNAME);
4668 sprintf(lpath, "%s" LOCKNAME, path);
4669 sprintf(dpath, "%s" DATANAME, path);
4673 flags |= env->me_flags;
4674 if (flags & MDB_RDONLY) {
4675 /* silently ignore WRITEMAP when we're only getting read access */
4676 flags &= ~MDB_WRITEMAP;
4678 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4679 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4682 env->me_flags = flags |= MDB_ENV_ACTIVE;
4686 env->me_path = strdup(path);
4687 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4688 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4689 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4690 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4695 /* For RDONLY, get lockfile after we know datafile exists */
4696 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4697 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4703 if (F_ISSET(flags, MDB_RDONLY)) {
4704 oflags = GENERIC_READ;
4705 len = OPEN_EXISTING;
4707 oflags = GENERIC_READ|GENERIC_WRITE;
4710 mode = FILE_ATTRIBUTE_NORMAL;
4711 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4712 NULL, len, mode, NULL);
4714 if (F_ISSET(flags, MDB_RDONLY))
4717 oflags = O_RDWR | O_CREAT;
4719 env->me_fd = open(dpath, oflags, mode);
4721 if (env->me_fd == INVALID_HANDLE_VALUE) {
4726 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4727 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4732 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4733 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4734 env->me_mfd = env->me_fd;
4736 /* Synchronous fd for meta writes. Needed even with
4737 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4740 len = OPEN_EXISTING;
4741 env->me_mfd = CreateFile(dpath, oflags,
4742 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4743 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4746 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4748 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4753 DPRINTF(("opened dbenv %p", (void *) env));
4755 rc = mdb_env_share_locks(env, &excl);
4759 if (!(flags & MDB_RDONLY)) {
4761 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4762 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4763 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4764 (txn = calloc(1, size)))
4766 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4767 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4768 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4769 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4771 txn->mt_dbxs = env->me_dbxs;
4781 mdb_env_close0(env, excl);
4787 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4789 mdb_env_close0(MDB_env *env, int excl)
4793 if (!(env->me_flags & MDB_ENV_ACTIVE))
4796 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4798 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4799 free(env->me_dbxs[i].md_name.mv_data);
4804 free(env->me_dbiseqs);
4805 free(env->me_dbflags);
4807 free(env->me_dirty_list);
4809 mdb_midl_free(env->me_free_pgs);
4811 if (env->me_flags & MDB_ENV_TXKEY) {
4812 pthread_key_delete(env->me_txkey);
4814 /* Delete our key from the global list */
4815 for (i=0; i<mdb_tls_nkeys; i++)
4816 if (mdb_tls_keys[i] == env->me_txkey) {
4817 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4825 munmap(env->me_map, env->me_mapsize);
4827 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4828 (void) close(env->me_mfd);
4829 if (env->me_fd != INVALID_HANDLE_VALUE)
4830 (void) close(env->me_fd);
4832 MDB_PID_T pid = env->me_pid;
4833 /* Clearing readers is done in this function because
4834 * me_txkey with its destructor must be disabled first.
4836 * We skip the the reader mutex, so we touch only
4837 * data owned by this process (me_close_readers and
4838 * our readers), and clear each reader atomically.
4840 for (i = env->me_close_readers; --i >= 0; )
4841 if (env->me_txns->mti_readers[i].mr_pid == pid)
4842 env->me_txns->mti_readers[i].mr_pid = 0;
4844 if (env->me_rmutex) {
4845 CloseHandle(env->me_rmutex);
4846 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4848 /* Windows automatically destroys the mutexes when
4849 * the last handle closes.
4851 #elif defined(MDB_USE_SYSV_SEM)
4852 if (env->me_rmutex.semid != -1) {
4853 /* If we have the filelock: If we are the
4854 * only remaining user, clean up semaphores.
4857 mdb_env_excl_lock(env, &excl);
4859 semctl(env->me_rmutex.semid, 0, IPC_RMID);
4862 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4864 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4867 /* Unlock the lockfile. Windows would have unlocked it
4868 * after closing anyway, but not necessarily at once.
4870 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4873 (void) close(env->me_lfd);
4876 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4880 mdb_env_close(MDB_env *env)
4887 VGMEMP_DESTROY(env);
4888 while ((dp = env->me_dpages) != NULL) {
4889 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4890 env->me_dpages = dp->mp_next;
4894 mdb_env_close0(env, 0);
4898 /** Compare two items pointing at aligned size_t's */
4900 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4902 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4903 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4906 /** Compare two items pointing at aligned unsigned int's.
4908 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4909 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4912 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4914 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4915 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4918 /** Compare two items pointing at unsigned ints of unknown alignment.
4919 * Nodes and keys are guaranteed to be 2-byte aligned.
4922 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4924 #if BYTE_ORDER == LITTLE_ENDIAN
4925 unsigned short *u, *c;
4928 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4929 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4932 } while(!x && u > (unsigned short *)a->mv_data);
4935 unsigned short *u, *c, *end;
4938 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4939 u = (unsigned short *)a->mv_data;
4940 c = (unsigned short *)b->mv_data;
4943 } while(!x && u < end);
4948 /** Compare two items lexically */
4950 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4957 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4963 diff = memcmp(a->mv_data, b->mv_data, len);
4964 return diff ? diff : len_diff<0 ? -1 : len_diff;
4967 /** Compare two items in reverse byte order */
4969 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4971 const unsigned char *p1, *p2, *p1_lim;
4975 p1_lim = (const unsigned char *)a->mv_data;
4976 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4977 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4979 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4985 while (p1 > p1_lim) {
4986 diff = *--p1 - *--p2;
4990 return len_diff<0 ? -1 : len_diff;
4993 /** Search for key within a page, using binary search.
4994 * Returns the smallest entry larger or equal to the key.
4995 * If exactp is non-null, stores whether the found entry was an exact match
4996 * in *exactp (1 or 0).
4997 * Updates the cursor index with the index of the found entry.
4998 * If no entry larger or equal to the key is found, returns NULL.
5001 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5003 unsigned int i = 0, nkeys;
5006 MDB_page *mp = mc->mc_pg[mc->mc_top];
5007 MDB_node *node = NULL;
5012 nkeys = NUMKEYS(mp);
5014 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5015 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5018 low = IS_LEAF(mp) ? 0 : 1;
5020 cmp = mc->mc_dbx->md_cmp;
5022 /* Branch pages have no data, so if using integer keys,
5023 * alignment is guaranteed. Use faster mdb_cmp_int.
5025 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5026 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5033 nodekey.mv_size = mc->mc_db->md_pad;
5034 node = NODEPTR(mp, 0); /* fake */
5035 while (low <= high) {
5036 i = (low + high) >> 1;
5037 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5038 rc = cmp(key, &nodekey);
5039 DPRINTF(("found leaf index %u [%s], rc = %i",
5040 i, DKEY(&nodekey), rc));
5049 while (low <= high) {
5050 i = (low + high) >> 1;
5052 node = NODEPTR(mp, i);
5053 nodekey.mv_size = NODEKSZ(node);
5054 nodekey.mv_data = NODEKEY(node);
5056 rc = cmp(key, &nodekey);
5059 DPRINTF(("found leaf index %u [%s], rc = %i",
5060 i, DKEY(&nodekey), rc));
5062 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5063 i, DKEY(&nodekey), NODEPGNO(node), rc));
5074 if (rc > 0) { /* Found entry is less than the key. */
5075 i++; /* Skip to get the smallest entry larger than key. */
5077 node = NODEPTR(mp, i);
5080 *exactp = (rc == 0 && nkeys > 0);
5081 /* store the key index */
5082 mc->mc_ki[mc->mc_top] = i;
5084 /* There is no entry larger or equal to the key. */
5087 /* nodeptr is fake for LEAF2 */
5093 mdb_cursor_adjust(MDB_cursor *mc, func)
5097 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5098 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5105 /** Pop a page off the top of the cursor's stack. */
5107 mdb_cursor_pop(MDB_cursor *mc)
5111 MDB_page *top = mc->mc_pg[mc->mc_top];
5117 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5118 DDBI(mc), (void *) mc));
5122 /** Push a page onto the top of the cursor's stack. */
5124 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5126 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5127 DDBI(mc), (void *) mc));
5129 if (mc->mc_snum >= CURSOR_STACK) {
5130 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5131 return MDB_CURSOR_FULL;
5134 mc->mc_top = mc->mc_snum++;
5135 mc->mc_pg[mc->mc_top] = mp;
5136 mc->mc_ki[mc->mc_top] = 0;
5141 /** Find the address of the page corresponding to a given page number.
5142 * @param[in] txn the transaction for this access.
5143 * @param[in] pgno the page number for the page to retrieve.
5144 * @param[out] ret address of a pointer where the page's address will be stored.
5145 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5146 * @return 0 on success, non-zero on failure.
5149 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5151 MDB_env *env = txn->mt_env;
5155 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5159 MDB_ID2L dl = tx2->mt_u.dirty_list;
5161 /* Spilled pages were dirtied in this txn and flushed
5162 * because the dirty list got full. Bring this page
5163 * back in from the map (but don't unspill it here,
5164 * leave that unless page_touch happens again).
5166 if (tx2->mt_spill_pgs) {
5167 MDB_ID pn = pgno << 1;
5168 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5169 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5170 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5175 unsigned x = mdb_mid2l_search(dl, pgno);
5176 if (x <= dl[0].mid && dl[x].mid == pgno) {
5182 } while ((tx2 = tx2->mt_parent) != NULL);
5185 if (pgno < txn->mt_next_pgno) {
5187 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5189 DPRINTF(("page %"Z"u not found", pgno));
5190 txn->mt_flags |= MDB_TXN_ERROR;
5191 return MDB_PAGE_NOTFOUND;
5201 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5202 * The cursor is at the root page, set up the rest of it.
5205 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5207 MDB_page *mp = mc->mc_pg[mc->mc_top];
5211 while (IS_BRANCH(mp)) {
5215 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5216 mdb_cassert(mc, NUMKEYS(mp) > 1);
5217 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5219 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5221 if (flags & MDB_PS_LAST)
5222 i = NUMKEYS(mp) - 1;
5225 node = mdb_node_search(mc, key, &exact);
5227 i = NUMKEYS(mp) - 1;
5229 i = mc->mc_ki[mc->mc_top];
5231 mdb_cassert(mc, i > 0);
5235 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5238 mdb_cassert(mc, i < NUMKEYS(mp));
5239 node = NODEPTR(mp, i);
5241 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5244 mc->mc_ki[mc->mc_top] = i;
5245 if ((rc = mdb_cursor_push(mc, mp)))
5248 if (flags & MDB_PS_MODIFY) {
5249 if ((rc = mdb_page_touch(mc)) != 0)
5251 mp = mc->mc_pg[mc->mc_top];
5256 DPRINTF(("internal error, index points to a %02X page!?",
5258 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5259 return MDB_CORRUPTED;
5262 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5263 key ? DKEY(key) : "null"));
5264 mc->mc_flags |= C_INITIALIZED;
5265 mc->mc_flags &= ~C_EOF;
5270 /** Search for the lowest key under the current branch page.
5271 * This just bypasses a NUMKEYS check in the current page
5272 * before calling mdb_page_search_root(), because the callers
5273 * are all in situations where the current page is known to
5277 mdb_page_search_lowest(MDB_cursor *mc)
5279 MDB_page *mp = mc->mc_pg[mc->mc_top];
5280 MDB_node *node = NODEPTR(mp, 0);
5283 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5286 mc->mc_ki[mc->mc_top] = 0;
5287 if ((rc = mdb_cursor_push(mc, mp)))
5289 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5292 /** Search for the page a given key should be in.
5293 * Push it and its parent pages on the cursor stack.
5294 * @param[in,out] mc the cursor for this operation.
5295 * @param[in] key the key to search for, or NULL for first/last page.
5296 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5297 * are touched (updated with new page numbers).
5298 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5299 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5300 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5301 * @return 0 on success, non-zero on failure.
5304 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5309 /* Make sure the txn is still viable, then find the root from
5310 * the txn's db table and set it as the root of the cursor's stack.
5312 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5313 DPUTS("transaction has failed, must abort");
5316 /* Make sure we're using an up-to-date root */
5317 if (*mc->mc_dbflag & DB_STALE) {
5319 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5321 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5322 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5329 MDB_node *leaf = mdb_node_search(&mc2,
5330 &mc->mc_dbx->md_name, &exact);
5332 return MDB_NOTFOUND;
5333 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5336 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5338 /* The txn may not know this DBI, or another process may
5339 * have dropped and recreated the DB with other flags.
