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 2011-2014 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)
93 #if defined(__sun) || defined(ANDROID)
94 /* Most platforms have posix_memalign, older may only have memalign */
95 #define HAVE_MEMALIGN 1
99 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
100 #include <netinet/in.h>
101 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
104 #if defined(__APPLE__) || defined (BSD)
105 # define MDB_USE_SYSV_SEM 1
106 # define MDB_FDATASYNC fsync
107 #elif defined(ANDROID)
108 # define MDB_FDATASYNC fsync
113 #ifdef MDB_USE_SYSV_SEM
116 #ifdef _SEM_SEMUN_UNDEFINED
119 struct semid_ds *buf;
120 unsigned short *array;
122 #endif /* _SEM_SEMUN_UNDEFINED */
123 #endif /* MDB_USE_SYSV_SEM */
127 #include <valgrind/memcheck.h>
128 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
129 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
130 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
131 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
132 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
134 #define VGMEMP_CREATE(h,r,z)
135 #define VGMEMP_ALLOC(h,a,s)
136 #define VGMEMP_FREE(h,a)
137 #define VGMEMP_DESTROY(h)
138 #define VGMEMP_DEFINED(a,s)
142 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
143 /* Solaris just defines one or the other */
144 # define LITTLE_ENDIAN 1234
145 # define BIG_ENDIAN 4321
146 # ifdef _LITTLE_ENDIAN
147 # define BYTE_ORDER LITTLE_ENDIAN
149 # define BYTE_ORDER BIG_ENDIAN
152 # define BYTE_ORDER __BYTE_ORDER
156 #ifndef LITTLE_ENDIAN
157 #define LITTLE_ENDIAN __LITTLE_ENDIAN
160 #define BIG_ENDIAN __BIG_ENDIAN
163 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
164 #define MISALIGNED_OK 1
170 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
171 # error "Unknown or unsupported endianness (BYTE_ORDER)"
172 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
173 # error "Two's complement, reasonably sized integer types, please"
177 /** Put infrequently used env functions in separate section */
179 # define ESECT __attribute__ ((section("__TEXT,text_env")))
181 # define ESECT __attribute__ ((section("text_env")))
187 /** @defgroup internal LMDB Internals
190 /** @defgroup compat Compatibility Macros
191 * A bunch of macros to minimize the amount of platform-specific ifdefs
192 * needed throughout the rest of the code. When the features this library
193 * needs are similar enough to POSIX to be hidden in a one-or-two line
194 * replacement, this macro approach is used.
198 /** Features under development */
203 #if defined(WIN32) || (defined(EOWNERDEAD) && !defined(MDB_USE_SYSV_SEM))
204 #define MDB_ROBUST_SUPPORTED 1
207 /** Wrapper around __func__, which is a C99 feature */
208 #if __STDC_VERSION__ >= 199901L
209 # define mdb_func_ __func__
210 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
211 # define mdb_func_ __FUNCTION__
213 /* If a debug message says <mdb_unknown>(), update the #if statements above */
214 # define mdb_func_ "<mdb_unknown>"
218 #define MDB_USE_HASH 1
219 #define MDB_PIDLOCK 0
220 #define THREAD_RET DWORD
221 #define pthread_t HANDLE
222 #define pthread_mutex_t HANDLE
223 #define pthread_cond_t HANDLE
224 typedef HANDLE mdb_mutex_t;
225 #define pthread_key_t DWORD
226 #define pthread_self() GetCurrentThreadId()
227 #define pthread_key_create(x,y) \
228 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
229 #define pthread_key_delete(x) TlsFree(x)
230 #define pthread_getspecific(x) TlsGetValue(x)
231 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
232 #define pthread_mutex_consistent(mutex) 0
233 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
234 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
235 #define pthread_cond_signal(x) SetEvent(*x)
236 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
237 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
238 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
239 #define MDB_MUTEX(env, rw) ((env)->me_##rw##mutex)
240 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
241 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
242 #define getpid() GetCurrentProcessId()
243 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
244 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
245 #define ErrCode() GetLastError()
246 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
247 #define close(fd) (CloseHandle(fd) ? 0 : -1)
248 #define munmap(ptr,len) UnmapViewOfFile(ptr)
249 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
250 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
252 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
256 #define THREAD_RET void *
257 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
258 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
259 #define Z "z" /**< printf format modifier for size_t */
261 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
262 #define MDB_PIDLOCK 1
264 #ifdef MDB_USE_SYSV_SEM
266 typedef struct mdb_mutex {
271 #define MDB_MUTEX(env, rw) (&(env)->me_##rw##mutex)
272 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
273 #define UNLOCK_MUTEX(mutex) do { struct sembuf sb = { mutex->semnum, 1, SEM_UNDO }; semop(mutex->semid, &sb, 1); } while(0)
276 mdb_sem_wait(mdb_mutex_t *sem)
279 struct sembuf sb = { sem->semnum, -1, SEM_UNDO };
280 while ((rc = semop(sem->semid, &sb, 1)) && (rc = errno) == EINTR) ;
285 /** Pointer/HANDLE type of shared mutex/semaphore.
287 typedef pthread_mutex_t mdb_mutex_t;
288 /** Mutex for the reader table (rw = r) or write transaction (rw = w).
290 #define MDB_MUTEX(env, rw) (&(env)->me_txns->mti_##rw##mutex)
291 /** Lock the reader or writer mutex.
292 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
294 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
295 /** Unlock the reader or writer mutex.
297 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
298 #endif /* MDB_USE_SYSV_SEM */
300 /** Get the error code for the last failed system function.
302 #define ErrCode() errno
304 /** An abstraction for a file handle.
305 * On POSIX systems file handles are small integers. On Windows
306 * they're opaque pointers.
310 /** A value for an invalid file handle.
311 * Mainly used to initialize file variables and signify that they are
314 #define INVALID_HANDLE_VALUE (-1)
316 /** Get the size of a memory page for the system.
317 * This is the basic size that the platform's memory manager uses, and is
318 * fundamental to the use of memory-mapped files.
320 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
325 #elif defined(MDB_USE_SYSV_SEM)
328 #define MNAME_LEN (sizeof(pthread_mutex_t))
333 #ifdef MDB_ROBUST_SUPPORTED
334 /** Lock mutex, handle any error, set rc = result.
335 * Return 0 on success, nonzero (not rc) on error.
337 #define LOCK_MUTEX(rc, env, mutex) \
338 (((rc) = LOCK_MUTEX0(mutex)) && \
339 ((rc) = mdb_mutex_failed(env, mutex, rc)))
340 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc);
342 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
343 #define mdb_mutex_failed(env, mutex, rc) (rc)
347 /** A flag for opening a file and requesting synchronous data writes.
348 * This is only used when writing a meta page. It's not strictly needed;
349 * we could just do a normal write and then immediately perform a flush.
350 * But if this flag is available it saves us an extra system call.
352 * @note If O_DSYNC is undefined but exists in /usr/include,
353 * preferably set some compiler flag to get the definition.
354 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
357 # define MDB_DSYNC O_DSYNC
361 /** Function for flushing the data of a file. Define this to fsync
362 * if fdatasync() is not supported.
364 #ifndef MDB_FDATASYNC
365 # define MDB_FDATASYNC fdatasync
369 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
380 /** A page number in the database.
381 * Note that 64 bit page numbers are overkill, since pages themselves
382 * already represent 12-13 bits of addressable memory, and the OS will
383 * always limit applications to a maximum of 63 bits of address space.
385 * @note In the #MDB_node structure, we only store 48 bits of this value,
386 * which thus limits us to only 60 bits of addressable data.
388 typedef MDB_ID pgno_t;
390 /** A transaction ID.
391 * See struct MDB_txn.mt_txnid for details.
393 typedef MDB_ID txnid_t;
395 /** @defgroup debug Debug Macros
399 /** Enable debug output. Needs variable argument macros (a C99 feature).
400 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
401 * read from and written to the database (used for free space management).
407 static int mdb_debug;
408 static txnid_t mdb_debug_start;
410 /** Print a debug message with printf formatting.
411 * Requires double parenthesis around 2 or more args.
413 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
414 # define DPRINTF0(fmt, ...) \
415 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
417 # define DPRINTF(args) ((void) 0)
419 /** Print a debug string.
420 * The string is printed literally, with no format processing.
422 #define DPUTS(arg) DPRINTF(("%s", arg))
423 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
425 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
428 /** @brief The maximum size of a database page.
430 * It is 32k or 64k, since value-PAGEBASE must fit in
431 * #MDB_page.%mp_upper.
433 * LMDB will use database pages < OS pages if needed.
434 * That causes more I/O in write transactions: The OS must
435 * know (read) the whole page before writing a partial page.
437 * Note that we don't currently support Huge pages. On Linux,
438 * regular data files cannot use Huge pages, and in general
439 * Huge pages aren't actually pageable. We rely on the OS
440 * demand-pager to read our data and page it out when memory
441 * pressure from other processes is high. So until OSs have
442 * actual paging support for Huge pages, they're not viable.
444 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
446 /** The minimum number of keys required in a database page.
447 * Setting this to a larger value will place a smaller bound on the
448 * maximum size of a data item. Data items larger than this size will
449 * be pushed into overflow pages instead of being stored directly in
450 * the B-tree node. This value used to default to 4. With a page size
451 * of 4096 bytes that meant that any item larger than 1024 bytes would
452 * go into an overflow page. That also meant that on average 2-3KB of
453 * each overflow page was wasted space. The value cannot be lower than
454 * 2 because then there would no longer be a tree structure. With this
455 * value, items larger than 2KB will go into overflow pages, and on
456 * average only 1KB will be wasted.
458 #define MDB_MINKEYS 2
460 /** A stamp that identifies a file as an LMDB file.
461 * There's nothing special about this value other than that it is easily
462 * recognizable, and it will reflect any byte order mismatches.
464 #define MDB_MAGIC 0xBEEFC0DE
466 /** The version number for a database's datafile format. */
467 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
468 /** The version number for a database's lockfile format. */
469 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
471 /** @brief The max size of a key we can write, or 0 for dynamic max.
473 * Define this as 0 to compute the max from the page size. 511
474 * is default for backwards compat: liblmdb <= 0.9.10 can break
475 * when modifying a DB with keys/dupsort data bigger than its max.
476 * #MDB_DEVEL sets the default to 0.
478 * Data items in an #MDB_DUPSORT database are also limited to
479 * this size, since they're actually keys of a sub-DB. Keys and
480 * #MDB_DUPSORT data items must fit on a node in a regular page.
482 #ifndef MDB_MAXKEYSIZE
483 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
486 /** The maximum size of a key we can write to the environment. */
488 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
490 #define ENV_MAXKEY(env) ((env)->me_maxkey)
493 /** @brief The maximum size of a data item.
495 * We only store a 32 bit value for node sizes.
497 #define MAXDATASIZE 0xffffffffUL
500 /** Key size which fits in a #DKBUF.
503 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
506 * This is used for printing a hex dump of a key's contents.
508 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
509 /** Display a key in hex.
511 * Invoke a function to display a key in hex.
513 #define DKEY(x) mdb_dkey(x, kbuf)
519 /** An invalid page number.
520 * Mainly used to denote an empty tree.
522 #define P_INVALID (~(pgno_t)0)
524 /** Test if the flags \b f are set in a flag word \b w. */
525 #define F_ISSET(w, f) (((w) & (f)) == (f))
527 /** Round \b n up to an even number. */
528 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
530 /** Used for offsets within a single page.
531 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
534 typedef uint16_t indx_t;
536 /** Default size of memory map.
537 * This is certainly too small for any actual applications. Apps should always set
538 * the size explicitly using #mdb_env_set_mapsize().
540 #define DEFAULT_MAPSIZE 1048576
542 /** @defgroup readers Reader Lock Table
543 * Readers don't acquire any locks for their data access. Instead, they
544 * simply record their transaction ID in the reader table. The reader
545 * mutex is needed just to find an empty slot in the reader table. The
546 * slot's address is saved in thread-specific data so that subsequent read
547 * transactions started by the same thread need no further locking to proceed.
549 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
551 * No reader table is used if the database is on a read-only filesystem, or
552 * if #MDB_NOLOCK is set.
554 * Since the database uses multi-version concurrency control, readers don't
555 * actually need any locking. This table is used to keep track of which
556 * readers are using data from which old transactions, so that we'll know
557 * when a particular old transaction is no longer in use. Old transactions
558 * that have discarded any data pages can then have those pages reclaimed
559 * for use by a later write transaction.
561 * The lock table is constructed such that reader slots are aligned with the
562 * processor's cache line size. Any slot is only ever used by one thread.
563 * This alignment guarantees that there will be no contention or cache
564 * thrashing as threads update their own slot info, and also eliminates
565 * any need for locking when accessing a slot.
567 * A writer thread will scan every slot in the table to determine the oldest
568 * outstanding reader transaction. Any freed pages older than this will be
569 * reclaimed by the writer. The writer doesn't use any locks when scanning
570 * this table. This means that there's no guarantee that the writer will
571 * see the most up-to-date reader info, but that's not required for correct
572 * operation - all we need is to know the upper bound on the oldest reader,
573 * we don't care at all about the newest reader. So the only consequence of
574 * reading stale information here is that old pages might hang around a
575 * while longer before being reclaimed. That's actually good anyway, because
576 * the longer we delay reclaiming old pages, the more likely it is that a
577 * string of contiguous pages can be found after coalescing old pages from
578 * many old transactions together.
581 /** Number of slots in the reader table.
582 * This value was chosen somewhat arbitrarily. 126 readers plus a
583 * couple mutexes fit exactly into 8KB on my development machine.
584 * Applications should set the table size using #mdb_env_set_maxreaders().
586 #define DEFAULT_READERS 126
588 /** The size of a CPU cache line in bytes. We want our lock structures
589 * aligned to this size to avoid false cache line sharing in the
591 * This value works for most CPUs. For Itanium this should be 128.
597 /** The information we store in a single slot of the reader table.
598 * In addition to a transaction ID, we also record the process and
599 * thread ID that owns a slot, so that we can detect stale information,
600 * e.g. threads or processes that went away without cleaning up.
601 * @note We currently don't check for stale records. We simply re-init
602 * the table when we know that we're the only process opening the
605 typedef struct MDB_rxbody {
606 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
607 * Multiple readers that start at the same time will probably have the
608 * same ID here. Again, it's not important to exclude them from
609 * anything; all we need to know is which version of the DB they
610 * started from so we can avoid overwriting any data used in that
611 * particular version.
613 volatile txnid_t mrb_txnid;
614 /** The process ID of the process owning this reader txn. */
615 volatile MDB_PID_T mrb_pid;
616 /** The thread ID of the thread owning this txn. */
617 volatile MDB_THR_T mrb_tid;
620 /** The actual reader record, with cacheline padding. */
621 typedef struct MDB_reader {
624 /** shorthand for mrb_txnid */
625 #define mr_txnid mru.mrx.mrb_txnid
626 #define mr_pid mru.mrx.mrb_pid
627 #define mr_tid mru.mrx.mrb_tid
628 /** cache line alignment */
629 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
633 /** The header for the reader table.
634 * The table resides in a memory-mapped file. (This is a different file
635 * than is used for the main database.)
637 * For POSIX the actual mutexes reside in the shared memory of this
638 * mapped file. On Windows, mutexes are named objects allocated by the
639 * kernel; we store the mutex names in this mapped file so that other
640 * processes can grab them. This same approach is also used on
641 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
642 * process-shared POSIX mutexes. For these cases where a named object
643 * is used, the object name is derived from a 64 bit FNV hash of the
644 * environment pathname. As such, naming collisions are extremely
645 * unlikely. If a collision occurs, the results are unpredictable.
647 typedef struct MDB_txbody {
648 /** Stamp identifying this as an LMDB file. It must be set
651 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
654 char mtb_rmname[MNAME_LEN];
655 #elif defined(MDB_USE_SYSV_SEM)
658 /** Mutex protecting access to this table.
659 * This is the #MDB_MUTEX(env,r) reader table lock.
661 pthread_mutex_t mtb_rmutex;
663 /** The ID of the last transaction committed to the database.
664 * This is recorded here only for convenience; the value can always
665 * be determined by reading the main database meta pages.
667 volatile txnid_t mtb_txnid;
668 /** The number of slots that have been used in the reader table.
669 * This always records the maximum count, it is not decremented
670 * when readers release their slots.
672 volatile unsigned mtb_numreaders;
675 /** The actual reader table definition. */
676 typedef struct MDB_txninfo {
679 #define mti_magic mt1.mtb.mtb_magic
680 #define mti_format mt1.mtb.mtb_format
681 #define mti_rmutex mt1.mtb.mtb_rmutex
682 #define mti_rmname mt1.mtb.mtb_rmname
683 #define mti_txnid mt1.mtb.mtb_txnid
684 #define mti_numreaders mt1.mtb.mtb_numreaders
685 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
689 char mt2_wmname[MNAME_LEN];
690 #define mti_wmname mt2.mt2_wmname
691 #elif defined(MDB_USE_SYSV_SEM)
692 #define mti_semid mt1.mtb.mtb_semid
694 pthread_mutex_t mt2_wmutex;
695 #define mti_wmutex mt2.mt2_wmutex
697 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
699 MDB_reader mti_readers[1];
702 /** Lockfile format signature: version, features and field layout */
703 #define MDB_LOCK_FORMAT \
705 ((MDB_LOCK_VERSION) \
706 /* Flags which describe functionality */ \
707 + (((MDB_PIDLOCK) != 0) << 16)))
710 /** Common header for all page types.
711 * Overflow records occupy a number of contiguous pages with no
712 * headers on any page after the first.
714 typedef struct MDB_page {
715 #define mp_pgno mp_p.p_pgno
716 #define mp_next mp_p.p_next
718 pgno_t p_pgno; /**< page number */
719 struct MDB_page *p_next; /**< for in-memory list of freed pages */
722 /** @defgroup mdb_page Page Flags
724 * Flags for the page headers.
727 #define P_BRANCH 0x01 /**< branch page */
728 #define P_LEAF 0x02 /**< leaf page */
729 #define P_OVERFLOW 0x04 /**< overflow page */
730 #define P_META 0x08 /**< meta page */
731 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
732 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
733 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
734 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
735 #define P_KEEP 0x8000 /**< leave this page alone during spill */
737 uint16_t mp_flags; /**< @ref mdb_page */
738 #define mp_lower mp_pb.pb.pb_lower
739 #define mp_upper mp_pb.pb.pb_upper
740 #define mp_pages mp_pb.pb_pages
743 indx_t pb_lower; /**< lower bound of free space */
744 indx_t pb_upper; /**< upper bound of free space */
746 uint32_t pb_pages; /**< number of overflow pages */
748 indx_t mp_ptrs[1]; /**< dynamic size */
751 /** Size of the page header, excluding dynamic data at the end */
752 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
754 /** Address of first usable data byte in a page, after the header */
755 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
757 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
758 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
760 /** Number of nodes on a page */
761 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
763 /** The amount of space remaining in the page */
764 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
766 /** The percentage of space used in the page, in tenths of a percent. */
767 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
768 ((env)->me_psize - PAGEHDRSZ))
769 /** The minimum page fill factor, in tenths of a percent.
770 * Pages emptier than this are candidates for merging.
772 #define FILL_THRESHOLD 250
774 /** Test if a page is a leaf page */
775 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
776 /** Test if a page is a LEAF2 page */
777 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
778 /** Test if a page is a branch page */
779 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
780 /** Test if a page is an overflow page */
781 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
782 /** Test if a page is a sub page */
783 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
785 /** The number of overflow pages needed to store the given size. */
786 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
788 /** Link in #MDB_txn.%mt_loose_pgs list */
789 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
791 /** Header for a single key/data pair within a page.
792 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
793 * We guarantee 2-byte alignment for 'MDB_node's.
795 typedef struct MDB_node {
796 /** lo and hi are used for data size on leaf nodes and for
797 * child pgno on branch nodes. On 64 bit platforms, flags
798 * is also used for pgno. (Branch nodes have no flags).
799 * They are in host byte order in case that lets some
800 * accesses be optimized into a 32-bit word access.
802 #if BYTE_ORDER == LITTLE_ENDIAN
803 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
805 unsigned short mn_hi, mn_lo;
807 /** @defgroup mdb_node Node Flags
809 * Flags for node headers.
812 #define F_BIGDATA 0x01 /**< data put on overflow page */
813 #define F_SUBDATA 0x02 /**< data is a sub-database */
814 #define F_DUPDATA 0x04 /**< data has duplicates */
816 /** valid flags for #mdb_node_add() */
817 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
820 unsigned short mn_flags; /**< @ref mdb_node */
821 unsigned short mn_ksize; /**< key size */
822 char mn_data[1]; /**< key and data are appended here */
825 /** Size of the node header, excluding dynamic data at the end */
826 #define NODESIZE offsetof(MDB_node, mn_data)
828 /** Bit position of top word in page number, for shifting mn_flags */
829 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
831 /** Size of a node in a branch page with a given key.
832 * This is just the node header plus the key, there is no data.
834 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
836 /** Size of a node in a leaf page with a given key and data.
837 * This is node header plus key plus data size.
839 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
841 /** Address of node \b i in page \b p */
842 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
844 /** Address of the key for the node */
845 #define NODEKEY(node) (void *)((node)->mn_data)
847 /** Address of the data for a node */
848 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
850 /** Get the page number pointed to by a branch node */
851 #define NODEPGNO(node) \
852 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
853 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
854 /** Set the page number in a branch node */
855 #define SETPGNO(node,pgno) do { \
856 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
857 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
859 /** Get the size of the data in a leaf node */
860 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
861 /** Set the size of the data for a leaf node */
862 #define SETDSZ(node,size) do { \
863 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
864 /** The size of a key in a node */
865 #define NODEKSZ(node) ((node)->mn_ksize)
867 /** Copy a page number from src to dst */
869 #define COPY_PGNO(dst,src) dst = src
871 #if SIZE_MAX > 4294967295UL
872 #define COPY_PGNO(dst,src) do { \
873 unsigned short *s, *d; \
874 s = (unsigned short *)&(src); \
875 d = (unsigned short *)&(dst); \
882 #define COPY_PGNO(dst,src) do { \
883 unsigned short *s, *d; \
884 s = (unsigned short *)&(src); \
885 d = (unsigned short *)&(dst); \
891 /** The address of a key in a LEAF2 page.
892 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
893 * There are no node headers, keys are stored contiguously.
895 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
897 /** Set the \b node's key into \b keyptr, if requested. */
898 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
899 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
901 /** Set the \b node's key into \b key. */
902 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
904 /** Information about a single database in the environment. */
905 typedef struct MDB_db {
906 uint32_t md_pad; /**< also ksize for LEAF2 pages */
907 uint16_t md_flags; /**< @ref mdb_dbi_open */
908 uint16_t md_depth; /**< depth of this tree */
909 pgno_t md_branch_pages; /**< number of internal pages */
910 pgno_t md_leaf_pages; /**< number of leaf pages */
911 pgno_t md_overflow_pages; /**< number of overflow pages */
912 size_t md_entries; /**< number of data items */
913 pgno_t md_root; /**< the root page of this tree */
916 /** mdb_dbi_open flags */
917 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
918 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
919 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
920 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
922 /** Handle for the DB used to track free pages. */
924 /** Handle for the default DB. */
927 /** Meta page content.
928 * A meta page is the start point for accessing a database snapshot.
929 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
931 typedef struct MDB_meta {
932 /** Stamp identifying this as an LMDB file. It must be set
935 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
937 void *mm_address; /**< address for fixed mapping */
938 size_t mm_mapsize; /**< size of mmap region */
939 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
940 /** The size of pages used in this DB */
941 #define mm_psize mm_dbs[0].md_pad
942 /** Any persistent environment flags. @ref mdb_env */
943 #define mm_flags mm_dbs[0].md_flags
944 pgno_t mm_last_pg; /**< last used page in file */
945 volatile txnid_t mm_txnid; /**< txnid that committed this page */
948 /** Buffer for a stack-allocated meta page.
949 * The members define size and alignment, and silence type
950 * aliasing warnings. They are not used directly; that could
951 * mean incorrectly using several union members in parallel.
953 typedef union MDB_metabuf {
956 char mm_pad[PAGEHDRSZ];
961 /** Auxiliary DB info.
962 * The information here is mostly static/read-only. There is
963 * only a single copy of this record in the environment.
965 typedef struct MDB_dbx {
966 MDB_val md_name; /**< name of the database */
967 MDB_cmp_func *md_cmp; /**< function for comparing keys */
968 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
969 MDB_rel_func *md_rel; /**< user relocate function */
970 void *md_relctx; /**< user-provided context for md_rel */
973 /** A database transaction.
974 * Every operation requires a transaction handle.
977 MDB_txn *mt_parent; /**< parent of a nested txn */
978 MDB_txn *mt_child; /**< nested txn under this txn */
979 pgno_t mt_next_pgno; /**< next unallocated page */
980 /** The ID of this transaction. IDs are integers incrementing from 1.
981 * Only committed write transactions increment the ID. If a transaction
982 * aborts, the ID may be re-used by the next writer.
985 MDB_env *mt_env; /**< the DB environment */
986 /** The list of pages that became unused during this transaction.
989 /** The list of loose pages that became unused and may be reused
990 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
992 MDB_page *mt_loose_pgs;
993 /* #Number of loose pages (#mt_loose_pgs) */
995 /** The sorted list of dirty pages we temporarily wrote to disk
996 * because the dirty list was full. page numbers in here are
997 * shifted left by 1, deleted slots have the LSB set.
999 MDB_IDL mt_spill_pgs;
1001 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1002 MDB_ID2L dirty_list;
1003 /** For read txns: This thread/txn's reader table slot, or NULL. */
1006 /** Array of records for each DB known in the environment. */
1008 /** Array of MDB_db records for each known DB */
1010 /** Array of sequence numbers for each DB handle */
1011 unsigned int *mt_dbiseqs;
1012 /** @defgroup mt_dbflag Transaction DB Flags
1016 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1017 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1018 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1019 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1021 /** In write txns, array of cursors for each DB */
1022 MDB_cursor **mt_cursors;
1023 /** Array of flags for each DB */
1024 unsigned char *mt_dbflags;
1025 /** Number of DB records in use. This number only ever increments;
1026 * we don't decrement it when individual DB handles are closed.
1030 /** @defgroup mdb_txn Transaction Flags
1034 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
1035 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1036 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1037 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1039 unsigned int mt_flags; /**< @ref mdb_txn */
1040 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1041 * Includes ancestor txns' dirty pages not hidden by other txns'
1042 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1043 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1045 unsigned int mt_dirty_room;
1048 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1049 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1050 * raise this on a 64 bit machine.
1052 #define CURSOR_STACK 32
1056 /** Cursors are used for all DB operations.
1057 * A cursor holds a path of (page pointer, key index) from the DB
1058 * root to a position in the DB, plus other state. #MDB_DUPSORT
1059 * cursors include an xcursor to the current data item. Write txns
1060 * track their cursors and keep them up to date when data moves.
1061 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1062 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1065 /** Next cursor on this DB in this txn */
1066 MDB_cursor *mc_next;
1067 /** Backup of the original cursor if this cursor is a shadow */
1068 MDB_cursor *mc_backup;
1069 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1070 struct MDB_xcursor *mc_xcursor;
1071 /** The transaction that owns this cursor */
1073 /** The database handle this cursor operates on */
1075 /** The database record for this cursor */
1077 /** The database auxiliary record for this cursor */
1079 /** The @ref mt_dbflag for this database */
1080 unsigned char *mc_dbflag;
1081 unsigned short mc_snum; /**< number of pushed pages */
1082 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1083 /** @defgroup mdb_cursor Cursor Flags
1085 * Cursor state flags.
1088 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1089 #define C_EOF 0x02 /**< No more data */
1090 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1091 #define C_DEL 0x08 /**< last op was a cursor_del */
1092 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1093 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1095 unsigned int mc_flags; /**< @ref mdb_cursor */
1096 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1097 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1100 /** Context for sorted-dup records.
1101 * We could have gone to a fully recursive design, with arbitrarily
1102 * deep nesting of sub-databases. But for now we only handle these
1103 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1105 typedef struct MDB_xcursor {
1106 /** A sub-cursor for traversing the Dup DB */
1107 MDB_cursor mx_cursor;
1108 /** The database record for this Dup DB */
1110 /** The auxiliary DB record for this Dup DB */
1112 /** The @ref mt_dbflag for this Dup DB */
1113 unsigned char mx_dbflag;
1116 /** State of FreeDB old pages, stored in the MDB_env */
1117 typedef struct MDB_pgstate {
1118 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1119 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1122 /** The database environment. */
1124 HANDLE me_fd; /**< The main data file */
1125 HANDLE me_lfd; /**< The lock file */
1126 HANDLE me_mfd; /**< just for writing the meta pages */
1127 /** Failed to update the meta page. Probably an I/O error. */
1128 #define MDB_FATAL_ERROR 0x80000000U
1129 /** Some fields are initialized. */
1130 #define MDB_ENV_ACTIVE 0x20000000U
1131 /** me_txkey is set */
1132 #define MDB_ENV_TXKEY 0x10000000U
1133 uint32_t me_flags; /**< @ref mdb_env */
1134 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1135 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1136 unsigned int me_maxreaders; /**< size of the reader table */
1137 unsigned int me_numreaders; /**< max numreaders set by this env */
1138 MDB_dbi me_numdbs; /**< number of DBs opened */
1139 MDB_dbi me_maxdbs; /**< size of the DB table */
1140 MDB_PID_T me_pid; /**< process ID of this env */
1141 char *me_path; /**< path to the DB files */
1142 char *me_map; /**< the memory map of the data file */
1143 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1144 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1145 void *me_pbuf; /**< scratch area for DUPSORT put() */
1146 MDB_txn *me_txn; /**< current write transaction */
1147 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1148 size_t me_mapsize; /**< size of the data memory map */
1149 off_t me_size; /**< current file size */
1150 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1151 MDB_dbx *me_dbxs; /**< array of static DB info */
1152 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1153 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1154 pthread_key_t me_txkey; /**< thread-key for readers */
1155 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1156 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1157 # define me_pglast me_pgstate.mf_pglast
1158 # define me_pghead me_pgstate.mf_pghead
1159 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1160 /** IDL of pages that became unused in a write txn */
1161 MDB_IDL me_free_pgs;
1162 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1163 MDB_ID2L me_dirty_list;
1164 /** Max number of freelist items that can fit in a single overflow page */
1166 /** Max size of a node on a page */
1167 unsigned int me_nodemax;
1168 #if !(MDB_MAXKEYSIZE)
1169 unsigned int me_maxkey; /**< max size of a key */
1171 int me_live_reader; /**< have liveness lock in reader table */
1173 int me_pidquery; /**< Used in OpenProcess */
1175 #if defined(_WIN32) || defined(MDB_USE_SYSV_SEM)
1176 /* Windows mutexes/SysV semaphores do not reside in shared mem */
1177 mdb_mutex_t me_rmutex;
1178 mdb_mutex_t me_wmutex;
1180 void *me_userctx; /**< User-settable context */
1181 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1184 /** Nested transaction */
1185 typedef struct MDB_ntxn {
1186 MDB_txn mnt_txn; /**< the transaction */
1187 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1190 /** max number of pages to commit in one writev() call */
1191 #define MDB_COMMIT_PAGES 64
1192 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1193 #undef MDB_COMMIT_PAGES
1194 #define MDB_COMMIT_PAGES IOV_MAX
1197 /** max bytes to write in one call */
1198 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1200 /** Check \b txn and \b dbi arguments to a function */
1201 #define TXN_DBI_EXIST(txn, dbi) \
1202 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1204 /** Check for misused \b dbi handles */
1205 #define TXN_DBI_CHANGED(txn, dbi) \
1206 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1208 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1209 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1210 static int mdb_page_touch(MDB_cursor *mc);
1212 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1213 static int mdb_page_search_root(MDB_cursor *mc,
1214 MDB_val *key, int modify);
1215 #define MDB_PS_MODIFY 1
1216 #define MDB_PS_ROOTONLY 2
1217 #define MDB_PS_FIRST 4
1218 #define MDB_PS_LAST 8
1219 static int mdb_page_search(MDB_cursor *mc,
1220 MDB_val *key, int flags);
1221 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1223 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1224 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1225 pgno_t newpgno, unsigned int nflags);
1227 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1228 static int mdb_env_pick_meta(const MDB_env *env);
1229 static int mdb_env_write_meta(MDB_txn *txn);
1230 #if !(defined(_WIN32) || defined(MDB_USE_SYSV_SEM)) /* Drop unused excl arg */
1231 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1233 static void mdb_env_close0(MDB_env *env, int excl);
1235 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1236 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1237 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1238 static void mdb_node_del(MDB_cursor *mc, int ksize);
1239 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1240 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1241 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1242 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1243 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1245 static int mdb_rebalance(MDB_cursor *mc);
1246 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1248 static void mdb_cursor_pop(MDB_cursor *mc);
1249 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1251 static int mdb_cursor_del0(MDB_cursor *mc);
1252 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1253 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1254 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1255 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1256 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1258 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1259 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1261 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1262 static void mdb_xcursor_init0(MDB_cursor *mc);
1263 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1265 static int mdb_drop0(MDB_cursor *mc, int subs);
1266 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1267 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1270 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1274 static SECURITY_DESCRIPTOR mdb_null_sd;
1275 static SECURITY_ATTRIBUTES mdb_all_sa;
1276 static int mdb_sec_inited;
1279 /** Return the library version info. */
1281 mdb_version(int *major, int *minor, int *patch)
1283 if (major) *major = MDB_VERSION_MAJOR;
1284 if (minor) *minor = MDB_VERSION_MINOR;
1285 if (patch) *patch = MDB_VERSION_PATCH;
1286 return MDB_VERSION_STRING;
1289 /** Table of descriptions for LMDB @ref errors */
1290 static char *const mdb_errstr[] = {
1291 "MDB_KEYEXIST: Key/data pair already exists",
1292 "MDB_NOTFOUND: No matching key/data pair found",
1293 "MDB_PAGE_NOTFOUND: Requested page not found",
1294 "MDB_CORRUPTED: Located page was wrong type",
1295 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1296 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1297 "MDB_INVALID: File is not an LMDB file",
1298 "MDB_MAP_FULL: Environment mapsize limit reached",
1299 "MDB_DBS_FULL: Environment maxdbs limit reached",
1300 "MDB_READERS_FULL: Environment maxreaders limit reached",
1301 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1302 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1303 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1304 "MDB_PAGE_FULL: Internal error - page has no more space",
1305 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1306 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1307 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1308 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1309 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1310 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1314 mdb_strerror(int err)
1317 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1318 * This works as long as no function between the call to mdb_strerror
1319 * and the actual use of the message uses more than 4K of stack.
1322 char buf[1024], *ptr = buf;
1326 return ("Successful return: 0");
1328 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1329 i = err - MDB_KEYEXIST;
1330 return mdb_errstr[i];
1334 /* These are the C-runtime error codes we use. The comment indicates
1335 * their numeric value, and the Win32 error they would correspond to
1336 * if the error actually came from a Win32 API. A major mess, we should
1337 * have used LMDB-specific error codes for everything.
1340 case ENOENT: /* 2, FILE_NOT_FOUND */
1341 case EIO: /* 5, ACCESS_DENIED */
1342 case ENOMEM: /* 12, INVALID_ACCESS */
1343 case EACCES: /* 13, INVALID_DATA */
1344 case EBUSY: /* 16, CURRENT_DIRECTORY */
1345 case EINVAL: /* 22, BAD_COMMAND */
1346 case ENOSPC: /* 28, OUT_OF_PAPER */
1347 return strerror(err);
1352 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1353 FORMAT_MESSAGE_IGNORE_INSERTS,
1354 NULL, err, 0, ptr, sizeof(buf), pad);
1357 return strerror(err);
1361 /** assert(3) variant in cursor context */
1362 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1363 /** assert(3) variant in transaction context */
1364 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1365 /** assert(3) variant in environment context */
1366 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1369 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1370 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1373 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1374 const char *func, const char *file, int line)
1377 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1378 file, line, expr_txt, func);
1379 if (env->me_assert_func)
1380 env->me_assert_func(env, buf);
1381 fprintf(stderr, "%s\n", buf);
1385 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1389 /** Return the page number of \b mp which may be sub-page, for debug output */
1391 mdb_dbg_pgno(MDB_page *mp)
1394 COPY_PGNO(ret, mp->mp_pgno);
1398 /** Display a key in hexadecimal and return the address of the result.
1399 * @param[in] key the key to display
1400 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1401 * @return The key in hexadecimal form.
1404 mdb_dkey(MDB_val *key, char *buf)
1407 unsigned char *c = key->mv_data;
1413 if (key->mv_size > DKBUF_MAXKEYSIZE)
1414 return "MDB_MAXKEYSIZE";
1415 /* may want to make this a dynamic check: if the key is mostly
1416 * printable characters, print it as-is instead of converting to hex.
1420 for (i=0; i<key->mv_size; i++)
1421 ptr += sprintf(ptr, "%02x", *c++);
1423 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1429 mdb_leafnode_type(MDB_node *n)
1431 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1432 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1433 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1436 /** Display all the keys in the page. */
1438 mdb_page_list(MDB_page *mp)
1440 pgno_t pgno = mdb_dbg_pgno(mp);
1441 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1443 unsigned int i, nkeys, nsize, total = 0;
1447 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1448 case P_BRANCH: type = "Branch page"; break;
1449 case P_LEAF: type = "Leaf page"; break;
1450 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1451 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1452 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1454 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1455 pgno, mp->mp_pages, state);
1458 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1459 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1462 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1466 nkeys = NUMKEYS(mp);
1467 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1469 for (i=0; i<nkeys; i++) {
1470 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1471 key.mv_size = nsize = mp->mp_pad;
1472 key.mv_data = LEAF2KEY(mp, i, nsize);
1474 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1477 node = NODEPTR(mp, i);
1478 key.mv_size = node->mn_ksize;
1479 key.mv_data = node->mn_data;
1480 nsize = NODESIZE + key.mv_size;
1481 if (IS_BRANCH(mp)) {
1482 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1486 if (F_ISSET(node->mn_flags, F_BIGDATA))
1487 nsize += sizeof(pgno_t);
1489 nsize += NODEDSZ(node);
1491 nsize += sizeof(indx_t);
1492 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1493 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1495 total = EVEN(total);
1497 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1498 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1502 mdb_cursor_chk(MDB_cursor *mc)
1508 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1509 for (i=0; i<mc->mc_top; i++) {
1511 node = NODEPTR(mp, mc->mc_ki[i]);
1512 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1515 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1521 /** Count all the pages in each DB and in the freelist
1522 * and make sure it matches the actual number of pages
1524 * All named DBs must be open for a correct count.
1526 static void mdb_audit(MDB_txn *txn)
1530 MDB_ID freecount, count;
1535 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1536 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1537 freecount += *(MDB_ID *)data.mv_data;
1538 mdb_tassert(txn, rc == MDB_NOTFOUND);
1541 for (i = 0; i<txn->mt_numdbs; i++) {
1543 if (!(txn->mt_dbflags[i] & DB_VALID))
1545 mdb_cursor_init(&mc, txn, i, &mx);
1546 if (txn->mt_dbs[i].md_root == P_INVALID)
1548 count += txn->mt_dbs[i].md_branch_pages +
1549 txn->mt_dbs[i].md_leaf_pages +
1550 txn->mt_dbs[i].md_overflow_pages;
1551 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1552 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1553 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1556 mp = mc.mc_pg[mc.mc_top];
1557 for (j=0; j<NUMKEYS(mp); j++) {
1558 MDB_node *leaf = NODEPTR(mp, j);
1559 if (leaf->mn_flags & F_SUBDATA) {
1561 memcpy(&db, NODEDATA(leaf), sizeof(db));
1562 count += db.md_branch_pages + db.md_leaf_pages +
1563 db.md_overflow_pages;
1567 mdb_tassert(txn, rc == MDB_NOTFOUND);
1570 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1571 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1572 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1578 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1580 return txn->mt_dbxs[dbi].md_cmp(a, b);
1584 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1586 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1589 /** Allocate memory for a page.
1590 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1593 mdb_page_malloc(MDB_txn *txn, unsigned num)
1595 MDB_env *env = txn->mt_env;
1596 MDB_page *ret = env->me_dpages;
1597 size_t psize = env->me_psize, sz = psize, off;
1598 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1599 * For a single page alloc, we init everything after the page header.
1600 * For multi-page, we init the final page; if the caller needed that
1601 * many pages they will be filling in at least up to the last page.
1605 VGMEMP_ALLOC(env, ret, sz);
1606 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1607 env->me_dpages = ret->mp_next;
1610 psize -= off = PAGEHDRSZ;
1615 if ((ret = malloc(sz)) != NULL) {
1616 VGMEMP_ALLOC(env, ret, sz);
1617 if (!(env->me_flags & MDB_NOMEMINIT)) {
1618 memset((char *)ret + off, 0, psize);
1622 txn->mt_flags |= MDB_TXN_ERROR;
1626 /** Free a single page.
1627 * Saves single pages to a list, for future reuse.
1628 * (This is not used for multi-page overflow pages.)
1631 mdb_page_free(MDB_env *env, MDB_page *mp)
1633 mp->mp_next = env->me_dpages;
1634 VGMEMP_FREE(env, mp);
1635 env->me_dpages = mp;
1638 /** Free a dirty page */
1640 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1642 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1643 mdb_page_free(env, dp);
1645 /* large pages just get freed directly */
1646 VGMEMP_FREE(env, dp);
1651 /** Return all dirty pages to dpage list */
1653 mdb_dlist_free(MDB_txn *txn)
1655 MDB_env *env = txn->mt_env;
1656 MDB_ID2L dl = txn->mt_u.dirty_list;
1657 unsigned i, n = dl[0].mid;
1659 for (i = 1; i <= n; i++) {
1660 mdb_dpage_free(env, dl[i].mptr);
1665 /** Loosen or free a single page.
1666 * Saves single pages to a list for future reuse
1667 * in this same txn. It has been pulled from the freeDB
1668 * and already resides on the dirty list, but has been
1669 * deleted. Use these pages first before pulling again
1672 * If the page wasn't dirtied in this txn, just add it
1673 * to this txn's free list.
1676 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1679 pgno_t pgno = mp->mp_pgno;
1680 MDB_txn *txn = mc->mc_txn;
1682 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1683 if (txn->mt_parent) {
1684 MDB_ID2 *dl = txn->mt_u.dirty_list;
1685 /* If txn has a parent, make sure the page is in our
1689 unsigned x = mdb_mid2l_search(dl, pgno);
1690 if (x <= dl[0].mid && dl[x].mid == pgno) {
1691 if (mp != dl[x].mptr) { /* bad cursor? */
1692 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1693 txn->mt_flags |= MDB_TXN_ERROR;
1694 return MDB_CORRUPTED;
1701 /* no parent txn, so it's just ours */
1706 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1708 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1709 txn->mt_loose_pgs = mp;
1710 txn->mt_loose_count++;
1711 mp->mp_flags |= P_LOOSE;
1713 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1721 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1722 * @param[in] mc A cursor handle for the current operation.
1723 * @param[in] pflags Flags of the pages to update:
1724 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1725 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1726 * @return 0 on success, non-zero on failure.
1729 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1731 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1732 MDB_txn *txn = mc->mc_txn;
1738 int rc = MDB_SUCCESS, level;
1740 /* Mark pages seen by cursors */
1741 if (mc->mc_flags & C_UNTRACK)
1742 mc = NULL; /* will find mc in mt_cursors */
1743 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1744 for (; mc; mc=mc->mc_next) {
1745 if (!(mc->mc_flags & C_INITIALIZED))
1747 for (m3 = mc;; m3 = &mx->mx_cursor) {
1749 for (j=0; j<m3->mc_snum; j++) {
1751 if ((mp->mp_flags & Mask) == pflags)
1752 mp->mp_flags ^= P_KEEP;
1754 mx = m3->mc_xcursor;
1755 /* Proceed to mx if it is at a sub-database */
1756 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1758 if (! (mp && (mp->mp_flags & P_LEAF)))
1760 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1761 if (!(leaf->mn_flags & F_SUBDATA))
1770 /* Mark dirty root pages */
1771 for (i=0; i<txn->mt_numdbs; i++) {
1772 if (txn->mt_dbflags[i] & DB_DIRTY) {
1773 pgno_t pgno = txn->mt_dbs[i].md_root;
1774 if (pgno == P_INVALID)
1776 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1778 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1779 dp->mp_flags ^= P_KEEP;
1787 static int mdb_page_flush(MDB_txn *txn, int keep);
1789 /** Spill pages from the dirty list back to disk.
1790 * This is intended to prevent running into #MDB_TXN_FULL situations,
1791 * but note that they may still occur in a few cases:
1792 * 1) our estimate of the txn size could be too small. Currently this
1793 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1794 * 2) child txns may run out of space if their parents dirtied a
1795 * lot of pages and never spilled them. TODO: we probably should do
1796 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1797 * the parent's dirty_room is below a given threshold.
1799 * Otherwise, if not using nested txns, it is expected that apps will
1800 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1801 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1802 * If the txn never references them again, they can be left alone.
1803 * If the txn only reads them, they can be used without any fuss.
1804 * If the txn writes them again, they can be dirtied immediately without
1805 * going thru all of the work of #mdb_page_touch(). Such references are
1806 * handled by #mdb_page_unspill().
1808 * Also note, we never spill DB root pages, nor pages of active cursors,
1809 * because we'll need these back again soon anyway. And in nested txns,
1810 * we can't spill a page in a child txn if it was already spilled in a
1811 * parent txn. That would alter the parent txns' data even though
1812 * the child hasn't committed yet, and we'd have no way to undo it if
1813 * the child aborted.
1815 * @param[in] m0 cursor A cursor handle identifying the transaction and
1816 * database for which we are checking space.
1817 * @param[in] key For a put operation, the key being stored.
1818 * @param[in] data For a put operation, the data being stored.
1819 * @return 0 on success, non-zero on failure.
1822 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1824 MDB_txn *txn = m0->mc_txn;
1826 MDB_ID2L dl = txn->mt_u.dirty_list;
1827 unsigned int i, j, need;
1830 if (m0->mc_flags & C_SUB)
1833 /* Estimate how much space this op will take */
1834 i = m0->mc_db->md_depth;
1835 /* Named DBs also dirty the main DB */
1836 if (m0->mc_dbi > MAIN_DBI)
1837 i += txn->mt_dbs[MAIN_DBI].md_depth;
1838 /* For puts, roughly factor in the key+data size */
1840 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1841 i += i; /* double it for good measure */
1844 if (txn->mt_dirty_room > i)
1847 if (!txn->mt_spill_pgs) {
1848 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1849 if (!txn->mt_spill_pgs)
1852 /* purge deleted slots */
1853 MDB_IDL sl = txn->mt_spill_pgs;
1854 unsigned int num = sl[0];
1856 for (i=1; i<=num; i++) {
1863 /* Preserve pages which may soon be dirtied again */
1864 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1867 /* Less aggressive spill - we originally spilled the entire dirty list,
1868 * with a few exceptions for cursor pages and DB root pages. But this
1869 * turns out to be a lot of wasted effort because in a large txn many
1870 * of those pages will need to be used again. So now we spill only 1/8th
1871 * of the dirty pages. Testing revealed this to be a good tradeoff,
1872 * better than 1/2, 1/4, or 1/10.
1874 if (need < MDB_IDL_UM_MAX / 8)
1875 need = MDB_IDL_UM_MAX / 8;
1877 /* Save the page IDs of all the pages we're flushing */
1878 /* flush from the tail forward, this saves a lot of shifting later on. */
1879 for (i=dl[0].mid; i && need; i--) {
1880 MDB_ID pn = dl[i].mid << 1;
1882 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1884 /* Can't spill twice, make sure it's not already in a parent's
1887 if (txn->mt_parent) {
1889 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1890 if (tx2->mt_spill_pgs) {
1891 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1892 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1893 dp->mp_flags |= P_KEEP;
1901 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1905 mdb_midl_sort(txn->mt_spill_pgs);
1907 /* Flush the spilled part of dirty list */
1908 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1911 /* Reset any dirty pages we kept that page_flush didn't see */
1912 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1915 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1919 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1921 mdb_find_oldest(MDB_txn *txn)
1924 txnid_t mr, oldest = txn->mt_txnid - 1;
1925 if (txn->mt_env->me_txns) {
1926 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1927 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1938 /** Add a page to the txn's dirty list */
1940 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1943 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1945 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1946 insert = mdb_mid2l_append;
1948 insert = mdb_mid2l_insert;
1950 mid.mid = mp->mp_pgno;
1952 rc = insert(txn->mt_u.dirty_list, &mid);
1953 mdb_tassert(txn, rc == 0);
1954 txn->mt_dirty_room--;
1957 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1958 * me_pghead and mt_next_pgno.
1960 * If there are free pages available from older transactions, they
1961 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1962 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1963 * and move me_pglast to say which records were consumed. Only this
1964 * function can create me_pghead and move me_pglast/mt_next_pgno.
1965 * @param[in] mc cursor A cursor handle identifying the transaction and
1966 * database for which we are allocating.
1967 * @param[in] num the number of pages to allocate.
1968 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1969 * will always be satisfied by a single contiguous chunk of memory.
1970 * @return 0 on success, non-zero on failure.
1973 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1975 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1976 /* Get at most <Max_retries> more freeDB records once me_pghead
1977 * has enough pages. If not enough, use new pages from the map.
1978 * If <Paranoid> and mc is updating the freeDB, only get new
1979 * records if me_pghead is empty. Then the freelist cannot play
1980 * catch-up with itself by growing while trying to save it.
1982 enum { Paranoid = 1, Max_retries = 500 };
1984 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1986 int rc, retry = num * 60;
1987 MDB_txn *txn = mc->mc_txn;
1988 MDB_env *env = txn->mt_env;
1989 pgno_t pgno, *mop = env->me_pghead;
1990 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1992 txnid_t oldest = 0, last;
1997 /* If there are any loose pages, just use them */
1998 if (num == 1 && txn->mt_loose_pgs) {
1999 np = txn->mt_loose_pgs;
2000 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2001 txn->mt_loose_count--;
2002 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2010 /* If our dirty list is already full, we can't do anything */
2011 if (txn->mt_dirty_room == 0) {
2016 for (op = MDB_FIRST;; op = MDB_NEXT) {
2021 /* Seek a big enough contiguous page range. Prefer
2022 * pages at the tail, just truncating the list.