5341 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5342 return MDB_INCOMPATIBLE;
5343 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5345 *mc->mc_dbflag &= ~DB_STALE;
5347 root = mc->mc_db->md_root;
5349 if (root == P_INVALID) { /* Tree is empty. */
5350 DPUTS("tree is empty");
5351 return MDB_NOTFOUND;
5355 mdb_cassert(mc, root > 1);
5356 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5357 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5363 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5364 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5366 if (flags & MDB_PS_MODIFY) {
5367 if ((rc = mdb_page_touch(mc)))
5371 if (flags & MDB_PS_ROOTONLY)
5374 return mdb_page_search_root(mc, key, flags);
5378 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5380 MDB_txn *txn = mc->mc_txn;
5381 pgno_t pg = mp->mp_pgno;
5382 unsigned x = 0, ovpages = mp->mp_pages;
5383 MDB_env *env = txn->mt_env;
5384 MDB_IDL sl = txn->mt_spill_pgs;
5385 MDB_ID pn = pg << 1;
5388 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5389 /* If the page is dirty or on the spill list we just acquired it,
5390 * so we should give it back to our current free list, if any.
5391 * Otherwise put it onto the list of pages we freed in this txn.
5393 * Won't create me_pghead: me_pglast must be inited along with it.
5394 * Unsupported in nested txns: They would need to hide the page
5395 * range in ancestor txns' dirty and spilled lists.
5397 if (env->me_pghead &&
5399 ((mp->mp_flags & P_DIRTY) ||
5400 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5404 MDB_ID2 *dl, ix, iy;
5405 rc = mdb_midl_need(&env->me_pghead, ovpages);
5408 if (!(mp->mp_flags & P_DIRTY)) {
5409 /* This page is no longer spilled */
5416 /* Remove from dirty list */
5417 dl = txn->mt_u.dirty_list;
5419 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5425 mdb_cassert(mc, x > 1);
5427 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5428 txn->mt_flags |= MDB_TXN_ERROR;
5429 return MDB_CORRUPTED;
5432 if (!(env->me_flags & MDB_WRITEMAP))
5433 mdb_dpage_free(env, mp);
5435 /* Insert in me_pghead */
5436 mop = env->me_pghead;
5437 j = mop[0] + ovpages;
5438 for (i = mop[0]; i && mop[i] < pg; i--)
5444 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5448 mc->mc_db->md_overflow_pages -= ovpages;
5452 /** Return the data associated with a given node.
5453 * @param[in] txn The transaction for this operation.
5454 * @param[in] leaf The node being read.
5455 * @param[out] data Updated to point to the node's data.
5456 * @return 0 on success, non-zero on failure.
5459 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5461 MDB_page *omp; /* overflow page */
5465 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5466 data->mv_size = NODEDSZ(leaf);
5467 data->mv_data = NODEDATA(leaf);
5471 /* Read overflow data.
5473 data->mv_size = NODEDSZ(leaf);
5474 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5475 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5476 DPRINTF(("read overflow page %"Z"u failed", pgno));
5479 data->mv_data = METADATA(omp);
5485 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5486 MDB_val *key, MDB_val *data)
5493 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5495 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5498 if (txn->mt_flags & MDB_TXN_ERROR)
5501 mdb_cursor_init(&mc, txn, dbi, &mx);
5502 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5505 /** Find a sibling for a page.
5506 * Replaces the page at the top of the cursor's stack with the
5507 * specified sibling, if one exists.
5508 * @param[in] mc The cursor for this operation.
5509 * @param[in] move_right Non-zero if the right sibling is requested,
5510 * otherwise the left sibling.
5511 * @return 0 on success, non-zero on failure.
5514 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5520 if (mc->mc_snum < 2) {
5521 return MDB_NOTFOUND; /* root has no siblings */
5525 DPRINTF(("parent page is page %"Z"u, index %u",
5526 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5528 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5529 : (mc->mc_ki[mc->mc_top] == 0)) {
5530 DPRINTF(("no more keys left, moving to %s sibling",
5531 move_right ? "right" : "left"));
5532 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5533 /* undo cursor_pop before returning */
5540 mc->mc_ki[mc->mc_top]++;
5542 mc->mc_ki[mc->mc_top]--;
5543 DPRINTF(("just moving to %s index key %u",
5544 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5546 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5548 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5549 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5550 /* mc will be inconsistent if caller does mc_snum++ as above */
5551 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5555 mdb_cursor_push(mc, mp);
5557 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5562 /** Move the cursor to the next data item. */
5564 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5570 if (mc->mc_flags & C_EOF) {
5571 return MDB_NOTFOUND;
5574 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5576 mp = mc->mc_pg[mc->mc_top];
5578 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5579 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5580 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5581 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5582 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5583 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5584 if (rc == MDB_SUCCESS)
5585 MDB_GET_KEY(leaf, key);
5590 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5591 if (op == MDB_NEXT_DUP)
5592 return MDB_NOTFOUND;
5596 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5597 mdb_dbg_pgno(mp), (void *) mc));
5598 if (mc->mc_flags & C_DEL)
5601 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5602 DPUTS("=====> move to next sibling page");
5603 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5604 mc->mc_flags |= C_EOF;
5607 mp = mc->mc_pg[mc->mc_top];
5608 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5610 mc->mc_ki[mc->mc_top]++;
5613 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5614 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5617 key->mv_size = mc->mc_db->md_pad;
5618 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5622 mdb_cassert(mc, IS_LEAF(mp));
5623 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5625 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5626 mdb_xcursor_init1(mc, leaf);
5629 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5632 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5633 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5634 if (rc != MDB_SUCCESS)
5639 MDB_GET_KEY(leaf, key);
5643 /** Move the cursor to the previous data item. */
5645 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5651 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5653 mp = mc->mc_pg[mc->mc_top];
5655 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5656 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5657 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5658 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5659 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5660 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5661 if (rc == MDB_SUCCESS) {
5662 MDB_GET_KEY(leaf, key);
5663 mc->mc_flags &= ~C_EOF;
5669 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5670 if (op == MDB_PREV_DUP)
5671 return MDB_NOTFOUND;
5675 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5676 mdb_dbg_pgno(mp), (void *) mc));
5678 if (mc->mc_ki[mc->mc_top] == 0) {
5679 DPUTS("=====> move to prev sibling page");
5680 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5683 mp = mc->mc_pg[mc->mc_top];
5684 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5685 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5687 mc->mc_ki[mc->mc_top]--;
5689 mc->mc_flags &= ~C_EOF;
5691 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5692 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5695 key->mv_size = mc->mc_db->md_pad;
5696 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5700 mdb_cassert(mc, IS_LEAF(mp));
5701 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5703 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5704 mdb_xcursor_init1(mc, leaf);
5707 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5710 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5711 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5712 if (rc != MDB_SUCCESS)
5717 MDB_GET_KEY(leaf, key);
5721 /** Set the cursor on a specific data item. */
5723 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5724 MDB_cursor_op op, int *exactp)
5728 MDB_node *leaf = NULL;
5731 if (key->mv_size == 0)
5732 return MDB_BAD_VALSIZE;
5735 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5737 /* See if we're already on the right page */
5738 if (mc->mc_flags & C_INITIALIZED) {
5741 mp = mc->mc_pg[mc->mc_top];
5743 mc->mc_ki[mc->mc_top] = 0;
5744 return MDB_NOTFOUND;
5746 if (mp->mp_flags & P_LEAF2) {
5747 nodekey.mv_size = mc->mc_db->md_pad;
5748 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5750 leaf = NODEPTR(mp, 0);
5751 MDB_GET_KEY2(leaf, nodekey);
5753 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5755 /* Probably happens rarely, but first node on the page
5756 * was the one we wanted.
5758 mc->mc_ki[mc->mc_top] = 0;
5765 unsigned int nkeys = NUMKEYS(mp);
5767 if (mp->mp_flags & P_LEAF2) {
5768 nodekey.mv_data = LEAF2KEY(mp,
5769 nkeys-1, nodekey.mv_size);
5771 leaf = NODEPTR(mp, nkeys-1);
5772 MDB_GET_KEY2(leaf, nodekey);
5774 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5776 /* last node was the one we wanted */
5777 mc->mc_ki[mc->mc_top] = nkeys-1;
5783 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5784 /* This is definitely the right page, skip search_page */
5785 if (mp->mp_flags & P_LEAF2) {
5786 nodekey.mv_data = LEAF2KEY(mp,
5787 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5789 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5790 MDB_GET_KEY2(leaf, nodekey);
5792 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5794 /* current node was the one we wanted */
5804 /* If any parents have right-sibs, search.
5805 * Otherwise, there's nothing further.
5807 for (i=0; i<mc->mc_top; i++)
5809 NUMKEYS(mc->mc_pg[i])-1)
5811 if (i == mc->mc_top) {
5812 /* There are no other pages */
5813 mc->mc_ki[mc->mc_top] = nkeys;
5814 return MDB_NOTFOUND;
5818 /* There are no other pages */
5819 mc->mc_ki[mc->mc_top] = 0;
5820 if (op == MDB_SET_RANGE && !exactp) {
5824 return MDB_NOTFOUND;
5828 rc = mdb_page_search(mc, key, 0);
5829 if (rc != MDB_SUCCESS)
5832 mp = mc->mc_pg[mc->mc_top];
5833 mdb_cassert(mc, IS_LEAF(mp));
5836 leaf = mdb_node_search(mc, key, exactp);
5837 if (exactp != NULL && !*exactp) {
5838 /* MDB_SET specified and not an exact match. */
5839 return MDB_NOTFOUND;
5843 DPUTS("===> inexact leaf not found, goto sibling");
5844 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5845 return rc; /* no entries matched */
5846 mp = mc->mc_pg[mc->mc_top];
5847 mdb_cassert(mc, IS_LEAF(mp));
5848 leaf = NODEPTR(mp, 0);
5852 mc->mc_flags |= C_INITIALIZED;
5853 mc->mc_flags &= ~C_EOF;
5856 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5857 key->mv_size = mc->mc_db->md_pad;
5858 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5863 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5864 mdb_xcursor_init1(mc, leaf);
5867 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5868 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5869 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5872 if (op == MDB_GET_BOTH) {
5878 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5879 if (rc != MDB_SUCCESS)
5882 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5885 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5887 dcmp = mc->mc_dbx->md_dcmp;
5888 #if UINT_MAX < SIZE_MAX
5889 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5890 dcmp = mdb_cmp_clong;
5892 rc = dcmp(data, &olddata);
5894 if (op == MDB_GET_BOTH || rc > 0)
5895 return MDB_NOTFOUND;
5902 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5903 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5908 /* The key already matches in all other cases */
5909 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5910 MDB_GET_KEY(leaf, key);
5911 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5916 /** Move the cursor to the first item in the database. */
5918 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5924 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5926 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5927 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5928 if (rc != MDB_SUCCESS)
5931 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5933 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5934 mc->mc_flags |= C_INITIALIZED;
5935 mc->mc_flags &= ~C_EOF;
5937 mc->mc_ki[mc->mc_top] = 0;
5939 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5940 key->mv_size = mc->mc_db->md_pad;
5941 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5946 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5947 mdb_xcursor_init1(mc, leaf);
5948 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5952 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5956 MDB_GET_KEY(leaf, key);
5960 /** Move the cursor to the last item in the database. */
5962 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5968 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5970 if (!(mc->mc_flags & C_EOF)) {
5972 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5973 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5974 if (rc != MDB_SUCCESS)
5977 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5980 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5981 mc->mc_flags |= C_INITIALIZED|C_EOF;
5982 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5984 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5985 key->mv_size = mc->mc_db->md_pad;
5986 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5991 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5992 mdb_xcursor_init1(mc, leaf);
5993 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5997 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6002 MDB_GET_KEY(leaf, key);
6007 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6012 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6017 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6021 case MDB_GET_CURRENT:
6022 if (!(mc->mc_flags & C_INITIALIZED)) {
6025 MDB_page *mp = mc->mc_pg[mc->mc_top];
6026 int nkeys = NUMKEYS(mp);
6027 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6028 mc->mc_ki[mc->mc_top] = nkeys;
6034 key->mv_size = mc->mc_db->md_pad;
6035 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6037 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6038 MDB_GET_KEY(leaf, key);
6040 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6041 if (mc->mc_flags & C_DEL)
6042 mdb_xcursor_init1(mc, leaf);
6043 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6045 rc = mdb_node_read(mc->mc_txn, leaf, data);
6052 case MDB_GET_BOTH_RANGE:
6057 if (mc->mc_xcursor == NULL) {
6058 rc = MDB_INCOMPATIBLE;
6068 rc = mdb_cursor_set(mc, key, data, op,
6069 op == MDB_SET_RANGE ? NULL : &exact);
6072 case MDB_GET_MULTIPLE:
6073 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6077 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6078 rc = MDB_INCOMPATIBLE;
6082 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6083 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6086 case MDB_NEXT_MULTIPLE:
6091 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6092 rc = MDB_INCOMPATIBLE;
6095 if (!(mc->mc_flags & C_INITIALIZED))
6096 rc = mdb_cursor_first(mc, key, data);
6098 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6099 if (rc == MDB_SUCCESS) {
6100 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6103 mx = &mc->mc_xcursor->mx_cursor;
6104 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6106 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6107 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6115 case MDB_NEXT_NODUP:
6116 if (!(mc->mc_flags & C_INITIALIZED))
6117 rc = mdb_cursor_first(mc, key, data);
6119 rc = mdb_cursor_next(mc, key, data, op);
6123 case MDB_PREV_NODUP:
6124 if (!(mc->mc_flags & C_INITIALIZED)) {
6125 rc = mdb_cursor_last(mc, key, data);
6128 mc->mc_flags |= C_INITIALIZED;
6129 mc->mc_ki[mc->mc_top]++;
6131 rc = mdb_cursor_prev(mc, key, data, op);
6134 rc = mdb_cursor_first(mc, key, data);
6137 mfunc = mdb_cursor_first;
6139 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6143 if (mc->mc_xcursor == NULL) {
6144 rc = MDB_INCOMPATIBLE;
6148 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6149 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6150 MDB_GET_KEY(leaf, key);
6151 rc = mdb_node_read(mc->mc_txn, leaf, data);
6155 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6159 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6162 rc = mdb_cursor_last(mc, key, data);
6165 mfunc = mdb_cursor_last;
6168 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6173 if (mc->mc_flags & C_DEL)
6174 mc->mc_flags ^= C_DEL;
6179 /** Touch all the pages in the cursor stack. Set mc_top.