2028 if (mop[i-n2] == pgno+n2)
2035 if (op == MDB_FIRST) { /* 1st iteration */
2036 /* Prepare to fetch more and coalesce */
2037 last = env->me_pglast;
2038 oldest = env->me_pgoldest;
2039 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2042 key.mv_data = &last; /* will look up last+1 */
2043 key.mv_size = sizeof(last);
2045 if (Paranoid && mc->mc_dbi == FREE_DBI)
2048 if (Paranoid && retry < 0 && mop_len)
2052 /* Do not fetch more if the record will be too recent */
2053 if (oldest <= last) {
2055 oldest = mdb_find_oldest(txn);
2056 env->me_pgoldest = oldest;
2062 rc = mdb_cursor_get(&m2, &key, NULL, op);
2064 if (rc == MDB_NOTFOUND)
2068 last = *(txnid_t*)key.mv_data;
2069 if (oldest <= last) {
2071 oldest = mdb_find_oldest(txn);
2072 env->me_pgoldest = oldest;
2078 np = m2.mc_pg[m2.mc_top];
2079 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2080 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2083 idl = (MDB_ID *) data.mv_data;
2086 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2091 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2093 mop = env->me_pghead;
2095 env->me_pglast = last;
2097 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2098 last, txn->mt_dbs[FREE_DBI].md_root, i));
2100 DPRINTF(("IDL %"Z"u", idl[j]));
2102 /* Merge in descending sorted order */
2103 mdb_midl_xmerge(mop, idl);
2107 /* Use new pages from the map when nothing suitable in the freeDB */
2109 pgno = txn->mt_next_pgno;
2110 if (pgno + num >= env->me_maxpg) {
2111 DPUTS("DB size maxed out");
2117 if (env->me_flags & MDB_WRITEMAP) {
2118 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2120 if (!(np = mdb_page_malloc(txn, num))) {
2126 mop[0] = mop_len -= num;
2127 /* Move any stragglers down */
2128 for (j = i-num; j < mop_len; )
2129 mop[++j] = mop[++i];
2131 txn->mt_next_pgno = pgno + num;
2134 mdb_page_dirty(txn, np);
2140 txn->mt_flags |= MDB_TXN_ERROR;
2144 /** Copy the used portions of a non-overflow page.
2145 * @param[in] dst page to copy into
2146 * @param[in] src page to copy from
2147 * @param[in] psize size of a page
2150 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2152 enum { Align = sizeof(pgno_t) };
2153 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2155 /* If page isn't full, just copy the used portion. Adjust
2156 * alignment so memcpy may copy words instead of bytes.
2158 if ((unused &= -Align) && !IS_LEAF2(src)) {
2159 upper = (upper + PAGEBASE) & -Align;
2160 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2161 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2164 memcpy(dst, src, psize - unused);
2168 /** Pull a page off the txn's spill list, if present.
2169 * If a page being referenced was spilled to disk in this txn, bring
2170 * it back and make it dirty/writable again.
2171 * @param[in] txn the transaction handle.
2172 * @param[in] mp the page being referenced. It must not be dirty.
2173 * @param[out] ret the writable page, if any. ret is unchanged if
2174 * mp wasn't spilled.
2177 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2179 MDB_env *env = txn->mt_env;
2182 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2184 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2185 if (!tx2->mt_spill_pgs)
2187 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2188 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2191 if (txn->mt_dirty_room == 0)
2192 return MDB_TXN_FULL;
2193 if (IS_OVERFLOW(mp))
2197 if (env->me_flags & MDB_WRITEMAP) {
2200 np = mdb_page_malloc(txn, num);
2204 memcpy(np, mp, num * env->me_psize);
2206 mdb_page_copy(np, mp, env->me_psize);
2209 /* If in current txn, this page is no longer spilled.
2210 * If it happens to be the last page, truncate the spill list.
2211 * Otherwise mark it as deleted by setting the LSB.
2213 if (x == txn->mt_spill_pgs[0])
2214 txn->mt_spill_pgs[0]--;
2216 txn->mt_spill_pgs[x] |= 1;
2217 } /* otherwise, if belonging to a parent txn, the
2218 * page remains spilled until child commits
2221 mdb_page_dirty(txn, np);
2222 np->mp_flags |= P_DIRTY;
2230 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2231 * @param[in] mc cursor pointing to the page to be touched
2232 * @return 0 on success, non-zero on failure.
2235 mdb_page_touch(MDB_cursor *mc)
2237 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2238 MDB_txn *txn = mc->mc_txn;
2239 MDB_cursor *m2, *m3;
2243 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2244 if (txn->mt_flags & MDB_TXN_SPILLS) {
2246 rc = mdb_page_unspill(txn, mp, &np);
2252 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2253 (rc = mdb_page_alloc(mc, 1, &np)))
2256 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2257 mp->mp_pgno, pgno));
2258 mdb_cassert(mc, mp->mp_pgno != pgno);
2259 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2260 /* Update the parent page, if any, to point to the new page */
2262 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2263 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2264 SETPGNO(node, pgno);
2266 mc->mc_db->md_root = pgno;
2268 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2269 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2271 /* If txn has a parent, make sure the page is in our
2275 unsigned x = mdb_mid2l_search(dl, pgno);
2276 if (x <= dl[0].mid && dl[x].mid == pgno) {
2277 if (mp != dl[x].mptr) { /* bad cursor? */
2278 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2279 txn->mt_flags |= MDB_TXN_ERROR;
2280 return MDB_CORRUPTED;
2285 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2287 np = mdb_page_malloc(txn, 1);
2292 rc = mdb_mid2l_insert(dl, &mid);
2293 mdb_cassert(mc, rc == 0);
2298 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2300 np->mp_flags |= P_DIRTY;
2303 /* Adjust cursors pointing to mp */
2304 mc->mc_pg[mc->mc_top] = np;
2305 m2 = txn->mt_cursors[mc->mc_dbi];
2306 if (mc->mc_flags & C_SUB) {
2307 for (; m2; m2=m2->mc_next) {
2308 m3 = &m2->mc_xcursor->mx_cursor;
2309 if (m3->mc_snum < mc->mc_snum) continue;
2310 if (m3->mc_pg[mc->mc_top] == mp)
2311 m3->mc_pg[mc->mc_top] = np;
2314 for (; m2; m2=m2->mc_next) {
2315 if (m2->mc_snum < mc->mc_snum) continue;
2316 if (m2->mc_pg[mc->mc_top] == mp) {
2317 m2->mc_pg[mc->mc_top] = np;
2318 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2320 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2322 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2323 if (!(leaf->mn_flags & F_SUBDATA))
2324 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2332 txn->mt_flags |= MDB_TXN_ERROR;
2337 mdb_env_sync(MDB_env *env, int force)
2340 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2341 if (env->me_flags & MDB_WRITEMAP) {
2342 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2343 ? MS_ASYNC : MS_SYNC;
2344 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2347 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2351 if (MDB_FDATASYNC(env->me_fd))
2358 /** Back up parent txn's cursors, then grab the originals for tracking */
2360 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2362 MDB_cursor *mc, *bk;
2367 for (i = src->mt_numdbs; --i >= 0; ) {
2368 if ((mc = src->mt_cursors[i]) != NULL) {
2369 size = sizeof(MDB_cursor);
2371 size += sizeof(MDB_xcursor);
2372 for (; mc; mc = bk->mc_next) {
2378 mc->mc_db = &dst->mt_dbs[i];
2379 /* Kill pointers into src - and dst to reduce abuse: The
2380 * user may not use mc until dst ends. Otherwise we'd...
2382 mc->mc_txn = NULL; /* ...set this to dst */
2383 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2384 if ((mx = mc->mc_xcursor) != NULL) {
2385 *(MDB_xcursor *)(bk+1) = *mx;
2386 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2388 mc->mc_next = dst->mt_cursors[i];
2389 dst->mt_cursors[i] = mc;
2396 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2397 * @param[in] txn the transaction handle.
2398 * @param[in] merge true to keep changes to parent cursors, false to revert.
2399 * @return 0 on success, non-zero on failure.
2402 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2404 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2408 for (i = txn->mt_numdbs; --i >= 0; ) {
2409 for (mc = cursors[i]; mc; mc = next) {
2411 if ((bk = mc->mc_backup) != NULL) {
2413 /* Commit changes to parent txn */
2414 mc->mc_next = bk->mc_next;
2415 mc->mc_backup = bk->mc_backup;
2416 mc->mc_txn = bk->mc_txn;
2417 mc->mc_db = bk->mc_db;
2418 mc->mc_dbflag = bk->mc_dbflag;
2419 if ((mx = mc->mc_xcursor) != NULL)
2420 mx->mx_cursor.mc_txn = bk->mc_txn;
2422 /* Abort nested txn */
2424 if ((mx = mc->mc_xcursor) != NULL)
2425 *mx = *(MDB_xcursor *)(bk+1);
2429 /* Only malloced cursors are permanently tracked. */
2437 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2440 mdb_txn_reset0(MDB_txn *txn, const char *act);
2442 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2448 Pidset = F_SETLK, Pidcheck = F_GETLK
2452 /** Set or check a pid lock. Set returns 0 on success.
2453 * Check returns 0 if the process is certainly dead, nonzero if it may
2454 * be alive (the lock exists or an error happened so we do not know).
2456 * On Windows Pidset is a no-op, we merely check for the existence
2457 * of the process with the given pid. On POSIX we use a single byte
2458 * lock on the lockfile, set at an offset equal to the pid.
2461 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2463 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2466 if (op == Pidcheck) {
2467 h = OpenProcess(env->me_pidquery, FALSE, pid);
2468 /* No documented "no such process" code, but other program use this: */
2470 return ErrCode() != ERROR_INVALID_PARAMETER;
2471 /* A process exists until all handles to it close. Has it exited? */
2472 ret = WaitForSingleObject(h, 0) != 0;
2479 struct flock lock_info;
2480 memset(&lock_info, 0, sizeof(lock_info));
2481 lock_info.l_type = F_WRLCK;
2482 lock_info.l_whence = SEEK_SET;
2483 lock_info.l_start = pid;
2484 lock_info.l_len = 1;
2485 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2486 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2488 } else if ((rc = ErrCode()) == EINTR) {
2496 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2497 * @param[in] txn the transaction handle to initialize
2498 * @return 0 on success, non-zero on failure.
2501 mdb_txn_renew0(MDB_txn *txn)
2503 MDB_env *env = txn->mt_env;
2504 MDB_txninfo *ti = env->me_txns;
2508 int rc, new_notls = 0;
2510 if (txn->mt_flags & MDB_TXN_RDONLY) {
2512 txn->mt_numdbs = env->me_numdbs;
2513 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2515 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2516 txn->mt_txnid = meta->mm_txnid;
2517 txn->mt_u.reader = NULL;
2519 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2520 pthread_getspecific(env->me_txkey);
2522 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2523 return MDB_BAD_RSLOT;
2525 MDB_PID_T pid = env->me_pid;
2526 MDB_THR_T tid = pthread_self();
2527 mdb_mutex_t *rmutex = MDB_MUTEX(env, r);
2529 if (!env->me_live_reader) {
2530 rc = mdb_reader_pid(env, Pidset, pid);
2533 env->me_live_reader = 1;
2536 if (LOCK_MUTEX(rc, env, rmutex))
2538 nr = ti->mti_numreaders;
2539 for (i=0; i<nr; i++)
2540 if (ti->mti_readers[i].mr_pid == 0)
2542 if (i == env->me_maxreaders) {
2543 UNLOCK_MUTEX(rmutex);
2544 return MDB_READERS_FULL;
2546 r = &ti->mti_readers[i];
2547 r->mr_txnid = (txnid_t)-1;
2549 r->mr_pid = pid; /* should be written last, see ITS#7971. */
2551 ti->mti_numreaders = ++nr;
2552 /* Save numreaders for un-mutexed mdb_env_close() */
2553 env->me_numreaders = nr;
2554 UNLOCK_MUTEX(rmutex);
2556 new_notls = (env->me_flags & MDB_NOTLS);
2557 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2562 do /* LY: Retry on a race, ITS#7970. */
2563 r->mr_txnid = ti->mti_txnid;
2564 while(r->mr_txnid != ti->mti_txnid);
2565 txn->mt_txnid = r->mr_txnid;
2566 txn->mt_u.reader = r;
2567 meta = env->me_metas[txn->mt_txnid & 1];
2571 if (LOCK_MUTEX(rc, env, MDB_MUTEX(env, w)))
2574 txn->mt_txnid = ti->mti_txnid;
2575 meta = env->me_metas[txn->mt_txnid & 1];
2577 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2578 txn->mt_txnid = meta->mm_txnid;
2581 txn->mt_numdbs = env->me_numdbs;
2584 if (txn->mt_txnid == mdb_debug_start)
2587 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2588 txn->mt_u.dirty_list = env->me_dirty_list;
2589 txn->mt_u.dirty_list[0].mid = 0;
2590 txn->mt_free_pgs = env->me_free_pgs;
2591 txn->mt_free_pgs[0] = 0;
2592 txn->mt_spill_pgs = NULL;
2594 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2597 /* Copy the DB info and flags */
2598 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2600 /* Moved to here to avoid a data race in read TXNs */
2601 txn->mt_next_pgno = meta->mm_last_pg+1;
2603 for (i=2; i<txn->mt_numdbs; i++) {
2604 x = env->me_dbflags[i];
2605 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2606 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2608 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2610 if (env->me_maxpg < txn->mt_next_pgno) {
2611 mdb_txn_reset0(txn, "renew0-mapfail");
2613 txn->mt_u.reader->mr_pid = 0;
2614 txn->mt_u.reader = NULL;
2616 return MDB_MAP_RESIZED;
2623 mdb_txn_renew(MDB_txn *txn)
2627 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2630 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2631 DPUTS("environment had fatal error, must shutdown!");
2635 rc = mdb_txn_renew0(txn);
2636 if (rc == MDB_SUCCESS) {
2637 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2638 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2639 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2645 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2649 int rc, size, tsize = sizeof(MDB_txn);
2651 if (env->me_flags & MDB_FATAL_ERROR) {
2652 DPUTS("environment had fatal error, must shutdown!");
2655 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2658 /* Nested transactions: Max 1 child, write txns only, no writemap */
2659 if (parent->mt_child ||
2660 (flags & MDB_RDONLY) ||
2661 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2662 (env->me_flags & MDB_WRITEMAP))
2664 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2666 tsize = sizeof(MDB_ntxn);
2669 if (!(flags & MDB_RDONLY)) {
2671 txn = env->me_txn0; /* just reuse preallocated write txn */
2675 /* child txns use own copy of cursors */
2676 size += env->me_maxdbs * sizeof(MDB_cursor *);
2678 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2680 if ((txn = calloc(1, size)) == NULL) {
2681 DPRINTF(("calloc: %s", strerror(errno)));
2684 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2685 if (flags & MDB_RDONLY) {
2686 txn->mt_flags |= MDB_TXN_RDONLY;
2687 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2688 txn->mt_dbiseqs = env->me_dbiseqs;
2690 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2692 txn->mt_dbiseqs = parent->mt_dbiseqs;
2693 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2695 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2696 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2704 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2705 if (!txn->mt_u.dirty_list ||
2706 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2708 free(txn->mt_u.dirty_list);
2712 txn->mt_txnid = parent->mt_txnid;
2713 txn->mt_dirty_room = parent->mt_dirty_room;
2714 txn->mt_u.dirty_list[0].mid = 0;
2715 txn->mt_spill_pgs = NULL;
2716 txn->mt_next_pgno = parent->mt_next_pgno;
2717 parent->mt_child = txn;
2718 txn->mt_parent = parent;
2719 txn->mt_numdbs = parent->mt_numdbs;
2720 txn->mt_flags = parent->mt_flags;
2721 txn->mt_dbxs = parent->mt_dbxs;
2722 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2723 /* Copy parent's mt_dbflags, but clear DB_NEW */
2724 for (i=0; i<txn->mt_numdbs; i++)
2725 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2727 ntxn = (MDB_ntxn *)txn;
2728 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2729 if (env->me_pghead) {
2730 size = MDB_IDL_SIZEOF(env->me_pghead);
2731 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2733 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2738 rc = mdb_cursor_shadow(parent, txn);
2740 mdb_txn_reset0(txn, "beginchild-fail");
2742 rc = mdb_txn_renew0(txn);
2745 if (txn != env->me_txn0)
2749 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2750 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2751 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2758 mdb_txn_env(MDB_txn *txn)
2760 if(!txn) return NULL;
2764 /** Export or close DBI handles opened in this txn. */
2766 mdb_dbis_update(MDB_txn *txn, int keep)
2769 MDB_dbi n = txn->mt_numdbs;
2770 MDB_env *env = txn->mt_env;
2771 unsigned char *tdbflags = txn->mt_dbflags;
2773 for (i = n; --i >= 2;) {
2774 if (tdbflags[i] & DB_NEW) {
2776 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2778 char *ptr = env->me_dbxs[i].md_name.mv_data;
2780 env->me_dbxs[i].md_name.mv_data = NULL;
2781 env->me_dbxs[i].md_name.mv_size = 0;
2782 env->me_dbflags[i] = 0;
2783 env->me_dbiseqs[i]++;
2789 if (keep && env->me_numdbs < n)
2793 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2794 * May be called twice for readonly txns: First reset it, then abort.
2795 * @param[in] txn the transaction handle to reset
2796 * @param[in] act why the transaction is being reset
2799 mdb_txn_reset0(MDB_txn *txn, const char *act)
2801 MDB_env *env = txn->mt_env;
2803 /* Close any DBI handles opened in this txn */
2804 mdb_dbis_update(txn, 0);
2806 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2807 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2808 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2810 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2811 if (txn->mt_u.reader) {
2812 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2813 if (!(env->me_flags & MDB_NOTLS))
2814 txn->mt_u.reader = NULL; /* txn does not own reader */
2816 txn->mt_numdbs = 0; /* close nothing if called again */
2817 txn->mt_dbxs = NULL; /* mark txn as reset */
2819 pgno_t *pghead = env->me_pghead;
2820 env->me_pghead = NULL;
2823 if (!(env->me_flags & MDB_WRITEMAP)) {
2824 mdb_dlist_free(txn);
2827 if (!txn->mt_parent) {
2828 if (mdb_midl_shrink(&txn->mt_free_pgs))
2829 env->me_free_pgs = txn->mt_free_pgs;
2832 /* The writer mutex was locked in mdb_txn_begin. */
2834 UNLOCK_MUTEX(MDB_MUTEX(env, w));
2837 mdb_cursors_close(txn, 0);
2839 mdb_midl_free(pghead);
2841 if (txn->mt_parent) {
2842 txn->mt_parent->mt_child = NULL;
2843 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2844 mdb_midl_free(txn->mt_free_pgs);
2845 mdb_midl_free(txn->mt_spill_pgs);
2846 free(txn->mt_u.dirty_list);
2852 mdb_txn_reset(MDB_txn *txn)
2857 /* This call is only valid for read-only txns */
2858 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2861 mdb_txn_reset0(txn, "reset");
2865 mdb_txn_abort(MDB_txn *txn)
2871 mdb_txn_abort(txn->mt_child);
2873 mdb_txn_reset0(txn, "abort");
2874 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2875 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2876 txn->mt_u.reader->mr_pid = 0;
2878 if (txn != txn->mt_env->me_txn0)
2882 /** Save the freelist as of this transaction to the freeDB.
2883 * This changes the freelist. Keep trying until it stabilizes.
2886 mdb_freelist_save(MDB_txn *txn)
2888 /* env->me_pghead[] can grow and shrink during this call.
2889 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2890 * Page numbers cannot disappear from txn->mt_free_pgs[].
2893 MDB_env *env = txn->mt_env;
2894 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2895 txnid_t pglast = 0, head_id = 0;
2896 pgno_t freecnt = 0, *free_pgs, *mop;
2897 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2899 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2901 if (env->me_pghead) {
2902 /* Make sure first page of freeDB is touched and on freelist */
2903 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2904 if (rc && rc != MDB_NOTFOUND)
2908 if (!env->me_pghead && txn->mt_loose_pgs) {
2909 /* Put loose page numbers in mt_free_pgs, since
2910 * we may be unable to return them to me_pghead.
2912 MDB_page *mp = txn->mt_loose_pgs;
2913 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2915 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2916 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2917 txn->mt_loose_pgs = NULL;
2918 txn->mt_loose_count = 0;
2921 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2922 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2923 ? SSIZE_MAX : maxfree_1pg;
2926 /* Come back here after each Put() in case freelist changed */
2931 /* If using records from freeDB which we have not yet
2932 * deleted, delete them and any we reserved for me_pghead.
2934 while (pglast < env->me_pglast) {
2935 rc = mdb_cursor_first(&mc, &key, NULL);
2938 pglast = head_id = *(txnid_t *)key.mv_data;
2939 total_room = head_room = 0;
2940 mdb_tassert(txn, pglast <= env->me_pglast);
2941 rc = mdb_cursor_del(&mc, 0);
2946 /* Save the IDL of pages freed by this txn, to a single record */
2947 if (freecnt < txn->mt_free_pgs[0]) {
2949 /* Make sure last page of freeDB is touched and on freelist */
2950 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2951 if (rc && rc != MDB_NOTFOUND)
2954 free_pgs = txn->mt_free_pgs;
2955 /* Write to last page of freeDB */
2956 key.mv_size = sizeof(txn->mt_txnid);
2957 key.mv_data = &txn->mt_txnid;
2959 freecnt = free_pgs[0];
2960 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2961 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2964 /* Retry if mt_free_pgs[] grew during the Put() */
2965 free_pgs = txn->mt_free_pgs;
2966 } while (freecnt < free_pgs[0]);
2967 mdb_midl_sort(free_pgs);
2968 memcpy(data.mv_data, free_pgs, data.mv_size);
2971 unsigned int i = free_pgs[0];
2972 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2973 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2975 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2981 mop = env->me_pghead;
2982 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2984 /* Reserve records for me_pghead[]. Split it if multi-page,
2985 * to avoid searching freeDB for a page range. Use keys in
2986 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2988 if (total_room >= mop_len) {
2989 if (total_room == mop_len || --more < 0)
2991 } else if (head_room >= maxfree_1pg && head_id > 1) {
2992 /* Keep current record (overflow page), add a new one */
2996 /* (Re)write {key = head_id, IDL length = head_room} */
2997 total_room -= head_room;
2998 head_room = mop_len - total_room;
2999 if (head_room > maxfree_1pg && head_id > 1) {
3000 /* Overflow multi-page for part of me_pghead */
3001 head_room /= head_id; /* amortize page sizes */
3002 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3003 } else if (head_room < 0) {
3004 /* Rare case, not bothering to delete this record */
3007 key.mv_size = sizeof(head_id);
3008 key.mv_data = &head_id;
3009 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3010 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3013 /* IDL is initially empty, zero out at least the length */
3014 pgs = (pgno_t *)data.mv_data;
3015 j = head_room > clean_limit ? head_room : 0;
3019 total_room += head_room;
3022 /* Return loose page numbers to me_pghead, though usually none are
3023 * left at this point. The pages themselves remain in dirty_list.
3025 if (txn->mt_loose_pgs) {
3026 MDB_page *mp = txn->mt_loose_pgs;
3027 unsigned count = txn->mt_loose_count;
3029 /* Room for loose pages + temp IDL with same */
3030 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3032 mop = env->me_pghead;
3033 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3034 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3035 loose[ ++count ] = mp->mp_pgno;
3037 mdb_midl_sort(loose);
3038 mdb_midl_xmerge(mop, loose);
3039 txn->mt_loose_pgs = NULL;
3040 txn->mt_loose_count = 0;
3044 /* Fill in the reserved me_pghead records */
3050 rc = mdb_cursor_first(&mc, &key, &data);
3051 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3052 txnid_t id = *(txnid_t *)key.mv_data;
3053 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3056 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3058 if (len > mop_len) {
3060 data.mv_size = (len + 1) * sizeof(MDB_ID);
3062 data.mv_data = mop -= len;
3065 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3067 if (rc || !(mop_len -= len))
3074 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3075 * @param[in] txn the transaction that's being committed
3076 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3077 * @return 0 on success, non-zero on failure.
3080 mdb_page_flush(MDB_txn *txn, int keep)
3082 MDB_env *env = txn->mt_env;
3083 MDB_ID2L dl = txn->mt_u.dirty_list;
3084 unsigned psize = env->me_psize, j;
3085 int i, pagecount = dl[0].mid, rc;
3086 size_t size = 0, pos = 0;
3088 MDB_page *dp = NULL;
3092 struct iovec iov[MDB_COMMIT_PAGES];
3093 ssize_t wpos = 0, wsize = 0, wres;
3094 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3100 if (env->me_flags & MDB_WRITEMAP) {
3101 /* Clear dirty flags */
3102 while (++i <= pagecount) {
3104 /* Don't flush this page yet */
3105 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3106 dp->mp_flags &= ~P_KEEP;
3110 dp->mp_flags &= ~P_DIRTY;
3115 /* Write the pages */
3117 if (++i <= pagecount) {
3119 /* Don't flush this page yet */
3120 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3121 dp->mp_flags &= ~P_KEEP;
3126 /* clear dirty flag */
3127 dp->mp_flags &= ~P_DIRTY;
3130 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3135 /* Windows actually supports scatter/gather I/O, but only on
3136 * unbuffered file handles. Since we're relying on the OS page
3137 * cache for all our data, that's self-defeating. So we just
3138 * write pages one at a time. We use the ov structure to set
3139 * the write offset, to at least save the overhead of a Seek
3142 DPRINTF(("committing page %"Z"u", pgno));
3143 memset(&ov, 0, sizeof(ov));
3144 ov.Offset = pos & 0xffffffff;
3145 ov.OffsetHigh = pos >> 16 >> 16;
3146 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3148 DPRINTF(("WriteFile: %d", rc));
3152 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3153 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3155 /* Write previous page(s) */
3156 #ifdef MDB_USE_PWRITEV
3157 wres = pwritev(env->me_fd, iov, n, wpos);
3160 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3162 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3164 DPRINTF(("lseek: %s", strerror(rc)));
3167 wres = writev(env->me_fd, iov, n);
3170 if (wres != wsize) {
3173 DPRINTF(("Write error: %s", strerror(rc)));
3175 rc = EIO; /* TODO: Use which error code? */
3176 DPUTS("short write, filesystem full?");
3187 DPRINTF(("committing page %"Z"u", pgno));
3188 next_pos = pos + size;
3189 iov[n].iov_len = size;
3190 iov[n].iov_base = (char *)dp;
3196 /* MIPS has cache coherency issues, this is a no-op everywhere else
3197 * Note: for any size >= on-chip cache size, entire on-chip cache is
3200 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3202 for (i = keep; ++i <= pagecount; ) {
3204 /* This is a page we skipped above */
3207 dl[j].mid = dp->mp_pgno;
3210 mdb_dpage_free(env, dp);
3215 txn->mt_dirty_room += i - j;
3221 mdb_txn_commit(MDB_txn *txn)
3227 if (txn == NULL || txn->mt_env == NULL)
3230 if (txn->mt_child) {
3231 rc = mdb_txn_commit(txn->mt_child);
3232 txn->mt_child = NULL;
3239 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3240 mdb_dbis_update(txn, 1);
3241 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3246 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3247 DPUTS("error flag is set, can't commit");
3249 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3254 if (txn->mt_parent) {
3255 MDB_txn *parent = txn->mt_parent;
3259 unsigned x, y, len, ps_len;
3261 /* Append our free list to parent's */
3262 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3265 mdb_midl_free(txn->mt_free_pgs);
3266 /* Failures after this must either undo the changes
3267 * to the parent or set MDB_TXN_ERROR in the parent.
3270 parent->mt_next_pgno = txn->mt_next_pgno;
3271 parent->mt_flags = txn->mt_flags;
3273 /* Merge our cursors into parent's and close them */
3274 mdb_cursors_close(txn, 1);
3276 /* Update parent's DB table. */
3277 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3278 parent->mt_numdbs = txn->mt_numdbs;
3279 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3280 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3281 for (i=2; i<txn->mt_numdbs; i++) {
3282 /* preserve parent's DB_NEW status */
3283 x = parent->mt_dbflags[i] & DB_NEW;
3284 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3287 dst = parent->mt_u.dirty_list;
3288 src = txn->mt_u.dirty_list;
3289 /* Remove anything in our dirty list from parent's spill list */
3290 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3292 pspill[0] = (pgno_t)-1;
3293 /* Mark our dirty pages as deleted in parent spill list */
3294 for (i=0, len=src[0].mid; ++i <= len; ) {
3295 MDB_ID pn = src[i].mid << 1;
3296 while (pn > pspill[x])
3298 if (pn == pspill[x]) {
3303 /* Squash deleted pagenums if we deleted any */
3304 for (x=y; ++x <= ps_len; )
3305 if (!(pspill[x] & 1))
3306 pspill[++y] = pspill[x];
3310 /* Find len = length of merging our dirty list with parent's */
3312 dst[0].mid = 0; /* simplify loops */
3313 if (parent->mt_parent) {
3314 len = x + src[0].mid;
3315 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3316 for (i = x; y && i; y--) {
3317 pgno_t yp = src[y].mid;
3318 while (yp < dst[i].mid)
3320 if (yp == dst[i].mid) {
3325 } else { /* Simplify the above for single-ancestor case */
3326 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3328 /* Merge our dirty list with parent's */
3330 for (i = len; y; dst[i--] = src[y--]) {
3331 pgno_t yp = src[y].mid;
3332 while (yp < dst[x].mid)
3333 dst[i--] = dst[x--];
3334 if (yp == dst[x].mid)
3335 free(dst[x--].mptr);
3337 mdb_tassert(txn, i == x);
3339 free(txn->mt_u.dirty_list);
3340 parent->mt_dirty_room = txn->mt_dirty_room;
3341 if (txn->mt_spill_pgs) {
3342 if (parent->mt_spill_pgs) {
3343 /* TODO: Prevent failure here, so parent does not fail */
3344 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3346 parent->mt_flags |= MDB_TXN_ERROR;
3347 mdb_midl_free(txn->mt_spill_pgs);
3348 mdb_midl_sort(parent->mt_spill_pgs);
3350 parent->mt_spill_pgs = txn->mt_spill_pgs;
3354 /* Append our loose page list to parent's */
3355 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3357 *lp = txn->mt_loose_pgs;
3358 parent->mt_loose_count += txn->mt_loose_count;
3360 parent->mt_child = NULL;
3361 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3366 if (txn != env->me_txn) {
3367 DPUTS("attempt to commit unknown transaction");
3372 mdb_cursors_close(txn, 0);
3374 if (!txn->mt_u.dirty_list[0].mid &&
3375 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3378 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3379 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3381 /* Update DB root pointers */
3382 if (txn->mt_numdbs > 2) {
3386 data.mv_size = sizeof(MDB_db);
3388 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3389 for (i = 2; i < txn->mt_numdbs; i++) {
3390 if (txn->mt_dbflags[i] & DB_DIRTY) {
3391 if (TXN_DBI_CHANGED(txn, i)) {
3395 data.mv_data = &txn->mt_dbs[i];
3396 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3403 rc = mdb_freelist_save(txn);
3407 mdb_midl_free(env->me_pghead);
3408 env->me_pghead = NULL;
3409 if (mdb_midl_shrink(&txn->mt_free_pgs))
3410 env->me_free_pgs = txn->mt_free_pgs;
3416 if ((rc = mdb_page_flush(txn, 0)) ||
3417 (rc = mdb_env_sync(env, 0)) ||
3418 (rc = mdb_env_write_meta(txn)))
3421 /* Free P_LOOSE pages left behind in dirty_list */
3422 if (!(env->me_flags & MDB_WRITEMAP))
3423 mdb_dlist_free(txn);
3428 mdb_dbis_update(txn, 1);
3431 UNLOCK_MUTEX(MDB_MUTEX(env, w));
3432 if (txn != env->me_txn0)
3442 /** Read the environment parameters of a DB environment before
3443 * mapping it into memory.
3444 * @param[in] env the environment handle
3445 * @param[out] meta address of where to store the meta information
3446 * @return 0 on success, non-zero on failure.
3449 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3455 enum { Size = sizeof(pbuf) };
3457 /* We don't know the page size yet, so use a minimum value.
3458 * Read both meta pages so we can use the latest one.
3461 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3465 memset(&ov, 0, sizeof(ov));
3467 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3468 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3471 rc = pread(env->me_fd, &pbuf, Size, off);
3474 if (rc == 0 && off == 0)
3476 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3477 DPRINTF(("read: %s", mdb_strerror(rc)));
3481 p = (MDB_page *)&pbuf;
3483 if (!F_ISSET(p->mp_flags, P_META)) {
3484 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3489 if (m->mm_magic != MDB_MAGIC) {
3490 DPUTS("meta has invalid magic");
3494 if (m->mm_version != MDB_DATA_VERSION) {
3495 DPRINTF(("database is version %u, expected version %u",
3496 m->mm_version, MDB_DATA_VERSION));
3497 return MDB_VERSION_MISMATCH;
3500 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3507 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3509 meta->mm_magic = MDB_MAGIC;
3510 meta->mm_version = MDB_DATA_VERSION;
3511 meta->mm_mapsize = env->me_mapsize;
3512 meta->mm_psize = env->me_psize;
3513 meta->mm_last_pg = 1;
3514 meta->mm_flags = env->me_flags & 0xffff;
3515 meta->mm_flags |= MDB_INTEGERKEY;
3516 meta->mm_dbs[0].md_root = P_INVALID;
3517 meta->mm_dbs[1].md_root = P_INVALID;
3520 /** Write the environment parameters of a freshly created DB environment.
3521 * @param[in] env the environment handle
3522 * @param[out] meta address of where to store the meta information
3523 * @return 0 on success, non-zero on failure.
3526 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3534 memset(&ov, 0, sizeof(ov));
3535 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3537 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3540 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3541 len = pwrite(fd, ptr, size, pos); \
3542 rc = (len >= 0); } while(0)
3545 DPUTS("writing new meta page");
3547 psize = env->me_psize;
3549 mdb_env_init_meta0(env, meta);
3551 p = calloc(2, psize);
3553 p->mp_flags = P_META;
3554 *(MDB_meta *)METADATA(p) = *meta;
3556 q = (MDB_page *)((char *)p + psize);
3558 q->mp_flags = P_META;
3559 *(MDB_meta *)METADATA(q) = *meta;
3561 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3564 else if ((unsigned) len == psize * 2)
3572 /** Update the environment info to commit a transaction.
3573 * @param[in] txn the transaction that's being committed
3574 * @return 0 on success, non-zero on failure.
3577 mdb_env_write_meta(MDB_txn *txn)
3580 MDB_meta meta, metab, *mp;
3583 int rc, len, toggle;
3592 toggle = txn->mt_txnid & 1;
3593 DPRINTF(("writing meta page %d for root page %"Z"u",
3594 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3597 mp = env->me_metas[toggle];
3598 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3599 /* Persist any increases of mapsize config */
3600 if (mapsize < env->me_mapsize)
3601 mapsize = env->me_mapsize;
3603 if (env->me_flags & MDB_WRITEMAP) {
3604 mp->mm_mapsize = mapsize;
3605 mp->mm_dbs[0] = txn->mt_dbs[0];
3606 mp->mm_dbs[1] = txn->mt_dbs[1];
3607 mp->mm_last_pg = txn->mt_next_pgno - 1;
3608 #if !(defined(_MSC_VER) || defined(__i386__) || defined(__x86_64__))
3609 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3610 __sync_synchronize();
3612 mp->mm_txnid = txn->mt_txnid;
3613 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3614 unsigned meta_size = env->me_psize;
3615 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3618 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3619 if (meta_size < env->me_os_psize)
3620 meta_size += meta_size;
3625 if (MDB_MSYNC(ptr, meta_size, rc)) {
3632 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3633 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3635 meta.mm_mapsize = mapsize;
3636 meta.mm_dbs[0] = txn->mt_dbs[0];
3637 meta.mm_dbs[1] = txn->mt_dbs[1];
3638 meta.mm_last_pg = txn->mt_next_pgno - 1;
3639 meta.mm_txnid = txn->mt_txnid;
3641 off = offsetof(MDB_meta, mm_mapsize);
3642 ptr = (char *)&meta + off;
3643 len = sizeof(MDB_meta) - off;
3645 off += env->me_psize;
3648 /* Write to the SYNC fd */
3649 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3650 env->me_fd : env->me_mfd;
3653 memset(&ov, 0, sizeof(ov));
3655 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3659 rc = pwrite(mfd, ptr, len, off);
3662 rc = rc < 0 ? ErrCode() : EIO;
3663 DPUTS("write failed, disk error?");
3664 /* On a failure, the pagecache still contains the new data.
3665 * Write some old data back, to prevent it from being used.
3666 * Use the non-SYNC fd; we know it will fail anyway.
3668 meta.mm_last_pg = metab.mm_last_pg;
3669 meta.mm_txnid = metab.mm_txnid;
3671 memset(&ov, 0, sizeof(ov));
3673 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3675 r2 = pwrite(env->me_fd, ptr, len, off);
3676 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3679 env->me_flags |= MDB_FATAL_ERROR;
3682 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3683 CACHEFLUSH(env->me_map + off, len, DCACHE);
3685 /* Memory ordering issues are irrelevant; since the entire writer
3686 * is wrapped by wmutex, all of these changes will become visible
3687 * after the wmutex is unlocked. Since the DB is multi-version,
3688 * readers will get consistent data regardless of how fresh or
3689 * how stale their view of these values is.
3692 env->me_txns->mti_txnid = txn->mt_txnid;
3697 /** Check both meta pages to see which one is newer.
3698 * @param[in] env the environment handle
3699 * @return meta toggle (0 or 1).
3702 mdb_env_pick_meta(const MDB_env *env)
3704 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3708 mdb_env_create(MDB_env **env)
3712 e = calloc(1, sizeof(MDB_env));
3716 e->me_maxreaders = DEFAULT_READERS;
3717 e->me_maxdbs = e->me_numdbs = 2;
3718 e->me_fd = INVALID_HANDLE_VALUE;
3719 e->me_lfd = INVALID_HANDLE_VALUE;
3720 e->me_mfd = INVALID_HANDLE_VALUE;
3721 #ifdef MDB_USE_SYSV_SEM
3722 e->me_rmutex.semid = -1;
3723 e->me_wmutex.semid = -1;
3725 e->me_pid = getpid();
3726 GET_PAGESIZE(e->me_os_psize);
3727 VGMEMP_CREATE(e,0,0);
3733 mdb_env_map(MDB_env *env, void *addr)
3736 unsigned int flags = env->me_flags;
3740 LONG sizelo, sizehi;
3743 if (flags & MDB_RDONLY) {
3744 /* Don't set explicit map size, use whatever exists */
3749 msize = env->me_mapsize;
3750 sizelo = msize & 0xffffffff;
3751 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3753 /* Windows won't create mappings for zero length files.
3754 * and won't map more than the file size.
3755 * Just set the maxsize right now.
3757 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3758 || !SetEndOfFile(env->me_fd)
3759 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3763 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3764 PAGE_READWRITE : PAGE_READONLY,
3765 sizehi, sizelo, NULL);
3768 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3769 FILE_MAP_WRITE : FILE_MAP_READ,
3771 rc = env->me_map ? 0 : ErrCode();
3776 int prot = PROT_READ;
3777 if (flags & MDB_WRITEMAP) {
3779 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3782 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3784 if (env->me_map == MAP_FAILED) {
3789 if (flags & MDB_NORDAHEAD) {
3790 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3792 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3794 #ifdef POSIX_MADV_RANDOM
3795 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3796 #endif /* POSIX_MADV_RANDOM */
3797 #endif /* MADV_RANDOM */
3801 /* Can happen because the address argument to mmap() is just a
3802 * hint. mmap() can pick another, e.g. if the range is in use.
3803 * The MAP_FIXED flag would prevent that, but then mmap could
3804 * instead unmap existing pages to make room for the new map.
3806 if (addr && env->me_map != addr)
3807 return EBUSY; /* TODO: Make a new MDB_* error code? */
3809 p = (MDB_page *)env->me_map;
3810 env->me_metas[0] = METADATA(p);
3811 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3817 mdb_env_set_mapsize(MDB_env *env, size_t size)
3819 /* If env is already open, caller is responsible for making
3820 * sure there are no active txns.
3828 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3829 else if (size < env->me_mapsize) {
3830 /* If the configured size is smaller, make sure it's
3831 * still big enough. Silently round up to minimum if not.
3833 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3837 munmap(env->me_map, env->me_mapsize);
3838 env->me_mapsize = size;
3839 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3840 rc = mdb_env_map(env, old);
3844 env->me_mapsize = size;
3846 env->me_maxpg = env->me_mapsize / env->me_psize;
3851 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3855 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3860 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3862 if (env->me_map || readers < 1)
3864 env->me_maxreaders = readers;
3869 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3871 if (!env || !readers)
3873 *readers = env->me_maxreaders;
3878 mdb_fsize(HANDLE fd, size_t *size)
3881 LARGE_INTEGER fsize;
3883 if (!GetFileSizeEx(fd, &fsize))
3886 *size = fsize.QuadPart;
3898 /** Further setup required for opening an LMDB environment
3901 mdb_env_open2(MDB_env *env)
3903 unsigned int flags = env->me_flags;
3904 int i, newenv = 0, rc;
3908 /* See if we should use QueryLimited */
3910 if ((rc & 0xff) > 5)
3911 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3913 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3916 memset(&meta, 0, sizeof(meta));
3918 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3921 DPUTS("new mdbenv");
3923 env->me_psize = env->me_os_psize;
3924 if (env->me_psize > MAX_PAGESIZE)
3925 env->me_psize = MAX_PAGESIZE;
3927 env->me_psize = meta.mm_psize;
3930 /* Was a mapsize configured? */
3931 if (!env->me_mapsize) {
3932 /* If this is a new environment, take the default,
3933 * else use the size recorded in the existing env.
3935 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3936 } else if (env->me_mapsize < meta.mm_mapsize) {
3937 /* If the configured size is smaller, make sure it's
3938 * still big enough. Silently round up to minimum if not.
3940 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3941 if (env->me_mapsize < minsize)
3942 env->me_mapsize = minsize;
3945 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3950 if (flags & MDB_FIXEDMAP)
3951 meta.mm_address = env->me_map;
3952 i = mdb_env_init_meta(env, &meta);
3953 if (i != MDB_SUCCESS) {
3958 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3959 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3961 #if !(MDB_MAXKEYSIZE)
3962 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3964 env->me_maxpg = env->me_mapsize / env->me_psize;
3968 int toggle = mdb_env_pick_meta(env);
3969 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3971 DPRINTF(("opened database version %u, pagesize %u",
3972 env->me_metas[0]->mm_version, env->me_psize));
3973 DPRINTF(("using meta page %d", toggle));
3974 DPRINTF(("depth: %u", db->md_depth));
3975 DPRINTF(("entries: %"Z"u", db->md_entries));
3976 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3977 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3978 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3979 DPRINTF(("root: %"Z"u", db->md_root));
3987 /** Release a reader thread's slot in the reader lock table.
3988 * This function is called automatically when a thread exits.
3989 * @param[in] ptr This points to the slot in the reader lock table.
3992 mdb_env_reader_dest(void *ptr)
3994 MDB_reader *reader = ptr;
4000 /** Junk for arranging thread-specific callbacks on Windows. This is
4001 * necessarily platform and compiler-specific. Windows supports up
4002 * to 1088 keys. Let's assume nobody opens more than 64 environments
4003 * in a single process, for now. They can override this if needed.
4005 #ifndef MAX_TLS_KEYS
4006 #define MAX_TLS_KEYS 64
4008 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4009 static int mdb_tls_nkeys;
4011 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4015 case DLL_PROCESS_ATTACH: break;
4016 case DLL_THREAD_ATTACH: break;
4017 case DLL_THREAD_DETACH:
4018 for (i=0; i<mdb_tls_nkeys; i++) {
4019 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4021 mdb_env_reader_dest(r);
4025 case DLL_PROCESS_DETACH: break;
4030 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4032 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4036 /* Force some symbol references.
4037 * _tls_used forces the linker to create the TLS directory if not already done
4038 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4040 #pragma comment(linker, "/INCLUDE:_tls_used")
4041 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4042 #pragma const_seg(".CRT$XLB")
4043 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4044 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4047 #pragma comment(linker, "/INCLUDE:__tls_used")
4048 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4049 #pragma data_seg(".CRT$XLB")
4050 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4052 #endif /* WIN 32/64 */
4053 #endif /* !__GNUC__ */
4056 /** Downgrade the exclusive lock on the region back to shared */
4058 mdb_env_share_locks(MDB_env *env, int *excl)
4060 int rc = 0, toggle = mdb_env_pick_meta(env);
4062 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4067 /* First acquire a shared lock. The Unlock will
4068 * then release the existing exclusive lock.
4070 memset(&ov, 0, sizeof(ov));
4071 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4074 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4080 struct flock lock_info;
4081 /* The shared lock replaces the existing lock */
4082 memset((void *)&lock_info, 0, sizeof(lock_info));
4083 lock_info.l_type = F_RDLCK;
4084 lock_info.l_whence = SEEK_SET;
4085 lock_info.l_start = 0;
4086 lock_info.l_len = 1;
4087 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4088 (rc = ErrCode()) == EINTR) ;
4089 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4096 /** Try to get exlusive lock, otherwise shared.
4097 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4100 mdb_env_excl_lock(MDB_env *env, int *excl)
4104 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4108 memset(&ov, 0, sizeof(ov));
4109 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4116 struct flock lock_info;
4117 memset((void *)&lock_info, 0, sizeof(lock_info));
4118 lock_info.l_type = F_WRLCK;
4119 lock_info.l_whence = SEEK_SET;
4120 lock_info.l_start = 0;
4121 lock_info.l_len = 1;
4122 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4123 (rc = ErrCode()) == EINTR) ;
4127 # ifdef MDB_USE_SYSV_SEM
4128 if (*excl < 0) /* always true when !MDB_USE_SYSV_SEM */
4131 lock_info.l_type = F_RDLCK;
4132 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4133 (rc = ErrCode()) == EINTR) ;
4143 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4145 * @(#) $Revision: 5.1 $
4146 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4147 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4149 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4153 * Please do not copyright this code. This code is in the public domain.
4155 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4156 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4157 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4158 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4159 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4160 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4161 * PERFORMANCE OF THIS SOFTWARE.
4164 * chongo <Landon Curt Noll> /\oo/\
4165 * http://www.isthe.com/chongo/
4167 * Share and Enjoy! :-)
4170 typedef unsigned long long mdb_hash_t;
4171 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4173 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4174 * @param[in] val value to hash
4175 * @param[in] hval initial value for hash
4176 * @return 64 bit hash
4178 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4179 * hval arg on the first call.
4182 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4184 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4185 unsigned char *end = s + val->mv_size;
4187 * FNV-1a hash each octet of the string
4190 /* xor the bottom with the current octet */
4191 hval ^= (mdb_hash_t)*s++;
4193 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4194 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4195 (hval << 7) + (hval << 8) + (hval << 40);
4197 /* return our new hash value */
4201 /** Hash the string and output the encoded hash.
4202 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4203 * very short name limits. We don't care about the encoding being reversible,
4204 * we just want to preserve as many bits of the input as possible in a
4205 * small printable string.
4206 * @param[in] str string to hash
4207 * @param[out] encbuf an array of 11 chars to hold the hash
4209 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4212 mdb_pack85(unsigned long l, char *out)
4216 for (i=0; i<5; i++) {
4217 *out++ = mdb_a85[l % 85];
4223 mdb_hash_enc(MDB_val *val, char *encbuf)
4225 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4227 mdb_pack85(h, encbuf);
4228 mdb_pack85(h>>32, encbuf+5);
4233 /** Open and/or initialize the lock region for the environment.
4234 * @param[in] env The LMDB environment.
4235 * @param[in] lpath The pathname of the file used for the lock region.
4236 * @param[in] mode The Unix permissions for the file, if we create it.
4237 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4238 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4239 * @return 0 on success, non-zero on failure.
4242 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4245 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4247 # define MDB_ERRCODE_ROFS EROFS
4248 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4249 # define MDB_CLOEXEC O_CLOEXEC
4252 # define MDB_CLOEXEC 0
4259 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4260 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4261 FILE_ATTRIBUTE_NORMAL, NULL);
4263 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4265 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4267 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4272 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4273 /* Lose record locks when exec*() */
4274 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4275 fcntl(env->me_lfd, F_SETFD, fdflags);
4278 if (!(env->me_flags & MDB_NOTLS)) {
4279 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4282 env->me_flags |= MDB_ENV_TXKEY;
4284 /* Windows TLS callbacks need help finding their TLS info. */
4285 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4289 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4293 /* Try to get exclusive lock. If we succeed, then
4294 * nobody is using the lock region and we should initialize it.
4296 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4299 size = GetFileSize(env->me_lfd, NULL);
4301 size = lseek(env->me_lfd, 0, SEEK_END);
4302 if (size == -1) goto fail_errno;
4304 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4305 if (size < rsize && *excl > 0) {
4307 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4308 || !SetEndOfFile(env->me_lfd))
4311 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4315 size = rsize - sizeof(MDB_txninfo);
4316 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4321 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4323 if (!mh) goto fail_errno;
4324 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4326 if (!env->me_txns) goto fail_errno;
4328 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4330 if (m == MAP_FAILED) goto fail_errno;
4336 BY_HANDLE_FILE_INFORMATION stbuf;
4345 if (!mdb_sec_inited) {
4346 InitializeSecurityDescriptor(&mdb_null_sd,
4347 SECURITY_DESCRIPTOR_REVISION);
4348 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4349 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4350 mdb_all_sa.bInheritHandle = FALSE;
4351 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4354 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4355 idbuf.volume = stbuf.dwVolumeSerialNumber;
4356 idbuf.nhigh = stbuf.nFileIndexHigh;
4357 idbuf.nlow = stbuf.nFileIndexLow;
4358 val.mv_data = &idbuf;
4359 val.mv_size = sizeof(idbuf);
4360 mdb_hash_enc(&val, encbuf);
4361 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4362 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4363 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4364 if (!env->me_rmutex) goto fail_errno;
4365 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4366 if (!env->me_wmutex) goto fail_errno;
4367 #elif defined(MDB_USE_SYSV_SEM)
4369 unsigned short vals[2] = {1, 1};
4370 int semid = semget(IPC_PRIVATE, 2, mode);
4374 env->me_rmutex.semid = semid;
4375 env->me_wmutex.semid = semid;
4376 env->me_rmutex.semnum = 0;
4377 env->me_wmutex.semnum = 1;
4380 if (semctl(semid, 0, SETALL, semu) < 0)
4382 env->me_txns->mti_semid = semid;
4383 #else /* MDB_USE_SYSV_SEM */
4384 pthread_mutexattr_t mattr;
4386 if ((rc = pthread_mutexattr_init(&mattr))
4387 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4388 #ifdef MDB_ROBUST_SUPPORTED
4389 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4391 || (rc = pthread_mutex_init(&env->me_txns->mti_rmutex, &mattr))
4392 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4394 pthread_mutexattr_destroy(&mattr);
4395 #endif /* _WIN32 || MDB_USE_SYSV_SEM */
4397 env->me_txns->mti_magic = MDB_MAGIC;
4398 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4399 env->me_txns->mti_txnid = 0;
4400 env->me_txns->mti_numreaders = 0;
4403 if (env->me_txns->mti_magic != MDB_MAGIC) {
4404 DPUTS("lock region has invalid magic");
4408 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4409 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4410 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4411 rc = MDB_VERSION_MISMATCH;
4415 if (rc && rc != EACCES && rc != EAGAIN) {
4419 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4420 if (!env->me_rmutex) goto fail_errno;
4421 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4422 if (!env->me_wmutex) goto fail_errno;
4423 #elif defined(MDB_USE_SYSV_SEM)
4424 struct semid_ds buf;
4426 int semid = env->me_txns->mti_semid;
4429 /* check for read access */
4430 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4432 /* check for write access */
4433 if (semctl(semid, 0, IPC_SET, semu) < 0)
4436 env->me_rmutex.semid = semid;
4437 env->me_wmutex.semid = semid;
4438 env->me_rmutex.semnum = 0;
4439 env->me_wmutex.semnum = 1;
4450 /** The name of the lock file in the DB environment */
4451 #define LOCKNAME "/lock.mdb"
4452 /** The name of the data file in the DB environment */
4453 #define DATANAME "/data.mdb"
4454 /** The suffix of the lock file when no subdir is used */
4455 #define LOCKSUFF "-lock"
4456 /** Only a subset of the @ref mdb_env flags can be changed
4457 * at runtime. Changing other flags requires closing the
4458 * environment and re-opening it with the new flags.