6180 * Makes sure all the pages are writable, before attempting a write operation.
6181 * @param[in] mc The cursor to operate on.
6184 mdb_cursor_touch(MDB_cursor *mc)
6186 int rc = MDB_SUCCESS;
6188 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6191 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6193 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6194 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6197 *mc->mc_dbflag |= DB_DIRTY;
6202 rc = mdb_page_touch(mc);
6203 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6204 mc->mc_top = mc->mc_snum-1;
6209 /** Do not spill pages to disk if txn is getting full, may fail instead */
6210 #define MDB_NOSPILL 0x8000
6213 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6217 MDB_node *leaf = NULL;
6220 MDB_val xdata, *rdata, dkey, olddata;
6222 int do_sub = 0, insert_key, insert_data;
6223 unsigned int mcount = 0, dcount = 0, nospill;
6226 unsigned int nflags;
6229 if (mc == NULL || key == NULL)
6232 env = mc->mc_txn->mt_env;
6234 /* Check this first so counter will always be zero on any
6237 if (flags & MDB_MULTIPLE) {
6238 dcount = data[1].mv_size;
6239 data[1].mv_size = 0;
6240 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6241 return MDB_INCOMPATIBLE;
6244 nospill = flags & MDB_NOSPILL;
6245 flags &= ~MDB_NOSPILL;
6247 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6248 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6250 if (key->mv_size-1 >= ENV_MAXKEY(env))
6251 return MDB_BAD_VALSIZE;
6253 #if SIZE_MAX > MAXDATASIZE
6254 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6255 return MDB_BAD_VALSIZE;
6257 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6258 return MDB_BAD_VALSIZE;
6261 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6262 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6266 if (flags == MDB_CURRENT) {
6267 if (!(mc->mc_flags & C_INITIALIZED))
6270 } else if (mc->mc_db->md_root == P_INVALID) {
6271 /* new database, cursor has nothing to point to */
6274 mc->mc_flags &= ~C_INITIALIZED;
6279 if (flags & MDB_APPEND) {
6281 rc = mdb_cursor_last(mc, &k2, &d2);
6283 rc = mc->mc_dbx->md_cmp(key, &k2);
6286 mc->mc_ki[mc->mc_top]++;
6288 /* new key is <= last key */
6293 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6295 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6296 DPRINTF(("duplicate key [%s]", DKEY(key)));
6298 return MDB_KEYEXIST;
6300 if (rc && rc != MDB_NOTFOUND)
6304 if (mc->mc_flags & C_DEL)
6305 mc->mc_flags ^= C_DEL;
6307 /* Cursor is positioned, check for room in the dirty list */
6309 if (flags & MDB_MULTIPLE) {
6311 xdata.mv_size = data->mv_size * dcount;
6315 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6319 if (rc == MDB_NO_ROOT) {
6321 /* new database, write a root leaf page */
6322 DPUTS("allocating new root leaf page");
6323 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6326 mdb_cursor_push(mc, np);
6327 mc->mc_db->md_root = np->mp_pgno;
6328 mc->mc_db->md_depth++;
6329 *mc->mc_dbflag |= DB_DIRTY;
6330 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6332 np->mp_flags |= P_LEAF2;
6333 mc->mc_flags |= C_INITIALIZED;
6335 /* make sure all cursor pages are writable */
6336 rc2 = mdb_cursor_touch(mc);
6341 insert_key = insert_data = rc;
6343 /* The key does not exist */
6344 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6345 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6346 LEAFSIZE(key, data) > env->me_nodemax)
6348 /* Too big for a node, insert in sub-DB. Set up an empty
6349 * "old sub-page" for prep_subDB to expand to a full page.
6351 fp_flags = P_LEAF|P_DIRTY;
6353 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6354 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6355 olddata.mv_size = PAGEHDRSZ;
6359 /* there's only a key anyway, so this is a no-op */
6360 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6362 unsigned int ksize = mc->mc_db->md_pad;
6363 if (key->mv_size != ksize)
6364 return MDB_BAD_VALSIZE;
6365 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6366 memcpy(ptr, key->mv_data, ksize);
6368 /* if overwriting slot 0 of leaf, need to
6369 * update branch key if there is a parent page
6371 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6372 unsigned short top = mc->mc_top;
6374 /* slot 0 is always an empty key, find real slot */
6375 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6377 if (mc->mc_ki[mc->mc_top])
6378 rc2 = mdb_update_key(mc, key);
6389 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6390 olddata.mv_size = NODEDSZ(leaf);
6391 olddata.mv_data = NODEDATA(leaf);
6394 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6395 /* Prepare (sub-)page/sub-DB to accept the new item,
6396 * if needed. fp: old sub-page or a header faking
6397 * it. mp: new (sub-)page. offset: growth in page
6398 * size. xdata: node data with new page or DB.
6400 unsigned i, offset = 0;
6401 mp = fp = xdata.mv_data = env->me_pbuf;
6402 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6404 /* Was a single item before, must convert now */
6405 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6407 /* Just overwrite the current item */
6408 if (flags == MDB_CURRENT)
6410 dcmp = mc->mc_dbx->md_dcmp;
6411 #if UINT_MAX < SIZE_MAX
6412 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6413 dcmp = mdb_cmp_clong;
6415 /* does data match? */
6416 if (!dcmp(data, &olddata)) {
6417 if (flags & MDB_NODUPDATA)
6418 return MDB_KEYEXIST;
6423 /* Back up original data item */
6424 dkey.mv_size = olddata.mv_size;
6425 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6427 /* Make sub-page header for the dup items, with dummy body */
6428 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6429 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6430 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6431 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6432 fp->mp_flags |= P_LEAF2;
6433 fp->mp_pad = data->mv_size;
6434 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6436 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6437 (dkey.mv_size & 1) + (data->mv_size & 1);
6439 fp->mp_upper = xdata.mv_size - PAGEBASE;
6440 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6441 } else if (leaf->mn_flags & F_SUBDATA) {
6442 /* Data is on sub-DB, just store it */
6443 flags |= F_DUPDATA|F_SUBDATA;
6446 /* Data is on sub-page */
6447 fp = olddata.mv_data;
6450 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6451 offset = EVEN(NODESIZE + sizeof(indx_t) +
6455 offset = fp->mp_pad;
6456 if (SIZELEFT(fp) < offset) {
6457 offset *= 4; /* space for 4 more */
6460 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6462 fp->mp_flags |= P_DIRTY;
6463 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6464 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6468 xdata.mv_size = olddata.mv_size + offset;
6471 fp_flags = fp->mp_flags;
6472 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6473 /* Too big for a sub-page, convert to sub-DB */
6474 fp_flags &= ~P_SUBP;
6476 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6477 fp_flags |= P_LEAF2;
6478 dummy.md_pad = fp->mp_pad;
6479 dummy.md_flags = MDB_DUPFIXED;
6480 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6481 dummy.md_flags |= MDB_INTEGERKEY;
6487 dummy.md_branch_pages = 0;
6488 dummy.md_leaf_pages = 1;
6489 dummy.md_overflow_pages = 0;
6490 dummy.md_entries = NUMKEYS(fp);
6491 xdata.mv_size = sizeof(MDB_db);
6492 xdata.mv_data = &dummy;
6493 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6495 offset = env->me_psize - olddata.mv_size;
6496 flags |= F_DUPDATA|F_SUBDATA;
6497 dummy.md_root = mp->mp_pgno;
6500 mp->mp_flags = fp_flags | P_DIRTY;
6501 mp->mp_pad = fp->mp_pad;
6502 mp->mp_lower = fp->mp_lower;
6503 mp->mp_upper = fp->mp_upper + offset;
6504 if (fp_flags & P_LEAF2) {
6505 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6507 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6508 olddata.mv_size - fp->mp_upper - PAGEBASE);
6509 for (i=0; i<NUMKEYS(fp); i++)
6510 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6518 mdb_node_del(mc, 0);
6522 /* overflow page overwrites need special handling */
6523 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6526 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6528 memcpy(&pg, olddata.mv_data, sizeof(pg));
6529 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6531 ovpages = omp->mp_pages;
6533 /* Is the ov page large enough? */
6534 if (ovpages >= dpages) {
6535 if (!(omp->mp_flags & P_DIRTY) &&
6536 (level || (env->me_flags & MDB_WRITEMAP)))
6538 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6541 level = 0; /* dirty in this txn or clean */
6544 if (omp->mp_flags & P_DIRTY) {
6545 /* yes, overwrite it. Note in this case we don't
6546 * bother to try shrinking the page if the new data
6547 * is smaller than the overflow threshold.
6550 /* It is writable only in a parent txn */
6551 size_t sz = (size_t) env->me_psize * ovpages, off;
6552 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6558 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6559 mdb_cassert(mc, rc2 == 0);
6560 if (!(flags & MDB_RESERVE)) {
6561 /* Copy end of page, adjusting alignment so
6562 * compiler may copy words instead of bytes.
6564 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6565 memcpy((size_t *)((char *)np + off),
6566 (size_t *)((char *)omp + off), sz - off);
6569 memcpy(np, omp, sz); /* Copy beginning of page */
6572 SETDSZ(leaf, data->mv_size);
6573 if (F_ISSET(flags, MDB_RESERVE))
6574 data->mv_data = METADATA(omp);
6576 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6580 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6582 } else if (data->mv_size == olddata.mv_size) {
6583 /* same size, just replace it. Note that we could
6584 * also reuse this node if the new data is smaller,
6585 * but instead we opt to shrink the node in that case.
6587 if (F_ISSET(flags, MDB_RESERVE))
6588 data->mv_data = olddata.mv_data;
6589 else if (!(mc->mc_flags & C_SUB))
6590 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6592 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6597 mdb_node_del(mc, 0);
6603 nflags = flags & NODE_ADD_FLAGS;
6604 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6605 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6606 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6607 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6609 nflags |= MDB_SPLIT_REPLACE;
6610 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6612 /* There is room already in this leaf page. */
6613 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6614 if (rc == 0 && insert_key) {
6615 /* Adjust other cursors pointing to mp */
6616 MDB_cursor *m2, *m3;
6617 MDB_dbi dbi = mc->mc_dbi;
6618 unsigned i = mc->mc_top;
6619 MDB_page *mp = mc->mc_pg[i];
6621 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6622 if (mc->mc_flags & C_SUB)
6623 m3 = &m2->mc_xcursor->mx_cursor;
6626 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6627 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6634 if (rc == MDB_SUCCESS) {
6635 /* Now store the actual data in the child DB. Note that we're
6636 * storing the user data in the keys field, so there are strict
6637 * size limits on dupdata. The actual data fields of the child
6638 * DB are all zero size.
6646 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6647 if (flags & MDB_CURRENT) {
6648 xflags = MDB_CURRENT|MDB_NOSPILL;
6650 mdb_xcursor_init1(mc, leaf);
6651 xflags = (flags & MDB_NODUPDATA) ?
6652 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6654 /* converted, write the original data first */
6656 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6660 /* Adjust other cursors pointing to mp */
6662 unsigned i = mc->mc_top;
6663 MDB_page *mp = mc->mc_pg[i];
6665 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6666 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6667 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6668 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6669 mdb_xcursor_init1(m2, leaf);
6673 /* we've done our job */
6676 ecount = mc->mc_xcursor->mx_db.md_entries;
6677 if (flags & MDB_APPENDDUP)
6678 xflags |= MDB_APPEND;
6679 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6680 if (flags & F_SUBDATA) {
6681 void *db = NODEDATA(leaf);
6682 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6684 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6686 /* Increment count unless we just replaced an existing item. */
6688 mc->mc_db->md_entries++;
6690 /* Invalidate txn if we created an empty sub-DB */
6693 /* If we succeeded and the key didn't exist before,
6694 * make sure the cursor is marked valid.