4460 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4461 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4462 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4464 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4465 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4469 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4471 int oflags, rc, len, excl = -1;
4472 char *lpath, *dpath;
4474 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4478 if (flags & MDB_NOSUBDIR) {
4479 rc = len + sizeof(LOCKSUFF) + len + 1;
4481 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4486 if (flags & MDB_NOSUBDIR) {
4487 dpath = lpath + len + sizeof(LOCKSUFF);
4488 sprintf(lpath, "%s" LOCKSUFF, path);
4489 strcpy(dpath, path);
4491 dpath = lpath + len + sizeof(LOCKNAME);
4492 sprintf(lpath, "%s" LOCKNAME, path);
4493 sprintf(dpath, "%s" DATANAME, path);
4497 flags |= env->me_flags;
4498 if (flags & MDB_RDONLY) {
4499 /* silently ignore WRITEMAP when we're only getting read access */
4500 flags &= ~MDB_WRITEMAP;
4502 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4503 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4506 env->me_flags = flags |= MDB_ENV_ACTIVE;
4510 env->me_path = strdup(path);
4511 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4512 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4513 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4514 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4519 /* For RDONLY, get lockfile after we know datafile exists */
4520 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4521 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4527 if (F_ISSET(flags, MDB_RDONLY)) {
4528 oflags = GENERIC_READ;
4529 len = OPEN_EXISTING;
4531 oflags = GENERIC_READ|GENERIC_WRITE;
4534 mode = FILE_ATTRIBUTE_NORMAL;
4535 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4536 NULL, len, mode, NULL);
4538 if (F_ISSET(flags, MDB_RDONLY))
4541 oflags = O_RDWR | O_CREAT;
4543 env->me_fd = open(dpath, oflags, mode);
4545 if (env->me_fd == INVALID_HANDLE_VALUE) {
4550 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4551 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4556 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4557 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4558 env->me_mfd = env->me_fd;
4560 /* Synchronous fd for meta writes. Needed even with
4561 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4564 len = OPEN_EXISTING;
4565 env->me_mfd = CreateFile(dpath, oflags,
4566 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4567 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4570 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4572 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4577 DPRINTF(("opened dbenv %p", (void *) env));
4579 rc = mdb_env_share_locks(env, &excl);
4583 if (!((flags & MDB_RDONLY) ||
4584 (env->me_pbuf = calloc(1, env->me_psize))))
4586 if (!(flags & MDB_RDONLY)) {
4588 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4589 (sizeof(MDB_db)+sizeof(MDB_cursor)+sizeof(unsigned int)+1);
4590 txn = calloc(1, size);
4592 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4593 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4594 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4595 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4597 txn->mt_dbxs = env->me_dbxs;
4607 mdb_env_close0(env, excl);
4613 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4615 mdb_env_close0(MDB_env *env, int excl)
4619 if (!(env->me_flags & MDB_ENV_ACTIVE))
4622 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4623 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4624 free(env->me_dbxs[i].md_name.mv_data);
4627 free(env->me_dbiseqs);
4628 free(env->me_dbflags);
4631 free(env->me_dirty_list);
4633 mdb_midl_free(env->me_free_pgs);
4635 if (env->me_flags & MDB_ENV_TXKEY) {
4636 pthread_key_delete(env->me_txkey);
4638 /* Delete our key from the global list */
4639 for (i=0; i<mdb_tls_nkeys; i++)
4640 if (mdb_tls_keys[i] == env->me_txkey) {
4641 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4649 munmap(env->me_map, env->me_mapsize);
4651 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4652 (void) close(env->me_mfd);
4653 if (env->me_fd != INVALID_HANDLE_VALUE)
4654 (void) close(env->me_fd);
4656 MDB_PID_T pid = env->me_pid;
4657 /* Clearing readers is done in this function because
4658 * me_txkey with its destructor must be disabled first.
4660 for (i = env->me_numreaders; --i >= 0; )
4661 if (env->me_txns->mti_readers[i].mr_pid == pid)
4662 env->me_txns->mti_readers[i].mr_pid = 0;
4664 if (env->me_rmutex) {
4665 CloseHandle(env->me_rmutex);
4666 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4668 /* Windows automatically destroys the mutexes when
4669 * the last handle closes.
4671 #elif defined(MDB_USE_SYSV_SEM)
4672 if (env->me_rmutex.semid != -1) {
4673 /* If we have the filelock: If we are the
4674 * only remaining user, clean up semaphores.
4677 mdb_env_excl_lock(env, &excl);
4679 semctl(env->me_rmutex.semid, 0, IPC_RMID);
4682 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4684 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4687 /* Unlock the lockfile. Windows would have unlocked it
4688 * after closing anyway, but not necessarily at once.
4690 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4693 (void) close(env->me_lfd);
4696 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4700 mdb_env_close(MDB_env *env)
4707 VGMEMP_DESTROY(env);
4708 while ((dp = env->me_dpages) != NULL) {
4709 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4710 env->me_dpages = dp->mp_next;
4714 mdb_env_close0(env, 0);
4718 /** Compare two items pointing at aligned size_t's */
4720 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4722 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4723 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4726 /** Compare two items pointing at aligned unsigned int's */
4728 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4730 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4731 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4734 /** Compare two items pointing at unsigned ints of unknown alignment.
4735 * Nodes and keys are guaranteed to be 2-byte aligned.
4738 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4740 #if BYTE_ORDER == LITTLE_ENDIAN
4741 unsigned short *u, *c;
4744 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4745 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4748 } while(!x && u > (unsigned short *)a->mv_data);
4751 unsigned short *u, *c, *end;
4754 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4755 u = (unsigned short *)a->mv_data;
4756 c = (unsigned short *)b->mv_data;
4759 } while(!x && u < end);
4764 /** Compare two items pointing at size_t's of unknown alignment. */
4765 #ifdef MISALIGNED_OK
4766 # define mdb_cmp_clong mdb_cmp_long
4768 # define mdb_cmp_clong mdb_cmp_cint
4771 /** Compare two items lexically */
4773 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4780 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4786 diff = memcmp(a->mv_data, b->mv_data, len);
4787 return diff ? diff : len_diff<0 ? -1 : len_diff;
4790 /** Compare two items in reverse byte order */
4792 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4794 const unsigned char *p1, *p2, *p1_lim;
4798 p1_lim = (const unsigned char *)a->mv_data;
4799 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4800 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4802 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4808 while (p1 > p1_lim) {
4809 diff = *--p1 - *--p2;
4813 return len_diff<0 ? -1 : len_diff;
4816 /** Search for key within a page, using binary search.
4817 * Returns the smallest entry larger or equal to the key.
4818 * If exactp is non-null, stores whether the found entry was an exact match
4819 * in *exactp (1 or 0).
4820 * Updates the cursor index with the index of the found entry.
4821 * If no entry larger or equal to the key is found, returns NULL.
4824 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4826 unsigned int i = 0, nkeys;
4829 MDB_page *mp = mc->mc_pg[mc->mc_top];
4830 MDB_node *node = NULL;
4835 nkeys = NUMKEYS(mp);
4837 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4838 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4841 low = IS_LEAF(mp) ? 0 : 1;
4843 cmp = mc->mc_dbx->md_cmp;
4845 /* Branch pages have no data, so if using integer keys,
4846 * alignment is guaranteed. Use faster mdb_cmp_int.
4848 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4849 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4856 nodekey.mv_size = mc->mc_db->md_pad;
4857 node = NODEPTR(mp, 0); /* fake */
4858 while (low <= high) {
4859 i = (low + high) >> 1;
4860 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4861 rc = cmp(key, &nodekey);
4862 DPRINTF(("found leaf index %u [%s], rc = %i",
4863 i, DKEY(&nodekey), rc));
4872 while (low <= high) {
4873 i = (low + high) >> 1;
4875 node = NODEPTR(mp, i);
4876 nodekey.mv_size = NODEKSZ(node);
4877 nodekey.mv_data = NODEKEY(node);
4879 rc = cmp(key, &nodekey);
4882 DPRINTF(("found leaf index %u [%s], rc = %i",
4883 i, DKEY(&nodekey), rc));
4885 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4886 i, DKEY(&nodekey), NODEPGNO(node), rc));
4897 if (rc > 0) { /* Found entry is less than the key. */
4898 i++; /* Skip to get the smallest entry larger than key. */
4900 node = NODEPTR(mp, i);
4903 *exactp = (rc == 0 && nkeys > 0);
4904 /* store the key index */
4905 mc->mc_ki[mc->mc_top] = i;
4907 /* There is no entry larger or equal to the key. */
4910 /* nodeptr is fake for LEAF2 */
4916 mdb_cursor_adjust(MDB_cursor *mc, func)
4920 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4921 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4928 /** Pop a page off the top of the cursor's stack. */
4930 mdb_cursor_pop(MDB_cursor *mc)
4934 MDB_page *top = mc->mc_pg[mc->mc_top];
4940 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4941 DDBI(mc), (void *) mc));
4945 /** Push a page onto the top of the cursor's stack. */
4947 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4949 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4950 DDBI(mc), (void *) mc));
4952 if (mc->mc_snum >= CURSOR_STACK) {
4953 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4954 return MDB_CURSOR_FULL;
4957 mc->mc_top = mc->mc_snum++;
4958 mc->mc_pg[mc->mc_top] = mp;
4959 mc->mc_ki[mc->mc_top] = 0;
4964 /** Find the address of the page corresponding to a given page number.
4965 * @param[in] txn the transaction for this access.
4966 * @param[in] pgno the page number for the page to retrieve.
4967 * @param[out] ret address of a pointer where the page's address will be stored.
4968 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4969 * @return 0 on success, non-zero on failure.
4972 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4974 MDB_env *env = txn->mt_env;
4978 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4982 MDB_ID2L dl = tx2->mt_u.dirty_list;
4984 /* Spilled pages were dirtied in this txn and flushed
4985 * because the dirty list got full. Bring this page
4986 * back in from the map (but don't unspill it here,
4987 * leave that unless page_touch happens again).
4989 if (tx2->mt_spill_pgs) {
4990 MDB_ID pn = pgno << 1;
4991 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4992 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4993 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4998 unsigned x = mdb_mid2l_search(dl, pgno);
4999 if (x <= dl[0].mid && dl[x].mid == pgno) {
5005 } while ((tx2 = tx2->mt_parent) != NULL);
5008 if (pgno < txn->mt_next_pgno) {
5010 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5012 DPRINTF(("page %"Z"u not found", pgno));
5013 txn->mt_flags |= MDB_TXN_ERROR;
5014 return MDB_PAGE_NOTFOUND;
5024 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5025 * The cursor is at the root page, set up the rest of it.
5028 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5030 MDB_page *mp = mc->mc_pg[mc->mc_top];
5034 while (IS_BRANCH(mp)) {
5038 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5039 mdb_cassert(mc, NUMKEYS(mp) > 1);
5040 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5042 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5044 if (flags & MDB_PS_LAST)
5045 i = NUMKEYS(mp) - 1;
5048 node = mdb_node_search(mc, key, &exact);
5050 i = NUMKEYS(mp) - 1;
5052 i = mc->mc_ki[mc->mc_top];
5054 mdb_cassert(mc, i > 0);
5058 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5061 mdb_cassert(mc, i < NUMKEYS(mp));
5062 node = NODEPTR(mp, i);
5064 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5067 mc->mc_ki[mc->mc_top] = i;
5068 if ((rc = mdb_cursor_push(mc, mp)))
5071 if (flags & MDB_PS_MODIFY) {
5072 if ((rc = mdb_page_touch(mc)) != 0)
5074 mp = mc->mc_pg[mc->mc_top];
5079 DPRINTF(("internal error, index points to a %02X page!?",
5081 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5082 return MDB_CORRUPTED;
5085 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5086 key ? DKEY(key) : "null"));
5087 mc->mc_flags |= C_INITIALIZED;
5088 mc->mc_flags &= ~C_EOF;
5093 /** Search for the lowest key under the current branch page.
5094 * This just bypasses a NUMKEYS check in the current page
5095 * before calling mdb_page_search_root(), because the callers
5096 * are all in situations where the current page is known to
5100 mdb_page_search_lowest(MDB_cursor *mc)
5102 MDB_page *mp = mc->mc_pg[mc->mc_top];
5103 MDB_node *node = NODEPTR(mp, 0);
5106 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5109 mc->mc_ki[mc->mc_top] = 0;
5110 if ((rc = mdb_cursor_push(mc, mp)))
5112 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5115 /** Search for the page a given key should be in.
5116 * Push it and its parent pages on the cursor stack.
5117 * @param[in,out] mc the cursor for this operation.
5118 * @param[in] key the key to search for, or NULL for first/last page.
5119 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5120 * are touched (updated with new page numbers).
5121 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5122 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5123 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5124 * @return 0 on success, non-zero on failure.
5127 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5132 /* Make sure the txn is still viable, then find the root from
5133 * the txn's db table and set it as the root of the cursor's stack.
5135 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5136 DPUTS("transaction has failed, must abort");
5139 /* Make sure we're using an up-to-date root */
5140 if (*mc->mc_dbflag & DB_STALE) {
5142 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5144 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5145 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5152 MDB_node *leaf = mdb_node_search(&mc2,
5153 &mc->mc_dbx->md_name, &exact);
5155 return MDB_NOTFOUND;
5156 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5159 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5161 /* The txn may not know this DBI, or another process may
5162 * have dropped and recreated the DB with other flags.
5164 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5165 return MDB_INCOMPATIBLE;
5166 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5168 *mc->mc_dbflag &= ~DB_STALE;
5170 root = mc->mc_db->md_root;
5172 if (root == P_INVALID) { /* Tree is empty. */
5173 DPUTS("tree is empty");
5174 return MDB_NOTFOUND;
5178 mdb_cassert(mc, root > 1);
5179 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5180 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5186 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5187 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5189 if (flags & MDB_PS_MODIFY) {
5190 if ((rc = mdb_page_touch(mc)))
5194 if (flags & MDB_PS_ROOTONLY)
5197 return mdb_page_search_root(mc, key, flags);
5201 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5203 MDB_txn *txn = mc->mc_txn;
5204 pgno_t pg = mp->mp_pgno;
5205 unsigned x = 0, ovpages = mp->mp_pages;
5206 MDB_env *env = txn->mt_env;
5207 MDB_IDL sl = txn->mt_spill_pgs;
5208 MDB_ID pn = pg << 1;
5211 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5212 /* If the page is dirty or on the spill list we just acquired it,
5213 * so we should give it back to our current free list, if any.
5214 * Otherwise put it onto the list of pages we freed in this txn.
5216 * Won't create me_pghead: me_pglast must be inited along with it.
5217 * Unsupported in nested txns: They would need to hide the page
5218 * range in ancestor txns' dirty and spilled lists.
5220 if (env->me_pghead &&
5222 ((mp->mp_flags & P_DIRTY) ||
5223 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5227 MDB_ID2 *dl, ix, iy;
5228 rc = mdb_midl_need(&env->me_pghead, ovpages);
5231 if (!(mp->mp_flags & P_DIRTY)) {
5232 /* This page is no longer spilled */
5239 /* Remove from dirty list */
5240 dl = txn->mt_u.dirty_list;
5242 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5248 mdb_cassert(mc, x > 1);
5250 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5251 txn->mt_flags |= MDB_TXN_ERROR;
5252 return MDB_CORRUPTED;
5255 if (!(env->me_flags & MDB_WRITEMAP))
5256 mdb_dpage_free(env, mp);
5258 /* Insert in me_pghead */
5259 mop = env->me_pghead;
5260 j = mop[0] + ovpages;
5261 for (i = mop[0]; i && mop[i] < pg; i--)
5267 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5271 mc->mc_db->md_overflow_pages -= ovpages;
5275 /** Return the data associated with a given node.
5276 * @param[in] txn The transaction for this operation.
5277 * @param[in] leaf The node being read.
5278 * @param[out] data Updated to point to the node's data.
5279 * @return 0 on success, non-zero on failure.
5282 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5284 MDB_page *omp; /* overflow page */
5288 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5289 data->mv_size = NODEDSZ(leaf);
5290 data->mv_data = NODEDATA(leaf);
5294 /* Read overflow data.
5296 data->mv_size = NODEDSZ(leaf);
5297 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5298 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5299 DPRINTF(("read overflow page %"Z"u failed", pgno));
5302 data->mv_data = METADATA(omp);
5308 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5309 MDB_val *key, MDB_val *data)
5316 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5318 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5321 if (txn->mt_flags & MDB_TXN_ERROR)
5324 mdb_cursor_init(&mc, txn, dbi, &mx);
5325 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5328 /** Find a sibling for a page.
5329 * Replaces the page at the top of the cursor's stack with the
5330 * specified sibling, if one exists.
5331 * @param[in] mc The cursor for this operation.
5332 * @param[in] move_right Non-zero if the right sibling is requested,
5333 * otherwise the left sibling.
5334 * @return 0 on success, non-zero on failure.
5337 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5343 if (mc->mc_snum < 2) {
5344 return MDB_NOTFOUND; /* root has no siblings */
5348 DPRINTF(("parent page is page %"Z"u, index %u",
5349 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5351 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5352 : (mc->mc_ki[mc->mc_top] == 0)) {
5353 DPRINTF(("no more keys left, moving to %s sibling",
5354 move_right ? "right" : "left"));
5355 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5356 /* undo cursor_pop before returning */
5363 mc->mc_ki[mc->mc_top]++;
5365 mc->mc_ki[mc->mc_top]--;
5366 DPRINTF(("just moving to %s index key %u",
5367 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5369 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5371 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5372 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5373 /* mc will be inconsistent if caller does mc_snum++ as above */
5374 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5378 mdb_cursor_push(mc, mp);
5380 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5385 /** Move the cursor to the next data item. */
5387 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5393 if (mc->mc_flags & C_EOF) {
5394 return MDB_NOTFOUND;
5397 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5399 mp = mc->mc_pg[mc->mc_top];
5401 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5402 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5403 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5404 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5405 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5406 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5407 if (rc == MDB_SUCCESS)
5408 MDB_GET_KEY(leaf, key);
5413 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5414 if (op == MDB_NEXT_DUP)
5415 return MDB_NOTFOUND;
5419 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5420 mdb_dbg_pgno(mp), (void *) mc));
5421 if (mc->mc_flags & C_DEL)
5424 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5425 DPUTS("=====> move to next sibling page");
5426 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5427 mc->mc_flags |= C_EOF;
5430 mp = mc->mc_pg[mc->mc_top];
5431 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5433 mc->mc_ki[mc->mc_top]++;
5436 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5437 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5440 key->mv_size = mc->mc_db->md_pad;
5441 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5445 mdb_cassert(mc, IS_LEAF(mp));
5446 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5448 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5449 mdb_xcursor_init1(mc, leaf);
5452 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5455 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5456 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5457 if (rc != MDB_SUCCESS)
5462 MDB_GET_KEY(leaf, key);
5466 /** Move the cursor to the previous data item. */
5468 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5474 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5476 mp = mc->mc_pg[mc->mc_top];
5478 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5479 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5480 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5481 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5482 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5483 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5484 if (rc == MDB_SUCCESS) {
5485 MDB_GET_KEY(leaf, key);
5486 mc->mc_flags &= ~C_EOF;
5492 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5493 if (op == MDB_PREV_DUP)
5494 return MDB_NOTFOUND;
5498 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5499 mdb_dbg_pgno(mp), (void *) mc));
5501 if (mc->mc_ki[mc->mc_top] == 0) {
5502 DPUTS("=====> move to prev sibling page");
5503 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5506 mp = mc->mc_pg[mc->mc_top];
5507 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5508 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5510 mc->mc_ki[mc->mc_top]--;
5512 mc->mc_flags &= ~C_EOF;
5514 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5515 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5518 key->mv_size = mc->mc_db->md_pad;
5519 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5523 mdb_cassert(mc, IS_LEAF(mp));
5524 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5526 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5527 mdb_xcursor_init1(mc, leaf);
5530 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5533 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5534 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5535 if (rc != MDB_SUCCESS)
5540 MDB_GET_KEY(leaf, key);
5544 /** Set the cursor on a specific data item. */
5546 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5547 MDB_cursor_op op, int *exactp)
5551 MDB_node *leaf = NULL;
5554 if (key->mv_size == 0)
5555 return MDB_BAD_VALSIZE;
5558 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5560 /* See if we're already on the right page */
5561 if (mc->mc_flags & C_INITIALIZED) {
5564 mp = mc->mc_pg[mc->mc_top];
5566 mc->mc_ki[mc->mc_top] = 0;
5567 return MDB_NOTFOUND;
5569 if (mp->mp_flags & P_LEAF2) {
5570 nodekey.mv_size = mc->mc_db->md_pad;
5571 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5573 leaf = NODEPTR(mp, 0);
5574 MDB_GET_KEY2(leaf, nodekey);
5576 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5578 /* Probably happens rarely, but first node on the page
5579 * was the one we wanted.
5581 mc->mc_ki[mc->mc_top] = 0;
5588 unsigned int nkeys = NUMKEYS(mp);
5590 if (mp->mp_flags & P_LEAF2) {
5591 nodekey.mv_data = LEAF2KEY(mp,
5592 nkeys-1, nodekey.mv_size);
5594 leaf = NODEPTR(mp, nkeys-1);
5595 MDB_GET_KEY2(leaf, nodekey);
5597 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5599 /* last node was the one we wanted */
5600 mc->mc_ki[mc->mc_top] = nkeys-1;
5606 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5607 /* This is definitely the right page, skip search_page */
5608 if (mp->mp_flags & P_LEAF2) {
5609 nodekey.mv_data = LEAF2KEY(mp,
5610 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5612 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5613 MDB_GET_KEY2(leaf, nodekey);
5615 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5617 /* current node was the one we wanted */
5627 /* If any parents have right-sibs, search.
5628 * Otherwise, there's nothing further.
5630 for (i=0; i<mc->mc_top; i++)
5632 NUMKEYS(mc->mc_pg[i])-1)
5634 if (i == mc->mc_top) {
5635 /* There are no other pages */
5636 mc->mc_ki[mc->mc_top] = nkeys;
5637 return MDB_NOTFOUND;
5641 /* There are no other pages */
5642 mc->mc_ki[mc->mc_top] = 0;
5643 if (op == MDB_SET_RANGE && !exactp) {
5647 return MDB_NOTFOUND;
5651 rc = mdb_page_search(mc, key, 0);
5652 if (rc != MDB_SUCCESS)
5655 mp = mc->mc_pg[mc->mc_top];
5656 mdb_cassert(mc, IS_LEAF(mp));
5659 leaf = mdb_node_search(mc, key, exactp);
5660 if (exactp != NULL && !*exactp) {
5661 /* MDB_SET specified and not an exact match. */
5662 return MDB_NOTFOUND;
5666 DPUTS("===> inexact leaf not found, goto sibling");
5667 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5668 return rc; /* no entries matched */
5669 mp = mc->mc_pg[mc->mc_top];
5670 mdb_cassert(mc, IS_LEAF(mp));
5671 leaf = NODEPTR(mp, 0);
5675 mc->mc_flags |= C_INITIALIZED;
5676 mc->mc_flags &= ~C_EOF;
5679 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5680 key->mv_size = mc->mc_db->md_pad;
5681 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5686 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5687 mdb_xcursor_init1(mc, leaf);
5690 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5691 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5692 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5695 if (op == MDB_GET_BOTH) {
5701 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5702 if (rc != MDB_SUCCESS)
5705 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5707 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5709 rc = mc->mc_dbx->md_dcmp(data, &d2);
5711 if (op == MDB_GET_BOTH || rc > 0)
5712 return MDB_NOTFOUND;
5719 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5720 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5725 /* The key already matches in all other cases */
5726 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5727 MDB_GET_KEY(leaf, key);
5728 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5733 /** Move the cursor to the first item in the database. */
5735 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5741 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5743 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5744 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5745 if (rc != MDB_SUCCESS)
5748 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5750 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5751 mc->mc_flags |= C_INITIALIZED;
5752 mc->mc_flags &= ~C_EOF;
5754 mc->mc_ki[mc->mc_top] = 0;
5756 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5757 key->mv_size = mc->mc_db->md_pad;
5758 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5763 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5764 mdb_xcursor_init1(mc, leaf);
5765 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5769 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5773 MDB_GET_KEY(leaf, key);
5777 /** Move the cursor to the last item in the database. */
5779 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5785 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5787 if (!(mc->mc_flags & C_EOF)) {
5789 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5790 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5791 if (rc != MDB_SUCCESS)
5794 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5797 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5798 mc->mc_flags |= C_INITIALIZED|C_EOF;
5799 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5801 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5802 key->mv_size = mc->mc_db->md_pad;
5803 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5808 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5809 mdb_xcursor_init1(mc, leaf);
5810 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5814 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5819 MDB_GET_KEY(leaf, key);
5824 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5829 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5834 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5838 case MDB_GET_CURRENT:
5839 if (!(mc->mc_flags & C_INITIALIZED)) {
5842 MDB_page *mp = mc->mc_pg[mc->mc_top];
5843 int nkeys = NUMKEYS(mp);
5844 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5845 mc->mc_ki[mc->mc_top] = nkeys;
5851 key->mv_size = mc->mc_db->md_pad;
5852 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5854 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5855 MDB_GET_KEY(leaf, key);
5857 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5858 if (mc->mc_flags & C_DEL)
5859 mdb_xcursor_init1(mc, leaf);
5860 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5862 rc = mdb_node_read(mc->mc_txn, leaf, data);
5869 case MDB_GET_BOTH_RANGE:
5874 if (mc->mc_xcursor == NULL) {
5875 rc = MDB_INCOMPATIBLE;
5885 rc = mdb_cursor_set(mc, key, data, op,
5886 op == MDB_SET_RANGE ? NULL : &exact);
5889 case MDB_GET_MULTIPLE:
5890 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5894 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5895 rc = MDB_INCOMPATIBLE;
5899 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5900 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5903 case MDB_NEXT_MULTIPLE:
5908 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5909 rc = MDB_INCOMPATIBLE;
5912 if (!(mc->mc_flags & C_INITIALIZED))
5913 rc = mdb_cursor_first(mc, key, data);
5915 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5916 if (rc == MDB_SUCCESS) {
5917 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5920 mx = &mc->mc_xcursor->mx_cursor;
5921 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5923 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5924 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5932 case MDB_NEXT_NODUP:
5933 if (!(mc->mc_flags & C_INITIALIZED))
5934 rc = mdb_cursor_first(mc, key, data);
5936 rc = mdb_cursor_next(mc, key, data, op);
5940 case MDB_PREV_NODUP:
5941 if (!(mc->mc_flags & C_INITIALIZED)) {
5942 rc = mdb_cursor_last(mc, key, data);
5945 mc->mc_flags |= C_INITIALIZED;
5946 mc->mc_ki[mc->mc_top]++;
5948 rc = mdb_cursor_prev(mc, key, data, op);
5951 rc = mdb_cursor_first(mc, key, data);
5954 mfunc = mdb_cursor_first;
5956 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5960 if (mc->mc_xcursor == NULL) {
5961 rc = MDB_INCOMPATIBLE;
5965 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5966 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5967 MDB_GET_KEY(leaf, key);
5968 rc = mdb_node_read(mc->mc_txn, leaf, data);
5972 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5976 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5979 rc = mdb_cursor_last(mc, key, data);
5982 mfunc = mdb_cursor_last;
5985 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5990 if (mc->mc_flags & C_DEL)
5991 mc->mc_flags ^= C_DEL;
5996 /** Touch all the pages in the cursor stack. Set mc_top.