6696 mc->mc_flags |= C_INITIALIZED;
6698 if (flags & MDB_MULTIPLE) {
6701 /* let caller know how many succeeded, if any */
6702 data[1].mv_size = mcount;
6703 if (mcount < dcount) {
6704 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6705 insert_key = insert_data = 0;
6712 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6715 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6720 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6726 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6727 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6729 if (!(mc->mc_flags & C_INITIALIZED))
6732 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6733 return MDB_NOTFOUND;
6735 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6738 rc = mdb_cursor_touch(mc);
6742 mp = mc->mc_pg[mc->mc_top];
6745 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6747 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6748 if (flags & MDB_NODUPDATA) {
6749 /* mdb_cursor_del0() will subtract the final entry */
6750 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6752 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6753 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6755 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6758 /* If sub-DB still has entries, we're done */
6759 if (mc->mc_xcursor->mx_db.md_entries) {
6760 if (leaf->mn_flags & F_SUBDATA) {
6761 /* update subDB info */
6762 void *db = NODEDATA(leaf);
6763 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6766 /* shrink fake page */
6767 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6768 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6769 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6770 /* fix other sub-DB cursors pointed at this fake page */
6771 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6772 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6773 if (m2->mc_pg[mc->mc_top] == mp &&
6774 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6775 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6778 mc->mc_db->md_entries--;
6779 mc->mc_flags |= C_DEL;
6782 /* otherwise fall thru and delete the sub-DB */
6785 if (leaf->mn_flags & F_SUBDATA) {
6786 /* add all the child DB's pages to the free list */
6787 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6793 /* add overflow pages to free list */
6794 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6798 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6799 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6800 (rc = mdb_ovpage_free(mc, omp)))
6805 return mdb_cursor_del0(mc);
6808 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6812 /** Allocate and initialize new pages for a database.
6813 * @param[in] mc a cursor on the database being added to.
6814 * @param[in] flags flags defining what type of page is being allocated.
6815 * @param[in] num the number of pages to allocate. This is usually 1,
6816 * unless allocating overflow pages for a large record.
6817 * @param[out] mp Address of a page, or NULL on failure.
6818 * @return 0 on success, non-zero on failure.
6821 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6826 if ((rc = mdb_page_alloc(mc, num, &np)))
6828 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6829 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6830 np->mp_flags = flags | P_DIRTY;
6831 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6832 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6835 mc->mc_db->md_branch_pages++;
6836 else if (IS_LEAF(np))
6837 mc->mc_db->md_leaf_pages++;
6838 else if (IS_OVERFLOW(np)) {
6839 mc->mc_db->md_overflow_pages += num;
6847 /** Calculate the size of a leaf node.
6848 * The size depends on the environment's page size; if a data item
6849 * is too large it will be put onto an overflow page and the node
6850 * size will only include the key and not the data. Sizes are always
6851 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6852 * of the #MDB_node headers.
6853 * @param[in] env The environment handle.
6854 * @param[in] key The key for the node.
6855 * @param[in] data The data for the node.
6856 * @return The number of bytes needed to store the node.
6859 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6863 sz = LEAFSIZE(key, data);
6864 if (sz > env->me_nodemax) {
6865 /* put on overflow page */
6866 sz -= data->mv_size - sizeof(pgno_t);
6869 return EVEN(sz + sizeof(indx_t));
6872 /** Calculate the size of a branch node.
6873 * The size should depend on the environment's page size but since
6874 * we currently don't support spilling large keys onto overflow
6875 * pages, it's simply the size of the #MDB_node header plus the
6876 * size of the key. Sizes are always rounded up to an even number
6877 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6878 * @param[in] env The environment handle.
6879 * @param[in] key The key for the node.
6880 * @return The number of bytes needed to store the node.
6883 mdb_branch_size(MDB_env *env, MDB_val *key)
6888 if (sz > env->me_nodemax) {
6889 /* put on overflow page */
6890 /* not implemented */
6891 /* sz -= key->size - sizeof(pgno_t); */
6894 return sz + sizeof(indx_t);
6897 /** Add a node to the page pointed to by the cursor.
6898 * @param[in] mc The cursor for this operation.
6899 * @param[in] indx The index on the page where the new node should be added.
6900 * @param[in] key The key for the new node.
6901 * @param[in] data The data for the new node, if any.
6902 * @param[in] pgno The page number, if adding a branch node.
6903 * @param[in] flags Flags for the node.
6904 * @return 0 on success, non-zero on failure. Possible errors are:
6906 * <li>ENOMEM - failed to allocate overflow pages for the node.
6907 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6908 * should never happen since all callers already calculate the
6909 * page's free space before calling this function.
6913 mdb_node_add(MDB_cursor *mc, indx_t indx,
6914 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6917 size_t node_size = NODESIZE;
6921 MDB_page *mp = mc->mc_pg[mc->mc_top];
6922 MDB_page *ofp = NULL; /* overflow page */
6925 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6927 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6928 IS_LEAF(mp) ? "leaf" : "branch",
6929 IS_SUBP(mp) ? "sub-" : "",
6930 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6931 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6934 /* Move higher keys up one slot. */
6935 int ksize = mc->mc_db->md_pad, dif;
6936 char *ptr = LEAF2KEY(mp, indx, ksize);
6937 dif = NUMKEYS(mp) - indx;
6939 memmove(ptr+ksize, ptr, dif*ksize);
6940 /* insert new key */
6941 memcpy(ptr, key->mv_data, ksize);
6943 /* Just using these for counting */
6944 mp->mp_lower += sizeof(indx_t);
6945 mp->mp_upper -= ksize - sizeof(indx_t);
6949 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6951 node_size += key->mv_size;
6953 mdb_cassert(mc, data);
6954 if (F_ISSET(flags, F_BIGDATA)) {
6955 /* Data already on overflow page. */
6956 node_size += sizeof(pgno_t);
6957 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6958 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6960 /* Put data on overflow page. */
6961 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6962 data->mv_size, node_size+data->mv_size));
6963 node_size = EVEN(node_size + sizeof(pgno_t));
6964 if ((ssize_t)node_size > room)
6966 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6968 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6972 node_size += data->mv_size;
6975 node_size = EVEN(node_size);
6976 if ((ssize_t)node_size > room)
6980 /* Move higher pointers up one slot. */
6981 for (i = NUMKEYS(mp); i > indx; i--)
6982 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6984 /* Adjust free space offsets. */
6985 ofs = mp->mp_upper - node_size;
6986 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6987 mp->mp_ptrs[indx] = ofs;
6989 mp->mp_lower += sizeof(indx_t);
6991 /* Write the node data. */
6992 node = NODEPTR(mp, indx);
6993 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6994 node->mn_flags = flags;
6996 SETDSZ(node,data->mv_size);
7001 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7004 mdb_cassert(mc, key);
7006 if (F_ISSET(flags, F_BIGDATA))
7007 memcpy(node->mn_data + key->mv_size, data->mv_data,
7009 else if (F_ISSET(flags, MDB_RESERVE))
7010 data->mv_data = node->mn_data + key->mv_size;
7012 memcpy(node->mn_data + key->mv_size, data->mv_data,
7015 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
7017 if (F_ISSET(flags, MDB_RESERVE))
7018 data->mv_data = METADATA(ofp);
7020 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
7027 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7028 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7029 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7030 DPRINTF(("node size = %"Z"u", node_size));
7031 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7032 return MDB_PAGE_FULL;
7035 /** Delete the specified node from a page.
7036 * @param[in] mc Cursor pointing to the node to delete.
7037 * @param[in] ksize The size of a node. Only used if the page is
7038 * part of a #MDB_DUPFIXED database.
7041 mdb_node_del(MDB_cursor *mc, int ksize)
7043 MDB_page *mp = mc->mc_pg[mc->mc_top];
7044 indx_t indx = mc->mc_ki[mc->mc_top];
7046 indx_t i, j, numkeys, ptr;
7050 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7051 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7052 numkeys = NUMKEYS(mp);
7053 mdb_cassert(mc, indx < numkeys);
7056 int x = numkeys - 1 - indx;
7057 base = LEAF2KEY(mp, indx, ksize);
7059 memmove(base, base + ksize, x * ksize);
7060 mp->mp_lower -= sizeof(indx_t);
7061 mp->mp_upper += ksize - sizeof(indx_t);
7065 node = NODEPTR(mp, indx);
7066 sz = NODESIZE + node->mn_ksize;
7068 if (F_ISSET(node->mn_flags, F_BIGDATA))
7069 sz += sizeof(pgno_t);
7071 sz += NODEDSZ(node);
7075 ptr = mp->mp_ptrs[indx];
7076 for (i = j = 0; i < numkeys; i++) {
7078 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7079 if (mp->mp_ptrs[i] < ptr)
7080 mp->mp_ptrs[j] += sz;
7085 base = (char *)mp + mp->mp_upper + PAGEBASE;
7086 memmove(base + sz, base, ptr - mp->mp_upper);
7088 mp->mp_lower -= sizeof(indx_t);
7092 /** Compact the main page after deleting a node on a subpage.
7093 * @param[in] mp The main page to operate on.
7094 * @param[in] indx The index of the subpage on the main page.
7097 mdb_node_shrink(MDB_page *mp, indx_t indx)
7103 indx_t i, numkeys, ptr;
7105 node = NODEPTR(mp, indx);
7106 sp = (MDB_page *)NODEDATA(node);
7107 delta = SIZELEFT(sp);
7108 xp = (MDB_page *)((char *)sp + delta);
7110 /* shift subpage upward */
7112 nsize = NUMKEYS(sp) * sp->mp_pad;
7114 return; /* do not make the node uneven-sized */
7115 memmove(METADATA(xp), METADATA(sp), nsize);
7118 numkeys = NUMKEYS(sp);
7119 for (i=numkeys-1; i>=0; i--)
7120 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7122 xp->mp_upper = sp->mp_lower;
7123 xp->mp_lower = sp->mp_lower;
7124 xp->mp_flags = sp->mp_flags;
7125 xp->mp_pad = sp->mp_pad;
7126 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7128 nsize = NODEDSZ(node) - delta;
7129 SETDSZ(node, nsize);
7131 /* shift lower nodes upward */
7132 ptr = mp->mp_ptrs[indx];
7133 numkeys = NUMKEYS(mp);
7134 for (i = 0; i < numkeys; i++) {
7135 if (mp->mp_ptrs[i] <= ptr)
7136 mp->mp_ptrs[i] += delta;
7139 base = (char *)mp + mp->mp_upper + PAGEBASE;
7140 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7141 mp->mp_upper += delta;
7144 /** Initial setup of a sorted-dups cursor.
7145 * Sorted duplicates are implemented as a sub-database for the given key.
7146 * The duplicate data items are actually keys of the sub-database.
7147 * Operations on the duplicate data items are performed using a sub-cursor
7148 * initialized when the sub-database is first accessed. This function does
7149 * the preliminary setup of the sub-cursor, filling in the fields that
7150 * depend only on the parent DB.
7151 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7154 mdb_xcursor_init0(MDB_cursor *mc)
7156 MDB_xcursor *mx = mc->mc_xcursor;
7158 mx->mx_cursor.mc_xcursor = NULL;
7159 mx->mx_cursor.mc_txn = mc->mc_txn;
7160 mx->mx_cursor.mc_db = &mx->mx_db;
7161 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7162 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7163 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7164 mx->mx_cursor.mc_snum = 0;
7165 mx->mx_cursor.mc_top = 0;
7166 mx->mx_cursor.mc_flags = C_SUB;
7167 mx->mx_dbx.md_name.mv_size = 0;
7168 mx->mx_dbx.md_name.mv_data = NULL;
7169 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7170 mx->mx_dbx.md_dcmp = NULL;
7171 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7174 /** Final setup of a sorted-dups cursor.
7175 * Sets up the fields that depend on the data from the main cursor.
7176 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7177 * @param[in] node The data containing the #MDB_db record for the
7178 * sorted-dup database.
7181 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7183 MDB_xcursor *mx = mc->mc_xcursor;
7185 if (node->mn_flags & F_SUBDATA) {
7186 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7187 mx->mx_cursor.mc_pg[0] = 0;
7188 mx->mx_cursor.mc_snum = 0;
7189 mx->mx_cursor.mc_top = 0;
7190 mx->mx_cursor.mc_flags = C_SUB;
7192 MDB_page *fp = NODEDATA(node);
7193 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7194 mx->mx_db.md_flags = 0;
7195 mx->mx_db.md_depth = 1;
7196 mx->mx_db.md_branch_pages = 0;
7197 mx->mx_db.md_leaf_pages = 1;
7198 mx->mx_db.md_overflow_pages = 0;
7199 mx->mx_db.md_entries = NUMKEYS(fp);
7200 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7201 mx->mx_cursor.mc_snum = 1;
7202 mx->mx_cursor.mc_top = 0;
7203 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7204 mx->mx_cursor.mc_pg[0] = fp;
7205 mx->mx_cursor.mc_ki[0] = 0;
7206 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7207 mx->mx_db.md_flags = MDB_DUPFIXED;
7208 mx->mx_db.md_pad = fp->mp_pad;
7209 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7210 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7213 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7214 mx->mx_db.md_root));
7215 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7216 #if UINT_MAX < SIZE_MAX
7217 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7218 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7222 /** Initialize a cursor for a given transaction and database. */
7224 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7227 mc->mc_backup = NULL;
7230 mc->mc_db = &txn->mt_dbs[dbi];
7231 mc->mc_dbx = &txn->mt_dbxs[dbi];
7232 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7238 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7239 mdb_tassert(txn, mx != NULL);
7240 mc->mc_xcursor = mx;
7241 mdb_xcursor_init0(mc);
7243 mc->mc_xcursor = NULL;
7245 if (*mc->mc_dbflag & DB_STALE) {
7246 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7251 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7254 size_t size = sizeof(MDB_cursor);
7256 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7259 if (txn->mt_flags & MDB_TXN_ERROR)
7262 /* Allow read access to the freelist */
7263 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7266 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7267 size += sizeof(MDB_xcursor);
7269 if ((mc = malloc(size)) != NULL) {
7270 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7271 if (txn->mt_cursors) {
7272 mc->mc_next = txn->mt_cursors[dbi];
7273 txn->mt_cursors[dbi] = mc;
7274 mc->mc_flags |= C_UNTRACK;
7286 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7288 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7291 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7294 if (txn->mt_flags & MDB_TXN_ERROR)
7297 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7301 /* Return the count of duplicate data items for the current key */
7303 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7307 if (mc == NULL || countp == NULL)
7310 if (mc->mc_xcursor == NULL)
7311 return MDB_INCOMPATIBLE;
7313 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7316 if (!(mc->mc_flags & C_INITIALIZED))
7319 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7320 return MDB_NOTFOUND;
7322 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7323 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7326 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7329 *countp = mc->mc_xcursor->mx_db.md_entries;
7335 mdb_cursor_close(MDB_cursor *mc)
7337 if (mc && !mc->mc_backup) {
7338 /* remove from txn, if tracked */
7339 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7340 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7341 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7343 *prev = mc->mc_next;
7350 mdb_cursor_txn(MDB_cursor *mc)
7352 if (!mc) return NULL;
7357 mdb_cursor_dbi(MDB_cursor *mc)
7362 /** Replace the key for a branch node with a new key.
7363 * @param[in] mc Cursor pointing to the node to operate on.
7364 * @param[in] key The new key to use.
7365 * @return 0 on success, non-zero on failure.
7368 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7374 int delta, ksize, oksize;
7375 indx_t ptr, i, numkeys, indx;
7378 indx = mc->mc_ki[mc->mc_top];
7379 mp = mc->mc_pg[mc->mc_top];
7380 node = NODEPTR(mp, indx);
7381 ptr = mp->mp_ptrs[indx];
7385 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7386 k2.mv_data = NODEKEY(node);
7387 k2.mv_size = node->mn_ksize;
7388 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7390 mdb_dkey(&k2, kbuf2),
7396 /* Sizes must be 2-byte aligned. */
7397 ksize = EVEN(key->mv_size);
7398 oksize = EVEN(node->mn_ksize);
7399 delta = ksize - oksize;
7401 /* Shift node contents if EVEN(key length) changed. */
7403 if (delta > 0 && SIZELEFT(mp) < delta) {
7405 /* not enough space left, do a delete and split */
7406 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7407 pgno = NODEPGNO(node);
7408 mdb_node_del(mc, 0);
7409 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7412 numkeys = NUMKEYS(mp);
7413 for (i = 0; i < numkeys; i++) {
7414 if (mp->mp_ptrs[i] <= ptr)
7415 mp->mp_ptrs[i] -= delta;
7418 base = (char *)mp + mp->mp_upper + PAGEBASE;
7419 len = ptr - mp->mp_upper + NODESIZE;
7420 memmove(base - delta, base, len);
7421 mp->mp_upper -= delta;
7423 node = NODEPTR(mp, indx);
7426 /* But even if no shift was needed, update ksize */
7427 if (node->mn_ksize != key->mv_size)
7428 node->mn_ksize = key->mv_size;
7431 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7437 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7439 /** Move a node from csrc to cdst.
7442 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7449 unsigned short flags;
7453 /* Mark src and dst as dirty. */
7454 if ((rc = mdb_page_touch(csrc)) ||
7455 (rc = mdb_page_touch(cdst)))
7458 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7459 key.mv_size = csrc->mc_db->md_pad;
7460 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7462 data.mv_data = NULL;
7466 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7467 mdb_cassert(csrc, !((size_t)srcnode & 1));
7468 srcpg = NODEPGNO(srcnode);
7469 flags = srcnode->mn_flags;
7470 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7471 unsigned int snum = csrc->mc_snum;
7473 /* must find the lowest key below src */
7474 rc = mdb_page_search_lowest(csrc);
7477 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7478 key.mv_size = csrc->mc_db->md_pad;
7479 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7481 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7482 key.mv_size = NODEKSZ(s2);
7483 key.mv_data = NODEKEY(s2);
7485 csrc->mc_snum = snum--;
7486 csrc->mc_top = snum;
7488 key.mv_size = NODEKSZ(srcnode);
7489 key.mv_data = NODEKEY(srcnode);
7491 data.mv_size = NODEDSZ(srcnode);
7492 data.mv_data = NODEDATA(srcnode);
7494 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7495 unsigned int snum = cdst->mc_snum;
7498 /* must find the lowest key below dst */
7499 mdb_cursor_copy(cdst, &mn);
7500 rc = mdb_page_search_lowest(&mn);
7503 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7504 bkey.mv_size = mn.mc_db->md_pad;
7505 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7507 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7508 bkey.mv_size = NODEKSZ(s2);
7509 bkey.mv_data = NODEKEY(s2);
7511 mn.mc_snum = snum--;
7514 rc = mdb_update_key(&mn, &bkey);
7519 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7520 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7521 csrc->mc_ki[csrc->mc_top],
7523 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7524 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7526 /* Add the node to the destination page.
7528 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7529 if (rc != MDB_SUCCESS)
7532 /* Delete the node from the source page.
7534 mdb_node_del(csrc, key.mv_size);
7537 /* Adjust other cursors pointing to mp */
7538 MDB_cursor *m2, *m3;
7539 MDB_dbi dbi = csrc->mc_dbi;
7540 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7542 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7543 if (csrc->mc_flags & C_SUB)
7544 m3 = &m2->mc_xcursor->mx_cursor;
7547 if (m3 == csrc) continue;
7548 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7549 csrc->mc_ki[csrc->mc_top]) {
7550 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7551 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7556 /* Update the parent separators.
7558 if (csrc->mc_ki[csrc->mc_top] == 0) {
7559 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7560 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7561 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7563 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7564 key.mv_size = NODEKSZ(srcnode);
7565 key.mv_data = NODEKEY(srcnode);
7567 DPRINTF(("update separator for source page %"Z"u to [%s]",
7568 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7569 mdb_cursor_copy(csrc, &mn);
7572 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7575 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7577 indx_t ix = csrc->mc_ki[csrc->mc_top];
7578 nullkey.mv_size = 0;
7579 csrc->mc_ki[csrc->mc_top] = 0;
7580 rc = mdb_update_key(csrc, &nullkey);
7581 csrc->mc_ki[csrc->mc_top] = ix;
7582 mdb_cassert(csrc, rc == MDB_SUCCESS);
7586 if (cdst->mc_ki[cdst->mc_top] == 0) {
7587 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7588 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7589 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7591 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7592 key.mv_size = NODEKSZ(srcnode);
7593 key.mv_data = NODEKEY(srcnode);
7595 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7596 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7597 mdb_cursor_copy(cdst, &mn);
7600 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7603 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7605 indx_t ix = cdst->mc_ki[cdst->mc_top];
7606 nullkey.mv_size = 0;
7607 cdst->mc_ki[cdst->mc_top] = 0;
7608 rc = mdb_update_key(cdst, &nullkey);
7609 cdst->mc_ki[cdst->mc_top] = ix;
7610 mdb_cassert(cdst, rc == MDB_SUCCESS);
7617 /** Merge one page into another.
7618 * The nodes from the page pointed to by \b csrc will
7619 * be copied to the page pointed to by \b cdst and then
7620 * the \b csrc page will be freed.
7621 * @param[in] csrc Cursor pointing to the source page.
7622 * @param[in] cdst Cursor pointing to the destination page.
7623 * @return 0 on success, non-zero on failure.
7626 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7628 MDB_page *psrc, *pdst;
7635 psrc = csrc->mc_pg[csrc->mc_top];
7636 pdst = cdst->mc_pg[cdst->mc_top];
7638 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7640 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7641 mdb_cassert(csrc, cdst->mc_snum > 1);
7643 /* Mark dst as dirty. */
7644 if ((rc = mdb_page_touch(cdst)))
7647 /* Move all nodes from src to dst.
7649 j = nkeys = NUMKEYS(pdst);
7650 if (IS_LEAF2(psrc)) {
7651 key.mv_size = csrc->mc_db->md_pad;
7652 key.mv_data = METADATA(psrc);
7653 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7654 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7655 if (rc != MDB_SUCCESS)
7657 key.mv_data = (char *)key.mv_data + key.mv_size;
7660 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7661 srcnode = NODEPTR(psrc, i);
7662 if (i == 0 && IS_BRANCH(psrc)) {
7665 mdb_cursor_copy(csrc, &mn);
7666 /* must find the lowest key below src */
7667 rc = mdb_page_search_lowest(&mn);
7670 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7671 key.mv_size = mn.mc_db->md_pad;
7672 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7674 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7675 key.mv_size = NODEKSZ(s2);
7676 key.mv_data = NODEKEY(s2);
7679 key.mv_size = srcnode->mn_ksize;
7680 key.mv_data = NODEKEY(srcnode);
7683 data.mv_size = NODEDSZ(srcnode);
7684 data.mv_data = NODEDATA(srcnode);
7685 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7686 if (rc != MDB_SUCCESS)
7691 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7692 pdst->mp_pgno, NUMKEYS(pdst),
7693 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7695 /* Unlink the src page from parent and add to free list.
7698 mdb_node_del(csrc, 0);
7699 if (csrc->mc_ki[csrc->mc_top] == 0) {
7701 rc = mdb_update_key(csrc, &key);
7709 psrc = csrc->mc_pg[csrc->mc_top];
7710 /* If not operating on FreeDB, allow this page to be reused
7711 * in this txn. Otherwise just add to free list.
7713 rc = mdb_page_loose(csrc, psrc);
7717 csrc->mc_db->md_leaf_pages--;
7719 csrc->mc_db->md_branch_pages--;
7721 /* Adjust other cursors pointing to mp */
7722 MDB_cursor *m2, *m3;
7723 MDB_dbi dbi = csrc->mc_dbi;
7725 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7726 if (csrc->mc_flags & C_SUB)
7727 m3 = &m2->mc_xcursor->mx_cursor;
7730 if (m3 == csrc) continue;
7731 if (m3->mc_snum < csrc->mc_snum) continue;
7732 if (m3->mc_pg[csrc->mc_top] == psrc) {
7733 m3->mc_pg[csrc->mc_top] = pdst;
7734 m3->mc_ki[csrc->mc_top] += nkeys;
7739 unsigned int snum = cdst->mc_snum;
7740 uint16_t depth = cdst->mc_db->md_depth;
7741 mdb_cursor_pop(cdst);
7742 rc = mdb_rebalance(cdst);
7743 /* Did the tree shrink? */
7744 if (depth > cdst->mc_db->md_depth)
7746 cdst->mc_snum = snum;
7747 cdst->mc_top = snum-1;
7752 /** Copy the contents of a cursor.
7753 * @param[in] csrc The cursor to copy from.
7754 * @param[out] cdst The cursor to copy to.
7757 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7761 cdst->mc_txn = csrc->mc_txn;
7762 cdst->mc_dbi = csrc->mc_dbi;
7763 cdst->mc_db = csrc->mc_db;
7764 cdst->mc_dbx = csrc->mc_dbx;
7765 cdst->mc_snum = csrc->mc_snum;
7766 cdst->mc_top = csrc->mc_top;
7767 cdst->mc_flags = csrc->mc_flags;
7769 for (i=0; i<csrc->mc_snum; i++) {
7770 cdst->mc_pg[i] = csrc->mc_pg[i];
7771 cdst->mc_ki[i] = csrc->mc_ki[i];
7775 /** Rebalance the tree after a delete operation.
7776 * @param[in] mc Cursor pointing to the page where rebalancing
7778 * @return 0 on success, non-zero on failure.