5997 * Makes sure all the pages are writable, before attempting a write operation.
5998 * @param[in] mc The cursor to operate on.
6001 mdb_cursor_touch(MDB_cursor *mc)
6003 int rc = MDB_SUCCESS;
6005 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6008 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6010 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6011 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6014 *mc->mc_dbflag |= DB_DIRTY;
6019 rc = mdb_page_touch(mc);
6020 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6021 mc->mc_top = mc->mc_snum-1;
6026 /** Do not spill pages to disk if txn is getting full, may fail instead */
6027 #define MDB_NOSPILL 0x8000
6030 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6033 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
6035 MDB_node *leaf = NULL;
6038 MDB_val xdata, *rdata, dkey, olddata;
6040 int do_sub = 0, insert_key, insert_data;
6041 unsigned int mcount = 0, dcount = 0, nospill;
6044 unsigned int nflags;
6047 if (mc == NULL || key == NULL)
6050 env = mc->mc_txn->mt_env;
6052 /* Check this first so counter will always be zero on any
6055 if (flags & MDB_MULTIPLE) {
6056 dcount = data[1].mv_size;
6057 data[1].mv_size = 0;
6058 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6059 return MDB_INCOMPATIBLE;
6062 nospill = flags & MDB_NOSPILL;
6063 flags &= ~MDB_NOSPILL;
6065 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6066 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6068 if (key->mv_size-1 >= ENV_MAXKEY(env))
6069 return MDB_BAD_VALSIZE;
6071 #if SIZE_MAX > MAXDATASIZE
6072 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6073 return MDB_BAD_VALSIZE;
6075 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6076 return MDB_BAD_VALSIZE;
6079 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6080 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6084 if (flags == MDB_CURRENT) {
6085 if (!(mc->mc_flags & C_INITIALIZED))
6088 } else if (mc->mc_db->md_root == P_INVALID) {
6089 /* new database, cursor has nothing to point to */
6092 mc->mc_flags &= ~C_INITIALIZED;
6097 if (flags & MDB_APPEND) {
6099 rc = mdb_cursor_last(mc, &k2, &d2);
6101 rc = mc->mc_dbx->md_cmp(key, &k2);
6104 mc->mc_ki[mc->mc_top]++;
6106 /* new key is <= last key */
6111 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6113 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6114 DPRINTF(("duplicate key [%s]", DKEY(key)));
6116 return MDB_KEYEXIST;
6118 if (rc && rc != MDB_NOTFOUND)
6122 if (mc->mc_flags & C_DEL)
6123 mc->mc_flags ^= C_DEL;
6125 /* Cursor is positioned, check for room in the dirty list */
6127 if (flags & MDB_MULTIPLE) {
6129 xdata.mv_size = data->mv_size * dcount;
6133 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6137 if (rc == MDB_NO_ROOT) {
6139 /* new database, write a root leaf page */
6140 DPUTS("allocating new root leaf page");
6141 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6144 mdb_cursor_push(mc, np);
6145 mc->mc_db->md_root = np->mp_pgno;
6146 mc->mc_db->md_depth++;
6147 *mc->mc_dbflag |= DB_DIRTY;
6148 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6150 np->mp_flags |= P_LEAF2;
6151 mc->mc_flags |= C_INITIALIZED;
6153 /* make sure all cursor pages are writable */
6154 rc2 = mdb_cursor_touch(mc);
6159 insert_key = insert_data = rc;
6161 /* The key does not exist */
6162 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6163 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6164 LEAFSIZE(key, data) > env->me_nodemax)
6166 /* Too big for a node, insert in sub-DB. Set up an empty
6167 * "old sub-page" for prep_subDB to expand to a full page.
6169 fp_flags = P_LEAF|P_DIRTY;
6171 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6172 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6173 olddata.mv_size = PAGEHDRSZ;
6177 /* there's only a key anyway, so this is a no-op */
6178 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6180 unsigned int ksize = mc->mc_db->md_pad;
6181 if (key->mv_size != ksize)
6182 return MDB_BAD_VALSIZE;
6183 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6184 memcpy(ptr, key->mv_data, ksize);
6186 /* if overwriting slot 0 of leaf, need to
6187 * update branch key if there is a parent page
6189 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6190 unsigned short top = mc->mc_top;
6192 /* slot 0 is always an empty key, find real slot */
6193 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6195 if (mc->mc_ki[mc->mc_top])
6196 rc2 = mdb_update_key(mc, key);
6207 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6208 olddata.mv_size = NODEDSZ(leaf);
6209 olddata.mv_data = NODEDATA(leaf);
6212 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6213 /* Prepare (sub-)page/sub-DB to accept the new item,
6214 * if needed. fp: old sub-page or a header faking
6215 * it. mp: new (sub-)page. offset: growth in page
6216 * size. xdata: node data with new page or DB.
6218 unsigned i, offset = 0;
6219 mp = fp = xdata.mv_data = env->me_pbuf;
6220 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6222 /* Was a single item before, must convert now */
6223 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6224 /* Just overwrite the current item */
6225 if (flags == MDB_CURRENT)
6228 #if UINT_MAX < SIZE_MAX
6229 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6230 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6232 /* does data match? */
6233 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6234 if (flags & MDB_NODUPDATA)
6235 return MDB_KEYEXIST;
6240 /* Back up original data item */
6241 dkey.mv_size = olddata.mv_size;
6242 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6244 /* Make sub-page header for the dup items, with dummy body */
6245 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6246 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6247 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6248 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6249 fp->mp_flags |= P_LEAF2;
6250 fp->mp_pad = data->mv_size;
6251 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6253 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6254 (dkey.mv_size & 1) + (data->mv_size & 1);
6256 fp->mp_upper = xdata.mv_size - PAGEBASE;
6257 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6258 } else if (leaf->mn_flags & F_SUBDATA) {
6259 /* Data is on sub-DB, just store it */
6260 flags |= F_DUPDATA|F_SUBDATA;
6263 /* Data is on sub-page */
6264 fp = olddata.mv_data;
6267 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6268 offset = EVEN(NODESIZE + sizeof(indx_t) +
6272 offset = fp->mp_pad;
6273 if (SIZELEFT(fp) < offset) {
6274 offset *= 4; /* space for 4 more */
6277 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6279 fp->mp_flags |= P_DIRTY;
6280 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6281 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6285 xdata.mv_size = olddata.mv_size + offset;
6288 fp_flags = fp->mp_flags;
6289 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6290 /* Too big for a sub-page, convert to sub-DB */
6291 fp_flags &= ~P_SUBP;
6293 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6294 fp_flags |= P_LEAF2;
6295 dummy.md_pad = fp->mp_pad;
6296 dummy.md_flags = MDB_DUPFIXED;
6297 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6298 dummy.md_flags |= MDB_INTEGERKEY;
6304 dummy.md_branch_pages = 0;
6305 dummy.md_leaf_pages = 1;
6306 dummy.md_overflow_pages = 0;
6307 dummy.md_entries = NUMKEYS(fp);
6308 xdata.mv_size = sizeof(MDB_db);
6309 xdata.mv_data = &dummy;
6310 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6312 offset = env->me_psize - olddata.mv_size;
6313 flags |= F_DUPDATA|F_SUBDATA;
6314 dummy.md_root = mp->mp_pgno;
6317 mp->mp_flags = fp_flags | P_DIRTY;
6318 mp->mp_pad = fp->mp_pad;
6319 mp->mp_lower = fp->mp_lower;
6320 mp->mp_upper = fp->mp_upper + offset;
6321 if (fp_flags & P_LEAF2) {
6322 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6324 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6325 olddata.mv_size - fp->mp_upper - PAGEBASE);
6326 for (i=0; i<NUMKEYS(fp); i++)
6327 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6335 mdb_node_del(mc, 0);
6339 /* overflow page overwrites need special handling */
6340 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6343 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6345 memcpy(&pg, olddata.mv_data, sizeof(pg));
6346 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6348 ovpages = omp->mp_pages;
6350 /* Is the ov page large enough? */
6351 if (ovpages >= dpages) {
6352 if (!(omp->mp_flags & P_DIRTY) &&
6353 (level || (env->me_flags & MDB_WRITEMAP)))
6355 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6358 level = 0; /* dirty in this txn or clean */
6361 if (omp->mp_flags & P_DIRTY) {
6362 /* yes, overwrite it. Note in this case we don't
6363 * bother to try shrinking the page if the new data
6364 * is smaller than the overflow threshold.
6367 /* It is writable only in a parent txn */
6368 size_t sz = (size_t) env->me_psize * ovpages, off;
6369 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6375 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6376 mdb_cassert(mc, rc2 == 0);
6377 if (!(flags & MDB_RESERVE)) {
6378 /* Copy end of page, adjusting alignment so
6379 * compiler may copy words instead of bytes.
6381 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6382 memcpy((size_t *)((char *)np + off),
6383 (size_t *)((char *)omp + off), sz - off);
6386 memcpy(np, omp, sz); /* Copy beginning of page */
6389 SETDSZ(leaf, data->mv_size);
6390 if (F_ISSET(flags, MDB_RESERVE))
6391 data->mv_data = METADATA(omp);
6393 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6397 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6399 } else if (data->mv_size == olddata.mv_size) {
6400 /* same size, just replace it. Note that we could
6401 * also reuse this node if the new data is smaller,
6402 * but instead we opt to shrink the node in that case.
6404 if (F_ISSET(flags, MDB_RESERVE))
6405 data->mv_data = olddata.mv_data;
6406 else if (!(mc->mc_flags & C_SUB))
6407 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6409 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6414 mdb_node_del(mc, 0);
6420 nflags = flags & NODE_ADD_FLAGS;
6421 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6422 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6423 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6424 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6426 nflags |= MDB_SPLIT_REPLACE;
6427 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6429 /* There is room already in this leaf page. */
6430 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6431 if (rc == 0 && insert_key) {
6432 /* Adjust other cursors pointing to mp */
6433 MDB_cursor *m2, *m3;
6434 MDB_dbi dbi = mc->mc_dbi;
6435 unsigned i = mc->mc_top;
6436 MDB_page *mp = mc->mc_pg[i];
6438 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6439 if (mc->mc_flags & C_SUB)
6440 m3 = &m2->mc_xcursor->mx_cursor;
6443 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6444 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6451 if (rc == MDB_SUCCESS) {
6452 /* Now store the actual data in the child DB. Note that we're
6453 * storing the user data in the keys field, so there are strict
6454 * size limits on dupdata. The actual data fields of the child
6455 * DB are all zero size.
6463 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6464 if (flags & MDB_CURRENT) {
6465 xflags = MDB_CURRENT|MDB_NOSPILL;
6467 mdb_xcursor_init1(mc, leaf);
6468 xflags = (flags & MDB_NODUPDATA) ?
6469 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6471 /* converted, write the original data first */
6473 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6477 /* Adjust other cursors pointing to mp */
6479 unsigned i = mc->mc_top;
6480 MDB_page *mp = mc->mc_pg[i];
6482 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6483 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6484 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6485 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6486 mdb_xcursor_init1(m2, leaf);
6490 /* we've done our job */
6493 ecount = mc->mc_xcursor->mx_db.md_entries;
6494 if (flags & MDB_APPENDDUP)
6495 xflags |= MDB_APPEND;
6496 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6497 if (flags & F_SUBDATA) {
6498 void *db = NODEDATA(leaf);
6499 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6501 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6503 /* Increment count unless we just replaced an existing item. */
6505 mc->mc_db->md_entries++;
6507 /* Invalidate txn if we created an empty sub-DB */
6510 /* If we succeeded and the key didn't exist before,
6511 * make sure the cursor is marked valid.
6513 mc->mc_flags |= C_INITIALIZED;
6515 if (flags & MDB_MULTIPLE) {
6518 /* let caller know how many succeeded, if any */
6519 data[1].mv_size = mcount;
6520 if (mcount < dcount) {
6521 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6522 insert_key = insert_data = 0;
6529 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6532 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6537 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6543 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6544 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6546 if (!(mc->mc_flags & C_INITIALIZED))
6549 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6550 return MDB_NOTFOUND;
6552 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6555 rc = mdb_cursor_touch(mc);
6559 mp = mc->mc_pg[mc->mc_top];
6562 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6564 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6565 if (flags & MDB_NODUPDATA) {
6566 /* mdb_cursor_del0() will subtract the final entry */
6567 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6569 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6570 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6572 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6575 /* If sub-DB still has entries, we're done */
6576 if (mc->mc_xcursor->mx_db.md_entries) {
6577 if (leaf->mn_flags & F_SUBDATA) {
6578 /* update subDB info */
6579 void *db = NODEDATA(leaf);
6580 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6583 /* shrink fake page */
6584 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6585 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6586 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6587 /* fix other sub-DB cursors pointed at this fake page */
6588 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6589 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6590 if (m2->mc_pg[mc->mc_top] == mp &&
6591 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6592 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6595 mc->mc_db->md_entries--;
6596 mc->mc_flags |= C_DEL;
6599 /* otherwise fall thru and delete the sub-DB */
6602 if (leaf->mn_flags & F_SUBDATA) {
6603 /* add all the child DB's pages to the free list */
6604 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6610 /* add overflow pages to free list */
6611 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6615 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6616 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6617 (rc = mdb_ovpage_free(mc, omp)))
6622 return mdb_cursor_del0(mc);
6625 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6629 /** Allocate and initialize new pages for a database.
6630 * @param[in] mc a cursor on the database being added to.
6631 * @param[in] flags flags defining what type of page is being allocated.
6632 * @param[in] num the number of pages to allocate. This is usually 1,
6633 * unless allocating overflow pages for a large record.
6634 * @param[out] mp Address of a page, or NULL on failure.
6635 * @return 0 on success, non-zero on failure.
6638 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6643 if ((rc = mdb_page_alloc(mc, num, &np)))
6645 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6646 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6647 np->mp_flags = flags | P_DIRTY;
6648 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6649 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6652 mc->mc_db->md_branch_pages++;
6653 else if (IS_LEAF(np))
6654 mc->mc_db->md_leaf_pages++;
6655 else if (IS_OVERFLOW(np)) {
6656 mc->mc_db->md_overflow_pages += num;
6664 /** Calculate the size of a leaf node.
6665 * The size depends on the environment's page size; if a data item
6666 * is too large it will be put onto an overflow page and the node
6667 * size will only include the key and not the data. Sizes are always
6668 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6669 * of the #MDB_node headers.
6670 * @param[in] env The environment handle.
6671 * @param[in] key The key for the node.
6672 * @param[in] data The data for the node.
6673 * @return The number of bytes needed to store the node.
6676 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6680 sz = LEAFSIZE(key, data);
6681 if (sz > env->me_nodemax) {
6682 /* put on overflow page */
6683 sz -= data->mv_size - sizeof(pgno_t);
6686 return EVEN(sz + sizeof(indx_t));
6689 /** Calculate the size of a branch node.
6690 * The size should depend on the environment's page size but since
6691 * we currently don't support spilling large keys onto overflow
6692 * pages, it's simply the size of the #MDB_node header plus the
6693 * size of the key. Sizes are always rounded up to an even number
6694 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6695 * @param[in] env The environment handle.
6696 * @param[in] key The key for the node.
6697 * @return The number of bytes needed to store the node.
6700 mdb_branch_size(MDB_env *env, MDB_val *key)
6705 if (sz > env->me_nodemax) {
6706 /* put on overflow page */
6707 /* not implemented */
6708 /* sz -= key->size - sizeof(pgno_t); */
6711 return sz + sizeof(indx_t);
6714 /** Add a node to the page pointed to by the cursor.
6715 * @param[in] mc The cursor for this operation.
6716 * @param[in] indx The index on the page where the new node should be added.
6717 * @param[in] key The key for the new node.
6718 * @param[in] data The data for the new node, if any.
6719 * @param[in] pgno The page number, if adding a branch node.
6720 * @param[in] flags Flags for the node.
6721 * @return 0 on success, non-zero on failure. Possible errors are:
6723 * <li>ENOMEM - failed to allocate overflow pages for the node.
6724 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6725 * should never happen since all callers already calculate the
6726 * page's free space before calling this function.
6730 mdb_node_add(MDB_cursor *mc, indx_t indx,
6731 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6734 size_t node_size = NODESIZE;
6738 MDB_page *mp = mc->mc_pg[mc->mc_top];
6739 MDB_page *ofp = NULL; /* overflow page */
6742 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6744 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6745 IS_LEAF(mp) ? "leaf" : "branch",
6746 IS_SUBP(mp) ? "sub-" : "",
6747 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6748 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6751 /* Move higher keys up one slot. */
6752 int ksize = mc->mc_db->md_pad, dif;
6753 char *ptr = LEAF2KEY(mp, indx, ksize);
6754 dif = NUMKEYS(mp) - indx;
6756 memmove(ptr+ksize, ptr, dif*ksize);
6757 /* insert new key */
6758 memcpy(ptr, key->mv_data, ksize);
6760 /* Just using these for counting */
6761 mp->mp_lower += sizeof(indx_t);
6762 mp->mp_upper -= ksize - sizeof(indx_t);
6766 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6768 node_size += key->mv_size;
6770 mdb_cassert(mc, data);
6771 if (F_ISSET(flags, F_BIGDATA)) {
6772 /* Data already on overflow page. */
6773 node_size += sizeof(pgno_t);
6774 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6775 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6777 /* Put data on overflow page. */
6778 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6779 data->mv_size, node_size+data->mv_size));
6780 node_size = EVEN(node_size + sizeof(pgno_t));
6781 if ((ssize_t)node_size > room)
6783 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6785 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6789 node_size += data->mv_size;
6792 node_size = EVEN(node_size);
6793 if ((ssize_t)node_size > room)
6797 /* Move higher pointers up one slot. */
6798 for (i = NUMKEYS(mp); i > indx; i--)
6799 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6801 /* Adjust free space offsets. */
6802 ofs = mp->mp_upper - node_size;
6803 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6804 mp->mp_ptrs[indx] = ofs;
6806 mp->mp_lower += sizeof(indx_t);
6808 /* Write the node data. */
6809 node = NODEPTR(mp, indx);
6810 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6811 node->mn_flags = flags;
6813 SETDSZ(node,data->mv_size);
6818 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6821 mdb_cassert(mc, key);
6823 if (F_ISSET(flags, F_BIGDATA))
6824 memcpy(node->mn_data + key->mv_size, data->mv_data,
6826 else if (F_ISSET(flags, MDB_RESERVE))
6827 data->mv_data = node->mn_data + key->mv_size;
6829 memcpy(node->mn_data + key->mv_size, data->mv_data,
6832 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6834 if (F_ISSET(flags, MDB_RESERVE))
6835 data->mv_data = METADATA(ofp);
6837 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6844 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6845 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6846 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6847 DPRINTF(("node size = %"Z"u", node_size));
6848 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6849 return MDB_PAGE_FULL;
6852 /** Delete the specified node from a page.
6853 * @param[in] mc Cursor pointing to the node to delete.
6854 * @param[in] ksize The size of a node. Only used if the page is
6855 * part of a #MDB_DUPFIXED database.
6858 mdb_node_del(MDB_cursor *mc, int ksize)
6860 MDB_page *mp = mc->mc_pg[mc->mc_top];
6861 indx_t indx = mc->mc_ki[mc->mc_top];
6863 indx_t i, j, numkeys, ptr;
6867 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6868 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6869 numkeys = NUMKEYS(mp);
6870 mdb_cassert(mc, indx < numkeys);
6873 int x = numkeys - 1 - indx;
6874 base = LEAF2KEY(mp, indx, ksize);
6876 memmove(base, base + ksize, x * ksize);
6877 mp->mp_lower -= sizeof(indx_t);
6878 mp->mp_upper += ksize - sizeof(indx_t);
6882 node = NODEPTR(mp, indx);
6883 sz = NODESIZE + node->mn_ksize;
6885 if (F_ISSET(node->mn_flags, F_BIGDATA))
6886 sz += sizeof(pgno_t);
6888 sz += NODEDSZ(node);
6892 ptr = mp->mp_ptrs[indx];
6893 for (i = j = 0; i < numkeys; i++) {
6895 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6896 if (mp->mp_ptrs[i] < ptr)
6897 mp->mp_ptrs[j] += sz;
6902 base = (char *)mp + mp->mp_upper + PAGEBASE;
6903 memmove(base + sz, base, ptr - mp->mp_upper);
6905 mp->mp_lower -= sizeof(indx_t);
6909 /** Compact the main page after deleting a node on a subpage.
6910 * @param[in] mp The main page to operate on.
6911 * @param[in] indx The index of the subpage on the main page.
6914 mdb_node_shrink(MDB_page *mp, indx_t indx)
6920 indx_t i, numkeys, ptr;
6922 node = NODEPTR(mp, indx);
6923 sp = (MDB_page *)NODEDATA(node);
6924 delta = SIZELEFT(sp);
6925 xp = (MDB_page *)((char *)sp + delta);
6927 /* shift subpage upward */
6929 nsize = NUMKEYS(sp) * sp->mp_pad;
6931 return; /* do not make the node uneven-sized */
6932 memmove(METADATA(xp), METADATA(sp), nsize);
6935 numkeys = NUMKEYS(sp);
6936 for (i=numkeys-1; i>=0; i--)
6937 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6939 xp->mp_upper = sp->mp_lower;
6940 xp->mp_lower = sp->mp_lower;
6941 xp->mp_flags = sp->mp_flags;
6942 xp->mp_pad = sp->mp_pad;
6943 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6945 nsize = NODEDSZ(node) - delta;
6946 SETDSZ(node, nsize);
6948 /* shift lower nodes upward */
6949 ptr = mp->mp_ptrs[indx];
6950 numkeys = NUMKEYS(mp);
6951 for (i = 0; i < numkeys; i++) {
6952 if (mp->mp_ptrs[i] <= ptr)
6953 mp->mp_ptrs[i] += delta;
6956 base = (char *)mp + mp->mp_upper + PAGEBASE;
6957 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6958 mp->mp_upper += delta;
6961 /** Initial setup of a sorted-dups cursor.
6962 * Sorted duplicates are implemented as a sub-database for the given key.
6963 * The duplicate data items are actually keys of the sub-database.
6964 * Operations on the duplicate data items are performed using a sub-cursor
6965 * initialized when the sub-database is first accessed. This function does
6966 * the preliminary setup of the sub-cursor, filling in the fields that
6967 * depend only on the parent DB.
6968 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6971 mdb_xcursor_init0(MDB_cursor *mc)
6973 MDB_xcursor *mx = mc->mc_xcursor;
6975 mx->mx_cursor.mc_xcursor = NULL;
6976 mx->mx_cursor.mc_txn = mc->mc_txn;
6977 mx->mx_cursor.mc_db = &mx->mx_db;
6978 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6979 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6980 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6981 mx->mx_cursor.mc_snum = 0;
6982 mx->mx_cursor.mc_top = 0;
6983 mx->mx_cursor.mc_flags = C_SUB;
6984 mx->mx_dbx.md_name.mv_size = 0;
6985 mx->mx_dbx.md_name.mv_data = NULL;
6986 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6987 mx->mx_dbx.md_dcmp = NULL;
6988 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6991 /** Final setup of a sorted-dups cursor.
6992 * Sets up the fields that depend on the data from the main cursor.
6993 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6994 * @param[in] node The data containing the #MDB_db record for the
6995 * sorted-dup database.