7781 mdb_rebalance(MDB_cursor *mc)
7785 unsigned int ptop, minkeys;
7789 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7790 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7791 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7792 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7793 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7795 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7796 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7797 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7798 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7802 if (mc->mc_snum < 2) {
7803 MDB_page *mp = mc->mc_pg[0];
7805 DPUTS("Can't rebalance a subpage, ignoring");
7808 if (NUMKEYS(mp) == 0) {
7809 DPUTS("tree is completely empty");
7810 mc->mc_db->md_root = P_INVALID;
7811 mc->mc_db->md_depth = 0;
7812 mc->mc_db->md_leaf_pages = 0;
7813 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7816 /* Adjust cursors pointing to mp */
7819 mc->mc_flags &= ~C_INITIALIZED;
7821 MDB_cursor *m2, *m3;
7822 MDB_dbi dbi = mc->mc_dbi;
7824 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7825 if (mc->mc_flags & C_SUB)
7826 m3 = &m2->mc_xcursor->mx_cursor;
7829 if (m3->mc_snum < mc->mc_snum) continue;
7830 if (m3->mc_pg[0] == mp) {
7833 m3->mc_flags &= ~C_INITIALIZED;
7837 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7839 DPUTS("collapsing root page!");
7840 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7843 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7844 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7847 mc->mc_db->md_depth--;
7848 mc->mc_db->md_branch_pages--;
7849 mc->mc_ki[0] = mc->mc_ki[1];
7850 for (i = 1; i<mc->mc_db->md_depth; i++) {
7851 mc->mc_pg[i] = mc->mc_pg[i+1];
7852 mc->mc_ki[i] = mc->mc_ki[i+1];
7855 /* Adjust other cursors pointing to mp */
7856 MDB_cursor *m2, *m3;
7857 MDB_dbi dbi = mc->mc_dbi;
7859 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7860 if (mc->mc_flags & C_SUB)
7861 m3 = &m2->mc_xcursor->mx_cursor;
7864 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7865 if (m3->mc_pg[0] == mp) {
7866 for (i=0; i<m3->mc_snum; i++) {
7867 m3->mc_pg[i] = m3->mc_pg[i+1];
7868 m3->mc_ki[i] = m3->mc_ki[i+1];
7876 DPUTS("root page doesn't need rebalancing");
7880 /* The parent (branch page) must have at least 2 pointers,
7881 * otherwise the tree is invalid.
7883 ptop = mc->mc_top-1;
7884 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7886 /* Leaf page fill factor is below the threshold.
7887 * Try to move keys from left or right neighbor, or
7888 * merge with a neighbor page.
7893 mdb_cursor_copy(mc, &mn);
7894 mn.mc_xcursor = NULL;
7896 oldki = mc->mc_ki[mc->mc_top];
7897 if (mc->mc_ki[ptop] == 0) {
7898 /* We're the leftmost leaf in our parent.
7900 DPUTS("reading right neighbor");
7902 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7903 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7906 mn.mc_ki[mn.mc_top] = 0;
7907 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7909 /* There is at least one neighbor to the left.
7911 DPUTS("reading left neighbor");
7913 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7914 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7917 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7918 mc->mc_ki[mc->mc_top] = 0;
7921 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7922 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7923 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7925 /* If the neighbor page is above threshold and has enough keys,
7926 * move one key from it. Otherwise we should try to merge them.
7927 * (A branch page must never have less than 2 keys.)
7929 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7930 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7931 rc = mdb_node_move(&mn, mc);
7932 if (mc->mc_ki[ptop]) {
7936 if (mc->mc_ki[ptop] == 0) {
7937 rc = mdb_page_merge(&mn, mc);
7940 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7941 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7942 /* We want mdb_rebalance to find mn when doing fixups */
7943 if (mc->mc_flags & C_SUB) {
7944 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7945 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
7946 dummy.mc_xcursor = (MDB_xcursor *)&mn;
7948 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7949 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
7951 rc = mdb_page_merge(mc, &mn);
7952 if (mc->mc_flags & C_SUB)
7953 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
7955 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
7956 mdb_cursor_copy(&mn, mc);
7958 mc->mc_flags &= ~C_EOF;
7960 mc->mc_ki[mc->mc_top] = oldki;
7964 /** Complete a delete operation started by #mdb_cursor_del(). */
7966 mdb_cursor_del0(MDB_cursor *mc)
7973 ki = mc->mc_ki[mc->mc_top];
7974 mdb_node_del(mc, mc->mc_db->md_pad);
7975 mc->mc_db->md_entries--;
7976 rc = mdb_rebalance(mc);
7978 if (rc == MDB_SUCCESS) {
7979 MDB_cursor *m2, *m3;
7980 MDB_dbi dbi = mc->mc_dbi;
7982 /* DB is totally empty now, just bail out.
7983 * Other cursors adjustments were already done
7984 * by mdb_rebalance and aren't needed here.
7989 mp = mc->mc_pg[mc->mc_top];
7990 nkeys = NUMKEYS(mp);
7992 /* if mc points past last node in page, find next sibling */
7993 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7994 rc = mdb_cursor_sibling(mc, 1);
7995 if (rc == MDB_NOTFOUND) {
7996 mc->mc_flags |= C_EOF;
8001 /* Adjust other cursors pointing to mp */
8002 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8003 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8004 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8006 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8008 if (m3->mc_pg[mc->mc_top] == mp) {
8009 if (m3->mc_ki[mc->mc_top] >= ki) {
8010 m3->mc_flags |= C_DEL;
8011 if (m3->mc_ki[mc->mc_top] > ki)
8012 m3->mc_ki[mc->mc_top]--;
8013 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8014 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8016 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8017 rc = mdb_cursor_sibling(m3, 1);
8018 if (rc == MDB_NOTFOUND) {
8019 m3->mc_flags |= C_EOF;
8025 mc->mc_flags |= C_DEL;
8029 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8034 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8035 MDB_val *key, MDB_val *data)
8037 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8040 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
8041 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8043 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8044 /* must ignore any data */
8048 return mdb_del0(txn, dbi, key, data, 0);
8052 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8053 MDB_val *key, MDB_val *data, unsigned flags)
8058 MDB_val rdata, *xdata;
8062 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8064 mdb_cursor_init(&mc, txn, dbi, &mx);
8073 flags |= MDB_NODUPDATA;
8075 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8077 /* let mdb_page_split know about this cursor if needed:
8078 * delete will trigger a rebalance; if it needs to move
8079 * a node from one page to another, it will have to
8080 * update the parent's separator key(s). If the new sepkey
8081 * is larger than the current one, the parent page may
8082 * run out of space, triggering a split. We need this
8083 * cursor to be consistent until the end of the rebalance.
8085 mc.mc_flags |= C_UNTRACK;
8086 mc.mc_next = txn->mt_cursors[dbi];
8087 txn->mt_cursors[dbi] = &mc;
8088 rc = mdb_cursor_del(&mc, flags);
8089 txn->mt_cursors[dbi] = mc.mc_next;
8094 /** Split a page and insert a new node.
8095 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8096 * The cursor will be updated to point to the actual page and index where
8097 * the node got inserted after the split.
8098 * @param[in] newkey The key for the newly inserted node.
8099 * @param[in] newdata The data for the newly inserted node.
8100 * @param[in] newpgno The page number, if the new node is a branch node.
8101 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8102 * @return 0 on success, non-zero on failure.
8105 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8106 unsigned int nflags)
8109 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8112 int i, j, split_indx, nkeys, pmax;
8113 MDB_env *env = mc->mc_txn->mt_env;
8115 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8116 MDB_page *copy = NULL;
8117 MDB_page *mp, *rp, *pp;
8122 mp = mc->mc_pg[mc->mc_top];
8123 newindx = mc->mc_ki[mc->mc_top];
8124 nkeys = NUMKEYS(mp);
8126 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8127 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8128 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8130 /* Create a right sibling. */
8131 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8133 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8135 if (mc->mc_snum < 2) {
8136 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8138 /* shift current top to make room for new parent */
8139 mc->mc_pg[1] = mc->mc_pg[0];
8140 mc->mc_ki[1] = mc->mc_ki[0];
8143 mc->mc_db->md_root = pp->mp_pgno;
8144 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8145 mc->mc_db->md_depth++;
8148 /* Add left (implicit) pointer. */
8149 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8150 /* undo the pre-push */
8151 mc->mc_pg[0] = mc->mc_pg[1];
8152 mc->mc_ki[0] = mc->mc_ki[1];
8153 mc->mc_db->md_root = mp->mp_pgno;
8154 mc->mc_db->md_depth--;
8161 ptop = mc->mc_top-1;
8162 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8165 mc->mc_flags |= C_SPLITTING;
8166 mdb_cursor_copy(mc, &mn);
8167 mn.mc_pg[mn.mc_top] = rp;
8168 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8170 if (nflags & MDB_APPEND) {
8171 mn.mc_ki[mn.mc_top] = 0;
8173 split_indx = newindx;
8177 split_indx = (nkeys+1) / 2;
8182 unsigned int lsize, rsize, ksize;
8183 /* Move half of the keys to the right sibling */
8184 x = mc->mc_ki[mc->mc_top] - split_indx;
8185 ksize = mc->mc_db->md_pad;
8186 split = LEAF2KEY(mp, split_indx, ksize);
8187 rsize = (nkeys - split_indx) * ksize;
8188 lsize = (nkeys - split_indx) * sizeof(indx_t);
8189 mp->mp_lower -= lsize;
8190 rp->mp_lower += lsize;
8191 mp->mp_upper += rsize - lsize;
8192 rp->mp_upper -= rsize - lsize;
8193 sepkey.mv_size = ksize;
8194 if (newindx == split_indx) {
8195 sepkey.mv_data = newkey->mv_data;
8197 sepkey.mv_data = split;
8200 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8201 memcpy(rp->mp_ptrs, split, rsize);
8202 sepkey.mv_data = rp->mp_ptrs;
8203 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8204 memcpy(ins, newkey->mv_data, ksize);
8205 mp->mp_lower += sizeof(indx_t);
8206 mp->mp_upper -= ksize - sizeof(indx_t);
8209 memcpy(rp->mp_ptrs, split, x * ksize);
8210 ins = LEAF2KEY(rp, x, ksize);
8211 memcpy(ins, newkey->mv_data, ksize);
8212 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8213 rp->mp_lower += sizeof(indx_t);
8214 rp->mp_upper -= ksize - sizeof(indx_t);
8215 mc->mc_ki[mc->mc_top] = x;
8216 mc->mc_pg[mc->mc_top] = rp;
8219 int psize, nsize, k;
8220 /* Maximum free space in an empty page */
8221 pmax = env->me_psize - PAGEHDRSZ;
8223 nsize = mdb_leaf_size(env, newkey, newdata);
8225 nsize = mdb_branch_size(env, newkey);
8226 nsize = EVEN(nsize);
8228 /* grab a page to hold a temporary copy */
8229 copy = mdb_page_malloc(mc->mc_txn, 1);
8234 copy->mp_pgno = mp->mp_pgno;
8235 copy->mp_flags = mp->mp_flags;
8236 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8237 copy->mp_upper = env->me_psize - PAGEBASE;
8239 /* prepare to insert */
8240 for (i=0, j=0; i<nkeys; i++) {
8242 copy->mp_ptrs[j++] = 0;
8244 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8247 /* When items are relatively large the split point needs
8248 * to be checked, because being off-by-one will make the
8249 * difference between success or failure in mdb_node_add.
8251 * It's also relevant if a page happens to be laid out
8252 * such that one half of its nodes are all "small" and
8253 * the other half of its nodes are "large." If the new
8254 * item is also "large" and falls on the half with
8255 * "large" nodes, it also may not fit.
8257 * As a final tweak, if the new item goes on the last
8258 * spot on the page (and thus, onto the new page), bias
8259 * the split so the new page is emptier than the old page.
8260 * This yields better packing during sequential inserts.
8262 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8263 /* Find split point */
8265 if (newindx <= split_indx || newindx >= nkeys) {
8267 k = newindx >= nkeys ? nkeys : split_indx+2;
8272 for (; i!=k; i+=j) {
8277 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8278 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8280 if (F_ISSET(node->mn_flags, F_BIGDATA))
8281 psize += sizeof(pgno_t);
8283 psize += NODEDSZ(node);
8285 psize = EVEN(psize);
8287 if (psize > pmax || i == k-j) {
8288 split_indx = i + (j<0);
8293 if (split_indx == newindx) {
8294 sepkey.mv_size = newkey->mv_size;
8295 sepkey.mv_data = newkey->mv_data;
8297 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8298 sepkey.mv_size = node->mn_ksize;
8299 sepkey.mv_data = NODEKEY(node);
8304 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8306 /* Copy separator key to the parent.
8308 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8312 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8317 if (mn.mc_snum == mc->mc_snum) {
8318 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8319 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8320 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8321 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8326 /* Right page might now have changed parent.
8327 * Check if left page also changed parent.
8329 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8330 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8331 for (i=0; i<ptop; i++) {
8332 mc->mc_pg[i] = mn.mc_pg[i];
8333 mc->mc_ki[i] = mn.mc_ki[i];
8335 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8336 if (mn.mc_ki[ptop]) {
8337 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8339 /* find right page's left sibling */
8340 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8341 mdb_cursor_sibling(mc, 0);
8346 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8349 mc->mc_flags ^= C_SPLITTING;
8350 if (rc != MDB_SUCCESS) {
8353 if (nflags & MDB_APPEND) {
8354 mc->mc_pg[mc->mc_top] = rp;
8355 mc->mc_ki[mc->mc_top] = 0;
8356 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8359 for (i=0; i<mc->mc_top; i++)
8360 mc->mc_ki[i] = mn.mc_ki[i];
8361 } else if (!IS_LEAF2(mp)) {
8363 mc->mc_pg[mc->mc_top] = rp;
8368 rkey.mv_data = newkey->mv_data;
8369 rkey.mv_size = newkey->mv_size;
8375 /* Update index for the new key. */
8376 mc->mc_ki[mc->mc_top] = j;
8378 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8379 rkey.mv_data = NODEKEY(node);
8380 rkey.mv_size = node->mn_ksize;
8382 xdata.mv_data = NODEDATA(node);
8383 xdata.mv_size = NODEDSZ(node);
8386 pgno = NODEPGNO(node);
8387 flags = node->mn_flags;
8390 if (!IS_LEAF(mp) && j == 0) {
8391 /* First branch index doesn't need key data. */
8395 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8401 mc->mc_pg[mc->mc_top] = copy;
8406 } while (i != split_indx);
8408 nkeys = NUMKEYS(copy);
8409 for (i=0; i<nkeys; i++)
8410 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8411 mp->mp_lower = copy->mp_lower;
8412 mp->mp_upper = copy->mp_upper;
8413 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8414 env->me_psize - copy->mp_upper - PAGEBASE);
8416 /* reset back to original page */
8417 if (newindx < split_indx) {
8418 mc->mc_pg[mc->mc_top] = mp;
8419 if (nflags & MDB_RESERVE) {
8420 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8421 if (!(node->mn_flags & F_BIGDATA))
8422 newdata->mv_data = NODEDATA(node);
8425 mc->mc_pg[mc->mc_top] = rp;
8427 /* Make sure mc_ki is still valid.
8429 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8430 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8431 for (i=0; i<=ptop; i++) {
8432 mc->mc_pg[i] = mn.mc_pg[i];
8433 mc->mc_ki[i] = mn.mc_ki[i];
8440 /* Adjust other cursors pointing to mp */
8441 MDB_cursor *m2, *m3;
8442 MDB_dbi dbi = mc->mc_dbi;
8443 int fixup = NUMKEYS(mp);
8445 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8446 if (mc->mc_flags & C_SUB)
8447 m3 = &m2->mc_xcursor->mx_cursor;
8452 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8454 if (m3->mc_flags & C_SPLITTING)
8459 for (k=m3->mc_top; k>=0; k--) {
8460 m3->mc_ki[k+1] = m3->mc_ki[k];
8461 m3->mc_pg[k+1] = m3->mc_pg[k];
8463 if (m3->mc_ki[0] >= split_indx) {
8468 m3->mc_pg[0] = mc->mc_pg[0];
8472 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8473 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8474 m3->mc_ki[mc->mc_top]++;
8475 if (m3->mc_ki[mc->mc_top] >= fixup) {
8476 m3->mc_pg[mc->mc_top] = rp;
8477 m3->mc_ki[mc->mc_top] -= fixup;
8478 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8480 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8481 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8486 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8489 if (copy) /* tmp page */
8490 mdb_page_free(env, copy);
8492 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8497 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8498 MDB_val *key, MDB_val *data, unsigned int flags)
8503 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8506 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8509 mdb_cursor_init(&mc, txn, dbi, &mx);
8510 return mdb_cursor_put(&mc, key, data, flags);
8514 #define MDB_WBUF (1024*1024)
8517 /** State needed for a compacting copy. */
8518 typedef struct mdb_copy {
8519 pthread_mutex_t mc_mutex;
8520 pthread_cond_t mc_cond;
8527 pgno_t mc_next_pgno;
8530 volatile int mc_new;
8535 /** Dedicated writer thread for compacting copy. */
8536 static THREAD_RET ESECT
8537 mdb_env_copythr(void *arg)
8541 int toggle = 0, wsize, rc;
8544 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8547 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8550 pthread_mutex_lock(&my->mc_mutex);
8552 pthread_cond_signal(&my->mc_cond);
8555 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8556 if (my->mc_new < 0) {
8561 wsize = my->mc_wlen[toggle];
8562 ptr = my->mc_wbuf[toggle];
8565 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8569 } else if (len > 0) {
8583 /* If there's an overflow page tail, write it too */
8584 if (my->mc_olen[toggle]) {
8585 wsize = my->mc_olen[toggle];
8586 ptr = my->mc_over[toggle];
8587 my->mc_olen[toggle] = 0;
8590 my->mc_wlen[toggle] = 0;
8592 pthread_cond_signal(&my->mc_cond);
8594 pthread_cond_signal(&my->mc_cond);
8595 pthread_mutex_unlock(&my->mc_mutex);
8596 return (THREAD_RET)0;
8600 /** Tell the writer thread there's a buffer ready to write */
8602 mdb_env_cthr_toggle(mdb_copy *my, int st)
8604 int toggle = my->mc_toggle ^ 1;
8605 pthread_mutex_lock(&my->mc_mutex);
8606 if (my->mc_status) {
8607 pthread_mutex_unlock(&my->mc_mutex);
8608 return my->mc_status;
8610 while (my->mc_new == 1)
8611 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8613 my->mc_toggle = toggle;
8614 pthread_cond_signal(&my->mc_cond);
8615 pthread_mutex_unlock(&my->mc_mutex);
8619 /** Depth-first tree traversal for compacting copy. */
8621 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8624 MDB_txn *txn = my->mc_txn;
8626 MDB_page *mo, *mp, *leaf;
8631 /* Empty DB, nothing to do */
8632 if (*pg == P_INVALID)
8639 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8642 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8646 /* Make cursor pages writable */
8647 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8651 for (i=0; i<mc.mc_top; i++) {
8652 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8653 mc.mc_pg[i] = (MDB_page *)ptr;
8654 ptr += my->mc_env->me_psize;
8657 /* This is writable space for a leaf page. Usually not needed. */
8658 leaf = (MDB_page *)ptr;
8660 toggle = my->mc_toggle;
8661 while (mc.mc_snum > 0) {
8663 mp = mc.mc_pg[mc.mc_top];
8667 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8668 for (i=0; i<n; i++) {
8669 ni = NODEPTR(mp, i);
8670 if (ni->mn_flags & F_BIGDATA) {
8674 /* Need writable leaf */
8676 mc.mc_pg[mc.mc_top] = leaf;
8677 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8679 ni = NODEPTR(mp, i);
8682 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8683 rc = mdb_page_get(txn, pg, &omp, NULL);
8686 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8687 rc = mdb_env_cthr_toggle(my, 1);
8690 toggle = my->mc_toggle;
8692 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8693 memcpy(mo, omp, my->mc_env->me_psize);
8694 mo->mp_pgno = my->mc_next_pgno;
8695 my->mc_next_pgno += omp->mp_pages;
8696 my->mc_wlen[toggle] += my->mc_env->me_psize;
8697 if (omp->mp_pages > 1) {
8698 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8699 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8700 rc = mdb_env_cthr_toggle(my, 1);
8703 toggle = my->mc_toggle;
8705 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8706 } else if (ni->mn_flags & F_SUBDATA) {
8709 /* Need writable leaf */
8711 mc.mc_pg[mc.mc_top] = leaf;
8712 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8714 ni = NODEPTR(mp, i);
8717 memcpy(&db, NODEDATA(ni), sizeof(db));
8718 my->mc_toggle = toggle;
8719 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8722 toggle = my->mc_toggle;
8723 memcpy(NODEDATA(ni), &db, sizeof(db));
8728 mc.mc_ki[mc.mc_top]++;
8729 if (mc.mc_ki[mc.mc_top] < n) {
8732 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8734 rc = mdb_page_get(txn, pg, &mp, NULL);
8739 mc.mc_ki[mc.mc_top] = 0;
8740 if (IS_BRANCH(mp)) {
8741 /* Whenever we advance to a sibling branch page,
8742 * we must proceed all the way down to its first leaf.
8744 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8747 mc.mc_pg[mc.mc_top] = mp;
8751 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8752 rc = mdb_env_cthr_toggle(my, 1);
8755 toggle = my->mc_toggle;
8757 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8758 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8759 mo->mp_pgno = my->mc_next_pgno++;
8760 my->mc_wlen[toggle] += my->mc_env->me_psize;
8762 /* Update parent if there is one */
8763 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8764 SETPGNO(ni, mo->mp_pgno);
8765 mdb_cursor_pop(&mc);
8767 /* Otherwise we're done */
8777 /** Copy environment with compaction. */
8779 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8784 MDB_txn *txn = NULL;
8789 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8790 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8791 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8792 if (my.mc_wbuf[0] == NULL)
8795 pthread_mutex_init(&my.mc_mutex, NULL);
8796 pthread_cond_init(&my.mc_cond, NULL);
8797 #ifdef HAVE_MEMALIGN
8798 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8799 if (my.mc_wbuf[0] == NULL)
8802 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8807 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8808 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8813 my.mc_next_pgno = 2;
8819 THREAD_CREATE(thr, mdb_env_copythr, &my);
8821 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8825 mp = (MDB_page *)my.mc_wbuf[0];
8826 memset(mp, 0, 2*env->me_psize);
8828 mp->mp_flags = P_META;
8829 mm = (MDB_meta *)METADATA(mp);
8830 mdb_env_init_meta0(env, mm);
8831 mm->mm_address = env->me_metas[0]->mm_address;
8833 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8835 mp->mp_flags = P_META;
8836 *(MDB_meta *)METADATA(mp) = *mm;
8837 mm = (MDB_meta *)METADATA(mp);
8839 /* Count the number of free pages, subtract from lastpg to find
8840 * number of active pages
8843 MDB_ID freecount = 0;
8846 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8847 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8848 freecount += *(MDB_ID *)data.mv_data;
8849 freecount += txn->mt_dbs[0].md_branch_pages +
8850 txn->mt_dbs[0].md_leaf_pages +
8851 txn->mt_dbs[0].md_overflow_pages;
8853 /* Set metapage 1 */
8854 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8855 mm->mm_dbs[1] = txn->mt_dbs[1];
8856 if (mm->mm_last_pg > 1) {
8857 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8860 mm->mm_dbs[1].md_root = P_INVALID;
8863 my.mc_wlen[0] = env->me_psize * 2;
8865 pthread_mutex_lock(&my.mc_mutex);
8867 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8868 pthread_mutex_unlock(&my.mc_mutex);
8869 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8870 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8871 rc = mdb_env_cthr_toggle(&my, 1);
8872 mdb_env_cthr_toggle(&my, -1);
8873 pthread_mutex_lock(&my.mc_mutex);
8875 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8876 pthread_mutex_unlock(&my.mc_mutex);
8881 CloseHandle(my.mc_cond);
8882 CloseHandle(my.mc_mutex);
8883 _aligned_free(my.mc_wbuf[0]);
8885 pthread_cond_destroy(&my.mc_cond);
8886 pthread_mutex_destroy(&my.mc_mutex);
8887 free(my.mc_wbuf[0]);
8892 /** Copy environment as-is. */
8894 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8896 MDB_txn *txn = NULL;
8897 mdb_mutex_t *wmutex = NULL;
8903 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8907 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8910 /* Do the lock/unlock of the reader mutex before starting the
8911 * write txn. Otherwise other read txns could block writers.
8913 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8918 /* We must start the actual read txn after blocking writers */
8919 mdb_txn_reset0(txn, "reset-stage1");
8921 /* Temporarily block writers until we snapshot the meta pages */
8922 wmutex = MDB_MUTEX(env, w);
8923 if (LOCK_MUTEX(rc, env, wmutex))
8926 rc = mdb_txn_renew0(txn);
8928 UNLOCK_MUTEX(wmutex);
8933 wsize = env->me_psize * 2;
8937 DO_WRITE(rc, fd, ptr, w2, len);
8941 } else if (len > 0) {
8947 /* Non-blocking or async handles are not supported */
8953 UNLOCK_MUTEX(wmutex);
8958 w2 = txn->mt_next_pgno * env->me_psize;
8961 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8968 if (wsize > MAX_WRITE)
8972 DO_WRITE(rc, fd, ptr, w2, len);
8976 } else if (len > 0) {
8993 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8995 if (flags & MDB_CP_COMPACT)
8996 return mdb_env_copyfd1(env, fd);
8998 return mdb_env_copyfd0(env, fd);
9002 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9004 return mdb_env_copyfd2(env, fd, 0);
9008 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9012 HANDLE newfd = INVALID_HANDLE_VALUE;
9014 if (env->me_flags & MDB_NOSUBDIR) {
9015 lpath = (char *)path;
9018 len += sizeof(DATANAME);
9019 lpath = malloc(len);
9022 sprintf(lpath, "%s" DATANAME, path);
9025 /* The destination path must exist, but the destination file must not.