6998 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7000 MDB_xcursor *mx = mc->mc_xcursor;
7002 if (node->mn_flags & F_SUBDATA) {
7003 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7004 mx->mx_cursor.mc_pg[0] = 0;
7005 mx->mx_cursor.mc_snum = 0;
7006 mx->mx_cursor.mc_top = 0;
7007 mx->mx_cursor.mc_flags = C_SUB;
7009 MDB_page *fp = NODEDATA(node);
7010 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7011 mx->mx_db.md_flags = 0;
7012 mx->mx_db.md_depth = 1;
7013 mx->mx_db.md_branch_pages = 0;
7014 mx->mx_db.md_leaf_pages = 1;
7015 mx->mx_db.md_overflow_pages = 0;
7016 mx->mx_db.md_entries = NUMKEYS(fp);
7017 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7018 mx->mx_cursor.mc_snum = 1;
7019 mx->mx_cursor.mc_top = 0;
7020 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7021 mx->mx_cursor.mc_pg[0] = fp;
7022 mx->mx_cursor.mc_ki[0] = 0;
7023 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7024 mx->mx_db.md_flags = MDB_DUPFIXED;
7025 mx->mx_db.md_pad = fp->mp_pad;
7026 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7027 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7030 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7031 mx->mx_db.md_root));
7032 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7033 #if UINT_MAX < SIZE_MAX
7034 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7035 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7039 /** Initialize a cursor for a given transaction and database. */
7041 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7044 mc->mc_backup = NULL;
7047 mc->mc_db = &txn->mt_dbs[dbi];
7048 mc->mc_dbx = &txn->mt_dbxs[dbi];
7049 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7054 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7055 mdb_tassert(txn, mx != NULL);
7056 mc->mc_xcursor = mx;
7057 mdb_xcursor_init0(mc);
7059 mc->mc_xcursor = NULL;
7061 if (*mc->mc_dbflag & DB_STALE) {
7062 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7067 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7070 size_t size = sizeof(MDB_cursor);
7072 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7075 if (txn->mt_flags & MDB_TXN_ERROR)
7078 /* Allow read access to the freelist */
7079 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7082 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7083 size += sizeof(MDB_xcursor);
7085 if ((mc = malloc(size)) != NULL) {
7086 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7087 if (txn->mt_cursors) {
7088 mc->mc_next = txn->mt_cursors[dbi];
7089 txn->mt_cursors[dbi] = mc;
7090 mc->mc_flags |= C_UNTRACK;
7102 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7104 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7107 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7110 if (txn->mt_flags & MDB_TXN_ERROR)
7113 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7117 /* Return the count of duplicate data items for the current key */
7119 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7123 if (mc == NULL || countp == NULL)
7126 if (mc->mc_xcursor == NULL)
7127 return MDB_INCOMPATIBLE;
7129 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7132 if (!(mc->mc_flags & C_INITIALIZED))
7135 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7136 return MDB_NOTFOUND;
7138 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7139 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7142 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7145 *countp = mc->mc_xcursor->mx_db.md_entries;
7151 mdb_cursor_close(MDB_cursor *mc)
7153 if (mc && !mc->mc_backup) {
7154 /* remove from txn, if tracked */
7155 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7156 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7157 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7159 *prev = mc->mc_next;
7166 mdb_cursor_txn(MDB_cursor *mc)
7168 if (!mc) return NULL;
7173 mdb_cursor_dbi(MDB_cursor *mc)
7178 /** Replace the key for a branch node with a new key.
7179 * @param[in] mc Cursor pointing to the node to operate on.
7180 * @param[in] key The new key to use.
7181 * @return 0 on success, non-zero on failure.
7184 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7190 int delta, ksize, oksize;
7191 indx_t ptr, i, numkeys, indx;
7194 indx = mc->mc_ki[mc->mc_top];
7195 mp = mc->mc_pg[mc->mc_top];
7196 node = NODEPTR(mp, indx);
7197 ptr = mp->mp_ptrs[indx];
7201 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7202 k2.mv_data = NODEKEY(node);
7203 k2.mv_size = node->mn_ksize;
7204 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7206 mdb_dkey(&k2, kbuf2),
7212 /* Sizes must be 2-byte aligned. */
7213 ksize = EVEN(key->mv_size);
7214 oksize = EVEN(node->mn_ksize);
7215 delta = ksize - oksize;
7217 /* Shift node contents if EVEN(key length) changed. */
7219 if (delta > 0 && SIZELEFT(mp) < delta) {
7221 /* not enough space left, do a delete and split */
7222 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7223 pgno = NODEPGNO(node);
7224 mdb_node_del(mc, 0);
7225 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7228 numkeys = NUMKEYS(mp);
7229 for (i = 0; i < numkeys; i++) {
7230 if (mp->mp_ptrs[i] <= ptr)
7231 mp->mp_ptrs[i] -= delta;
7234 base = (char *)mp + mp->mp_upper + PAGEBASE;
7235 len = ptr - mp->mp_upper + NODESIZE;
7236 memmove(base - delta, base, len);
7237 mp->mp_upper -= delta;
7239 node = NODEPTR(mp, indx);
7242 /* But even if no shift was needed, update ksize */
7243 if (node->mn_ksize != key->mv_size)
7244 node->mn_ksize = key->mv_size;
7247 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7253 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7255 /** Move a node from csrc to cdst.
7258 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7265 unsigned short flags;
7269 /* Mark src and dst as dirty. */
7270 if ((rc = mdb_page_touch(csrc)) ||
7271 (rc = mdb_page_touch(cdst)))
7274 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7275 key.mv_size = csrc->mc_db->md_pad;
7276 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7278 data.mv_data = NULL;
7282 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7283 mdb_cassert(csrc, !((size_t)srcnode & 1));
7284 srcpg = NODEPGNO(srcnode);
7285 flags = srcnode->mn_flags;
7286 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7287 unsigned int snum = csrc->mc_snum;
7289 /* must find the lowest key below src */
7290 rc = mdb_page_search_lowest(csrc);
7293 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7294 key.mv_size = csrc->mc_db->md_pad;
7295 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7297 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7298 key.mv_size = NODEKSZ(s2);
7299 key.mv_data = NODEKEY(s2);
7301 csrc->mc_snum = snum--;
7302 csrc->mc_top = snum;
7304 key.mv_size = NODEKSZ(srcnode);
7305 key.mv_data = NODEKEY(srcnode);
7307 data.mv_size = NODEDSZ(srcnode);
7308 data.mv_data = NODEDATA(srcnode);
7310 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7311 unsigned int snum = cdst->mc_snum;
7314 /* must find the lowest key below dst */
7315 mdb_cursor_copy(cdst, &mn);
7316 rc = mdb_page_search_lowest(&mn);
7319 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7320 bkey.mv_size = mn.mc_db->md_pad;
7321 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7323 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7324 bkey.mv_size = NODEKSZ(s2);
7325 bkey.mv_data = NODEKEY(s2);
7327 mn.mc_snum = snum--;
7330 rc = mdb_update_key(&mn, &bkey);
7335 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7336 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7337 csrc->mc_ki[csrc->mc_top],
7339 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7340 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7342 /* Add the node to the destination page.
7344 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7345 if (rc != MDB_SUCCESS)
7348 /* Delete the node from the source page.
7350 mdb_node_del(csrc, key.mv_size);
7353 /* Adjust other cursors pointing to mp */
7354 MDB_cursor *m2, *m3;
7355 MDB_dbi dbi = csrc->mc_dbi;
7356 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7358 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7359 if (csrc->mc_flags & C_SUB)
7360 m3 = &m2->mc_xcursor->mx_cursor;
7363 if (m3 == csrc) continue;
7364 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7365 csrc->mc_ki[csrc->mc_top]) {
7366 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7367 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7372 /* Update the parent separators.
7374 if (csrc->mc_ki[csrc->mc_top] == 0) {
7375 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7376 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7377 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7379 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7380 key.mv_size = NODEKSZ(srcnode);
7381 key.mv_data = NODEKEY(srcnode);
7383 DPRINTF(("update separator for source page %"Z"u to [%s]",
7384 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7385 mdb_cursor_copy(csrc, &mn);
7388 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7391 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7393 indx_t ix = csrc->mc_ki[csrc->mc_top];
7394 nullkey.mv_size = 0;
7395 csrc->mc_ki[csrc->mc_top] = 0;
7396 rc = mdb_update_key(csrc, &nullkey);
7397 csrc->mc_ki[csrc->mc_top] = ix;
7398 mdb_cassert(csrc, rc == MDB_SUCCESS);
7402 if (cdst->mc_ki[cdst->mc_top] == 0) {
7403 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7404 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7405 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7407 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7408 key.mv_size = NODEKSZ(srcnode);
7409 key.mv_data = NODEKEY(srcnode);
7411 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7412 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7413 mdb_cursor_copy(cdst, &mn);
7416 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7419 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7421 indx_t ix = cdst->mc_ki[cdst->mc_top];
7422 nullkey.mv_size = 0;
7423 cdst->mc_ki[cdst->mc_top] = 0;
7424 rc = mdb_update_key(cdst, &nullkey);
7425 cdst->mc_ki[cdst->mc_top] = ix;
7426 mdb_cassert(csrc, rc == MDB_SUCCESS);
7433 /** Merge one page into another.
7434 * The nodes from the page pointed to by \b csrc will
7435 * be copied to the page pointed to by \b cdst and then
7436 * the \b csrc page will be freed.
7437 * @param[in] csrc Cursor pointing to the source page.
7438 * @param[in] cdst Cursor pointing to the destination page.
7439 * @return 0 on success, non-zero on failure.
7442 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7444 MDB_page *psrc, *pdst;
7451 psrc = csrc->mc_pg[csrc->mc_top];
7452 pdst = cdst->mc_pg[cdst->mc_top];
7454 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7456 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7457 mdb_cassert(csrc, cdst->mc_snum > 1);
7459 /* Mark dst as dirty. */
7460 if ((rc = mdb_page_touch(cdst)))
7463 /* Move all nodes from src to dst.
7465 j = nkeys = NUMKEYS(pdst);
7466 if (IS_LEAF2(psrc)) {
7467 key.mv_size = csrc->mc_db->md_pad;
7468 key.mv_data = METADATA(psrc);
7469 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7470 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7471 if (rc != MDB_SUCCESS)
7473 key.mv_data = (char *)key.mv_data + key.mv_size;
7476 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7477 srcnode = NODEPTR(psrc, i);
7478 if (i == 0 && IS_BRANCH(psrc)) {
7481 mdb_cursor_copy(csrc, &mn);
7482 /* must find the lowest key below src */
7483 rc = mdb_page_search_lowest(&mn);
7486 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7487 key.mv_size = mn.mc_db->md_pad;
7488 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7490 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7491 key.mv_size = NODEKSZ(s2);
7492 key.mv_data = NODEKEY(s2);
7495 key.mv_size = srcnode->mn_ksize;
7496 key.mv_data = NODEKEY(srcnode);
7499 data.mv_size = NODEDSZ(srcnode);
7500 data.mv_data = NODEDATA(srcnode);
7501 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7502 if (rc != MDB_SUCCESS)
7507 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7508 pdst->mp_pgno, NUMKEYS(pdst),
7509 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7511 /* Unlink the src page from parent and add to free list.
7514 mdb_node_del(csrc, 0);
7515 if (csrc->mc_ki[csrc->mc_top] == 0) {
7517 rc = mdb_update_key(csrc, &key);
7525 psrc = csrc->mc_pg[csrc->mc_top];
7526 /* If not operating on FreeDB, allow this page to be reused
7527 * in this txn. Otherwise just add to free list.
7529 rc = mdb_page_loose(csrc, psrc);
7533 csrc->mc_db->md_leaf_pages--;
7535 csrc->mc_db->md_branch_pages--;
7537 /* Adjust other cursors pointing to mp */
7538 MDB_cursor *m2, *m3;
7539 MDB_dbi dbi = csrc->mc_dbi;
7541 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7542 if (csrc->mc_flags & C_SUB)
7543 m3 = &m2->mc_xcursor->mx_cursor;
7546 if (m3 == csrc) continue;
7547 if (m3->mc_snum < csrc->mc_snum) continue;
7548 if (m3->mc_pg[csrc->mc_top] == psrc) {
7549 m3->mc_pg[csrc->mc_top] = pdst;
7550 m3->mc_ki[csrc->mc_top] += nkeys;
7555 unsigned int snum = cdst->mc_snum;
7556 uint16_t depth = cdst->mc_db->md_depth;
7557 mdb_cursor_pop(cdst);
7558 rc = mdb_rebalance(cdst);
7559 /* Did the tree shrink? */
7560 if (depth > cdst->mc_db->md_depth)
7562 cdst->mc_snum = snum;
7563 cdst->mc_top = snum-1;
7568 /** Copy the contents of a cursor.
7569 * @param[in] csrc The cursor to copy from.
7570 * @param[out] cdst The cursor to copy to.
7573 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7577 cdst->mc_txn = csrc->mc_txn;
7578 cdst->mc_dbi = csrc->mc_dbi;
7579 cdst->mc_db = csrc->mc_db;
7580 cdst->mc_dbx = csrc->mc_dbx;
7581 cdst->mc_snum = csrc->mc_snum;
7582 cdst->mc_top = csrc->mc_top;
7583 cdst->mc_flags = csrc->mc_flags;
7585 for (i=0; i<csrc->mc_snum; i++) {
7586 cdst->mc_pg[i] = csrc->mc_pg[i];
7587 cdst->mc_ki[i] = csrc->mc_ki[i];
7591 /** Rebalance the tree after a delete operation.
7592 * @param[in] mc Cursor pointing to the page where rebalancing
7594 * @return 0 on success, non-zero on failure.
7597 mdb_rebalance(MDB_cursor *mc)
7601 unsigned int ptop, minkeys;
7605 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7606 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7607 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7608 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7609 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7611 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7612 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7613 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7614 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7618 if (mc->mc_snum < 2) {
7619 MDB_page *mp = mc->mc_pg[0];
7621 DPUTS("Can't rebalance a subpage, ignoring");
7624 if (NUMKEYS(mp) == 0) {
7625 DPUTS("tree is completely empty");
7626 mc->mc_db->md_root = P_INVALID;
7627 mc->mc_db->md_depth = 0;
7628 mc->mc_db->md_leaf_pages = 0;
7629 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7632 /* Adjust cursors pointing to mp */
7635 mc->mc_flags &= ~C_INITIALIZED;
7637 MDB_cursor *m2, *m3;
7638 MDB_dbi dbi = mc->mc_dbi;
7640 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7641 if (mc->mc_flags & C_SUB)
7642 m3 = &m2->mc_xcursor->mx_cursor;
7645 if (m3->mc_snum < mc->mc_snum) continue;
7646 if (m3->mc_pg[0] == mp) {
7649 m3->mc_flags &= ~C_INITIALIZED;
7653 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7655 DPUTS("collapsing root page!");
7656 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7659 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7660 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7663 mc->mc_db->md_depth--;
7664 mc->mc_db->md_branch_pages--;
7665 mc->mc_ki[0] = mc->mc_ki[1];
7666 for (i = 1; i<mc->mc_db->md_depth; i++) {
7667 mc->mc_pg[i] = mc->mc_pg[i+1];
7668 mc->mc_ki[i] = mc->mc_ki[i+1];
7671 /* Adjust other cursors pointing to mp */
7672 MDB_cursor *m2, *m3;
7673 MDB_dbi dbi = mc->mc_dbi;
7675 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7676 if (mc->mc_flags & C_SUB)
7677 m3 = &m2->mc_xcursor->mx_cursor;
7680 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7681 if (m3->mc_pg[0] == mp) {
7684 for (i=0; i<m3->mc_snum; i++) {
7685 m3->mc_pg[i] = m3->mc_pg[i+1];
7686 m3->mc_ki[i] = m3->mc_ki[i+1];
7692 DPUTS("root page doesn't need rebalancing");
7696 /* The parent (branch page) must have at least 2 pointers,
7697 * otherwise the tree is invalid.
7699 ptop = mc->mc_top-1;
7700 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7702 /* Leaf page fill factor is below the threshold.
7703 * Try to move keys from left or right neighbor, or
7704 * merge with a neighbor page.
7709 mdb_cursor_copy(mc, &mn);
7710 mn.mc_xcursor = NULL;
7712 oldki = mc->mc_ki[mc->mc_top];
7713 if (mc->mc_ki[ptop] == 0) {
7714 /* We're the leftmost leaf in our parent.
7716 DPUTS("reading right neighbor");
7718 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7719 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7722 mn.mc_ki[mn.mc_top] = 0;
7723 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7725 /* There is at least one neighbor to the left.
7727 DPUTS("reading left neighbor");
7729 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7730 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7733 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7734 mc->mc_ki[mc->mc_top] = 0;
7737 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7738 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7739 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7741 /* If the neighbor page is above threshold and has enough keys,
7742 * move one key from it. Otherwise we should try to merge them.
7743 * (A branch page must never have less than 2 keys.)
7745 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7746 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7747 rc = mdb_node_move(&mn, mc);
7748 if (mc->mc_ki[ptop]) {
7752 if (mc->mc_ki[ptop] == 0) {
7753 rc = mdb_page_merge(&mn, mc);
7755 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7756 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7757 rc = mdb_page_merge(mc, &mn);
7758 mdb_cursor_copy(&mn, mc);
7760 mc->mc_flags &= ~C_EOF;
7762 mc->mc_ki[mc->mc_top] = oldki;
7766 /** Complete a delete operation started by #mdb_cursor_del(). */
7768 mdb_cursor_del0(MDB_cursor *mc)
7775 ki = mc->mc_ki[mc->mc_top];
7776 mdb_node_del(mc, mc->mc_db->md_pad);
7777 mc->mc_db->md_entries--;
7778 rc = mdb_rebalance(mc);
7780 if (rc == MDB_SUCCESS) {
7781 MDB_cursor *m2, *m3;
7782 MDB_dbi dbi = mc->mc_dbi;
7784 mp = mc->mc_pg[mc->mc_top];
7785 nkeys = NUMKEYS(mp);
7787 /* if mc points past last node in page, find next sibling */
7788 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7789 rc = mdb_cursor_sibling(mc, 1);
7790 if (rc == MDB_NOTFOUND) {
7791 mc->mc_flags |= C_EOF;
7796 /* Adjust other cursors pointing to mp */
7797 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7798 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7799 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7801 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7803 if (m3->mc_pg[mc->mc_top] == mp) {
7804 if (m3->mc_ki[mc->mc_top] >= ki) {
7805 m3->mc_flags |= C_DEL;
7806 if (m3->mc_ki[mc->mc_top] > ki)
7807 m3->mc_ki[mc->mc_top]--;
7808 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7809 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7811 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7812 rc = mdb_cursor_sibling(m3, 1);
7813 if (rc == MDB_NOTFOUND) {
7814 m3->mc_flags |= C_EOF;
7820 mc->mc_flags |= C_DEL;
7824 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7829 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7830 MDB_val *key, MDB_val *data)
7832 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7835 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7836 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7838 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7839 /* must ignore any data */
7843 return mdb_del0(txn, dbi, key, data, 0);
7847 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7848 MDB_val *key, MDB_val *data, unsigned flags)
7853 MDB_val rdata, *xdata;
7857 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7859 mdb_cursor_init(&mc, txn, dbi, &mx);
7868 flags |= MDB_NODUPDATA;
7870 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7872 /* let mdb_page_split know about this cursor if needed:
7873 * delete will trigger a rebalance; if it needs to move
7874 * a node from one page to another, it will have to
7875 * update the parent's separator key(s). If the new sepkey
7876 * is larger than the current one, the parent page may
7877 * run out of space, triggering a split. We need this
7878 * cursor to be consistent until the end of the rebalance.
7880 mc.mc_flags |= C_UNTRACK;
7881 mc.mc_next = txn->mt_cursors[dbi];
7882 txn->mt_cursors[dbi] = &mc;
7883 rc = mdb_cursor_del(&mc, flags);
7884 txn->mt_cursors[dbi] = mc.mc_next;
7889 /** Split a page and insert a new node.
7890 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7891 * The cursor will be updated to point to the actual page and index where
7892 * the node got inserted after the split.
7893 * @param[in] newkey The key for the newly inserted node.
7894 * @param[in] newdata The data for the newly inserted node.
7895 * @param[in] newpgno The page number, if the new node is a branch node.
7896 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7897 * @return 0 on success, non-zero on failure.
7900 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7901 unsigned int nflags)
7904 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7907 int i, j, split_indx, nkeys, pmax;
7908 MDB_env *env = mc->mc_txn->mt_env;
7910 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7911 MDB_page *copy = NULL;
7912 MDB_page *mp, *rp, *pp;
7917 mp = mc->mc_pg[mc->mc_top];
7918 newindx = mc->mc_ki[mc->mc_top];
7919 nkeys = NUMKEYS(mp);
7921 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7922 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7923 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7925 /* Create a right sibling. */
7926 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7928 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7930 if (mc->mc_snum < 2) {
7931 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7933 /* shift current top to make room for new parent */
7934 mc->mc_pg[1] = mc->mc_pg[0];
7935 mc->mc_ki[1] = mc->mc_ki[0];
7938 mc->mc_db->md_root = pp->mp_pgno;
7939 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7940 mc->mc_db->md_depth++;
7943 /* Add left (implicit) pointer. */
7944 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7945 /* undo the pre-push */
7946 mc->mc_pg[0] = mc->mc_pg[1];
7947 mc->mc_ki[0] = mc->mc_ki[1];
7948 mc->mc_db->md_root = mp->mp_pgno;
7949 mc->mc_db->md_depth--;
7956 ptop = mc->mc_top-1;
7957 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7960 mc->mc_flags |= C_SPLITTING;
7961 mdb_cursor_copy(mc, &mn);
7962 mn.mc_pg[mn.mc_top] = rp;
7963 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7965 if (nflags & MDB_APPEND) {
7966 mn.mc_ki[mn.mc_top] = 0;
7968 split_indx = newindx;
7972 split_indx = (nkeys+1) / 2;
7977 unsigned int lsize, rsize, ksize;
7978 /* Move half of the keys to the right sibling */
7979 x = mc->mc_ki[mc->mc_top] - split_indx;
7980 ksize = mc->mc_db->md_pad;
7981 split = LEAF2KEY(mp, split_indx, ksize);
7982 rsize = (nkeys - split_indx) * ksize;
7983 lsize = (nkeys - split_indx) * sizeof(indx_t);
7984 mp->mp_lower -= lsize;
7985 rp->mp_lower += lsize;
7986 mp->mp_upper += rsize - lsize;
7987 rp->mp_upper -= rsize - lsize;
7988 sepkey.mv_size = ksize;
7989 if (newindx == split_indx) {
7990 sepkey.mv_data = newkey->mv_data;
7992 sepkey.mv_data = split;
7995 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7996 memcpy(rp->mp_ptrs, split, rsize);
7997 sepkey.mv_data = rp->mp_ptrs;
7998 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7999 memcpy(ins, newkey->mv_data, ksize);
8000 mp->mp_lower += sizeof(indx_t);
8001 mp->mp_upper -= ksize - sizeof(indx_t);
8004 memcpy(rp->mp_ptrs, split, x * ksize);
8005 ins = LEAF2KEY(rp, x, ksize);
8006 memcpy(ins, newkey->mv_data, ksize);
8007 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8008 rp->mp_lower += sizeof(indx_t);
8009 rp->mp_upper -= ksize - sizeof(indx_t);
8010 mc->mc_ki[mc->mc_top] = x;
8011 mc->mc_pg[mc->mc_top] = rp;
8014 int psize, nsize, k;
8015 /* Maximum free space in an empty page */
8016 pmax = env->me_psize - PAGEHDRSZ;
8018 nsize = mdb_leaf_size(env, newkey, newdata);
8020 nsize = mdb_branch_size(env, newkey);
8021 nsize = EVEN(nsize);
8023 /* grab a page to hold a temporary copy */
8024 copy = mdb_page_malloc(mc->mc_txn, 1);
8029 copy->mp_pgno = mp->mp_pgno;
8030 copy->mp_flags = mp->mp_flags;
8031 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8032 copy->mp_upper = env->me_psize - PAGEBASE;
8034 /* prepare to insert */
8035 for (i=0, j=0; i<nkeys; i++) {
8037 copy->mp_ptrs[j++] = 0;
8039 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8042 /* When items are relatively large the split point needs
8043 * to be checked, because being off-by-one will make the
8044 * difference between success or failure in mdb_node_add.
8046 * It's also relevant if a page happens to be laid out
8047 * such that one half of its nodes are all "small" and
8048 * the other half of its nodes are "large." If the new
8049 * item is also "large" and falls on the half with
8050 * "large" nodes, it also may not fit.
8052 * As a final tweak, if the new item goes on the last
8053 * spot on the page (and thus, onto the new page), bias
8054 * the split so the new page is emptier than the old page.
8055 * This yields better packing during sequential inserts.
8057 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8058 /* Find split point */
8060 if (newindx <= split_indx || newindx >= nkeys) {
8062 k = newindx >= nkeys ? nkeys : split_indx+2;
8067 for (; i!=k; i+=j) {
8072 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8073 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8075 if (F_ISSET(node->mn_flags, F_BIGDATA))
8076 psize += sizeof(pgno_t);
8078 psize += NODEDSZ(node);
8080 psize = EVEN(psize);
8082 if (psize > pmax || i == k-j) {
8083 split_indx = i + (j<0);
8088 if (split_indx == newindx) {
8089 sepkey.mv_size = newkey->mv_size;
8090 sepkey.mv_data = newkey->mv_data;
8092 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8093 sepkey.mv_size = node->mn_ksize;
8094 sepkey.mv_data = NODEKEY(node);
8099 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8101 /* Copy separator key to the parent.
8103 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8107 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8112 if (mn.mc_snum == mc->mc_snum) {
8113 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8114 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8115 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8116 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8121 /* Right page might now have changed parent.
8122 * Check if left page also changed parent.