9026 * We don't want the OS to cache the writes, since the source data is
9027 * already in the OS cache.
9030 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9031 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9033 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9035 if (newfd == INVALID_HANDLE_VALUE) {
9040 if (env->me_psize >= env->me_os_psize) {
9042 /* Set O_DIRECT if the file system supports it */
9043 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9044 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9046 #ifdef F_NOCACHE /* __APPLE__ */
9047 rc = fcntl(newfd, F_NOCACHE, 1);
9055 rc = mdb_env_copyfd2(env, newfd, flags);
9058 if (!(env->me_flags & MDB_NOSUBDIR))
9060 if (newfd != INVALID_HANDLE_VALUE)
9061 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9068 mdb_env_copy(MDB_env *env, const char *path)
9070 return mdb_env_copy2(env, path, 0);
9074 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9076 if ((flag & CHANGEABLE) != flag)
9079 env->me_flags |= flag;
9081 env->me_flags &= ~flag;
9086 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9091 *arg = env->me_flags;
9096 mdb_env_set_userctx(MDB_env *env, void *ctx)
9100 env->me_userctx = ctx;
9105 mdb_env_get_userctx(MDB_env *env)
9107 return env ? env->me_userctx : NULL;
9111 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9116 env->me_assert_func = func;
9122 mdb_env_get_path(MDB_env *env, const char **arg)
9127 *arg = env->me_path;
9132 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9141 /** Common code for #mdb_stat() and #mdb_env_stat().
9142 * @param[in] env the environment to operate in.
9143 * @param[in] db the #MDB_db record containing the stats to return.
9144 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9145 * @return 0, this function always succeeds.
9148 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9150 arg->ms_psize = env->me_psize;
9151 arg->ms_depth = db->md_depth;
9152 arg->ms_branch_pages = db->md_branch_pages;
9153 arg->ms_leaf_pages = db->md_leaf_pages;
9154 arg->ms_overflow_pages = db->md_overflow_pages;
9155 arg->ms_entries = db->md_entries;
9161 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9165 if (env == NULL || arg == NULL)
9168 toggle = mdb_env_pick_meta(env);
9170 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9174 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9178 if (env == NULL || arg == NULL)
9181 toggle = mdb_env_pick_meta(env);
9182 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9183 arg->me_mapsize = env->me_mapsize;
9184 arg->me_maxreaders = env->me_maxreaders;
9185 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9187 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9188 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9192 /** Set the default comparison functions for a database.
9193 * Called immediately after a database is opened to set the defaults.
9194 * The user can then override them with #mdb_set_compare() or
9195 * #mdb_set_dupsort().
9196 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9197 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9200 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9202 uint16_t f = txn->mt_dbs[dbi].md_flags;
9204 txn->mt_dbxs[dbi].md_cmp =
9205 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9206 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9208 txn->mt_dbxs[dbi].md_dcmp =
9209 !(f & MDB_DUPSORT) ? 0 :
9210 ((f & MDB_INTEGERDUP)
9211 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9212 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9215 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9221 int rc, dbflag, exact;
9222 unsigned int unused = 0, seq;
9225 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9226 mdb_default_cmp(txn, FREE_DBI);
9229 if ((flags & VALID_FLAGS) != flags)
9231 if (txn->mt_flags & MDB_TXN_ERROR)
9237 if (flags & PERSISTENT_FLAGS) {
9238 uint16_t f2 = flags & PERSISTENT_FLAGS;
9239 /* make sure flag changes get committed */
9240 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9241 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9242 txn->mt_flags |= MDB_TXN_DIRTY;
9245 mdb_default_cmp(txn, MAIN_DBI);
9249 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9250 mdb_default_cmp(txn, MAIN_DBI);
9253 /* Is the DB already open? */
9255 for (i=2; i<txn->mt_numdbs; i++) {
9256 if (!txn->mt_dbxs[i].md_name.mv_size) {
9257 /* Remember this free slot */
9258 if (!unused) unused = i;
9261 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9262 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9268 /* If no free slot and max hit, fail */
9269 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9270 return MDB_DBS_FULL;
9272 /* Cannot mix named databases with some mainDB flags */
9273 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9274 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9276 /* Find the DB info */
9277 dbflag = DB_NEW|DB_VALID;
9280 key.mv_data = (void *)name;
9281 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9282 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9283 if (rc == MDB_SUCCESS) {
9284 /* make sure this is actually a DB */
9285 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9286 if (!(node->mn_flags & F_SUBDATA))
9287 return MDB_INCOMPATIBLE;
9288 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9289 /* Create if requested */
9290 data.mv_size = sizeof(MDB_db);
9291 data.mv_data = &dummy;
9292 memset(&dummy, 0, sizeof(dummy));
9293 dummy.md_root = P_INVALID;
9294 dummy.md_flags = flags & PERSISTENT_FLAGS;
9295 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9299 /* OK, got info, add to table */
9300 if (rc == MDB_SUCCESS) {
9301 unsigned int slot = unused ? unused : txn->mt_numdbs;
9302 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9303 txn->mt_dbxs[slot].md_name.mv_size = len;
9304 txn->mt_dbxs[slot].md_rel = NULL;
9305 txn->mt_dbflags[slot] = dbflag;
9306 /* txn-> and env-> are the same in read txns, use
9307 * tmp variable to avoid undefined assignment
9309 seq = ++txn->mt_env->me_dbiseqs[slot];
9310 txn->mt_dbiseqs[slot] = seq;
9312 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9314 mdb_default_cmp(txn, slot);
9323 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9325 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9328 if (txn->mt_flags & MDB_TXN_ERROR)
9331 if (txn->mt_dbflags[dbi] & DB_STALE) {
9334 /* Stale, must read the DB's root. cursor_init does it for us. */
9335 mdb_cursor_init(&mc, txn, dbi, &mx);
9337 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9340 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9343 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9345 ptr = env->me_dbxs[dbi].md_name.mv_data;
9346 /* If there was no name, this was already closed */
9348 env->me_dbxs[dbi].md_name.mv_data = NULL;
9349 env->me_dbxs[dbi].md_name.mv_size = 0;
9350 env->me_dbflags[dbi] = 0;
9351 env->me_dbiseqs[dbi]++;
9356 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9358 /* We could return the flags for the FREE_DBI too but what's the point? */
9359 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9361 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9365 /** Add all the DB's pages to the free list.
9366 * @param[in] mc Cursor on the DB to free.
9367 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9368 * @return 0 on success, non-zero on failure.
9371 mdb_drop0(MDB_cursor *mc, int subs)
9375 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9376 if (rc == MDB_SUCCESS) {
9377 MDB_txn *txn = mc->mc_txn;
9382 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9383 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9386 mdb_cursor_copy(mc, &mx);
9387 while (mc->mc_snum > 0) {
9388 MDB_page *mp = mc->mc_pg[mc->mc_top];
9389 unsigned n = NUMKEYS(mp);
9391 for (i=0; i<n; i++) {
9392 ni = NODEPTR(mp, i);
9393 if (ni->mn_flags & F_BIGDATA) {
9396 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9397 rc = mdb_page_get(txn, pg, &omp, NULL);
9400 mdb_cassert(mc, IS_OVERFLOW(omp));
9401 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9405 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9406 mdb_xcursor_init1(mc, ni);
9407 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9413 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9415 for (i=0; i<n; i++) {
9417 ni = NODEPTR(mp, i);
9420 mdb_midl_xappend(txn->mt_free_pgs, pg);
9425 mc->mc_ki[mc->mc_top] = i;
9426 rc = mdb_cursor_sibling(mc, 1);
9428 if (rc != MDB_NOTFOUND)
9430 /* no more siblings, go back to beginning
9431 * of previous level.
9435 for (i=1; i<mc->mc_snum; i++) {
9437 mc->mc_pg[i] = mx.mc_pg[i];
9442 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9445 txn->mt_flags |= MDB_TXN_ERROR;
9446 } else if (rc == MDB_NOTFOUND) {
9452 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9454 MDB_cursor *mc, *m2;
9457 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9460 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9463 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9466 rc = mdb_cursor_open(txn, dbi, &mc);
9470 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9471 /* Invalidate the dropped DB's cursors */
9472 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9473 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9477 /* Can't delete the main DB */
9478 if (del && dbi > MAIN_DBI) {
9479 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9481 txn->mt_dbflags[dbi] = DB_STALE;
9482 mdb_dbi_close(txn->mt_env, dbi);
9484 txn->mt_flags |= MDB_TXN_ERROR;
9487 /* reset the DB record, mark it dirty */
9488 txn->mt_dbflags[dbi] |= DB_DIRTY;
9489 txn->mt_dbs[dbi].md_depth = 0;
9490 txn->mt_dbs[dbi].md_branch_pages = 0;
9491 txn->mt_dbs[dbi].md_leaf_pages = 0;
9492 txn->mt_dbs[dbi].md_overflow_pages = 0;
9493 txn->mt_dbs[dbi].md_entries = 0;
9494 txn->mt_dbs[dbi].md_root = P_INVALID;
9496 txn->mt_flags |= MDB_TXN_DIRTY;
9499 mdb_cursor_close(mc);
9503 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9505 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9508 txn->mt_dbxs[dbi].md_cmp = cmp;
9512 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9514 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9517 txn->mt_dbxs[dbi].md_dcmp = cmp;
9521 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9523 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9526 txn->mt_dbxs[dbi].md_rel = rel;
9530 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9532 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9535 txn->mt_dbxs[dbi].md_relctx = ctx;
9540 mdb_env_get_maxkeysize(MDB_env *env)
9542 return ENV_MAXKEY(env);
9546 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9548 unsigned int i, rdrs;
9551 int rc = 0, first = 1;
9555 if (!env->me_txns) {
9556 return func("(no reader locks)\n", ctx);
9558 rdrs = env->me_txns->mti_numreaders;
9559 mr = env->me_txns->mti_readers;
9560 for (i=0; i<rdrs; i++) {
9562 txnid_t txnid = mr[i].mr_txnid;
9563 sprintf(buf, txnid == (txnid_t)-1 ?
9564 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9565 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9568 rc = func(" pid thread txnid\n", ctx);
9572 rc = func(buf, ctx);
9578 rc = func("(no active readers)\n", ctx);
9583 /** Insert pid into list if not already present.
9584 * return -1 if already present.
9587 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9589 /* binary search of pid in list */
9591 unsigned cursor = 1;
9593 unsigned n = ids[0];
9596 unsigned pivot = n >> 1;
9597 cursor = base + pivot + 1;
9598 val = pid - ids[cursor];
9603 } else if ( val > 0 ) {
9608 /* found, so it's a duplicate */
9617 for (n = ids[0]; n > cursor; n--)
9624 mdb_reader_check(MDB_env *env, int *dead)
9630 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9633 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9634 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9636 mdb_mutex_t *rmutex = rlocked ? NULL : MDB_MUTEX(env, r);
9637 unsigned int i, j, rdrs;
9639 MDB_PID_T *pids, pid;
9640 int rc = MDB_SUCCESS, count = 0;
9642 rdrs = env->me_txns->mti_numreaders;
9643 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9647 mr = env->me_txns->mti_readers;
9648 for (i=0; i<rdrs; i++) {
9650 if (pid && pid != env->me_pid) {
9651 if (mdb_pid_insert(pids, pid) == 0) {
9652 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9653 /* Stale reader found */
9656 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9657 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9659 rdrs = 0; /* the above checked all readers */
9661 /* Recheck, a new process may have reused pid */
9662 if (mdb_reader_pid(env, Pidcheck, pid))
9667 if (mr[j].mr_pid == pid) {
9668 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9669 (unsigned) pid, mr[j].mr_txnid));
9674 UNLOCK_MUTEX(rmutex);
9685 #ifdef MDB_ROBUST_SUPPORTED
9686 /** Handle #LOCK_MUTEX0() failure.
9687 * Try to repair the lock file if the mutex owner died.
9688 * @param[in] env the environment handle
9689 * @param[in] mutex LOCK_MUTEX0() mutex
9690 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9691 * @return 0 on success with the mutex locked, or an error code on failure.
9693 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc)
9695 int toggle, rlocked, rc2;
9697 if (rc == MDB_OWNERDEAD) {
9698 /* We own the mutex. Clean up after dead previous owner. */
9700 rlocked = (mutex == MDB_MUTEX(env, r));
9702 /* Keep mti_txnid updated, otherwise next writer can
9703 * overwrite data which latest meta page refers to.
9705 toggle = mdb_env_pick_meta(env);
9706 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9707 /* env is hosed if the dead thread was ours */
9709 env->me_flags |= MDB_FATAL_ERROR;
9714 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9715 (rc ? "this process' env is hosed" : "recovering")));
9716 rc2 = mdb_reader_check0(env, rlocked, NULL);
9718 rc2 = mdb_mutex_consistent(mutex);
9719 if (rc || (rc = rc2)) {
9720 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9721 UNLOCK_MUTEX(mutex);
9727 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9732 #endif /* MDB_ROBUST_SUPPORTED */