8124 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8125 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8126 for (i=0; i<ptop; i++) {
8127 mc->mc_pg[i] = mn.mc_pg[i];
8128 mc->mc_ki[i] = mn.mc_ki[i];
8130 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8131 if (mn.mc_ki[ptop]) {
8132 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8134 /* find right page's left sibling */
8135 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8136 mdb_cursor_sibling(mc, 0);
8141 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8144 mc->mc_flags ^= C_SPLITTING;
8145 if (rc != MDB_SUCCESS) {
8148 if (nflags & MDB_APPEND) {
8149 mc->mc_pg[mc->mc_top] = rp;
8150 mc->mc_ki[mc->mc_top] = 0;
8151 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8154 for (i=0; i<mc->mc_top; i++)
8155 mc->mc_ki[i] = mn.mc_ki[i];
8156 } else if (!IS_LEAF2(mp)) {
8158 mc->mc_pg[mc->mc_top] = rp;
8163 rkey.mv_data = newkey->mv_data;
8164 rkey.mv_size = newkey->mv_size;
8170 /* Update index for the new key. */
8171 mc->mc_ki[mc->mc_top] = j;
8173 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8174 rkey.mv_data = NODEKEY(node);
8175 rkey.mv_size = node->mn_ksize;
8177 xdata.mv_data = NODEDATA(node);
8178 xdata.mv_size = NODEDSZ(node);
8181 pgno = NODEPGNO(node);
8182 flags = node->mn_flags;
8185 if (!IS_LEAF(mp) && j == 0) {
8186 /* First branch index doesn't need key data. */
8190 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8196 mc->mc_pg[mc->mc_top] = copy;
8201 } while (i != split_indx);
8203 nkeys = NUMKEYS(copy);
8204 for (i=0; i<nkeys; i++)
8205 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8206 mp->mp_lower = copy->mp_lower;
8207 mp->mp_upper = copy->mp_upper;
8208 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8209 env->me_psize - copy->mp_upper - PAGEBASE);
8211 /* reset back to original page */
8212 if (newindx < split_indx) {
8213 mc->mc_pg[mc->mc_top] = mp;
8214 if (nflags & MDB_RESERVE) {
8215 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8216 if (!(node->mn_flags & F_BIGDATA))
8217 newdata->mv_data = NODEDATA(node);
8220 mc->mc_pg[mc->mc_top] = rp;
8222 /* Make sure mc_ki is still valid.
8224 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8225 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8226 for (i=0; i<=ptop; i++) {
8227 mc->mc_pg[i] = mn.mc_pg[i];
8228 mc->mc_ki[i] = mn.mc_ki[i];
8235 /* Adjust other cursors pointing to mp */
8236 MDB_cursor *m2, *m3;
8237 MDB_dbi dbi = mc->mc_dbi;
8238 int fixup = NUMKEYS(mp);
8240 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8241 if (mc->mc_flags & C_SUB)
8242 m3 = &m2->mc_xcursor->mx_cursor;
8247 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8249 if (m3->mc_flags & C_SPLITTING)
8254 for (k=m3->mc_top; k>=0; k--) {
8255 m3->mc_ki[k+1] = m3->mc_ki[k];
8256 m3->mc_pg[k+1] = m3->mc_pg[k];
8258 if (m3->mc_ki[0] >= split_indx) {
8263 m3->mc_pg[0] = mc->mc_pg[0];
8267 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8268 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8269 m3->mc_ki[mc->mc_top]++;
8270 if (m3->mc_ki[mc->mc_top] >= fixup) {
8271 m3->mc_pg[mc->mc_top] = rp;
8272 m3->mc_ki[mc->mc_top] -= fixup;
8273 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8275 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8276 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8281 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8284 if (copy) /* tmp page */
8285 mdb_page_free(env, copy);
8287 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8292 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8293 MDB_val *key, MDB_val *data, unsigned int flags)
8298 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8301 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8304 mdb_cursor_init(&mc, txn, dbi, &mx);
8305 return mdb_cursor_put(&mc, key, data, flags);
8309 #define MDB_WBUF (1024*1024)
8312 /** State needed for a compacting copy. */
8313 typedef struct mdb_copy {
8314 pthread_mutex_t mc_mutex;
8315 pthread_cond_t mc_cond;
8322 pgno_t mc_next_pgno;
8325 volatile int mc_new;
8330 /** Dedicated writer thread for compacting copy. */
8331 static THREAD_RET ESECT
8332 mdb_env_copythr(void *arg)
8336 int toggle = 0, wsize, rc;
8339 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8342 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8345 pthread_mutex_lock(&my->mc_mutex);
8347 pthread_cond_signal(&my->mc_cond);
8350 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8351 if (my->mc_new < 0) {
8356 wsize = my->mc_wlen[toggle];
8357 ptr = my->mc_wbuf[toggle];
8360 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8364 } else if (len > 0) {
8378 /* If there's an overflow page tail, write it too */
8379 if (my->mc_olen[toggle]) {
8380 wsize = my->mc_olen[toggle];
8381 ptr = my->mc_over[toggle];
8382 my->mc_olen[toggle] = 0;
8385 my->mc_wlen[toggle] = 0;
8387 pthread_cond_signal(&my->mc_cond);
8389 pthread_cond_signal(&my->mc_cond);
8390 pthread_mutex_unlock(&my->mc_mutex);
8391 return (THREAD_RET)0;
8395 /** Tell the writer thread there's a buffer ready to write */
8397 mdb_env_cthr_toggle(mdb_copy *my, int st)
8399 int toggle = my->mc_toggle ^ 1;
8400 pthread_mutex_lock(&my->mc_mutex);
8401 if (my->mc_status) {
8402 pthread_mutex_unlock(&my->mc_mutex);
8403 return my->mc_status;
8405 while (my->mc_new == 1)
8406 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8408 my->mc_toggle = toggle;
8409 pthread_cond_signal(&my->mc_cond);
8410 pthread_mutex_unlock(&my->mc_mutex);
8414 /** Depth-first tree traversal for compacting copy. */
8416 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8419 MDB_txn *txn = my->mc_txn;
8421 MDB_page *mo, *mp, *leaf;
8426 /* Empty DB, nothing to do */
8427 if (*pg == P_INVALID)
8434 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8437 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8441 /* Make cursor pages writable */
8442 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8446 for (i=0; i<mc.mc_top; i++) {
8447 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8448 mc.mc_pg[i] = (MDB_page *)ptr;
8449 ptr += my->mc_env->me_psize;
8452 /* This is writable space for a leaf page. Usually not needed. */
8453 leaf = (MDB_page *)ptr;
8455 toggle = my->mc_toggle;
8456 while (mc.mc_snum > 0) {
8458 mp = mc.mc_pg[mc.mc_top];
8462 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8463 for (i=0; i<n; i++) {
8464 ni = NODEPTR(mp, i);
8465 if (ni->mn_flags & F_BIGDATA) {
8469 /* Need writable leaf */
8471 mc.mc_pg[mc.mc_top] = leaf;
8472 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8474 ni = NODEPTR(mp, i);
8477 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8478 rc = mdb_page_get(txn, pg, &omp, NULL);
8481 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8482 rc = mdb_env_cthr_toggle(my, 1);
8485 toggle = my->mc_toggle;
8487 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8488 memcpy(mo, omp, my->mc_env->me_psize);
8489 mo->mp_pgno = my->mc_next_pgno;
8490 my->mc_next_pgno += omp->mp_pages;
8491 my->mc_wlen[toggle] += my->mc_env->me_psize;
8492 if (omp->mp_pages > 1) {
8493 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8494 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8495 rc = mdb_env_cthr_toggle(my, 1);
8498 toggle = my->mc_toggle;
8500 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8501 } else if (ni->mn_flags & F_SUBDATA) {
8504 /* Need writable leaf */
8506 mc.mc_pg[mc.mc_top] = leaf;
8507 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8509 ni = NODEPTR(mp, i);
8512 memcpy(&db, NODEDATA(ni), sizeof(db));
8513 my->mc_toggle = toggle;
8514 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8517 toggle = my->mc_toggle;
8518 memcpy(NODEDATA(ni), &db, sizeof(db));
8523 mc.mc_ki[mc.mc_top]++;
8524 if (mc.mc_ki[mc.mc_top] < n) {
8527 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8529 rc = mdb_page_get(txn, pg, &mp, NULL);
8534 mc.mc_ki[mc.mc_top] = 0;
8535 if (IS_BRANCH(mp)) {
8536 /* Whenever we advance to a sibling branch page,
8537 * we must proceed all the way down to its first leaf.
8539 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8542 mc.mc_pg[mc.mc_top] = mp;
8546 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8547 rc = mdb_env_cthr_toggle(my, 1);
8550 toggle = my->mc_toggle;
8552 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8553 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8554 mo->mp_pgno = my->mc_next_pgno++;
8555 my->mc_wlen[toggle] += my->mc_env->me_psize;
8557 /* Update parent if there is one */
8558 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8559 SETPGNO(ni, mo->mp_pgno);
8560 mdb_cursor_pop(&mc);
8562 /* Otherwise we're done */
8572 /** Copy environment with compaction. */
8574 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8579 MDB_txn *txn = NULL;
8584 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8585 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8586 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8587 if (my.mc_wbuf[0] == NULL)
8590 pthread_mutex_init(&my.mc_mutex, NULL);
8591 pthread_cond_init(&my.mc_cond, NULL);
8592 #ifdef HAVE_MEMALIGN
8593 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8594 if (my.mc_wbuf[0] == NULL)
8597 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8602 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8603 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8608 my.mc_next_pgno = 2;
8614 THREAD_CREATE(thr, mdb_env_copythr, &my);
8616 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8620 mp = (MDB_page *)my.mc_wbuf[0];
8621 memset(mp, 0, 2*env->me_psize);
8623 mp->mp_flags = P_META;
8624 mm = (MDB_meta *)METADATA(mp);
8625 mdb_env_init_meta0(env, mm);
8626 mm->mm_address = env->me_metas[0]->mm_address;
8628 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8630 mp->mp_flags = P_META;
8631 *(MDB_meta *)METADATA(mp) = *mm;
8632 mm = (MDB_meta *)METADATA(mp);
8634 /* Count the number of free pages, subtract from lastpg to find
8635 * number of active pages
8638 MDB_ID freecount = 0;
8641 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8642 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8643 freecount += *(MDB_ID *)data.mv_data;
8644 freecount += txn->mt_dbs[0].md_branch_pages +
8645 txn->mt_dbs[0].md_leaf_pages +
8646 txn->mt_dbs[0].md_overflow_pages;
8648 /* Set metapage 1 */
8649 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8650 mm->mm_dbs[1] = txn->mt_dbs[1];
8651 if (mm->mm_last_pg > 1) {
8652 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8655 mm->mm_dbs[1].md_root = P_INVALID;
8658 my.mc_wlen[0] = env->me_psize * 2;
8660 pthread_mutex_lock(&my.mc_mutex);
8662 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8663 pthread_mutex_unlock(&my.mc_mutex);
8664 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8665 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8666 rc = mdb_env_cthr_toggle(&my, 1);
8667 mdb_env_cthr_toggle(&my, -1);
8668 pthread_mutex_lock(&my.mc_mutex);
8670 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8671 pthread_mutex_unlock(&my.mc_mutex);
8676 CloseHandle(my.mc_cond);
8677 CloseHandle(my.mc_mutex);
8678 _aligned_free(my.mc_wbuf[0]);
8680 pthread_cond_destroy(&my.mc_cond);
8681 pthread_mutex_destroy(&my.mc_mutex);
8682 free(my.mc_wbuf[0]);
8687 /** Copy environment as-is. */
8689 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8691 MDB_txn *txn = NULL;
8692 mdb_mutex_t *wmutex = NULL;
8698 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8702 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8705 /* Do the lock/unlock of the reader mutex before starting the
8706 * write txn. Otherwise other read txns could block writers.
8708 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8713 /* We must start the actual read txn after blocking writers */
8714 mdb_txn_reset0(txn, "reset-stage1");
8716 /* Temporarily block writers until we snapshot the meta pages */
8717 wmutex = MDB_MUTEX(env, w);
8718 if (LOCK_MUTEX(rc, env, wmutex))
8721 rc = mdb_txn_renew0(txn);
8723 UNLOCK_MUTEX(wmutex);
8728 wsize = env->me_psize * 2;
8732 DO_WRITE(rc, fd, ptr, w2, len);
8736 } else if (len > 0) {
8742 /* Non-blocking or async handles are not supported */
8748 UNLOCK_MUTEX(wmutex);
8753 w2 = txn->mt_next_pgno * env->me_psize;
8756 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8763 if (wsize > MAX_WRITE)
8767 DO_WRITE(rc, fd, ptr, w2, len);
8771 } else if (len > 0) {
8788 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8790 if (flags & MDB_CP_COMPACT)
8791 return mdb_env_copyfd1(env, fd);
8793 return mdb_env_copyfd0(env, fd);
8797 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8799 return mdb_env_copyfd2(env, fd, 0);
8803 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8807 HANDLE newfd = INVALID_HANDLE_VALUE;
8809 if (env->me_flags & MDB_NOSUBDIR) {
8810 lpath = (char *)path;
8813 len += sizeof(DATANAME);
8814 lpath = malloc(len);
8817 sprintf(lpath, "%s" DATANAME, path);
8820 /* The destination path must exist, but the destination file must not.
8821 * We don't want the OS to cache the writes, since the source data is
8822 * already in the OS cache.
8825 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8826 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8828 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8830 if (newfd == INVALID_HANDLE_VALUE) {
8835 if (env->me_psize >= env->me_os_psize) {
8837 /* Set O_DIRECT if the file system supports it */
8838 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8839 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8841 #ifdef F_NOCACHE /* __APPLE__ */
8842 rc = fcntl(newfd, F_NOCACHE, 1);
8850 rc = mdb_env_copyfd2(env, newfd, flags);
8853 if (!(env->me_flags & MDB_NOSUBDIR))
8855 if (newfd != INVALID_HANDLE_VALUE)
8856 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8863 mdb_env_copy(MDB_env *env, const char *path)
8865 return mdb_env_copy2(env, path, 0);
8869 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8871 if (flag & (env->me_map ? ~CHANGEABLE : ~(CHANGEABLE|CHANGELESS)))
8874 env->me_flags |= flag;
8876 env->me_flags &= ~flag;
8881 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8886 *arg = env->me_flags;
8891 mdb_env_set_userctx(MDB_env *env, void *ctx)
8895 env->me_userctx = ctx;
8900 mdb_env_get_userctx(MDB_env *env)
8902 return env ? env->me_userctx : NULL;
8906 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8911 env->me_assert_func = func;
8917 mdb_env_get_path(MDB_env *env, const char **arg)
8922 *arg = env->me_path;
8927 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8936 /** Common code for #mdb_stat() and #mdb_env_stat().
8937 * @param[in] env the environment to operate in.
8938 * @param[in] db the #MDB_db record containing the stats to return.
8939 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8940 * @return 0, this function always succeeds.
8943 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8945 arg->ms_psize = env->me_psize;
8946 arg->ms_depth = db->md_depth;
8947 arg->ms_branch_pages = db->md_branch_pages;
8948 arg->ms_leaf_pages = db->md_leaf_pages;
8949 arg->ms_overflow_pages = db->md_overflow_pages;
8950 arg->ms_entries = db->md_entries;
8956 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8960 if (env == NULL || arg == NULL)
8963 toggle = mdb_env_pick_meta(env);
8965 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8969 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8973 if (env == NULL || arg == NULL)
8976 toggle = mdb_env_pick_meta(env);
8977 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8978 arg->me_mapsize = env->me_mapsize;
8979 arg->me_maxreaders = env->me_maxreaders;
8981 /* me_numreaders may be zero if this process never used any readers. Use
8982 * the shared numreader count if it exists.
8984 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8986 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8987 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8991 /** Set the default comparison functions for a database.
8992 * Called immediately after a database is opened to set the defaults.
8993 * The user can then override them with #mdb_set_compare() or
8994 * #mdb_set_dupsort().
8995 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8996 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8999 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9001 uint16_t f = txn->mt_dbs[dbi].md_flags;
9003 txn->mt_dbxs[dbi].md_cmp =
9004 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9005 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9007 txn->mt_dbxs[dbi].md_dcmp =
9008 !(f & MDB_DUPSORT) ? 0 :
9009 ((f & MDB_INTEGERDUP)
9010 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9011 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9014 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9020 int rc, dbflag, exact;
9021 unsigned int unused = 0, seq;
9024 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9025 mdb_default_cmp(txn, FREE_DBI);
9028 if ((flags & VALID_FLAGS) != flags)
9030 if (txn->mt_flags & MDB_TXN_ERROR)
9036 if (flags & PERSISTENT_FLAGS) {
9037 uint16_t f2 = flags & PERSISTENT_FLAGS;
9038 /* make sure flag changes get committed */
9039 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9040 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9041 txn->mt_flags |= MDB_TXN_DIRTY;
9044 mdb_default_cmp(txn, MAIN_DBI);
9048 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9049 mdb_default_cmp(txn, MAIN_DBI);
9052 /* Is the DB already open? */
9054 for (i=2; i<txn->mt_numdbs; i++) {
9055 if (!txn->mt_dbxs[i].md_name.mv_size) {
9056 /* Remember this free slot */
9057 if (!unused) unused = i;
9060 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9061 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9067 /* If no free slot and max hit, fail */
9068 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9069 return MDB_DBS_FULL;
9071 /* Cannot mix named databases with some mainDB flags */
9072 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9073 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9075 /* Find the DB info */
9076 dbflag = DB_NEW|DB_VALID;
9079 key.mv_data = (void *)name;
9080 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9081 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9082 if (rc == MDB_SUCCESS) {
9083 /* make sure this is actually a DB */
9084 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9085 if (!(node->mn_flags & F_SUBDATA))
9086 return MDB_INCOMPATIBLE;
9087 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9088 /* Create if requested */
9089 data.mv_size = sizeof(MDB_db);
9090 data.mv_data = &dummy;
9091 memset(&dummy, 0, sizeof(dummy));
9092 dummy.md_root = P_INVALID;
9093 dummy.md_flags = flags & PERSISTENT_FLAGS;
9094 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9098 /* OK, got info, add to table */
9099 if (rc == MDB_SUCCESS) {
9100 unsigned int slot = unused ? unused : txn->mt_numdbs;
9101 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9102 txn->mt_dbxs[slot].md_name.mv_size = len;
9103 txn->mt_dbxs[slot].md_rel = NULL;
9104 txn->mt_dbflags[slot] = dbflag;
9105 /* txn-> and env-> are the same in read txns, use
9106 * tmp variable to avoid undefined assignment
9108 seq = ++txn->mt_env->me_dbiseqs[slot];
9109 txn->mt_dbiseqs[slot] = seq;
9111 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9113 mdb_default_cmp(txn, slot);
9122 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9124 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9127 if (txn->mt_flags & MDB_TXN_ERROR)
9130 if (txn->mt_dbflags[dbi] & DB_STALE) {
9133 /* Stale, must read the DB's root. cursor_init does it for us. */
9134 mdb_cursor_init(&mc, txn, dbi, &mx);
9136 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9139 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9142 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9144 ptr = env->me_dbxs[dbi].md_name.mv_data;
9145 /* If there was no name, this was already closed */
9147 env->me_dbxs[dbi].md_name.mv_data = NULL;
9148 env->me_dbxs[dbi].md_name.mv_size = 0;
9149 env->me_dbflags[dbi] = 0;
9150 env->me_dbiseqs[dbi]++;
9155 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9157 /* We could return the flags for the FREE_DBI too but what's the point? */
9158 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9160 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9164 /** Add all the DB's pages to the free list.
9165 * @param[in] mc Cursor on the DB to free.
9166 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9167 * @return 0 on success, non-zero on failure.
9170 mdb_drop0(MDB_cursor *mc, int subs)
9174 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9175 if (rc == MDB_SUCCESS) {
9176 MDB_txn *txn = mc->mc_txn;
9181 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9182 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9185 mdb_cursor_copy(mc, &mx);
9186 while (mc->mc_snum > 0) {
9187 MDB_page *mp = mc->mc_pg[mc->mc_top];
9188 unsigned n = NUMKEYS(mp);
9190 for (i=0; i<n; i++) {
9191 ni = NODEPTR(mp, i);
9192 if (ni->mn_flags & F_BIGDATA) {
9195 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9196 rc = mdb_page_get(txn, pg, &omp, NULL);
9199 mdb_cassert(mc, IS_OVERFLOW(omp));
9200 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9204 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9205 mdb_xcursor_init1(mc, ni);
9206 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9212 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9214 for (i=0; i<n; i++) {
9216 ni = NODEPTR(mp, i);
9219 mdb_midl_xappend(txn->mt_free_pgs, pg);
9224 mc->mc_ki[mc->mc_top] = i;
9225 rc = mdb_cursor_sibling(mc, 1);
9227 if (rc != MDB_NOTFOUND)
9229 /* no more siblings, go back to beginning
9230 * of previous level.
9234 for (i=1; i<mc->mc_snum; i++) {
9236 mc->mc_pg[i] = mx.mc_pg[i];
9241 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9244 txn->mt_flags |= MDB_TXN_ERROR;
9245 } else if (rc == MDB_NOTFOUND) {
9251 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9253 MDB_cursor *mc, *m2;
9256 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9259 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9262 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9265 rc = mdb_cursor_open(txn, dbi, &mc);
9269 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9270 /* Invalidate the dropped DB's cursors */
9271 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9272 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9276 /* Can't delete the main DB */
9277 if (del && dbi > MAIN_DBI) {
9278 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9280 txn->mt_dbflags[dbi] = DB_STALE;
9281 mdb_dbi_close(txn->mt_env, dbi);
9283 txn->mt_flags |= MDB_TXN_ERROR;
9286 /* reset the DB record, mark it dirty */
9287 txn->mt_dbflags[dbi] |= DB_DIRTY;
9288 txn->mt_dbs[dbi].md_depth = 0;
9289 txn->mt_dbs[dbi].md_branch_pages = 0;
9290 txn->mt_dbs[dbi].md_leaf_pages = 0;
9291 txn->mt_dbs[dbi].md_overflow_pages = 0;
9292 txn->mt_dbs[dbi].md_entries = 0;
9293 txn->mt_dbs[dbi].md_root = P_INVALID;
9295 txn->mt_flags |= MDB_TXN_DIRTY;
9298 mdb_cursor_close(mc);
9302 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9304 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9307 txn->mt_dbxs[dbi].md_cmp = cmp;
9311 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9313 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9316 txn->mt_dbxs[dbi].md_dcmp = cmp;
9320 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9322 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9325 txn->mt_dbxs[dbi].md_rel = rel;
9329 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9331 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9334 txn->mt_dbxs[dbi].md_relctx = ctx;
9339 mdb_env_get_maxkeysize(MDB_env *env)
9341 return ENV_MAXKEY(env);
9345 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9347 unsigned int i, rdrs;
9350 int rc = 0, first = 1;
9354 if (!env->me_txns) {
9355 return func("(no reader locks)\n", ctx);
9357 rdrs = env->me_txns->mti_numreaders;
9358 mr = env->me_txns->mti_readers;
9359 for (i=0; i<rdrs; i++) {
9361 txnid_t txnid = mr[i].mr_txnid;
9362 sprintf(buf, txnid == (txnid_t)-1 ?
9363 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9364 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9367 rc = func(" pid thread txnid\n", ctx);
9371 rc = func(buf, ctx);
9377 rc = func("(no active readers)\n", ctx);
9382 /** Insert pid into list if not already present.
9383 * return -1 if already present.
9386 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9388 /* binary search of pid in list */
9390 unsigned cursor = 1;
9392 unsigned n = ids[0];
9395 unsigned pivot = n >> 1;
9396 cursor = base + pivot + 1;
9397 val = pid - ids[cursor];
9402 } else if ( val > 0 ) {
9407 /* found, so it's a duplicate */
9416 for (n = ids[0]; n > cursor; n--)
9423 mdb_reader_check(MDB_env *env, int *dead)
9429 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9432 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9433 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9435 mdb_mutex_t *rmutex = rlocked ? NULL : MDB_MUTEX(env, r);
9436 unsigned int i, j, rdrs;
9438 MDB_PID_T *pids, pid;
9439 int rc = MDB_SUCCESS, count = 0;
9441 rdrs = env->me_txns->mti_numreaders;
9442 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9446 mr = env->me_txns->mti_readers;
9447 for (i=0; i<rdrs; i++) {
9449 if (pid && pid != env->me_pid) {
9450 if (mdb_pid_insert(pids, pid) == 0) {
9451 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9452 /* Stale reader found */
9455 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9456 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9458 rdrs = 0; /* the above checked all readers */
9460 /* Recheck, a new process may have reused pid */
9461 if (mdb_reader_pid(env, Pidcheck, pid))
9466 if (mr[j].mr_pid == pid) {
9467 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9468 (unsigned) pid, mr[j].mr_txnid));
9473 UNLOCK_MUTEX(rmutex);
9484 #ifdef MDB_ROBUST_SUPPORTED
9485 /** Handle #LOCK_MUTEX0() failure.
9486 * With #MDB_ROBUST, try to repair the lock file if the mutex owner died.
9487 * @param[in] env the environment handle
9488 * @param[in] mutex LOCK_MUTEX0() mutex
9489 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9490 * @return 0 on success with the mutex locked, or an error code on failure.
9492 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc)
9496 enum { WAIT_ABANDONED = EOWNERDEAD };
9499 if (rc == (int) WAIT_ABANDONED) {
9500 /* We own the mutex. Clean up after dead previous owner. */
9502 rlocked = (mutex == MDB_MUTEX(env, r));
9504 /* env is hosed if the dead thread was ours */
9506 env->me_flags |= MDB_FATAL_ERROR;
9511 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9512 (rc ? "this process' env is hosed" : "recovering")));
9513 rc2 = mdb_reader_check0(env, rlocked, NULL);
9515 rc2 = pthread_mutex_consistent(mutex);
9516 if (rc || (rc = rc2)) {
9517 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9518 UNLOCK_MUTEX(mutex);
9524 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9529 #endif /* MDB_ROBUST_SUPPORTED */